diff --git a/CHANGELOG.md b/CHANGELOG.md
index 326678d19..b8aad6ef9 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,6 +9,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Breaking
 
+- aimux's public JSON equals the AI SDK's JSON. Struct fields serialize
+  camelCase (`toolCallId`, `providerOptions`, `inputTokens`); unions carry the
+  AI SDK's tag and strings (`{"type":"text-delta"}`, `"tool-call"`,
+  `"reasoning-file"`, Source `sourceType`, Warning/FileData/ResponseFormat/
+  stream parts no longer externally tagged); byte payloads are untagged
+  (array of numbers or base64 string); optional fields are omitted instead of
+  `null` and generate as `field?: T`; `AiMuxError` is tagged by `name`
+  (`AI_APICallError`, `AI_RetryError`, ...). Vendor request and response types
+  and the OpenAI chat-completion output keep their vendor wire formats.
+  Persisted recordings and trace records change with the same rule.
+
 - Middleware wraps generation and streaming with callbacks for both operations,
   parameter transforms and provider/model/supported-URL overrides. Explicit
   model and provider ids take precedence over middleware overrides.
diff --git a/aimux-core/src/content.rs b/aimux-core/src/content.rs
index 25681a97c..398108ade 100644
--- a/aimux-core/src/content.rs
+++ b/aimux-core/src/content.rs
@@ -4,19 +4,25 @@ use serde::{Deserialize, Serialize};
 use serde_json::Value;
 use ts_rs::TS;
 
+use crate::language_model_message::ToolResultOutput;
 use crate::shared::SharedProviderOptions;
 
 /// A part of a multi-part message.
 ///
 /// Shared between `ModelMessage` (user-facing) and `LanguageModelPrompt` (provider-facing).
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
-#[serde(tag = "type", rename_all = "snake_case")]
+#[serde(
+    tag = "type",
+    rename_all = "kebab-case",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum ContentPart {
     /// A text segment.
     Text {
         text: String,
         /// Provider-specific options for this part (e.g. `openai.promptCacheBreakpoint`).
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -24,6 +30,7 @@ pub enum ContentPart {
     /// Provider-specific custom content.
     Custom {
         kind: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -32,6 +39,7 @@ pub enum ContentPart {
     ReasoningFile {
         data: crate::shared::GeneratedFileData,
         media_type: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -40,12 +48,16 @@ pub enum ContentPart {
     ToolApprovalRequest {
         approval_id: String,
         tool_call_id: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         reason: Option<String>,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         is_automatic: Option<bool>,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         signature: Option<String>,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         input_schema_input: Option<Value>,
     },
@@ -54,6 +66,7 @@ pub enum ContentPart {
     Image {
         image: Vec<u8>,
         media_type: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -62,8 +75,10 @@ pub enum ContentPart {
     File {
         data: Vec<u8>,
         media_type: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         filename: Option<String>,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -77,8 +92,10 @@ pub enum ContentPart {
     FileBase64 {
         data: String,
         media_type: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         filename: Option<String>,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -89,6 +106,7 @@ pub enum ContentPart {
     FileUrl {
         url: String,
         media_type: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -101,8 +119,10 @@ pub enum ContentPart {
     FileReference {
         media_type: String,
         reference: Value,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         filename: Option<String>,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -116,10 +136,13 @@ pub enum ContentPart {
     /// `redacted_thinking` block is emitted).
     Reasoning {
         text: String,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         signature: Option<String>,
         /// Provider-specific options for this part (e.g.
         /// `anthropic.signature`, `anthropic.redactedData`,
         /// `anthropic.cacheControl`).
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -133,41 +156,24 @@ pub enum ContentPart {
         /// Whether the tool call is executed by the provider rather than by
         /// the client. This is part of the prompt contract because providers
         /// need it to replay server tool calls on a later turn.
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_executed: Option<bool>,
         /// Provider-specific options for this part (e.g.
         /// `anthropic.cacheControl`).
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
 
-    /// The result of executing a tool call.
+    /// The result of executing a tool call (upstream `ToolResultPart`).
     ToolResult {
         tool_call_id: String,
-        /// The tool's output (usually a JSON value or plain text).
-        ///
-        /// Accepts the legacy `output` field name (used by the Vercel AI SDK
-        /// and the 0.1.0 TypeScript bindings) as an alias during deserialization,
-        /// so multi-part tool messages constructed either way round-trip
-        /// correctly. Serialization always emits `result`.
-        #[serde(alias = "output")]
-        result: Value,
-        /// The name of the tool that produced this result (optional on the
-        /// user-input side; providers that need it can look it up from the
-        /// preceding tool call).
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        tool_name: Option<String>,
-        /// Whether the result is an error or error message.
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        is_error: Option<bool>,
-        /// Whether the result is preliminary (replaces prior, e.g. image previews).
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        preliminary: Option<bool>,
-        /// Whether the tool is dynamic (defined at runtime, e.g. MCP tools).
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        dynamic: Option<bool>,
+        tool_name: String,
+        output: ToolResultOutput,
         /// Provider-specific options for this part (e.g.
         /// `anthropic.cacheControl`).
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -197,19 +203,6 @@ impl ContentPart {
         }
     }
 
-    /// Convenience constructor for a tool-result part (no provider options).
-    pub fn tool_result(tool_call_id: impl Into<String>, result: Value) -> Self {
-        ContentPart::ToolResult {
-            tool_call_id: tool_call_id.into(),
-            result,
-            tool_name: None,
-            is_error: None,
-            preliminary: None,
-            dynamic: None,
-            provider_options: None,
-        }
-    }
-
     /// Convenience constructor for a reasoning/thinking part (no provider
     /// options, no signature).
     pub fn reasoning(text: impl Into<String>) -> Self {
diff --git a/aimux-core/src/embedding_model.rs b/aimux-core/src/embedding_model.rs
index e97ff1dab..44d74a405 100644
--- a/aimux-core/src/embedding_model.rs
+++ b/aimux-core/src/embedding_model.rs
@@ -25,6 +25,7 @@ pub type Embedding = Vec<f32>;
 ///
 /// Aligned with V4 `EmbeddingModelV4CallOptions`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct EmbeddingCallOptions {
     /// List of text values to generate embeddings for.
@@ -36,15 +37,23 @@ pub struct EmbeddingCallOptions {
     pub abort_signal: Option<AbortSignal>,
 
     /// Per-call retry override. `None` uses the model default.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_retries: Option<u32>,
 
     /// Per-call operation timeout.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timeout: Option<crate::options::TimeoutConfiguration>,
 
     /// Additional provider-specific options, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 
     /// Additional HTTP headers to send with the request.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
 }
 
@@ -64,6 +73,7 @@ impl EmbeddingCallOptions {
 
 /// Token usage for an embedding call. Embeddings only report input tokens.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct EmbeddingUsage {
     /// Number of input tokens consumed.
@@ -74,18 +84,25 @@ pub struct EmbeddingUsage {
 ///
 /// Aligned with V4 `EmbeddingModelV4Result`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct EmbeddingResult {
     /// Generated embeddings, in the same order as the input `values`.
     pub embeddings: Vec<Embedding>,
 
     /// Token usage (input tokens only).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub usage: Option<EmbeddingUsage>,
 
     /// Additional provider-specific metadata.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
 
     /// Optional response information for debugging.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response: Option<EmbeddingResponse>,
 
     /// Warnings for the call, e.g. unsupported settings.
@@ -94,11 +111,16 @@ pub struct EmbeddingResult {
 
 /// Debugging response info specific to embedding calls.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct EmbeddingResponse {
     /// Response headers.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
     /// The response body (opaque JSON).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<serde_json::Value>,
 }
 
diff --git a/aimux-core/src/error.rs b/aimux-core/src/error.rs
index d9c7bfd9b..9fd4097b0 100644
--- a/aimux-core/src/error.rs
+++ b/aimux-core/src/error.rs
@@ -29,6 +29,7 @@ use ts_rs::TS;
 /// This keeps transport and response failures self-contained and matches the
 /// AI SDK's `APICallError` contract.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ApiCallError {
     /// Sanitized request URL. Required for every API-derived failure.
@@ -39,12 +40,14 @@ pub struct ApiCallError {
     /// (`APICallError.statusCode`). Always the *observed* status: the HTTP
     /// layer fills it for every response-derived error; errors built without
     /// an HTTP exchange leave it `None`.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub status_code: Option<u16>,
     /// Provider's machine-readable code (OpenAI's `code`/`type`,
     /// e.g. `"rate_limit_exceeded"` vs `"insufficient_quota"`). Never an HTTP
     /// status. Our normalized take on `APICallError.data`.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_code: Option<String>,
     /// Human-readable failure text. Provider text is verbatim when available;
     /// locally detected transport/parse failures include their source detail.
@@ -53,13 +56,16 @@ pub struct ApiCallError {
     /// The raw response body, verbatim (`APICallError.responseBody`) — the
     /// lossless evidence when `message`/`provider_code` are extractions.
     /// `None` when the error did not come from a response body.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response_body: Option<String>,
     /// Sanitized response headers.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response_headers: Option<HashMap<String, String>>,
     /// Parsed provider error data.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub data: Option<serde_json::Value>,
     /// Whether retrying can help (`APICallError.isRetryable`) — stored at
     /// construction, exactly like the AI SDK: the response path computes it
@@ -103,6 +109,7 @@ pub enum RetryErrorReason {
 
 /// Complete error history for a retried model operation.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct RetryError {
     pub reason: RetryErrorReason,
@@ -164,7 +171,8 @@ impl std::fmt::Display for ApiCallError {
 /// failure came from. `ApiCallError` is boxed only to keep the Rust enum
 /// compact; serde and every binding still observe the same object shape.
 #[derive(Debug, Clone, Serialize, Deserialize, TS, Error)]
-#[ts(export)]
+#[serde(from = "AiMuxErrorWire", into = "AiMuxErrorWire")]
+#[ts(export, as = "AiMuxErrorWire")]
 pub enum AiMuxError {
     #[error("API call error: {0}")]
     ApiCall(Box<ApiCallError>),
@@ -195,10 +203,12 @@ pub enum AiMuxError {
     #[error("Model tried to call unavailable tool '{tool_name}'. {}", no_such_tool_availability(.available_tools))]
     NoSuchTool {
         tool_name: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         available_tools: Option<Vec<String>>,
         /// Original argument text, when supplied by the provider. Absent in
         /// older serialized errors and errors constructed without a call.
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         tool_input: Option<String>,
     },
@@ -280,7 +290,11 @@ pub enum AiMuxError {
 
     /// Registry-level "provider name does not resolve" (the AI SDK's
     /// `NoSuchProviderError`).
-    #[error("No such provider: {provider_id} (available providers: {})", .available_providers.join(","))]
+    ///
+    /// The message adds up to three close names from `available_providers`
+    /// (an aimux addition to the upstream text; the fields are unchanged).
+    #[error("No such provider: {provider_id}{} (available providers: {})",
+        did_you_mean(.provider_id, .available_providers), .available_providers.join(","))]
     NoSuchProvider {
         provider_id: String,
         model_id: String,
@@ -298,6 +312,197 @@ pub enum AiMuxError {
     Other(String),
 }
 
+/// The JSON form of [`AiMuxError`]: internally tagged by `name`, with the AI
+/// SDK's error names (`AI_APICallError`, ...). `AiMuxError` serializes through
+/// this mirror because its tuple variants cannot carry an internal tag.
+#[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(tag = "name", rename_all_fields = "camelCase")]
+pub enum AiMuxErrorWire {
+    #[serde(rename = "AI_APICallError")]
+    ApiCall(Box<ApiCallError>),
+    #[serde(rename = "AI_RetryError")]
+    Retry(RetryError),
+    #[serde(rename = "AI_JSONParseError")]
+    JsonParse { message: String },
+    #[serde(rename = "AI_NoOutputGeneratedError")]
+    NoOutputGenerated { message: String },
+    #[serde(rename = "AI_InvalidResponseDataError")]
+    InvalidResponseData { message: String },
+    #[serde(rename = "AI_ToolCallNotFoundForApprovalError")]
+    ToolCallNotFoundForApproval {
+        tool_call_id: String,
+        approval_id: String,
+    },
+    #[serde(rename = "AI_NoSuchToolError")]
+    NoSuchTool {
+        tool_name: String,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        available_tools: Option<Vec<String>>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        tool_input: Option<String>,
+    },
+    #[serde(rename = "AI_InvalidToolInputError")]
+    InvalidToolInput {
+        tool_name: String,
+        tool_input: String,
+        cause: String,
+    },
+    #[serde(rename = "AI_ToolCallRepairError")]
+    ToolCallRepair {
+        original_error: Box<AiMuxError>,
+        cause: Box<AiMuxError>,
+    },
+    #[serde(rename = "AI_InvalidArgumentError")]
+    InvalidArgument { message: String },
+    #[serde(rename = "AI_InvalidPromptError")]
+    InvalidPrompt { message: String },
+    #[serde(rename = "AI_LoadAPIKeyError")]
+    LoadApiKey {
+        env_var: String,
+        description: String,
+    },
+    #[serde(rename = "AI_LoadSettingError")]
+    LoadSetting {
+        env_var: String,
+        /// `name` is the union's tag key, so the setting's name is `settingName`.
+        #[serde(rename = "settingName")]
+        name: String,
+    },
+    #[serde(rename = "AI_TokenExpiredError")]
+    TokenExpired { message: String },
+    #[serde(rename = "AI_UnsupportedFunctionalityError")]
+    UnsupportedFunctionality { message: String },
+    #[serde(rename = "AI_NoSuchModelError")]
+    NoSuchModel {
+        model_id: String,
+        #[serde(default)]
+        model_type: String,
+    },
+    #[serde(rename = "AI_NoSuchProviderError")]
+    NoSuchProvider {
+        provider_id: String,
+        model_id: String,
+        model_type: String,
+        available_providers: Vec<String>,
+    },
+    #[serde(rename = "TimeoutError")]
+    Timeout { message: String },
+    #[serde(rename = "AbortError")]
+    Aborted { message: String },
+    #[serde(rename = "AI_Error")]
+    Other { message: String },
+}
+
+/// Maps each `AiMuxError` variant to its `AiMuxErrorWire` twin and back:
+/// tuple variants carry their payload as `message`, struct variants keep
+/// their fields.
+macro_rules! wire_variants {
+    (
+        newtype: $($nv:ident),* ;
+        message: $($mv:ident),* ;
+        fields: $($fv:ident { $($f:ident),* }),*
+    ) => {
+        impl From<AiMuxError> for AiMuxErrorWire {
+            fn from(e: AiMuxError) -> Self {
+                match e {
+                    $(AiMuxError::$nv(v) => Self::$nv(v),)*
+                    $(AiMuxError::$mv(message) => Self::$mv { message },)*
+                    $(AiMuxError::$fv { $($f),* } => Self::$fv { $($f),* },)*
+                }
+            }
+        }
+        impl From<AiMuxErrorWire> for AiMuxError {
+            fn from(e: AiMuxErrorWire) -> Self {
+                match e {
+                    $(AiMuxErrorWire::$nv(v) => Self::$nv(v),)*
+                    $(AiMuxErrorWire::$mv { message } => Self::$mv(message),)*
+                    $(AiMuxErrorWire::$fv { $($f),* } => Self::$fv { $($f),* },)*
+                }
+            }
+        }
+    };
+}
+
+wire_variants! {
+    newtype: ApiCall, Retry;
+    message: JsonParse, NoOutputGenerated, InvalidResponseData, InvalidArgument,
+        InvalidPrompt, TokenExpired, UnsupportedFunctionality, Timeout, Aborted, Other;
+    fields: ToolCallNotFoundForApproval { tool_call_id, approval_id },
+        NoSuchTool { tool_name, available_tools, tool_input },
+        InvalidToolInput { tool_name, tool_input, cause },
+        ToolCallRepair { original_error, cause },
+        LoadApiKey { env_var, description },
+        LoadSetting { env_var, name },
+        NoSuchModel { model_id, model_type },
+        NoSuchProvider { provider_id, model_id, model_type, available_providers }
+}
+
+/// `". Did you mean 'a', 'b'?"` for the names in `candidates` closest to
+/// `name`, or `""` when none is close. Names compare case-insensitively with
+/// `-` read as `_` (the built-in names use `_`); a candidate is close when its
+/// edit distance (adjacent swaps count once) is at most 2 and at most a third
+/// of the longer name. At most three names, nearest first.
+fn did_you_mean(name: &str, candidates: &[String]) -> String {
+    fn normalize(name: &str) -> Vec<char> {
+        name.chars()
+            .map(|c| {
+                if c == '-' {
+                    '_'
+                } else {
+                    c.to_ascii_lowercase()
+                }
+            })
+            .collect()
+    }
+    let wanted = normalize(name);
+    let mut close: Vec<(usize, &str)> = candidates
+        .iter()
+        .filter_map(|candidate| {
+            let distance = edit_distance(&wanted, &normalize(candidate));
+            let longer = wanted.len().max(candidate.chars().count());
+            (distance <= 2 && distance * 3 <= longer).then_some((distance, candidate.as_str()))
+        })
+        .collect();
+    close.sort_unstable();
+    close.dedup_by(|a, b| a.1 == b.1);
+    if close.is_empty() {
+        return String::new();
+    }
+    let names: Vec<String> = close
+        .iter()
+        .take(3)
+        .map(|(_, name)| format!("'{name}'"))
+        .collect();
+    format!(". Did you mean {}?", names.join(", "))
+}
+
+/// Optimal-string-alignment distance: insertions, deletions, substitutions
+/// and swaps of two adjacent characters each cost one.
+fn edit_distance(a: &[char], b: &[char]) -> usize {
+    let mut rows = vec![vec![0usize; b.len() + 1]; a.len() + 1];
+    for (i, row) in rows.iter_mut().enumerate() {
+        row[0] = i;
+    }
+    for (j, cell) in rows[0].iter_mut().enumerate() {
+        *cell = j;
+    }
+    for i in 1..=a.len() {
+        for j in 1..=b.len() {
+            let cost = usize::from(a[i - 1] != b[j - 1]);
+            let mut best = (rows[i - 1][j] + 1)
+                .min(rows[i][j - 1] + 1)
+                .min(rows[i - 1][j - 1] + cost);
+            if i > 1 && j > 1 && a[i - 1] == b[j - 2] && a[i - 2] == b[j - 1] {
+                best = best.min(rows[i - 2][j - 2] + 1);
+            }
+            rows[i][j] = best;
+        }
+    }
+    rows[a.len()][b.len()]
+}
+
 fn no_such_tool_availability(available_tools: &Option<Vec<String>>) -> String {
     match available_tools {
         Some(tools) => format!("Available tools: {}.", tools.join(", ")),
@@ -435,6 +640,62 @@ impl AiMuxError {
 mod tests {
     use super::*;
 
+    fn no_such_provider(name: &str, available: &[&str]) -> String {
+        AiMuxError::NoSuchProvider {
+            provider_id: name.into(),
+            model_id: name.into(),
+            model_type: String::new(),
+            available_providers: available.iter().map(|n| (*n).to_string()).collect(),
+        }
+        .to_string()
+    }
+
+    /// A misspelled provider name gets the close built-in names first; the
+    /// full list still follows.
+    #[test]
+    fn no_such_provider_suggests_close_names() {
+        let available = ["anthropic", "google_vertex", "groq", "openai", "openrouter"];
+        assert_eq!(
+            no_such_provider("opeanai", &available),
+            "No such provider: opeanai. Did you mean 'openai'? \
+             (available providers: anthropic,google_vertex,groq,openai,openrouter)"
+        );
+        // An adjacent swap counts once.
+        assert!(no_such_provider("gorq", &available).contains("Did you mean 'groq'?"));
+        // Case and `-` versus `_` do not matter.
+        assert!(
+            no_such_provider("Google-Vertex", &available).contains("Did you mean 'google_vertex'?")
+        );
+        assert!(no_such_provider("antropic", &available).contains("Did you mean 'anthropic'?"));
+    }
+
+    /// No suggestion when nothing is close or there is nothing to suggest.
+    #[test]
+    fn no_such_provider_without_a_close_name_keeps_the_plain_message() {
+        assert_eq!(
+            no_such_provider("acme", &["anthropic", "openai"]),
+            "No such provider: acme (available providers: anthropic,openai)"
+        );
+        assert_eq!(
+            no_such_provider("p", &[]),
+            "No such provider: p (available providers: )"
+        );
+    }
+
+    /// At most three names, nearest first, then by name.
+    #[test]
+    fn no_such_provider_suggests_at_most_three_names() {
+        let candidates = ["abcdefxy", "abcdefgz", "abcdefgy", "abcdefgx"];
+        assert!(
+            no_such_provider("abcdefgh", &candidates)
+                .contains("Did you mean 'abcdefgx', 'abcdefgy', 'abcdefgz'?")
+        );
+        assert!(
+            no_such_provider("abcdefgh", &["abcdefxy", "abcdefgz"])
+                .contains("Did you mean 'abcdefgz', 'abcdefxy'?")
+        );
+    }
+
     fn api_error(message: &str) -> ApiCallError {
         ApiCallError::new(message, "https://example.test", serde_json::json!({}))
     }
diff --git a/aimux-core/src/evaluation_model.rs b/aimux-core/src/evaluation_model.rs
index 809305fdb..f07e1fdd2 100644
--- a/aimux-core/src/evaluation_model.rs
+++ b/aimux-core/src/evaluation_model.rs
@@ -56,6 +56,7 @@ impl EvaluationQuestionType {
 /// The optional descriptions of a Boolean question's outcomes. An absent key
 /// and an explicit `null` differ: `Some(None)` is `null` (no description).
 #[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct BooleanCriteria {
     #[serde(
@@ -86,7 +87,11 @@ where
 
 /// A judgment to make about the shared state (`EvaluationModelV4Question`).
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
-#[serde(tag = "type", rename_all = "lowercase")]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum EvaluationQuestion {
     Choice {
@@ -103,6 +108,7 @@ pub enum EvaluationQuestion {
     },
     Boolean {
         instructions: EvaluationInput,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         criteria: Option<BooleanCriteria>,
     },
@@ -123,6 +129,7 @@ impl EvaluationQuestion {
 /// Options passed to [`EvaluationModel::do_evaluate`]
 /// (`EvaluationModelV4CallOptions`).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct EvaluationCallOptions {
     /// One shared state, even when the value is an array.
@@ -134,13 +141,21 @@ pub struct EvaluationCallOptions {
     #[serde(skip)]
     #[ts(skip)]
     pub abort_signal: Option<AbortSignal>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 }
 
 /// The answer to one question (`EvaluationModelV4Answer`).
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
-#[serde(tag = "type", rename_all = "lowercase")]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum EvaluationAnswer {
     Choice {
@@ -148,6 +163,7 @@ pub enum EvaluationAnswer {
         /// exists.
         choice: String,
         /// Complete distribution over the question's options, when available.
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         probabilities: Option<HashMap<String, f64>>,
     },
@@ -156,6 +172,7 @@ pub enum EvaluationAnswer {
         score: f64,
         /// Complete distribution keyed by zero-based level indices as
         /// strings; when supplied, `score` is its probability-weighted mean.
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         probabilities: Option<HashMap<String, f64>>,
     },
@@ -168,31 +185,50 @@ pub enum EvaluationAnswer {
 
 /// Decimal places the provider rounded its output to; omit for full precision.
 #[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct EvaluationRounding {
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub probability_decimals: Option<u32>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub score_decimals: Option<u32>,
 }
 
 /// Token usage of an evaluation.
 #[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct EvaluationUsage {
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub input_tokens: Option<u32>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub output_tokens: Option<u32>,
 }
 
 /// The result of [`EvaluationModel::do_evaluate`] (`EvaluationModelV4Result`).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct EvaluationResult {
     /// Exactly one answer per question, under the original question ids.
     #[ts(as = "std::collections::HashMap<String, EvaluationAnswer>")]
     pub answers: IndexMap<String, EvaluationAnswer>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub rounding: Option<EvaluationRounding>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub usage: Option<EvaluationUsage>,
     pub warnings: Vec<Warning>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response: Option<ResponseInfo>,
 }
 
diff --git a/aimux-core/src/files_model.rs b/aimux-core/src/files_model.rs
index 091488838..9aeb7e64b 100644
--- a/aimux-core/src/files_model.rs
+++ b/aimux-core/src/files_model.rs
@@ -22,6 +22,11 @@ use crate::shared::{
 /// `SharedV4FileData` (URLs and provider references are not valid uploads), so
 /// this is a dedicated two-variant enum rather than the full [`crate::shared::FileData`].
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum UploadFileData {
     /// Raw bytes (`Uint8Array`) or a base64-encoded string.
@@ -34,6 +39,7 @@ pub enum UploadFileData {
 ///
 /// Aligned with V4 `FilesV4UploadFileCallOptions`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct UploadFileCallOptions {
     /// The file data (raw bytes/base64 or inline text).
@@ -43,9 +49,13 @@ pub struct UploadFileCallOptions {
     pub media_type: String,
 
     /// The filename of the file.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub filename: Option<String>,
 
     /// Additional provider-specific options, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 
     /// Abort signal for cancelling the upload operation.
@@ -71,6 +81,7 @@ impl UploadFileCallOptions {
 ///
 /// Aligned with V4 `FilesV4UploadFileResult`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct UploadFileResult {
     /// A provider reference mapping provider names to provider-specific file
@@ -79,12 +90,18 @@ pub struct UploadFileResult {
 
     /// The IANA media type of the uploaded file, if available from the
     /// provider.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub media_type: Option<String>,
 
     /// The filename of the uploaded file, if available from the provider.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub filename: Option<String>,
 
     /// Additional provider-specific metadata, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
 
     /// Warnings from the provider.
diff --git a/aimux-core/src/generate.rs b/aimux-core/src/generate.rs
index 3a71543df..0df13af5d 100644
--- a/aimux-core/src/generate.rs
+++ b/aimux-core/src/generate.rs
@@ -82,35 +82,70 @@ async fn resolve_file(
 /// Unlike `CallOptions` (provider-facing), this does not include `prompt`
 /// (passed separately) and defaults are more ergonomic.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct GenerateTextOptions {
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_output_tokens: Option<u32>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub temperature: Option<f64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub stop_sequences: Option<Vec<String>>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub top_p: Option<f64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub top_k: Option<f64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub presence_penalty: Option<f64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub frequency_penalty: Option<f64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response_format: Option<ResponseFormat>,
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub seed: Option<u64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub tools: Option<Vec<Tool>>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub tool_choice: Option<ToolChoice>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<HashMap<String, String>>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
     /// Top-level reasoning effort.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub reasoning: Option<ReasoningEffort>,
     /// System instructions prepended to the prompt.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub instructions: Option<String>,
     /// Per-call retry count override. `None` = provider default, `Some(0)` = disable.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_retries: Option<u32>,
     /// Per-call timeout configuration (total / first-chunk / chunk idle).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timeout: Option<crate::options::TimeoutConfiguration>,
     /// Session identifier, for grouping consecutive calls into a session
     /// (observability, see RFC-0024). Orthogonal to RFC-0019 session-affinity
     /// headers. When `None` and the optional session inferer is enabled, one
     /// may be inferred from prompt-prefix continuation.
-    #[serde(skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub session_id: Option<String>,
     /// Abort signal for cancelling the call.
     ///
@@ -120,6 +155,8 @@ pub struct GenerateTextOptions {
     #[ts(skip)]
     pub abort_signal: Option<AbortSignal>,
     /// Emit raw provider stream chunks as `StreamPart::Raw` (debugging aid).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub include_raw_chunks: Option<bool>,
     /// Optional one-shot repair callback for unknown, malformed, or
     /// schema-invalid tool calls.
@@ -174,6 +211,7 @@ impl GenerateTextOptions {
 
 /// Result of `generate_text` (user-facing).
 #[derive(Debug, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct GenerateTextResult {
     /// Ordered generated content with parsed tool calls.
@@ -210,11 +248,13 @@ pub struct GenerateTextResult {
     pub response_messages: Vec<ModelMessage>,
     /// The raw provider-specific finish reason string (e.g. "stop",
     /// "end_turn", "safety"). Useful when `finish_reason.unified` is `Other`.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub raw_finish_reason: Option<String>,
     /// Provider-specific metadata (e.g. Anthropic cache info). Mirrored from
     /// `raw.provider_metadata` for top-level convenience.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<ProviderMetadata>,
     /// Request information from the provider.
     #[serde(default)]
@@ -231,6 +271,7 @@ pub struct GenerateTextResult {
 /// Result of `generate_object` (user-facing, M12). The parsed JSON object plus
 /// convenience fields from the underlying `generate_text` call.
 #[derive(Debug, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct GenerateObjectResult {
     /// The parsed JSON object returned by the model.
@@ -238,17 +279,20 @@ pub struct GenerateObjectResult {
     /// Why generation stopped.
     pub finish_reason: FinishReason,
     /// Raw provider-specific finish reason string.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub raw_finish_reason: Option<String>,
     /// Token usage.
     pub usage: Usage,
     /// Warnings from the provider.
     pub warnings: Vec<Warning>,
     /// Concatenated reasoning text (if the model produced reasoning/thinking).
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub reasoning: Option<String>,
     /// Provider-specific metadata (e.g. Anthropic cache info).
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<ProviderMetadata>,
     /// Response metadata (id, timestamp, model_id).
     #[serde(default)]
diff --git a/aimux-core/src/image_model.rs b/aimux-core/src/image_model.rs
index 159de19df..8bb910d1e 100644
--- a/aimux-core/src/image_model.rs
+++ b/aimux-core/src/image_model.rs
@@ -17,6 +17,11 @@ use crate::{AbortSignal, retry, timeout};
 ///
 /// Aligned with V4 `ImageModelV4File`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum ImageFile {
     /// Inline file data (base64 or binary) with an explicit media type.
@@ -35,6 +40,7 @@ pub enum ImageFile {
 
 /// File payload for an [`ImageFile::File`]: base64 string or raw bytes.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(untagged)]
 #[ts(export)]
 pub enum ImageFileData {
     /// Base64-encoded string.
@@ -51,6 +57,7 @@ pub enum ImageFileData {
 /// upstream API returns base64, return [`ImageOutputs::Base64`]; if binary,
 /// return [`ImageOutputs::Binary`].
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(untagged)]
 #[ts(export)]
 pub enum ImageOutputs {
     /// Images as base64-encoded strings.
@@ -63,13 +70,20 @@ pub enum ImageOutputs {
 ///
 /// Aligned with V4 `ImageModelV4Usage`.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ImageUsage {
     /// Number of input (prompt) tokens used.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub input_tokens: Option<u32>,
     /// Number of output tokens used, if reported.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub output_tokens: Option<u32>,
     /// Total tokens as reported by the provider.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub total_tokens: Option<u32>,
 }
 
@@ -77,10 +91,13 @@ pub struct ImageUsage {
 ///
 /// Aligned with V4 `ImageModelV4CallOptions`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ImageCallOptions {
     /// Prompt for the image generation. `None` for operations (e.g. upscaling)
     /// that do not require a prompt.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub prompt: Option<String>,
 
     /// Number of images to generate.
@@ -88,20 +105,29 @@ pub struct ImageCallOptions {
 
     /// Size of the images, in `{width}x{height}` format.
     /// `None` uses the provider's default size.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub size: Option<Size>,
 
     /// Aspect ratio of the images, in `{width}:{height}` format.
     /// `None` uses the provider's default aspect ratio.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub aspect_ratio: Option<AspectRatio>,
 
     /// Seed for deterministic generation. `None` uses the provider's default.
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub seed: Option<u64>,
 
     /// Images for image editing or variation generation.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub files: Option<Vec<ImageFile>>,
 
     /// Mask image for inpainting operations.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub mask: Option<ImageFile>,
 
     /// Additional provider-specific options, keyed by provider name.
@@ -113,12 +139,18 @@ pub struct ImageCallOptions {
     pub abort_signal: Option<AbortSignal>,
 
     /// Per-call retry override. `None` uses the model default.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_retries: Option<u32>,
 
     /// Per-call operation timeout.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timeout: Option<crate::options::TimeoutConfiguration>,
 
     /// Additional HTTP headers to send with the request.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
 }
 
@@ -146,6 +178,7 @@ impl ImageCallOptions {
 ///
 /// Aligned with V4 `ImageModelV4Result`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ImageResult {
     /// Generated images (base64 strings or binary data).
@@ -155,24 +188,35 @@ pub struct ImageResult {
     pub warnings: Vec<Warning>,
 
     /// Additional provider-specific metadata, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
 
     /// Response information for telemetry and debugging.
     pub response: ImageResponse,
 
     /// Optional token usage for the call (if the provider reports it).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub usage: Option<ImageUsage>,
 }
 
 /// Response information for an image generation call.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ImageResponse {
     /// Timestamp for the start of the generated response (ISO 8601 string).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timestamp: Option<String>,
     /// The ID of the model that was used to generate the response.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub model_id: Option<String>,
     /// Response headers.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
 }
 
diff --git a/aimux-core/src/language_model_message.rs b/aimux-core/src/language_model_message.rs
index 2f8f920ee..fbaef20ef 100644
--- a/aimux-core/src/language_model_message.rs
+++ b/aimux-core/src/language_model_message.rs
@@ -11,7 +11,6 @@ use crate::error::AiMuxError;
 use crate::message::{MessageContent, ModelMessage, Role};
 use crate::shared::{FileBytes, FileData, GeneratedFileData, SharedProviderOptions};
 use std::borrow::Cow;
-use std::collections::HashMap;
 
 use serde::{Deserialize, Serialize};
 use serde_json::Value;
@@ -25,110 +24,144 @@ pub type LanguageModelPrompt = Vec<LanguageModelMessage>;
 /// `provider_options` is message-level (e.g. `anthropic.cacheControl`),
 /// mirroring `LanguageModelV4Message.providerOptions`.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
-#[serde(tag = "role", rename_all = "snake_case")]
+#[serde(
+    tag = "role",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum LanguageModelMessage {
     System {
         content: String,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
     User {
         content: Vec<UserPart>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
     Assistant {
         content: Vec<AssistantPart>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
     Tool {
         content: Vec<ToolPart>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
 }
 
 /// Text content part.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TextPart {
     pub text: String,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 }
 
 /// File content part (the only file part; `data` is a tagged `FileData`).
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct FilePart {
     pub data: FileData,
     pub media_type: String,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub filename: Option<String>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 }
 
 /// Reasoning / thinking content part.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 #[ts(rename = "LanguageModelReasoningPart")]
 pub struct ReasoningPart {
     pub text: String,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 }
 
 /// Tool call content part.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ToolCallPart {
     pub tool_call_id: String,
     pub tool_name: String,
     pub input: Value,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_executed: Option<bool>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 }
 
 /// Tool result content part.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ToolResultPart {
     pub tool_call_id: String,
     pub tool_name: String,
     pub output: ToolResultOutput,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 }
 
 /// The provider-facing result of a tool call.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
-#[serde(tag = "type", rename_all = "kebab-case")]
+#[serde(
+    tag = "type",
+    rename_all = "kebab-case",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum ToolResultOutput {
     Text {
         value: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
     Json {
         value: Value,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
     ExecutionDenied {
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         reason: Option<String>,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
     ErrorText {
         value: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
     ErrorJson {
         value: Value,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -139,12 +172,17 @@ pub enum ToolResultOutput {
 
 /// Content within a tool result.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
-#[serde(tag = "type", rename_all = "kebab-case")]
+#[serde(
+    tag = "type",
+    rename_all = "kebab-case",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum ToolResultContent {
     Text(TextPart),
     File(FilePart),
     Custom {
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_options: Option<SharedProviderOptions>,
     },
@@ -152,31 +190,38 @@ pub enum ToolResultContent {
 
 /// A file generated as part of reasoning.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ReasoningFilePart {
     pub data: GeneratedFileData,
     pub media_type: String,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 }
 
 /// Provider-specific content identified by its provider and kind.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct CustomPart {
     pub kind: String,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 }
 
 /// The user's decision for a provider-executed tool approval request.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ToolApprovalResponsePart {
     pub approval_id: String,
     pub approved: bool,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub reason: Option<String>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 }
@@ -229,10 +274,7 @@ impl LanguageModelMessage {
 
 /// Map a user-facing part onto the widest provider part (the assistant union);
 /// the role then narrows it.
-fn to_assistant_part(
-    part: &ContentPart,
-    tool_names: &HashMap<String, String>,
-) -> Result<AssistantPart, AiMuxError> {
+fn to_assistant_part(part: &ContentPart) -> Result<AssistantPart, AiMuxError> {
     let file = |data, media_type: &String, filename: &Option<String>, po: &Option<_>| {
         AssistantPart::File(FilePart {
             data,
@@ -338,42 +380,13 @@ fn to_assistant_part(
         }),
         ContentPart::ToolResult {
             tool_call_id,
-            result,
             tool_name,
-            is_error,
+            output,
             provider_options,
-            ..
         } => AssistantPart::ToolResult(ToolResultPart {
             tool_call_id: tool_call_id.clone(),
-            tool_name: tool_name
-                .as_ref()
-                .or_else(|| tool_names.get(tool_call_id))
-                .cloned()
-                .ok_or_else(|| {
-                    AiMuxError::InvalidPrompt(format!(
-                        "tool result {tool_call_id} has no tool name or preceding tool call"
-                    ))
-                })?,
-            output: if *is_error == Some(true) {
-                ToolResultOutput::ErrorText {
-                    value: match result {
-                        Value::Null => "unknown error".into(),
-                        Value::String(value) => value.clone(),
-                        _ => result.to_string(),
-                    },
-                    provider_options: None,
-                }
-            } else if let Value::String(value) = result {
-                ToolResultOutput::Text {
-                    value: value.clone(),
-                    provider_options: None,
-                }
-            } else {
-                ToolResultOutput::Json {
-                    value: result.clone(),
-                    provider_options: None,
-                }
-            },
+            tool_name: tool_name.clone(),
+            output: output.clone(),
             provider_options: provider_options.clone(),
         }),
         ContentPart::Custom {
@@ -423,7 +436,6 @@ pub fn convert_to_language_model_prompt(
         });
     }
 
-    let mut tool_names = HashMap::new();
     for msg in messages {
         let bad = |what: &str| AiMuxError::InvalidPrompt(format!("{:?} message: {what}", msg.role));
         if msg.role == Role::System {
@@ -440,18 +452,6 @@ pub fn convert_to_language_model_prompt(
             MessageContent::Text(text) => Cow::Owned(vec![ContentPart::text(text)]),
             MessageContent::Parts(parts) => Cow::Borrowed(parts),
         };
-        if msg.role == Role::Assistant {
-            for part in parts.iter() {
-                if let ContentPart::ToolCall {
-                    tool_call_id,
-                    tool_name,
-                    ..
-                } = part
-                {
-                    tool_names.insert(tool_call_id.clone(), tool_name.clone());
-                }
-            }
-        }
         let parts = parts
             .iter()
             .filter(|part| {
@@ -464,7 +464,7 @@ pub fn convert_to_language_model_prompt(
                         if text.is_empty() && (msg.role == Role::User
                             || (msg.role == Role::Assistant && provider_options.is_none())))
             })
-            .map(|part| to_assistant_part(part, &tool_names))
+            .map(to_assistant_part)
             .collect::<Result<Vec<_>, _>>()?;
         let provider_options = None;
         result.push(match msg.role {
diff --git a/aimux-core/src/message.rs b/aimux-core/src/message.rs
index 00ee5622c..48c9ff3e1 100644
--- a/aimux-core/src/message.rs
+++ b/aimux-core/src/message.rs
@@ -34,6 +34,7 @@ pub enum MessageContent {
 
 /// A single user-facing chat message.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ModelMessage {
     pub role: Role,
diff --git a/aimux-core/src/moa.rs b/aimux-core/src/moa.rs
index 6123c1924..3a3f08c37 100644
--- a/aimux-core/src/moa.rs
+++ b/aimux-core/src/moa.rs
@@ -34,6 +34,7 @@ pub enum MoaFailMode {
 }
 
 #[derive(Debug, Clone, Deserialize)]
+#[serde(rename_all = "camelCase")]
 pub struct MoaConfig {
     /// `provider()` value (default `"moa"`).
     #[serde(default = "default_moa_provider")]
diff --git a/aimux-core/src/model_catalogue.rs b/aimux-core/src/model_catalogue.rs
index a7ef0d8e9..3eadfc4ab 100644
--- a/aimux-core/src/model_catalogue.rs
+++ b/aimux-core/src/model_catalogue.rs
@@ -24,16 +24,18 @@ use ts_rs::TS;
 /// comes from [`ModelSpec`] (community knowledge); the host merges the two as
 /// needed (RFC-0027). aimux itself does not combine them.
 #[derive(Debug, Clone, Serialize, Deserialize, TS, PartialEq)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct RuntimeModel {
     /// Model identifier, e.g. `"gpt-4o"`.
     pub id: String,
     /// `owned_by` field from OpenAI-compatible `/models` responses.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub owned_by: Option<String>,
     /// `created` unix timestamp from OpenAI-compatible `/models` responses.
+    #[ts(optional, type = "number")]
     #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[ts(type = "number | null")]
     pub created: Option<u64>,
 }
 
@@ -93,24 +95,26 @@ pub enum Modality {
 
 /// Token / request limits for a model.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS, PartialEq)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ModelLimits {
     /// Maximum total context window (input + output), in tokens.
+    #[ts(optional, type = "number")]
     #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[ts(type = "number | null")]
     pub context: Option<u64>,
     /// Maximum output tokens.
+    #[ts(optional, type = "number")]
     #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[ts(type = "number | null")]
     pub output: Option<u64>,
     /// Maximum input tokens (when distinct from `context`).
+    #[ts(optional, type = "number")]
     #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[ts(type = "number | null")]
     pub input: Option<u64>,
 }
 
 /// Input/output modality sets.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS, PartialEq)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ModelModalities {
     #[serde(default, skip_serializing_if = "Vec::is_empty")]
@@ -121,13 +125,14 @@ pub struct ModelModalities {
 
 /// Model-level capability flags (more precise than provider-level `supports_*`).
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS, PartialEq)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ModelCapabilities {
     /// Supports function/tool calling.
     #[serde(default)]
     pub tool_call: bool,
     /// Supports structured / JSON-schema output.
-    #[serde(default, rename = "structured_output")]
+    #[serde(default)]
     pub structured_output: bool,
     /// Accepts a `temperature` parameter.
     #[serde(default)]
@@ -172,6 +177,7 @@ pub enum ReasoningVisibility {
 
 /// Reasoning / thinking portrait for a model (advisory).
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS, PartialEq)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ReasoningSpec {
     /// Whether the model family supports reasoning at all.
@@ -181,37 +187,45 @@ pub struct ReasoningSpec {
     #[serde(default)]
     pub default_enabled: bool,
     /// Primary control mode.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub mode: Option<ReasoningMode>,
     /// Default effort level (when `mode` is Effort/Mixed).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub effort_default: Option<String>,
     /// Supported effort levels.
     #[serde(default, skip_serializing_if = "Vec::is_empty")]
     pub effort_options: Vec<String>,
     /// Minimum token budget (when `mode` is Budget/Mixed).
+    #[ts(optional, type = "number")]
     #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[ts(type = "number | null")]
     pub budget_min: Option<u64>,
     /// Interleaved reasoning (thinking blocks between tool calls).
     #[serde(default)]
     pub interleaved: bool,
     /// Output visibility.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub visibility: Option<ReasoningVisibility>,
 }
 
 /// Pricing (advisory, read-only metadata). Units follow the source catalog
 /// (anya2a uses USD per 1M tokens).
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS, PartialEq)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ModelCost {
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub input: Option<f64>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub output: Option<f64>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub cache_read: Option<f64>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub cache_write: Option<f64>,
 }
@@ -221,8 +235,10 @@ pub struct ModelCost {
 /// All fields are advisory — aimux never auto-applies them in the request path.
 /// The caller reads them to decide how to configure `GenerateTextOptions`.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS, PartialEq)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ModelSpec {
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub display_name: Option<String>,
     #[serde(default)]
@@ -233,16 +249,20 @@ pub struct ModelSpec {
     pub modalities: ModelModalities,
     #[serde(default)]
     pub capabilities: ModelCapabilities,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub reasoning: Option<ReasoningSpec>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub cost: Option<ModelCost>,
     #[serde(default)]
     pub source: CatalogueSource,
     /// Provider name (normalized to aimux registry spelling).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider: Option<String>,
     /// Verbatim source entry, preserved for forward compatibility / debugging.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub raw: Option<serde_json::Value>,
 }
diff --git a/aimux-core/src/model_id.rs b/aimux-core/src/model_id.rs
index a319c1635..932828f2c 100644
--- a/aimux-core/src/model_id.rs
+++ b/aimux-core/src/model_id.rs
@@ -6,6 +6,7 @@ use ts_rs::TS;
 
 /// A model identifier in `"provider/model-name"` format, e.g. `"openai/gpt-4o"`.
 #[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ModelId {
     provider: String,
diff --git a/aimux-core/src/options.rs b/aimux-core/src/options.rs
index d43fbb159..ceb9e53a3 100644
--- a/aimux-core/src/options.rs
+++ b/aimux-core/src/options.rs
@@ -16,14 +16,25 @@ use crate::types::ReasoningEffort;
 
 /// How the model should format its response.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum ResponseFormat {
     /// Plain text (default).
     Text,
     /// JSON output, optionally constrained to a schema.
     Json {
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         schema: Option<Value>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         name: Option<String>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         description: Option<String>,
     },
 }
@@ -35,6 +46,7 @@ pub enum ResponseFormat {
 /// corresponding limit. A `total` timeout also covers retry backoff and the
 /// whole streamed response.
 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TimeoutConfiguration {
     /// Overall timeout for the entire call (including retries and, for
@@ -42,21 +54,25 @@ pub struct TimeoutConfiguration {
     // `number`, not the `bigint` ts-rs infers from u64: the JS bindings pass
     // options through `JSON.stringify`, which throws on BigInt. Milliseconds
     // cannot reach 2^53 (~285k years), so precision is never at stake.
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub total_ms: Option<u64>,
     /// Timeout for one generation step, including that step's attempts and
     /// retry backoff, in milliseconds. Aimux currently has one step.
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub step_ms: Option<u64>,
     /// Timeout waiting for the first stream chunk (streaming only).
     ///
     /// Counted from operation start, so it also bounds stream establishment
     /// and any retries before the first semantic output: it is the
     /// user-perceived time-to-first-output budget, not a per-attempt timer.
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub first_chunk_ms: Option<u64>,
     /// Maximum idle time between consecutive stream chunks (streaming only).
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub chunk_ms: Option<u64>,
 }
 
@@ -66,61 +82,93 @@ pub struct TimeoutConfiguration {
 /// `GenerateTextOptions` (user-facing) which is converted to `CallOptions`
 /// by the `generate_text` / `stream_text` functions.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct CallOptions {
     /// The standardized prompt (message array). Required.
     pub prompt: LanguageModelPrompt,
 
     /// Maximum tokens to generate.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_output_tokens: Option<u32>,
 
     /// Sampling temperature.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub temperature: Option<f64>,
 
     /// Stop sequences.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub stop_sequences: Option<Vec<String>>,
 
     /// Nucleus sampling `top_p`.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub top_p: Option<f64>,
 
     /// Top-k sampling.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub top_k: Option<f64>,
 
     /// Presence penalty.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub presence_penalty: Option<f64>,
 
     /// Frequency penalty.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub frequency_penalty: Option<f64>,
 
     /// Response format (text or JSON).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response_format: Option<ResponseFormat>,
 
     /// Seed for reproducibility.
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub seed: Option<u64>,
 
     /// Tools available to the model (function tools and/or provider-defined tools).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub tools: Option<Vec<Tool>>,
 
     /// How the model should choose tools.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub tool_choice: Option<ToolChoice>,
 
     /// Extra HTTP headers.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<HashMap<String, String>>,
 
     /// Provider-specific options (keyed by provider name).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 
     /// Top-level reasoning effort. Maps to OpenAI `reasoning_effort` and
     /// Anthropic `thinking` config.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub reasoning: Option<ReasoningEffort>,
 
     /// Per-call retry count override. `None` uses the provider's configured
     /// Core operation retry. `Some(0)` disables retries.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_retries: Option<u32>,
 
     /// Per-call timeout configuration (total / first-chunk / chunk idle).
     /// `None` = no timeouts (provider defaults still apply at the HTTP layer).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timeout: Option<TimeoutConfiguration>,
 
     /// Session identifier, for grouping consecutive calls into a session
@@ -129,7 +177,8 @@ pub struct CallOptions {
     /// Orthogonal to RFC-0019 session-affinity headers: this field is for
     /// local grouping, while session headers in `headers` are for upstream
     /// routing — both may share an id value but travel different paths.
-    #[serde(skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub session_id: Option<String>,
 
     /// Abort signal for cancelling the call.
@@ -159,6 +208,8 @@ pub struct CallOptions {
     /// Currently honored by the OpenAI-compatible family (openai / azure /
     /// openai-compatible registry providers); other provider families ignore
     /// it for now (RFC-0016 M2).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub include_raw_chunks: Option<bool>,
 }
 
diff --git a/aimux-core/src/parse_tool_call.rs b/aimux-core/src/parse_tool_call.rs
index a784bbc6e..79be0bacd 100644
--- a/aimux-core/src/parse_tool_call.rs
+++ b/aimux-core/src/parse_tool_call.rs
@@ -163,7 +163,9 @@ fn apply_repair_outcome(
     let cause = match outcome {
         RepairOutcome::Repaired(repaired) => match parse_and_validate_tool_call(&repaired, tools) {
             Ok((input, dynamic)) => return parsed_tool_call(repaired, input, dynamic),
-            Err(repaired_error) => repaired_error,
+            // Upstream rethrows the parse error of the repaired call as is;
+            // only a failing repair function is wrapped.
+            Err(repaired_error) => return invalid_tool_call(tool_call, repaired_error),
         },
         RepairOutcome::Unchanged => return invalid_tool_call(tool_call, original_error),
         RepairOutcome::Failed(repair_error) => repair_error,
@@ -317,10 +319,15 @@ fn invalid_tool_call(tool_call: RawToolCall, error: AiMuxError) -> ToolCall {
 /// The bindings cannot carry a [`ToolCallRepair`] closure, so a host instead
 /// reads the invalid call off the result, repairs it in its own language, and
 /// hands the answer back as this value. Wire shape:
-/// `{"type":"repaired","tool_call":{…}}`, `{"type":"unchanged"}`, or
+/// `{"type":"repaired","toolCall":{…}}`, `{"type":"unchanged"}`, or
 /// `{"type":"failed","message":"…"}`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
-#[serde(tag = "type", rename_all = "snake_case", deny_unknown_fields)]
+#[serde(
+    tag = "type",
+    rename_all = "kebab-case",
+    rename_all_fields = "camelCase",
+    deny_unknown_fields
+)]
 #[ts(export)]
 pub enum ToolCallRepairReply {
     /// Re-validate this replacement call, exactly as the closure path does
@@ -435,9 +442,9 @@ pub fn tool_call_repair_inputs(
 /// Build the repair argument a host needs, as JSON, or `None` when the call
 /// was made without a tool set.
 ///
-/// Mirrors the AI SDK `repairToolCall` argument: `{tool_call, error,
-/// input_schema, tools, messages, instructions}`. `tool_call` is the
-/// provider-facing call (raw argument *text*), `input_schema` the schema of
+/// Mirrors the AI SDK `repairToolCall` argument: `{toolCall, error,
+/// inputSchema, tools, messages, instructions}`. `toolCall` is the
+/// provider-facing call (raw argument *text*), `inputSchema` the schema of
 /// the named tool (an empty-object schema when the name does not resolve),
 /// and `error` the serialized failure the call already carries.
 ///
@@ -468,9 +475,9 @@ pub fn tool_call_repair_context(
         error,
     };
     Ok(Some(serde_json::json!({
-        "tool_call": context.tool_call,
+        "toolCall": context.tool_call,
         "error": context.error,
-        "input_schema": context.input_schema(&context.tool_call.tool_name),
+        "inputSchema": context.input_schema(&context.tool_call.tool_name),
         "tools": context.tools,
         "messages": context.messages,
         "instructions": context.instructions,
@@ -502,7 +509,7 @@ pub fn apply_tool_call_repair(
 /// `GenerateObjectResult`, or `StreamTextResultAggregated`, returning the
 /// patched result.
 ///
-/// Both `tool_calls[]` and the matching `response_messages[]` tool-call part
+/// Both `toolCalls[]` and the matching `responseMessages[]` tool-call part
 /// are rewritten, the latter under the same rule `to_response_messages` uses
 /// (malformed primitive input is not replayed as a prompt input). A
 /// `GenerateObjectResult` is recognised by its nested `raw` result; `object`
@@ -521,7 +528,7 @@ pub fn apply_tool_call_repair(
 /// # Errors
 ///
 /// [`AiMuxError::InvalidArgument`] when `tools` is `None`, when the document
-/// has no `tool_calls` array, when `tool_call_id` is missing or ambiguous,
+/// has no `toolCalls` array, when `toolCallId` is missing or ambiguous,
 /// when the matched call is not invalid, or when a replacement id conflicts
 /// with another call.
 pub fn apply_tool_call_repair_to_result(
@@ -532,43 +539,43 @@ pub fn apply_tool_call_repair_to_result(
 ) -> Result<Value, AiMuxError> {
     let mut patched = result.clone();
     // `GenerateObjectResult` carries the whole text result under `raw`.
-    let target = if patched.get("tool_calls").is_some() {
+    let target = if patched.get("toolCalls").is_some() {
         &mut patched
     } else {
         patched.get_mut("raw").ok_or_else(|| {
-            AiMuxError::InvalidArgument("result: neither tool_calls nor raw".to_string())
+            AiMuxError::InvalidArgument("result: neither toolCalls nor raw".to_string())
         })?
     };
 
     let calls = target
-        .get_mut("tool_calls")
+        .get_mut("toolCalls")
         .and_then(Value::as_array_mut)
-        .ok_or_else(|| AiMuxError::InvalidArgument("result: no tool_calls array".to_string()))?;
-    // Both `tool_calls` and the transcript are keyed by id here, so a
+        .ok_or_else(|| AiMuxError::InvalidArgument("result: no toolCalls array".to_string()))?;
+    // Both `toolCalls` and the transcript are keyed by id here, so a
     // duplicated id cannot be patched unambiguously; refuse rather than
     // rewrite the first entry twice.
     let target_index = calls
         .iter()
-        .position(|call| call.get("tool_call_id").and_then(Value::as_str) == Some(tool_call_id))
+        .position(|call| call.get("toolCallId").and_then(Value::as_str) == Some(tool_call_id))
         .ok_or_else(|| {
             AiMuxError::InvalidArgument(format!("result: no tool call '{tool_call_id}'"))
         })?;
     if calls[target_index + 1..]
         .iter()
-        .any(|call| call.get("tool_call_id").and_then(Value::as_str) == Some(tool_call_id))
+        .any(|call| call.get("toolCallId").and_then(Value::as_str) == Some(tool_call_id))
     {
         return Err(AiMuxError::InvalidArgument(format!(
             "result: tool call id '{tool_call_id}' is not unique"
         )));
     }
     let original: ToolCall = serde_json::from_value(calls[target_index].clone())
-        .map_err(|error| AiMuxError::InvalidArgument(format!("result.tool_calls: {error}")))?;
+        .map_err(|error| AiMuxError::InvalidArgument(format!("result.toolCalls: {error}")))?;
 
     let repaired = apply_tool_call_repair(&original, tools, reply)?;
     if repaired.tool_call_id != tool_call_id
         && calls.iter().enumerate().any(|(index, call)| {
             index != target_index
-                && call.get("tool_call_id").and_then(Value::as_str)
+                && call.get("toolCallId").and_then(Value::as_str)
                     == Some(repaired.tool_call_id.as_str())
         })
     {
@@ -580,20 +587,20 @@ pub fn apply_tool_call_repair_to_result(
     calls[target_index] = serde_json::to_value(&repaired)
         .map_err(|error| AiMuxError::InvalidArgument(format!("tool call: {error}")))?;
 
-    // The replayed transcript must agree with `tool_calls`, or the next turn
+    // The replayed transcript must agree with `toolCalls`, or the next turn
     // sends the model the unrepaired arguments.
     let replay_input =
         crate::response_messages::response_tool_call_input(&repaired.input, repaired.invalid);
     for part in target
-        .get_mut("response_messages")
+        .get_mut("responseMessages")
         .and_then(Value::as_array_mut)
         .into_iter()
         .flatten()
         .filter_map(|message| message.get_mut("content")?.as_array_mut())
         .flatten()
     {
-        if part.get("type").and_then(Value::as_str) != Some("tool_call")
-            || part.get("tool_call_id").and_then(Value::as_str) != Some(tool_call_id)
+        if part.get("type").and_then(Value::as_str) != Some("tool-call")
+            || part.get("toolCallId").and_then(Value::as_str) != Some(tool_call_id)
         {
             continue;
         }
@@ -601,17 +608,17 @@ pub fn apply_tool_call_repair_to_result(
             continue;
         };
         // A repair may rename the call; the transcript must keep pointing at
-        // the same entry as `tool_calls`.
+        // the same entry as `toolCalls`.
         part.insert(
-            "tool_call_id".to_string(),
+            "toolCallId".to_string(),
             repaired.tool_call_id.clone().into(),
         );
-        part.insert("tool_name".to_string(), repaired.tool_name.clone().into());
+        part.insert("toolName".to_string(), repaired.tool_name.clone().into());
         part.insert("input".to_string(), replay_input.clone());
     }
 
     // `content` carries the same call, and an approval request embeds it, so
-    // both must agree with `tool_calls` too. `raw` stays as is: it is the
+    // both must agree with `toolCalls` too. `raw` stays as is: it is the
     // provider's output, which a repair does not change.
     for entry in target
         .get_mut("content")
@@ -636,7 +643,7 @@ pub fn apply_tool_call_repair_to_result(
     }
     if repaired.tool_call_id != tool_call_id {
         for part in target
-            .get_mut("response_messages")
+            .get_mut("responseMessages")
             .and_then(Value::as_array_mut)
             .into_iter()
             .flatten()
diff --git a/aimux-core/src/recording.rs b/aimux-core/src/recording.rs
index c7282c597..785cf8df1 100644
--- a/aimux-core/src/recording.rs
+++ b/aimux-core/src/recording.rs
@@ -65,6 +65,7 @@ impl RecordingOperation {
 
 /// 一次完整调用的录制记录(三层 + call_id 关联)。
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct Recording {
     /// 格式版本。读取时只接受当前 [`RECORDING_SCHEMA`];旧 schema 的录制带有
@@ -99,9 +100,11 @@ pub struct Recording {
     /// 会话归组(RFC-0024 P3):所在会话 id。None = 未归组(无 session_id 且
     /// 推断关闭)。由 `Recorder::record_session` 填充,写入 InputRecord 之前
     /// 或之后均可(writer 端按 call_id 合并)。
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub session_id: Option<String>,
     /// 会话内步号(0 起,由 SessionStore 分配)。与 `session_id` 同生命周期。
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub step: Option<u32>,
 }
@@ -150,6 +153,7 @@ impl Recording {
 
 /// ① 输入侧:完整调用参数,足以重建 generate_text 调用。
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct InputRecord {
     /// Decision contract used for validation and deterministic offline replay.
@@ -197,6 +201,7 @@ impl InputRecord {
 /// 只记身份,不记配置:回放时按 `provider_id` + `model_id` 重新构建 model,
 /// 凭证与 base URL 只来自环境/registry(RFC-0036 §3.3)。
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ProviderRecord {
     /// provider 名(registry 名),回放重建的查找键,如 "openai"、"deepseek"。
@@ -230,10 +235,12 @@ impl ProviderRecord {
 
 /// ③ HTTP 侧:单次 attempt 的 wire 交换。
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct HttpExchange {
     /// Composite step this exchange belongs to (e.g. `router[0]:openai/gpt-4o`
     /// or `moa.ref[1]:...`). `None` for a plain, non-composite operation.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub step: Option<String>,
     /// Core operation attempt, starting at 1. Attempt numbers are unique
@@ -245,8 +252,12 @@ pub struct HttpExchange {
     pub exchange_index: u32,
     pub request: HttpRecord,
     /// None = 请求失败未获响应。
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response: Option<ResponseRecord>,
     pub timing: TimingRecord,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub error: Option<String>,
     /// 流式:该 exchange 是否已终结(收到 response 补全)。非流式恒 true。
     #[serde(default = "default_finalized")]
@@ -258,6 +269,7 @@ fn default_finalized() -> bool {
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct HttpRecord {
     pub method: String,
@@ -265,30 +277,39 @@ pub struct HttpRecord {
     /// 敏感头(authorization/cookie/含 api-key/key/x-amz-security-token 等)已脱敏为 "\[REDACTED\]"。
     pub headers: Vec<(String, String)>,
     /// 明文(脱敏后);None = 无 body。
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<String>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ResponseRecord {
     pub status: u16,
     pub headers: Vec<(String, String)>,
     /// 非流式:完整 JSON;流式:原始 SSE 拼接文本(上限截断)。
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub stream_chunks: Option<usize>,
     /// 首字节延迟(流式)。
+    #[ts(optional, type = "number")]
     #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[ts(type = "number | null")]
     pub ttfb_ms: Option<u64>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TimingRecord {
     /// ms 级延迟,TS 用 number(远小于 2^53,避免 bigint)。
     #[ts(type = "number")]
     pub latency_ms: u64,
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub ttfb_ms: Option<u64>,
 }
 
@@ -311,19 +332,27 @@ pub enum OutcomeStatus {
 
 /// 最终结果摘要。
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct OutcomeRecord {
     /// Normalized decision result, independent of a provider's wire format.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub decision_result: Option<serde_json::Value>,
     pub status: OutcomeStatus,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub finish_reason: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub error: Option<String>,
     /// Lossless structured domain error. In particular, `RetryError.errors`
     /// keeps the complete attempt history rather than only its display text.
-    #[serde(default)]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub error_value: Option<serde_json::Value>,
     /// 序列化的 Usage。
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub usage: Option<serde_json::Value>,
 }
 
@@ -1957,8 +1986,8 @@ mod tests {
         opts.abort_signal = None;
         let rec = InputRecord::from_call_options(&opts);
         let v = &rec.options;
-        assert!(v.get("call_id").is_none());
-        assert!(v.get("abort_signal").is_none());
+        assert!(v.get("callId").is_none());
+        assert!(v.get("abortSignal").is_none());
     }
 
     #[test]
@@ -1979,7 +2008,7 @@ mod tests {
             assert!(is_sensitive_key(key), "{key} must be redacted");
         }
         // Usage counters carry no credentials and must stay loggable.
-        for key in ["max_tokens", "input_tokens", "output_tokens", "model"] {
+        for key in ["maxTokens", "inputTokens", "outputTokens", "model"] {
             assert!(!is_sensitive_key(key), "{key} must stay visible");
         }
     }
@@ -1993,7 +2022,7 @@ mod tests {
                 "x-goog-api-key": "gkey",
                 "x-amz-security-token": "sts-tok"
             },
-            "provider_options": {
+            "providerOptions": {
                 "headers": { "Cookie": "s=1" },
                 "openai": {
                     "api_key": "sk-secret",
@@ -2003,9 +2032,9 @@ mod tests {
             },
             // 用量字段名含 "token" 子串,但非凭据——contains("token") 曾误伤。
             "usage": {
-                "max_output_tokens": 4096,
-                "prompt_tokens": 10,
-                "completion_tokens": 20
+                "maxOutputTokens": 4096,
+                "inputTokens": 10,
+                "outputTokens": 20
             },
             "temperature": 0.5,
         });
@@ -2014,17 +2043,14 @@ mod tests {
         assert_eq!(r["headers"]["x-goog-api-key"], "[REDACTED]");
         assert_eq!(r["headers"]["x-amz-security-token"], "[REDACTED]");
         assert_eq!(r["headers"]["X-Key"], "ok");
-        assert_eq!(r["provider_options"]["headers"]["Cookie"], "[REDACTED]");
-        assert_eq!(r["provider_options"]["openai"]["api_key"], "[REDACTED]");
-        assert_eq!(r["provider_options"]["openai"]["token"], "[REDACTED]");
-        assert_eq!(
-            r["provider_options"]["openai"]["access_token"],
-            "[REDACTED]"
-        );
+        assert_eq!(r["providerOptions"]["headers"]["Cookie"], "[REDACTED]");
+        assert_eq!(r["providerOptions"]["openai"]["api_key"], "[REDACTED]");
+        assert_eq!(r["providerOptions"]["openai"]["token"], "[REDACTED]");
+        assert_eq!(r["providerOptions"]["openai"]["access_token"], "[REDACTED]");
         // 含 "token" 子串的用量字段不再被误脱敏(回归 contains("token"))。
-        assert_eq!(r["usage"]["max_output_tokens"], 4096);
-        assert_eq!(r["usage"]["prompt_tokens"], 10);
-        assert_eq!(r["usage"]["completion_tokens"], 20);
+        assert_eq!(r["usage"]["maxOutputTokens"], 4096);
+        assert_eq!(r["usage"]["inputTokens"], 10);
+        assert_eq!(r["usage"]["outputTokens"], 20);
         assert_eq!(r["temperature"], 0.5);
     }
 
@@ -2228,7 +2254,7 @@ mod tests {
         assert_eq!(parsed.session_id.as_deref(), Some("sess-abc"));
         assert_eq!(parsed.step, Some(3));
         // 序列化时 null 省略:只有归组的调用才带这两个字段。
-        assert!(content.contains("\"session_id\":\"sess-abc\""));
+        assert!(content.contains("\"sessionId\":\"sess-abc\""));
         assert!(content.contains("\"step\":3"));
     }
 
@@ -2241,7 +2267,7 @@ mod tests {
             ProviderRecord::from_model("openai", "gpt-4o"),
         );
         let json = serde_json::to_string(&rec).unwrap();
-        assert!(!json.contains("session_id"), "ungrouped must omit: {json}");
+        assert!(!json.contains("sessionId"), "ungrouped must omit: {json}");
         assert!(!json.contains("\"step\""), "ungrouped must omit: {json}");
 
         // 旧 jsonl(无 session 字段)仍能反序列化(serde default)。
@@ -2249,7 +2275,7 @@ mod tests {
             let mut r = serde_json::to_value(&rec).unwrap();
             r.as_object_mut()
                 .unwrap()
-                .insert("transport_closed".into(), serde_json::json!(true));
+                .insert("transportClosed".into(), serde_json::json!(true));
             r
         })
         .unwrap();
@@ -2740,7 +2766,7 @@ mod tests {
         rec.record_transport_closed("c");
         assert!(rec.try_flush().is_ok());
         let content = std::fs::read_to_string(rec.path()).unwrap();
-        assert!(content.contains("\"call_id\":\"c\""));
+        assert!(content.contains("\"callId\":\"c\""));
         let _ = std::fs::remove_dir_all(&dir);
 
         // 非法目录 → new 降级(tx=None)→ try_flush 返回 WriterGone。
diff --git a/aimux-core/src/replay.rs b/aimux-core/src/replay.rs
index 934aadc83..1c11bb1ed 100644
--- a/aimux-core/src/replay.rs
+++ b/aimux-core/src/replay.rs
@@ -104,13 +104,13 @@ fn canonical_call_key(opts: &CallOptions) -> serde_json::Value {
     serde_json::json!({
         "prompt": serde_json::to_value(&opts.prompt).unwrap_or_default(),
         "temperature": opts.temperature,
-        "max_output_tokens": opts.max_output_tokens,
+        "maxOutputTokens": opts.max_output_tokens,
         "seed": opts.seed,
-        "response_format": serde_json::to_value(&opts.response_format).unwrap_or_default(),
+        "responseFormat": serde_json::to_value(&opts.response_format).unwrap_or_default(),
         "tools": serde_json::to_value(&opts.tools).unwrap_or_default(),
-        "tool_choice": serde_json::to_value(&opts.tool_choice).unwrap_or_default(),
+        "toolChoice": serde_json::to_value(&opts.tool_choice).unwrap_or_default(),
         "headers": serde_json::to_value(&opts.headers).unwrap_or_default(),
-        "provider_options": serde_json::to_value(&opts.provider_options).unwrap_or_default(),
+        "providerOptions": serde_json::to_value(&opts.provider_options).unwrap_or_default(),
     })
 }
 
@@ -120,13 +120,13 @@ fn canonical_recording_key(rec: &Recording) -> serde_json::Value {
     serde_json::json!({
         "prompt": serde_json::to_value(&rec.input.prompt).unwrap_or_default(),
         "temperature": o.get("temperature").cloned().unwrap_or_default(),
-        "max_output_tokens": o.get("max_output_tokens").cloned().unwrap_or_default(),
+        "maxOutputTokens": o.get("maxOutputTokens").cloned().unwrap_or_default(),
         "seed": o.get("seed").cloned().unwrap_or_default(),
-        "response_format": o.get("response_format").cloned().unwrap_or_default(),
+        "responseFormat": o.get("responseFormat").cloned().unwrap_or_default(),
         "tools": o.get("tools").cloned().unwrap_or_default(),
-        "tool_choice": o.get("tool_choice").cloned().unwrap_or_default(),
+        "toolChoice": o.get("toolChoice").cloned().unwrap_or_default(),
         "headers": o.get("headers").cloned().unwrap_or_default(),
-        "provider_options": o.get("provider_options").cloned().unwrap_or_default(),
+        "providerOptions": o.get("providerOptions").cloned().unwrap_or_default(),
     })
 }
 
@@ -180,7 +180,7 @@ pub(crate) fn redaction_aware_eq(rec: &serde_json::Value, req: &serde_json::Valu
 /// 被脱敏为 `"[REDACTED]"` 而请求值为数值,精确比较会 miss(保守:宁可 miss
 /// 不可误 hit,因为 max_output_tokens 影响响应,通配会引入伪命中)。
 fn canonical_keys_match(rec_key: &serde_json::Value, call_key: &serde_json::Value) -> bool {
-    const REDACTED_FIELDS: [&str; 2] = ["headers", "provider_options"];
+    const REDACTED_FIELDS: [&str; 2] = ["headers", "providerOptions"];
     let (ro, co) = match (rec_key.as_object(), call_key.as_object()) {
         (Some(a), Some(b)) => (a, b),
         _ => return rec_key == call_key,
diff --git a/aimux-core/src/reranking_model.rs b/aimux-core/src/reranking_model.rs
index cf677ebc6..e386c242a 100644
--- a/aimux-core/src/reranking_model.rs
+++ b/aimux-core/src/reranking_model.rs
@@ -15,6 +15,11 @@ use crate::{AbortSignal, retry, timeout};
 ///
 /// Aligned with V4 `RerankingModelV4CallOptions.documents`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum RerankingDocuments {
     /// A list of plain-text documents.
@@ -27,6 +32,7 @@ pub enum RerankingDocuments {
 ///
 /// Aligned with V4 `RerankingModelV4CallOptions`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct RerankingCallOptions {
     /// Documents to rerank.
@@ -36,6 +42,8 @@ pub struct RerankingCallOptions {
     pub query: String,
 
     /// Optional limit: return only the top `n` documents.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub top_n: Option<u32>,
 
     /// Abort signal for cancelling the operation.
@@ -44,15 +52,23 @@ pub struct RerankingCallOptions {
     pub abort_signal: Option<AbortSignal>,
 
     /// Per-call retry override. `None` uses the model default.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_retries: Option<u32>,
 
     /// Per-call operation timeout.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timeout: Option<crate::options::TimeoutConfiguration>,
 
     /// Additional provider-specific options, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 
     /// Additional HTTP headers to send with the request.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
 }
 
@@ -74,6 +90,7 @@ impl RerankingCallOptions {
 
 /// A single reranked entry: the original index and its relevance score.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct RerankingRank {
     /// The index of the document in the original list (before reranking).
@@ -86,6 +103,7 @@ pub struct RerankingRank {
 ///
 /// Aligned with V4 `RerankingModelV4Result`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct RerankingResult {
     /// Ordered list of reranked documents, sorted by descending relevance
@@ -94,28 +112,45 @@ pub struct RerankingResult {
     pub ranking: Vec<RerankingRank>,
 
     /// Additional provider-specific metadata, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
 
     /// Warnings for the call, e.g. unsupported settings.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub warnings: Option<Vec<Warning>>,
 
     /// Optional response information for debugging.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response: Option<RerankingResponse>,
 }
 
 /// Optional response information for a reranking call.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct RerankingResponse {
     /// ID for the generated response, if the provider sends one.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub id: Option<String>,
     /// Timestamp for the start of the generated response (ISO 8601 string).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timestamp: Option<String>,
     /// The ID of the model that was used to generate the response.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub model_id: Option<String>,
     /// Response headers.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
     /// Response body (opaque JSON).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<serde_json::Value>,
 }
 
diff --git a/aimux-core/src/response_messages.rs b/aimux-core/src/response_messages.rs
index c3239b6e3..6d89336f5 100644
--- a/aimux-core/src/response_messages.rs
+++ b/aimux-core/src/response_messages.rs
@@ -10,6 +10,7 @@ use base64::Engine;
 use serde_json::Value;
 
 use crate::content::ContentPart;
+use crate::language_model_message::ToolResultOutput;
 use crate::message::{MessageContent, ModelMessage, Role};
 use crate::result::{
     GenerateContent, GeneratedFile, ReasoningFileOutput, ReasoningOutput, ReasoningPart, Source,
@@ -254,11 +255,8 @@ impl ResponseMessageBuilder {
                 },
                 GenerateContent::ToolResult(result) => ContentPart::ToolResult {
                     tool_call_id: result.tool_call_id.clone(),
-                    tool_name: Some(result.tool_name.clone()),
-                    result: result.result.clone(),
-                    is_error: result.is_error,
-                    preliminary: result.preliminary,
-                    dynamic: result.dynamic,
+                    tool_name: result.tool_name.clone(),
+                    output: tool_model_output(&result.result, result.is_error == Some(true)),
                     provider_options: result.provider_metadata.clone(),
                 },
                 GenerateContent::Source(_) => continue,
@@ -280,6 +278,34 @@ impl ResponseMessageBuilder {
     }
 }
 
+/// TS: `createToolModelOutput` without a `toModelOutput` hook: errors become
+/// `error-text` with the error message, other strings `text`, the rest `json`.
+fn tool_model_output(result: &Value, is_error: bool) -> ToolResultOutput {
+    let provider_options = None;
+    match (result, is_error) {
+        (Value::String(value), false) => ToolResultOutput::Text {
+            value: value.clone(),
+            provider_options,
+        },
+        (_, false) => ToolResultOutput::Json {
+            value: result.clone(),
+            provider_options,
+        },
+        (Value::Null, true) => ToolResultOutput::ErrorText {
+            value: "unknown error".into(),
+            provider_options,
+        },
+        (Value::String(value), true) => ToolResultOutput::ErrorText {
+            value: value.clone(),
+            provider_options,
+        },
+        (_, true) => ToolResultOutput::ErrorText {
+            value: result.to_string(),
+            provider_options,
+        },
+    }
+}
+
 fn file_base64(data: &FileBytes) -> String {
     match data {
         FileBytes::Base64(data) => data.clone(),
diff --git a/aimux-core/src/result.rs b/aimux-core/src/result.rs
index 5f0d17c8b..a3cd58f55 100644
--- a/aimux-core/src/result.rs
+++ b/aimux-core/src/result.rs
@@ -15,11 +15,17 @@ use crate::types::{FinishReason, ProviderMetadata, Usage, Warning};
 
 /// A content item in the generation result.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "kebab-case",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum GenerateContent<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFile> {
     /// Generated text.
     Text {
         text: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -37,6 +43,7 @@ pub enum GenerateContent<C = RawToolCall, A = RawToolApprovalRequest, R = Genera
     /// Provider-specific content.
     Custom {
         kind: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -60,25 +67,31 @@ struct TextContentBinding;
 
 /// Provider approval request for a provider-executed call.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct RawToolApprovalRequest {
     pub approval_id: String,
     pub tool_call_id: String,
     #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub provider_metadata: Option<ProviderMetadata>,
 }
 
 /// Approval request exposed by text generation.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ToolApprovalRequestOutput {
     pub approval_id: String,
     pub tool_call: crate::tool::ToolCall,
     #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub reason: Option<String>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub is_automatic: Option<bool>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub signature: Option<String>,
 }
 
@@ -86,15 +99,21 @@ pub struct ToolApprovalRequestOutput {
 /// payload of both `GenerateContent::Source` and `StreamPart::Source`, and the
 /// element of the text results' `sources`.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "sourceType",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
-#[serde(tag = "source_type", rename_all = "lowercase")]
 pub enum Source {
     Url {
         id: String,
         url: String,
         #[serde(default, skip_serializing_if = "Option::is_none")]
+        #[ts(optional)]
         title: Option<String>,
         #[serde(default, skip_serializing_if = "Option::is_none")]
+        #[ts(optional)]
         provider_metadata: Option<ProviderMetadata>,
     },
     Document {
@@ -102,8 +121,10 @@ pub enum Source {
         media_type: String,
         title: String,
         #[serde(default, skip_serializing_if = "Option::is_none")]
+        #[ts(optional)]
         filename: Option<String>,
         #[serde(default, skip_serializing_if = "Option::is_none")]
+        #[ts(optional)]
         provider_metadata: Option<ProviderMetadata>,
     },
 }
@@ -113,11 +134,13 @@ pub enum Source {
 /// or URL. The payload of both `GenerateContent::File` and
 /// `StreamPart::File`, and the element of the text results' `files`.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct GeneratedFile {
     pub data: GeneratedFileData,
     pub media_type: String,
     #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub provider_metadata: Option<ProviderMetadata>,
 }
 
@@ -127,26 +150,33 @@ pub struct GeneratedFile {
 /// (`providerMetadata.anthropic.signature`). The payload of
 /// `GenerateContent::Reasoning` and the element of the text results' `reasoning`.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ReasoningOutput {
     pub text: String,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub provider_metadata: Option<ProviderMetadata>,
 }
 
 /// Text or a generated file produced during reasoning.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
-#[serde(untagged)]
+#[serde(tag = "type")]
 #[ts(export)]
 pub enum ReasoningPart {
+    #[serde(rename = "reasoning")]
     Text(ReasoningOutput),
+    #[serde(rename = "reasoning-file")]
     File(ReasoningFileOutput),
 }
 
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ReasoningFileOutput {
     pub file: GeneratedFile,
     #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub provider_metadata: Option<ProviderMetadata>,
 }
 
@@ -154,6 +184,7 @@ pub struct ReasoningFileOutput {
 /// `GenerateTextResult`'s user-facing fields (without `raw`, since streaming
 /// has no `GenerateResult` equivalent).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct StreamTextResultAggregated {
     /// Ordered content from the last generation step.
@@ -179,7 +210,8 @@ pub struct StreamTextResultAggregated {
     /// Why generation stopped.
     pub finish_reason: crate::types::FinishReason,
     /// Raw provider-specific finish reason string.
-    #[serde(default)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub raw_finish_reason: Option<String>,
     /// Token usage.
     #[serde(default)]
@@ -192,7 +224,8 @@ pub struct StreamTextResultAggregated {
     #[serde(default)]
     pub warnings: Vec<crate::types::Warning>,
     /// Provider-specific metadata from the Finish chunk.
-    #[serde(default)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub provider_metadata: Option<ProviderMetadata>,
     /// Request information from the provider.
     #[serde(default)]
@@ -207,6 +240,7 @@ pub struct StreamTextResultAggregated {
 
 /// Result of `LanguageModel::do_generate` (non-streaming).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct GenerateResult {
     /// Generated content items (text, tool calls, etc.).
@@ -218,10 +252,16 @@ pub struct GenerateResult {
     /// Warnings issued by the provider.
     pub warnings: Vec<Warning>,
     /// Provider-specific metadata.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub provider_metadata: Option<ProviderMetadata>,
     /// Optional request information for telemetry and debugging.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub request: Option<RequestInfo>,
     /// Optional response metadata, headers, and body.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
     pub response: Option<ResponseInfo>,
 }
 
diff --git a/aimux-core/src/search_model.rs b/aimux-core/src/search_model.rs
index 4c94c65dd..4f1ee9b71 100644
--- a/aimux-core/src/search_model.rs
+++ b/aimux-core/src/search_model.rs
@@ -16,28 +16,39 @@ use crate::{AbortSignal, retry, timeout};
 
 /// Options passed to [`SearchModel::do_search`].
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SearchCallOptions {
     /// The search query string.
     pub query: String,
 
     /// Maximum number of results to return.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_results: Option<u32>,
 
     /// Whether to include raw page content in results.
     ///
     /// Provider support varies; providers that cannot honor this should
     /// issue a warning rather than erroring.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub include_raw_content: Option<bool>,
 
     /// Optional time range filter (e.g. `"day"`, `"week"`, `"month"`,
     /// `"year"`). Provider support varies.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub time_range: Option<String>,
 
     /// Optional list of domains to include in results.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub include_domains: Option<Vec<String>>,
 
     /// Optional list of domains to exclude from results.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub exclude_domains: Option<Vec<String>>,
 
     /// Abort signal for cancelling the operation.
@@ -46,15 +57,23 @@ pub struct SearchCallOptions {
     pub abort_signal: Option<AbortSignal>,
 
     /// Per-call retry override. `None` uses the model default.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_retries: Option<u32>,
 
     /// Per-call operation timeout.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timeout: Option<crate::options::TimeoutConfiguration>,
 
     /// Additional provider-specific options, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 
     /// Additional HTTP headers to send with the request.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
 }
 
@@ -79,30 +98,44 @@ impl SearchCallOptions {
 
 /// A single search result item.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SearchResultItem {
     /// The title of the result (e.g. page title).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub title: Option<String>,
 
     /// The URL of the result.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub url: Option<String>,
 
     /// A snippet/summary of the result content.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub content: Option<String>,
 
     /// Raw page content (when `include_raw_content` is requested and
     /// supported by the provider).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub raw_content: Option<String>,
 
     /// A relevance score (0.0–1.0) if the provider returns one.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub score: Option<f64>,
 
     /// Provider-specific metadata for this result.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
 }
 
 /// The result of [`SearchModel::do_search`].
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SearchResult {
     /// Ordered list of search results.
@@ -110,25 +143,36 @@ pub struct SearchResult {
 
     /// An optional direct answer / summary (some providers return
     /// an AI-generated answer alongside results).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub answer: Option<String>,
 
     /// Additional provider-specific metadata.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
 
     /// Warnings for the call.
     pub warnings: Vec<Warning>,
 
     /// Optional response information for debugging.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response: Option<SearchResponse>,
 }
 
 /// Optional response information for a search call.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SearchResponse {
     /// Response headers.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
     /// The response body (opaque JSON).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<serde_json::Value>,
 }
 
diff --git a/aimux-core/src/session.rs b/aimux-core/src/session.rs
index 38a3869a2..6b37d46dc 100644
--- a/aimux-core/src/session.rs
+++ b/aimux-core/src/session.rs
@@ -44,6 +44,7 @@ pub enum SessionSource {
 /// A single call within a session (an index entry only — call content lives
 /// in Recording (RFC-0023) / TraceRecord (RFC-0015)).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SessionCall {
     /// Call-level unique id. This is the association key for the future
@@ -58,6 +59,7 @@ pub struct SessionCall {
 
 /// Aggregated view of one session (session_id → ordered calls).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SessionView {
     pub session_id: String,
diff --git a/aimux-core/src/shared.rs b/aimux-core/src/shared.rs
index 24cddc49d..0b8218e24 100644
--- a/aimux-core/src/shared.rs
+++ b/aimux-core/src/shared.rs
@@ -82,6 +82,7 @@ pub type SharedProviderReference = HashMap<String, String>;
 /// the upstream API returns base64, return [`FileBytes::Base64`]; if it
 /// returns binary, return [`FileBytes::Binary`].
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(untagged)]
 #[ts(export)]
 pub enum FileBytes {
     /// Raw binary bytes.
@@ -94,6 +95,11 @@ pub enum FileBytes {
 ///
 /// Aligned with V4 `SharedV4FileData`.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum FileData {
     /// Raw bytes (`Uint8Array`) or a base64-encoded string.
@@ -101,6 +107,7 @@ pub enum FileData {
     /// A URL that points to the file.
     Url {
         url: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         original_url: Option<String>,
     },
@@ -112,6 +119,11 @@ pub enum FileData {
 
 /// Data or a URL returned for a generated file.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum GeneratedFileData {
     Data {
@@ -119,6 +131,7 @@ pub enum GeneratedFileData {
     },
     Url {
         url: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         original_url: Option<String>,
     },
@@ -130,7 +143,8 @@ pub enum GeneratedFileData {
 /// `resolution` options. Aligned with the TS template-literal type
 /// `` `${number}x${number}` ``.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, TS)]
-#[ts(export)]
+#[serde(into = "String", try_from = "String")]
+#[ts(export, type = "`${number}x${number}`")]
 pub struct Size {
     width: u32,
     height: u32,
@@ -181,6 +195,19 @@ impl std::fmt::Display for Size {
     }
 }
 
+impl From<Size> for String {
+    fn from(value: Size) -> Self {
+        value.to_string()
+    }
+}
+
+impl TryFrom<String> for Size {
+    type Error = String;
+    fn try_from(value: String) -> Result<Self, Self::Error> {
+        Self::parse(&value)
+    }
+}
+
 impl std::str::FromStr for Size {
     type Err = String;
     fn from_str(s: &str) -> Result<Self, Self::Err> {
@@ -194,7 +221,8 @@ impl std::str::FromStr for Size {
 /// `aspectRatio` options. Aligned with the TS template-literal type
 /// `` `${number}:${number}` ``.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, TS)]
-#[ts(export)]
+#[serde(into = "String", try_from = "String")]
+#[ts(export, type = "`${number}:${number}`")]
 pub struct AspectRatio {
     width: u32,
     height: u32,
@@ -246,6 +274,19 @@ impl std::fmt::Display for AspectRatio {
     }
 }
 
+impl From<AspectRatio> for String {
+    fn from(value: AspectRatio) -> Self {
+        value.to_string()
+    }
+}
+
+impl TryFrom<String> for AspectRatio {
+    type Error = String;
+    fn try_from(value: String) -> Result<Self, Self::Error> {
+        Self::parse(&value)
+    }
+}
+
 impl std::str::FromStr for AspectRatio {
     type Err = String;
     fn from_str(s: &str) -> Result<Self, Self::Err> {
@@ -269,18 +310,28 @@ fn parse_pair(s: &str, sep: char) -> Option<(u32, u32)> {
 /// `ResponseInfo` directly, while those that mark it optional use
 /// `Option<ResponseInfo>`.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ResponseInfo {
     /// ID of the generated response, when supplied by the provider.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub id: Option<String>,
     /// Timestamp for the start of the generated response (ISO 8601 string).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timestamp: Option<String>,
     /// The ID of the model that was used to generate the response.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub model_id: Option<String>,
     /// Response headers.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
     /// The response body (opaque JSON).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<Value>,
 }
 
@@ -288,16 +339,22 @@ pub struct ResponseInfo {
 ///
 /// Aligned with the `request?` object that appears on some V4 model results.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct RequestInfo {
     /// The request body that was sent (opaque JSON).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<Value>,
 }
 
 /// Response information returned when starting a language-model stream.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct StreamResponseInfo {
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
 }
 
diff --git a/aimux-core/src/skills_model.rs b/aimux-core/src/skills_model.rs
index deb5171c0..c913fb17c 100644
--- a/aimux-core/src/skills_model.rs
+++ b/aimux-core/src/skills_model.rs
@@ -11,6 +11,7 @@ use crate::shared::{
 };
 
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SkillFile {
     pub path: String,
@@ -18,21 +19,37 @@ pub struct SkillFile {
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct UploadSkillCallOptions {
     pub files: Vec<SkillFile>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub display_title: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct UploadSkillResult {
     pub provider_reference: SharedProviderReference,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub display_title: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub name: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub description: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub latest_version: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
     pub warnings: Vec<Warning>,
 }
diff --git a/aimux-core/src/speech_model.rs b/aimux-core/src/speech_model.rs
index d83d38c8f..30512a840 100644
--- a/aimux-core/src/speech_model.rs
+++ b/aimux-core/src/speech_model.rs
@@ -16,6 +16,7 @@ use crate::{AbortSignal, retry, timeout};
 /// The TS result type is `string | Uint8Array`. Providers should return data
 /// without unnecessary conversion.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(untagged)]
 #[ts(export)]
 pub enum AudioData {
     /// Audio as a base64-encoded string.
@@ -28,6 +29,7 @@ pub enum AudioData {
 ///
 /// Aligned with V4 `SpeechModelV4CallOptions`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SpeechCallOptions {
     /// Text to convert to speech.
@@ -35,23 +37,35 @@ pub struct SpeechCallOptions {
 
     /// The voice to use (provider-specific voice ID, name, or other
     /// identifier).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub voice: Option<String>,
 
     /// The desired output format for the audio, e.g. `"mp3"`, `"wav"`.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub output_format: Option<String>,
 
     /// Instructions for the speech generation, e.g.
     /// `"Speak in a slow and steady tone"`.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub instructions: Option<String>,
 
     /// The speed of the speech generation.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub speed: Option<f64>,
 
     /// The language for speech generation (ISO 639-1 code, e.g. `"en"`, or
     /// `"auto"` for automatic detection). Provider support varies.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub language: Option<String>,
 
     /// Additional provider-specific options, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 
     /// Abort signal for cancelling the operation.
@@ -60,12 +74,18 @@ pub struct SpeechCallOptions {
     pub abort_signal: Option<AbortSignal>,
 
     /// Per-call retry override. `None` uses the model default.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_retries: Option<u32>,
 
     /// Per-call operation timeout.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timeout: Option<crate::options::TimeoutConfiguration>,
 
     /// Additional HTTP headers to send with the request.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
 }
 
@@ -92,6 +112,7 @@ impl SpeechCallOptions {
 ///
 /// Aligned with V4 `SpeechModelV4Result`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SpeechResult {
     /// Generated audio (base64 string or binary data).
@@ -101,34 +122,50 @@ pub struct SpeechResult {
     pub warnings: Vec<Warning>,
 
     /// Optional request information for telemetry and debugging.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub request: Option<SpeechRequest>,
 
     /// Response information for telemetry and debugging.
     pub response: SpeechResponse,
 
     /// Additional provider-specific metadata, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
 }
 
 /// Optional request information for a speech call.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SpeechRequest {
     /// Response body (HTTP providers only).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<serde_json::Value>,
 }
 
 /// Response information for a speech call.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SpeechResponse {
     /// Timestamp for the start of the generated response (ISO 8601 string).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timestamp: Option<String>,
     /// The ID of the model that was used to generate the response.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub model_id: Option<String>,
     /// Response headers.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
     /// Response body (opaque JSON).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<serde_json::Value>,
 }
 
diff --git a/aimux-core/src/stream_part.rs b/aimux-core/src/stream_part.rs
index 16c5bbb16..1e226b2bb 100644
--- a/aimux-core/src/stream_part.rs
+++ b/aimux-core/src/stream_part.rs
@@ -20,12 +20,18 @@ use crate::types::{FinishReason, ProviderMetadata, ResponseMetadata, Usage, Warn
 /// [`StreamPart`] (raw argument text); `stream_text` emits [`TextStreamPart`]
 /// (parsed and validated [`ToolCall`]).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "kebab-case",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export, concrete(C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFile))]
 pub enum StreamPart<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFile> {
     // ── P0: text ──
     /// Start of a text segment.
     TextStart {
         id: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -33,12 +39,14 @@ pub enum StreamPart<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFi
     TextDelta {
         id: String,
         delta: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
     /// End of a text segment.
     TextEnd {
         id: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -52,6 +60,8 @@ pub enum StreamPart<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFi
     Finish {
         finish_reason: FinishReason,
         usage: Usage,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
     /// An error occurred mid-stream.
@@ -65,14 +75,18 @@ pub enum StreamPart<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFi
         id: String,
         tool_name: String,
         /// Whether the tool call will be executed by the provider.
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_executed: Option<bool>,
         /// Whether the tool is dynamic (defined at runtime, e.g. MCP tools).
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         dynamic: Option<bool>,
         /// Optional title for the tool call.
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         title: Option<String>,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -80,12 +94,14 @@ pub enum StreamPart<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFi
     ToolInputDelta {
         id: String,
         delta: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
     /// End of a tool call's input streaming.
     ToolInputEnd {
         id: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -103,6 +119,7 @@ pub enum StreamPart<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFi
     ReasoningFile(R),
     Custom {
         kind: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -113,6 +130,8 @@ pub enum StreamPart<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFi
         id: String,
         /// Provider-specific metadata (e.g. xAI `itemId`).
         #[allow(unused)]
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
     ReasoningDelta {
@@ -120,12 +139,16 @@ pub enum StreamPart<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFi
         delta: String,
         /// Provider-specific metadata (e.g. xAI `itemId`).
         #[allow(unused)]
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
     ReasoningEnd {
         id: String,
         /// Provider-specific metadata (e.g. xAI `itemId`, `reasoningEncryptedContent`).
         #[allow(unused)]
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
 
@@ -144,12 +167,18 @@ pub enum StreamPart<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFi
 
 /// A public call-layer stream chunk, with parsed tool calls.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "kebab-case",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum TextStreamPart {
     // ── P0: text ──
     /// Start of a text segment.
     TextStart {
         id: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -157,12 +186,14 @@ pub enum TextStreamPart {
     TextDelta {
         id: String,
         delta: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
     /// End of a text segment.
     TextEnd {
         id: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -176,12 +207,16 @@ pub enum TextStreamPart {
     Finish {
         finish_reason: FinishReason,
         usage: Usage,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
     /// A completed model call, including its response metadata.
     FinishStep {
         finish_reason: FinishReason,
         usage: Usage,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
         response: crate::shared::ResponseInfo,
     },
@@ -196,14 +231,18 @@ pub enum TextStreamPart {
         id: String,
         tool_name: String,
         /// Whether the tool call will be executed by the provider.
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_executed: Option<bool>,
         /// Whether the tool is dynamic (defined at runtime, e.g. MCP tools).
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         dynamic: Option<bool>,
         /// Optional title for the tool call.
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         title: Option<String>,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -211,12 +250,14 @@ pub enum TextStreamPart {
     ToolInputDelta {
         id: String,
         delta: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
     /// End of a tool call's input streaming.
     ToolInputEnd {
         id: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -234,6 +275,7 @@ pub enum TextStreamPart {
     ReasoningFile(ReasoningFileOutput),
     Custom {
         kind: String,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
@@ -244,6 +286,8 @@ pub enum TextStreamPart {
         id: String,
         /// Provider-specific metadata (e.g. xAI `itemId`).
         #[allow(unused)]
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
     ReasoningDelta {
@@ -251,12 +295,16 @@ pub enum TextStreamPart {
         delta: String,
         /// Provider-specific metadata (e.g. xAI `itemId`).
         #[allow(unused)]
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
     ReasoningEnd {
         id: String,
         /// Provider-specific metadata (e.g. xAI `itemId`, `reasoningEncryptedContent`).
         #[allow(unused)]
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<ProviderMetadata>,
     },
 
diff --git a/aimux-core/src/tool.rs b/aimux-core/src/tool.rs
index 2c95fb05c..e345a8fcf 100644
--- a/aimux-core/src/tool.rs
+++ b/aimux-core/src/tool.rs
@@ -10,31 +10,37 @@ use crate::types::ProviderMetadata;
 
 /// A tool definition passed to the model in `CallOptions.tools`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct FunctionTool {
     /// Unique tool name.
     pub name: String,
     /// Human-readable description.
-    #[serde(skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub description: Option<String>,
     /// JSON Schema describing the tool's parameters.
     pub input_schema: Value,
     /// Whether the tool supports strict schema enforcement (OpenAI strict mode).
-    #[serde(skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub strict: Option<bool>,
     /// Provider-specific options, keyed by provider name
     /// (e.g. `{"anthropic": {"eagerInputStreaming": true}}`). Aligned with the
     /// V4 `providerOptions` field on function tools.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
     /// Example inputs for the tool (V4 `inputExamples`), used by some providers
     /// to emit `input_examples` in the request body.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub input_examples: Option<Vec<FunctionToolInputExample>>,
 }
 
 /// An example input for a function tool.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct FunctionToolInputExample {
     pub input: Map<String, Value>,
@@ -77,6 +83,7 @@ impl FunctionTool {
 /// schemas are defined by the provider, and some are executed on the provider's
 /// servers (e.g. web search, code execution).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ProviderTool {
     /// The ID of the tool, following the format `<provider-id>.<unique-tool-name>`.
@@ -92,7 +99,7 @@ pub struct ProviderTool {
 /// provider-defined tool. Mirrors the V4
 /// `Array<LanguageModelV4FunctionTool | LanguageModelV4ProviderTool>`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
-#[serde(tag = "type", rename_all = "snake_case")]
+#[serde(tag = "type", rename_all = "lowercase")]
 #[ts(export)]
 pub enum Tool {
     /// A user-defined function tool.
@@ -112,21 +119,26 @@ impl From<FunctionTool> for Tool {
 /// The wire shape matches [`ToolCall`] field for field, except that `input`
 /// is the provider's raw argument *text* rather than a parsed value.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct RawToolCall {
     pub tool_call_id: String,
     pub tool_name: String,
     pub input: String,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_executed: Option<bool>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub dynamic: Option<bool>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<ProviderMetadata>,
 }
 
 /// A tool call requested by the model.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ToolCall {
     /// Provider-assigned call id.
@@ -137,19 +149,24 @@ pub struct ToolCall {
     /// provider input was not valid JSON.
     pub input: Value,
     /// Whether the tool call will be executed by the provider.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_executed: Option<bool>,
     /// Whether the tool is dynamic (defined at runtime, e.g. MCP tools).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub dynamic: Option<bool>,
     /// Additional provider-specific metadata associated with this call.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<ProviderMetadata>,
     /// Set when lookup, JSON parsing, or schema validation still failed after
     /// the optional repair attempt.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub invalid: Option<bool>,
     /// Typed failure associated with an invalid tool call.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub error: Option<AiMuxError>,
 }
@@ -157,6 +174,7 @@ pub struct ToolCall {
 /// A tool result produced by a provider-executed tool (e.g. xAI `file_search`,
 /// web search), carried next to the [`RawToolCall`] it answers.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ToolResult {
     /// Must match the corresponding call's `tool_call_id`.
@@ -165,15 +183,19 @@ pub struct ToolResult {
     /// The tool's output (usually a JSON-serializable value or plain text).
     pub result: Value,
     /// Whether the result is an error or error message.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub is_error: Option<bool>,
     /// Whether the result is preliminary (replaces prior, e.g. image previews).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub preliminary: Option<bool>,
     /// Whether the tool is dynamic (defined at runtime, e.g. MCP tools).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub dynamic: Option<bool>,
     /// Additional provider-specific metadata for the tool result.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<ProviderMetadata>,
 }
diff --git a/aimux-core/src/trace/fingerprint.rs b/aimux-core/src/trace/fingerprint.rs
index 516fc511f..04652aee1 100644
--- a/aimux-core/src/trace/fingerprint.rs
+++ b/aimux-core/src/trace/fingerprint.rs
@@ -126,6 +126,7 @@ fn denoise_at(value: &Value, depth: usize) -> Value {
 /// Serialized-wire fingerprint stored in a `TraceRecord` (hex strings —
 /// `u128` is not JSON-serializable).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct Fingerprint {
     /// 128-bit hash of the whole denoised body (hex).
@@ -240,6 +241,6 @@ mod tests {
         assert_eq!(f.body_hash.len(), 32); // 128-bit hex
         assert!(f.block_hashes.iter().all(|h| h.len() == 32));
         let json = serde_json::to_string(&f).unwrap();
-        assert!(json.contains("\"body_hash\":\""));
+        assert!(json.contains("\"bodyHash\":\""));
     }
 }
diff --git a/aimux-core/src/trace/record.rs b/aimux-core/src/trace/record.rs
index a41834d3e..d6c3986ee 100644
--- a/aimux-core/src/trace/record.rs
+++ b/aimux-core/src/trace/record.rs
@@ -13,23 +13,40 @@ use super::verdict::Verdict;
 
 /// Token usage snapshot (7 flat fields + raw passthrough).
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct UsageSnapshot {
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub input_total: Option<u64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub input_no_cache: Option<u64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub cache_read: Option<u64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub cache_write: Option<u64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub output_total: Option<u64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub output_text: Option<u64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub output_reasoning: Option<u64>,
     /// Raw provider usage payload (opaque passthrough; numbers only — no
     /// prompt text).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub raw: Option<Value>,
 }
 
 /// Request-side cache hints (best-effort, e.g. Anthropic `cache_control`).
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct RequestCacheHints {
     /// Whether any message requested `cache_control` (write) this call.
@@ -38,13 +55,18 @@ pub struct RequestCacheHints {
 
 /// One probed call. Plaintext never persists — fingerprints only.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TraceRecord {
     pub provider: String,
     pub model: String,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub request_id: Option<String>,
     /// From `CallOptions.session_id` (RFC-0024) — explicit first, inference
     /// fallback; may also be the layer's default session.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub session_id: Option<String>,
     /// Call-level unique id (association key for Recording / replay,
     /// RFC-0023).
@@ -60,23 +82,29 @@ pub struct TraceRecord {
     /// Client-side LCP token upper bound (block upper bound, byte-proxy
     /// len/4). `None` when no history matched. Consumed by `TraceStats`
     /// aggregation.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub lcp_token_upper: Option<u64>,
     /// Time to first streamed token (ms). Non-streaming: `None`.
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub ttft_ms: Option<u64>,
     pub fingerprint: Fingerprint,
     pub usage: UsageSnapshot,
     /// Response-side cache headers (e.g. `x-openrouter-cache-status`).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response_cache_headers: Option<HashMap<String, String>>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub request_cache_hints: Option<RequestCacheHints>,
     /// Audit verdict (`None` when no auditor is attached — the default).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub verdict: Option<Verdict>,
     /// Error string when the call failed (still recorded: failures are part
     /// of the session).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub error: Option<String>,
     /// Internal scope key (index only; never crosses the wire contract).
diff --git a/aimux-core/src/trace/store.rs b/aimux-core/src/trace/store.rs
index 09685cd73..2a329a22e 100644
--- a/aimux-core/src/trace/store.rs
+++ b/aimux-core/src/trace/store.rs
@@ -280,17 +280,27 @@ impl TraceStore {
 
 /// Query filter (RFC-0015 §5.3).
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TraceFilter {
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub model: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub session_id: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub since_unix_ms: Option<i64>,
 }
 
 /// Aggregated statistics (RFC-0015 §5.2 — the two hit rates are reported
 /// side by side, never merged into a single number).
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TraceStats {
     pub provider: String,
@@ -299,16 +309,25 @@ pub struct TraceStats {
     pub input_tokens_total: u64,
     pub claimed_cache_read_total: u64,
     pub claimed_cache_write_total: u64,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub reported_hit_rate: Option<f64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub client_upper_bound_hit_rate: Option<f64>,
     pub verdict_counts: BTreeMap<String, u64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub ttft_p50_ms: Option<u64>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub ttft_p95_ms: Option<u64>,
     pub errors: u64,
 }
 
 /// Session chain view (append-only order + prefix stability, RFC-0015 §5.3).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SessionChainView {
     pub session_id: String,
@@ -319,6 +338,7 @@ pub struct SessionChainView {
 
 /// One step of a session's cache-hit trajectory (RFC-0024 §4.3).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct SessionStepStat {
     /// Step within the session (0-based, session order).
@@ -327,18 +347,22 @@ pub struct SessionStepStat {
     pub call_id: String,
     /// Reported hit rate for this step (`cache_read / input_total`); `None`
     /// when the call carried no input/cache-read usage.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub hit_rate: Option<f64>,
     /// Audit verdict (present only when an auditor is attached).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub verdict: Option<Verdict>,
     /// Error string for failed calls (failures are part of the session).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub error: Option<String>,
 }
 
 /// A prefix break between two consecutive records of a session.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct PrefixBreak {
     pub at_record_id: String,
diff --git a/aimux-core/src/trace/verdict.rs b/aimux-core/src/trace/verdict.rs
index d23939e7a..029dcecc6 100644
--- a/aimux-core/src/trace/verdict.rs
+++ b/aimux-core/src/trace/verdict.rs
@@ -26,6 +26,7 @@ pub enum VerdictConfidence {
 
 /// A verdict for one call.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct Verdict {
     pub kind: VerdictKind,
@@ -59,11 +60,14 @@ impl Verdict {
 /// Provider audit parameters (RFC-0015 §7 matrix — the part that lives in
 /// core; display logic belongs to the CLI).
 #[derive(Debug, Clone, Copy, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ProviderAuditSpec {
     /// Provider family for family-specific invariants.
     pub family: ProviderFamily,
     /// Token quantization granularity; `None` = no quantization.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub gran: Option<u64>,
     /// Minimum cache threshold (tokens): below it `claimed` must be 0.
     pub threshold: u64,
@@ -74,6 +78,8 @@ pub struct ProviderAuditSpec {
     pub model56plus: bool,
     /// Claimed-hit ceiling: above this multiple of the client upper bound
     /// the verdict is overclaim (W) even in shared mode. `None` = use U+τ.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub shared_ceiling_mult: Option<f64>,
 }
 
diff --git a/aimux-core/src/transcription_model.rs b/aimux-core/src/transcription_model.rs
index 516d345cf..bac827ae4 100644
--- a/aimux-core/src/transcription_model.rs
+++ b/aimux-core/src/transcription_model.rs
@@ -22,6 +22,7 @@ use crate::{AbortSignal, retry, timeout};
 
 /// Audio input: raw bytes or a base64-encoded string.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(untagged)]
 #[ts(export)]
 pub enum AudioInput {
     /// Raw binary bytes.
@@ -32,6 +33,7 @@ pub enum AudioInput {
 
 /// A chunk of audio in a streaming transcription request.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(untagged)]
 #[ts(export)]
 pub enum AudioChunk {
     /// Raw binary bytes.
@@ -42,6 +44,7 @@ pub enum AudioChunk {
 
 /// A transcript segment with timing information.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TranscriptionSegment {
     /// The text content of this segment.
@@ -56,6 +59,7 @@ pub struct TranscriptionSegment {
 ///
 /// Aligned with V4 `TranscriptionModelV4CallOptions`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TranscriptionCallOptions {
     /// Audio data to transcribe (raw bytes or a base64-encoded string).
@@ -65,6 +69,8 @@ pub struct TranscriptionCallOptions {
     pub media_type: String,
 
     /// Additional provider-specific options, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_options: Option<SharedProviderOptions>,
 
     /// Abort signal for cancelling the operation.
@@ -73,12 +79,18 @@ pub struct TranscriptionCallOptions {
     pub abort_signal: Option<AbortSignal>,
 
     /// Per-call retry override. `None` uses the model default.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_retries: Option<u32>,
 
     /// Per-call operation timeout.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timeout: Option<crate::options::TimeoutConfiguration>,
 
     /// Additional HTTP headers to send with the request.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
 }
 
@@ -101,6 +113,7 @@ impl TranscriptionCallOptions {
 ///
 /// Aligned with V4 `TranscriptionModelV4Result`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TranscriptionResult {
     /// The complete transcribed text from the audio.
@@ -110,53 +123,76 @@ pub struct TranscriptionResult {
     pub segments: Vec<TranscriptionSegment>,
 
     /// The detected language (ISO 639-1 code, e.g. `"en"`), if detected.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub language: Option<String>,
 
     /// The total duration of the audio in seconds, if determined.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub duration_in_seconds: Option<f64>,
 
     /// Warnings for the call, e.g. unsupported settings.
     pub warnings: Vec<Warning>,
 
     /// Optional request information for telemetry and debugging.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub request: Option<TranscriptionRequest>,
 
     /// Response information for telemetry and debugging.
     pub response: TranscriptionResponse,
 
     /// Additional provider-specific metadata, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
 }
 
 /// Optional request information for a transcription call.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TranscriptionRequest {
     /// Raw request HTTP body that was sent (JSON stringified).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<String>,
 }
 
 /// Response information for a transcription call.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct TranscriptionResponse {
     /// Timestamp for the start of the generated response (ISO 8601 string).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timestamp: Option<String>,
     /// The ID of the model that was used to generate the response.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub model_id: Option<String>,
     /// Response headers.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
     /// Response body (opaque JSON).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub body: Option<serde_json::Value>,
 }
 
 /// The input audio format for a streaming transcription request.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct InputAudioFormat {
     /// Audio format type, e.g. `"audio/pcm"`, `"audio/pcmu"`, `"audio/pcma"`.
     pub format_type: String,
     /// Sample rate in Hz. Only applicable for formats that require a rate.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub rate: Option<u32>,
 }
 
@@ -208,47 +244,90 @@ impl std::fmt::Debug for TranscriptionStreamOptions {
 ///
 /// Aligned with V4 `TranscriptionModelV4StreamPart`.
 #[derive(Debug, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "kebab-case",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum TranscriptionStreamPart {
     /// Stream start event, carrying warnings.
     StreamStart { warnings: Vec<Warning> },
     /// Append-only transcript delta.
     TranscriptDelta {
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         id: Option<String>,
         delta: String,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<SharedProviderMetadata>,
     },
     /// Non-final transcript text (may be revised by later parts).
     TranscriptPartial {
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         id: Option<String>,
         text: String,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         start_second: Option<f64>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         duration_in_seconds: Option<f64>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         channel_index: Option<u32>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<SharedProviderMetadata>,
     },
     /// Final transcript text for a provider-defined segment or utterance.
     TranscriptFinal {
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         id: Option<String>,
         text: String,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         start_second: Option<f64>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         end_second: Option<f64>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         channel_index: Option<u32>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<SharedProviderMetadata>,
     },
     /// Response metadata, emitted once available.
     ResponseMetadata {
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         timestamp: Option<String>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         model_id: Option<String>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         headers: Option<SharedHeaders>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         body: Option<serde_json::Value>,
     },
     /// Metadata available after the stream finishes.
     Finish {
         text: String,
         segments: Vec<TranscriptionSegment>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         language: Option<String>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         duration_in_seconds: Option<f64>,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_metadata: Option<SharedProviderMetadata>,
     },
     /// Raw provider chunk (only when `include_raw_chunks` is `true`).
diff --git a/aimux-core/src/types.rs b/aimux-core/src/types.rs
index 0f9e54730..d6c0a5505 100644
--- a/aimux-core/src/types.rs
+++ b/aimux-core/src/types.rs
@@ -6,11 +6,14 @@ use ts_rs::TS;
 
 /// Unified finish reason (why the model stopped generating).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct FinishReason {
     /// The unified reason.
     pub unified: FinishReasonUnified,
     /// The raw provider-specific reason string.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub raw: Option<String>,
 }
 
@@ -29,43 +32,55 @@ pub enum FinishReasonUnified {
 
 /// Token usage statistics.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct Usage {
     pub input_tokens: InputTokenUsage,
     pub output_tokens: OutputTokenUsage,
     /// Raw usage information from the provider (opaque, provider-specific).
+    #[ts(optional, type = "Record<string, unknown>")]
     #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[ts(type = "Record<string, unknown> | null")]
     pub raw: Option<serde_json::Map<String, Value>>,
 }
 
 /// Input token usage detail.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct InputTokenUsage {
     /// Total tokens.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub total: Option<u32>,
     /// Tokens that were not served from cache.
-    #[serde(skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub no_cache: Option<u32>,
     /// Tokens served from cache (read).
-    #[serde(skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub cache_read: Option<u32>,
     /// Tokens written to cache.
-    #[serde(skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub cache_write: Option<u32>,
 }
 
 /// Output token usage detail.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct OutputTokenUsage {
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub total: Option<u32>,
     /// Text (non-reasoning) tokens (output side).
-    #[serde(skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub text: Option<u32>,
     /// Reasoning tokens (output side).
-    #[serde(skip_serializing_if = "Option::is_none")]
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub reasoning: Option<u32>,
 }
 
@@ -139,16 +154,25 @@ impl std::str::FromStr for ReasoningEffort {
 
 /// A warning issued by the provider (e.g. unsupported parameter).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum Warning {
     /// A feature is not supported by this provider.
     Unsupported {
         feature: String,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         details: Option<String>,
     },
     /// A feature has compatibility issues.
     Compatibility {
         feature: String,
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         details: Option<String>,
     },
     /// A deprecated setting was used.
@@ -159,10 +183,17 @@ pub enum Warning {
 
 /// Metadata about the API response.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct ResponseMetadata {
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub id: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timestamp: Option<String>,
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub model_id: Option<String>,
 }
 
diff --git a/aimux-core/src/video_model.rs b/aimux-core/src/video_model.rs
index 647f3625a..eaeb413f7 100644
--- a/aimux-core/src/video_model.rs
+++ b/aimux-core/src/video_model.rs
@@ -19,6 +19,11 @@ use crate::{AbortSignal, retry, timeout};
 ///
 /// Aligned with V4 `VideoModelV4File`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum VideoFile {
     /// Inline file data (base64 or binary) with an explicit media type.
@@ -33,12 +38,15 @@ pub enum VideoFile {
         /// The URL of the video or image file.
         url: String,
         /// The media type of the referenced file, when known.
+        #[ts(optional)]
+        #[serde(default, skip_serializing_if = "Option::is_none")]
         media_type: Option<String>,
     },
 }
 
 /// File payload for a [`VideoFile::File`]: base64 string or raw bytes.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(untagged)]
 #[ts(export)]
 pub enum VideoFileData {
     /// Base64-encoded string.
@@ -51,6 +59,7 @@ pub enum VideoFileData {
 ///
 /// Aligned with V4 `VideoModelV4FrameType`.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, TS)]
+#[serde(rename_all = "snake_case")]
 #[ts(export)]
 pub enum VideoFrameType {
     /// The starting frame the model animates from.
@@ -63,6 +72,7 @@ pub enum VideoFrameType {
 ///
 /// Aligned with V4 `VideoModelV4FrameImage`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct VideoFrameImage {
     /// The image file used for this frame.
@@ -76,6 +86,11 @@ pub struct VideoFrameImage {
 /// Aligned with V4 `VideoModelV4VideoData`. Most providers return URLs due to
 /// large file sizes.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(
+    tag = "type",
+    rename_all = "lowercase",
+    rename_all_fields = "camelCase"
+)]
 #[ts(export)]
 pub enum VideoData {
     /// Video available as a URL (most common).
@@ -90,9 +105,12 @@ pub enum VideoData {
 ///
 /// Aligned with V4 `VideoModelV4CallOptions`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct VideoCallOptions {
     /// Text prompt for the video generation. `None` when not required.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub prompt: Option<String>,
 
     /// Number of videos to generate. Default `1`; most models only support
@@ -103,31 +121,48 @@ pub struct VideoCallOptions {
     pub n: u32,
 
     /// Aspect ratio, in `{width}:{height}` format (e.g. `"16:9"`).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub aspect_ratio: Option<AspectRatio>,
 
     /// Resolution, in `{width}x{height}` format (e.g. `"1280x720"`).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub resolution: Option<Size>,
 
     /// Duration of the video in seconds. Typically 3–10 seconds.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub duration: Option<u32>,
 
     /// Frames per second. Common values: 24, 30, 60.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub fps: Option<u32>,
 
     /// Seed for deterministic generation. `None` uses a random seed.
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub seed: Option<u64>,
 
     /// Input image for image-to-video generation (the starting frame).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub image: Option<VideoFile>,
 
     /// Role-tagged image inputs for first-last-frame generation.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub frame_images: Option<Vec<VideoFrameImage>>,
 
     /// Reference inputs for reference-to-video generation (images or videos).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub input_references: Option<Vec<VideoFile>>,
 
     /// Whether the model should generate audio alongside the video.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub generate_audio: Option<bool>,
 
     /// Additional provider-specific options, keyed by provider name.
@@ -140,16 +175,24 @@ pub struct VideoCallOptions {
     pub abort_signal: Option<AbortSignal>,
 
     /// Per-call retry override. `None` uses the model default.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_retries: Option<u32>,
 
     /// Per-call poll pacing override for the start/status flow. Unset fields
     /// fall back to the model's [`VideoModel::poll_config`].
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub poll: Option<VideoPollOptions>,
 
     /// Per-call operation timeout.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timeout: Option<crate::options::TimeoutConfiguration>,
 
     /// Additional HTTP headers to send with the request.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
 }
 
@@ -184,15 +227,18 @@ impl VideoCallOptions {
 
 /// Per-call poll pacing for video generation (AI SDK `generateVideo` `poll`).
 #[derive(Debug, Clone, Copy, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct VideoPollOptions {
     /// Delay between consecutive status checks, in milliseconds.
     // `number`, not the `bigint` ts-rs infers from u64: the JS bindings pass
     // options through `JSON.stringify`, which throws on BigInt.
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub interval_ms: Option<u64>,
     /// Maximum total time to wait for completion, in milliseconds.
-    #[ts(type = "number | null")]
+    #[ts(optional, type = "number")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timeout_ms: Option<u64>,
 }
 
@@ -200,6 +246,7 @@ pub struct VideoPollOptions {
 ///
 /// Aligned with V4 `VideoModelV4Result`.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct VideoResult {
     /// Generated videos as URLs, base64 strings, or binary data.
@@ -209,6 +256,8 @@ pub struct VideoResult {
     pub warnings: Vec<Warning>,
 
     /// Additional provider-specific metadata, keyed by provider name.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub provider_metadata: Option<SharedProviderMetadata>,
 
     /// Response information for telemetry and debugging.
@@ -217,13 +266,20 @@ pub struct VideoResult {
 
 /// Response information for a video generation call.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
+#[serde(rename_all = "camelCase")]
 #[ts(export)]
 pub struct VideoResponse {
     /// Timestamp for the start of the generated response (ISO 8601 string).
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub timestamp: Option<String>,
     /// The ID of the model that was used to generate the response.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub model_id: Option<String>,
     /// Response headers.
+    #[ts(optional)]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<SharedHeaders>,
 }
 
diff --git a/aimux-core/tests/error_value_golden_test.rs b/aimux-core/tests/error_value_golden_test.rs
index fa62d5598..b86beef36 100644
--- a/aimux-core/tests/error_value_golden_test.rs
+++ b/aimux-core/tests/error_value_golden_test.rs
@@ -1,5 +1,6 @@
-//! Golden snapshots of the `error_value` payload — the externally-tagged serde
-//! JSON that PR #91 ships across the FFI boundary to all eight languages.
+//! Golden snapshots of the `error_value` payload — the `name`-tagged serde
+//! JSON (the AI SDK's error names) that ships across the FFI boundary to all
+//! eight languages.
 //!
 //! These assert the *exact* JSON for every variant. The payload is a public
 //! cross-language contract: a field rename, a variant rename or a shape change
@@ -32,7 +33,7 @@ fn error_value_snapshots_api_call_shapes() {
             is_retryable: true,
             ..api_error("boom")
         })),
-        r#"{"ApiCall":{"url":"https://example.test/v1","request_body_values":{},"status_code":500,"provider_code":"server_error","message":"boom","response_body":null,"response_headers":null,"data":null,"is_retryable":true}}"#,
+        r#"{"name":"AI_APICallError","url":"https://example.test/v1","requestBodyValues":{},"statusCode":500,"providerCode":"server_error","message":"boom","isRetryable":true}"#,
     );
     golden(
         &AiMuxError::ApiCall(Box::new(ApiCallError {
@@ -42,7 +43,7 @@ fn error_value_snapshots_api_call_shapes() {
             is_retryable: true,
             ..api_error("boom")
         })),
-        r#"{"ApiCall":{"url":"https://example.test/v1","request_body_values":{},"status_code":500,"provider_code":"server_error","message":"boom","response_body":"{\"error\":{\"message\":\"boom\",\"type\":\"server_error\"}}","response_headers":null,"data":null,"is_retryable":true}}"#,
+        r#"{"name":"AI_APICallError","url":"https://example.test/v1","requestBodyValues":{},"statusCode":500,"providerCode":"server_error","message":"boom","responseBody":"{\"error\":{\"message\":\"boom\",\"type\":\"server_error\"}}","isRetryable":true}"#,
     );
     // A transport failure (no response arrived): no status, retryable —
     // exactly the AI SDK's handleFetchError shape.
@@ -51,7 +52,7 @@ fn error_value_snapshots_api_call_shapes() {
             is_retryable: true,
             ..api_error("connection reset")
         })),
-        r#"{"ApiCall":{"url":"https://example.test/v1","request_body_values":{},"status_code":null,"provider_code":null,"message":"connection reset","response_body":null,"response_headers":null,"data":null,"is_retryable":true}}"#,
+        r#"{"name":"AI_APICallError","url":"https://example.test/v1","requestBodyValues":{},"message":"connection reset","isRetryable":true}"#,
     );
     // A 429 is an ApiCall error whose classification is the status field —
     // there is no RateLimited variant; the hint remains in response headers.
@@ -66,11 +67,11 @@ fn error_value_snapshots_api_call_shapes() {
             is_retryable: true,
             ..api_error("slow down")
         })),
-        r#"{"ApiCall":{"url":"https://example.test/v1","request_body_values":{},"status_code":429,"provider_code":"rate_limit_exceeded","message":"slow down","response_body":null,"response_headers":{"retry-after-ms":"2500"},"data":null,"is_retryable":true}}"#,
+        r#"{"name":"AI_APICallError","url":"https://example.test/v1","requestBodyValues":{},"statusCode":429,"providerCode":"rate_limit_exceeded","message":"slow down","responseHeaders":{"retry-after-ms":"2500"},"isRetryable":true}"#,
     );
     golden(
         &AiMuxError::TokenExpired("expired".into()),
-        r#"{"TokenExpired":"expired"}"#,
+        r#"{"name":"AI_TokenExpiredError","message":"expired"}"#,
     );
 }
 
@@ -90,7 +91,7 @@ fn error_value_snapshot_retry_history() {
                 })),
             ],
         }),
-        r#"{"Retry":{"reason":"maxRetriesExceeded","errors":[{"ApiCall":{"url":"https://example.test/v1","request_body_values":{},"status_code":null,"provider_code":null,"message":"first","response_body":null,"response_headers":null,"data":null,"is_retryable":true}},{"ApiCall":{"url":"https://example.test/v1","request_body_values":{},"status_code":null,"provider_code":null,"message":"second","response_body":null,"response_headers":null,"data":null,"is_retryable":true}}]}}"#,
+        r#"{"name":"AI_RetryError","reason":"maxRetriesExceeded","errors":[{"name":"AI_APICallError","url":"https://example.test/v1","requestBodyValues":{},"message":"first","isRetryable":true},{"name":"AI_APICallError","url":"https://example.test/v1","requestBodyValues":{},"message":"second","isRetryable":true}]}"#,
     );
 }
 
@@ -99,11 +100,11 @@ fn error_value_snapshots_plain_variants() {
     for (err, expected) in [
         (
             AiMuxError::JsonParse("bad json".into()),
-            r#"{"JsonParse":"bad json"}"#,
+            r#"{"name":"AI_JSONParseError","message":"bad json"}"#,
         ),
         (
             AiMuxError::InvalidResponseData("eof".into()),
-            r#"{"InvalidResponseData":"eof"}"#,
+            r#"{"name":"AI_InvalidResponseDataError","message":"eof"}"#,
         ),
         // NoSuchTool is pinned in both shapes: `skip_serializing_if` makes
         // the wire payload vary with `available_tools`.
@@ -113,7 +114,7 @@ fn error_value_snapshots_plain_variants() {
                 available_tools: Some(vec!["weather".into(), "search".into()]),
                 tool_input: Some(r#""hello""#.into()),
             },
-            r#"{"NoSuchTool":{"tool_name":"weathr","available_tools":["weather","search"],"tool_input":"\"hello\""}}"#,
+            r#"{"name":"AI_NoSuchToolError","toolName":"weathr","availableTools":["weather","search"],"toolInput":"\"hello\""}"#,
         ),
         (
             AiMuxError::NoSuchTool {
@@ -121,7 +122,7 @@ fn error_value_snapshots_plain_variants() {
                 available_tools: None,
                 tool_input: None,
             },
-            r#"{"NoSuchTool":{"tool_name":"weathr"}}"#,
+            r#"{"name":"AI_NoSuchToolError","toolName":"weathr"}"#,
         ),
         (
             AiMuxError::InvalidToolInput {
@@ -129,7 +130,7 @@ fn error_value_snapshots_plain_variants() {
                 tool_input: "{".into(),
                 cause: "JSON parsing failed".into(),
             },
-            r#"{"InvalidToolInput":{"tool_name":"weather","tool_input":"{","cause":"JSON parsing failed"}}"#,
+            r#"{"name":"AI_InvalidToolInputError","toolName":"weather","toolInput":"{","cause":"JSON parsing failed"}"#,
         ),
         (
             AiMuxError::ToolCallRepair {
@@ -140,44 +141,44 @@ fn error_value_snapshots_plain_variants() {
                 }),
                 cause: Box::new(AiMuxError::Other("repair model failed".into())),
             },
-            r#"{"ToolCallRepair":{"original_error":{"NoSuchTool":{"tool_name":"weathr"}},"cause":{"Other":"repair model failed"}}}"#,
+            r#"{"name":"AI_ToolCallRepairError","originalError":{"name":"AI_NoSuchToolError","toolName":"weathr"},"cause":{"name":"AI_Error","message":"repair model failed"}}"#,
         ),
         (
             AiMuxError::InvalidArgument("bad arg".into()),
-            r#"{"InvalidArgument":"bad arg"}"#,
+            r#"{"name":"AI_InvalidArgumentError","message":"bad arg"}"#,
         ),
         (
             AiMuxError::InvalidPrompt("bad prompt".into()),
-            r#"{"InvalidPrompt":"bad prompt"}"#,
+            r#"{"name":"AI_InvalidPromptError","message":"bad prompt"}"#,
         ),
         (
             AiMuxError::TokenExpired("expired".into()),
-            r#"{"TokenExpired":"expired"}"#,
+            r#"{"name":"AI_TokenExpiredError","message":"expired"}"#,
         ),
         (
             AiMuxError::UnsupportedFunctionality("no audio".into()),
-            r#"{"UnsupportedFunctionality":"no audio"}"#,
+            r#"{"name":"AI_UnsupportedFunctionalityError","message":"no audio"}"#,
         ),
         (
             AiMuxError::LoadApiKey {
                 env_var: "OPENAI_API_KEY".into(),
                 description: "OpenAI".into(),
             },
-            r#"{"LoadApiKey":{"env_var":"OPENAI_API_KEY","description":"OpenAI"}}"#,
+            r#"{"name":"AI_LoadAPIKeyError","envVar":"OPENAI_API_KEY","description":"OpenAI"}"#,
         ),
         (
             AiMuxError::LoadSetting {
                 env_var: "AWS_REGION".into(),
                 name: "region".into(),
             },
-            r#"{"LoadSetting":{"env_var":"AWS_REGION","name":"region"}}"#,
+            r#"{"name":"AI_LoadSettingError","envVar":"AWS_REGION","settingName":"region"}"#,
         ),
         (
             AiMuxError::NoSuchModel {
                 model_id: "gpt-9".into(),
                 model_type: "languageModel".into(),
             },
-            r#"{"NoSuchModel":{"model_id":"gpt-9","model_type":"languageModel"}}"#,
+            r#"{"name":"AI_NoSuchModelError","modelId":"gpt-9","modelType":"languageModel"}"#,
         ),
         (
             AiMuxError::NoSuchProvider {
@@ -186,52 +187,31 @@ fn error_value_snapshots_plain_variants() {
                 model_type: "languageModel".into(),
                 available_providers: vec!["other".into()],
             },
-            r#"{"NoSuchProvider":{"provider_id":"acme","model_id":"acme","model_type":"languageModel","available_providers":["other"]}}"#,
+            r#"{"name":"AI_NoSuchProviderError","providerId":"acme","modelId":"acme","modelType":"languageModel","availableProviders":["other"]}"#,
         ),
         (
             AiMuxError::Timeout("total timeout".into()),
-            r#"{"Timeout":"total timeout"}"#,
+            r#"{"name":"TimeoutError","message":"total timeout"}"#,
         ),
         (
             AiMuxError::Aborted("request aborted".into()),
-            r#"{"Aborted":"request aborted"}"#,
+            r#"{"name":"AbortError","message":"request aborted"}"#,
+        ),
+        (
+            AiMuxError::Other("misc".into()),
+            r#"{"name":"AI_Error","message":"misc"}"#,
         ),
-        (AiMuxError::Other("misc".into()), r#"{"Other":"misc"}"#),
     ] {
         golden(&err, expected);
     }
 }
 
-/// Request context is required. Payloads from the pre-context schema are
-/// deliberately rejected instead of silently fabricating an empty URL/body.
+/// Request context is required: a payload without `url` / `requestBodyValues`
+/// is rejected instead of silently fabricating an empty URL/body.
 #[test]
 fn api_call_requires_request_context() {
-    let old = r#"{"ApiCall":{"message":"boom"}}"#;
-    assert!(serde_json::from_str::<AiMuxError>(old).is_err());
-
-    // The removed variants no longer deserialize — a deliberate breaking
-    // change pinned here so it cannot happen silently a second time.
-    assert!(
-        serde_json::from_str::<AiMuxError>(r#"{"RateLimited":{"retry_after_ms":5000}}"#).is_err()
-    );
-    assert!(serde_json::from_str::<AiMuxError>(r#"{"Auth":"bad key"}"#).is_err());
-    assert!(serde_json::from_str::<AiMuxError>(r#"{"Http":"connection reset"}"#).is_err());
-    assert!(serde_json::from_str::<AiMuxError>(r#"{"Provider":{"message":"boom"}}"#).is_err());
-    // NoSuchModel went from a plain string to a struct.
-    assert!(serde_json::from_str::<AiMuxError>(r#"{"NoSuchModel":"gpt-9"}"#).is_err());
-    // The renamed variants' OLD wire names no longer deserialize:
-    // Json/Stream/Unsupported/UnknownProvider became
-    // JsonParse/InvalidResponseData/UnsupportedFunctionality/NoSuchProvider.
-    assert!(serde_json::from_str::<AiMuxError>(r#"{"Json":"bad json"}"#).is_err());
-    assert!(serde_json::from_str::<AiMuxError>(r#"{"Stream":"eof"}"#).is_err());
-    assert!(serde_json::from_str::<AiMuxError>(r#"{"Unsupported":"no audio"}"#).is_err());
-    assert!(serde_json::from_str::<AiMuxError>(r#"{"UnknownProvider":"acme"}"#).is_err());
-    assert!(
-        serde_json::from_str::<AiMuxError>(
-            r#"{"UnknownProvider":{"provider_id":"acme","available":[]}}"#
-        )
-        .is_err()
-    );
+    let incomplete = r#"{"name":"AI_APICallError","message":"boom","isRetryable":false}"#;
+    assert!(serde_json::from_str::<AiMuxError>(incomplete).is_err());
 }
 
 /// The status lives in the field and *only* there. `Display` composes the
diff --git a/aimux-core/tests/tool_choice_test.rs b/aimux-core/tests/tool_choice_test.rs
index c7a38463f..4db0092bd 100644
--- a/aimux-core/tests/tool_choice_test.rs
+++ b/aimux-core/tests/tool_choice_test.rs
@@ -1,97 +1,36 @@
-//! Wire-format tests for `ToolChoice`.
-//!
-//! The serialized shape aligns with Vercel AI SDK `toolChoice`:
-//! `"auto" | "none" | "required" | { "type": "tool", "toolName": "..." }`.
-//! These tests lock the cross-language wire contract so refactors on the
-//! Rust side cannot silently break the bindings.
+//! Ported from upstream `ai/src/prompt/prepare-tool-choice.test.ts`: the user
+//! facing `toolChoice` values (`"auto" | "none" | "required" | { type: "tool",
+//! toolName }`) and their wire shape.
 
 use aimux_core::tool::ToolChoice;
 use serde_json::json;
 
-#[test]
-fn serialize_unit_variants_as_bare_strings() {
-    assert_eq!(
-        serde_json::to_value(ToolChoice::Auto).unwrap(),
-        json!("auto")
-    );
-    assert_eq!(
-        serde_json::to_value(ToolChoice::None).unwrap(),
-        json!("none")
-    );
-    assert_eq!(
-        serde_json::to_value(ToolChoice::Required).unwrap(),
-        json!("required")
-    );
+fn assert_wire(choice: ToolChoice, wire: serde_json::Value) {
+    assert_eq!(serde_json::to_value(&choice).unwrap(), wire);
+    assert_eq!(serde_json::from_value::<ToolChoice>(wire).unwrap(), choice);
 }
 
+/// TS: returns auto when tool choice is not provided
 #[test]
-fn serialize_tool_variant_as_tagged_object_with_camelcase_field() {
-    let tc = ToolChoice::Tool {
-        tool_name: "get_weather".into(),
-    };
-    assert_eq!(
-        serde_json::to_value(tc).unwrap(),
-        json!({ "type": "tool", "toolName": "get_weather" })
-    );
+fn defaults_to_auto() {
+    assert_eq!(ToolChoice::default(), ToolChoice::Auto);
 }
 
+/// TS: handles string tool choice: auto / none / required
 #[test]
-fn deserialize_bare_strings() {
-    assert_eq!(
-        serde_json::from_str::<ToolChoice>("\"auto\"").unwrap(),
-        ToolChoice::Auto
-    );
-    assert_eq!(
-        serde_json::from_str::<ToolChoice>("\"none\"").unwrap(),
-        ToolChoice::None
-    );
-    assert_eq!(
-        serde_json::from_str::<ToolChoice>("\"required\"").unwrap(),
-        ToolChoice::Required
-    );
+fn string_tool_choices() {
+    assert_wire(ToolChoice::Auto, json!("auto"));
+    assert_wire(ToolChoice::None, json!("none"));
+    assert_wire(ToolChoice::Required, json!("required"));
 }
 
+/// TS: handles object tool choice
 #[test]
-fn deserialize_tool_object() {
-    let v = json!({ "type": "tool", "toolName": "get_weather" });
-    assert_eq!(
-        serde_json::from_value::<ToolChoice>(v).unwrap(),
+fn object_tool_choice() {
+    assert_wire(
         ToolChoice::Tool {
-            tool_name: "get_weather".into()
-        }
-    );
-}
-
-#[test]
-fn round_trip_all_variants() {
-    let cases = [
-        ToolChoice::Auto,
-        ToolChoice::None,
-        ToolChoice::Required,
-        ToolChoice::Tool {
-            tool_name: "search".into(),
+            tool_name: "tool2".into(),
         },
-    ];
-    for tc in cases {
-        let j = serde_json::to_string(&tc).unwrap();
-        let back: ToolChoice = serde_json::from_str(&j).unwrap();
-        assert_eq!(back, tc, "round-trip failed for {j}");
-    }
-}
-
-#[test]
-fn reject_unknown_string() {
-    assert!(serde_json::from_str::<ToolChoice>("\"always\"").is_err());
-}
-
-#[test]
-fn reject_tool_object_missing_tool_name() {
-    let v = json!({ "type": "tool" });
-    assert!(serde_json::from_value::<ToolChoice>(v).is_err());
-}
-
-#[test]
-fn reject_unknown_tool_type() {
-    let v = json!({ "type": "function", "toolName": "x" });
-    assert!(serde_json::from_value::<ToolChoice>(v).is_err());
+        json!({ "type": "tool", "toolName": "tool2" }),
+    );
 }
diff --git a/aimux-core/tests/tool_input_test.rs b/aimux-core/tests/tool_input_test.rs
index 3360f1bca..b36cda348 100644
--- a/aimux-core/tests/tool_input_test.rs
+++ b/aimux-core/tests/tool_input_test.rs
@@ -1,3 +1,5 @@
+//! Ported from upstream `ai/src/generate-text/parse-tool-call.test.ts`.
+
 use std::sync::Arc;
 use std::sync::atomic::{AtomicUsize, Ordering};
 
@@ -21,6 +23,7 @@ fn raw(name: &str, input: &str) -> RawToolCall {
     }
 }
 
+/// TS: should successfully parse a valid tool call
 #[tokio::test]
 async fn parses_and_validates_exact_json() {
     let mut tool_call = raw("weather", r#"{"city":"Singapore","days":3}"#);
@@ -33,6 +36,7 @@ async fn parses_and_validates_exact_json() {
     assert!(call.error.is_none());
 }
 
+/// TS: should successfully process empty tool calls for tools that have no inputSchema
 #[tokio::test]
 async fn validates_empty_input_as_an_empty_object() {
     let no_arg_tool = FunctionTool::new(
@@ -52,47 +56,7 @@ async fn validates_empty_input_as_an_empty_object() {
     assert_eq!(call.invalid, None);
 }
 
-#[tokio::test]
-async fn parses_a_json_string_without_confusing_it_with_the_raw_carrier() {
-    let echo_tool = FunctionTool::new("echo", json!({ "type": "string" }));
-    let call = parse_tool_call(
-        raw("echo", r#""hello""#),
-        Some(&[echo_tool.into()]),
-        None,
-        &[],
-        None,
-    )
-    .await;
-
-    assert_eq!(call.input, json!("hello"));
-    assert_eq!(call.invalid, None);
-}
-
-#[tokio::test]
-async fn does_not_apply_partial_json_repair_to_final_tool_calls() {
-    for input in [
-        r#"{"city":"Singapore"#,
-        r#"{"city":"Singapore",}"#,
-        r#"{"city":"Singapore"},"days":3}"#,
-    ] {
-        let call = parse_tool_call(
-            raw("weather", input),
-            Some(&[weather_tool()]),
-            None,
-            &[],
-            None,
-        )
-        .await;
-
-        assert_eq!(call.input, json!(input));
-        assert_eq!(call.invalid, Some(true));
-        assert!(matches!(
-            call.error,
-            Some(AiMuxError::InvalidToolInput { .. })
-        ));
-    }
-}
-
+/// TS: should throw InvalidToolInputError when args are invalid
 #[tokio::test]
 async fn preserves_parsed_input_when_schema_validation_fails() {
     let input = r#"{"city":7}"#;
@@ -113,6 +77,7 @@ async fn preserves_parsed_input_when_schema_validation_fails() {
     ));
 }
 
+/// TS: should throw NoSuchToolError when tool is not found
 #[tokio::test]
 async fn unknown_tool_is_an_invalid_dynamic_call_with_available_tools() {
     let call = parse_tool_call(
@@ -137,6 +102,7 @@ async fn unknown_tool_is_an_invalid_dynamic_call_with_available_tools() {
     ));
 }
 
+/// TS: should invoke repairTool when provided and use its result / should pass instructions to repairToolCall
 #[tokio::test]
 async fn repair_runs_once_and_the_replacement_is_fully_revalidated() {
     let calls = Arc::new(AtomicUsize::new(0));
@@ -169,40 +135,7 @@ async fn repair_runs_once_and_the_replacement_is_fully_revalidated() {
     assert_eq!(call.invalid, None);
 }
 
-#[tokio::test]
-async fn an_invalid_repair_is_not_repaired_again() {
-    let calls = Arc::new(AtomicUsize::new(0));
-    let repair_calls = Arc::clone(&calls);
-    let repair = ToolCallRepair::new(move |context| {
-        repair_calls.fetch_add(1, Ordering::SeqCst);
-        async move {
-            Ok(Some(RawToolCall {
-                input: r#"{"city":7}"#.into(),
-                ..context.tool_call
-            }))
-        }
-    });
-
-    let call = parse_tool_call(
-        raw("weather", "{"),
-        Some(&[weather_tool()]),
-        Some(&repair),
-        &[],
-        None,
-    )
-    .await;
-
-    assert_eq!(calls.load(Ordering::SeqCst), 1);
-    assert_eq!(call.input, json!("{"));
-    assert_eq!(call.invalid, Some(true));
-    assert!(matches!(
-        call.error,
-        Some(AiMuxError::ToolCallRepair { original_error, cause })
-            if matches!(*original_error, AiMuxError::InvalidToolInput { ref tool_input, .. } if tool_input == "{")
-                && matches!(*cause, AiMuxError::InvalidToolInput { ref tool_input, .. } if tool_input == r#"{"city":7}"#)
-    ));
-}
-
+/// TS: should throw NoSuchToolError when tools is null
 #[tokio::test]
 async fn missing_tools_bypasses_repair_like_ai_sdk() {
     let calls = Arc::new(AtomicUsize::new(0));
@@ -224,6 +157,7 @@ async fn missing_tools_bypasses_repair_like_ai_sdk() {
     ));
 }
 
+/// TS: should re-throw error if tool call repair returns null
 #[tokio::test]
 async fn repair_returning_none_keeps_the_original_failure() {
     let repair = ToolCallRepair::new(|_| async { Ok(None) });
@@ -245,6 +179,7 @@ async fn repair_returning_none_keeps_the_original_failure() {
     ));
 }
 
+/// TS: should throw ToolCallRepairError if repairToolCall throws
 #[tokio::test]
 async fn repair_failure_keeps_both_typed_errors() {
     let repair = ToolCallRepair::new(|context| async move {
@@ -269,6 +204,7 @@ async fn repair_failure_keeps_both_typed_errors() {
     ));
 }
 
+/// TS: should successfully parse a valid provider-executed dynamic tool call
 #[tokio::test]
 async fn provider_executed_dynamic_calls_do_not_require_a_local_tool() {
     let mut tool_call = raw("provider_search", r#"{"query":"rust"}"#);
diff --git a/aimux-ffi/aimux-ffi.h b/aimux-ffi/aimux-ffi.h
index d694ef1ae..fe17d31bc 100644
--- a/aimux-ffi/aimux-ffi.h
+++ b/aimux-ffi/aimux-ffi.h
@@ -48,7 +48,7 @@
  * ## Concurrency
  *
  * All functions are synchronous (block until the operation completes).
- * `opts_json.timeout` (`{"total_ms","step_ms","first_chunk_ms","chunk_ms"}`,
+ * `opts_json.timeout` (`{"totalMs","stepMs","firstChunkMs","chunkMs"}`,
  * milliseconds; the latter two streaming-only) bounds a call from the callee
  * side; `aimux_abort_signal_*` plus the `*_with_abort` entry points cancel a
  * running call from another thread.
@@ -67,7 +67,7 @@
  * - Malformed JSON text is a C ABI failure; well-formed JSON of the wrong
  *   shape is AiMuxError AIMUX_E_INVALID_ARGUMENT.
  * - Results: serialized JSON of GenerateTextResult on success
- * - Stream parts: serialized JSON of StreamPart
+ * - Stream parts: serialized JSON of TextStreamPart
  */
 
 #ifndef AIMUX_FFI_H
@@ -206,7 +206,7 @@ aimux_error_t *aimux_provider_handle_new(const char *name, const char *api_key,
 
 /**
  * [AiMuxError] List models on a provider handle (RFC-0027 runtime discovery).
- * Writes a JSON array of sparse RuntimeModel (id / owned_by / created) —
+ * Writes a JSON array of sparse RuntimeModel (id / ownedBy / created) —
  * no community enrichment. To supplement with model specs, call
  * aimux_get_model_specs separately and merge in the host.
  */
@@ -262,7 +262,7 @@ aimux_error_t *aimux_consume_stream_text(uint64_t handle,
  * @param handle      Model handle from aimux_*_new.
  * @param prompt_json JSON prompt (see wire format above).
  * @param opts_json   JSON options (NULL or empty for defaults).
- * @param on_part     Called for each StreamPart (JSON string, valid during
+ * @param on_part     Called for each TextStreamPart (JSON string, valid during
  *                    the call only). Required: NULL is a C ABI failure
  *                    ("on_part: must not be NULL").
  * @param on_done     Called once when the stream ends normally. Required.
@@ -310,7 +310,7 @@ void aimux_abort_signal_drop(uint64_t abort_handle);
  * @param abort_handle Handle from aimux_abort_signal_new.
  * @param prompt_json  JSON prompt.
  * @param opts_json    JSON options. NULL or empty uses defaults.
- * @param on_part      Called for each StreamPart.
+ * @param on_part      Called for each TextStreamPart.
  * @param on_done      Called once after normal completion.
  * @param stream_ctx   Opaque pointer passed through to callbacks.
  */
@@ -393,7 +393,7 @@ aimux_error_t *aimux_generate_text_result_as_openai(uint64_t handle, const char
  * defaults.
  *
  * reply_json is one of:
- *   {"type":"repaired","tool_call":{"tool_call_id":…,"tool_name":…,"input":"<raw text>"}}
+ *   {"type":"repaired","toolCall":{"toolCallId":…,"toolName":…,"input":"<raw text>"}}
  *   {"type":"unchanged"}
  *   {"type":"failed","message":"…"}
  *
@@ -521,9 +521,9 @@ typedef enum aimux_transcription_next_part_state {
    parts with aimux_transcription_next_part. model_handle must support
    streaming (models without do_stream fail on the first next_part).
    abort_handle: 0 = no cancellation, or an aimux_abort_signal_new handle.
-   opts_json (all optional): { "input_audio_format": {"format_type","rate"},
-   "provider_options", "headers", "include_raw_chunks", "timeout":
-   {"total_ms","first_chunk_ms","chunk_ms"} }; NULL/empty/"null" = defaults
+   opts_json (all optional): { "inputAudioFormat": {"formatType","rate"},
+   "providerOptions", "headers", "includeRawChunks", "timeout":
+   {"totalMs","firstChunkMs","chunkMs"} }; NULL/empty/"null" = defaults
    (wrong shape → AiMuxError AIMUX_E_INVALID_ARGUMENT). Writes the session handle. */
 aimux_error_t *aimux_transcription_session_new(uint64_t model_handle, uint64_t abort_handle,
                                                    const char *opts_json, uint64_t *out_handle);
@@ -749,7 +749,7 @@ aimux_error_t *aimux_mock_replay_new(const char *recordings_jsonl, uint64_t *out
    dropped; len == 0 (no children) is a C ABI failure too; NULL with len > 0
    is a C ABI failure. config_json selects the router + fallback policy:
    { "router": "rule"|"weighted", "weights": [..], "fallback": "on_error"|"none",
-     "provider_name": "router", "model_id": "router" } — all optional; defaults
+     "providerName": "router", "modelId": "router" } — all optional; defaults
    are rule / on_error / "router" / "router" (wrong shape → AiMuxError
    AIMUX_E_INVALID_ARGUMENT). */
 aimux_error_t *aimux_router_new(const uint64_t *handles, size_t len,
@@ -764,8 +764,8 @@ aimux_error_t *aimux_router_new(const uint64_t *handles, size_t len,
    with ref_len > 0 is a C ABI failure). aggregator: a single live handle.
    Any dead reference or aggregator handle is a C ABI failure — nothing is
    silently dropped. config_json is a serialized MoaConfig (all fields
-   optional): { "provider_name", "model_id", "aggregator_instructions",
-   "strip_reference_tools", "fail_mode": "best_effort"|"fail_fast" } (wrong shape →
+   optional): { "providerName", "modelId", "aggregatorInstructions",
+   "stripReferenceTools", "failMode": "best_effort"|"fail_fast" } (wrong shape →
    AiMuxError AIMUX_E_INVALID_ARGUMENT). */
 aimux_error_t *aimux_moa_new(const uint64_t *reference_handles, size_t ref_len,
                                  uint64_t aggregator, const char *config_json,
diff --git a/aimux-ffi/src/lib.rs b/aimux-ffi/src/lib.rs
index 49fb34140..5e9a5b9a0 100644
--- a/aimux-ffi/src/lib.rs
+++ b/aimux-ffi/src/lib.rs
@@ -1660,7 +1660,7 @@ pub extern "C" fn aimux_provider_handle_new(
 /// List models on a provider handle (RFC-0027 runtime discovery).
 ///
 /// `handle` is from `aimux_provider_handle_new`. Writes a JSON array of
-/// sparse `RuntimeModel` (id / owned_by / created) — **no community
+/// sparse `RuntimeModel` (id / ownedBy / created) — **no community
 /// enrichment**. To supplement with model specs, call `aimux_get_model_specs`
 /// separately and merge in the host.
 #[unsafe(no_mangle)]
@@ -1755,8 +1755,8 @@ pub extern "C" fn aimux_generate_text(
 /// Generate a structured JSON object from the model (M12, RFC-0016).
 ///
 /// Same signature as [`aimux_generate_text`]; writes the serialized
-/// `GenerateObjectResult`. The caller passes `response_format: { "Json": {
-/// ... } }` via `opts_json` to control the schema; the function applies JSON
+/// `GenerateObjectResult`. The caller passes `responseFormat: { "type": "json",
+/// ... }` via `opts_json` to control the schema; the function applies JSON
 /// repair before parsing.
 #[unsafe(no_mangle)]
 pub extern "C" fn aimux_generate_object(
@@ -1898,13 +1898,13 @@ pub extern "C" fn aimux_stream_text_with_abort(
 
 /// Build the repair argument for one invalid tool call (RFC-0035).
 ///
-/// `tool_call_json` is one entry of `GenerateTextResult.tool_calls` with
+/// `tool_call_json` is one entry of `GenerateTextResult.toolCalls` with
 /// `"invalid": true`. `prompt_json` and `opts_json` are **the same two
 /// strings the call was generated with** ([`aimux_generate_text`] /
 /// [`aimux_stream_text`]); messages, instructions, and the tool set are
 /// derived from them here, so no host repeats that derivation.
 ///
-/// Writes `{tool_call, error, input_schema, tools, messages, instructions}` —
+/// Writes `{toolCall, error, inputSchema, tools, messages, instructions}` —
 /// the AI SDK `repairToolCall` argument — to `*out_json`, or the JSON literal
 /// `null` when `opts_json` carried no tools: as in the AI SDK, a call made
 /// without a tool set is never repaired, and the host skips it. `null` is a
@@ -1936,7 +1936,7 @@ pub extern "C" fn aimux_tool_call_repair_context(
 /// Resolve one invalid tool call against a host's repair reply (RFC-0035).
 ///
 /// `opts_json` is the same string the call was generated with; the tool set
-/// comes from it. `reply_json` is `{"type":"repaired","tool_call":{…}}`,
+/// comes from it. `reply_json` is `{"type":"repaired","toolCall":{…}}`,
 /// `{"type":"unchanged"}`, or `{"type":"failed","message":"…"}`. Writes the
 /// resulting `ToolCall` — valid, or invalid with the AI SDK's nested
 /// `ToolCallRepairError` — to `*out_json`. Same purity as
@@ -1968,7 +1968,7 @@ pub extern "C" fn aimux_apply_tool_call_repair(
 ///
 /// `result_json` is a serialized `GenerateTextResult` or
 /// `GenerateObjectResult`; `opts_json` is the same string the call was
-/// generated with. Both `tool_calls` and the matching `response_messages`
+/// generated with. Both `toolCalls` and the matching `responseMessages`
 /// tool-call part are rewritten. The OpenAI-shaped
 /// `ChatCompletion` is not supported — it carries no `invalid`/`error`, so
 /// repair is driven from the native result, which
@@ -2004,7 +2004,7 @@ pub extern "C" fn aimux_apply_tool_call_repair_to_result(
 ///
 /// The conversion half of [`aimux_generate_text_as_openai`], for hosts that
 /// repair a native result themselves (RFC-0035) and then want the OpenAI
-/// shape: tool calls come from `result_json`'s `tool_calls`, so a patched
+/// shape: tool calls come from `result_json`'s `toolCalls`, so a patched
 /// result converts to a completion that reflects the repair. `handle` only
 /// supplies the model id fallback for `ChatCompletion.model`; no runtime work,
 /// so this is safe from inside a stream callback.
@@ -2624,9 +2624,10 @@ pub extern "C" fn aimux_transcription_generate(
 // ─────────────────────────────────────────────────────────────────────────────
 
 /// JSON options for `aimux_transcription_session_new` (all fields optional):
-/// `{ "input_audio_format": { "format_type": "audio/pcm", "rate": 24000 },
-///    "provider_options": { … }, "headers": { … }, "include_raw_chunks": false }`.
+/// `{ "inputAudioFormat": { "formatType": "audio/pcm", "rate": 24000 },
+///    "providerOptions": { … }, "headers": { … }, "includeRawChunks": false }`.
 #[derive(serde::Deserialize, Default)]
+#[serde(rename_all = "camelCase")]
 struct TranscriptionSessionFfiOptions {
     input_audio_format: Option<aimux_core::transcription_model::InputAudioFormat>,
     provider_options: Option<aimux_core::shared::SharedProviderOptions>,
@@ -3589,7 +3590,7 @@ fn model_handles(
 /// `handles` is an array of `len` existing model handles (e.g. from
 /// `aimux_openai_new`). `config_json` selects the router + fallback policy:
 /// `{ "router": "rule" | "weighted", "weights": [..], "fallback": "on_error" |
-/// "none", "provider_name": "router", "model_id": "router" }`. All keys are
+/// "none", "providerName": "router", "modelId": "router" }`. All keys are
 /// optional; defaults are `rule` / `on_error` / `"router"` / `"router"`.
 ///
 /// Fails on: NULL `handles` with `len > 0`, bad JSON, zero-length `handles`,
@@ -3646,9 +3647,9 @@ pub extern "C" fn aimux_router_new(
 /// `reference_handles` is an array of `ref_len` existing model handles (may be
 /// 0 — MoaModel then degrades to running just the aggregator). `aggregator`
 /// is a single existing model handle. `config_json` is a serialized `MoaConfig`
-/// (all fields optional): `{ "provider_name": "moa", "model_id": "moa",
-/// "aggregator_instructions": null, "strip_reference_tools": true,
-/// "fail_mode": "best_effort" | "fail_fast" }`.
+/// (all fields optional): `{ "providerName": "moa", "modelId": "moa",
+/// "aggregatorInstructions": null, "stripReferenceTools": true,
+/// "failMode": "best_effort" | "fail_fast" }`.
 ///
 /// Fails on: bad JSON, an unknown aggregator handle (references may be empty;
 /// aggregator may not).
@@ -3682,6 +3683,7 @@ pub extern "C" fn aimux_moa_new(
 
 /// FFI-side router config (lenient: all fields optional).
 #[derive(Default, serde::Deserialize)]
+#[serde(rename_all = "camelCase")]
 struct RouterFfiConfig {
     router: Option<String>,
     weights: Option<Vec<f64>>,
@@ -4537,15 +4539,15 @@ mod tests {
         // Pull parts: StreamStart, delta("3 "), delta("2 "), Final, Finish,
         // then ENDED.
         let p1 = pull_part(session);
-        assert!(p1.contains("StreamStart"), "part 1: {p1}");
+        assert!(p1.contains("\"type\":\"stream-start\""), "part 1: {p1}");
         let p2 = pull_part(session);
         assert!(p2.contains("3 "), "part 2: {p2}");
         let p3 = pull_part(session);
         assert!(p3.contains("2 "), "part 3: {p3}");
         let p4 = pull_part(session);
-        assert!(p4.contains("TranscriptFinal"), "part 4: {p4}");
+        assert!(p4.contains("\"type\":\"transcript-final\""), "part 4: {p4}");
         let p5 = pull_part(session);
-        assert!(p5.contains("Finish"), "part 5: {p5}");
+        assert!(p5.contains("\"type\":\"finish\""), "part 5: {p5}");
         assert!(p5.contains("3 2 "), "finish text: {p5}");
 
         // Stream ended: NULL error, ENDED state, no part.
diff --git a/aimux-ffi/tests/exports_smoke_test.rs b/aimux-ffi/tests/exports_smoke_test.rs
index 4a5bf8701..b151e269f 100644
--- a/aimux-ffi/tests/exports_smoke_test.rs
+++ b/aimux-ffi/tests/exports_smoke_test.rs
@@ -7,7 +7,7 @@
 //! the ABI contract:
 //!
 //! - **Constructor class** (`*_new` / `*_new_with_base` / composites): fake
-//!   API key + `base_url` pointing at a loopback port nothing listens on
+//!   API key + `baseUrl` pointing at a loopback port nothing listens on
 //!   (`http://127.0.0.1:1`) → must return NULL and write a non-zero handle,
 //!   which is then released with [`aimux_ffi::aimux_drop_handle`].
 //!   Constructors that cannot take a base URL (e.g. `aimux_provider_from_env`)
@@ -18,7 +18,7 @@
 //!   transcription / rerank / video / search / files): handle + minimal
 //!   options against the unreachable base URL → must return a returned error
 //!   with the AiMuxError view and leave the out-param NULL (a clean error path).
-//!   Text calls pass `{"max_retries":0}` to skip the retry backoff; the
+//!   Text calls pass `{"maxRetries":0}` to skip the retry backoff; the
 //!   multimodal option structs expose no retry override, so those tests run
 //!   as separate `#[test]`s to parallelise the default backoff.
 //! - Every returned error and every returned JSON string is released
@@ -495,7 +495,7 @@ fn constructor_exports_build_and_release_handles() {
     );
 
     // Registry-backed constructors (RFC-0017 / RFC-0027).
-    let provider_opts = c(r#"{"base_url":"http://127.0.0.1:1"}"#);
+    let provider_opts = c(r#"{"baseUrl":"http://127.0.0.1:1"}"#);
     ctor!(
         "provider_new",
         aimux_provider_new(
@@ -575,7 +575,7 @@ fn constructor_exports_build_and_release_handles() {
 fn text_generation_exports_fail_cleanly_on_unreachable_host() {
     let h = unreachable_model();
     let prompt = c("\"ffi smoke\"");
-    let opts = c(r#"{"max_retries":0}"#);
+    let opts = c(r#"{"maxRetries":0}"#);
     let mut out: *mut c_char = ptr::null_mut();
 
     let e = aimux_generate_text(h, prompt.as_ptr(), opts.as_ptr(), &mut out);
@@ -599,7 +599,7 @@ fn text_generation_exports_fail_cleanly_on_unreachable_host() {
 fn streaming_exports_fail_cleanly_on_unreachable_host() {
     let h = unreachable_model();
     let prompt = c("\"ffi smoke\"");
-    let opts = c(r#"{"max_retries":0}"#);
+    let opts = c(r#"{"maxRetries":0}"#);
     let done_before = DONE_SEEN.load(Ordering::Relaxed);
 
     let e = aimux_stream_text(
@@ -702,7 +702,7 @@ fn image_generate_fails_cleanly_on_unreachable_host() {
         &mut h,
     );
     expect_handle(e, h, "image handle");
-    let opts = c(r#"{"prompt":"a rust crab","n":1,"provider_options":{}}"#);
+    let opts = c(r#"{"prompt":"a rust crab","n":1,"providerOptions":{}}"#);
     let mut out: *mut c_char = ptr::null_mut();
     let e = aimux_image_generate(h, opts.as_ptr(), &mut out);
     expect_ptr_aimux_failure(e, out, "image_generate");
@@ -758,7 +758,7 @@ fn rerank_fails_cleanly_on_unreachable_host() {
         &mut h,
     );
     expect_handle(e, h, "reranking handle");
-    let opts = c(r#"{"query":"rust","documents":{"Text":{"values":["a","b"]}}}"#);
+    let opts = c(r#"{"query":"rust","documents":{"type":"text","values":["a","b"]}}"#);
     let mut out: *mut c_char = ptr::null_mut();
     let e = aimux_rerank(h, opts.as_ptr(), &mut out);
     expect_ptr_aimux_failure(e, out, "rerank");
@@ -779,7 +779,7 @@ fn video_generate_fails_cleanly_on_unreachable_host() {
     // pins the serde defaults and the poll wire keys at the ABI boundary —
     // typed binding structs omit unset fields, so a strict parse here would
     // regress every C-ABI language at once.
-    let opts = c(r#"{"prompt":"waves","poll":{"interval_ms":1,"timeout_ms":1}}"#);
+    let opts = c(r#"{"prompt":"waves","poll":{"intervalMs":1,"timeoutMs":1}}"#);
     let mut out: *mut c_char = ptr::null_mut();
     let e = aimux_video_generate(h, opts.as_ptr(), &mut out);
     assert!(out.is_null(), "video_generate: expected NULL out-param");
@@ -886,7 +886,7 @@ fn provider_discovery_and_catalogue_exports_fail_cleanly() {
     let e = aimux_provider_handle_new(
         c("groq").as_ptr(),
         c(FAKE_KEY).as_ptr(),
-        c(r#"{"base_url":"http://127.0.0.1:1"}"#).as_ptr(),
+        c(r#"{"baseUrl":"http://127.0.0.1:1"}"#).as_ptr(),
         &mut provider,
     );
     expect_handle(e, provider, "provider_handle_new");
@@ -1099,9 +1099,9 @@ fn recording_exports_lifecycle_cleanly() {
 
 // ── host-side tool-call repair (RFC-0035, 3 exports) ───────────────────────
 
-const REPAIR_TOOLS: &str = r#"[{"type":"function","name":"weather","input_schema":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}}]"#;
-const REPAIR_INVALID: &str = r#"{"tool_call_id":"call-1","tool_name":"weather","input":{"town":"SG"},"dynamic":true,"invalid":true,"error":{"InvalidToolInput":{"tool_name":"weather","tool_input":"{\"town\":\"SG\"}","cause":"missing city"}}}"#;
-const REPAIR_REPLY: &str = r#"{"type":"repaired","tool_call":{"tool_call_id":"call-1","tool_name":"weather","input":"{\"city\":\"SG\"}"}}"#;
+const REPAIR_TOOLS: &str = r#"[{"type":"function","name":"weather","inputSchema":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}}]"#;
+const REPAIR_INVALID: &str = r#"{"toolCallId":"call-1","toolName":"weather","input":{"town":"SG"},"dynamic":true,"invalid":true,"error":{"name":"AI_InvalidToolInputError","toolName":"weather","toolInput":"{\"town\":\"SG\"}","cause":"missing city"}}"#;
+const REPAIR_REPLY: &str = r#"{"type":"repaired","toolCall":{"toolCallId":"call-1","toolName":"weather","input":"{\"city\":\"SG\"}"}}"#;
 
 /// The three repair exports are pure: they take JSON, return JSON, and never
 /// touch a handle or the runtime. They also accept the very `prompt_json` /
@@ -1120,7 +1120,7 @@ fn tool_call_repair_exports_round_trip() {
     ok(e, "tool_call_repair_context");
     let context = read_and_free_json(out, "tool_call_repair_context");
     assert!(
-        context.contains("\"input_schema\"") && context.contains("{\\\"town\\\":\\\"SG\\\"}"),
+        context.contains("\"inputSchema\"") && context.contains("{\\\"town\\\":\\\"SG\\\"}"),
         "context should carry the schema and the raw argument text: {context}"
     );
     // The `{"prompt": …}` wrapper unwrapped into the one user message.
@@ -1143,7 +1143,7 @@ fn tool_call_repair_exports_round_trip() {
         "repaired call should be valid: {repaired}"
     );
 
-    let result = format!(r#"{{"tool_calls":[{REPAIR_INVALID}],"response_messages":[]}}"#);
+    let result = format!(r#"{{"toolCalls":[{REPAIR_INVALID}],"responseMessages":[]}}"#);
     let e = aimux_apply_tool_call_repair_to_result(
         c(&result).as_ptr(),
         c(&opts).as_ptr(),
@@ -1174,10 +1174,10 @@ fn tool_call_repair_exports_round_trip() {
 #[test]
 fn a_repaired_result_converts_to_a_chat_completion() {
     let finish = r#"{"unified":"tool-calls","raw":null}"#;
-    let usage = r#"{"input_tokens":{"total":1},"output_tokens":{"total":1}}"#;
-    let repaired = r#"{"tool_call_id":"call-1","tool_name":"weather","input":{"city":"SG"}}"#;
+    let usage = r#"{"inputTokens":{"total":1},"outputTokens":{"total":1}}"#;
+    let repaired = r#"{"toolCallId":"call-1","toolName":"weather","input":{"city":"SG"}}"#;
     let result = format!(
-        r#"{{"text":"","tool_calls":[{repaired}],"finish_reason":{finish},"usage":{usage},"warnings":[],"raw":{{"content":[{{"ToolCall":{{"tool_call_id":"call-1","tool_name":"weather","input":"{{\"town\":\"SG\"}}"}}}}],"finish_reason":{finish},"usage":{usage},"warnings":[],"provider_metadata":null,"response":{{"id":null,"timestamp":null,"model_id":null}},"request_body":null,"response_headers":null}}}}"#
+        r#"{{"text":"","toolCalls":[{repaired}],"finishReason":{finish},"usage":{usage},"warnings":[],"raw":{{"content":[{{"type":"tool-call","toolCallId":"call-1","toolName":"weather","input":"{{\"town\":\"SG\"}}"}}],"finishReason":{finish},"usage":{usage},"warnings":[],"response":{{}}}}}}"#
     );
     let h = unreachable_model();
 
diff --git a/aimux-ffi/tests/provider_handle_test.rs b/aimux-ffi/tests/provider_handle_test.rs
index 3acc32e7d..f8d26a736 100644
--- a/aimux-ffi/tests/provider_handle_test.rs
+++ b/aimux-ffi/tests/provider_handle_test.rs
@@ -124,7 +124,7 @@ fn provider_handle_new_with_config_json() {
     let e = aimux_provider_handle_new(
         c("deepseek").as_ptr(),
         c("sk-test-fake").as_ptr(),
-        c(r#"{"base_url":"https://example.com/v1"}"#).as_ptr(),
+        c(r#"{"baseUrl":"https://example.com/v1"}"#).as_ptr(),
         &mut h,
     );
     aimux_drop_handle(expect_handle(e, h, "provider_handle_new(deepseek, config)"));
diff --git a/aimux-providers/src/catalogue.rs b/aimux-providers/src/catalogue.rs
index 4b5815e93..de926e87f 100644
--- a/aimux-providers/src/catalogue.rs
+++ b/aimux-providers/src/catalogue.rs
@@ -86,6 +86,7 @@ pub fn normalize_provider_name(anya2a_id: &str) -> String {
 /// In-memory catalogue: per-provider model specs, keyed by aimux provider
 /// name then model id.
 #[derive(Debug, Clone, Default, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
 pub struct Catalogue {
     #[serde(default)]
     pub updated_at: u64,
diff --git a/aimux-providers/tests/decision_recording_test.rs b/aimux-providers/tests/decision_recording_test.rs
index 08cb9d732..0fef961ed 100644
--- a/aimux-providers/tests/decision_recording_test.rs
+++ b/aimux-providers/tests/decision_recording_test.rs
@@ -85,7 +85,7 @@ async fn ring_records_one_complete_call_and_replays_exact_inputs_with_redaction(
     assert_eq!(
         serde_json::to_value(&records[0].provider).unwrap(),
         json!({
-            "provider_id": "jev", "provider": "jev", "model_id": "jev-latest"
+            "providerId": "jev", "provider": "jev", "modelId": "jev-latest"
         })
     );
     assert_eq!(records[0].exchanges.len(), 1);
diff --git a/aimux-providers/tests/e2e_test.rs b/aimux-providers/tests/e2e_test.rs
index e2301cbbd..5e023300e 100644
--- a/aimux-providers/tests/e2e_test.rs
+++ b/aimux-providers/tests/e2e_test.rs
@@ -9,6 +9,7 @@
 use aimux_core::content::ContentPart;
 use aimux_core::error::AiMuxError;
 use aimux_core::generate::{GenerateTextOptions, generate_text, stream_text};
+use aimux_core::language_model_message::ToolResultOutput;
 use aimux_core::message::{MessageContent, ModelMessage, Role};
 use aimux_core::stream_part::TextStreamPart;
 use aimux_core::tool::{FunctionTool, Tool, ToolCall, ToolChoice};
@@ -607,10 +608,15 @@ async fn e2e_openai_tool_call_round_trip() {
         },
         ModelMessage {
             role: Role::Tool,
-            content: MessageContent::Parts(vec![ContentPart::tool_result(
-                "call_abc",
-                json!({"temperature": 22, "condition": "sunny"}),
-            )]),
+            content: MessageContent::Parts(vec![ContentPart::ToolResult {
+                tool_call_id: "call_abc".into(),
+                tool_name: "get_weather".into(),
+                output: ToolResultOutput::Json {
+                    value: json!({"temperature": 22, "condition": "sunny"}),
+                    provider_options: None,
+                },
+                provider_options: None,
+            }]),
         },
     ];
 
diff --git a/aimux-providers/tests/fixtures/jev_systemone_live.jsonl b/aimux-providers/tests/fixtures/jev_systemone_live.jsonl
index b2fff1e5d..a8a84f561 100644
--- a/aimux-providers/tests/fixtures/jev_systemone_live.jsonl
+++ b/aimux-providers/tests/fixtures/jev_systemone_live.jsonl
@@ -1,2 +1,2 @@
-{"schema":3,"call_id":"call-1791293057374663321-0","recorded_at":"2026-10-06T13:24:17.375Z","input":{"operation":"decision","prompt":[],"options":{"state":"I was billed twice and payments are blocked. Please fix this urgently.","questions":[{"type":"boolean","id":"urgent","instructions":"Does the customer request urgent action?"},{"type":"choice","id":"department","instructions":"Choose the responsible department.","options":[{"label":"billing"},{"label":"technical"},{"label":"sales"}]},{"type":"score","id":"severity","instructions":"Rate the impact of the issue.","levels":["Minor inconvenience","Significant issue","Payments completely blocked"]}],"max_retries":0,"timeout":{"total_ms":30000,"step_ms":null,"first_chunk_ms":null,"chunk_ms":null}},"decision_capabilities":{"rounding":{"probability_decimals":2,"score_decimals":2},"probability_source":"native","supports_boolean":true,"supports_choice":true,"supports_score":true,"returns_distributions":true,"max_questions":null,"min_choices":1,"max_choices":255,"max_score_levels":10}},"provider":{"provider":"jev","model_id":"jev-latest","provider_id":"jev"},"exchanges":[{"attempt":1,"exchange_index":1,"request":{"method":"POST","url":"https://api.typesafe.ai/v1/systemone","headers":[["Authorization","[REDACTED]"]],"body":"{\"model\":\"jev-latest\",\"state\":\"I was billed twice and payments are blocked. Please fix this urgently.\",\"questions\":{\"urgent\":{\"type\":\"noul\",\"instructions\":\"Does the customer request urgent action?\"},\"department\":{\"type\":\"choice\",\"instructions\":\"Choose the responsible department.\",\"criteria\":{\"billing\":null,\"technical\":null,\"sales\":null}},\"severity\":{\"type\":\"score\",\"instructions\":\"Rate the impact of the issue.\",\"criteria\":[\"Minor inconvenience\",\"Significant issue\",\"Payments completely blocked\"]}}}"},"response":{"status":200,"headers":[["date","Tue, 06 Oct 2026 13:24:17 GMT"],["content-type","application/json"],["content-length","440"],["connection","keep-alive"],["server","cloudflare"],["x-typesafe-request-id","req_01a11162eaba71f5b6825f0466bca790"],["cf-cache-status","DYNAMIC"],["set-cookie","[REDACTED]"],["cf-ray","a464fc4aaee03760-LAX"],["alt-svc","h3=\":443\"; ma=86400"]],"body":"{\"model\":\"jev-1.13.0\",\"answers\":{\"urgent\":{\"type\":\"noul\",\"noul\":0.99},\"department\":{\"type\":\"choice\",\"choice\":\"billing\",\"confidence\":1.0,\"probabilities\":{\"billing\":1.0,\"technical\":0.0,\"sales\":0.0}},\"severity\":{\"type\":\"score\",\"score\":2.0,\"confidence\":1.0,\"legend\":{\"0\":\"Minor inconvenience\",\"1\":\"Significant issue\",\"2\":\"Payments completely blocked\"},\"probabilities\":{\"0\":0.0,\"1\":0.0,\"2\":1.0}}},\"usage\":{\"input_tokens\":365,\"output_tokens\":68}}","stream_chunks":1,"ttfb_ms":486},"timing":{"latency_ms":486,"ttfb_ms":486},"error":null,"finalized":true}],"outcome":{"decision_result":{"rounding":{"probability_decimals":2,"score_decimals":2},"answers":{"department":{"type":"choice","selected":"billing","probabilities":{"billing":1.0,"sales":0.0,"technical":0.0},"confidence":1.0},"severity":{"type":"score","expected_value":2.0,"levels":["Minor inconvenience","Significant issue","Payments completely blocked"],"probabilities":[0.0,0.0,1.0],"confidence":1.0},"urgent":{"type":"boolean","probability_true":0.99}},"provider":"jev","model":"jev-1.13.0","model_version":null,"probability_source":"native","usage":{"input_tokens":{"total":365},"output_tokens":{"total":68},"raw":{"input_tokens":365,"output_tokens":68}},"latency_ms":null,"provider_metadata":{"jev":{"model":"jev-1.13.0","answers":{"urgent":{"type":"noul","noul":0.99},"department":{"type":"choice","choice":"billing","confidence":1.0,"probabilities":{"billing":1.0,"technical":0.0,"sales":0.0}},"severity":{"type":"score","score":2.0,"confidence":1.0,"legend":{"0":"Minor inconvenience","1":"Significant issue","2":"Payments completely blocked"},"probabilities":{"0":0.0,"1":0.0,"2":1.0}}},"usage":{"input_tokens":365,"output_tokens":68}}},"response":{"headers":{"connection":"keep-alive","date":"Tue, 06 Oct 2026 13:24:17 GMT","set-cookie":"[REDACTED]","alt-svc":"h3=\":443\"; ma=86400","x-typesafe-request-id":"req_01a11162eaba71f5b6825f0466bca790","cf-cache-status":"DYNAMIC","server":"cloudflare","content-length":"440","content-type":"application/json","cf-ray":"a464fc4aaee03760-LAX"},"body":{"model":"jev-1.13.0","answers":{"urgent":{"type":"noul","noul":0.99},"department":{"type":"choice","choice":"billing","confidence":1.0,"probabilities":{"billing":1.0,"technical":0.0,"sales":0.0}},"severity":{"type":"score","score":2.0,"confidence":1.0,"legend":{"0":"Minor inconvenience","1":"Significant issue","2":"Payments completely blocked"},"probabilities":{"0":0.0,"1":0.0,"2":1.0}}},"usage":{"input_tokens":365,"output_tokens":68}}}},"status":"success","finish_reason":null,"error":null,"error_value":null,"usage":{"input_tokens":{"total":365},"output_tokens":{"total":68},"raw":{"input_tokens":365,"output_tokens":68}}},"complete":true,"transport_closed":true}
-{"schema":3,"call_id":"call-1791293057868486266-1","recorded_at":"2026-10-06T13:24:17.868Z","input":{"operation":"decision","prompt":[],"options":{"state":{"ticket":"Payments blocked"},"questions":[{"type":"boolean","id":"urgent","instructions":{"question":"Is action urgent?","context":["payments"]},"criteria":{"true":{"includes":["blocked"]},"false":["routine","can wait"]}},{"type":"choice","id":"department","instructions":["Route the ticket",{"priority":"payments"}],"options":[{"label":"billing","description":{"includes":["payments"],"examples":["payout failure"]}},{"label":"technical"}]},{"type":"score","id":"severity","instructions":"Rate severity","levels":[{"label":"low","includes":["minor inconvenience"]},["blocking","No workaround"]]}],"max_retries":0,"timeout":{"total_ms":30000,"step_ms":null,"first_chunk_ms":null,"chunk_ms":null}},"decision_capabilities":{"rounding":{"probability_decimals":2,"score_decimals":2},"probability_source":"native","supports_boolean":true,"supports_choice":true,"supports_score":true,"returns_distributions":true,"max_questions":null,"min_choices":1,"max_choices":255,"max_score_levels":10}},"provider":{"provider":"jev","model_id":"jev-latest","provider_id":"jev"},"exchanges":[{"attempt":1,"exchange_index":1,"request":{"method":"POST","url":"https://api.typesafe.ai/v1/systemone","headers":[["Authorization","[REDACTED]"]],"body":"{\"model\":\"jev-latest\",\"state\":{\"ticket\":\"Payments blocked\"},\"questions\":{\"urgent\":{\"type\":\"noul\",\"instructions\":{\"question\":\"Is action urgent?\",\"context\":[\"payments\"]},\"criteria\":{\"true\":{\"includes\":[\"blocked\"]},\"false\":[\"routine\",\"can wait\"]}},\"department\":{\"type\":\"choice\",\"instructions\":[\"Route the ticket\",{\"priority\":\"payments\"}],\"criteria\":{\"billing\":{\"includes\":[\"payments\"],\"examples\":[\"payout failure\"]},\"technical\":null}},\"severity\":{\"type\":\"score\",\"instructions\":\"Rate severity\",\"criteria\":[{\"label\":\"low\",\"includes\":[\"minor inconvenience\"]},[\"blocking\",\"No workaround\"]]}}}"},"response":{"status":200,"headers":[["date","Tue, 06 Oct 2026 13:24:18 GMT"],["content-type","application/json"],["content-length","431"],["connection","keep-alive"],["server","cloudflare"],["set-cookie","[REDACTED]"],["x-typesafe-request-id","req_01a11162ebc076ee849d8e67ae47f547"],["cf-cache-status","DYNAMIC"],["cf-ray","a464fc4c4b1c3760-LAX"],["alt-svc","h3=\":443\"; ma=86400"]],"body":"{\"model\":\"jev-1.13.0\",\"answers\":{\"urgent\":{\"type\":\"noul\",\"noul\":0.92},\"department\":{\"type\":\"choice\",\"choice\":\"billing\",\"confidence\":0.98,\"probabilities\":{\"technical\":0.01,\"billing\":0.99}},\"severity\":{\"type\":\"score\",\"score\":0.99,\"confidence\":0.98,\"legend\":{\"0\":{\"label\":\"low\",\"includes\":[\"minor inconvenience\"]},\"1\":[\"blocking\",\"No workaround\"]},\"probabilities\":{\"0\":0.01,\"1\":0.99}}},\"usage\":{\"input_tokens\":448,\"output_tokens\":61}}","stream_chunks":1,"ttfb_ms":257},"timing":{"latency_ms":257,"ttfb_ms":257},"error":null,"finalized":true}],"outcome":{"decision_result":{"rounding":{"probability_decimals":2,"score_decimals":2},"answers":{"department":{"type":"choice","selected":"billing","probabilities":{"billing":0.99,"technical":0.01},"confidence":0.98},"severity":{"type":"score","expected_value":0.99,"levels":[{"label":"low","includes":["minor inconvenience"]},["blocking","No workaround"]],"probabilities":[0.01,0.99],"confidence":0.98},"urgent":{"type":"boolean","probability_true":0.92}},"provider":"jev","model":"jev-1.13.0","model_version":null,"probability_source":"native","usage":{"input_tokens":{"total":448},"output_tokens":{"total":61},"raw":{"input_tokens":448,"output_tokens":61}},"latency_ms":null,"provider_metadata":{"jev":{"model":"jev-1.13.0","answers":{"urgent":{"type":"noul","noul":0.92},"department":{"type":"choice","choice":"billing","confidence":0.98,"probabilities":{"technical":0.01,"billing":0.99}},"severity":{"type":"score","score":0.99,"confidence":0.98,"legend":{"0":{"label":"low","includes":["minor inconvenience"]},"1":["blocking","No workaround"]},"probabilities":{"0":0.01,"1":0.99}}},"usage":{"input_tokens":448,"output_tokens":61}}},"response":{"headers":{"date":"Tue, 06 Oct 2026 13:24:18 GMT","content-length":"431","set-cookie":"[REDACTED]","connection":"keep-alive","server":"cloudflare","cf-ray":"a464fc4c4b1c3760-LAX","alt-svc":"h3=\":443\"; ma=86400","x-typesafe-request-id":"req_01a11162ebc076ee849d8e67ae47f547","cf-cache-status":"DYNAMIC","content-type":"application/json"},"body":{"model":"jev-1.13.0","answers":{"urgent":{"type":"noul","noul":0.92},"department":{"type":"choice","choice":"billing","confidence":0.98,"probabilities":{"technical":0.01,"billing":0.99}},"severity":{"type":"score","score":0.99,"confidence":0.98,"legend":{"0":{"label":"low","includes":["minor inconvenience"]},"1":["blocking","No workaround"]},"probabilities":{"0":0.01,"1":0.99}}},"usage":{"input_tokens":448,"output_tokens":61}}}},"status":"success","finish_reason":null,"error":null,"error_value":null,"usage":{"input_tokens":{"total":448},"output_tokens":{"total":61},"raw":{"input_tokens":448,"output_tokens":61}}},"complete":true,"transport_closed":true}
+{"schema":3,"input":{"operation":"decision","prompt":[],"options":{"state":"I was billed twice and payments are blocked. Please fix this urgently.","questions":[{"type":"boolean","id":"urgent","instructions":"Does the customer request urgent action?"},{"type":"choice","id":"department","instructions":"Choose the responsible department.","options":[{"label":"billing"},{"label":"technical"},{"label":"sales"}]},{"type":"score","id":"severity","instructions":"Rate the impact of the issue.","levels":["Minor inconvenience","Significant issue","Payments completely blocked"]}],"max_retries":0,"timeout":{"totalMs":30000,"stepMs":null,"firstChunkMs":null,"chunkMs":null}},"decisionCapabilities":{"rounding":{"probability_decimals":2,"score_decimals":2},"probability_source":"native","supports_boolean":true,"supports_choice":true,"supports_score":true,"returns_distributions":true,"max_questions":null,"min_choices":1,"max_choices":255,"max_score_levels":10}},"provider":{"provider":"jev","modelId":"jev-latest","providerId":"jev"},"exchanges":[{"attempt":1,"request":{"method":"POST","url":"https://api.typesafe.ai/v1/systemone","headers":[["Authorization","[REDACTED]"]],"body":"{\"model\":\"jev-latest\",\"state\":\"I was billed twice and payments are blocked. Please fix this urgently.\",\"questions\":{\"urgent\":{\"type\":\"noul\",\"instructions\":\"Does the customer request urgent action?\"},\"department\":{\"type\":\"choice\",\"instructions\":\"Choose the responsible department.\",\"criteria\":{\"billing\":null,\"technical\":null,\"sales\":null}},\"severity\":{\"type\":\"score\",\"instructions\":\"Rate the impact of the issue.\",\"criteria\":[\"Minor inconvenience\",\"Significant issue\",\"Payments completely blocked\"]}}}"},"response":{"status":200,"headers":[["date","Tue, 06 Oct 2026 13:24:17 GMT"],["content-type","application/json"],["content-length","440"],["connection","keep-alive"],["server","cloudflare"],["x-typesafe-request-id","req_01a11162eaba71f5b6825f0466bca790"],["cf-cache-status","DYNAMIC"],["set-cookie","[REDACTED]"],["cf-ray","a464fc4aaee03760-LAX"],["alt-svc","h3=\":443\"; ma=86400"]],"body":"{\"model\":\"jev-1.13.0\",\"answers\":{\"urgent\":{\"type\":\"noul\",\"noul\":0.99},\"department\":{\"type\":\"choice\",\"choice\":\"billing\",\"confidence\":1.0,\"probabilities\":{\"billing\":1.0,\"technical\":0.0,\"sales\":0.0}},\"severity\":{\"type\":\"score\",\"score\":2.0,\"confidence\":1.0,\"legend\":{\"0\":\"Minor inconvenience\",\"1\":\"Significant issue\",\"2\":\"Payments completely blocked\"},\"probabilities\":{\"0\":0.0,\"1\":0.0,\"2\":1.0}}},\"usage\":{\"input_tokens\":365,\"output_tokens\":68}}","streamChunks":1,"ttfbMs":486},"timing":{"latencyMs":486,"ttfbMs":486},"error":null,"finalized":true,"exchangeIndex":1}],"outcome":{"status":"success","error":null,"usage":{"raw":{"input_tokens":365,"output_tokens":68},"inputTokens":{"total":365},"outputTokens":{"total":68}},"decisionResult":{"rounding":{"probability_decimals":2,"score_decimals":2},"answers":{"department":{"type":"choice","selected":"billing","probabilities":{"billing":1.0,"sales":0.0,"technical":0.0},"confidence":1.0},"severity":{"type":"score","expected_value":2.0,"levels":["Minor inconvenience","Significant issue","Payments completely blocked"],"probabilities":[0.0,0.0,1.0],"confidence":1.0},"urgent":{"type":"boolean","probability_true":0.99}},"provider":"jev","model":"jev-1.13.0","model_version":null,"probability_source":"native","usage":{"raw":{"input_tokens":365,"output_tokens":68},"inputTokens":{"total":365},"outputTokens":{"total":68}},"latency_ms":null,"provider_metadata":{"jev":{"model":"jev-1.13.0","answers":{"urgent":{"type":"noul","noul":0.99},"department":{"type":"choice","choice":"billing","confidence":1.0,"probabilities":{"billing":1.0,"technical":0.0,"sales":0.0}},"severity":{"type":"score","score":2.0,"confidence":1.0,"legend":{"0":"Minor inconvenience","1":"Significant issue","2":"Payments completely blocked"},"probabilities":{"0":0.0,"1":0.0,"2":1.0}}},"usage":{"input_tokens":365,"output_tokens":68}}},"response":{"headers":{"connection":"keep-alive","date":"Tue, 06 Oct 2026 13:24:17 GMT","set-cookie":"[REDACTED]","alt-svc":"h3=\":443\"; ma=86400","x-typesafe-request-id":"req_01a11162eaba71f5b6825f0466bca790","cf-cache-status":"DYNAMIC","server":"cloudflare","content-length":"440","content-type":"application/json","cf-ray":"a464fc4aaee03760-LAX"},"body":{"model":"jev-1.13.0","answers":{"urgent":{"type":"noul","noul":0.99},"department":{"type":"choice","choice":"billing","confidence":1.0,"probabilities":{"billing":1.0,"technical":0.0,"sales":0.0}},"severity":{"type":"score","score":2.0,"confidence":1.0,"legend":{"0":"Minor inconvenience","1":"Significant issue","2":"Payments completely blocked"},"probabilities":{"0":0.0,"1":0.0,"2":1.0}}},"usage":{"input_tokens":365,"output_tokens":68}}}},"finishReason":null,"errorValue":null},"complete":true,"callId":"call-1791293057374663321-0","recordedAt":"2026-10-06T13:24:17.375Z","transportClosed":true}
+{"schema":3,"input":{"operation":"decision","prompt":[],"options":{"state":{"ticket":"Payments blocked"},"questions":[{"type":"boolean","id":"urgent","instructions":{"question":"Is action urgent?","context":["payments"]},"criteria":{"true":{"includes":["blocked"]},"false":["routine","can wait"]}},{"type":"choice","id":"department","instructions":["Route the ticket",{"priority":"payments"}],"options":[{"label":"billing","description":{"includes":["payments"],"examples":["payout failure"]}},{"label":"technical"}]},{"type":"score","id":"severity","instructions":"Rate severity","levels":[{"label":"low","includes":["minor inconvenience"]},["blocking","No workaround"]]}],"max_retries":0,"timeout":{"totalMs":30000,"stepMs":null,"firstChunkMs":null,"chunkMs":null}},"decisionCapabilities":{"rounding":{"probability_decimals":2,"score_decimals":2},"probability_source":"native","supports_boolean":true,"supports_choice":true,"supports_score":true,"returns_distributions":true,"max_questions":null,"min_choices":1,"max_choices":255,"max_score_levels":10}},"provider":{"provider":"jev","modelId":"jev-latest","providerId":"jev"},"exchanges":[{"attempt":1,"request":{"method":"POST","url":"https://api.typesafe.ai/v1/systemone","headers":[["Authorization","[REDACTED]"]],"body":"{\"model\":\"jev-latest\",\"state\":{\"ticket\":\"Payments blocked\"},\"questions\":{\"urgent\":{\"type\":\"noul\",\"instructions\":{\"question\":\"Is action urgent?\",\"context\":[\"payments\"]},\"criteria\":{\"true\":{\"includes\":[\"blocked\"]},\"false\":[\"routine\",\"can wait\"]}},\"department\":{\"type\":\"choice\",\"instructions\":[\"Route the ticket\",{\"priority\":\"payments\"}],\"criteria\":{\"billing\":{\"includes\":[\"payments\"],\"examples\":[\"payout failure\"]},\"technical\":null}},\"severity\":{\"type\":\"score\",\"instructions\":\"Rate severity\",\"criteria\":[{\"label\":\"low\",\"includes\":[\"minor inconvenience\"]},[\"blocking\",\"No workaround\"]]}}}"},"response":{"status":200,"headers":[["date","Tue, 06 Oct 2026 13:24:18 GMT"],["content-type","application/json"],["content-length","431"],["connection","keep-alive"],["server","cloudflare"],["set-cookie","[REDACTED]"],["x-typesafe-request-id","req_01a11162ebc076ee849d8e67ae47f547"],["cf-cache-status","DYNAMIC"],["cf-ray","a464fc4c4b1c3760-LAX"],["alt-svc","h3=\":443\"; ma=86400"]],"body":"{\"model\":\"jev-1.13.0\",\"answers\":{\"urgent\":{\"type\":\"noul\",\"noul\":0.92},\"department\":{\"type\":\"choice\",\"choice\":\"billing\",\"confidence\":0.98,\"probabilities\":{\"technical\":0.01,\"billing\":0.99}},\"severity\":{\"type\":\"score\",\"score\":0.99,\"confidence\":0.98,\"legend\":{\"0\":{\"label\":\"low\",\"includes\":[\"minor inconvenience\"]},\"1\":[\"blocking\",\"No workaround\"]},\"probabilities\":{\"0\":0.01,\"1\":0.99}}},\"usage\":{\"input_tokens\":448,\"output_tokens\":61}}","streamChunks":1,"ttfbMs":257},"timing":{"latencyMs":257,"ttfbMs":257},"error":null,"finalized":true,"exchangeIndex":1}],"outcome":{"status":"success","error":null,"usage":{"raw":{"input_tokens":448,"output_tokens":61},"inputTokens":{"total":448},"outputTokens":{"total":61}},"decisionResult":{"rounding":{"probability_decimals":2,"score_decimals":2},"answers":{"department":{"type":"choice","selected":"billing","probabilities":{"billing":0.99,"technical":0.01},"confidence":0.98},"severity":{"type":"score","expected_value":0.99,"levels":[{"label":"low","includes":["minor inconvenience"]},["blocking","No workaround"]],"probabilities":[0.01,0.99],"confidence":0.98},"urgent":{"type":"boolean","probability_true":0.92}},"provider":"jev","model":"jev-1.13.0","model_version":null,"probability_source":"native","usage":{"raw":{"input_tokens":448,"output_tokens":61},"inputTokens":{"total":448},"outputTokens":{"total":61}},"latency_ms":null,"provider_metadata":{"jev":{"model":"jev-1.13.0","answers":{"urgent":{"type":"noul","noul":0.92},"department":{"type":"choice","choice":"billing","confidence":0.98,"probabilities":{"technical":0.01,"billing":0.99}},"severity":{"type":"score","score":0.99,"confidence":0.98,"legend":{"0":{"label":"low","includes":["minor inconvenience"]},"1":["blocking","No workaround"]},"probabilities":{"0":0.01,"1":0.99}}},"usage":{"input_tokens":448,"output_tokens":61}}},"response":{"headers":{"date":"Tue, 06 Oct 2026 13:24:18 GMT","content-length":"431","set-cookie":"[REDACTED]","connection":"keep-alive","server":"cloudflare","cf-ray":"a464fc4c4b1c3760-LAX","alt-svc":"h3=\":443\"; ma=86400","x-typesafe-request-id":"req_01a11162ebc076ee849d8e67ae47f547","cf-cache-status":"DYNAMIC","content-type":"application/json"},"body":{"model":"jev-1.13.0","answers":{"urgent":{"type":"noul","noul":0.92},"department":{"type":"choice","choice":"billing","confidence":0.98,"probabilities":{"technical":0.01,"billing":0.99}},"severity":{"type":"score","score":0.99,"confidence":0.98,"legend":{"0":{"label":"low","includes":["minor inconvenience"]},"1":["blocking","No workaround"]},"probabilities":{"0":0.01,"1":0.99}}},"usage":{"input_tokens":448,"output_tokens":61}}}},"finishReason":null,"errorValue":null},"complete":true,"callId":"call-1791293057868486266-1","recordedAt":"2026-10-06T13:24:17.868Z","transportClosed":true}
diff --git a/aimux-providers/tests/fixtures/jev_systemone_live.md b/aimux-providers/tests/fixtures/jev_systemone_live.md
index 092850b56..5d08ea3c5 100644
--- a/aimux-providers/tests/fixtures/jev_systemone_live.md
+++ b/aimux-providers/tests/fixtures/jev_systemone_live.md
@@ -16,3 +16,5 @@ normalized result. Offline provider tests replay the wire exchange via the
 shared replay helper; Core and the CLI replay the normalized result.
 
 Stack integration migrates the recording envelope to schema 3: provider identity only, with decision capabilities under input. Recorded requests, responses and normalized outcomes are unchanged.
+
+The wire-format layer uses camelCase for the recording envelope and shared Usage/Timeout types. The original HTTP request and response body strings remain unchanged.
diff --git a/aimux-providers/tests/google_evaluation_test.rs b/aimux-providers/tests/google_evaluation_test.rs
index 54f0b2f75..dd1891123 100644
--- a/aimux-providers/tests/google_evaluation_test.rs
+++ b/aimux-providers/tests/google_evaluation_test.rs
@@ -58,9 +58,7 @@ fn options() -> EvaluationCallOptions {
                 "instructions": ["Rate severity."],
                 "criteria": ["Low", "Medium", "High"]
             }
-        },
-        "headers": null,
-        "provider_options": null
+        }
     }))
     .unwrap()
 }
diff --git a/aimux-providers/tests/mistral_upstream_e2e_test.rs b/aimux-providers/tests/mistral_upstream_e2e_test.rs
index 768f8e82d..247f05d66 100644
--- a/aimux-providers/tests/mistral_upstream_e2e_test.rs
+++ b/aimux-providers/tests/mistral_upstream_e2e_test.rs
@@ -86,8 +86,8 @@ async fn generate_preserves_complete_usage() {
     assert_eq!(
         serde_json::to_value(result.usage).unwrap(),
         json!({
-            "input_tokens": {"total": 20, "no_cache": 20},
-            "output_tokens": {"total": 2, "text": 2},
+            "inputTokens": {"total": 20, "noCache": 20},
+            "outputTokens": {"total": 2, "text": 2},
             "raw": body["usage"]
         })
     );
@@ -171,8 +171,8 @@ async fn stream_text_and_usage() {
             assert_eq!(
                 serde_json::to_value(usage).unwrap(),
                 json!({
-                    "input_tokens": {"total": 13, "no_cache": 13},
-                    "output_tokens": {"total": 8, "text": 8},
+                    "inputTokens": {"total": 13, "noCache": 13},
+                    "outputTokens": {"total": 8, "text": 8},
                     "raw": {"prompt_tokens": 13, "completion_tokens": 8, "total_tokens": 21}
                 })
             );
diff --git a/aimux-providers/tests/openai_decision_test.rs b/aimux-providers/tests/openai_decision_test.rs
index b9a35a6df..a19b33837 100644
--- a/aimux-providers/tests/openai_decision_test.rs
+++ b/aimux-providers/tests/openai_decision_test.rs
@@ -474,7 +474,7 @@ async fn bad_media_and_duplicate_native_values_are_rejected_before_http() {
     .unwrap();
     let server = MockServer::start().await;
     for (pointer, value) in [
-        ("/images/0/data", json!({"Base64":"bad!"})),
+        ("/images/0/data", json!("bad!")),
         ("/images/0/detail", json!("unknown")),
         ("/images/0/media_type", json!("text/plain")),
         ("/questions/0/options/1/value", json!(true)),
diff --git a/aimux-providers/tests/openai_upstream_e2e_test.rs b/aimux-providers/tests/openai_upstream_e2e_test.rs
index 29c399312..e74567530 100644
--- a/aimux-providers/tests/openai_upstream_e2e_test.rs
+++ b/aimux-providers/tests/openai_upstream_e2e_test.rs
@@ -257,28 +257,32 @@ async fn responses_stream() {
     assert_eq!(
         serde_json::to_value(parts).unwrap(),
         json!([
-            {"StreamStart": {"warnings": []}},
-            {"ResponseMetadata": {
+            {"type": "stream-start", "warnings": []},
+            {
+                "type": "response-metadata",
                 "id": "resp_67c9a81b6a048190a9ee441c5755a4e8",
-                "model_id": "test-model",
+                "modelId": "test-model",
                 "timestamp": "2025-03-06T13:50:19+00:00"
-            }},
-            {"TextStart": {
+            },
+            {
+                "type": "text-start",
                 "id": "msg_67c9a81dea8c8190b79651a2b3adf91e",
-                "provider_metadata": {"openai": {"itemId": "msg_67c9a81dea8c8190b79651a2b3adf91e"}}
-            }},
-            {"TextDelta": {"id": "msg_67c9a81dea8c8190b79651a2b3adf91e", "delta": "Hello,"}},
-            {"TextDelta": {"id": "msg_67c9a81dea8c8190b79651a2b3adf91e", "delta": " World!"}},
-            {"TextEnd": {
+                "providerMetadata": {"openai": {"itemId": "msg_67c9a81dea8c8190b79651a2b3adf91e"}}
+            },
+            {"type": "text-delta", "id": "msg_67c9a81dea8c8190b79651a2b3adf91e", "delta": "Hello,"},
+            {"type": "text-delta", "id": "msg_67c9a81dea8c8190b79651a2b3adf91e", "delta": " World!"},
+            {
+                "type": "text-end",
                 "id": "msg_67c9a81dea8c8190b79651a2b3adf91e",
-                "provider_metadata": {"openai": {"itemId": "msg_67c9a81dea8c8190b79651a2b3adf91e"}}
-            }},
-            {"Finish": {
-                "finish_reason": {"unified": "stop", "raw": null},
-                "provider_metadata": {"openai": {"responseId": "resp_67c9a81b6a048190a9ee441c5755a4e8"}},
+                "providerMetadata": {"openai": {"itemId": "msg_67c9a81dea8c8190b79651a2b3adf91e"}}
+            },
+            {
+                "type": "finish",
+                "finishReason": {"unified": "stop"},
+                "providerMetadata": {"openai": {"responseId": "resp_67c9a81b6a048190a9ee441c5755a4e8"}},
                 "usage": {
-                    "input_tokens": {"total": 543, "no_cache": 309, "cache_read": 234},
-                    "output_tokens": {"total": 478, "reasoning": 123, "text": 355},
+                    "inputTokens": {"total": 543, "noCache": 309, "cacheRead": 234},
+                    "outputTokens": {"total": 478, "reasoning": 123, "text": 355},
                     "raw": {
                         "input_tokens": 543,
                         "input_tokens_details": {"cached_tokens": 234},
@@ -287,7 +291,7 @@ async fn responses_stream() {
                         "total_tokens": 512
                     }
                 }
-            }}
+            }
         ])
     );
 }
@@ -562,13 +566,13 @@ async fn responses_web_search_stream() {
     assert_eq!(
         serde_json::to_value(parts).unwrap(),
         json!([
-            {"StreamStart":{"warnings":[]}},
-            {"ResponseMetadata":{"id":"resp_missing_action","model_id":"test-model","timestamp":"2025-03-10T18:10:55+00:00"}},
-            {"ToolInputStart":{"id":"ws_missing_action","tool_name":"webSearch","provider_executed":true}},
-            {"ToolInputEnd":{"id":"ws_missing_action"}},
-            {"ToolCall":{"tool_call_id":"ws_missing_action","tool_name":"webSearch","input":"{}","provider_executed":true}},
-            {"ToolResult":{"tool_call_id":"ws_missing_action","tool_name":"webSearch","result":{}}},
-            {"Finish":{"finish_reason":{"unified":"stop","raw":null},"usage":{"input_tokens":{"total":10,"no_cache":10,"cache_read":0},"output_tokens":{"total":2,"reasoning":0,"text":2},"raw":{"input_tokens":10,"input_tokens_details":{"cached_tokens":0},"output_tokens":2,"output_tokens_details":{"reasoning_tokens":0},"total_tokens":12}},"provider_metadata":{"openai":{"responseId":"resp_missing_action"}}}}
+            {"type":"stream-start","warnings":[]},
+            {"type":"response-metadata","id":"resp_missing_action","modelId":"test-model","timestamp":"2025-03-10T18:10:55+00:00"},
+            {"type":"tool-input-start","id":"ws_missing_action","toolName":"webSearch","providerExecuted":true},
+            {"type":"tool-input-end","id":"ws_missing_action"},
+            {"type":"tool-call","toolCallId":"ws_missing_action","toolName":"webSearch","input":"{}","providerExecuted":true},
+            {"type":"tool-result","toolCallId":"ws_missing_action","toolName":"webSearch","result":{}},
+            {"type":"finish","finishReason":{"unified":"stop"},"usage":{"inputTokens":{"total":10,"noCache":10,"cacheRead":0},"outputTokens":{"total":2,"reasoning":0,"text":2},"raw":{"input_tokens":10,"input_tokens_details":{"cached_tokens":0},"output_tokens":2,"output_tokens_details":{"reasoning_tokens":0},"total_tokens":12}},"providerMetadata":{"openai":{"responseId":"resp_missing_action"}}}
         ])
     );
 }
diff --git a/aimux-providers/tests/runtime_decision_test.rs b/aimux-providers/tests/runtime_decision_test.rs
index 8872e8d4b..da3bbf645 100644
--- a/aimux-providers/tests/runtime_decision_test.rs
+++ b/aimux-providers/tests/runtime_decision_test.rs
@@ -180,7 +180,7 @@ async fn media_is_native_per_runtime_and_recorded_for_replay() {
         let mut request = options();
         request.images = vec![serde_json::from_value(image.clone()).unwrap()];
         let mut expected = fixture()["request"].clone();
-        let encoded = image["data"]["Base64"].as_str().unwrap();
+        let encoded = image["data"].as_str().unwrap();
         let data_url = format!("data:image/png;base64,{encoded}");
         expected["images"] = match provider {
             "ollama" => json!([encoded]),
diff --git a/aimux-providers/tests/vertex_upstream_e2e_test.rs b/aimux-providers/tests/vertex_upstream_e2e_test.rs
index 1bd3feff0..1e6038bc2 100644
--- a/aimux-providers/tests/vertex_upstream_e2e_test.rs
+++ b/aimux-providers/tests/vertex_upstream_e2e_test.rs
@@ -289,8 +289,8 @@ async fn transcription_generate_request_and_result() {
     assert_eq!(
         serde_json::to_value(result.segments).unwrap(),
         json!([
-            {"text": "hello", "start_second": 0.0, "end_second": 0.5},
-            {"text": "world", "start_second": 0.5, "end_second": 1.0}
+            {"text": "hello", "startSecond": 0.0, "endSecond": 0.5},
+            {"text": "world", "startSecond": 0.5, "endSecond": 1.0}
         ])
     );
     assert_eq!(result.language.as_deref(), Some("en"));
diff --git a/aimux-providers/tests/xai_transcription_test.rs b/aimux-providers/tests/xai_transcription_test.rs
index 08c14ba5b..da99088e4 100644
--- a/aimux-providers/tests/xai_transcription_test.rs
+++ b/aimux-providers/tests/xai_transcription_test.rs
@@ -307,21 +307,26 @@ fn done(text: &str, duration: f64) -> Value {
 }
 
 fn kinds(parts: &[Value]) -> Vec<&str> {
-    parts
-        .iter()
-        .map(|p| p.as_object().unwrap().keys().next().unwrap().as_str())
-        .collect()
+    parts.iter().map(|p| p["type"].as_str().unwrap()).collect()
 }
 
 fn of_kind(parts: &[Value], kind: &str) -> Vec<Value> {
-    parts.iter().filter_map(|p| p.get(kind).cloned()).collect()
+    parts
+        .iter()
+        .filter(|p| p["type"] == kind)
+        .cloned()
+        .collect()
 }
 
 fn finish(text: &str, language: Value, duration: Option<f64>) -> Value {
-    json!({"Finish": {
-        "text": text, "segments": [], "language": language,
-        "duration_in_seconds": duration, "provider_metadata": null
-    }})
+    let mut part = json!({"type": "finish", "text": text, "segments": []});
+    if !language.is_null() {
+        part["language"] = language;
+    }
+    if let Some(duration) = duration {
+        part["durationInSeconds"] = json!(duration);
+    }
+    part
 }
 
 // ── tests: model information ────────────────────────────────────────────────
@@ -579,15 +584,15 @@ async fn should_stream_xai_stt_over_websocket() {
     assert_eq!(
         run.values(),
         vec![
-            json!({"StreamStart": {"warnings": []}}),
-            json!({"TranscriptPartial": {
-                "id": null, "text": "Hel", "start_second": 0.0, "duration_in_seconds": 0.5,
-                "channel_index": null, "provider_metadata": null
-            }}),
-            json!({"TranscriptFinal": {
-                "id": null, "text": "Hello", "start_second": 0.0, "end_second": 1.0,
-                "channel_index": null, "provider_metadata": null
-            }}),
+            json!({"type": "stream-start", "warnings": []}),
+            json!({
+                "type": "transcript-partial", "text": "Hel", "startSecond": 0.0,
+                "durationInSeconds": 0.5
+            }),
+            json!({
+                "type": "transcript-final", "text": "Hello", "startSecond": 0.0,
+                "endSecond": 1.0
+            }),
             finish("Hello", json!("en"), Some(1.0)),
         ]
     );
@@ -628,14 +633,14 @@ async fn should_emit_one_final_per_utterance_and_reconstruct_finish_text() {
     let parts = run(None, events).await.values();
 
     assert_eq!(
-        of_kind(&parts, "TranscriptFinal"),
+        of_kind(&parts, "transcript-final"),
         vec![json!({
-            "id": null, "text": "Hello world.", "start_second": 0.0, "end_second": 1.0,
-            "channel_index": null, "provider_metadata": null
+            "type": "transcript-final", "text": "Hello world.", "startSecond": 0.0,
+            "endSecond": 1.0
         })]
     );
     // the speech_final:false re-send surfaces as a partial
-    assert_eq!(of_kind(&parts, "TranscriptPartial").len(), 2);
+    assert_eq!(of_kind(&parts, "transcript-partial").len(), 2);
     assert_eq!(
         parts.last().unwrap(),
         &finish("Hello world.", Value::Null, Some(1.0))
@@ -656,14 +661,14 @@ async fn should_treat_is_final_fragments_as_partials_and_use_speech_final_text()
     assert_eq!(
         kinds(&parts),
         [
-            "StreamStart",
-            "TranscriptPartial",
-            "TranscriptPartial",
-            "TranscriptFinal",
-            "Finish"
+            "stream-start",
+            "transcript-partial",
+            "transcript-partial",
+            "transcript-final",
+            "finish"
         ]
     );
-    assert_eq!(parts.last().unwrap()["Finish"]["text"], "No, I'm not.");
+    assert_eq!(parts.last().unwrap()["text"], "No, I'm not.");
 }
 
 /// TS: doStream > should fall back to the latest pending text when no speech_final arrived before transcript.done
@@ -677,8 +682,8 @@ async fn should_fall_back_to_latest_pending_text_without_speech_final() {
     let parts = run(None, events).await.values();
 
     let last = parts.last().unwrap();
-    assert_eq!(kinds(&parts).last(), Some(&"Finish"));
-    assert_eq!(last["Finish"]["text"], "Hello world");
+    assert_eq!(kinds(&parts).last(), Some(&"finish"));
+    assert_eq!(last["text"], "Hello world");
 }
 
 /// TS: doStream > should join finalized utterances per channel when transcript.done is empty
@@ -692,7 +697,7 @@ async fn should_join_finalized_utterances_when_transcript_done_is_empty() {
     let parts = run(None, events).await.values();
 
     assert_eq!(
-        parts.last().unwrap()["Finish"]["text"],
+        parts.last().unwrap()["text"],
         "First utterance. Second utterance."
     );
 }
diff --git a/aimux-providers/tests/xai_upstream_e2e_test.rs b/aimux-providers/tests/xai_upstream_e2e_test.rs
index 89adc16f6..6ca24de7c 100644
--- a/aimux-providers/tests/xai_upstream_e2e_test.rs
+++ b/aimux-providers/tests/xai_upstream_e2e_test.rs
@@ -89,8 +89,8 @@ async fn generate_converts_usage() {
     assert_eq!(
         serde_json::to_value(result.usage).unwrap(),
         json!({
-            "input_tokens": {"total": 345, "no_cache": 345, "cache_read": 0},
-            "output_tokens": {"total": 538, "text": 415, "reasoning": 123},
+            "inputTokens": {"total": 345, "noCache": 345, "cacheRead": 0},
+            "outputTokens": {"total": 538, "text": 415, "reasoning": 123},
             "raw": usage
         })
     );
diff --git a/bindings/flutter/CHANGELOG.md b/bindings/flutter/CHANGELOG.md
index 3327a9844..4f7772016 100644
--- a/bindings/flutter/CHANGELOG.md
+++ b/bindings/flutter/CHANGELOG.md
@@ -1,3 +1,21 @@
+## Unreleased
+
+- Breaking: the typed layer and the maps `Model` returns follow the AI SDK
+  JSON of the Rust core: camelCase fields, `type`-tagged unions with
+  kebab-case values (`text-delta`, `tool-call`, ...), optional fields absent
+  instead of `null`, `FileBytes` untagged, errors keyed by `name`.
+- Breaking: `ContentPartToolResult` carries `toolCallId`, `toolName` and a
+  typed `ToolResultOutput`; `result`, `isError`, `preliminary`, `dynamic` are
+  gone from the message part (provider-executed `ToolResult` keeps them).
+- Breaking: `StreamPart` models the stream's `TextStreamPart` (adds
+  `finish-step`; `StreamPartToolCall` / `StreamPartToolResult` / `StreamPartFile`
+  wrap `ToolCall` / `ToolResult` / `GeneratedFile`; no `response-metadata`).
+- Breaking: `registerProviders` is removed together with
+  `aimux_register_providers`; `config_json` keys are camelCase (`baseUrl`) and
+  unknown keys are rejected.
+- `types.dart` no longer uses `json_serializable` except for the OpenAI Chat
+  Completions family; `types.g.dart` is regenerated for it.
+
 ## 0.5.0
 
 - Breaking release aligning the whole aimux family — see the
diff --git a/bindings/flutter/README.md b/bindings/flutter/README.md
index 828bec400..c231335eb 100644
--- a/bindings/flutter/README.md
+++ b/bindings/flutter/README.md
@@ -35,7 +35,8 @@ Desktop platforms resolve the library from the platform library path
 
 ```bash
 cargo build -p aimux-ffi --release
-LD_LIBRARY_PATH=../../target/release flutter test   # Linux/macOS
+LD_LIBRARY_PATH=../../target/release dart test       # Linux
+DYLD_LIBRARY_PATH=../../target/release dart test     # macOS
 ```
 
 ## License
diff --git a/bindings/flutter/lib/aimux.dart b/bindings/flutter/lib/aimux.dart
index 39428eccd..4785784d4 100644
--- a/bindings/flutter/lib/aimux.dart
+++ b/bindings/flutter/lib/aimux.dart
@@ -118,7 +118,7 @@ typedef _DropHandleC = Void Function(Uint64);
 typedef _DropHandleDart = void Function(int);
 
 // (const char *) → error: aimux_init_logging, aimux_init_recording,
-// aimux_register_providers, aimux_init_proxy.
+// aimux_init_proxy.
 typedef _StrC = _Err Function(Pointer<Utf8>);
 typedef _StrDart = _Err Function(Pointer<Utf8>);
 
@@ -292,8 +292,6 @@ final class _AimuxFFI {
           'aimux_recording_try_flush');
   late final mockReplayNew = _lib
       .lookupFunction<_MockReplayC, _MockReplayDart>('aimux_mock_replay_new');
-  late final registerProviders =
-      _lib.lookupFunction<_StrC, _StrDart>('aimux_register_providers');
   late final initProxy =
       _lib.lookupFunction<_StrC, _StrDart>('aimux_init_proxy');
   late final routerNew = _lib
@@ -499,9 +497,10 @@ class Model implements Finalizable {
   ///
   /// [apiKey] may be null to read the provider's env var from the registry
   /// entry. [configJson] is an optional JSON object of ProviderOptions
-  /// (`{"base_url": "...", "headers": {...}, "organization": "...",
-  /// "project": "...", "params": {...}}`). `max_retries` (a per-call option)
-  /// and `body_overrides` (removed) are rejected as [InvalidArgumentError].
+  /// (`{"baseUrl": "...", "headers": {...}, "organization": "...",
+  /// "project": "...", "params": {...}}`). Any other key (`maxRetries` is a
+  /// per-call option, `bodyOverrides` is gone) is rejected as
+  /// [InvalidArgumentError].
   factory Model.provider(String name, String modelId,
       {String? apiKey, String? configJson}) {
     checkJson(configJson, 'config_json', emptyIsDefault: true);
@@ -537,7 +536,7 @@ class Model implements Finalizable {
   /// Generate a structured JSON object (M12, RFC-0016).
   ///
   /// Same signature as [generateText]; returns the parsed
-  /// GenerateObjectResult as a Map. Pass `response_format: { "Json": { ... } }`
+  /// GenerateObjectResult as a Map. Pass `responseFormat: {"type": "json", ...}`
   /// via [options] for schema control; aimux-core applies JSON repair before
   /// parsing.
   ///
@@ -623,7 +622,7 @@ class Model implements Finalizable {
   /// Stream text from the model.
   ///
   /// Returns a Stream of StreamPart maps. Recoverable frame errors arrive
-  /// as `Error` stream-part maps and the stream continues; only
+  /// as `error` stream-part maps and the stream continues; only
   /// transport/Core failures throw [AimuxException] via [Stream.addError]
   /// (and do not call on_done).
   ///
@@ -646,7 +645,7 @@ class Model implements Finalizable {
   /// Returns a Stream of ChatCompletionChunk maps (OpenAI
   /// `chat.completion.chunk` objects). Stream options (`include_usage`,
   /// `include_reasoning`) are passed via
-  /// `options['provider_options']['openai']['stream_options']`.
+  /// `options['providerOptions']['openai']['stream_options']`.
   ///
   /// NOTE: like [streamText], this blocks the calling isolate until the
   /// stream completes and effectively buffers all parts until then.
@@ -748,8 +747,8 @@ class ProviderConfig {
   /// only the ones the preset declares are accepted.
   ///
   /// Retry is a per-call setting ([GenerateTextOptions.maxRetries]) and
-  /// request-body overrides no longer exist: a `max_retries` or
-  /// `body_overrides` key in the config is rejected as [InvalidArgumentError],
+  /// request-body overrides no longer exist: a `maxRetries` or
+  /// `bodyOverrides` key in the config is rejected as [InvalidArgumentError],
   /// so this class has neither.
   final Map<String, String>? params;
 
@@ -761,12 +760,12 @@ class ProviderConfig {
     this.params,
   });
 
-  /// Serialize to the `ProviderOptions` JSON wire format (snake_case keys,
-  /// null fields omitted). Returns null when no field is set, so callers can
+  /// Serialize to the `config_json` wire format (camelCase keys, null fields
+  /// omitted). Returns null when no field is set, so callers can
   /// pass null to the FFI for defaults.
   String? toJson() {
     final map = <String, dynamic>{};
-    if (baseUrl != null) map['base_url'] = baseUrl;
+    if (baseUrl != null) map['baseUrl'] = baseUrl;
     if (headers != null) map['headers'] = headers;
     if (organization != null) map['organization'] = organization;
     if (project != null) map['project'] = project;
@@ -817,7 +816,7 @@ class ProviderHandle implements Finalizable {
   /// (anya2a) when available.
   ///
   /// Returns a JSON array string of `ResolvedModel`
-  /// (`[{"id":"...","owned_by":"...","created":<u64>,"spec":{...}}, ...]`).
+  /// (`[{"id":"...","ownedBy":"...","created":<u64>,"spec":{...}}, ...]`).
   /// Synchronous: blocks the calling isolate until the network call completes.
   String listModels() {
     _checkOpen();
@@ -1003,19 +1002,6 @@ Model mockReplay(String recordingsJsonl) {
   return Model._(handle);
 }
 
-/// Register external OpenAI-compatible providers from a JSON config string
-/// (RFC-0020).
-///
-/// `configJSON` is `{ "providers": [ { "name", "base_url", ... } ] }`. Entries
-/// override same-named built-ins or add new ones. Like `initRecording`, this
-/// mutates process-global registry state. A well-formed document the registry
-/// rejects throws [InvalidArgumentError].
-void registerProviders(String configJSON) {
-  checkJson(configJSON, 'config_json');
-  withUtf8(configJSON,
-      (p) => expectAimuxError(_ffi.registerProviders(p), 'register_providers'));
-}
-
 /// Set the global proxy configuration (M6, RFC-0016). Must be called before
 /// the first `generateText` / `streamText` call; a no-op if the shared HTTP
 /// client is already initialised.
@@ -1032,7 +1018,7 @@ void initProxy(String configJSON) {
 /// (per `configJson`).
 ///
 /// `configJson` (optional): `{"router": "rule"|"weighted", "weights": [...],
-/// "fallback": "on_error"|"none", "provider_name", "model_id"}`.
+/// "fallback": "on_error"|"none", "providerName", "modelId"}`.
 Model router(List<Model> models, {String? configJson}) {
   if (models.isEmpty) {
     throw ArgumentError('router: models must be non-empty');
diff --git a/bindings/flutter/lib/errors.dart b/bindings/flutter/lib/errors.dart
index 470ab8b09..50443813f 100644
--- a/bindings/flutter/lib/errors.dart
+++ b/bindings/flutter/lib/errors.dart
@@ -720,8 +720,8 @@ class InvalidToolInputError extends AimuxException {
 /// A `repairToolCall` hook itself failed.
 class ToolCallRepairError extends AimuxException {
   /// The original lookup/parse/validation error the hook was repairing,
-  /// decoded from its externally-tagged wire JSON (the same shape as
-  /// `ToolCall.error`).
+  /// decoded from its wire JSON (an `AiMuxError` object keyed by `name`, the
+  /// same shape as `ToolCall.error`).
   final dynamic originalError;
 
   ToolCallRepairError(super.message,
diff --git a/bindings/flutter/lib/multimodal.dart b/bindings/flutter/lib/multimodal.dart
index df455a322..fedc79263 100644
--- a/bindings/flutter/lib/multimodal.dart
+++ b/bindings/flutter/lib/multimodal.dart
@@ -381,8 +381,8 @@ class TranscriptionModel {
   /// gpt-realtime-whisper).
   ///
   /// [optsJson] — optional session options JSON
-  /// (`{"input_audio_format": {...}, "provider_options", "headers",
-  /// "include_raw_chunks"}`).
+  /// (`{"inputAudioFormat": {...}, "providerOptions", "headers",
+  /// "includeRawChunks"}`).
   /// [abortHandle] — optional abort handle; firing it aborts the session.
   TranscriptionSession startStream({String? optsJson, int abortHandle = 0}) {
     _checkOpen();
diff --git a/bindings/flutter/lib/tool_call_repair.dart b/bindings/flutter/lib/tool_call_repair.dart
index d46bc932c..982d9cc75 100644
--- a/bindings/flutter/lib/tool_call_repair.dart
+++ b/bindings/flutter/lib/tool_call_repair.dart
@@ -98,7 +98,7 @@ ToolCallRepairContext? _contextFor(
 
 Map<String, dynamic> _replyFor(RawToolCall? replacement) => replacement == null
     ? const {'type': 'unchanged'}
-    : {'type': 'repaired', 'tool_call': replacement.toJson()};
+    : {'type': 'repaired', 'toolCall': replacement.toJson()};
 
 /// A thrown object has no typed counterpart in the core, so it becomes
 /// `failed` — the core turns that into `ToolCallRepair{cause: Other(message)}`.
@@ -129,7 +129,7 @@ Never _asyncHookInSyncCall() => throw StateError(
 /// derives the tool set, messages and instructions from them.
 ///
 /// Call this on the raw result map *before* decoding it into typed classes —
-/// both `tool_calls` and the matching `response_messages` part are rewritten.
+/// both `toolCalls` and the matching `responseMessages` part are rewritten.
 Map<String, dynamic> repairResultToolCalls(
   Map<String, dynamic> result,
   Object prompt,
@@ -168,7 +168,7 @@ Map<String, dynamic> repairResultToolCalls(
         _applyRepairToResult,
         jsonEncode(current),
         optsJson,
-        call['tool_call_id'] as String,
+        call['toolCallId'] as String,
         encodeJson(reply, 'repairToolCall reply'),
         'apply_tool_call_repair_to_result')) as Map<String, dynamic>;
   }
@@ -178,8 +178,8 @@ Map<String, dynamic> repairResultToolCalls(
 /// The invalid calls of a result document. `generate_object` nests the
 /// `generate_text` result under `raw`.
 List<Map<String, dynamic>> _invalidToolCalls(Map<String, dynamic> result) {
-  final calls = result['tool_calls'] ??
-      (result['raw'] as Map<String, dynamic>?)?['tool_calls'];
+  final calls = result['toolCalls'] ??
+      (result['raw'] as Map<String, dynamic>?)?['toolCalls'];
   if (calls is! List) return const [];
   return calls
       .whereType<Map<String, dynamic>>()
@@ -191,7 +191,7 @@ List<Map<String, dynamic>> _invalidToolCalls(Map<String, dynamic> result) {
 // Streaming: patch the tool-call part in flight
 // ─────────────────────────────────────────────────────────────────────────────
 
-/// Replace every invalid `ToolCall` part of [parts] with its repaired version.
+/// Replace every invalid `tool-call` part of [parts] with its repaired version.
 ///
 /// Returns [parts] unchanged when `options.repairToolCall` is null. Every other
 /// part — tool-input deltas included — passes through untouched and in order
@@ -209,9 +209,9 @@ Stream<Map<String, dynamic>> repairStreamToolCalls(
   final optsJson = encodeJson(options!.toJson(), 'options');
 
   return parts.asyncMap((part) async {
-    // Externally tagged: {"ToolCall": {...}}.
-    final call = part['ToolCall'];
-    if (call is! Map<String, dynamic> || call['invalid'] != true) return part;
+    // `{"type": "tool-call", ...ToolCall}`: the call is the part minus its tag.
+    if (part['type'] != 'tool-call' || part['invalid'] != true) return part;
+    final call = {...part}..remove('type');
 
     final context = _contextFor(call, promptJson, optsJson);
     if (context == null) return part;
@@ -225,6 +225,6 @@ Stream<Map<String, dynamic>> repairStreamToolCalls(
 
     final patched = _call3(_applyRepair, jsonEncode(call), optsJson,
         encodeJson(reply, 'repairToolCall reply'), 'apply_tool_call_repair');
-    return {'ToolCall': jsonDecode(patched)};
+    return {'type': 'tool-call', ...jsonDecode(patched) as Map<String, dynamic>};
   });
 }
diff --git a/bindings/flutter/lib/typed_model.dart b/bindings/flutter/lib/typed_model.dart
index 6c924baa3..ee36d71d2 100644
--- a/bindings/flutter/lib/typed_model.dart
+++ b/bindings/flutter/lib/typed_model.dart
@@ -67,7 +67,7 @@ class TypedModel {
   /// Generate a structured JSON object from a string [prompt] (M12).
   ///
   /// Same signature as [generateText]; returns a typed
-  /// [GenerateObjectResult]. Pass `response_format: { "Json": { ... } }`
+  /// [GenerateObjectResult]. Pass `responseFormat: {"type": "json", ...}`
   /// via [options] for schema control; the engine applies JSON repair
   /// before parsing.
   GenerateObjectResult generateObject(
diff --git a/bindings/flutter/lib/types.dart b/bindings/flutter/lib/types.dart
index 7293725f7..2b8885754 100644
--- a/bindings/flutter/lib/types.dart
+++ b/bindings/flutter/lib/types.dart
@@ -1,22 +1,15 @@
-// types.dart — typed models mirroring bindings/node/src/types/*.ts (ts-rs).
+// types.dart — typed models for the JSON the Rust core exchanges across the
+// C ABI. Mirrors the ts-rs generated types in `bindings/node/src/types/*.ts`:
+// camelCase field names, unions tagged with a `type` key (kebab-case values;
+// `sourceType` for sources), optional fields **absent** rather than `null`.
 //
-// Hand-written Dart classes with json_serializable. Field names are
-// camelCase; `@JsonKey(name: ...)` maps to the snake_case wire format
-// produced by the Rust core (serde). These wrap the raw
-// `Map<String, dynamic>` boundary exposed by the dart:ffi binding in
-// `aimux.dart`, so callers get compile-time types instead of dynamic maps.
+// The classes are hand-written: every `fromJson` reads exactly that JSON and
+// every `toJson` writes it back (nested values are plain maps, ready for
+// `jsonEncode`). Only the OpenAI Chat Completions family at the bottom keeps
+// OpenAI's own snake_case and uses `json_serializable`.
 //
-// Wire shapes are derived from the Rust structs in aimux-core
-// (`types.rs`, `tool.rs`, `generate.rs`, `result.rs`, `stream_part.rs`,
-// `shared.rs`):
-//   - GenerateTextResult  (generate.rs:90)
-//   - GenerateContent      (result.rs)        — sealed, externally-tagged
-//   - StreamPart           (stream_part.rs)   — sealed, externally-tagged
-//   - ContentPart          (content.rs)       — sealed, internally-tagged (type)
-//   - FileData/FileBytes   (shared.rs)        — sealed, externally-tagged
-//   - ToolCall             (tool.rs:102)
-//   - Usage / InputTokenUsage / OutputTokenUsage   (types.rs:33, types.rs:44)
-//   - FinishReason         (types.rs:10)
+// Weak-typed fields (`Map<String, dynamic>`) are the ones the TS types leave
+// open (`JsonValue`, provider metadata, warnings, errors).
 
 import 'dart:async';
 
@@ -24,6 +17,15 @@ import 'package:json_annotation/json_annotation.dart';
 
 part 'types.g.dart';
 
+typedef Json = Map<String, dynamic>;
+
+List<T> _list<T>(Object? v, T Function(Json) f) =>
+    [for (final e in (v as List? ?? const <Object?>[])) f(e as Json)];
+List<Json> _maps(Object? v) => _list(v, (m) => m);
+Json? _opt(Object? v) => v as Json?;
+List<T>? _optList<T>(Object? v, T Function(Json) f) =>
+    v == null ? null : _list(v, f);
+
 // ─────────────────────────────────────────────────────────────────────────────
 // Shared enums
 // ─────────────────────────────────────────────────────────────────────────────
@@ -73,98 +75,103 @@ enum ReasoningEffort {
 }
 
 // ─────────────────────────────────────────────────────────────────────────────
-// Token usage
+// Token usage, finish reason
 // ─────────────────────────────────────────────────────────────────────────────
 
-/// Input token usage detail with cache breakdown.
-///
-/// All fields are nullable: `total` is always serialized by the Rust core
-/// (null when unknown), the rest are omitted when `None`.
-@JsonSerializable()
+/// Input token usage detail with cache breakdown. Mirrors `InputTokenUsage.ts`.
 class InputTokenUsage {
   final int? total;
-  @JsonKey(name: 'no_cache') final int? noCache;
-  @JsonKey(name: 'cache_read') final int? cacheRead;
-  @JsonKey(name: 'cache_write') final int? cacheWrite;
-  InputTokenUsage({this.total, this.noCache, this.cacheRead, this.cacheWrite});
-  factory InputTokenUsage.fromJson(Map<String, dynamic> json) => _$InputTokenUsageFromJson(json);
-  Map<String, dynamic> toJson() => _$InputTokenUsageToJson(this);
+  final int? noCache;
+  final int? cacheRead;
+  final int? cacheWrite;
+  const InputTokenUsage({this.total, this.noCache, this.cacheRead, this.cacheWrite});
+  factory InputTokenUsage.fromJson(Json j) => InputTokenUsage(
+      total: j['total'] as int?,
+      noCache: j['noCache'] as int?,
+      cacheRead: j['cacheRead'] as int?,
+      cacheWrite: j['cacheWrite'] as int?);
+  Json toJson() => {
+        if (total != null) 'total': total,
+        if (noCache != null) 'noCache': noCache,
+        if (cacheRead != null) 'cacheRead': cacheRead,
+        if (cacheWrite != null) 'cacheWrite': cacheWrite,
+      };
 }
 
-@JsonSerializable()
+/// Mirrors `OutputTokenUsage.ts`.
 class OutputTokenUsage {
   final int? total;
   final int? text;
   final int? reasoning;
-  OutputTokenUsage({this.total, this.text, this.reasoning});
-  factory OutputTokenUsage.fromJson(Map<String, dynamic> json) => _$OutputTokenUsageFromJson(json);
-  Map<String, dynamic> toJson() => _$OutputTokenUsageToJson(this);
+  const OutputTokenUsage({this.total, this.text, this.reasoning});
+  factory OutputTokenUsage.fromJson(Json j) => OutputTokenUsage(
+      total: j['total'] as int?,
+      text: j['text'] as int?,
+      reasoning: j['reasoning'] as int?);
+  Json toJson() => {
+        if (total != null) 'total': total,
+        if (text != null) 'text': text,
+        if (reasoning != null) 'reasoning': reasoning,
+      };
 }
 
 /// Token usage statistics. Mirrors `Usage.ts`.
-@JsonSerializable()
 class Usage {
-  @JsonKey(name: 'input_tokens')
   final InputTokenUsage inputTokens;
-  @JsonKey(name: 'output_tokens')
   final OutputTokenUsage outputTokens;
-  /// Raw usage information from the provider (opaque JSON).
-  final Map<String, dynamic>? raw;
-
-  Usage({required this.inputTokens, required this.outputTokens, this.raw});
 
-  factory Usage.fromJson(Map<String, dynamic> json) => _$UsageFromJson(json);
-  Map<String, dynamic> toJson() => _$UsageToJson(this);
+  /// Raw usage information from the provider (opaque JSON).
+  final Json? raw;
+
+  const Usage({required this.inputTokens, required this.outputTokens, this.raw});
+  factory Usage.fromJson(Json j) => Usage(
+      inputTokens: InputTokenUsage.fromJson(j['inputTokens'] as Json),
+      outputTokens: OutputTokenUsage.fromJson(j['outputTokens'] as Json),
+      raw: _opt(j['raw']));
+  Json toJson() => {
+        'inputTokens': inputTokens.toJson(),
+        'outputTokens': outputTokens.toJson(),
+        if (raw != null) 'raw': raw,
+      };
 }
 
-// ─────────────────────────────────────────────────────────────────────────────
-// Finish reason
-// ─────────────────────────────────────────────────────────────────────────────
-
 /// Why generation stopped. Mirrors `FinishReason.ts`.
 ///
-/// `unified` is the kebab-case unified reason (`"stop"`, `"length"`,
-/// `"content-filter"`, `"tool-calls"`, `"error"`, `"other"`); `raw` is the
-/// provider-specific reason string (nullable).
-@JsonSerializable()
+/// [unified] is the kebab-case unified reason ([FinishReasonUnified]);
+/// [raw] is the provider-specific reason string.
 class FinishReason {
   final String unified;
   final String? raw;
-
-  FinishReason({required this.unified, this.raw});
-
-  factory FinishReason.fromJson(Map<String, dynamic> json) =>
-      _$FinishReasonFromJson(json);
-  Map<String, dynamic> toJson() => _$FinishReasonToJson(this);
+  const FinishReason({required this.unified, this.raw});
+  factory FinishReason.fromJson(Json j) =>
+      FinishReason(unified: j['unified'] as String, raw: j['raw'] as String?);
+  Json toJson() => {'unified': unified, if (raw != null) 'raw': raw};
 }
 
 // ─────────────────────────────────────────────────────────────────────────────
 // Tool calls
 // ─────────────────────────────────────────────────────────────────────────────
 
-/// A tool call requested by the model. Mirrors `ToolCall.ts`.
-@JsonSerializable()
+/// A parsed tool call requested by the model. Mirrors `ToolCall.ts`.
+///
+/// [error] is the typed lookup, parse, schema or repair failure of an invalid
+/// call: an `AiMuxError` object keyed by `name` (`{"name":"AI_NoSuchToolError",
+/// "toolName":...}`).
 class ToolCall {
-  @JsonKey(name: 'tool_call_id')
   final String toolCallId;
-  @JsonKey(name: 'tool_name')
   final String toolName;
   final dynamic input;
-  @JsonKey(name: 'provider_executed')
   final bool? providerExecuted;
-  // `dynamic` is a Dart built-in identifier — field is named `isDynamic`
-  // and mapped to the `dynamic` JSON key.
-  @JsonKey(name: 'dynamic')
+
+  /// The wire key is `dynamic`, a Dart built-in identifier, hence the name.
   final bool? isDynamic;
-  /// Additional provider-specific metadata associated with this call.
-  @JsonKey(name: 'provider_metadata', includeIfNull: false)
-  final dynamic providerMetadata;
-  /// Set by Core when the tool call stays invalid after optional repair.
+  final Json? providerMetadata;
+
+  /// Set by the core when the call stays invalid after optional repair.
   final bool? invalid;
-  /// The typed lookup, parse, schema, or repair failure for an invalid call.
-  final dynamic error;
+  final Json? error;
 
-  ToolCall({
+  const ToolCall({
     required this.toolCallId,
     required this.toolName,
     required this.input,
@@ -175,9 +182,27 @@ class ToolCall {
     this.error,
   });
 
-  factory ToolCall.fromJson(Map<String, dynamic> json) =>
-      _$ToolCallFromJson(json);
-  Map<String, dynamic> toJson() => _$ToolCallToJson(this);
+  factory ToolCall.fromJson(Json j) => ToolCall(
+        toolCallId: j['toolCallId'] as String,
+        toolName: j['toolName'] as String,
+        input: j['input'],
+        providerExecuted: j['providerExecuted'] as bool?,
+        isDynamic: j['dynamic'] as bool?,
+        providerMetadata: _opt(j['providerMetadata']),
+        invalid: j['invalid'] as bool?,
+        error: _opt(j['error']),
+      );
+
+  Json toJson() => {
+        'toolCallId': toolCallId,
+        'toolName': toolName,
+        'input': input,
+        if (providerExecuted != null) 'providerExecuted': providerExecuted,
+        if (isDynamic != null) 'dynamic': isDynamic,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+        if (invalid != null) 'invalid': invalid,
+        if (error != null) 'error': error,
+      };
 }
 
 // ─────────────────────────────────────────────────────────────────────────────
@@ -185,14 +210,14 @@ class ToolCall {
 // ─────────────────────────────────────────────────────────────────────────────
 
 /// A tool call as the model emitted it: [input] is the provider's raw argument
-/// **text**, not a decoded object.
+/// **text**, not a decoded object. Mirrors `RawToolCall.ts`.
 ///
 /// Both the call handed to a [RepairToolCall] hook
 /// ([ToolCallRepairContext.toolCall]) and the replacement it returns. The
 /// replacement is re-parsed and re-validated against the tool's schema, so
 /// [input] must be the argument text the tool expects (usually a JSON object
 /// literal); returning text that still fails validation leaves the call
-/// invalid, now carrying a [ToolCallRepairError].
+/// invalid, now carrying a `ToolCallRepairError`.
 class RawToolCall {
   final String toolCallId;
   final String toolName;
@@ -202,15 +227,12 @@ class RawToolCall {
 
   /// Whether the provider executes the call itself. The core adopts the
   /// replacement as returned, so a hook must carry this over (as
-  /// `copyWith` does) or the call turns into a client-executed one.
+  /// [copyWith] does) or the call turns into a client-executed one.
   final bool? providerExecuted;
 
-  /// Whether the call targets a dynamic tool. `dynamic` is a Dart built-in
-  /// identifier, so the field is `isDynamic` and maps to the `dynamic` key.
+  /// Whether the call targets a dynamic tool (the wire key is `dynamic`).
   final bool? isDynamic;
-
-  /// Additional provider-specific metadata associated with this call.
-  final dynamic providerMetadata;
+  final Json? providerMetadata;
 
   const RawToolCall({
     required this.toolCallId,
@@ -221,13 +243,13 @@ class RawToolCall {
     this.providerMetadata,
   });
 
-  factory RawToolCall.fromJson(Map<String, dynamic> json) => RawToolCall(
-        toolCallId: json['tool_call_id'] as String,
-        toolName: json['tool_name'] as String,
-        input: json['input'] as String,
-        providerExecuted: json['provider_executed'] as bool?,
-        isDynamic: json['dynamic'] as bool?,
-        providerMetadata: json['provider_metadata'],
+  factory RawToolCall.fromJson(Json j) => RawToolCall(
+        toolCallId: j['toolCallId'] as String,
+        toolName: j['toolName'] as String,
+        input: j['input'] as String,
+        providerExecuted: j['providerExecuted'] as bool?,
+        isDynamic: j['dynamic'] as bool?,
+        providerMetadata: _opt(j['providerMetadata']),
       );
 
   /// A copy with [input] (and optionally the id or name) replaced and every
@@ -242,13 +264,13 @@ class RawToolCall {
         providerMetadata: providerMetadata,
       );
 
-  Map<String, dynamic> toJson() => {
-        'tool_call_id': toolCallId,
-        'tool_name': toolName,
+  Json toJson() => {
+        'toolCallId': toolCallId,
+        'toolName': toolName,
         'input': input,
-        if (providerExecuted != null) 'provider_executed': providerExecuted,
+        if (providerExecuted != null) 'providerExecuted': providerExecuted,
         if (isDynamic != null) 'dynamic': isDynamic,
-        if (providerMetadata != null) 'provider_metadata': providerMetadata,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
@@ -262,13 +284,12 @@ class ToolCallRepairContext {
   /// The invalid call, with the provider's raw argument text.
   final RawToolCall toolCall;
 
-  /// The lookup, parse or validation failure, externally tagged
-  /// (`{"InvalidToolInput": {...}}`). Weak type — same shape as
-  /// [ToolCall.error].
-  final Object? error;
+  /// The lookup, parse or validation failure as an `AiMuxError` object keyed by
+  /// `name` — the same shape as [ToolCall.error].
+  final Json? error;
 
-  /// The schema [toolCall] failed to satisfy; null when the tool is unknown.
-  final Map<String, dynamic>? inputSchema;
+  /// The schema [toolCall] failed to satisfy.
+  final Json? inputSchema;
 
   /// The full tool set the call was generated with.
   final List<Tool> tools;
@@ -288,19 +309,13 @@ class ToolCallRepairContext {
     this.instructions,
   });
 
-  factory ToolCallRepairContext.fromJson(Map<String, dynamic> json) =>
-      ToolCallRepairContext(
-        toolCall:
-            RawToolCall.fromJson(json['tool_call'] as Map<String, dynamic>),
-        error: json['error'],
-        inputSchema: json['input_schema'] as Map<String, dynamic>?,
-        tools: (json['tools'] as List<dynamic>? ?? const [])
-            .map((t) => Tool.fromJson(t as Map<String, dynamic>))
-            .toList(),
-        messages: (json['messages'] as List<dynamic>? ?? const [])
-            .map((m) => ModelMessage.fromJson(m as Map<String, dynamic>))
-            .toList(),
-        instructions: json['instructions'] as String?,
+  factory ToolCallRepairContext.fromJson(Json j) => ToolCallRepairContext(
+        toolCall: RawToolCall.fromJson(j['toolCall'] as Json),
+        error: _opt(j['error']),
+        inputSchema: _opt(j['inputSchema']),
+        tools: _list(j['tools'], Tool.fromJson),
+        messages: _list(j['messages'], ModelMessage.fromJson),
+        instructions: j['instructions'] as String?,
       );
 }
 
@@ -309,7 +324,7 @@ class ToolCallRepairContext {
 ///
 /// Return a [RawToolCall] to replace the call, or null to leave it as it is.
 /// Throwing means the repair failed: the call stays invalid and carries a
-/// [ToolCallRepairError] whose cause is the thrown object's `toString()`.
+/// `ToolCallRepairError` whose cause is the thrown object's `toString()`.
 ///
 /// The hook runs on the caller's isolate, after the native call has returned,
 /// so it may itself call back into aimux (including another model call).
@@ -324,27 +339,23 @@ typedef RepairToolCall = FutureOr<RawToolCall?> Function(
 // ─────────────────────────────────────────────────────────────────────────────
 
 class FunctionToolInputExample {
-  final Map<String, dynamic> input;
-  FunctionToolInputExample({required this.input});
-  factory FunctionToolInputExample.fromJson(Map<String, dynamic> json) =>
-      FunctionToolInputExample(input: json['input'] as Map<String, dynamic>);
-  Map<String, dynamic> toJson() => {'input': input};
+  final Json input;
+  const FunctionToolInputExample({required this.input});
+  factory FunctionToolInputExample.fromJson(Json j) =>
+      FunctionToolInputExample(input: j['input'] as Json);
+  Json toJson() => {'input': input};
 }
 
 /// A function tool definition. Mirrors `FunctionTool.ts`.
-@JsonSerializable()
 class FunctionTool {
   final String name;
   final String? description;
-  @JsonKey(name: 'input_schema')
-  final Map<String, dynamic> inputSchema;
+  final Json inputSchema;
   final bool? strict;
-  @JsonKey(name: 'provider_options')
-  final Map<String, dynamic>? providerOptions;
-  @JsonKey(name: 'input_examples')
+  final Json? providerOptions;
   final List<FunctionToolInputExample>? inputExamples;
 
-  FunctionTool({
+  const FunctionTool({
     required this.name,
     this.description,
     required this.inputSchema,
@@ -353,36 +364,52 @@ class FunctionTool {
     this.inputExamples,
   });
 
-  factory FunctionTool.fromJson(Map<String, dynamic> json) =>
-      _$FunctionToolFromJson(json);
-  Map<String, dynamic> toJson() => _$FunctionToolToJson(this);
+  factory FunctionTool.fromJson(Json j) => FunctionTool(
+        name: j['name'] as String,
+        description: j['description'] as String?,
+        inputSchema: j['inputSchema'] as Json,
+        strict: j['strict'] as bool?,
+        providerOptions: _opt(j['providerOptions']),
+        inputExamples:
+            _optList(j['inputExamples'], FunctionToolInputExample.fromJson),
+      );
+
+  Json toJson() => {
+        'name': name,
+        if (description != null) 'description': description,
+        'inputSchema': inputSchema,
+        if (strict != null) 'strict': strict,
+        if (providerOptions != null) 'providerOptions': providerOptions,
+        if (inputExamples != null)
+          'inputExamples': [for (final e in inputExamples!) e.toJson()],
+      };
 }
 
-/// A tool definition: function or provider tool. Mirrors `Tool.ts`.
+/// A tool definition: function or provider tool. Mirrors `Tool.ts`
+/// (`{"type":"function", ...FunctionTool}` / `{"type":"provider", id, name,
+/// args}`).
 class Tool {
   final String type;
   final FunctionTool? function;
   final String? id;
   final String? name;
-  final Map<String, dynamic>? args;
+  final Json? args;
 
-  Tool._({required this.type, this.function, this.id, this.name, this.args});
+  const Tool._({required this.type, this.function, this.id, this.name, this.args});
 
-  factory Tool.function(FunctionTool fn) =>
-      Tool._(type: 'function', function: fn);
-  factory Tool.provider({required String id, required String name, required Map<String, dynamic> args}) =>
+  factory Tool.function(FunctionTool fn) => Tool._(type: 'function', function: fn);
+  factory Tool.provider(
+          {required String id, required String name, required Json args}) =>
       Tool._(type: 'provider', id: id, name: name, args: args);
 
-  Map<String, dynamic> toJson() {
-    if (function != null) return {'type': 'function', ...function!.toJson()};
-    return {'type': 'provider', 'id': id!, 'name': name!, 'args': args ?? {}};
-  }
-  factory Tool.fromJson(Map<String, dynamic> json) {
-    if (json['type'] == 'function') {
-      return Tool._(type: 'function', function: FunctionTool.fromJson(json));
-    }
-    return Tool._(type: 'provider', id: json['id'] as String, name: json['name'] as String, args: json['args'] as Map<String, dynamic>?);
-  }
+  Json toJson() => function != null
+      ? {'type': 'function', ...function!.toJson()}
+      : {'type': 'provider', 'id': id!, 'name': name!, 'args': args ?? {}};
+
+  factory Tool.fromJson(Json j) => j['type'] == 'function'
+      ? Tool.function(FunctionTool.fromJson(j))
+      : Tool.provider(
+          id: j['id'] as String, name: j['name'] as String, args: j['args'] as Json);
 }
 
 /// How the model should choose tools. Mirrors `ToolChoice.ts`.
@@ -395,202 +422,231 @@ class ToolChoice {
   static const required = ToolChoice._('required', null);
   factory ToolChoice.tool(String toolName) => ToolChoice._('tool', toolName);
 
-  dynamic toJson() => _kind == 'tool' ? {'type': 'tool', 'toolName': toolName} : _kind;
-  factory ToolChoice.fromJson(dynamic json) =>
-    json is String ? ToolChoice._(json, null) : ToolChoice._('tool', (json as Map<String, dynamic>)['toolName'] as String);
+  dynamic toJson() =>
+      _kind == 'tool' ? {'type': 'tool', 'toolName': toolName} : _kind;
+  factory ToolChoice.fromJson(dynamic json) => json is String
+      ? ToolChoice._(json, null)
+      : ToolChoice._('tool', (json as Json)['toolName'] as String);
 }
 
 // ─────────────────────────────────────────────────────────────────────────────
-// GenerateResult (raw provider result)
+// Request / response metadata
 // ─────────────────────────────────────────────────────────────────────────────
 
-@JsonSerializable()
-class ResponseMetadata {
-  final String? id;
-  final String? timestamp;
-  @JsonKey(name: 'model_id') final String? modelId;
-  ResponseMetadata({this.id, this.timestamp, this.modelId});
-  factory ResponseMetadata.fromJson(Map<String, dynamic> json) => _$ResponseMetadataFromJson(json);
-  Map<String, dynamic> toJson() => _$ResponseMetadataToJson(this);
-}
-
-class GenerateRequest {
+/// Mirrors `RequestInfo.ts`.
+class RequestInfo {
   final dynamic body;
-  GenerateRequest({this.body});
-  factory GenerateRequest.fromJson(Map<String, dynamic> json) => GenerateRequest(body: json['body']);
-  Map<String, dynamic> toJson() => {'body': body};
+  const RequestInfo({this.body});
+  factory RequestInfo.fromJson(Json j) => RequestInfo(body: j['body']);
+  Json toJson() => {if (body != null) 'body': body};
 }
 
-class GenerateResponse {
+/// Mirrors `ResponseInfo.ts`.
+class ResponseInfo {
   final String? id;
   final String? timestamp;
   final String? modelId;
   final Map<String, String>? headers;
   final dynamic body;
-  GenerateResponse({this.id, this.timestamp, this.modelId, this.headers, this.body});
-  factory GenerateResponse.fromJson(Map<String, dynamic> json) => GenerateResponse(id: json['id'] as String?, timestamp: json['timestamp'] as String?, modelId: json['model_id'] as String?, headers: (json['headers'] as Map<String, dynamic>?)?.map((key, value) => MapEntry(key, value as String)), body: json['body']);
-  Map<String, dynamic> toJson() => {'id': id, 'timestamp': timestamp, 'model_id': modelId, if (headers != null) 'headers': headers, if (body != null) 'body': body};
+  const ResponseInfo({this.id, this.timestamp, this.modelId, this.headers, this.body});
+  factory ResponseInfo.fromJson(Json j) => ResponseInfo(
+      id: j['id'] as String?,
+      timestamp: j['timestamp'] as String?,
+      modelId: j['modelId'] as String?,
+      headers: (j['headers'] as Json?)?.cast<String, String>(),
+      body: j['body']);
+  Json toJson() => {
+        if (id != null) 'id': id,
+        if (timestamp != null) 'timestamp': timestamp,
+        if (modelId != null) 'modelId': modelId,
+        if (headers != null) 'headers': headers,
+        if (body != null) 'body': body,
+      };
 }
 
-class GenerateRequest {
-  final dynamic body;
-  GenerateRequest({this.body});
-  factory GenerateRequest.fromJson(Map<String, dynamic> json) => GenerateRequest(body: json['body']);
-  Map<String, dynamic> toJson() => {if (body != null) 'body': body};
+/// Mirrors `ResponseMetadata.ts`.
+class ResponseMetadata {
+  final String? id;
+  final String? timestamp;
+  final String? modelId;
+  const ResponseMetadata({this.id, this.timestamp, this.modelId});
+  factory ResponseMetadata.fromJson(Json j) => ResponseMetadata(
+      id: j['id'] as String?,
+      timestamp: j['timestamp'] as String?,
+      modelId: j['modelId'] as String?);
+  Json toJson() => {
+        if (id != null) 'id': id,
+        if (timestamp != null) 'timestamp': timestamp,
+        if (modelId != null) 'modelId': modelId,
+      };
 }
 
-/// A content item in a `GenerateResult`. Mirrors `GenerateContent.ts`
-/// (`aimux-core/src/result.rs`).
-///
-/// Externally-tagged union — each item is a single-key map whose key is the
-/// variant tag and whose value is the variant payload. Modeled as a sealed
-/// class hierarchy so callers get compile-time types: every known variant is a
-/// typed subclass with named fields, and [GenerateContentUnknown] is the
-/// fallback for tags added by newer core versions so they pass through verbatim
-/// instead of crashing.
-///
-/// Known variants: [GenerateContentText], [GenerateContentToolCall],
-/// [GenerateContentSource], [GenerateContentReasoning], [GenerateContentFile],
-/// [GenerateContentToolResult]. The `File` variant carries model-generated
-/// files (e.g. images or documents) as a `FileData` tagged union under `data`,
-/// alongside a `media_type` (there is **no** `filename` field). Narrow via
-/// `switch`/`is`/`whereType` to read variant-specific fields.
-sealed class GenerateContent {
-  const GenerateContent();
+// ─────────────────────────────────────────────────────────────────────────────
+// File data, generated files, sources
+// ─────────────────────────────────────────────────────────────────────────────
 
-  /// The variant tag — the single top-level key on the wire, e.g. `'Text'`,
-  /// `'ToolCall'`. Useful for logging / generic dispatch.
-  String get tag;
-
-  /// Re-encode to the externally-tagged wire shape (`{tag: payload}`).
-  Map<String, dynamic> toJson();
-
-  /// Decode an externally-tagged content map. Unknown tags fall back to
-  /// [GenerateContentUnknown] instead of throwing.
-  factory GenerateContent.fromJson(Map<String, dynamic> json) {
-    final e = json.entries.first;
-    final payload = e.value as Map<String, dynamic>;
-    return switch (e.key) {
-      'Text' => GenerateContentText.fromJson(payload),
-      'ToolCall' => GenerateContentToolCall.fromJson(payload),
-      'Source' => GenerateContentSource.fromJson(payload),
-      'Reasoning' => GenerateContentReasoning.fromJson(payload),
-      'Custom' => GenerateContentCustom.fromJson(payload),
-      'ReasoningFile' => GenerateContentReasoningFile.fromJson(payload),
-      'ToolApprovalRequest' => GenerateContentToolApprovalRequest.fromJson(payload),
-      'File' => GenerateContentFile.fromJson(payload),
-      'ToolResult' => GenerateContentToolResult.fromJson(payload),
-      _ => GenerateContentUnknown(tag: e.key, data: payload),
-    };
-  }
+/// Raw bytes (a JSON array of ints) or a base64 string — `FileBytes.ts` is the
+/// untagged union `Array<number> | string`.
+sealed class FileBytes {
+  const FileBytes();
+  factory FileBytes.fromJson(Object json) => json is String
+      ? FileBytesBase64(data: json)
+      : FileBytesBinary(data: (json as List).cast<int>());
+  Object toJson();
 }
 
-/// Generated text.
-final class GenerateContentText extends GenerateContent {
-  final String text;
-  final Map<String, dynamic>? providerMetadata;
-
-  GenerateContentText({required this.text, this.providerMetadata});
+final class FileBytesBinary extends FileBytes {
+  final List<int> data;
+  const FileBytesBinary({required this.data});
+  @override
+  Object toJson() => data;
+}
 
+final class FileBytesBase64 extends FileBytes {
+  final String data;
+  const FileBytesBase64({required this.data});
   @override
-  String get tag => 'Text';
+  Object toJson() => data;
+}
 
-  factory GenerateContentText.fromJson(Map<String, dynamic> json) =>
-      GenerateContentText(
-        text: json['text'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
+/// File data of a file part. Mirrors `FileData.ts`:
+/// `{"type":"data"|"url"|"reference"|"text", ...}`.
+sealed class FileData {
+  const FileData();
+  factory FileData.fromJson(Json j) => switch (j['type']) {
+        'data' => FileDataData(data: FileBytes.fromJson(j['data'] as Object)),
+        'url' => FileDataUrl(
+            url: j['url'] as String, originalUrl: j['originalUrl'] as String?),
+        'reference' => FileDataReference(reference: j['reference'] as Json),
+        'text' => FileDataText(text: j['text'] as String),
+        _ => throw FormatException('unknown FileData type: ${j['type']}'),
+      };
+  Json toJson();
+}
+
+final class FileDataData extends FileData {
+  final FileBytes data;
+  const FileDataData({required this.data});
+  @override
+  Json toJson() => {'type': 'data', 'data': data.toJson()};
+}
 
+final class FileDataUrl extends FileData {
+  final String url;
+  final String? originalUrl;
+  const FileDataUrl({required this.url, this.originalUrl});
   @override
-  Map<String, dynamic> toJson() => {
-        'Text': {
-          'text': text,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'url',
+        'url': url,
+        if (originalUrl != null) 'originalUrl': originalUrl,
       };
 }
 
-/// A tool call requested by the model.
-final class GenerateContentToolCall extends GenerateContent {
-  final String toolCallId;
-  final String toolName;
-  final dynamic input;
-  final bool? providerExecuted;
-  final bool? isDynamic;
-  final Map<String, dynamic>? providerMetadata;
+final class FileDataReference extends FileData {
+  final Json reference;
+  const FileDataReference({required this.reference});
+  @override
+  Json toJson() => {'type': 'reference', 'reference': reference};
+}
 
-  GenerateContentToolCall({
-    required this.toolCallId,
-    required this.toolName,
-    required this.input,
-    this.providerExecuted,
-    this.isDynamic,
-    this.providerMetadata,
-  });
+final class FileDataText extends FileData {
+  final String text;
+  const FileDataText({required this.text});
+  @override
+  Json toJson() => {'type': 'text', 'text': text};
+}
+
+/// Data or a URL returned for a generated file. Mirrors `GeneratedFileData.ts`.
+sealed class GeneratedFileData {
+  const GeneratedFileData();
+  factory GeneratedFileData.fromJson(Json j) => switch (j['type']) {
+        'data' =>
+          GeneratedFileDataData(data: FileBytes.fromJson(j['data'] as Object)),
+        'url' => GeneratedFileDataUrl(
+            url: j['url'] as String, originalUrl: j['originalUrl'] as String?),
+        _ => throw FormatException('unknown GeneratedFileData type: ${j['type']}'),
+      };
+  Json toJson();
+}
 
+final class GeneratedFileDataData extends GeneratedFileData {
+  final FileBytes data;
+  const GeneratedFileDataData({required this.data});
   @override
-  String get tag => 'ToolCall';
-
-  factory GenerateContentToolCall.fromJson(Map<String, dynamic> json) =>
-      GenerateContentToolCall(
-        toolCallId: json['tool_call_id'] as String,
-        toolName: json['tool_name'] as String,
-        input: json['input'],
-        providerExecuted: json['provider_executed'] as bool?,
-        isDynamic: json['dynamic'] as bool?,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
+  Json toJson() => {'type': 'data', 'data': data.toJson()};
+}
 
+final class GeneratedFileDataUrl extends GeneratedFileData {
+  final String url;
+  final String? originalUrl;
+  const GeneratedFileDataUrl({required this.url, this.originalUrl});
   @override
-  Map<String, dynamic> toJson() => {
-        'ToolCall': {
-          'tool_call_id': toolCallId,
-          'tool_name': toolName,
-          'input': input,
-          if (providerExecuted != null) 'provider_executed': providerExecuted,
-          if (isDynamic != null) 'dynamic': isDynamic,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'url',
+        'url': url,
+        if (originalUrl != null) 'originalUrl': originalUrl,
+      };
+}
+
+/// A file generated by the model. Mirrors `GeneratedFile.ts`.
+class GeneratedFile {
+  final GeneratedFileData data;
+  final String mediaType;
+  final Json? providerMetadata;
+  const GeneratedFile(
+      {required this.data, required this.mediaType, this.providerMetadata});
+  factory GeneratedFile.fromJson(Json j) => GeneratedFile(
+      data: GeneratedFileData.fromJson(j['data'] as Json),
+      mediaType: j['mediaType'] as String,
+      providerMetadata: _opt(j['providerMetadata']));
+  Json toJson() => {
+        'data': data.toJson(),
+        'mediaType': mediaType,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
+/// A source / citation. Mirrors `Source.ts`, tagged by `sourceType`
+/// (`"url"` / `"document"`).
 sealed class Source {
   const Source();
 
-  factory Source.fromJson(Map<String, dynamic> json) => switch (json['source_type']) {
-    'url' => UrlSource(
-        id: json['id'] as String,
-        url: json['url'] as String,
-        title: json['title'] as String?,
-        providerMetadata: json['provider_metadata'] as Map<String, dynamic>?),
-    'document' => DocumentSource(
-        id: json['id'] as String,
-        mediaType: json['media_type'] as String,
-        title: json['title'] as String,
-        filename: json['filename'] as String?,
-        providerMetadata: json['provider_metadata'] as Map<String, dynamic>?),
-    _ => throw FormatException('invalid source type: ${json['source_type']}'),
-  };
+  factory Source.fromJson(Json j) => switch (j['sourceType']) {
+        'url' => UrlSource(
+            id: j['id'] as String,
+            url: j['url'] as String,
+            title: j['title'] as String?,
+            providerMetadata: _opt(j['providerMetadata'])),
+        'document' => DocumentSource(
+            id: j['id'] as String,
+            mediaType: j['mediaType'] as String,
+            title: j['title'] as String,
+            filename: j['filename'] as String?,
+            providerMetadata: _opt(j['providerMetadata'])),
+        _ => throw FormatException('invalid source type: ${j['sourceType']}'),
+      };
 
-  Map<String, dynamic> toJson();
+  Json toJson();
 }
 
 final class UrlSource extends Source {
   final String id;
   final String url;
   final String? title;
-  final Map<String, dynamic>? providerMetadata;
+  final Json? providerMetadata;
 
-  const UrlSource({required this.id, required this.url, this.title, this.providerMetadata});
+  const UrlSource(
+      {required this.id, required this.url, this.title, this.providerMetadata});
 
   @override
-  Map<String, dynamic> toJson() => {
-    'source_type': 'url', 'id': id, 'url': url,
-    if (title != null) 'title': title,
-    if (providerMetadata != null) 'provider_metadata': providerMetadata,
-  };
+  Json toJson() => {
+        'sourceType': 'url',
+        'id': id,
+        'url': url,
+        if (title != null) 'title': title,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+      };
 }
 
 final class DocumentSource extends Source {
@@ -598,250 +654,637 @@ final class DocumentSource extends Source {
   final String mediaType;
   final String title;
   final String? filename;
-  final Map<String, dynamic>? providerMetadata;
+  final Json? providerMetadata;
 
-  const DocumentSource({required this.id, required this.mediaType, required this.title,
-      this.filename, this.providerMetadata});
+  const DocumentSource(
+      {required this.id,
+      required this.mediaType,
+      required this.title,
+      this.filename,
+      this.providerMetadata});
 
   @override
-  Map<String, dynamic> toJson() => {
-    'source_type': 'document', 'id': id, 'media_type': mediaType, 'title': title,
-    if (filename != null) 'filename': filename,
-    if (providerMetadata != null) 'provider_metadata': providerMetadata,
-  };
+  Json toJson() => {
+        'sourceType': 'document',
+        'id': id,
+        'mediaType': mediaType,
+        'title': title,
+        if (filename != null) 'filename': filename,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+      };
 }
 
-/// A source / citation (e.g. URL citation from search-preview models).
-final class GenerateContentSource extends GenerateContent {
-  final Source source;
-
-  GenerateContentSource({required this.source});
+// ─────────────────────────────────────────────────────────────────────────────
+// ContentPart (message content) — `{"type": "text", ...}`
+// ─────────────────────────────────────────────────────────────────────────────
 
-  @override
-  String get tag => 'Source';
+/// The output of a tool call in a `tool-result` part. Mirrors
+/// `ToolResultOutput.ts`: `{"type":"text"|"error-text","value":"…"}`,
+/// `{"type":"json"|"error-json","value":<json>}`,
+/// `{"type":"execution-denied","reason"?}` or `{"type":"content","value":[…]}`
+/// (`value` is then a list of text / file / custom content maps, kept as
+/// weak-typed maps).
+class ToolResultOutput {
+  final String type;
+  final dynamic value;
+  final String? reason;
+  final Json? providerOptions;
+
+  const ToolResultOutput._(this.type, {this.value, this.reason, this.providerOptions});
+
+  factory ToolResultOutput.text(String value, {Json? providerOptions}) =>
+      ToolResultOutput._('text', value: value, providerOptions: providerOptions);
+  factory ToolResultOutput.json(Object? value, {Json? providerOptions}) =>
+      ToolResultOutput._('json', value: value, providerOptions: providerOptions);
+  factory ToolResultOutput.errorText(String value, {Json? providerOptions}) =>
+      ToolResultOutput._('error-text', value: value, providerOptions: providerOptions);
+  factory ToolResultOutput.errorJson(Object? value, {Json? providerOptions}) =>
+      ToolResultOutput._('error-json', value: value, providerOptions: providerOptions);
+  factory ToolResultOutput.executionDenied({String? reason, Json? providerOptions}) =>
+      ToolResultOutput._('execution-denied',
+          reason: reason, providerOptions: providerOptions);
+  factory ToolResultOutput.content(List<Json> value) =>
+      ToolResultOutput._('content', value: value);
+
+  factory ToolResultOutput.fromJson(Json j) => ToolResultOutput._(
+        j['type'] as String,
+        value: j['value'],
+        reason: j['reason'] as String?,
+        providerOptions: _opt(j['providerOptions']),
+      );
 
-  factory GenerateContentSource.fromJson(Map<String, dynamic> json) =>
-      GenerateContentSource(source: Source.fromJson(json));
+  Json toJson() => {
+        'type': type,
+        if (type != 'execution-denied') 'value': value,
+        if (reason != null) 'reason': reason,
+        if (providerOptions != null) 'providerOptions': providerOptions,
+      };
+}
 
-  @override
-  Map<String, dynamic> toJson() => {'Source': source.toJson()};
+/// One part of a message's content. Mirrors `ContentPart.ts`; unknown `type`
+/// values fall back to [ContentPartUnknown] and pass through verbatim.
+sealed class ContentPart {
+  const ContentPart();
+  factory ContentPart.fromJson(Json j) => switch (j['type']) {
+        'text' => ContentPartText(
+            text: j['text'] as String, providerOptions: _opt(j['providerOptions'])),
+        'custom' => ContentPartCustom(
+            kind: j['kind'] as String, providerOptions: _opt(j['providerOptions'])),
+        'reasoning-file' => ContentPartReasoningFile(
+            data: GeneratedFileData.fromJson(j['data'] as Json),
+            mediaType: j['mediaType'] as String,
+            providerOptions: _opt(j['providerOptions'])),
+        'tool-approval-request' => ContentPartToolApprovalRequest(
+            approvalId: j['approvalId'] as String,
+            toolCallId: j['toolCallId'] as String,
+            reason: j['reason'] as String?,
+            isAutomatic: j['isAutomatic'] as bool?,
+            signature: j['signature'] as String?,
+            inputSchemaInput: j['inputSchemaInput']),
+        'image' => ContentPartImage(
+            image: (j['image'] as List).cast<int>(),
+            mediaType: j['mediaType'] as String,
+            providerOptions: _opt(j['providerOptions'])),
+        'file' => ContentPartFile(
+            data: (j['data'] as List).cast<int>(),
+            mediaType: j['mediaType'] as String,
+            filename: j['filename'] as String?,
+            providerOptions: _opt(j['providerOptions'])),
+        'file-base64' => ContentPartFileBase64(
+            data: j['data'] as String,
+            mediaType: j['mediaType'] as String,
+            filename: j['filename'] as String?,
+            providerOptions: _opt(j['providerOptions'])),
+        'file-url' => ContentPartFileUrl(
+            url: j['url'] as String,
+            mediaType: j['mediaType'] as String,
+            providerOptions: _opt(j['providerOptions'])),
+        'file-reference' => ContentPartFileReference(
+            mediaType: j['mediaType'] as String,
+            reference: j['reference'],
+            filename: j['filename'] as String?,
+            providerOptions: _opt(j['providerOptions'])),
+        'reasoning' => ContentPartReasoning(
+            text: j['text'] as String,
+            signature: j['signature'] as String?,
+            providerOptions: _opt(j['providerOptions'])),
+        'tool-call' => ContentPartToolCall(
+            toolCallId: j['toolCallId'] as String,
+            toolName: j['toolName'] as String,
+            input: j['input'],
+            providerExecuted: j['providerExecuted'] as bool?,
+            providerOptions: _opt(j['providerOptions'])),
+        'tool-result' => ContentPartToolResult(
+            toolCallId: j['toolCallId'] as String,
+            toolName: j['toolName'] as String,
+            output: ToolResultOutput.fromJson(j['output'] as Json),
+            providerOptions: _opt(j['providerOptions'])),
+        _ => ContentPartUnknown(j),
+      };
+  Json toJson();
 }
 
-/// A reasoning / thinking segment produced by the model.
-final class GenerateContentReasoning extends GenerateContent {
+final class ContentPartText extends ContentPart {
   final String text;
-  final Map<String, dynamic>? providerMetadata;
-
-  GenerateContentReasoning({required this.text, this.providerMetadata});
-
+  final Json? providerOptions;
+  const ContentPartText({required this.text, this.providerOptions});
   @override
-  String get tag => 'Reasoning';
-
-  factory GenerateContentReasoning.fromJson(Map<String, dynamic> json) =>
-      GenerateContentReasoning(
-        text: json['text'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
+  Json toJson() => {
+        'type': 'text',
+        'text': text,
+        if (providerOptions != null) 'providerOptions': providerOptions,
+      };
+}
 
+final class ContentPartCustom extends ContentPart {
+  final String kind;
+  final Json? providerOptions;
+  const ContentPartCustom({required this.kind, this.providerOptions});
   @override
-  Map<String, dynamic> toJson() => {
-        'Reasoning': {
-          'text': text,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'custom',
+        'kind': kind,
+        if (providerOptions != null) 'providerOptions': providerOptions,
       };
 }
 
-/// A file generated by the model (e.g. an image or document). The `data`
-/// field is a `FileData` tagged union; there is **no** `filename` field.
-final class GenerateContentFile extends GenerateContent {
+final class ContentPartReasoningFile extends ContentPart {
   final GeneratedFileData data;
   final String mediaType;
-  final Map<String, dynamic>? providerMetadata;
-
-  GenerateContentFile({
-    required this.data,
-    required this.mediaType,
-    this.providerMetadata,
-  });
-
-  @override
-  String get tag => 'File';
-
-  factory GenerateContentFile.fromJson(Map<String, dynamic> json) =>
-      GenerateContentFile(
-        data: GeneratedFileData.fromJson(json['data'] as Map<String, dynamic>),
-        mediaType: json['media_type'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
+  final Json? providerOptions;
+  const ContentPartReasoningFile(
+      {required this.data, required this.mediaType, this.providerOptions});
   @override
-  Map<String, dynamic> toJson() => {
-        'File': {
-          'data': data.toJson(),
-          'media_type': mediaType,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'reasoning-file',
+        'data': data.toJson(),
+        'mediaType': mediaType,
+        if (providerOptions != null) 'providerOptions': providerOptions,
       };
 }
 
-/// A tool result from a provider-executed tool (e.g. xAI file_search,
-/// web_search). Emitted alongside the preceding `ToolCall`.
-final class GenerateContentToolResult extends GenerateContent {
+final class ContentPartToolApprovalRequest extends ContentPart {
+  final String approvalId;
   final String toolCallId;
-  final String toolName;
-  final dynamic result;
-  final bool? isError;
-  final bool? preliminary;
-  final bool? isDynamic;
-  final Map<String, dynamic>? providerMetadata;
-
-  GenerateContentToolResult({
-    required this.toolCallId,
-    required this.toolName,
-    required this.result,
-    this.isError,
-    this.preliminary,
-    this.isDynamic,
-    this.providerMetadata,
-  });
-
-  @override
-  String get tag => 'ToolResult';
-
-  factory GenerateContentToolResult.fromJson(Map<String, dynamic> json) =>
-      GenerateContentToolResult(
-        toolCallId: json['tool_call_id'] as String,
-        toolName: json['tool_name'] as String,
-        result: json['result'],
-        isError: json['is_error'] as bool?,
-        preliminary: json['preliminary'] as bool?,
-        isDynamic: json['dynamic'] as bool?,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
+  final String? reason;
+  final bool? isAutomatic;
+  final String? signature;
+  final dynamic inputSchemaInput;
+  const ContentPartToolApprovalRequest(
+      {required this.approvalId,
+      required this.toolCallId,
+      this.reason,
+      this.isAutomatic,
+      this.signature,
+      this.inputSchemaInput});
   @override
-  Map<String, dynamic> toJson() => {
-        'ToolResult': {
-          'tool_call_id': toolCallId,
-          'tool_name': toolName,
-          'result': result,
-          if (isError != null) 'is_error': isError,
-          if (preliminary != null) 'preliminary': preliminary,
-          if (isDynamic != null) 'dynamic': isDynamic,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'tool-approval-request',
+        'approvalId': approvalId,
+        'toolCallId': toolCallId,
+        if (reason != null) 'reason': reason,
+        if (isAutomatic != null) 'isAutomatic': isAutomatic,
+        if (signature != null) 'signature': signature,
+        if (inputSchemaInput != null) 'inputSchemaInput': inputSchemaInput,
       };
 }
 
-/// Fallback for unknown/forward-compatible `GenerateContent` variants.
-///
-/// Newer core versions may emit tags this binding does not model yet; this
-/// class passes them through verbatim (`{tag: data}`) instead of discarding
-/// or crashing.
-final class GenerateContentCustom extends GenerateContent {
-  final String kind;
-  final Map<String, dynamic>? providerMetadata;
-  GenerateContentCustom({required this.kind, this.providerMetadata});
-  @override String get tag => 'Custom';
-  factory GenerateContentCustom.fromJson(Map<String, dynamic> json) => GenerateContentCustom(kind: json['kind'] as String, providerMetadata: json['provider_metadata'] as Map<String, dynamic>?);
-  @override Map<String, dynamic> toJson() => {'Custom': {'kind': kind, if (providerMetadata != null) 'provider_metadata': providerMetadata}};
-}
-
-final class GenerateContentReasoningFile extends GenerateContent {
-  final GeneratedFileData data;
+final class ContentPartImage extends ContentPart {
+  final List<int> image;
   final String mediaType;
-  final Map<String, dynamic>? providerMetadata;
-  GenerateContentReasoningFile({required this.data, required this.mediaType, this.providerMetadata});
-  @override String get tag => 'ReasoningFile';
-  factory GenerateContentReasoningFile.fromJson(Map<String, dynamic> json) => GenerateContentReasoningFile(data: GeneratedFileData.fromJson(json['data'] as Map<String, dynamic>), mediaType: json['media_type'] as String, providerMetadata: json['provider_metadata'] as Map<String, dynamic>?);
-  @override Map<String, dynamic> toJson() => {'ReasoningFile': {'data': data.toJson(), 'media_type': mediaType, if (providerMetadata != null) 'provider_metadata': providerMetadata}};
+  final Json? providerOptions;
+  const ContentPartImage(
+      {required this.image, required this.mediaType, this.providerOptions});
+  @override
+  Json toJson() => {
+        'type': 'image',
+        'image': image,
+        'mediaType': mediaType,
+        if (providerOptions != null) 'providerOptions': providerOptions,
+      };
 }
 
-final class GenerateContentToolApprovalRequest extends GenerateContent {
-  final String approvalId;
-  final String toolCallId;
-  final Map<String, dynamic>? providerMetadata;
-  GenerateContentToolApprovalRequest({required this.approvalId, required this.toolCallId, this.providerMetadata});
-  @override String get tag => 'ToolApprovalRequest';
-  factory GenerateContentToolApprovalRequest.fromJson(Map<String, dynamic> json) => GenerateContentToolApprovalRequest(approvalId: json['approval_id'] as String, toolCallId: json['tool_call_id'] as String, providerMetadata: json['provider_metadata'] as Map<String, dynamic>?);
-  @override Map<String, dynamic> toJson() => {'ToolApprovalRequest': {'approval_id': approvalId, 'tool_call_id': toolCallId, if (providerMetadata != null) 'provider_metadata': providerMetadata}};
+final class ContentPartFile extends ContentPart {
+  final List<int> data;
+  final String mediaType;
+  final String? filename;
+  final Json? providerOptions;
+  const ContentPartFile(
+      {required this.data,
+      required this.mediaType,
+      this.filename,
+      this.providerOptions});
+  @override
+  Json toJson() => {
+        'type': 'file',
+        'data': data,
+        'mediaType': mediaType,
+        if (filename != null) 'filename': filename,
+        if (providerOptions != null) 'providerOptions': providerOptions,
+      };
 }
 
-final class GenerateContentUnknown extends GenerateContent {
+final class ContentPartFileBase64 extends ContentPart {
+  final String data;
+  final String mediaType;
+  final String? filename;
+  final Json? providerOptions;
+  const ContentPartFileBase64(
+      {required this.data,
+      required this.mediaType,
+      this.filename,
+      this.providerOptions});
   @override
-  final String tag;
-  final Map<String, dynamic> data;
-
-  GenerateContentUnknown({required this.tag, required this.data});
+  Json toJson() => {
+        'type': 'file-base64',
+        'data': data,
+        'mediaType': mediaType,
+        if (filename != null) 'filename': filename,
+        if (providerOptions != null) 'providerOptions': providerOptions,
+      };
+}
 
+final class ContentPartFileUrl extends ContentPart {
+  final String url;
+  final String mediaType;
+  final Json? providerOptions;
+  const ContentPartFileUrl(
+      {required this.url, required this.mediaType, this.providerOptions});
   @override
-  Map<String, dynamic> toJson() => {tag: data};
+  Json toJson() => {
+        'type': 'file-url',
+        'url': url,
+        'mediaType': mediaType,
+        if (providerOptions != null) 'providerOptions': providerOptions,
+      };
 }
 
-class GenerateResult {
-  final List<GenerateContent> content;
-  final FinishReason finishReason;
-  final Usage usage;
-  final List<Map<String, dynamic>> warnings;
-  final Map<String, dynamic>? providerMetadata;
-  final GenerateResponse? response;
-  final GenerateRequest? request;
-  GenerateResult({required this.content, required this.finishReason, required this.usage, this.warnings = const [], this.providerMetadata, this.response, this.request});
-  factory GenerateResult.fromJson(Map<String, dynamic> json) => GenerateResult(
-    content: (json['content'] as List<dynamic>? ?? []).map((e) => GenerateContent.fromJson(e as Map<String, dynamic>)).toList(),
-    finishReason: FinishReason.fromJson(json['finish_reason'] as Map<String, dynamic>),
-    usage: Usage.fromJson(json['usage'] as Map<String, dynamic>),
-    warnings: (json['warnings'] as List<dynamic>? ?? []).map((e) => e as Map<String, dynamic>).toList(),
-    providerMetadata: json['provider_metadata'] as Map<String, dynamic>?,
-    response: json['response'] == null ? null : GenerateResponse.fromJson(json['response'] as Map<String, dynamic>),
-    request: json['request'] == null ? null : GenerateRequest.fromJson(json['request'] as Map<String, dynamic>),
-  );
-  Map<String, dynamic> toJson() => {'content': content.map((c) => c.toJson()).toList(), 'finish_reason': finishReason.toJson(), 'usage': usage.toJson(), 'warnings': warnings, 'response': response?.toJson(), if (providerMetadata != null) 'provider_metadata': providerMetadata, if (request != null) 'request': request?.toJson()};
+final class ContentPartFileReference extends ContentPart {
+  final String mediaType;
+  final dynamic reference;
+  final String? filename;
+  final Json? providerOptions;
+  const ContentPartFileReference(
+      {required this.mediaType,
+      required this.reference,
+      this.filename,
+      this.providerOptions});
+  @override
+  Json toJson() => {
+        'type': 'file-reference',
+        'mediaType': mediaType,
+        'reference': reference,
+        if (filename != null) 'filename': filename,
+        if (providerOptions != null) 'providerOptions': providerOptions,
+      };
+}
+
+final class ContentPartReasoning extends ContentPart {
+  final String text;
+  final String? signature;
+  final Json? providerOptions;
+  const ContentPartReasoning(
+      {required this.text, this.signature, this.providerOptions});
+  @override
+  Json toJson() => {
+        'type': 'reasoning',
+        'text': text,
+        if (signature != null) 'signature': signature,
+        if (providerOptions != null) 'providerOptions': providerOptions,
+      };
+}
+
+final class ContentPartToolCall extends ContentPart {
+  final String toolCallId;
+  final String toolName;
+  final dynamic input;
+  final bool? providerExecuted;
+  final Json? providerOptions;
+  const ContentPartToolCall(
+      {required this.toolCallId,
+      required this.toolName,
+      required this.input,
+      this.providerExecuted,
+      this.providerOptions});
+  @override
+  Json toJson() => {
+        'type': 'tool-call',
+        'toolCallId': toolCallId,
+        'toolName': toolName,
+        'input': input,
+        if (providerExecuted != null) 'providerExecuted': providerExecuted,
+        if (providerOptions != null) 'providerOptions': providerOptions,
+      };
+}
+
+/// The call-layer tool result a host appends to the transcript: the call's
+/// [toolCallId] and [toolName] plus a typed [output].
+final class ContentPartToolResult extends ContentPart {
+  final String toolCallId;
+  final String toolName;
+  final ToolResultOutput output;
+  final Json? providerOptions;
+  const ContentPartToolResult(
+      {required this.toolCallId,
+      required this.toolName,
+      required this.output,
+      this.providerOptions});
+  @override
+  Json toJson() => {
+        'type': 'tool-result',
+        'toolCallId': toolCallId,
+        'toolName': toolName,
+        'output': output.toJson(),
+        if (providerOptions != null) 'providerOptions': providerOptions,
+      };
+}
+
+/// Forward-compatibility: a `type` this binding does not model, verbatim.
+final class ContentPartUnknown extends ContentPart {
+  final Json data;
+  const ContentPartUnknown(this.data);
+  String get tag => data['type'] as String;
+  @override
+  Json toJson() => data;
 }
 
 // ─────────────────────────────────────────────────────────────────────────────
-// Result / options
+// Messages
 // ─────────────────────────────────────────────────────────────────────────────
 
-/// Result of `generate_text` (user-facing). Mirrors `GenerateTextResult.ts`.
-@JsonSerializable()
+/// A single chat message. Mirrors `ModelMessage.ts`.
+///
+/// `content` is either a plain `String` or a `List` of content-part maps
+/// (see [ContentPart]); it is kept as `Object` so both shapes pass through
+/// verbatim.
+class ModelMessage {
+  final String role;
+  final Object content;
+
+  const ModelMessage({required this.role, required this.content});
+
+  factory ModelMessage.fromJson(Json j) =>
+      ModelMessage(role: j['role'] as String, content: j['content'] as Object);
+  Json toJson() => {'role': role, 'content': content};
+
+  /// [content] as a list of content-part maps: a `String` becomes a single
+  /// `{'type': 'text', 'text': content}` part.
+  List<Json> get contentParts {
+    final c = content;
+    if (c is String) {
+      return [
+        {'type': 'text', 'text': c},
+      ];
+    }
+    if (c is List) return c.whereType<Json>().toList();
+    return const [];
+  }
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// GenerateResult (provider result) and its content
+// ─────────────────────────────────────────────────────────────────────────────
+
+/// A content item in a [GenerateResult]. Mirrors `GenerateContent.ts` with the
+/// provider-level parameters (`RawToolCall`, `RawToolApprovalRequest`,
+/// `GeneratedFile`): `{"type": "text" | "tool-call" | "source" | "reasoning" |
+/// "file" | "reasoning-file" | "custom" | "tool-approval-request" |
+/// "tool-result", ...}`. Unknown types fall back to [GenerateContentUnknown].
+sealed class GenerateContent {
+  const GenerateContent();
+
+  /// The wire `type` value, e.g. `'text'`, `'tool-call'`.
+  String get tag => toJson()['type'] as String;
+
+  Json toJson();
+
+  factory GenerateContent.fromJson(Json j) => switch (j['type']) {
+        'text' => GenerateContentText(
+            text: j['text'] as String,
+            providerMetadata: _opt(j['providerMetadata'])),
+        'tool-call' => GenerateContentToolCall(RawToolCall.fromJson(j)),
+        'source' => GenerateContentSource(source: Source.fromJson(j)),
+        'reasoning' => GenerateContentReasoning(
+            text: j['text'] as String,
+            providerMetadata: _opt(j['providerMetadata'])),
+        'file' => GenerateContentFile(GeneratedFile.fromJson(j)),
+        'reasoning-file' => GenerateContentReasoningFile(GeneratedFile.fromJson(j)),
+        'custom' => GenerateContentCustom(
+            kind: j['kind'] as String,
+            providerMetadata: _opt(j['providerMetadata'])),
+        'tool-approval-request' => GenerateContentToolApprovalRequest(
+            approvalId: j['approvalId'] as String,
+            toolCallId: j['toolCallId'] as String,
+            providerMetadata: _opt(j['providerMetadata'])),
+        'tool-result' => GenerateContentToolResult(ToolResult.fromJson(j)),
+        _ => GenerateContentUnknown(j),
+      };
+}
+
+final class GenerateContentText extends GenerateContent {
+  final String text;
+  final Json? providerMetadata;
+  const GenerateContentText({required this.text, this.providerMetadata});
+  @override
+  Json toJson() => {
+        'type': 'text',
+        'text': text,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+      };
+}
+
+/// A tool call as the provider emitted it (raw argument text).
+final class GenerateContentToolCall extends GenerateContent {
+  final RawToolCall call;
+  const GenerateContentToolCall(this.call);
+  @override
+  Json toJson() => {'type': 'tool-call', ...call.toJson()};
+}
+
+final class GenerateContentSource extends GenerateContent {
+  final Source source;
+  const GenerateContentSource({required this.source});
+  @override
+  Json toJson() => {'type': 'source', ...source.toJson()};
+}
+
+final class GenerateContentReasoning extends GenerateContent {
+  final String text;
+  final Json? providerMetadata;
+  const GenerateContentReasoning({required this.text, this.providerMetadata});
+  @override
+  Json toJson() => {
+        'type': 'reasoning',
+        'text': text,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+      };
+}
+
+final class GenerateContentFile extends GenerateContent {
+  final GeneratedFile file;
+  const GenerateContentFile(this.file);
+  @override
+  Json toJson() => {'type': 'file', ...file.toJson()};
+}
+
+final class GenerateContentReasoningFile extends GenerateContent {
+  final GeneratedFile file;
+  const GenerateContentReasoningFile(this.file);
+  @override
+  Json toJson() => {'type': 'reasoning-file', ...file.toJson()};
+}
+
+final class GenerateContentCustom extends GenerateContent {
+  final String kind;
+  final Json? providerMetadata;
+  const GenerateContentCustom({required this.kind, this.providerMetadata});
+  @override
+  Json toJson() => {
+        'type': 'custom',
+        'kind': kind,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+      };
+}
+
+final class GenerateContentToolApprovalRequest extends GenerateContent {
+  final String approvalId;
+  final String toolCallId;
+  final Json? providerMetadata;
+  const GenerateContentToolApprovalRequest(
+      {required this.approvalId, required this.toolCallId, this.providerMetadata});
+  @override
+  Json toJson() => {
+        'type': 'tool-approval-request',
+        'approvalId': approvalId,
+        'toolCallId': toolCallId,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+      };
+}
+
+/// The result of a provider-executed tool. Mirrors `ToolResult.ts`: [result] is
+/// the tool's output, [isError] / [preliminary] / [isDynamic] (wire key
+/// `dynamic`) its flags.
+class ToolResult {
+  final String toolCallId;
+  final String toolName;
+  final dynamic result;
+  final bool? isError;
+  final bool? preliminary;
+  final bool? isDynamic;
+  final Json? providerMetadata;
+  const ToolResult({
+    required this.toolCallId,
+    required this.toolName,
+    required this.result,
+    this.isError,
+    this.preliminary,
+    this.isDynamic,
+    this.providerMetadata,
+  });
+  factory ToolResult.fromJson(Json j) => ToolResult(
+        toolCallId: j['toolCallId'] as String,
+        toolName: j['toolName'] as String,
+        result: j['result'],
+        isError: j['isError'] as bool?,
+        preliminary: j['preliminary'] as bool?,
+        isDynamic: j['dynamic'] as bool?,
+        providerMetadata: _opt(j['providerMetadata']),
+      );
+  Json toJson() => {
+        'toolCallId': toolCallId,
+        'toolName': toolName,
+        'result': result,
+        if (isError != null) 'isError': isError,
+        if (preliminary != null) 'preliminary': preliminary,
+        if (isDynamic != null) 'dynamic': isDynamic,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+      };
+}
+
+final class GenerateContentToolResult extends GenerateContent {
+  final ToolResult result;
+  const GenerateContentToolResult(this.result);
+  @override
+  Json toJson() => {'type': 'tool-result', ...result.toJson()};
+}
+
+/// Forward-compatibility: a `type` this binding does not model, verbatim.
+final class GenerateContentUnknown extends GenerateContent {
+  final Json data;
+  const GenerateContentUnknown(this.data);
+  @override
+  Json toJson() => data;
+}
+
+/// The raw provider result. Mirrors `GenerateResult.ts`.
+class GenerateResult {
+  final List<GenerateContent> content;
+  final FinishReason finishReason;
+  final Usage usage;
+  final List<Json> warnings;
+  final Json? providerMetadata;
+  final RequestInfo? request;
+  final ResponseInfo? response;
+  const GenerateResult({
+    required this.content,
+    required this.finishReason,
+    required this.usage,
+    required this.warnings,
+    this.providerMetadata,
+    this.request,
+    this.response,
+  });
+  factory GenerateResult.fromJson(Json j) => GenerateResult(
+        content: _list(j['content'], GenerateContent.fromJson),
+        finishReason: FinishReason.fromJson(j['finishReason'] as Json),
+        usage: Usage.fromJson(j['usage'] as Json),
+        warnings: _maps(j['warnings']),
+        providerMetadata: _opt(j['providerMetadata']),
+        request: j['request'] == null ? null : RequestInfo.fromJson(j['request'] as Json),
+        response:
+            j['response'] == null ? null : ResponseInfo.fromJson(j['response'] as Json),
+      );
+  Json toJson() => {
+        'content': [for (final c in content) c.toJson()],
+        'finishReason': finishReason.toJson(),
+        'usage': usage.toJson(),
+        'warnings': warnings,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+        if (request != null) 'request': request!.toJson(),
+        if (response != null) 'response': response!.toJson(),
+      };
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Results / options
+// ─────────────────────────────────────────────────────────────────────────────
+
+/// Result of `generate_text`. Mirrors `GenerateTextResult.ts`. The typed-union
+/// fields (`content`, `reasoning`, `files`, `warnings`) stay weak-typed maps.
 class GenerateTextResult {
   final String text;
-  final List<Map<String, dynamic>> content;
-  @JsonKey(name: 'tool_calls')
+  final List<Json> content;
   final List<ToolCall> toolCalls;
-  @JsonKey(name: 'finish_reason')
   final FinishReason finishReason;
   final Usage usage;
-  final List<Map<String, dynamic>> warnings;
+  final List<Json> warnings;
   final GenerateResult raw;
-  // M7: top-level aggregation fields
-  final List<Map<String, dynamic>> reasoning;
-  @JsonKey(name: 'reasoning_text')
+  final List<Json> reasoning;
   final String reasoningText;
   final List<Source> sources;
-  final List<Map<String, dynamic>> files;
-  @JsonKey(name: 'response_messages')
+  final List<Json> files;
+
+  /// Assistant messages ready to append to the prompt for the next turn.
   final List<ModelMessage> responseMessages;
-  /// Raw provider-specific finish reason string (M12, e.g. "stop", "end_turn").
-  @JsonKey(name: 'raw_finish_reason')
   final String? rawFinishReason;
-  /// Provider-specific metadata (e.g. Anthropic cache info). Mirrored from
-  /// raw.provider_metadata for top-level convenience. Weak type.
-  @JsonKey(name: 'provider_metadata')
-  final Map<String, dynamic>? providerMetadata;
-  /// Response metadata (id, timestamp, model_id). Mirrored from raw.response.
-  final ResponseMetadata response;
-  /// Total token usage across all steps. In single-step mode (aimux's
-  /// default), equals usage. Provided for AI SDK parity.
-  @JsonKey(name: 'total_usage')
+  final Json? providerMetadata;
+  final RequestInfo request;
+  final ResponseInfo response;
+
+  /// Total token usage across all steps (equals [usage] in single-step mode).
   final Usage totalUsage;
 
-  GenerateTextResult({
+  const GenerateTextResult({
     required this.text,
     this.content = const [],
     required this.toolCalls,
@@ -856,40 +1299,68 @@ class GenerateTextResult {
     this.responseMessages = const [],
     this.rawFinishReason,
     this.providerMetadata,
+    required this.request,
     required this.response,
     required this.totalUsage,
   });
 
-  factory GenerateTextResult.fromJson(Map<String, dynamic> json) =>
-      _$GenerateTextResultFromJson(json);
-  Map<String, dynamic> toJson() => _$GenerateTextResultToJson(this);
+  factory GenerateTextResult.fromJson(Json j) => GenerateTextResult(
+        text: j['text'] as String,
+        content: _maps(j['content']),
+        toolCalls: _list(j['toolCalls'], ToolCall.fromJson),
+        finishReason: FinishReason.fromJson(j['finishReason'] as Json),
+        usage: Usage.fromJson(j['usage'] as Json),
+        warnings: _maps(j['warnings']),
+        raw: GenerateResult.fromJson(j['raw'] as Json),
+        reasoning: _maps(j['reasoning']),
+        reasoningText: j['reasoningText'] as String,
+        sources: _list(j['sources'], Source.fromJson),
+        files: _maps(j['files']),
+        responseMessages: _list(j['responseMessages'], ModelMessage.fromJson),
+        rawFinishReason: j['rawFinishReason'] as String?,
+        providerMetadata: _opt(j['providerMetadata']),
+        request: RequestInfo.fromJson(j['request'] as Json),
+        response: ResponseInfo.fromJson(j['response'] as Json),
+        totalUsage: Usage.fromJson(j['totalUsage'] as Json),
+      );
+
+  Json toJson() => {
+        'content': content,
+        'text': text,
+        'toolCalls': [for (final c in toolCalls) c.toJson()],
+        'finishReason': finishReason.toJson(),
+        'usage': usage.toJson(),
+        'warnings': warnings,
+        'raw': raw.toJson(),
+        'reasoning': reasoning,
+        'reasoningText': reasoningText,
+        'sources': [for (final s in sources) s.toJson()],
+        'files': files,
+        'responseMessages': [for (final m in responseMessages) m.toJson()],
+        if (rawFinishReason != null) 'rawFinishReason': rawFinishReason,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+        'request': request.toJson(),
+        'response': response.toJson(),
+        'totalUsage': totalUsage.toJson(),
+      };
 }
 
-/// Result of `generate_object` (user-facing, M12). The parsed JSON object plus
-/// convenience fields from the underlying `generate_text` call.
-///
-/// Mirrors `GenerateObjectResult.ts`. `object` is an arbitrary JSON value
-/// (weak type — `Map<String, dynamic>` / primitive).
-@JsonSerializable()
+/// Result of `generate_object`. Mirrors `GenerateObjectResult.ts`; [object] is
+/// an arbitrary JSON value.
 class GenerateObjectResult {
-  /// The parsed JSON object returned by the model (arbitrary JSON, weak type).
   final Object? object;
-  @JsonKey(name: 'finish_reason')
   final FinishReason finishReason;
-  @JsonKey(name: 'raw_finish_reason')
   final String? rawFinishReason;
   final Usage usage;
-  final List<Map<String, dynamic>> warnings;
-  /// Concatenated reasoning text (if the model produced reasoning/thinking).
+  final List<Json> warnings;
+
+  /// Concatenated reasoning text, if the model produced any.
   final String? reasoning;
-  /// Provider-specific metadata (e.g. Anthropic cache info). Weak type.
-  @JsonKey(name: 'provider_metadata')
-  final Map<String, dynamic>? providerMetadata;
-  /// Response metadata (id, timestamp, model_id).
+  final Json? providerMetadata;
   final ResponseMetadata response;
   final GenerateTextResult raw;
 
-  GenerateObjectResult({
+  const GenerateObjectResult({
     this.object,
     required this.finishReason,
     this.rawFinishReason,
@@ -901,49 +1372,55 @@ class GenerateObjectResult {
     required this.raw,
   });
 
-  factory GenerateObjectResult.fromJson(Map<String, dynamic> json) =>
-      _$GenerateObjectResultFromJson(json);
-  Map<String, dynamic> toJson() => _$GenerateObjectResultToJson(this);
+  factory GenerateObjectResult.fromJson(Json j) => GenerateObjectResult(
+        object: j['object'],
+        finishReason: FinishReason.fromJson(j['finishReason'] as Json),
+        rawFinishReason: j['rawFinishReason'] as String?,
+        usage: Usage.fromJson(j['usage'] as Json),
+        warnings: _maps(j['warnings']),
+        reasoning: j['reasoning'] as String?,
+        providerMetadata: _opt(j['providerMetadata']),
+        response: ResponseMetadata.fromJson(j['response'] as Json),
+        raw: GenerateTextResult.fromJson(j['raw'] as Json),
+      );
+
+  Json toJson() => {
+        'object': object,
+        'finishReason': finishReason.toJson(),
+        if (rawFinishReason != null) 'rawFinishReason': rawFinishReason,
+        'usage': usage.toJson(),
+        'warnings': warnings,
+        if (reasoning != null) 'reasoning': reasoning,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+        'response': response.toJson(),
+        'raw': raw.toJson(),
+      };
 }
 
-/// Aggregated result of `stream_text().consume()` (M11). Mirrors
-/// `GenerateTextResult`'s user-facing fields (without `raw`, since streaming
-/// has no `GenerateResult` equivalent).
-///
-/// Mirrors `StreamTextResultAggregated.ts`. reasoning/files use weak
-/// types (`Map<String, dynamic>`) — same strategy as `GenerateTextResult`.
-@JsonSerializable()
+/// Aggregated result of `consume_stream_text`. Mirrors
+/// `StreamTextResultAggregated.ts` (the fields of [GenerateTextResult] except
+/// `raw`).
 class StreamTextResultAggregated {
   final String text;
-  final List<Map<String, dynamic>> content;
-  // reasoning/files use weak types — same strategy as GenerateTextResult.
-  final List<Map<String, dynamic>> reasoning;
-  @JsonKey(name: 'reasoning_text')
+  final List<Json> content;
+  final List<Json> reasoning;
   final String reasoningText;
-  @JsonKey(name: 'tool_calls')
   final List<ToolCall> toolCalls;
   final List<Source> sources;
-  final List<Map<String, dynamic>> files;
-  @JsonKey(name: 'finish_reason')
+  final List<Json> files;
   final FinishReason finishReason;
-  @JsonKey(name: 'raw_finish_reason')
   final String? rawFinishReason;
   final Usage usage;
-  /// Total token usage across all steps. In single-step mode (aimux's
-  /// default), equals usage. Provided for AI SDK parity.
-  @JsonKey(name: 'total_usage')
   final Usage totalUsage;
-  final List<Map<String, dynamic>> warnings;
-  /// Provider-specific metadata from the Finish chunk. Weak type.
-  @JsonKey(name: 'provider_metadata')
-  final Map<String, dynamic>? providerMetadata;
-  /// Response metadata (id, timestamp, model_id) if emitted by the stream.
-  final ResponseMetadata? response;
-  @JsonKey(name: 'response_messages')
+  final List<Json> warnings;
+  final Json? providerMetadata;
+  final RequestInfo request;
+  final ResponseInfo response;
   final List<ModelMessage> responseMessages;
 
-  StreamTextResultAggregated({
-    this.text = '',
+  const StreamTextResultAggregated({
+    required this.text,
+    this.content = const [],
     this.reasoning = const [],
     this.reasoningText = '',
     this.toolCalls = const [],
@@ -955,80 +1432,97 @@ class StreamTextResultAggregated {
     required this.totalUsage,
     this.warnings = const [],
     this.providerMetadata,
-    this.response,
+    required this.request,
+    required this.response,
     this.responseMessages = const [],
   });
 
-  factory StreamTextResultAggregated.fromJson(Map<String, dynamic> json) =>
-      _$StreamTextResultAggregatedFromJson(json);
-  Map<String, dynamic> toJson() => _$StreamTextResultAggregatedToJson(this);
+  factory StreamTextResultAggregated.fromJson(Json j) => StreamTextResultAggregated(
+        text: j['text'] as String,
+        content: _maps(j['content']),
+        reasoning: _maps(j['reasoning']),
+        reasoningText: j['reasoningText'] as String,
+        toolCalls: _list(j['toolCalls'], ToolCall.fromJson),
+        sources: _list(j['sources'], Source.fromJson),
+        files: _maps(j['files']),
+        finishReason: FinishReason.fromJson(j['finishReason'] as Json),
+        rawFinishReason: j['rawFinishReason'] as String?,
+        usage: Usage.fromJson(j['usage'] as Json),
+        totalUsage: Usage.fromJson(j['totalUsage'] as Json),
+        warnings: _maps(j['warnings']),
+        providerMetadata: _opt(j['providerMetadata']),
+        request: RequestInfo.fromJson(j['request'] as Json),
+        response: ResponseInfo.fromJson(j['response'] as Json),
+        responseMessages: _list(j['responseMessages'], ModelMessage.fromJson),
+      );
+
+  Json toJson() => {
+        'content': content,
+        'text': text,
+        'reasoning': reasoning,
+        'reasoningText': reasoningText,
+        'toolCalls': [for (final c in toolCalls) c.toJson()],
+        'sources': [for (final s in sources) s.toJson()],
+        'files': files,
+        'finishReason': finishReason.toJson(),
+        if (rawFinishReason != null) 'rawFinishReason': rawFinishReason,
+        'usage': usage.toJson(),
+        'totalUsage': totalUsage.toJson(),
+        'warnings': warnings,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+        'request': request.toJson(),
+        'response': response.toJson(),
+        'responseMessages': [for (final m in responseMessages) m.toJson()],
+      };
 }
 
-/// Per-call timeout configuration. Mirrors `TimeoutConfiguration.ts`.
-///
-/// All values are milliseconds; `null` disables the corresponding limit.
-@JsonSerializable()
+/// Per-call timeout configuration, in milliseconds; null disables a limit.
+/// Mirrors `TimeoutConfiguration.ts`.
 class TimeoutConfiguration {
-  @JsonKey(name: 'total_ms')
   final int? totalMs;
-  @JsonKey(name: 'step_ms')
   final int? stepMs;
-  @JsonKey(name: 'first_chunk_ms')
   final int? firstChunkMs;
-  @JsonKey(name: 'chunk_ms')
   final int? chunkMs;
 
-  TimeoutConfiguration({this.totalMs, this.stepMs, this.firstChunkMs, this.chunkMs});
+  const TimeoutConfiguration({this.totalMs, this.stepMs, this.firstChunkMs, this.chunkMs});
 
-  factory TimeoutConfiguration.fromJson(Map<String, dynamic> json) =>
-      TimeoutConfiguration(
-        totalMs: json['total_ms'] as int?,
-        stepMs: json['step_ms'] as int?,
-        firstChunkMs: json['first_chunk_ms'] as int?,
-        chunkMs: json['chunk_ms'] as int?,
+  factory TimeoutConfiguration.fromJson(Json j) => TimeoutConfiguration(
+        totalMs: j['totalMs'] as int?,
+        stepMs: j['stepMs'] as int?,
+        firstChunkMs: j['firstChunkMs'] as int?,
+        chunkMs: j['chunkMs'] as int?,
       );
-  Map<String, dynamic> toJson() => {
-        if (totalMs != null) 'total_ms': totalMs,
-        if (stepMs != null) 'step_ms': stepMs,
-        if (firstChunkMs != null) 'first_chunk_ms': firstChunkMs,
-        if (chunkMs != null) 'chunk_ms': chunkMs,
+  Json toJson() => {
+        if (totalMs != null) 'totalMs': totalMs,
+        if (stepMs != null) 'stepMs': stepMs,
+        if (firstChunkMs != null) 'firstChunkMs': firstChunkMs,
+        if (chunkMs != null) 'chunkMs': chunkMs,
       };
 }
 
 /// User-facing options for `generate_text` / `stream_text`. Mirrors
-/// `GenerateTextOptions.ts`.
-@JsonSerializable()
+/// `GenerateTextOptions.ts`; unset fields are omitted from the JSON.
 class GenerateTextOptions {
-  @JsonKey(name: 'max_output_tokens')
   final int? maxOutputTokens;
   final double? temperature;
-  @JsonKey(name: 'stop_sequences')
   final List<String>? stopSequences;
-  @JsonKey(name: 'top_p')
   final double? topP;
-  @JsonKey(name: 'top_k')
   final double? topK;
-  @JsonKey(name: 'presence_penalty')
   final double? presencePenalty;
-  @JsonKey(name: 'frequency_penalty')
   final double? frequencyPenalty;
-  @JsonKey(name: 'response_format')
-  final Map<String, dynamic>? responseFormat;
+
+  /// `{"type": "text"}` or `{"type": "json", "schema"?, "name"?, "description"?}`.
+  final Json? responseFormat;
   final int? seed;
   final List<Tool>? tools;
-  @JsonKey(name: 'tool_choice')
   final ToolChoice? toolChoice;
   final Map<String, String>? headers;
-  @JsonKey(name: 'provider_options')
-  final Map<String, dynamic>? providerOptions;
+  final Json? providerOptions;
   final ReasoningEffort? reasoning;
   final String? instructions;
-  @JsonKey(name: 'max_retries')
   final int? maxRetries;
   final TimeoutConfiguration? timeout;
-  @JsonKey(name: 'include_raw_chunks')
   final bool? includeRawChunks;
-  @JsonKey(name: 'session_id')
   final String? sessionId;
 
   /// Repair a tool call the model got wrong (RFC-0035), mirroring the AI SDK's
@@ -1046,10 +1540,9 @@ class GenerateTextOptions {
   /// argument deltas verbatim, as in the AI SDK.
   ///
   /// Never serialized — it is a host closure, not part of the wire options.
-  @JsonKey(includeFromJson: false, includeToJson: false)
   final RepairToolCall? repairToolCall;
 
-  GenerateTextOptions({
+  const GenerateTextOptions({
     this.maxOutputTokens,
     this.temperature,
     this.stopSequences,
@@ -1072,182 +1565,125 @@ class GenerateTextOptions {
     this.repairToolCall,
   });
 
-  factory GenerateTextOptions.fromJson(Map<String, dynamic> json) {
-    return GenerateTextOptions(
-      maxOutputTokens: json['max_output_tokens'] as int?,
-      temperature: (json['temperature'] as num?)?.toDouble(),
-      stopSequences: (json['stop_sequences'] as List<dynamic>?)?.cast<String>(),
-      topP: (json['top_p'] as num?)?.toDouble(),
-      topK: (json['top_k'] as num?)?.toDouble(),
-      presencePenalty: (json['presence_penalty'] as num?)?.toDouble(),
-      frequencyPenalty: (json['frequency_penalty'] as num?)?.toDouble(),
-      responseFormat: json['response_format'] as Map<String, dynamic>?,
-      seed: json['seed'] as int?,
-      tools: (json['tools'] as List<dynamic>?)
-          ?.map((e) => Tool.fromJson(e as Map<String, dynamic>))
-          .toList(),
-      toolChoice: json['tool_choice'] != null
-          ? ToolChoice.fromJson(json['tool_choice'])
-          : null,
-      headers: (json['headers'] as Map<String, dynamic>?)?.cast<String, String>(),
-      providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      reasoning: json['reasoning'] != null
-          ? ReasoningEffort.fromJson(json['reasoning'] as String)
-          : null,
-      instructions: json['instructions'] as String?,
-      maxRetries: json['max_retries'] as int?,
-      timeout: json['timeout'] != null
-          ? TimeoutConfiguration.fromJson(json['timeout'] as Map<String, dynamic>)
-          : null,
-      includeRawChunks: json['include_raw_chunks'] as bool?,
-      sessionId: json['session_id'] as String?,
-    );
-  }
-  Map<String, dynamic> toJson() => {
-        if (maxOutputTokens != null) 'max_output_tokens': maxOutputTokens,
+  factory GenerateTextOptions.fromJson(Json j) => GenerateTextOptions(
+        maxOutputTokens: j['maxOutputTokens'] as int?,
+        temperature: (j['temperature'] as num?)?.toDouble(),
+        stopSequences: (j['stopSequences'] as List?)?.cast<String>(),
+        topP: (j['topP'] as num?)?.toDouble(),
+        topK: (j['topK'] as num?)?.toDouble(),
+        presencePenalty: (j['presencePenalty'] as num?)?.toDouble(),
+        frequencyPenalty: (j['frequencyPenalty'] as num?)?.toDouble(),
+        responseFormat: _opt(j['responseFormat']),
+        seed: j['seed'] as int?,
+        tools: _optList(j['tools'], Tool.fromJson),
+        toolChoice:
+            j['toolChoice'] == null ? null : ToolChoice.fromJson(j['toolChoice']),
+        headers: (j['headers'] as Json?)?.cast<String, String>(),
+        providerOptions: _opt(j['providerOptions']),
+        reasoning: j['reasoning'] == null
+            ? null
+            : ReasoningEffort.fromJson(j['reasoning'] as String),
+        instructions: j['instructions'] as String?,
+        maxRetries: j['maxRetries'] as int?,
+        timeout: j['timeout'] == null
+            ? null
+            : TimeoutConfiguration.fromJson(j['timeout'] as Json),
+        includeRawChunks: j['includeRawChunks'] as bool?,
+        sessionId: j['sessionId'] as String?,
+      );
+
+  Json toJson() => {
+        if (maxOutputTokens != null) 'maxOutputTokens': maxOutputTokens,
         if (temperature != null) 'temperature': temperature,
-        if (stopSequences != null) 'stop_sequences': stopSequences,
-        if (topP != null) 'top_p': topP,
-        if (topK != null) 'top_k': topK,
-        if (presencePenalty != null) 'presence_penalty': presencePenalty,
-        if (frequencyPenalty != null) 'frequency_penalty': frequencyPenalty,
-        if (responseFormat != null) 'response_format': responseFormat,
+        if (stopSequences != null) 'stopSequences': stopSequences,
+        if (topP != null) 'topP': topP,
+        if (topK != null) 'topK': topK,
+        if (presencePenalty != null) 'presencePenalty': presencePenalty,
+        if (frequencyPenalty != null) 'frequencyPenalty': frequencyPenalty,
+        if (responseFormat != null) 'responseFormat': responseFormat,
         if (seed != null) 'seed': seed,
-        if (tools != null) 'tools': tools!.map((t) => t.toJson()).toList(),
-        if (toolChoice != null) 'tool_choice': toolChoice!.toJson(),
+        if (tools != null) 'tools': [for (final t in tools!) t.toJson()],
+        if (toolChoice != null) 'toolChoice': toolChoice!.toJson(),
         if (headers != null) 'headers': headers,
-        if (providerOptions != null) 'provider_options': providerOptions,
+        if (providerOptions != null) 'providerOptions': providerOptions,
         if (reasoning != null) 'reasoning': reasoning!.toJson(),
         if (instructions != null) 'instructions': instructions,
-        if (maxRetries != null) 'max_retries': maxRetries,
+        if (maxRetries != null) 'maxRetries': maxRetries,
         if (timeout != null) 'timeout': timeout!.toJson(),
-        if (includeRawChunks != null) 'include_raw_chunks': includeRawChunks,
-        if (sessionId != null) 'session_id': sessionId,
+        if (includeRawChunks != null) 'includeRawChunks': includeRawChunks,
+        if (sessionId != null) 'sessionId': sessionId,
       };
 }
 
-// ─────────────────────────────────────────────────────────────────────────────
-// Messages
-// ─────────────────────────────────────────────────────────────────────────────
-
-/// A single user-facing chat message. Mirrors `ModelMessage.ts`.
-///
-/// `content` is either a plain `String` or a `List` of content-part maps
-/// (`[{"type":"text","text":"..."}, ...]`); it is kept as `Object` so both
-/// shapes pass through verbatim.
-@JsonSerializable()
-class ModelMessage {
-  final String role;
-  final Object content;
-
-  ModelMessage({required this.role, required this.content});
-
-  factory ModelMessage.fromJson(Map<String, dynamic> json) =>
-      _$ModelMessageFromJson(json);
-  Map<String, dynamic> toJson() => _$ModelMessageToJson(this);
-
-  /// Convenience view of [content] as a list of content-part maps.
-  ///
-  /// If [content] is a `String`, returns `[{'type': 'text', 'text': content}]`;
-  /// if it is a `List`, returns it cast to `List<Map<String, dynamic>>`;
-  /// otherwise an empty list. `content` itself stays `Object` so both the
-  /// string and the part-list wire shapes pass through verbatim.
-  List<Map<String, dynamic>> get contentParts {
-    final c = content;
-    if (c is String) {
-      return <Map<String, dynamic>>[
-        {'type': 'text', 'text': c},
-      ];
-    }
-    if (c is List) {
-      return c.whereType<Map<String, dynamic>>().toList();
-    }
-    return const <Map<String, dynamic>>[];
-  }
-}
-
 // ─────────────────────────────────────────────────────────────────────────────
 // Streaming
 // ─────────────────────────────────────────────────────────────────────────────
 
-/// A single chunk in the stream returned by `stream_text`. Mirrors
-/// `StreamPart.ts` (`aimux-core/src/stream_part.rs`).
-///
-/// `StreamPart` is an externally-tagged union: each part is a single-key map
-/// like `{"TextDelta": {"id": "...", "delta": "..."}}`. Modeled as a sealed
-/// class hierarchy so callers get compile-time types: every known variant is a
-/// typed subclass with named fields, and [StreamPartUnknown] is the fallback
-/// for tags added by newer core versions.
-///
-/// Known variants: [StreamPartTextStart]/[StreamPartTextDelta]/[StreamPartTextEnd],
-/// [StreamPartStreamStart], [StreamPartFinish], [StreamPartError],
-/// [StreamPartToolInputStart]/[StreamPartToolInputDelta]/[StreamPartToolInputEnd],
-/// [StreamPartToolCall], [StreamPartToolResult], [StreamPartFile],
-/// [StreamPartReasoningStart]/[StreamPartReasoningDelta]/[StreamPartReasoningEnd],
-/// [StreamPartResponseMetadata], [StreamPartSource], [StreamPartRaw]. Narrow
-/// via `switch`/`is`/`whereType` to read variant-specific fields.
+/// A single part of the stream returned by `stream_text`. Mirrors
+/// `TextStreamPart.ts`: `{"type": "text-delta", "id": ..., "delta": ...}` —
+/// the tag values are `text-start`, `text-delta`, `text-end`, `stream-start`,
+/// `finish`, `finish-step`, `error`, `tool-input-start`, `tool-input-delta`,
+/// `tool-input-end`, `tool-call`, `tool-result`, `file`, `reasoning-file`,
+/// `custom`, `tool-approval-request`, `reasoning-start`, `reasoning-delta`,
+/// `reasoning-end`, `source` and `raw`. A `type` this binding does not model
+/// becomes [StreamPartUnknown] and passes through verbatim. Narrow via
+/// `switch` / `is` / `whereType`.
 sealed class StreamPart {
   const StreamPart();
 
-  /// The variant tag — the single top-level key on the wire (e.g.
-  /// `'TextDelta'`, `'ToolCall'`, `'ToolResult'`, `'File'`, `'Finish'`,
-  /// `'Error'`). Empty only for a malformed part.
-  String get type => switch (this) {
-        StreamPartTextStart() => 'TextStart',
-        StreamPartTextDelta() => 'TextDelta',
-        StreamPartTextEnd() => 'TextEnd',
-        StreamPartStreamStart() => 'StreamStart',
-        StreamPartFinish() => 'Finish',
-        StreamPartError() => 'Error',
-        StreamPartToolInputStart() => 'ToolInputStart',
-        StreamPartToolInputDelta() => 'ToolInputDelta',
-        StreamPartToolInputEnd() => 'ToolInputEnd',
-        StreamPartToolCall() => 'ToolCall',
-        StreamPartToolResult() => 'ToolResult',
-        StreamPartCustom() => 'Custom',
-        StreamPartReasoningFile() => 'ReasoningFile',
-        StreamPartToolApprovalRequest() => 'ToolApprovalRequest',
-        StreamPartFile() => 'File',
-        StreamPartReasoningStart() => 'ReasoningStart',
-        StreamPartReasoningDelta() => 'ReasoningDelta',
-        StreamPartReasoningEnd() => 'ReasoningEnd',
-        StreamPartResponseMetadata() => 'ResponseMetadata',
-        StreamPartSource() => 'Source',
-        StreamPartRaw() => 'Raw',
-        StreamPartUnknown(:final tag) => tag,
-      };
-
-  /// Re-encode to the externally-tagged wire shape (`{tag: payload}`).
-  Map<String, dynamic> toJson();
-
-  /// Decode an externally-tagged stream-part map. Unknown tags fall back to
-  /// [StreamPartUnknown] instead of throwing.
-  factory StreamPart.fromJson(Map<String, dynamic> json) {
-    final e = json.entries.first;
-    final payload = e.value as Map<String, dynamic>;
-    return switch (e.key) {
-      'TextStart' => StreamPartTextStart.fromJson(payload),
-      'TextDelta' => StreamPartTextDelta.fromJson(payload),
-      'TextEnd' => StreamPartTextEnd.fromJson(payload),
-      'StreamStart' => StreamPartStreamStart.fromJson(payload),
-      'Finish' => StreamPartFinish.fromJson(payload),
-      'Error' => StreamPartError.fromJson(payload),
-      'ToolInputStart' => StreamPartToolInputStart.fromJson(payload),
-      'ToolInputDelta' => StreamPartToolInputDelta.fromJson(payload),
-      'ToolInputEnd' => StreamPartToolInputEnd.fromJson(payload),
-      'ToolCall' => StreamPartToolCall.fromJson(payload),
-      'ToolResult' => StreamPartToolResult.fromJson(payload),
-      'Custom' => StreamPartCustom.fromJson(payload),
-      'ReasoningFile' => StreamPartReasoningFile.fromJson(payload),
-      'ToolApprovalRequest' => StreamPartToolApprovalRequest.fromJson(payload),
-      'File' => StreamPartFile.fromJson(payload),
-      'ReasoningStart' => StreamPartReasoningStart.fromJson(payload),
-      'ReasoningDelta' => StreamPartReasoningDelta.fromJson(payload),
-      'ReasoningEnd' => StreamPartReasoningEnd.fromJson(payload),
-      'ResponseMetadata' => StreamPartResponseMetadata.fromJson(payload),
-      'Source' => StreamPartSource.fromJson(payload),
-      'Raw' => StreamPartRaw.fromJson(payload),
-      _ => StreamPartUnknown(tag: e.key, data: payload),
+  /// The wire `type` value, e.g. `'text-delta'`.
+  String get type => toJson()['type'] as String;
+
+  Json toJson();
+
+  factory StreamPart.fromJson(Json j) {
+    String id() => j['id'] as String;
+    Json? meta() => _opt(j['providerMetadata']);
+    return switch (j['type']) {
+      'text-start' => StreamPartTextStart(id: id(), providerMetadata: meta()),
+      'text-delta' => StreamPartTextDelta(
+          id: id(), delta: j['delta'] as String, providerMetadata: meta()),
+      'text-end' => StreamPartTextEnd(id: id(), providerMetadata: meta()),
+      'stream-start' => StreamPartStreamStart(warnings: _maps(j['warnings'])),
+      'finish' => StreamPartFinish(
+          finishReason: FinishReason.fromJson(j['finishReason'] as Json),
+          usage: Usage.fromJson(j['usage'] as Json),
+          providerMetadata: meta()),
+      'finish-step' => StreamPartFinishStep(
+          finishReason: FinishReason.fromJson(j['finishReason'] as Json),
+          usage: Usage.fromJson(j['usage'] as Json),
+          providerMetadata: meta(),
+          response: ResponseInfo.fromJson(j['response'] as Json)),
+      'error' => StreamPartError(error: j['error'] as Json),
+      'tool-input-start' => StreamPartToolInputStart(
+          id: id(),
+          toolName: j['toolName'] as String,
+          providerExecuted: j['providerExecuted'] as bool?,
+          isDynamic: j['dynamic'] as bool?,
+          title: j['title'] as String?,
+          providerMetadata: meta()),
+      'tool-input-delta' => StreamPartToolInputDelta(
+          id: id(), delta: j['delta'] as String, providerMetadata: meta()),
+      'tool-input-end' => StreamPartToolInputEnd(id: id(), providerMetadata: meta()),
+      'tool-call' => StreamPartToolCall(ToolCall.fromJson(j)),
+      'tool-result' => StreamPartToolResult(ToolResult.fromJson(j)),
+      'file' => StreamPartFile(GeneratedFile.fromJson(j)),
+      'reasoning-file' => StreamPartReasoningFile(
+          file: GeneratedFile.fromJson(j['file'] as Json), providerMetadata: meta()),
+      'custom' => StreamPartCustom(kind: j['kind'] as String, providerMetadata: meta()),
+      'tool-approval-request' => StreamPartToolApprovalRequest(
+          approvalId: j['approvalId'] as String,
+          toolCall: ToolCall.fromJson(j['toolCall'] as Json),
+          reason: j['reason'] as String?,
+          isAutomatic: j['isAutomatic'] as bool?,
+          signature: j['signature'] as String?),
+      'reasoning-start' => StreamPartReasoningStart(id: id(), providerMetadata: meta()),
+      'reasoning-delta' => StreamPartReasoningDelta(
+          id: id(), delta: j['delta'] as String, providerMetadata: meta()),
+      'reasoning-end' => StreamPartReasoningEnd(id: id(), providerMetadata: meta()),
+      'source' => StreamPartSource(source: Source.fromJson(j)),
+      'raw' => StreamPartRaw(rawValue: j['rawValue']),
+      _ => StreamPartUnknown(j),
     };
   }
 
@@ -1255,153 +1691,101 @@ sealed class StreamPart {
   String toString() => 'StreamPart($type)';
 }
 
-// ── Text variants ────────────────────────────────────────────────────────────
-
-/// Start of a text segment.
+/// Start / delta / end of a text segment.
 final class StreamPartTextStart extends StreamPart {
   final String id;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartTextStart({required this.id, this.providerMetadata});
-
-  factory StreamPartTextStart.fromJson(Map<String, dynamic> json) =>
-      StreamPartTextStart(
-        id: json['id'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
+  final Json? providerMetadata;
+  const StreamPartTextStart({required this.id, this.providerMetadata});
   @override
-  Map<String, dynamic> toJson() => {
-        'TextStart': {
-          'id': id,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'text-start',
+        'id': id,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
-/// A delta of generated text.
 final class StreamPartTextDelta extends StreamPart {
   final String id;
   final String delta;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartTextDelta({
-    required this.id,
-    required this.delta,
-    this.providerMetadata,
-  });
-
-  factory StreamPartTextDelta.fromJson(Map<String, dynamic> json) =>
-      StreamPartTextDelta(
-        id: json['id'] as String,
-        delta: json['delta'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
+  final Json? providerMetadata;
+  const StreamPartTextDelta(
+      {required this.id, required this.delta, this.providerMetadata});
   @override
-  Map<String, dynamic> toJson() => {
-        'TextDelta': {
-          'id': id,
-          'delta': delta,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'text-delta',
+        'id': id,
+        'delta': delta,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
-/// End of a text segment.
 final class StreamPartTextEnd extends StreamPart {
   final String id;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartTextEnd({required this.id, this.providerMetadata});
-
-  factory StreamPartTextEnd.fromJson(Map<String, dynamic> json) =>
-      StreamPartTextEnd(
-        id: json['id'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
+  final Json? providerMetadata;
+  const StreamPartTextEnd({required this.id, this.providerMetadata});
   @override
-  Map<String, dynamic> toJson() => {
-        'TextEnd': {
-          'id': id,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'text-end',
+        'id': id,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
-// ── Stream lifecycle ──────────────────────────────────────────────────────────
-
-/// First chunk — carries warnings from the provider.
+/// First part — carries the provider's warnings.
 final class StreamPartStreamStart extends StreamPart {
-  final List<Map<String, dynamic>> warnings;
-
-  StreamPartStreamStart({this.warnings = const <Map<String, dynamic>>[]});
-
-  factory StreamPartStreamStart.fromJson(Map<String, dynamic> json) =>
-      StreamPartStreamStart(
-        warnings: ((json['warnings'] as List<dynamic>?) ?? const <dynamic>[])
-            .map((e) => e as Map<String, dynamic>)
-            .toList(),
-      );
-
+  final List<Json> warnings;
+  const StreamPartStreamStart({this.warnings = const []});
   @override
-  Map<String, dynamic> toJson() => {
-        'StreamStart': {'warnings': warnings},
-      };
+  Json toJson() => {'type': 'stream-start', 'warnings': warnings};
 }
 
-/// Final chunk — carries usage, finish reason, and metadata.
+/// Final part — carries usage, finish reason and metadata.
 final class StreamPartFinish extends StreamPart {
   final FinishReason finishReason;
   final Usage usage;
-  final Map<String, dynamic>? providerMetadata;
+  final Json? providerMetadata;
+  const StreamPartFinish(
+      {required this.finishReason, required this.usage, this.providerMetadata});
+  @override
+  Json toJson() => {
+        'type': 'finish',
+        'finishReason': finishReason.toJson(),
+        'usage': usage.toJson(),
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+      };
+}
 
-  StreamPartFinish({
+/// End of one generation step.
+final class StreamPartFinishStep extends StreamPart {
+  final FinishReason finishReason;
+  final Usage usage;
+  final Json? providerMetadata;
+  final ResponseInfo response;
+  const StreamPartFinishStep({
     required this.finishReason,
     required this.usage,
     this.providerMetadata,
+    required this.response,
   });
-
-  factory StreamPartFinish.fromJson(Map<String, dynamic> json) =>
-      StreamPartFinish(
-        finishReason: FinishReason.fromJson(
-            json['finish_reason'] as Map<String, dynamic>),
-        usage: Usage.fromJson(json['usage'] as Map<String, dynamic>),
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
   @override
-  Map<String, dynamic> toJson() => {
-        'Finish': {
-          'finish_reason': finishReason.toJson(),
-          'usage': usage.toJson(),
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'finish-step',
+        'finishReason': finishReason.toJson(),
+        'usage': usage.toJson(),
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+        'response': response.toJson(),
       };
 }
 
-/// An error occurred mid-stream.
+/// An error mid-stream: an `AiMuxError` object keyed by `name`
+/// (`{"name": "AI_APICallError", ...}`).
 final class StreamPartError extends StreamPart {
-  final dynamic error;
-
-  StreamPartError({required this.error});
-
-  factory StreamPartError.fromJson(Map<String, dynamic> json) =>
-      StreamPartError(error: json['error']);
-
+  final Json error;
+  const StreamPartError({required this.error});
   @override
-  Map<String, dynamic> toJson() => {
-        'Error': {'error': error},
-      };
+  Json toJson() => {'type': 'error', 'error': error};
 }
 
-// ── Tool calls ─────────────────────────────────────────────────────────────────
-
 /// Start of a tool call's input streaming.
 final class StreamPartToolInputStart extends StreamPart {
   final String id;
@@ -1409,9 +1793,8 @@ final class StreamPartToolInputStart extends StreamPart {
   final bool? providerExecuted;
   final bool? isDynamic;
   final String? title;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartToolInputStart({
+  final Json? providerMetadata;
+  const StreamPartToolInputStart({
     required this.id,
     required this.toolName,
     this.providerExecuted,
@@ -1419,772 +1802,183 @@ final class StreamPartToolInputStart extends StreamPart {
     this.title,
     this.providerMetadata,
   });
-
-  factory StreamPartToolInputStart.fromJson(Map<String, dynamic> json) =>
-      StreamPartToolInputStart(
-        id: json['id'] as String,
-        toolName: json['tool_name'] as String,
-        providerExecuted: json['provider_executed'] as bool?,
-        isDynamic: json['dynamic'] as bool?,
-        title: json['title'] as String?,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
   @override
-  Map<String, dynamic> toJson() => {
-        'ToolInputStart': {
-          'id': id,
-          'tool_name': toolName,
-          if (providerExecuted != null) 'provider_executed': providerExecuted,
-          if (isDynamic != null) 'dynamic': isDynamic,
-          if (title != null) 'title': title,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'tool-input-start',
+        'id': id,
+        'toolName': toolName,
+        if (providerExecuted != null) 'providerExecuted': providerExecuted,
+        if (isDynamic != null) 'dynamic': isDynamic,
+        if (title != null) 'title': title,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
-/// A delta of tool call input (partial JSON).
+/// A delta of tool call input (partial JSON text).
 final class StreamPartToolInputDelta extends StreamPart {
   final String id;
   final String delta;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartToolInputDelta({
-    required this.id,
-    required this.delta,
-    this.providerMetadata,
-  });
-
-  factory StreamPartToolInputDelta.fromJson(Map<String, dynamic> json) =>
-      StreamPartToolInputDelta(
-        id: json['id'] as String,
-        delta: json['delta'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
+  final Json? providerMetadata;
+  const StreamPartToolInputDelta(
+      {required this.id, required this.delta, this.providerMetadata});
   @override
-  Map<String, dynamic> toJson() => {
-        'ToolInputDelta': {
-          'id': id,
-          'delta': delta,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'tool-input-delta',
+        'id': id,
+        'delta': delta,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
-/// End of a tool call's input streaming.
 final class StreamPartToolInputEnd extends StreamPart {
   final String id;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartToolInputEnd({required this.id, this.providerMetadata});
-
-  factory StreamPartToolInputEnd.fromJson(Map<String, dynamic> json) =>
-      StreamPartToolInputEnd(
-        id: json['id'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
+  final Json? providerMetadata;
+  const StreamPartToolInputEnd({required this.id, this.providerMetadata});
   @override
-  Map<String, dynamic> toJson() => {
-        'ToolInputEnd': {
-          'id': id,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
+  Json toJson() => {
+        'type': 'tool-input-end',
+        'id': id,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
-/// A complete tool call (alternative to the start/delta/end flow).
+/// A complete, parsed tool call ([ToolCall.invalid] marks one that stayed
+/// invalid after optional repair).
 final class StreamPartToolCall extends StreamPart {
-  final String toolCallId;
-  final String toolName;
-  final dynamic input;
-  final bool? providerExecuted;
-  final bool? isDynamic;
-  final Map<String, dynamic>? providerMetadata;
-  /// Set by Core when the tool call stays invalid after optional repair.
-  final bool? invalid;
-  /// The typed lookup, parse, schema, or repair failure for an invalid call.
-  final dynamic error;
-
-  StreamPartToolCall({
-    required this.toolCallId,
-    required this.toolName,
-    required this.input,
-    this.providerExecuted,
-    this.isDynamic,
-    this.providerMetadata,
-    this.invalid,
-    this.error,
-  });
-
-  factory StreamPartToolCall.fromJson(Map<String, dynamic> json) =>
-      StreamPartToolCall(
-        toolCallId: json['tool_call_id'] as String,
-        toolName: json['tool_name'] as String,
-        input: json['input'],
-        providerExecuted: json['provider_executed'] as bool?,
-        isDynamic: json['dynamic'] as bool?,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-        invalid: json['invalid'] as bool?,
-        error: json['error'],
-      );
-
+  final ToolCall call;
+  const StreamPartToolCall(this.call);
   @override
-  Map<String, dynamic> toJson() => {
-        'ToolCall': {
-          'tool_call_id': toolCallId,
-          'tool_name': toolName,
-          'input': input,
-          if (providerExecuted != null) 'provider_executed': providerExecuted,
-          if (isDynamic != null) 'dynamic': isDynamic,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-          if (invalid != null) 'invalid': invalid,
-          if (error != null) 'error': error,
-        },
-      };
+  Json toJson() => {'type': 'tool-call', ...call.toJson()};
 }
 
-/// A tool result (provider-executed tools).
+/// A provider-executed tool's result.
 final class StreamPartToolResult extends StreamPart {
-  final String toolCallId;
-  final String toolName;
-  final dynamic result;
-  final bool? isError;
-  final bool? preliminary;
-  final bool? isDynamic;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartToolResult({
-    required this.toolCallId,
-    required this.toolName,
-    required this.result,
-    this.isError,
-    this.preliminary,
-    this.isDynamic,
-    this.providerMetadata,
-  });
-
-  factory StreamPartToolResult.fromJson(Map<String, dynamic> json) =>
-      StreamPartToolResult(
-        toolCallId: json['tool_call_id'] as String,
-        toolName: json['tool_name'] as String,
-        result: json['result'],
-        isError: json['is_error'] as bool?,
-        preliminary: json['preliminary'] as bool?,
-        isDynamic: json['dynamic'] as bool?,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
+  final ToolResult result;
+  const StreamPartToolResult(this.result);
   @override
-  Map<String, dynamic> toJson() => {
-        'ToolResult': {
-          'tool_call_id': toolCallId,
-          'tool_name': toolName,
-          'result': result,
-          if (isError != null) 'is_error': isError,
-          if (preliminary != null) 'preliminary': preliminary,
-          if (isDynamic != null) 'dynamic': isDynamic,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
-      };
+  Json toJson() => {'type': 'tool-result', ...result.toJson()};
 }
 
-// ── File ──────────────────────────────────────────────────────────────────────
-
-/// A file generated by the model (e.g. an image or document). The `data`
-/// field is a `FileData` tagged union; there is **no** `filename` field.
+/// A file generated by the model.
 final class StreamPartFile extends StreamPart {
-  final GeneratedFileData data;
-  final String mediaType;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartFile({
-    required this.data,
-    required this.mediaType,
-    this.providerMetadata,
-  });
-
-  factory StreamPartFile.fromJson(Map<String, dynamic> json) => StreamPartFile(
-        data: GeneratedFileData.fromJson(json['data'] as Map<String, dynamic>),
-        mediaType: json['media_type'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
+  final GeneratedFile file;
+  const StreamPartFile(this.file);
   @override
-  Map<String, dynamic> toJson() => {
-        'File': {
-          'data': data.toJson(),
-          'media_type': mediaType,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
-      };
+  Json toJson() => {'type': 'file', ...file.toJson()};
 }
 
-// ── Reasoning ──────────────────────────────────────────────────────────────────
-
-/// Start of a reasoning / thinking segment.
-final class StreamPartReasoningStart extends StreamPart {
-  final String id;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartReasoningStart({required this.id, this.providerMetadata});
-
-  factory StreamPartReasoningStart.fromJson(Map<String, dynamic> json) =>
-      StreamPartReasoningStart(
-        id: json['id'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
-  @override
-  Map<String, dynamic> toJson() => {
-        'ReasoningStart': {
-          'id': id,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
-      };
-}
-
-/// A delta of a reasoning / thinking segment.
-final class StreamPartReasoningDelta extends StreamPart {
-  final String id;
-  final String delta;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartReasoningDelta({
-    required this.id,
-    required this.delta,
-    this.providerMetadata,
-  });
-
-  factory StreamPartReasoningDelta.fromJson(Map<String, dynamic> json) =>
-      StreamPartReasoningDelta(
-        id: json['id'] as String,
-        delta: json['delta'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
-  @override
-  Map<String, dynamic> toJson() => {
-        'ReasoningDelta': {
-          'id': id,
-          'delta': delta,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
-      };
-}
-
-/// End of a reasoning / thinking segment.
-final class StreamPartReasoningEnd extends StreamPart {
-  final String id;
-  final Map<String, dynamic>? providerMetadata;
-
-  StreamPartReasoningEnd({required this.id, this.providerMetadata});
-
-  factory StreamPartReasoningEnd.fromJson(Map<String, dynamic> json) =>
-      StreamPartReasoningEnd(
-        id: json['id'] as String,
-        providerMetadata:
-            json['provider_metadata'] as Map<String, dynamic>?,
-      );
-
-  @override
-  Map<String, dynamic> toJson() => {
-        'ReasoningEnd': {
-          'id': id,
-          if (providerMetadata != null) 'provider_metadata': providerMetadata,
-        },
-      };
-}
-
-// ── Metadata / sources / raw ──────────────────────────────────────────────────
-
-/// Response metadata (id, timestamp, model_id).
-final class StreamPartResponseMetadata extends StreamPart {
-  final String? id;
-  final String? timestamp;
-  final String? modelId;
-
-  StreamPartResponseMetadata({this.id, this.timestamp, this.modelId});
-
-  factory StreamPartResponseMetadata.fromJson(Map<String, dynamic> json) =>
-      StreamPartResponseMetadata(
-        id: json['id'] as String?,
-        timestamp: json['timestamp'] as String?,
-        modelId: json['model_id'] as String?,
-      );
-
-  @override
-  Map<String, dynamic> toJson() => {
-        'ResponseMetadata': {
-          if (id != null) 'id': id,
-          if (timestamp != null) 'timestamp': timestamp,
-          if (modelId != null) 'model_id': modelId,
-        },
-      };
-}
-
-/// A source / citation (e.g. URL citation from search-preview models).
-final class StreamPartSource extends StreamPart {
-  final Source source;
-
-  StreamPartSource({required this.source});
-
-  factory StreamPartSource.fromJson(Map<String, dynamic> json) =>
-      StreamPartSource(source: Source.fromJson(json));
-
-  @override
-  Map<String, dynamic> toJson() => {'Source': source.toJson()};
-}
-
-/// A raw chunk from the provider (for debugging, when `include_raw_chunks`
-/// is set).
-final class StreamPartRaw extends StreamPart {
-  final dynamic rawValue;
-
-  StreamPartRaw({required this.rawValue});
-
-  factory StreamPartRaw.fromJson(Map<String, dynamic> json) =>
-      StreamPartRaw(rawValue: json['raw_value']);
-
+/// A file generated as part of reasoning (`ReasoningFileOutput.ts`).
+final class StreamPartReasoningFile extends StreamPart {
+  final GeneratedFile file;
+  final Json? providerMetadata;
+  const StreamPartReasoningFile({required this.file, this.providerMetadata});
   @override
-  Map<String, dynamic> toJson() => {
-        'Raw': {'raw_value': rawValue},
+  Json toJson() => {
+        'type': 'reasoning-file',
+        'file': file.toJson(),
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
-/// Fallback for unknown/forward-compatible `StreamPart` variants.
-///
-/// Newer core versions may emit tags this binding does not model yet; this
-/// class passes them through verbatim (`{tag: data}`) instead of discarding
-/// or crashing.
 final class StreamPartCustom extends StreamPart {
   final String kind;
-  final Map<String, dynamic>? providerMetadata;
-  StreamPartCustom({required this.kind, this.providerMetadata});
-  factory StreamPartCustom.fromJson(Map<String, dynamic> json) => StreamPartCustom(kind: json['kind'] as String, providerMetadata: json['provider_metadata'] as Map<String, dynamic>?);
-  @override Map<String, dynamic> toJson() => {'Custom': {'kind': kind, if (providerMetadata != null) 'provider_metadata': providerMetadata}};
-}
-
-final class StreamPartReasoningFile extends StreamPart {
-  final GenerateContentFile file;
-  final Map<String, dynamic>? providerMetadata;
-  StreamPartReasoningFile({required this.file, this.providerMetadata});
-  factory StreamPartReasoningFile.fromJson(Map<String, dynamic> json) => StreamPartReasoningFile(file: GenerateContentFile.fromJson(json['file'] as Map<String, dynamic>), providerMetadata: json['provider_metadata'] as Map<String, dynamic>?);
-  @override Map<String, dynamic> toJson() => {'ReasoningFile': {'file': file.toJson(), if (providerMetadata != null) 'provider_metadata': providerMetadata}};
+  final Json? providerMetadata;
+  const StreamPartCustom({required this.kind, this.providerMetadata});
+  @override
+  Json toJson() => {
+        'type': 'custom',
+        'kind': kind,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
+      };
 }
 
+/// A tool call waiting for approval (`ToolApprovalRequestOutput.ts`).
 final class StreamPartToolApprovalRequest extends StreamPart {
   final String approvalId;
   final ToolCall toolCall;
   final String? reason;
   final bool? isAutomatic;
   final String? signature;
-  StreamPartToolApprovalRequest({required this.approvalId, required this.toolCall, this.reason, this.isAutomatic, this.signature});
-  factory StreamPartToolApprovalRequest.fromJson(Map<String, dynamic> json) => StreamPartToolApprovalRequest(approvalId: json['approval_id'] as String, toolCall: ToolCall.fromJson(json['tool_call'] as Map<String, dynamic>), reason: json['reason'] as String?, isAutomatic: json['is_automatic'] as bool?, signature: json['signature'] as String?);
-  @override Map<String, dynamic> toJson() => {'ToolApprovalRequest': {'approval_id': approvalId, 'tool_call': toolCall.toJson(), if (reason != null) 'reason': reason, if (isAutomatic != null) 'is_automatic': isAutomatic, if (signature != null) 'signature': signature}};
-}
-
-final class StreamPartUnknown extends StreamPart {
-  final String tag;
-  final Map<String, dynamic> data;
-
-  StreamPartUnknown({required this.tag, required this.data});
-
-  @override
-  Map<String, dynamic> toJson() => {tag: data};
-}
-
-// ─────────────────────────────────────────────────────────────────────────────
-// FileBytes (shared.rs) — externally-tagged: {"Binary": [...]} / {"Base64": "..."}
-// ─────────────────────────────────────────────────────────────────────────────
-
-sealed class FileBytes {
-  const FileBytes();
-  factory FileBytes.fromJson(Map<String, dynamic> json) {
-    final e = json.entries.first;
-    return switch (e.key) {
-      'Binary' => FileBytesBinary(data: (e.value as List).cast<int>()),
-      'Base64' => FileBytesBase64(data: e.value as String),
-      _ => throw FormatException('unknown FileBytes tag: $e'),
-    };
-  }
-  Map<String, dynamic> toJson();
-}
-
-final class FileBytesBinary extends FileBytes {
-  final List<int> data;
-  const FileBytesBinary({required this.data});
-  @override
-  Map<String, dynamic> toJson() => {'Binary': data};
-}
-
-final class FileBytesBase64 extends FileBytes {
-  final String data;
-  const FileBytesBase64({required this.data});
-  @override
-  Map<String, dynamic> toJson() => {'Base64': data};
-}
-
-// ─────────────────────────────────────────────────────────────────────────────
-// FileData (shared.rs) — externally-tagged: {"Data": {...}} / {"Url": {...}} / ...
-// ─────────────────────────────────────────────────────────────────────────────
-
-sealed class FileData {
-  const FileData();
-  factory FileData.fromJson(Map<String, dynamic> json) {
-    final e = json.entries.first;
-    return switch (e.key) {
-      'Data' => FileDataData(
-          data: FileBytes.fromJson((e.value as Map)['data'] as Map<String, dynamic>)),
-      'Url' => FileDataUrl(url: (e.value as Map)['url'] as String, originalUrl: (e.value as Map)['original_url'] as String?),
-      'Reference' => FileDataReference(
-          reference: (e.value as Map)['reference'] as Map<String, dynamic>),
-      'Text' => FileDataText(text: (e.value as Map)['text'] as String),
-      _ => throw FormatException('unknown FileData tag: $e'),
-    };
-  }
-  Map<String, dynamic> toJson();
-}
-
-final class FileDataData extends FileData {
-  final FileBytes data;
-  const FileDataData({required this.data});
-  @override
-  Map<String, dynamic> toJson() => {'Data': {'data': data.toJson()}};
-}
-
-final class FileDataUrl extends FileData {
-  final String url;
-  final String? originalUrl;
-  const FileDataUrl({required this.url, this.originalUrl});
-  @override
-  Map<String, dynamic> toJson() => {'Url': {'url': url, if (originalUrl != null) 'original_url': originalUrl}};
-}
-
-final class FileDataReference extends FileData {
-  final Map<String, dynamic> reference;
-  const FileDataReference({required this.reference});
-  @override
-  Map<String, dynamic> toJson() => {'Reference': {'reference': reference}};
-}
-
-final class FileDataText extends FileData {
-  final String text;
-  const FileDataText({required this.text});
-  @override
-  Map<String, dynamic> toJson() => {'Text': {'text': text}};
-}
-
-sealed class GeneratedFileData {
-  const GeneratedFileData();
-  factory GeneratedFileData.fromJson(Map<String, dynamic> json) {
-    final e = json.entries.first;
-    return switch (e.key) {
-      'Data' => GeneratedFileDataData(
-          data: FileBytes.fromJson((e.value as Map)['data'] as Map<String, dynamic>)),
-      'Url' => GeneratedFileDataUrl(url: (e.value as Map)['url'] as String, originalUrl: (e.value as Map)['original_url'] as String?),
-      _ => throw FormatException('unknown GeneratedFileData tag: $e'),
-    };
-  }
-  Map<String, dynamic> toJson();
-}
-
-final class GeneratedFileDataData extends GeneratedFileData {
-  final FileBytes data;
-  const GeneratedFileDataData({required this.data});
-  @override
-  Map<String, dynamic> toJson() => {'Data': {'data': data.toJson()}};
-}
-
-final class GeneratedFileDataUrl extends GeneratedFileData {
-  final String url;
-  final String? originalUrl;
-  const GeneratedFileDataUrl({required this.url, this.originalUrl});
-  @override
-  Map<String, dynamic> toJson() => {'Url': {'url': url, if (originalUrl != null) 'original_url': originalUrl}};
-}
-
-// ─────────────────────────────────────────────────────────────────────────────
-// ContentPart (content.rs) — internally-tagged: {"type": "text", "text": "..."}
-// ─────────────────────────────────────────────────────────────────────────────
-
-sealed class ContentPart {
-  const ContentPart();
-  factory ContentPart.fromJson(Map<String, dynamic> json) {
-    final tag = json['type'] as String;
-    return switch (tag) {
-      'custom' => ContentPartCustom.fromJson(json),
-      'reasoning_file' => ContentPartReasoningFile.fromJson(json),
-      'tool_approval_request' => ContentPartToolApprovalRequest.fromJson(json),
-      'text' => ContentPartText.fromJson(json),
-      'image' => ContentPartImage.fromJson(json),
-      'file' => ContentPartFile.fromJson(json),
-      'file_base64' => ContentPartFileBase64.fromJson(json),
-      'file_url' => ContentPartFileUrl.fromJson(json),
-      'file_reference' => ContentPartFileReference.fromJson(json),
-      'reasoning' => ContentPartReasoning.fromJson(json),
-      'tool_call' => ContentPartToolCall.fromJson(json),
-      'tool_result' => ContentPartToolResult.fromJson(json),
-      _ => ContentPartUnknown(tag: tag, data: json),
-    };
-  }
-  Map<String, dynamic> toJson();
-}
-
-final class ContentPartText extends ContentPart {
-  final String text;
-  final Map<String, dynamic>? providerOptions;
-  const ContentPartText({required this.text, this.providerOptions});
-  static ContentPartText fromJson(Map<String, dynamic> json) => ContentPartText(
-        text: json['text'] as String,
-        providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      );
-  @override
-  Map<String, dynamic> toJson() => {
-        'type': 'text',
-        'text': text,
-        if (providerOptions != null) 'provider_options': providerOptions,
-      };
-}
-
-final class ContentPartImage extends ContentPart {
-  final List<int> image;
-  final String mediaType;
-  final Map<String, dynamic>? providerOptions;
-  const ContentPartImage(
-      {required this.image, required this.mediaType, this.providerOptions});
-  static ContentPartImage fromJson(Map<String, dynamic> json) =>
-      ContentPartImage(
-        image: (json['image'] as List).cast<int>(),
-        mediaType: json['media_type'] as String,
-        providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      );
-  @override
-  Map<String, dynamic> toJson() => {
-        'type': 'image',
-        'image': image,
-        'media_type': mediaType,
-        if (providerOptions != null) 'provider_options': providerOptions,
-      };
-}
-
-final class ContentPartFile extends ContentPart {
-  final List<int> data;
-  final String mediaType;
-  final String? filename;
-  final Map<String, dynamic>? providerOptions;
-  const ContentPartFile(
-      {required this.data, required this.mediaType, this.filename, this.providerOptions});
-  static ContentPartFile fromJson(Map<String, dynamic> json) => ContentPartFile(
-        data: (json['data'] as List).cast<int>(),
-        mediaType: json['media_type'] as String,
-        filename: json['filename'] as String?,
-        providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      );
-  @override
-  Map<String, dynamic> toJson() => {
-        'type': 'file',
-        'data': data,
-        'media_type': mediaType,
-        if (filename != null) 'filename': filename,
-        if (providerOptions != null) 'provider_options': providerOptions,
-      };
-}
-
-final class ContentPartFileBase64 extends ContentPart {
-  final String data;
-  final String mediaType;
-  final String? filename;
-  final Map<String, dynamic>? providerOptions;
-  const ContentPartFileBase64(
-      {required this.data, required this.mediaType, this.filename, this.providerOptions});
-  static ContentPartFileBase64 fromJson(Map<String, dynamic> json) =>
-      ContentPartFileBase64(
-        data: json['data'] as String,
-        mediaType: json['media_type'] as String,
-        filename: json['filename'] as String?,
-        providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      );
+  const StreamPartToolApprovalRequest({
+    required this.approvalId,
+    required this.toolCall,
+    this.reason,
+    this.isAutomatic,
+    this.signature,
+  });
   @override
-  Map<String, dynamic> toJson() => {
-        'type': 'file_base64',
-        'data': data,
-        'media_type': mediaType,
-        if (filename != null) 'filename': filename,
-        if (providerOptions != null) 'provider_options': providerOptions,
+  Json toJson() => {
+        'type': 'tool-approval-request',
+        'approvalId': approvalId,
+        'toolCall': toolCall.toJson(),
+        if (reason != null) 'reason': reason,
+        if (isAutomatic != null) 'isAutomatic': isAutomatic,
+        if (signature != null) 'signature': signature,
       };
 }
 
-final class ContentPartFileUrl extends ContentPart {
-  final String url;
-  final String mediaType;
-  final Map<String, dynamic>? providerOptions;
-  const ContentPartFileUrl(
-      {required this.url, required this.mediaType, this.providerOptions});
-  static ContentPartFileUrl fromJson(Map<String, dynamic> json) =>
-      ContentPartFileUrl(
-        url: json['url'] as String,
-        mediaType: json['media_type'] as String,
-        providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      );
+/// Start / delta / end of a reasoning segment.
+final class StreamPartReasoningStart extends StreamPart {
+  final String id;
+  final Json? providerMetadata;
+  const StreamPartReasoningStart({required this.id, this.providerMetadata});
   @override
-  Map<String, dynamic> toJson() => {
-        'type': 'file_url',
-        'url': url,
-        'media_type': mediaType,
-        if (providerOptions != null) 'provider_options': providerOptions,
+  Json toJson() => {
+        'type': 'reasoning-start',
+        'id': id,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
-final class ContentPartFileReference extends ContentPart {
-  final String mediaType;
-  final Map<String, dynamic> reference;
-  final String? filename;
-  final Map<String, dynamic>? providerOptions;
-  const ContentPartFileReference(
-      {required this.mediaType, required this.reference, this.filename, this.providerOptions});
-  static ContentPartFileReference fromJson(Map<String, dynamic> json) =>
-      ContentPartFileReference(
-        mediaType: json['media_type'] as String,
-        reference: json['reference'] as Map<String, dynamic>,
-        filename: json['filename'] as String?,
-        providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      );
+final class StreamPartReasoningDelta extends StreamPart {
+  final String id;
+  final String delta;
+  final Json? providerMetadata;
+  const StreamPartReasoningDelta(
+      {required this.id, required this.delta, this.providerMetadata});
   @override
-  Map<String, dynamic> toJson() => {
-        'type': 'file_reference',
-        'media_type': mediaType,
-        'reference': reference,
-        if (filename != null) 'filename': filename,
-        if (providerOptions != null) 'provider_options': providerOptions,
+  Json toJson() => {
+        'type': 'reasoning-delta',
+        'id': id,
+        'delta': delta,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
-final class ContentPartReasoning extends ContentPart {
-  final String text;
-  final String? signature;
-  final Map<String, dynamic>? providerOptions;
-  const ContentPartReasoning(
-      {required this.text, this.signature, this.providerOptions});
-  static ContentPartReasoning fromJson(Map<String, dynamic> json) =>
-      ContentPartReasoning(
-        text: json['text'] as String,
-        signature: json['signature'] as String?,
-        providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      );
+final class StreamPartReasoningEnd extends StreamPart {
+  final String id;
+  final Json? providerMetadata;
+  const StreamPartReasoningEnd({required this.id, this.providerMetadata});
   @override
-  Map<String, dynamic> toJson() => {
-        'type': 'reasoning',
-        'text': text,
-        if (signature != null) 'signature': signature,
-        if (providerOptions != null) 'provider_options': providerOptions,
+  Json toJson() => {
+        'type': 'reasoning-end',
+        'id': id,
+        if (providerMetadata != null) 'providerMetadata': providerMetadata,
       };
 }
 
-final class ContentPartToolCall extends ContentPart {
-  final String toolCallId;
-  final String toolName;
-  final dynamic input;
-  final bool? providerExecuted;
-  final Map<String, dynamic>? providerOptions;
-  const ContentPartToolCall(
-      {required this.toolCallId, required this.toolName, required this.input, this.providerExecuted, this.providerOptions});
-  static ContentPartToolCall fromJson(Map<String, dynamic> json) =>
-      ContentPartToolCall(
-        toolCallId: json['tool_call_id'] as String,
-        toolName: json['tool_name'] as String,
-        input: json['input'],
-        providerExecuted: json['provider_executed'] as bool?,
-        providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      );
+/// A source / citation.
+final class StreamPartSource extends StreamPart {
+  final Source source;
+  const StreamPartSource({required this.source});
   @override
-  Map<String, dynamic> toJson() => {
-        'type': 'tool_call',
-        'tool_call_id': toolCallId,
-        'tool_name': toolName,
-        'input': input,
-        if (providerExecuted != null) 'provider_executed': providerExecuted,
-        if (providerOptions != null) 'provider_options': providerOptions,
-      };
+  Json toJson() => {'type': 'source', ...source.toJson()};
 }
 
-final class ContentPartToolResult extends ContentPart {
-  final String toolCallId;
-  final dynamic result;
-  final String? toolName;
-  final bool? isError;
-  final bool? preliminary;
-  final bool? isDynamic;
-  final Map<String, dynamic>? providerOptions;
-  const ContentPartToolResult(
-      {required this.toolCallId, required this.result, this.toolName, this.isError, this.preliminary, this.isDynamic, this.providerOptions});
-  static ContentPartToolResult fromJson(Map<String, dynamic> json) =>
-      ContentPartToolResult(
-        toolCallId: json['tool_call_id'] as String,
-        result: json['result'],
-        toolName: json['tool_name'] as String?,
-        isError: json['is_error'] as bool?,
-        preliminary: json['preliminary'] as bool?,
-        isDynamic: json['dynamic'] as bool?,
-        providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      );
+/// A raw provider chunk (when `includeRawChunks` is set).
+final class StreamPartRaw extends StreamPart {
+  final dynamic rawValue;
+  const StreamPartRaw({required this.rawValue});
   @override
-  Map<String, dynamic> toJson() => {
-        'type': 'tool_result',
-        'tool_call_id': toolCallId,
-        'result': result,
-        if (toolName != null) 'tool_name': toolName,
-        if (isError != null) 'is_error': isError,
-        if (preliminary != null) 'preliminary': preliminary,
-        if (isDynamic != null) 'dynamic': isDynamic,
-        if (providerOptions != null) 'provider_options': providerOptions,
-      };
-}
-
-final class ContentPartCustom extends ContentPart {
-  final String kind;
-  final Map<String, dynamic>? providerOptions;
-  ContentPartCustom({required this.kind, this.providerOptions});
-  factory ContentPartCustom.fromJson(Map<String, dynamic> json) => ContentPartCustom(kind: json['kind'] as String, providerOptions: json['provider_options'] as Map<String, dynamic>?);
-  @override Map<String, dynamic> toJson() => {'type': 'custom', 'kind': kind, if (providerOptions != null) 'provider_options': providerOptions};
+  Json toJson() => {'type': 'raw', 'rawValue': rawValue};
 }
 
-final class ContentPartReasoningFile extends ContentPart {
-  final GeneratedFileData data;
-  final String mediaType;
-  final Map<String, dynamic>? providerOptions;
-  ContentPartReasoningFile({required this.data, required this.mediaType, this.providerOptions});
-  factory ContentPartReasoningFile.fromJson(Map<String, dynamic> json) => ContentPartReasoningFile(data: GeneratedFileData.fromJson(json['data'] as Map<String, dynamic>), mediaType: json['media_type'] as String, providerOptions: json['provider_options'] as Map<String, dynamic>?);
-  @override Map<String, dynamic> toJson() => {'type': 'reasoning_file', 'data': data.toJson(), 'media_type': mediaType, if (providerOptions != null) 'provider_options': providerOptions};
-}
-
-final class ContentPartToolApprovalRequest extends ContentPart {
-  final String approvalId;
-  final String toolCallId;
-  final String? reason;
-  final bool? isAutomatic;
-  final String? signature;
-  final dynamic inputSchemaInput;
-  ContentPartToolApprovalRequest({required this.approvalId, required this.toolCallId, this.reason, this.isAutomatic, this.signature, this.inputSchemaInput});
-  factory ContentPartToolApprovalRequest.fromJson(Map<String, dynamic> json) => ContentPartToolApprovalRequest(approvalId: json['approval_id'] as String, toolCallId: json['tool_call_id'] as String, reason: json['reason'] as String?, isAutomatic: json['is_automatic'] as bool?, signature: json['signature'] as String?, inputSchemaInput: json['input_schema_input']);
-  @override Map<String, dynamic> toJson() => {'type': 'tool_approval_request', 'approval_id': approvalId, 'tool_call_id': toolCallId, if (reason != null) 'reason': reason, if (isAutomatic != null) 'is_automatic': isAutomatic, if (signature != null) 'signature': signature, if (inputSchemaInput != null) 'input_schema_input': inputSchemaInput};
-}
-
-final class ContentPartUnknown extends ContentPart {
-  final String tag;
-  final Map<String, dynamic> data;
-  const ContentPartUnknown({required this.tag, required this.data});
+/// Forward-compatibility: a `type` this binding does not model, verbatim.
+final class StreamPartUnknown extends StreamPart {
+  final Json data;
+  const StreamPartUnknown(this.data);
   @override
-  Map<String, dynamic> toJson() => data;
+  Json toJson() => data;
 }
 
 // ─────────────────────────────────────────────────────────────────────────────
diff --git a/bindings/flutter/lib/types.g.dart b/bindings/flutter/lib/types.g.dart
index 5de8fc8cb..db681d685 100644
--- a/bindings/flutter/lib/types.g.dart
+++ b/bindings/flutter/lib/types.g.dart
@@ -1,5 +1,4 @@
 // GENERATED CODE - DO NOT MODIFY BY HAND
-// ignore_for_file: type=lint, unused_element
 
 part of 'types.dart';
 
@@ -7,356 +6,6 @@ part of 'types.dart';
 // JsonSerializableGenerator
 // **************************************************************************
 
-InputTokenUsage _$InputTokenUsageFromJson(Map<String, dynamic> json) => InputTokenUsage(total: (json['total'] as num?)?.toInt(), noCache: (json['no_cache'] as num?)?.toInt(), cacheRead: (json['cache_read'] as num?)?.toInt(), cacheWrite: (json['cache_write'] as num?)?.toInt());
-Map<String, dynamic> _$InputTokenUsageToJson(InputTokenUsage instance) => {'total': instance.total, 'no_cache': instance.noCache, 'cache_read': instance.cacheRead, 'cache_write': instance.cacheWrite};
-
-OutputTokenUsage _$OutputTokenUsageFromJson(Map<String, dynamic> json) => OutputTokenUsage(total: (json['total'] as num?)?.toInt(), text: (json['text'] as num?)?.toInt(), reasoning: (json['reasoning'] as num?)?.toInt());
-Map<String, dynamic> _$OutputTokenUsageToJson(OutputTokenUsage instance) => {'total': instance.total, 'text': instance.text, 'reasoning': instance.reasoning};
-
-Usage _$UsageFromJson(Map<String, dynamic> json) => Usage(
-      inputTokens:
-          InputTokenUsage.fromJson(json['input_tokens'] as Map<String, dynamic>),
-      outputTokens:
-          OutputTokenUsage.fromJson(json['output_tokens'] as Map<String, dynamic>),
-      raw: json['raw'] as Map<String, dynamic>?,
-    );
-
-Map<String, dynamic> _$UsageToJson(Usage instance) => <String, dynamic>{
-      'input_tokens': instance.inputTokens,
-      'output_tokens': instance.outputTokens,
-      'raw': instance.raw,
-    };
-
-FinishReason _$FinishReasonFromJson(Map<String, dynamic> json) => FinishReason(
-      unified: json['unified'] as String,
-      raw: json['raw'] as String?,
-    );
-
-Map<String, dynamic> _$FinishReasonToJson(FinishReason instance) =>
-    <String, dynamic>{
-      'unified': instance.unified,
-      'raw': instance.raw,
-    };
-
-ToolCall _$ToolCallFromJson(Map<String, dynamic> json) => ToolCall(
-      toolCallId: json['tool_call_id'] as String,
-      toolName: json['tool_name'] as String,
-      input: json['input'],
-      providerExecuted: json['provider_executed'] as bool?,
-      isDynamic: json['dynamic'] as bool?,
-      providerMetadata: json['provider_metadata'],
-      invalid: json['invalid'] as bool?,
-      error: json['error'],
-    );
-
-Map<String, dynamic> _$ToolCallToJson(ToolCall instance) => <String, dynamic>{
-      'tool_call_id': instance.toolCallId,
-      'tool_name': instance.toolName,
-      'input': instance.input,
-      'provider_executed': instance.providerExecuted,
-      'dynamic': instance.isDynamic,
-      if (instance.providerMetadata != null)
-        'provider_metadata': instance.providerMetadata,
-      'invalid': instance.invalid,
-      'error': instance.error,
-    };
-
-FunctionTool _$FunctionToolFromJson(Map<String, dynamic> json) => FunctionTool(
-      name: json['name'] as String,
-      description: json['description'] as String?,
-      inputSchema: json['input_schema'] as Map<String, dynamic>,
-      strict: json['strict'] as bool?,
-      providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      inputExamples: (json['input_examples'] as List<dynamic>?)
-          ?.map((e) => FunctionToolInputExample.fromJson(e as Map<String, dynamic>))
-          .toList(),
-    );
-
-Map<String, dynamic> _$FunctionToolToJson(FunctionTool instance) =>
-    <String, dynamic>{
-      'name': instance.name,
-      'description': instance.description,
-      'input_schema': instance.inputSchema,
-      'strict': instance.strict,
-      'provider_options': instance.providerOptions,
-      'input_examples': instance.inputExamples?.map((e) => e.toJson()).toList(),
-    };
-
-ResponseMetadata _$ResponseMetadataFromJson(Map<String, dynamic> json) =>
-    ResponseMetadata(
-      id: json['id'] as String?,
-      timestamp: json['timestamp'] as String?,
-      modelId: json['model_id'] as String?,
-    );
-
-Map<String, dynamic> _$ResponseMetadataToJson(ResponseMetadata instance) =>
-    <String, dynamic>{
-      'id': instance.id,
-      'timestamp': instance.timestamp,
-      'model_id': instance.modelId,
-    };
-
-GenerateTextResult _$GenerateTextResultFromJson(Map<String, dynamic> json) =>
-    GenerateTextResult(
-      content: (json['content'] as List<dynamic>? ?? []).map((item) => item as Map<String, dynamic>).toList(),
-      text: json['text'] as String,
-      toolCalls: (json['tool_calls'] as List<dynamic>)
-          .map((e) => ToolCall.fromJson(e as Map<String, dynamic>))
-          .toList(),
-      finishReason:
-          FinishReason.fromJson(json['finish_reason'] as Map<String, dynamic>),
-      usage: Usage.fromJson(json['usage'] as Map<String, dynamic>),
-      warnings: (json['warnings'] as List<dynamic>?)
-              ?.map((e) => e as Map<String, dynamic>)
-              .toList() ??
-          const [],
-      raw: GenerateResult.fromJson(json['raw'] as Map<String, dynamic>),
-      reasoning: (json['reasoning'] as List<dynamic>?)
-              ?.map((e) => e as Map<String, dynamic>)
-              .toList() ??
-          const [],
-      reasoningText: json['reasoning_text'] as String? ?? '',
-      sources: (json['sources'] as List<dynamic>?)
-              ?.map((e) => Source.fromJson(e as Map<String, dynamic>))
-              .toList() ??
-          const [],
-      files: (json['files'] as List<dynamic>?)
-              ?.map((e) => e as Map<String, dynamic>)
-              .toList() ??
-          const [],
-      responseMessages: (json['response_messages'] as List<dynamic>?)
-              ?.map((e) => ModelMessage.fromJson(e as Map<String, dynamic>))
-              .toList() ??
-          const [],
-      rawFinishReason: json['raw_finish_reason'] as String?,
-      providerMetadata: json['provider_metadata'] as Map<String, dynamic>?,
-      response: json['response'] != null
-          ? ResponseMetadata.fromJson(json['response'] as Map<String, dynamic>)
-          : ResponseMetadata(),
-      totalUsage: json['total_usage'] != null
-          ? Usage.fromJson(json['total_usage'] as Map<String, dynamic>)
-          : Usage(inputTokens: InputTokenUsage(), outputTokens: OutputTokenUsage()),
-    );
-
-Map<String, dynamic> _$GenerateTextResultToJson(GenerateTextResult instance) =>
-    <String, dynamic>{
-      'text': instance.text,
-      'content': instance.content,
-      'tool_calls': instance.toolCalls,
-      'finish_reason': instance.finishReason,
-      'usage': instance.usage,
-      'warnings': instance.warnings,
-      'raw': instance.raw,
-      'reasoning': instance.reasoning,
-      'reasoning_text': instance.reasoningText,
-      'sources': instance.sources,
-      'files': instance.files,
-      'response_messages': instance.responseMessages,
-      'raw_finish_reason': instance.rawFinishReason,
-      'provider_metadata': instance.providerMetadata,
-      'response': instance.response,
-      'total_usage': instance.totalUsage,
-    };
-
-GenerateObjectResult _$GenerateObjectResultFromJson(
-        Map<String, dynamic> json) =>
-    GenerateObjectResult(
-      object: json['object'],
-      finishReason:
-          FinishReason.fromJson(json['finish_reason'] as Map<String, dynamic>),
-      rawFinishReason: json['raw_finish_reason'] as String?,
-      usage: json['usage'] != null
-          ? Usage.fromJson(json['usage'] as Map<String, dynamic>)
-          : Usage(inputTokens: InputTokenUsage(), outputTokens: OutputTokenUsage()),
-      warnings: (json['warnings'] as List<dynamic>?)
-              ?.map((e) => e as Map<String, dynamic>)
-              .toList() ??
-          const [],
-      reasoning: json['reasoning'] as String?,
-      providerMetadata: json['provider_metadata'] as Map<String, dynamic>?,
-      response: json['response'] != null
-          ? ResponseMetadata.fromJson(json['response'] as Map<String, dynamic>)
-          : ResponseMetadata(),
-      raw: GenerateTextResult.fromJson(json['raw'] as Map<String, dynamic>),
-    );
-
-Map<String, dynamic> _$GenerateObjectResultToJson(
-        GenerateObjectResult instance) =>
-    <String, dynamic>{
-      'object': instance.object,
-      'finish_reason': instance.finishReason,
-      'raw_finish_reason': instance.rawFinishReason,
-      'usage': instance.usage,
-      'warnings': instance.warnings,
-      'reasoning': instance.reasoning,
-      'provider_metadata': instance.providerMetadata,
-      'response': instance.response,
-      'raw': instance.raw,
-    };
-
-StreamTextResultAggregated _$StreamTextResultAggregatedFromJson(
-        Map<String, dynamic> json) =>
-    StreamTextResultAggregated(
-      content: (json['content'] as List<dynamic>? ?? []).map((item) => item as Map<String, dynamic>).toList(),
-      text: json['text'] as String? ?? '',
-      reasoning: (json['reasoning'] as List<dynamic>?)
-              ?.map((e) => e as Map<String, dynamic>)
-              .toList() ??
-          const [],
-      reasoningText: json['reasoning_text'] as String? ?? '',
-      toolCalls: (json['tool_calls'] as List<dynamic>?)
-              ?.map((e) => ToolCall.fromJson(e as Map<String, dynamic>))
-              .toList() ??
-          const [],
-      sources: (json['sources'] as List<dynamic>?)
-              ?.map((e) => Source.fromJson(e as Map<String, dynamic>))
-              .toList() ??
-          const [],
-      files: (json['files'] as List<dynamic>?)
-              ?.map((e) => e as Map<String, dynamic>)
-              .toList() ??
-          const [],
-      finishReason:
-          FinishReason.fromJson(json['finish_reason'] as Map<String, dynamic>),
-      rawFinishReason: json['raw_finish_reason'] as String?,
-      usage: json['usage'] != null
-          ? Usage.fromJson(json['usage'] as Map<String, dynamic>)
-          : Usage(inputTokens: InputTokenUsage(), outputTokens: OutputTokenUsage()),
-      totalUsage: json['total_usage'] != null
-          ? Usage.fromJson(json['total_usage'] as Map<String, dynamic>)
-          : Usage(inputTokens: InputTokenUsage(), outputTokens: OutputTokenUsage()),
-      warnings: (json['warnings'] as List<dynamic>?)
-              ?.map((e) => e as Map<String, dynamic>)
-              .toList() ??
-          const [],
-      providerMetadata: json['provider_metadata'] as Map<String, dynamic>?,
-      response: json['response'] == null
-          ? null
-          : ResponseMetadata.fromJson(json['response'] as Map<String, dynamic>),
-      responseMessages: (json['response_messages'] as List<dynamic>?)
-              ?.map((e) => ModelMessage.fromJson(e as Map<String, dynamic>))
-              .toList() ??
-          const [],
-    );
-
-Map<String, dynamic> _$StreamTextResultAggregatedToJson(
-        StreamTextResultAggregated instance) =>
-    <String, dynamic>{
-      'text': instance.text,
-      'content': instance.content,
-      'reasoning': instance.reasoning,
-      'reasoning_text': instance.reasoningText,
-      'tool_calls': instance.toolCalls,
-      'sources': instance.sources,
-      'files': instance.files,
-      'finish_reason': instance.finishReason,
-      'raw_finish_reason': instance.rawFinishReason,
-      'usage': instance.usage,
-      'total_usage': instance.totalUsage,
-      'warnings': instance.warnings,
-      'provider_metadata': instance.providerMetadata,
-      'response': instance.response,
-      'response_messages': instance.responseMessages,
-    };
-
-TimeoutConfiguration _$TimeoutConfigurationFromJson(
-        Map<String, dynamic> json) =>
-    TimeoutConfiguration(
-      totalMs: (json['total_ms'] as num?)?.toInt(),
-      stepMs: (json['step_ms'] as num?)?.toInt(),
-      firstChunkMs: (json['first_chunk_ms'] as num?)?.toInt(),
-      chunkMs: (json['chunk_ms'] as num?)?.toInt(),
-    );
-
-Map<String, dynamic> _$TimeoutConfigurationToJson(
-        TimeoutConfiguration instance) =>
-    <String, dynamic>{
-      'total_ms': instance.totalMs,
-      'step_ms': instance.stepMs,
-      'first_chunk_ms': instance.firstChunkMs,
-      'chunk_ms': instance.chunkMs,
-    };
-
-GenerateTextOptions _$GenerateTextOptionsFromJson(Map<String, dynamic> json) =>
-    GenerateTextOptions(
-      maxOutputTokens: (json['max_output_tokens'] as num?)?.toInt(),
-      temperature: (json['temperature'] as num?)?.toDouble(),
-      stopSequences: (json['stop_sequences'] as List<dynamic>?)
-          ?.map((e) => e as String)
-          .toList(),
-      topP: (json['top_p'] as num?)?.toDouble(),
-      topK: (json['top_k'] as num?)?.toDouble(),
-      presencePenalty: (json['presence_penalty'] as num?)?.toDouble(),
-      frequencyPenalty: (json['frequency_penalty'] as num?)?.toDouble(),
-      responseFormat: json['response_format'] as Map<String, dynamic>?,
-      seed: (json['seed'] as num?)?.toInt(),
-      tools: (json['tools'] as List<dynamic>?)
-          ?.map((e) => Tool.fromJson(e as Map<String, dynamic>))
-          .toList(),
-      toolChoice: json['tool_choice'] == null
-          ? null
-          : ToolChoice.fromJson(json['tool_choice']),
-      headers: (json['headers'] as Map<String, dynamic>?)?.map(
-        (k, e) => MapEntry(k, e as String),
-      ),
-      providerOptions: json['provider_options'] as Map<String, dynamic>?,
-      reasoning:
-          $enumDecodeNullable(_$ReasoningEffortEnumMap, json['reasoning']),
-      instructions: json['instructions'] as String?,
-      maxRetries: (json['max_retries'] as num?)?.toInt(),
-      timeout: json['timeout'] == null
-          ? null
-          : TimeoutConfiguration.fromJson(
-              json['timeout'] as Map<String, dynamic>),
-    );
-
-Map<String, dynamic> _$GenerateTextOptionsToJson(
-        GenerateTextOptions instance) =>
-    <String, dynamic>{
-      'max_output_tokens': instance.maxOutputTokens,
-      'temperature': instance.temperature,
-      'stop_sequences': instance.stopSequences,
-      'top_p': instance.topP,
-      'top_k': instance.topK,
-      'presence_penalty': instance.presencePenalty,
-      'frequency_penalty': instance.frequencyPenalty,
-      'response_format': instance.responseFormat,
-      'seed': instance.seed,
-      'tools': instance.tools,
-      'tool_choice': instance.toolChoice,
-      'headers': instance.headers,
-      'provider_options': instance.providerOptions,
-      'reasoning': instance.reasoning,
-      'instructions': instance.instructions,
-      'max_retries': instance.maxRetries,
-      'timeout': instance.timeout,
-    };
-
-const _$ReasoningEffortEnumMap = {
-  ReasoningEffort.providerDefault: 'providerDefault',
-  ReasoningEffort.none: 'none',
-  ReasoningEffort.minimal: 'minimal',
-  ReasoningEffort.low: 'low',
-  ReasoningEffort.medium: 'medium',
-  ReasoningEffort.high: 'high',
-  ReasoningEffort.xhigh: 'xhigh',
-};
-
-ModelMessage _$ModelMessageFromJson(Map<String, dynamic> json) => ModelMessage(
-      role: json['role'] as String,
-      content: json['content'] as Object,
-    );
-
-Map<String, dynamic> _$ModelMessageToJson(ModelMessage instance) =>
-    <String, dynamic>{
-      'role': instance.role,
-      'content': instance.content,
-    };
-
-// ─────────────────────────────────────────────────────────────────────────────
-// OpenAI Chat Completions output (RFC-0026).
-// ─────────────────────────────────────────────────────────────────────────────
-
 ChatCompletion _$ChatCompletionFromJson(Map<String, dynamic> json) =>
     ChatCompletion(
       id: json['id'] as String,
@@ -366,7 +15,8 @@ ChatCompletion _$ChatCompletionFromJson(Map<String, dynamic> json) =>
       choices: (json['choices'] as List<dynamic>)
           .map((e) => ChatCompletionChoice.fromJson(e as Map<String, dynamic>))
           .toList(),
-      usage: ChatCompletionUsage.fromJson(json['usage'] as Map<String, dynamic>),
+      usage:
+          ChatCompletionUsage.fromJson(json['usage'] as Map<String, dynamic>),
       systemFingerprint: json['system_fingerprint'] as String?,
     );
 
@@ -381,7 +31,8 @@ Map<String, dynamic> _$ChatCompletionToJson(ChatCompletion instance) =>
       'system_fingerprint': instance.systemFingerprint,
     };
 
-ChatCompletionChoice _$ChatCompletionChoiceFromJson(Map<String, dynamic> json) =>
+ChatCompletionChoice _$ChatCompletionChoiceFromJson(
+        Map<String, dynamic> json) =>
     ChatCompletionChoice(
       index: (json['index'] as num).toInt(),
       message: ChatCompletionMessage.fromJson(
@@ -406,7 +57,8 @@ ChatCompletionMessage _$ChatCompletionMessageFromJson(
       content: json['content'] as String?,
       reasoningContent: json['reasoning_content'] as String?,
       toolCalls: (json['tool_calls'] as List<dynamic>?)
-          ?.map((e) => ChatCompletionToolCall.fromJson(e as Map<String, dynamic>))
+          ?.map(
+              (e) => ChatCompletionToolCall.fromJson(e as Map<String, dynamic>))
           .toList(),
       annotations: json['annotations'] as List<dynamic>?,
     );
@@ -532,8 +184,8 @@ ChatCompletionChunkChoice _$ChatCompletionChunkChoiceFromJson(
         Map<String, dynamic> json) =>
     ChatCompletionChunkChoice(
       index: (json['index'] as num).toInt(),
-      delta: ChatCompletionDelta.fromJson(
-          json['delta'] as Map<String, dynamic>),
+      delta:
+          ChatCompletionDelta.fromJson(json['delta'] as Map<String, dynamic>),
       finishReason: json['finish_reason'] as String?,
       logprobs: json['logprobs'],
     );
diff --git a/bindings/flutter/test/aimux_test.dart b/bindings/flutter/test/aimux_test.dart
index aaa09384d..27be6cd13 100644
--- a/bindings/flutter/test/aimux_test.dart
+++ b/bindings/flutter/test/aimux_test.dart
@@ -154,14 +154,6 @@ void main() {
     );
   });
 
-  test('registerProviders validates raw config_json in Dart', () {
-    expect(
-      () => registerProviders('{not json'),
-      throwsA(isA<FormatException>()
-          .having((e) => e.source, 'source', 'config_json')),
-    );
-  });
-
   test('multimodal use-after-close throws StateError', () {
     final m = EmbeddingModel.openai('sk-test-fake-key', 'text-embedding-3-small');
     m.close();
diff --git a/bindings/flutter/test/content_part_test.dart b/bindings/flutter/test/content_part_test.dart
index 898211e9e..accff16f3 100644
--- a/bindings/flutter/test/content_part_test.dart
+++ b/bindings/flutter/test/content_part_test.dart
@@ -1,222 +1,144 @@
-import 'package:test/test.dart';
+// Round trips for the message-content types against the generated wire shape:
+// `bindings/node/src/types/{ContentPart,FileData,FileBytes,ToolResultOutput}.ts`.
 import 'package:aimux/types.dart';
+import 'package:test/test.dart';
 
 void main() {
-  group('FileBytes round-trip', () {
-    test('Binary variant', () {
-      final original = FileBytesBinary(data: [104, 105]);
-      final json = original.toJson();
-      expect(json, {'Binary': [104, 105]});
-      final decoded = FileBytes.fromJson(json);
-      expect(decoded, isA<FileBytesBinary>());
-      expect((decoded as FileBytesBinary).data, [104, 105]);
+  group('FileBytes (untagged: number[] | string)', () {
+    test('binary is a bare array', () {
+      expect(FileBytesBinary(data: [104, 105]).toJson(), [104, 105]);
+      expect(FileBytes.fromJson([104, 105]), isA<FileBytesBinary>());
     });
 
-    test('Base64 variant', () {
-      final original = FileBytesBase64(data: 'aGk=');
-      final json = original.toJson();
-      expect(json, {'Base64': 'aGk='});
-      final decoded = FileBytes.fromJson(json);
-      expect(decoded, isA<FileBytesBase64>());
-      expect((decoded as FileBytesBase64).data, 'aGk=');
+    test('base64 is a bare string', () {
+      expect(FileBytesBase64(data: 'aGk=').toJson(), 'aGk=');
+      expect((FileBytes.fromJson('aGk=') as FileBytesBase64).data, 'aGk=');
     });
   });
 
-  group('FileData round-trip', () {
-    test('Data variant', () {
-      final original = FileDataData(data: FileBytesBase64(data: 'aGk='));
-      final json = original.toJson();
-      expect(json, {
-        'Data': {'data': {'Base64': 'aGk='}}
-      });
-      final decoded = FileData.fromJson(json);
-      expect(decoded, isA<FileDataData>());
-      final d = decoded as FileDataData;
-      expect(d.data, isA<FileBytesBase64>());
-      expect((d.data as FileBytesBase64).data, 'aGk=');
-    });
-
-    test('Url variant', () {
-      final original = FileDataUrl(url: 'https://example.com/f.png');
-      final json = original.toJson();
-      expect(json, {
-        'Url': {'url': 'https://example.com/f.png'}
-      });
-      final decoded = FileData.fromJson(json);
-      expect(decoded, isA<FileDataUrl>());
-      expect((decoded as FileDataUrl).url, 'https://example.com/f.png');
-    });
-
-    test('Reference variant', () {
-      final original = FileDataReference(reference: {'openai': 'file-abc'});
-      final json = original.toJson();
-      expect(json, {
-        'Reference': {
-          'reference': {'openai': 'file-abc'}
-        }
-      });
-      final decoded = FileData.fromJson(json);
-      expect(decoded, isA<FileDataReference>());
-      expect((decoded as FileDataReference).reference, {'openai': 'file-abc'});
-    });
-
-    test('Text variant', () {
-      final original = FileDataText(text: 'hello world');
-      final json = original.toJson();
-      expect(json, {
-        'Text': {'text': 'hello world'}
-      });
-      final decoded = FileData.fromJson(json);
-      expect(decoded, isA<FileDataText>());
-      expect((decoded as FileDataText).text, 'hello world');
+  group('FileData ("type"-tagged)', () {
+    test('data / url / reference / text round-trip', () {
+      final cases = <Map<String, dynamic>>[
+        {'type': 'data', 'data': 'aGk='},
+        {'type': 'url', 'url': 'https://example.com/f.png', 'originalUrl': 'x'},
+        {'type': 'reference', 'reference': {'openai': 'file-abc'}},
+        {'type': 'text', 'text': 'hello world'},
+      ];
+      for (final json in cases) {
+        expect(FileData.fromJson(json).toJson(), json);
+      }
+      expect(FileData.fromJson(cases[0]), isA<FileDataData>());
+      expect(FileData.fromJson(cases[1]), isA<FileDataUrl>());
+      expect(FileData.fromJson(cases[2]), isA<FileDataReference>());
+      expect(FileData.fromJson(cases[3]), isA<FileDataText>());
     });
   });
 
   group('ContentPart round-trip', () {
-    test('Text variant', () {
-      final original = ContentPartText(text: 'hello');
-      final json = original.toJson();
-      expect(json['type'], 'text');
-      expect(json['text'], 'hello');
-      final decoded = ContentPart.fromJson(json);
-      expect(decoded, isA<ContentPartText>());
-      expect((decoded as ContentPartText).text, 'hello');
-    });
-
-    test('Image variant', () {
-      final original = ContentPartImage(image: [1, 2, 3], mediaType: 'image/png');
-      final json = original.toJson();
-      expect(json['type'], 'image');
-      expect(json['image'], [1, 2, 3]);
-      expect(json['media_type'], 'image/png');
-      final decoded = ContentPart.fromJson(json);
-      expect(decoded, isA<ContentPartImage>());
-      final d = decoded as ContentPartImage;
-      expect(d.image, [1, 2, 3]);
-      expect(d.mediaType, 'image/png');
-    });
-
-    test('File variant — has filename', () {
-      final original = ContentPartFile(
-        data: [104, 105],
-        mediaType: 'application/pdf',
-        filename: 'doc.pdf',
-      );
-      final json = original.toJson();
-      expect(json['type'], 'file');
-      expect(json, contains('filename'));
-      expect(json['filename'], 'doc.pdf');
-      final decoded = ContentPart.fromJson(json);
-      expect(decoded, isA<ContentPartFile>());
-      final d = decoded as ContentPartFile;
-      expect(d.data, [104, 105]);
-      expect(d.mediaType, 'application/pdf');
-      expect(d.filename, 'doc.pdf');
-    });
-
-    test('FileBase64 variant', () {
-      final original = ContentPartFileBase64(
-        data: 'aGk=',
-        mediaType: 'image/png',
-        filename: 'img.png',
-      );
-      final json = original.toJson();
-      expect(json['type'], 'file_base64');
-      expect(json['data'], 'aGk=');
-      final decoded = ContentPart.fromJson(json);
-      expect(decoded, isA<ContentPartFileBase64>());
-      expect((decoded as ContentPartFileBase64).data, 'aGk=');
-    });
-
-    test('FileUrl variant', () {
-      final original = ContentPartFileUrl(
-        url: 'https://example.com/f.png',
-        mediaType: 'image/png',
-      );
-      final json = original.toJson();
-      expect(json['type'], 'file_url');
-      expect(json['url'], 'https://example.com/f.png');
-      final decoded = ContentPart.fromJson(json);
-      expect(decoded, isA<ContentPartFileUrl>());
-      expect((decoded as ContentPartFileUrl).url, 'https://example.com/f.png');
-    });
-
-    test('FileReference variant', () {
-      final original = ContentPartFileReference(
-        mediaType: 'application/pdf',
-        reference: {'openai': 'file-abc'},
-        filename: 'doc.pdf',
-      );
-      final json = original.toJson();
-      expect(json['type'], 'file_reference');
-      expect(json['reference'], {'openai': 'file-abc'});
-      final decoded = ContentPart.fromJson(json);
-      expect(decoded, isA<ContentPartFileReference>());
-      expect((decoded as ContentPartFileReference).reference, {'openai': 'file-abc'});
-    });
-
-    test('Reasoning variant', () {
-      final original = ContentPartReasoning(text: 'thinking...', signature: 'sig123');
-      final json = original.toJson();
-      expect(json['type'], 'reasoning');
-      expect(json['text'], 'thinking...');
-      expect(json['signature'], 'sig123');
-      final decoded = ContentPart.fromJson(json);
-      expect(decoded, isA<ContentPartReasoning>());
-      expect((decoded as ContentPartReasoning).signature, 'sig123');
-    });
-
-    test('ToolCall variant', () {
-      final original = ContentPartToolCall(
-        toolCallId: 'call_1',
-        toolName: 'get_weather',
-        input: {'location': 'Tokyo'},
-        providerExecuted: true,
-      );
-      final json = original.toJson();
-      expect(json['type'], 'tool_call');
-      expect(json['tool_call_id'], 'call_1');
-      expect(json['tool_name'], 'get_weather');
-      expect(json['provider_executed'], true);
-      final decoded = ContentPart.fromJson(json);
-      expect(decoded, isA<ContentPartToolCall>());
-      final d = decoded as ContentPartToolCall;
-      expect(d.toolCallId, 'call_1');
-      expect(d.toolName, 'get_weather');
-      expect(d.input, {'location': 'Tokyo'});
-      expect(d.providerExecuted, true);
-    });
+    // One wire document per variant; optional fields are absent, not null.
+    final cases = <String, Map<String, dynamic>>{
+      'text': {'type': 'text', 'text': 'hello'},
+      'image': {'type': 'image', 'image': [1, 2, 3], 'mediaType': 'image/png'},
+      'file': {
+        'type': 'file',
+        'data': [104, 105],
+        'mediaType': 'application/pdf',
+        'filename': 'doc.pdf',
+      },
+      'file-base64': {
+        'type': 'file-base64',
+        'data': 'aGk=',
+        'mediaType': 'image/png',
+      },
+      'file-url': {
+        'type': 'file-url',
+        'url': 'https://example.com/f.png',
+        'mediaType': 'image/png',
+      },
+      'file-reference': {
+        'type': 'file-reference',
+        'mediaType': 'application/pdf',
+        'reference': {'openai': 'file-abc'},
+      },
+      'reasoning': {'type': 'reasoning', 'text': 'thinking...', 'signature': 'sig'},
+      'reasoning-file': {
+        'type': 'reasoning-file',
+        'data': {'type': 'data', 'data': 'aGk='},
+        'mediaType': 'image/png',
+      },
+      'custom': {
+        'type': 'custom',
+        'kind': 'k',
+        'providerOptions': {'a': {'b': 1}},
+      },
+      'tool-approval-request': {
+        'type': 'tool-approval-request',
+        'approvalId': 'ap1',
+        'toolCallId': 'call_1',
+        'isAutomatic': true,
+      },
+      'tool-call': {
+        'type': 'tool-call',
+        'toolCallId': 'call_1',
+        'toolName': 'get_weather',
+        'input': {'location': 'Tokyo'},
+        'providerExecuted': true,
+      },
+      'tool-result': {
+        'type': 'tool-result',
+        'toolCallId': 'call_1',
+        'toolName': 'get_weather',
+        'output': {
+          'type': 'json',
+          'value': {'temp': 22},
+        },
+      },
+    };
+
+    for (final entry in cases.entries) {
+      test(entry.key, () {
+        final part = ContentPart.fromJson(entry.value);
+        expect(part, isNot(isA<ContentPartUnknown>()));
+        expect(part.toJson(), entry.value);
+      });
+    }
 
-    test('ToolResult variant — uses result, not output', () {
-      final original = ContentPartToolResult(
+    test('tool-result carries a typed output, not result/isError', () {
+      final part = ContentPartToolResult(
         toolCallId: 'call_1',
-        result: {'temp': 22},
         toolName: 'get_weather',
-        isError: false,
-        preliminary: null,
-        isDynamic: null,
+        output: ToolResultOutput.errorText('boom'),
       );
-      final json = original.toJson();
-      expect(json['type'], 'tool_result');
-      expect(json, contains('result'));
-      expect(json, isNot(contains('output')));
-      expect(json['tool_name'], 'get_weather');
-      expect(json['is_error'], false);
-      final decoded = ContentPart.fromJson(json);
-      expect(decoded, isA<ContentPartToolResult>());
-      final d = decoded as ContentPartToolResult;
-      expect(d.result, {'temp': 22});
-      expect(d.toolName, 'get_weather');
-      expect(d.isError, false);
-    });
-
-    test('Unknown variant — forward compat', () {
-      final json = {'type': 'future_variant', 'data': 'something'};
-      final decoded = ContentPart.fromJson(json);
-      expect(decoded, isA<ContentPartUnknown>());
-      final unknown = decoded as ContentPartUnknown;
-      expect(unknown.tag, 'future_variant');
-      expect(unknown.data['data'], 'something');
+      final json = part.toJson();
+      expect(json, isNot(contains('result')));
+      expect(json, isNot(contains('isError')));
+      expect(json['output'], {'type': 'error-text', 'value': 'boom'});
+    });
+
+    test('ToolResultOutput variants', () {
+      for (final json in <Map<String, dynamic>>[
+        {'type': 'text', 'value': 'ok'},
+        {'type': 'json', 'value': {'a': 1}},
+        {'type': 'execution-denied', 'reason': 'no'},
+        {'type': 'execution-denied'},
+        {'type': 'error-text', 'value': 'e'},
+        {'type': 'error-json', 'value': [1]},
+        {
+          'type': 'content',
+          'value': [
+            {'type': 'text', 'text': 't'}
+          ]
+        },
+      ]) {
+        expect(ToolResultOutput.fromJson(json).toJson(), json);
+      }
+    });
+
+    test('unknown type passes through verbatim', () {
+      final json = {'type': 'future-variant', 'data': 'something'};
+      final part = ContentPart.fromJson(json);
+      expect(part, isA<ContentPartUnknown>());
+      expect(part.toJson(), json);
     });
   });
 }
diff --git a/bindings/flutter/test/errors_test.dart b/bindings/flutter/test/errors_test.dart
index bd9546cf3..6e801cabb 100644
--- a/bindings/flutter/test/errors_test.dart
+++ b/bindings/flutter/test/errors_test.dart
@@ -90,10 +90,10 @@ void main() {
     test('unknown code is rejected with StateError', () {
       // A code outside the published table is an ABI mismatch, not an error
       // kind. 1 is AIMUX_E_OTHER now because Other inherited the old UNKNOWN
-      // slot, so it resolves; 4 is retired and 20 is the first unassigned value.
+      // slot, so it resolves; 4 is retired and 21 is the first unassigned value.
       expect(() => AimuxException.fromCode(999, 'future'), throwsStateError);
       expect(() => AimuxException.fromCode(4, 'retired'), throwsStateError);
-      expect(() => AimuxException.fromCode(20, 'unassigned'), throwsStateError);
+      expect(() => AimuxException.fromCode(21, 'unassigned'), throwsStateError);
     });
 
     test('bare retry code synthesizes a single-attempt RetryError', () {
diff --git a/bindings/flutter/test/multimodal_e2e_test.dart b/bindings/flutter/test/multimodal_e2e_test.dart
index 9737941ea..0bab55efe 100644
--- a/bindings/flutter/test/multimodal_e2e_test.dart
+++ b/bindings/flutter/test/multimodal_e2e_test.dart
@@ -118,7 +118,7 @@ Future<MockOpenAIServer> startMockServer(
 // the worker so the main isolate's event loop stays free to serve the mock.
 //
 // Call options that the Rust core deserializes (speech/image/rerank/search)
-// always carry `"provider_options":{}` — matching the Go wire format, where the
+// always carry `"providerOptions":{}` — matching the Go wire format, where the
 // SharedProviderOptions map must be present (not null). embed/transcription/
 // uploadFile pass no options (null pointer), matching the Go nil-opts path.
 
@@ -251,7 +251,7 @@ const String speechAudioBody = 'SGVsbG8gd29ybGQ=';
 
 /// Parsed-only fixture for the video result wire format.
 const String videoResultJson = r'''
-{"videos":[{"Url":{"url":"https://example.com/v.mp4","media_type":"video/mp4"}}]}
+{"videos":[{"type":"url","url":"https://example.com/v.mp4","mediaType":"video/mp4"}]}
 ''';
 
 // ─────────────────────────────────────────────────────────────────────────────
@@ -293,8 +293,8 @@ void main() {
     final optsJson = jsonEncode({
       'text': 'Hi',
       'voice': 'alloy',
-      'output_format': 'mp3',
-      'provider_options': <String, dynamic>{},
+      'outputFormat': 'mp3',
+      'providerOptions': <String, dynamic>{},
     });
 
     final result =
@@ -302,9 +302,8 @@ void main() {
 
     final decoded = jsonDecode(result) as Map<String, dynamic>;
     expect(decoded, contains('audio'));
-    final audio = decoded['audio'] as Map<String, dynamic>;
-    expect(audio, contains('Binary'));
-    final binary = audio['Binary'] as List;
+    // AudioData is untagged: a base64 string or a byte array.
+    final binary = decoded['audio'] as List;
     expect(binary, isNotEmpty);
 
     expect(server.recorded, hasLength(1));
@@ -320,7 +319,7 @@ void main() {
     final optsJson = jsonEncode({
       'prompt': 'otter',
       'n': 1,
-      'provider_options': <String, dynamic>{},
+      'providerOptions': <String, dynamic>{},
     });
 
     final result =
@@ -328,9 +327,8 @@ void main() {
 
     final decoded = jsonDecode(result) as Map<String, dynamic>;
     expect(decoded, contains('images'));
-    final images = decoded['images'] as Map<String, dynamic>;
-    expect(images, contains('Base64'));
-    final base64List = images['Base64'] as List;
+    // ImageOutputs is untagged: an array of base64 strings or of byte arrays.
+    final base64List = decoded['images'] as List;
     expect(base64List, hasLength(1));
     expect(base64List[0], 'aW1hZ2Ux');
 
@@ -362,9 +360,9 @@ void main() {
 
     final optsJson = jsonEncode({
       'query': 'which?',
-      'documents': {'Text': {'values': ['doc1', 'doc2']}},
-      'top_n': 2,
-      'provider_options': <String, dynamic>{},
+      'documents': {'type': 'text', 'values': ['doc1', 'doc2']},
+      'topN': 2,
+      'providerOptions': <String, dynamic>{},
     });
 
     final result =
@@ -374,7 +372,7 @@ void main() {
     expect(decoded, contains('ranking'));
     final ranking = decoded['ranking'] as List;
     expect(ranking, hasLength(2));
-    expect((ranking[0] as Map)['relevance_score'], 0.95);
+    expect((ranking[0] as Map)['relevanceScore'], 0.95);
     expect((ranking[0] as Map)['index'], 1);
 
     expect(server.recorded, hasLength(1));
@@ -389,8 +387,8 @@ void main() {
 
     final optsJson = jsonEncode({
       'query': 'What is Rust?',
-      'max_results': 5,
-      'provider_options': <String, dynamic>{},
+      'maxResults': 5,
+      'providerOptions': <String, dynamic>{},
     });
 
     final result = await runSearch(server.baseUrl, apiKey, optsJson);
@@ -417,8 +415,8 @@ void main() {
         server.baseUrl, apiKey, 'dGVzdA==', 'application/pdf');
 
     final decoded = jsonDecode(result) as Map<String, dynamic>;
-    expect(decoded, contains('provider_reference'));
-    final providerRef = decoded['provider_reference'] as Map<String, dynamic>;
+    expect(decoded, contains('providerReference'));
+    final providerRef = decoded['providerReference'] as Map<String, dynamic>;
     expect(providerRef['openai'], 'file-abc');
 
     expect(server.recorded, hasLength(1));
@@ -435,8 +433,9 @@ void main() {
     expect(decoded, contains('videos'));
     final videos = decoded['videos'] as List;
     expect(videos, hasLength(1));
-    final urlData = (videos[0] as Map)['Url'] as Map<String, dynamic>;
-    expect(urlData['url'], 'https://example.com/v.mp4');
-    expect(urlData['media_type'], 'video/mp4');
+    final video = videos[0] as Map<String, dynamic>;
+    expect(video['type'], 'url');
+    expect(video['url'], 'https://example.com/v.mp4');
+    expect(video['mediaType'], 'video/mp4');
   });
 }
diff --git a/bindings/flutter/test/structured_e2e_test.dart b/bindings/flutter/test/structured_e2e_test.dart
index d5a9a3fba..b1ce1bf4b 100644
--- a/bindings/flutter/test/structured_e2e_test.dart
+++ b/bindings/flutter/test/structured_e2e_test.dart
@@ -28,9 +28,10 @@ import 'dart:convert';
 import 'dart:io';
 import 'dart:isolate';
 
-import 'package:aimux/aimux.dart';
 import 'package:test/test.dart';
 
+import 'typed_model_test.dart' show mockChatModel;
+
 // ─────────────────────────────────────────────────────────────────────────────
 // Mock server
 // ─────────────────────────────────────────────────────────────────────────────
@@ -115,7 +116,7 @@ class _GenerateArgs {
 /// Run a non-streaming `generateText` in a worker isolate.
 Future<Map<String, dynamic>> runGenerateText(_GenerateArgs args) {
   return Isolate.run(() {
-    final model = Model.openai(args.apiKey, args.modelId, baseUrl: args.baseUrl);
+    final model = mockChatModel(args.baseUrl, args.apiKey, args.modelId);
     try {
       return model.generateText(args.prompt, args.options);
     } finally {
@@ -127,7 +128,7 @@ Future<Map<String, dynamic>> runGenerateText(_GenerateArgs args) {
 /// Run a streaming `streamText` in a worker isolate, returning all parts.
 Future<List<Map<String, dynamic>>> runStreamText(_GenerateArgs args) {
   return Isolate.run(() async {
-    final model = Model.openai(args.apiKey, args.modelId, baseUrl: args.baseUrl);
+    final model = mockChatModel(args.baseUrl, args.apiKey, args.modelId);
     try {
       // streamText blocks synchronously until the stream completes, then
       // returns an already-closed stream containing the buffered parts.
@@ -217,7 +218,7 @@ Map<String, dynamic> _weatherTool() {
   return {
     'type': 'function',
     'name': 'get_weather',
-    'input_schema': {
+    'inputSchema': {
       'type': 'object',
       'properties': {
         'location': {'type': 'string'},
@@ -235,7 +236,7 @@ void main() {
   const apiKey = 'sk-test-mock-key';
   const modelId = 'gpt-4o';
 
-  test('generateText parses tool_calls', () async {
+  test('generateText parses toolCalls', () async {
     final server = await startMockServer([
       {'body': toolCallOpenAIResponse},
     ]);
@@ -255,18 +256,18 @@ void main() {
     expect(server.recorded.first.path, '/chat/completions');
 
     // Tool call parsed into the user-facing result.
-    expect(result, contains('tool_calls'));
-    final toolCalls = result['tool_calls'] as List;
+    expect(result, contains('toolCalls'));
+    final toolCalls = result['toolCalls'] as List;
     expect(toolCalls, hasLength(1));
-    expect(toolCalls[0]['tool_name'], 'get_weather');
-    expect(toolCalls[0]['tool_call_id'], 'call_abc');
+    expect(toolCalls[0]['toolName'], 'get_weather');
+    expect(toolCalls[0]['toolCallId'], 'call_abc');
     expect((toolCalls[0]['input'] as Map)['location'], 'Tokyo');
 
     // The raw provider result contains a ToolCall content variant.
     final raw = result['raw'] as Map<String, dynamic>;
     final content = raw['content'] as List;
     final hasToolCallVariant =
-        content.any((c) => (c as Map).containsKey('ToolCall'));
+        content.any((c) => (c as Map)['type'] == 'tool-call');
     expect(hasToolCallVariant, isTrue);
   });
 
@@ -317,7 +318,7 @@ void main() {
       'What is the weather in Tokyo?',
       {
         'tools': [_weatherTool()],
-        'tool_choice': 'required',
+        'toolChoice': 'required',
       },
     ));
 
@@ -350,18 +351,17 @@ void main() {
     // The stream delivered tool-call parts.
     expect(parts, isNotEmpty);
     final hasToolInputDelta =
-        parts.any((p) => (p as Map).containsKey('ToolInputDelta'));
+        parts.any((p) => (p as Map)['type'] == 'tool-input-delta');
     final hasToolCall =
-        parts.any((p) => (p as Map).containsKey('ToolCall'));
+        parts.any((p) => (p as Map)['type'] == 'tool-call');
     expect(hasToolInputDelta || hasToolCall, isTrue,
         reason: 'expected a tool-call stream part; got: $parts');
 
     // The assembled ToolCall should carry the get_weather tool + Tokyo input.
-    final toolCallPart = parts
-        .cast<Map<String, dynamic>?>()
-        .firstWhere((p) => p!.containsKey('ToolCall'));
-    final tc = toolCallPart!['ToolCall'] as Map<String, dynamic>;
-    expect(tc['tool_name'], 'get_weather');
+    final tc = parts
+        .cast<Map<String, dynamic>>()
+        .firstWhere((p) => p['type'] == 'tool-call');
+    expect(tc['toolName'], 'get_weather');
     expect((tc['input'] as Map)['location'], 'Tokyo');
   });
 
@@ -394,11 +394,11 @@ void main() {
       {'tools': [_weatherTool()]},
     ));
 
-    expect(firstResult, contains('tool_calls'));
-    final toolCalls = firstResult['tool_calls'] as List;
+    expect(firstResult, contains('toolCalls'));
+    final toolCalls = firstResult['toolCalls'] as List;
     expect(toolCalls, hasLength(1));
-    expect(toolCalls[0]['tool_name'], 'get_weather');
-    expect(toolCalls[0]['tool_call_id'], 'call_abc');
+    expect(toolCalls[0]['toolName'], 'get_weather');
+    expect(toolCalls[0]['toolCallId'], 'call_abc');
 
     // 2. Build the follow-up messages: original user question, the assistant's
     //    tool call, and the tool result we "executed" and are filling back in.
@@ -408,9 +408,9 @@ void main() {
         'role': 'assistant',
         'content': [
           {
-            'type': 'tool_call',
-            'tool_call_id': 'call_abc',
-            'tool_name': 'get_weather',
+            'type': 'tool-call',
+            'toolCallId': 'call_abc',
+            'toolName': 'get_weather',
             'input': {'location': 'Tokyo'},
           }
         ],
@@ -419,9 +419,13 @@ void main() {
         'role': 'tool',
         'content': [
           {
-            'type': 'tool_result',
-            'tool_call_id': 'call_abc',
-            'output': {'temperature': 22, 'condition': 'sunny'},
+            'type': 'tool-result',
+            'toolCallId': 'call_abc',
+            'toolName': 'get_weather',
+            'output': {
+              'type': 'json',
+              'value': {'temperature': 22, 'condition': 'sunny'},
+            },
           }
         ],
       },
diff --git a/bindings/flutter/test/tool_call_repair_typed_test.dart b/bindings/flutter/test/tool_call_repair_typed_test.dart
index 181cb5c5b..273e6438c 100644
--- a/bindings/flutter/test/tool_call_repair_typed_test.dart
+++ b/bindings/flutter/test/tool_call_repair_typed_test.dart
@@ -10,13 +10,12 @@
 import 'dart:convert';
 import 'dart:isolate';
 
-import 'package:aimux/aimux.dart';
 import 'package:aimux/typed_model.dart';
 import 'package:aimux/types.dart';
 import 'package:test/test.dart';
 
 import 'typed_model_test.dart'
-    show buildToolCallSse, startMockServer, toolCallOpenAIResponse;
+    show buildToolCallSse, mockChatModel, startMockServer, toolCallOpenAIResponse;
 
 /// Accepts `city` only, so the canned `{"location":"Tokyo"}` is invalid.
 GenerateTextOptions _strictOptions() => GenerateTextOptions(
@@ -48,7 +47,7 @@ void main() {
     final baseUrl = server.baseUrl;
 
     final result = await Isolate.run(() {
-      final typed = TypedModel(Model.openai('sk-test', 'gpt-4o', baseUrl: baseUrl));
+      final typed = TypedModel(mockChatModel(baseUrl, 'sk-test', 'gpt-4o'));
       try {
         return jsonDecode(jsonEncode(typed.generateText('weather?', _strictOptions())))
             as Map<String, dynamic>;
@@ -57,13 +56,13 @@ void main() {
       }
     });
 
-    final call = (result['tool_calls'] as List).single as Map<String, dynamic>;
+    final call = (result['toolCalls'] as List).single as Map<String, dynamic>;
     expect(call['invalid'], isNull);
     expect(call['input'], {'city': 'Tokyo'});
-    final part = (result['response_messages'] as List)
+    final part = (result['responseMessages'] as List)
         .expand((m) => (m as Map<String, dynamic>)['content'] as List)
         .whereType<Map<String, dynamic>>()
-        .singleWhere((p) => p['type'] == 'tool_call');
+        .singleWhere((p) => p['type'] == 'tool-call');
     expect(part['input'], {'city': 'Tokyo'});
   });
 
@@ -75,7 +74,7 @@ void main() {
     final baseUrl = server.baseUrl;
 
     final arguments = await Isolate.run(() {
-      final typed = TypedModel(Model.openai('sk-test', 'gpt-4o', baseUrl: baseUrl));
+      final typed = TypedModel(mockChatModel(baseUrl, 'sk-test', 'gpt-4o'));
       try {
         final completion = typed.generateTextAsOpenAI('weather?', _strictOptions());
         return completion.choices.first.message.toolCalls!.single.function.arguments;
@@ -95,10 +94,10 @@ void main() {
     final baseUrl = server.baseUrl;
 
     final input = await Isolate.run(() async {
-      final typed = TypedModel(Model.openai('sk-test', 'gpt-4o', baseUrl: baseUrl));
+      final typed = TypedModel(mockChatModel(baseUrl, 'sk-test', 'gpt-4o'));
       try {
         final parts = await typed.streamText('weather?', _strictOptions()).toList();
-        final call = parts.whereType<StreamPartToolCall>().single;
+        final call = parts.whereType<StreamPartToolCall>().single.call;
         return {'invalid': call.invalid, 'input': call.input};
       } finally {
         typed.close();
@@ -108,27 +107,4 @@ void main() {
     expect(input['invalid'], isNull);
     expect(input['input'], {'city': 'Tokyo'});
   });
-
-  test('RawToolCall keeps every provider field through a repair', () {
-    final call = RawToolCall.fromJson({
-      'tool_call_id': 'call-1',
-      'tool_name': 'weather',
-      'input': '{"town":"Tokyo"}',
-      'provider_executed': true,
-      'dynamic': true,
-      'provider_metadata': {
-        'google': {'x': 1},
-      },
-    });
-    expect(call.copyWith(input: '{"city":"Tokyo"}').toJson(), {
-      'tool_call_id': 'call-1',
-      'tool_name': 'weather',
-      'input': '{"city":"Tokyo"}',
-      'provider_executed': true,
-      'dynamic': true,
-      'provider_metadata': {
-        'google': {'x': 1},
-      },
-    });
-  });
 }
diff --git a/bindings/flutter/test/typed_model_test.dart b/bindings/flutter/test/typed_model_test.dart
index 838deb5f3..0c52a9f2c 100644
--- a/bindings/flutter/test/typed_model_test.dart
+++ b/bindings/flutter/test/typed_model_test.dart
@@ -37,6 +37,12 @@ import 'package:test/test.dart';
 // Mock server (mirrors structured_e2e_test.dart)
 // ─────────────────────────────────────────────────────────────────────────────
 
+/// A chat-completions model against [baseUrl]: the registry's `groq` entry with
+/// a `baseUrl` setting (`Model.openai` speaks the Responses API).
+Model mockChatModel(String baseUrl, String apiKey, String modelId) =>
+    Model.provider('groq', modelId,
+        apiKey: apiKey, configJson: jsonEncode({'baseUrl': baseUrl}));
+
 class MockOpenAIServer {
   final HttpServer _server;
   final List<Map<String, dynamic>> _responses;
@@ -134,7 +140,7 @@ class _TypedArgs {
 /// instances (not plain maps), which cannot be re-parsed by `fromJson`.
 Future<Map<String, dynamic>> runTypedGenerateText(_TypedArgs args) {
   return Isolate.run(() {
-    final model = Model.openai(args.apiKey, args.modelId, baseUrl: args.baseUrl);
+    final model = mockChatModel(args.baseUrl, args.apiKey, args.modelId);
     final typed = TypedModel(model);
     try {
       final hasOptions = args.tools != null ||
@@ -160,7 +166,7 @@ Future<Map<String, dynamic>> runTypedGenerateText(_TypedArgs args) {
 /// Run a typed `generateTextMessages` (multi-turn) in a worker isolate.
 Future<Map<String, dynamic>> runTypedGenerateTextMessages(_TypedArgs args) {
   return Isolate.run(() {
-    final model = Model.openai(args.apiKey, args.modelId, baseUrl: args.baseUrl);
+    final model = mockChatModel(args.baseUrl, args.apiKey, args.modelId);
     final typed = TypedModel(model);
     try {
       final messages = args.messages!
@@ -177,7 +183,7 @@ Future<Map<String, dynamic>> runTypedGenerateTextMessages(_TypedArgs args) {
 /// Run a typed `streamText` in a worker isolate, returning the parsed parts.
 Future<List<StreamPart>> runTypedStreamText(_TypedArgs args) {
   return Isolate.run(() async {
-    final model = Model.openai(args.apiKey, args.modelId, baseUrl: args.baseUrl);
+    final model = mockChatModel(args.baseUrl, args.apiKey, args.modelId);
     final typed = TypedModel(model);
     try {
       final options = (args.tools != null || args.includeRawChunks != null)
@@ -415,22 +421,21 @@ void main() {
     ));
 
     expect(parts, isNotEmpty);
-    // Every part is a single-key tagged union with a non-empty tag.
+    // Every part is a `type`-tagged union member the binding models.
     expect(parts.every((p) => p.type.isNotEmpty), isTrue);
+    expect(parts.whereType<StreamPartUnknown>(), isEmpty);
 
-    // The ToolCall part is a typed subclass with named fields.
+    // The tool-call part carries the parsed call.
     final toolCall = parts.whereType<StreamPartToolCall>().first;
-    expect(toolCall.type, 'ToolCall');
-    expect(toolCall.toolName, 'get_weather');
-    expect(toolCall.toolCallId, 'call_xyz');
-    expect((toolCall.input as Map<String, dynamic>)['location'], 'Tokyo');
+    expect(toolCall.type, 'tool-call');
+    expect(toolCall.call.toolName, 'get_weather');
+    expect(toolCall.call.toolCallId, 'call_xyz');
+    expect((toolCall.call.input as Map<String, dynamic>)['location'], 'Tokyo');
 
     // A Finish part carries typed usage + finish reason.
     final finish = parts.whereType<StreamPartFinish>().first;
-    expect(finish.type, 'Finish');
+    expect(finish.type, 'finish');
     expect(finish.usage, isA<Usage>());
-    expect(finish.usage.inputTokens.total, 5);
-    expect(finish.usage.outputTokens.total, 2);
 
     // The request was a streaming POST.
     expect(server.recorded, hasLength(1));
@@ -439,7 +444,7 @@ void main() {
   });
 
   test('streamText emits Raw parts when includeRawChunks enabled', () async {
-    // RFC-0016 M2: include_raw_chunks surfaces one Raw part per JSON SSE
+    // RFC-0016 M2: includeRawChunks surfaces one Raw part per JSON SSE
     // event (parsed payload), before the parsed parts; [DONE] excluded.
     final server = await startMockServer([
       {'sse': true, 'body': buildTextSse()},
@@ -476,8 +481,8 @@ void main() {
     expect(text, 'Hello world');
   });
 
-  test('streamText parses non-empty StreamStart warnings', () async {
-    // RFC-0016 M9: StreamStart carries a warnings array; a non-empty array
+  test('streamText decodes the stream-start part', () async {
+    // RFC-0016 M9: stream-start carries a warnings array; a non-empty array
     // (e.g. from a provider that rejects top_k) must decode without breaking
     // the stream. The openai full profile produces no warning, so the
     // assertion is decode-ability + field presence (the warning itself is
@@ -495,7 +500,7 @@ void main() {
     ));
 
     final start = parts.whereType<StreamPartStreamStart>().first;
-    expect(start.type, 'StreamStart');
+    expect(start.type, 'stream-start');
     expect(start.warnings, isA<List<dynamic>>());
   });
 }
diff --git a/bindings/flutter/test/typed_round_trip_test.dart b/bindings/flutter/test/typed_round_trip_test.dart
index 807b2b39d..558ab2031 100644
--- a/bindings/flutter/test/typed_round_trip_test.dart
+++ b/bindings/flutter/test/typed_round_trip_test.dart
@@ -1,458 +1,270 @@
-// Round-trip serialization tests for the typed `GenerateContent` /
-// `StreamPart` sealed class hierarchies in `types.dart`.
-//
-// These tests are pure Dart — no FFI, no Rust core, no mock server. They
-// verify that every known variant (and the `Unknown` fallback) survives a
-// `toJson()` -> `fromJson()` round-trip with all fields intact, and that the
-// externally-tagged wire shape matches the Rust core contract:
-//
-//   - `File` variants serialize `data` + `media_type` and **never** `filename`.
-//   - `ToolResult` variants use the `result` key, **not** `output`.
-//   - `Unknown` variants pass the tag + payload through verbatim.
-//
-// Run:
-//   export PATH="/home/eric8810/.local/dart-sdk/bin:$PATH"
-//   cd bindings/flutter && dart test test/typed_round_trip_test.dart
-
+// Pure-Dart round trips for the typed `GenerateContent` / `StreamPart` unions,
+// tool calls, options and results against the generated wire shapes in
+// `bindings/node/src/types/`: camelCase fields, `type`-tagged unions with
+// kebab-case values, optional fields absent (never null).
 import 'dart:convert';
 
 import 'package:aimux/types.dart';
 import 'package:test/test.dart';
 
-/// Deep-flatten a `toJson()` result across the JSON boundary.
-///
-/// Mirrors the workaround used in `typed_model_test.dart`: `Usage` /
-/// `TokenUsage` are `@JsonSerializable()` with the default
-/// `explicitToJson: false`, so their generated `toJson()` keeps nested
-/// `TokenUsage` instances instead of recursing to plain maps. The only variant
-/// affected is `StreamPartFinish` (the only one nesting `@JsonSerializable`
-/// types). In the real pipeline the value always crosses a JSON-string
-/// boundary (Rust core → JSON → Dart), which flattens it — so we do the same
-/// here before re-parsing.
-Map<String, dynamic> deepFlatten(Map<String, dynamic> json) =>
-    jsonDecode(jsonEncode(json)) as Map<String, dynamic>;
+const _meta = {
+  'openai': {'itemId': 'item_1'}
+};
+const _file = {
+  'data': {'type': 'data', 'data': 'aGk='},
+  'mediaType': 'image/png',
+};
+const _usage = {
+  'inputTokens': {'total': 5, 'noCache': 5},
+  'outputTokens': {'total': 2},
+};
+const _finish = {'unified': 'stop', 'raw': 'stop'};
 
 void main() {
-  group('ToolCall round-trip', () {
-    test('preserves provider_metadata', () {
-      final original = ToolCall(
-        toolCallId: 'call_1',
-        toolName: 'get_weather',
-        input: {'city': 'Paris'},
-        providerMetadata: {
-          'openai': {'itemId': 'item_1'},
-        },
-      );
-      final json = original.toJson();
-      expect(json['provider_metadata'], {
-        'openai': {'itemId': 'item_1'},
-      });
-
-      final decoded = ToolCall.fromJson(json);
-      expect(decoded.providerMetadata, original.providerMetadata);
+  group('ToolCall', () {
+    test('round-trips and omits absent optionals', () {
+      final json = {
+        'toolCallId': 'call_1',
+        'toolName': 'get_weather',
+        'input': {'city': 'Paris'},
+        'providerExecuted': true,
+        'dynamic': false,
+        'providerMetadata': _meta,
+        'invalid': true,
+        'error': {'name': 'AI_NoSuchToolError', 'toolName': 'x'},
+      };
+      expect(ToolCall.fromJson(json).toJson(), json);
+
+      final minimal = ToolCall(toolCallId: 'c', toolName: 't', input: const {});
+      expect(minimal.toJson().keys, ['toolCallId', 'toolName', 'input']);
     });
 
-    test('provider_metadata accepts arbitrary JSON and omits null', () {
-      final scalar = ToolCall(
-        toolCallId: 'call_scalar',
-        toolName: 'tool',
-        input: const {},
-        providerMetadata: 'opaque-provider-token',
-      );
-      expect(scalar.toJson()['provider_metadata'], 'opaque-provider-token');
-      expect(ToolCall.fromJson(scalar.toJson()).providerMetadata,
-          'opaque-provider-token');
-
-      final absent = ToolCall(
-        toolCallId: 'call_absent',
-        toolName: 'tool',
-        input: const {},
-      );
-      expect(absent.toJson(), isNot(contains('provider_metadata')));
+    test('RawToolCall keeps every provider field through copyWith', () {
+      final call = RawToolCall.fromJson({
+        'toolCallId': 'call-1',
+        'toolName': 'weather',
+        'input': '{"town":"Tokyo"}',
+        'providerExecuted': true,
+        'dynamic': true,
+        'providerMetadata': _meta,
+      });
+      expect(call.copyWith(input: '{"city":"Tokyo"}').toJson(), {
+        'toolCallId': 'call-1',
+        'toolName': 'weather',
+        'input': '{"city":"Tokyo"}',
+        'providerExecuted': true,
+        'dynamic': true,
+        'providerMetadata': _meta,
+      });
     });
   });
 
-  // ─────────────────────────────────────────────────────────────────────────
-  // GenerateContent (6 variants + Unknown)
-  // ─────────────────────────────────────────────────────────────────────────
-  group('GenerateContent round-trip', () {
-    test('Text variant', () {
-      final original =
-          GenerateContentText(text: 'hello', providerMetadata: null);
-      final decoded = GenerateContent.fromJson(original.toJson());
-
-      expect(decoded, isA<GenerateContentText>());
-      expect(decoded.tag, 'Text');
-      final t = decoded as GenerateContentText;
-      expect(t.text, 'hello');
-      expect(t.providerMetadata, isNull);
-    });
-
-    test('Text variant preserves provider_metadata', () {
-      final original = GenerateContentText(
-        text: 'hi',
-        providerMetadata: {'model': 'gpt-4o'},
-      );
-      final decoded = GenerateContent.fromJson(original.toJson());
-
-      final t = decoded as GenerateContentText;
-      expect(t.providerMetadata, {'model': 'gpt-4o'});
-    });
-
-    test('ToolCall variant', () {
-      final original = GenerateContentToolCall(
-        toolCallId: 'call_1',
-        toolName: 'get_weather',
-        input: {'location': 'Tokyo'},
-        providerExecuted: true,
-        isDynamic: false,
-        providerMetadata: null,
-      );
-      final decoded = GenerateContent.fromJson(original.toJson());
-
-      expect(decoded, isA<GenerateContentToolCall>());
-      expect(decoded.tag, 'ToolCall');
-      final tc = decoded as GenerateContentToolCall;
-      expect(tc.toolCallId, 'call_1');
-      expect(tc.toolName, 'get_weather');
-      expect((tc.input as Map<String, dynamic>)['location'], 'Tokyo');
-      expect(tc.providerExecuted, true);
-      expect(tc.isDynamic, false);
-    });
-
-    test('Reasoning variant', () {
-      final original =
-          GenerateContentReasoning(text: 'thinking...', providerMetadata: null);
-      final decoded = GenerateContent.fromJson(original.toJson());
-
-      expect(decoded, isA<GenerateContentReasoning>());
-      expect(decoded.tag, 'Reasoning');
-      final r = decoded as GenerateContentReasoning;
-      expect(r.text, 'thinking...');
-    });
-
-
-
-    test('ToolResult variant — uses result, not output', () {
-      final original = GenerateContentToolResult(
-        toolCallId: 'call_1',
-        toolName: 'web_search',
-        result: {'count': 3, 'hits': ['a', 'b', 'c']},
-        isError: false,
-        preliminary: true,
-        isDynamic: false,
-        providerMetadata: null,
-      );
-      final json = original.toJson();
-
-      expect(json.keys, ['ToolResult']);
-      final payload = json['ToolResult'] as Map<String, dynamic>;
-
-      // Contract: `result` key, never `output`.
-      expect(payload, contains('result'));
-      expect(payload, isNot(contains('output')));
-      expect(payload, contains('tool_call_id'));
-      expect(payload, contains('tool_name'));
-      expect(payload, contains('is_error'));
-      expect(payload, contains('preliminary'));
-      expect(payload, contains('dynamic'));
-
-      final decoded = GenerateContent.fromJson(json);
-      expect(decoded, isA<GenerateContentToolResult>());
-      expect(decoded.tag, 'ToolResult');
-      final tr = decoded as GenerateContentToolResult;
-      expect(tr.toolCallId, 'call_1');
-      expect(tr.toolName, 'web_search');
-      expect((tr.result as Map<String, dynamic>)['count'], 3);
-      expect(tr.isError, false);
-      expect(tr.preliminary, true);
-      expect(tr.isDynamic, false);
-    });
-
-    test('Unknown variant — tag + data pass through verbatim', () {
-      final original = GenerateContentUnknown(
-        tag: 'FutureVariant',
-        data: {'foo': 'bar', 'n': 42},
-      );
-      final json = original.toJson();
-
-      // Unknown re-encodes as `{tag: data}`.
-      expect(json.keys, ['FutureVariant']);
-      expect(json['FutureVariant'], {'foo': 'bar', 'n': 42});
-
-      final decoded = GenerateContent.fromJson(json);
-      expect(decoded, isA<GenerateContentUnknown>());
-      final u = decoded as GenerateContentUnknown;
-      expect(u.tag, 'FutureVariant');
-      expect(u.data, {'foo': 'bar', 'n': 42});
+  group('GenerateContent', () {
+    final cases = <Map<String, dynamic>>[
+      {'type': 'text', 'text': 'hi', 'providerMetadata': _meta},
+      {
+        'type': 'tool-call',
+        'toolCallId': 'c1',
+        'toolName': 'w',
+        'input': '{"a":1}',
+        'dynamic': true,
+      },
+      {'type': 'source', 'sourceType': 'url', 'id': 's', 'url': 'https://x'},
+      {
+        'type': 'source',
+        'sourceType': 'document',
+        'id': 's',
+        'mediaType': 'application/pdf',
+        'title': 'T',
+        'filename': 'f.pdf',
+      },
+      {'type': 'reasoning', 'text': 'hmm'},
+      {'type': 'file', ..._file},
+      {'type': 'reasoning-file', ..._file},
+      {'type': 'custom', 'kind': 'k'},
+      {'type': 'tool-approval-request', 'approvalId': 'a', 'toolCallId': 'c'},
+      {
+        'type': 'tool-result',
+        'toolCallId': 'c1',
+        'toolName': 'web_search',
+        'result': {'count': 3},
+        'isError': false,
+        'preliminary': true,
+        'dynamic': false,
+      },
+    ];
+    for (final json in cases) {
+      test(json['type'] as String, () {
+        final c = GenerateContent.fromJson(json);
+        expect(c, isNot(isA<GenerateContentUnknown>()));
+        expect(c.tag, json['type']);
+        expect(c.toJson(), json);
+      });
+    }
+
+    test('unknown type passes through verbatim', () {
+      final json = {'type': 'future', 'n': 42};
+      final c = GenerateContent.fromJson(json);
+      expect(c, isA<GenerateContentUnknown>());
+      expect(c.tag, 'future');
+      expect(c.toJson(), json);
     });
   });
 
-  // ─────────────────────────────────────────────────────────────────────────
-  // StreamPart (18 variants + Unknown)
-  // ─────────────────────────────────────────────────────────────────────────
-  group('StreamPart round-trip', () {
-    // ── Text variants ──
-    test('TextStart variant', () {
-      final original = StreamPartTextStart(id: 't1', providerMetadata: null);
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartTextStart>());
-      expect(decoded.type, 'TextStart');
-      expect((decoded as StreamPartTextStart).id, 't1');
-    });
-
-    test('TextDelta variant', () {
-      final original =
-          StreamPartTextDelta(id: 't1', delta: 'hel', providerMetadata: null);
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartTextDelta>());
-      expect(decoded.type, 'TextDelta');
-      final d = decoded as StreamPartTextDelta;
-      expect(d.id, 't1');
-      expect(d.delta, 'hel');
-    });
-
-    test('TextEnd variant', () {
-      final original = StreamPartTextEnd(id: 't1', providerMetadata: null);
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartTextEnd>());
-      expect(decoded.type, 'TextEnd');
-      expect((decoded as StreamPartTextEnd).id, 't1');
-    });
-
-    // ── Stream lifecycle ──
-    test('StreamStart variant', () {
-      final original = StreamPartStreamStart(warnings: [
-        {'code': 'deprecation', 'message': 'soon'},
-      ]);
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartStreamStart>());
-      expect(decoded.type, 'StreamStart');
-      final s = decoded as StreamPartStreamStart;
-      expect(s.warnings, hasLength(1));
-      expect(s.warnings.first['code'], 'deprecation');
-    });
-
-
-
-    test('Error variant', () {
-      final original = StreamPartError(error: {'message': 'boom', 'code': 500});
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartError>());
-      expect(decoded.type, 'Error');
-      final e = decoded as StreamPartError;
-      expect((e.error as Map<String, dynamic>)['message'], 'boom');
-    });
-
-    // ── Tool calls ──
-    test('ToolInputStart variant — has title field', () {
-      final original = StreamPartToolInputStart(
-        id: 'tc1',
-        toolName: 'get_weather',
-        providerExecuted: true,
-        isDynamic: false,
-        title: 'Weather lookup',
-        providerMetadata: null,
-      );
-      final json = original.toJson();
-
-      final payload = json['ToolInputStart'] as Map<String, dynamic>;
-      expect(payload, contains('title'));
-      expect(payload['title'], 'Weather lookup');
-
-      final decoded = StreamPart.fromJson(json);
-      expect(decoded, isA<StreamPartToolInputStart>());
-      expect(decoded.type, 'ToolInputStart');
-      final t = decoded as StreamPartToolInputStart;
-      expect(t.id, 'tc1');
-      expect(t.toolName, 'get_weather');
-      expect(t.title, 'Weather lookup');
-      expect(t.providerExecuted, true);
-      expect(t.isDynamic, false);
-    });
-
-    test('ToolInputDelta variant', () {
-      final original =
-          StreamPartToolInputDelta(id: 'tc1', delta: '{"loc', providerMetadata: null);
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartToolInputDelta>());
-      expect(decoded.type, 'ToolInputDelta');
-      final d = decoded as StreamPartToolInputDelta;
-      expect(d.id, 'tc1');
-      expect(d.delta, '{"loc');
-    });
-
-    test('ToolInputEnd variant', () {
-      final original =
-          StreamPartToolInputEnd(id: 'tc1', providerMetadata: null);
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartToolInputEnd>());
-      expect(decoded.type, 'ToolInputEnd');
-      expect((decoded as StreamPartToolInputEnd).id, 'tc1');
-    });
-
-    test('ToolCall variant — has provider_metadata field', () {
-      final original = StreamPartToolCall(
-        toolCallId: 'call_1',
-        toolName: 'get_weather',
-        input: {'location': 'Tokyo'},
-        providerExecuted: true,
-        isDynamic: false,
-        providerMetadata: {'provider': 'openai'},
-      );
-      final json = original.toJson();
-
-      final payload = json['ToolCall'] as Map<String, dynamic>;
-      expect(payload, contains('provider_metadata'));
-      expect(payload['provider_metadata'], {'provider': 'openai'});
-
-      final decoded = StreamPart.fromJson(json);
-      expect(decoded, isA<StreamPartToolCall>());
-      expect(decoded.type, 'ToolCall');
-      final tc = decoded as StreamPartToolCall;
-      expect(tc.toolCallId, 'call_1');
-      expect(tc.toolName, 'get_weather');
-      expect((tc.input as Map<String, dynamic>)['location'], 'Tokyo');
-      expect(tc.providerMetadata, {'provider': 'openai'});
-    });
-
-    test('ToolResult variant — uses result, not output', () {
-      final original = StreamPartToolResult(
-        toolCallId: 'call_1',
-        toolName: 'web_search',
-        result: {'count': 3},
-        isError: false,
-        preliminary: true,
-        isDynamic: false,
-        providerMetadata: null,
-      );
-      final json = original.toJson();
-
-      final payload = json['ToolResult'] as Map<String, dynamic>;
-      expect(payload, contains('result'));
-      expect(payload, isNot(contains('output')));
-
-      final decoded = StreamPart.fromJson(json);
-      expect(decoded, isA<StreamPartToolResult>());
-      expect(decoded.type, 'ToolResult');
-      final tr = decoded as StreamPartToolResult;
-      expect(tr.toolCallId, 'call_1');
-      expect(tr.toolName, 'web_search');
-      expect((tr.result as Map<String, dynamic>)['count'], 3);
-      expect(tr.isError, false);
-      expect(tr.preliminary, true);
-      expect(tr.isDynamic, false);
-    });
-
-    // ── File ──
-
-
-    // ── Reasoning ──
-    test('ReasoningStart variant', () {
-      final original =
-          StreamPartReasoningStart(id: 'r1', providerMetadata: null);
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartReasoningStart>());
-      expect(decoded.type, 'ReasoningStart');
-      expect((decoded as StreamPartReasoningStart).id, 'r1');
-    });
-
-    test('ReasoningDelta variant', () {
-      final original =
-          StreamPartReasoningDelta(id: 'r1', delta: 'thi', providerMetadata: null);
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartReasoningDelta>());
-      expect(decoded.type, 'ReasoningDelta');
-      final d = decoded as StreamPartReasoningDelta;
-      expect(d.id, 'r1');
-      expect(d.delta, 'thi');
-    });
-
-    test('ReasoningEnd variant', () {
-      final original =
-          StreamPartReasoningEnd(id: 'r1', providerMetadata: null);
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartReasoningEnd>());
-      expect(decoded.type, 'ReasoningEnd');
-      expect((decoded as StreamPartReasoningEnd).id, 'r1');
+  group('StreamPart', () {
+    final call = {
+      'toolCallId': 'c1',
+      'toolName': 'w',
+      'input': {'a': 1},
+      'invalid': true,
+    };
+    final cases = <Map<String, dynamic>>[
+      {'type': 'text-start', 'id': 't'},
+      {'type': 'text-delta', 'id': 't', 'delta': 'hel', 'providerMetadata': _meta},
+      {'type': 'text-end', 'id': 't'},
+      {
+        'type': 'stream-start',
+        'warnings': [
+          {'type': 'other', 'message': 'm'}
+        ]
+      },
+      {'type': 'finish', 'finishReason': _finish, 'usage': _usage},
+      {
+        'type': 'finish-step',
+        'finishReason': _finish,
+        'usage': _usage,
+        'response': {'id': 'r', 'modelId': 'gpt-4o'},
+      },
+      {
+        'type': 'error',
+        'error': {'name': 'AI_APICallError', 'message': 'boom', 'isRetryable': false},
+      },
+      {
+        'type': 'tool-input-start',
+        'id': 'tc',
+        'toolName': 'w',
+        'providerExecuted': true,
+        'dynamic': false,
+        'title': 'Weather',
+      },
+      {'type': 'tool-input-delta', 'id': 'tc', 'delta': '{"lo'},
+      {'type': 'tool-input-end', 'id': 'tc'},
+      {'type': 'tool-call', ...call},
+      {
+        'type': 'tool-result',
+        'toolCallId': 'c1',
+        'toolName': 'w',
+        'result': {'count': 3},
+      },
+      {'type': 'file', ..._file},
+      {'type': 'reasoning-file', 'file': _file},
+      {'type': 'custom', 'kind': 'k'},
+      {
+        'type': 'tool-approval-request',
+        'approvalId': 'a',
+        'toolCall': call,
+        'reason': 'r',
+      },
+      {'type': 'reasoning-start', 'id': 'r1'},
+      {'type': 'reasoning-delta', 'id': 'r1', 'delta': 'thi'},
+      {'type': 'reasoning-end', 'id': 'r1'},
+      {'type': 'source', 'sourceType': 'url', 'id': 's', 'url': 'https://x'},
+      {'type': 'raw', 'rawValue': {'raw': 'chunk'}},
+    ];
+    for (final json in cases) {
+      test(json['type'] as String, () {
+        final p = StreamPart.fromJson(json);
+        expect(p, isNot(isA<StreamPartUnknown>()));
+        expect(p.type, json['type']);
+        expect(p.toJson(), json);
+      });
+    }
+
+    test('unknown type passes through verbatim', () {
+      final json = {'type': 'future-part', 'n': 7};
+      final p = StreamPart.fromJson(json);
+      expect(p, isA<StreamPartUnknown>());
+      expect(p.type, 'future-part');
+      expect(p.toJson(), json);
     });
+  });
 
-    // ── Metadata / sources / raw ──
-    test('ResponseMetadata variant', () {
-      final original = StreamPartResponseMetadata(
-        id: 'resp_1',
-        timestamp: '2024-01-01T00:00:00Z',
-        modelId: 'gpt-4o',
+  group('options and results', () {
+    test('GenerateTextOptions uses camelCase keys and omits unset fields', () {
+      final options = GenerateTextOptions(
+        maxOutputTokens: 10,
+        stopSequences: ['x'],
+        toolChoice: ToolChoice.tool('w'),
+        timeout: TimeoutConfiguration(totalMs: 1000),
+        includeRawChunks: true,
+        sessionId: 'sess-1',
+        tools: [
+          Tool.function(FunctionTool(name: 'w', inputSchema: {'type': 'object'})),
+        ],
       );
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartResponseMetadata>());
-      expect(decoded.type, 'ResponseMetadata');
-      final m = decoded as StreamPartResponseMetadata;
-      expect(m.id, 'resp_1');
-      expect(m.timestamp, '2024-01-01T00:00:00Z');
-      expect(m.modelId, 'gpt-4o');
+      final json = options.toJson();
+      expect(json, {
+        'maxOutputTokens': 10,
+        'stopSequences': ['x'],
+        'toolChoice': {'type': 'tool', 'toolName': 'w'},
+        'timeout': {'totalMs': 1000},
+        'includeRawChunks': true,
+        'sessionId': 'sess-1',
+        'tools': [
+          {'type': 'function', 'name': 'w', 'inputSchema': {'type': 'object'}}
+        ],
+      });
+      expect(GenerateTextOptions.fromJson(json).toJson(), json);
+      expect(GenerateTextOptions().toJson(), isEmpty);
     });
 
-    test('Raw variant', () {
-      final original = StreamPartRaw(rawValue: {'raw': 'chunk'});
-      final decoded = StreamPart.fromJson(original.toJson());
-
-      expect(decoded, isA<StreamPartRaw>());
-      expect(decoded.type, 'Raw');
-      final r = decoded as StreamPartRaw;
-      expect((r.rawValue as Map<String, dynamic>)['raw'], 'chunk');
+    test('provider tool', () {
+      final json = {
+        'type': 'provider',
+        'id': 'anthropic.web_search_20250305',
+        'name': 'web_search',
+        'args': {'maxUses': 1},
+      };
+      expect(Tool.fromJson(json).toJson(), json);
     });
 
-    test('Unknown variant — tag + data pass through verbatim', () {
-      final original =
-          StreamPartUnknown(tag: 'FuturePart', data: {'foo': 'bar', 'n': 7});
-      final json = original.toJson();
-
-      expect(json.keys, ['FuturePart']);
-      expect(json['FuturePart'], {'foo': 'bar', 'n': 7});
-
-      final decoded = StreamPart.fromJson(json);
-      expect(decoded, isA<StreamPartUnknown>());
-      expect(decoded.type, 'FuturePart');
-      final u = decoded as StreamPartUnknown;
-      expect(u.tag, 'FuturePart');
-      expect(u.data, {'foo': 'bar', 'n': 7});
+    test('GenerateTextResult round-trips', () {
+      final json = jsonDecode(jsonEncode({
+        'content': [
+          {'type': 'text', 'text': 'hi'}
+        ],
+        'text': 'hi',
+        'toolCalls': [
+          {'toolCallId': 'c', 'toolName': 't', 'input': {}}
+        ],
+        'finishReason': _finish,
+        'usage': _usage,
+        'warnings': [],
+        'raw': {
+          'content': [
+            {'type': 'text', 'text': 'hi'}
+          ],
+          'finishReason': _finish,
+          'usage': _usage,
+          'warnings': [],
+        },
+        'reasoning': [],
+        'reasoningText': '',
+        'sources': [],
+        'files': [],
+        'responseMessages': [
+          {'role': 'assistant', 'content': 'hi'}
+        ],
+        'rawFinishReason': 'stop',
+        'request': {},
+        'response': {'id': 'r'},
+        'totalUsage': _usage,
+      })) as Map<String, dynamic>;
+      final result = GenerateTextResult.fromJson(json);
+      expect(result.finishReason.unified, 'stop');
+      expect(result.totalUsage.inputTokens.total, 5);
+      expect(result.raw.content.single, isA<GenerateContentText>());
+      expect(jsonDecode(jsonEncode(result)), json);
     });
   });
-
-  test('GenerateTextOptions includeRawChunks round-trips', () {
-    // RFC-0016 M2 true-case through the Dart typed options.
-    final original = GenerateTextOptions(includeRawChunks: true);
-    final json = original.toJson();
-    expect(json['include_raw_chunks'], isTrue);
-    final back = GenerateTextOptions.fromJson(json);
-    expect(back.includeRawChunks, isTrue);
-
-    // Default omits the field (toJson only emits non-null).
-    final defaults = GenerateTextOptions().toJson();
-    expect(defaults.containsKey('include_raw_chunks'), isFalse);
-  });
-
-  test('GenerateTextOptions sessionId round-trips', () {
-    // RFC-0024: explicit session_id crosses the wire as snake_case.
-    final original = GenerateTextOptions(sessionId: 'sess-1');
-    final json = original.toJson();
-    expect(json['session_id'], 'sess-1');
-    final back = GenerateTextOptions.fromJson(json);
-    expect(back.sessionId, 'sess-1');
-
-    final defaults = GenerateTextOptions().toJson();
-    expect(defaults.containsKey('session_id'), isFalse);
-  });
 }
diff --git a/bindings/go/README.md b/bindings/go/README.md
index 56a7996c7..7db593930 100644
--- a/bindings/go/README.md
+++ b/bindings/go/README.md
@@ -43,14 +43,20 @@ func main() {
     }
     defer p.Close()
 
-    text, err := p.GenerateText(`"Explain Rust ownership in one sentence."`, "")
+    result, err := p.Generate("Explain Rust ownership in one sentence.", nil)
     if err != nil {
         log.Fatal(err)
     }
-    fmt.Println(text)
+    fmt.Println(result.Text)
 }
 ```
 
+`GenerateText` / `StreamText` take and return raw JSON strings; `Generate` /
+`Stream` are the typed wrappers. The JSON is the AI SDK JSON: camelCase field
+names, unions tagged by `type` (`"tool-call"`, `"text-delta"`, ...), optional
+fields absent. A tool result is a `tool-result` content part
+(`toolCallId`, `toolName`, `output`). See [docs/api/go.md](../../docs/api/go.md).
+
 ## Features
 
 - Typed API for text, streaming, embeddings, tools, images, speech, and more
diff --git a/bindings/go/aimux.go b/bindings/go/aimux.go
index 3fb7f88cd..afbc66119 100644
--- a/bindings/go/aimux.go
+++ b/bindings/go/aimux.go
@@ -352,18 +352,6 @@ func MockReplay(recordingsJsonl string) (*Model, error) {
 	return wrapHandle(&h, C.aimux_mock_replay_new(cJsonl, &h))
 }
 
-// RegisterProviders registers external OpenAI-compatible providers from a JSON
-// config string (RFC-0020). Entries override same-named built-ins or add new
-// ones. configJSON shape: { "providers": [ { "name", "base_url", ... } ] }.
-func RegisterProviders(configJSON string) error {
-	if err := requireJSON("config_json", configJSON); err != nil {
-		return err
-	}
-	cJSON := C.CString(configJSON)
-	defer C.free(unsafe.Pointer(cJSON))
-	return expectAimuxError(C.aimux_register_providers(cJSON))
-}
-
 // InitProxy sets the global proxy configuration (M6, RFC-0016). Must be called
 // before the first GenerateText / StreamText call; a no-op if the shared HTTP
 // client is already initialised. configJSON is required (pass "{}" for
@@ -384,7 +372,7 @@ func InitProxy(configJSON string) error {
 // on error (per configJSON). models must be non-empty.
 //
 // configJSON (optional): {"router": "rule"|"weighted", "weights": [...],
-// "fallback": "on_error"|"none", "provider_name", "model_id"}.
+// "fallback": "on_error"|"none", "providerName", "modelId"}.
 //
 // The same model may appear several times. A nil entry returns an error
 // ("aimux: router: models[i] is nil"), never a panic. Concurrent Close is
@@ -661,13 +649,13 @@ func ProviderWithBase(name, apiKey, modelID, baseURL string) (*Model, error) {
 	return ProviderWithConfig(name, apiKey, modelID, &ProviderConfig{BaseURL: baseURL})
 }
 
-// ProviderConfig mirrors the Rust ProviderOptions accepted by
-// aimux_provider_new's config_json (RFC-0017 §3.4). Zero-value fields are
-// omitted. Retry is a per-call setting (CallOptions.MaxRetries), and request-body
-// overrides no longer exist: a config_json carrying max_retries or
-// body_overrides is rejected as CodeInvalidArgument, so this struct has neither.
+// ProviderConfig is the config_json accepted by aimux_provider_new and
+// aimux_provider_handle_new (RFC-0017 §3.4). Zero-value fields are omitted.
+// Retry is a per-call setting (GenerateTextOptions.MaxRetries), and request-body
+// overrides no longer exist: any other key in config_json is rejected as
+// CodeInvalidArgument, so this struct has neither.
 type ProviderConfig struct {
-	BaseURL      string            `json:"base_url,omitempty"`
+	BaseURL      string            `json:"baseUrl,omitempty"`
 	Headers      map[string]string `json:"headers,omitempty"`
 	Organization string            `json:"organization,omitempty"`
 	Project      string            `json:"project,omitempty"`
@@ -861,8 +849,8 @@ func (m *Model) GenerateText(promptJson, optsJson string) (string, error) {
 // GenerateObject generates a structured JSON object (M12, RFC-0016).
 //
 // Same signature as GenerateText; returns the JSON-serialized
-// GenerateObjectResult. Pass response_format: { "Json": { ... } } via
-// optsJson for schema control; the function applies JSON repair before
+// GenerateObjectResult. Pass responseFormat: {"type":"json","schema":{...}}
+// via optsJson for schema control; the function applies JSON repair before
 // parsing.
 func (m *Model) GenerateObject(promptJson, optsJson string) (string, error) {
 	if err := checkPromptOpts(promptJson, optsJson); err != nil {
@@ -1050,8 +1038,8 @@ func (m *Model) StreamTextContext(ctx context.Context, promptJson, optsJson stri
 // ChatCompletionChunk (OpenAI "chat.completion.chunk" object). Works with any
 // provider.
 //
-// Opts may carry providerOptions.openai.stream_options with include_usage
-// (bool, default true) and include_reasoning (bool, default true).
+// Opts may carry providerOptions.openai.streamOptions with includeUsage
+// (bool, default true) and includeReasoning (bool, default true).
 func (m *Model) StreamTextAsOpenAI(promptJson, optsJson string) *Stream {
 	return m.StreamTextAsOpenAIContext(context.Background(), promptJson, optsJson)
 }
diff --git a/bindings/go/aimux_test.go b/bindings/go/aimux_test.go
index 804941e37..87fb53493 100644
--- a/bindings/go/aimux_test.go
+++ b/bindings/go/aimux_test.go
@@ -242,13 +242,19 @@ func TestProviderWithConfigParams(t *testing.T) {
 }
 
 // A registry provider with no key and no env var fails with CodeLoadAPIKey and
-// names the variable it consulted.
+// names the variable it consulted. The key is read when a call is made (as in
+// the AI SDK), not when the model is built.
 func TestMissingAPIKeyIsLoadAPIKey(t *testing.T) {
 	if old, ok := os.LookupEnv("GROQ_API_KEY"); ok {
 		os.Unsetenv("GROQ_API_KEY")
 		defer os.Setenv("GROQ_API_KEY", old)
 	}
-	_, err := Provider("groq", "", "llama-3.3-70b")
+	m, err := Provider("groq", "", "llama-3.3-70b")
+	if err != nil {
+		t.Fatalf("Provider: %v", err)
+	}
+	defer m.Close()
+	_, err = m.GenerateText(`"hi"`, "")
 	var e *Error
 	if !errors.As(err, &e) || e.Code != CodeLoadAPIKey {
 		t.Fatalf("want CodeLoadAPIKey, got %v", err)
diff --git a/bindings/go/batch1_e2e_test.go b/bindings/go/batch1_e2e_test.go
index 8cc575733..95ec964f7 100644
--- a/bindings/go/batch1_e2e_test.go
+++ b/bindings/go/batch1_e2e_test.go
@@ -1,7 +1,7 @@
 package aimux
 
 // RFC-0016 第一批 e2e tests (Go shell):
-//   M2 include_raw_chunks -> StreamPart.Raw emission
+//   M2 includeRawChunks -> "raw" stream part emission
 //   M10 Usage.raw preservation
 //   M9 StreamStart non-empty warnings
 //
@@ -25,7 +25,7 @@ func TestE2E_StreamTextEmitsRawWhenEnabled(t *testing.T) {
 	srv.SetContentType("text/event-stream")
 	srv.SetResponse(buildTextDeltaSSE())
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	stream := m.StreamText(`"Say hello"`, mustMarshalOptions(t, &GenerateTextOptions{
@@ -40,15 +40,15 @@ func TestE2E_StreamTextEmitsRawWhenEnabled(t *testing.T) {
 		if err != nil {
 			t.Fatalf("failed to parse: %v", err)
 		}
-		switch sp.Tag {
-		case "Raw":
-			raws = append(raws, sp.Payload)
+		switch sp.Type {
+		case "raw":
+			raws = append(raws, sp.Raw)
 			if !rawBeforeFirstDelta {
 				rawBeforeFirstDelta = true
 			}
-		case "TextDelta":
+		case "text-delta":
 			var td TextDeltaPayload
-			json.Unmarshal(sp.Payload, &td)
+			json.Unmarshal(sp.Raw, &td)
 			textBuilder.WriteString(td.Delta)
 		}
 	}
@@ -65,10 +65,10 @@ func TestE2E_StreamTextEmitsRawWhenEnabled(t *testing.T) {
 	if err := json.Unmarshal(raws[0], &first); err != nil {
 		t.Fatalf("Raw payload not an object: %v", err)
 	}
-	// Raw payload is wrapped: {"raw_value": {chunk}}.
-	inner, ok := first["raw_value"].(map[string]any)
+	// A raw part is {"type":"raw","rawValue": {chunk}}.
+	inner, ok := first["rawValue"].(map[string]any)
 	if !ok {
-		t.Fatalf("Raw payload missing raw_value, got %#v", first)
+		t.Fatalf("raw part missing rawValue, got %#v", first)
 	}
 	delta, ok := inner["choices"].([]any)[0].(map[string]any)["delta"].(map[string]any)["content"].(string)
 	if !ok || delta != "Hello" {
@@ -88,12 +88,12 @@ func TestE2E_StreamStartCarriesNonEmptyWarnings(t *testing.T) {
 	srv.SetContentType("text/event-stream")
 	srv.SetResponse(buildTextDeltaSSE())
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	// top_k on a provider that supports it produces no warning on the openai
 	// full profile; the Go assertion here is that a non-empty warnings array
-	// in StreamStart decodes without breaking the stream (the warning itself
+	// in stream-start decodes without breaking the stream (the warning itself
 	// is produced by the groq profile — covered in groq_test.rs).
 	stream := m.StreamText(`"Say hello"`, mustMarshalOptions(t, &GenerateTextOptions{
 		TopK: f64Ptr(0.5),
@@ -105,13 +105,13 @@ func TestE2E_StreamStartCarriesNonEmptyWarnings(t *testing.T) {
 		if err != nil {
 			t.Fatalf("failed to parse: %v", err)
 		}
-		if sp.Tag == "StreamStart" {
+		if sp.Type == "stream-start" {
 			sawStreamStart = true
 			var start struct {
 				Warnings []json.RawMessage `json:"warnings"`
 			}
-			if err := json.Unmarshal(sp.Payload, &start); err != nil {
-				t.Fatalf("failed to decode StreamStart payload: %v", err)
+			if err := json.Unmarshal(sp.Raw, &start); err != nil {
+				t.Fatalf("failed to decode stream-start part: %v", err)
 			}
 			// warnings may be empty on the full profile; the point is that a
 			// non-empty array would decode here without error.
@@ -122,7 +122,7 @@ func TestE2E_StreamStartCarriesNonEmptyWarnings(t *testing.T) {
 		t.Fatalf("stream error: %v", err)
 	}
 	if !sawStreamStart {
-		t.Fatal("no StreamStart part received")
+		t.Fatal("no stream-start part received")
 	}
 }
 
diff --git a/bindings/go/e2e_test.go b/bindings/go/e2e_test.go
index 40ef89b5c..c1e1c2f4b 100644
--- a/bindings/go/e2e_test.go
+++ b/bindings/go/e2e_test.go
@@ -1,18 +1,16 @@
 // Structured end-to-end tests for the Go binding.
 //
 // These spin up a local mock HTTP server speaking the OpenAI chat-completions
-// wire format, point an OpenAI Model at it via the base_url constructor,
+// wire format, point a chat-completions Model at it via a base URL,
 // and assert that the binding correctly:
-//   1. parses text from a provider response,
-//   2. parses tool_calls out of a provider response,
-//   3. forwards multi-role messages,
-//   4. forwards tool_choice,
-//   5. parses tool-call stream parts from an SSE stream,
-//   6. yields text-delta stream parts.
+//  1. parses text from a provider response,
+//  2. parses tool_calls out of a provider response,
+//  3. forwards multi-role messages,
+//  4. forwards tool_choice,
+//  5. parses tool-call stream parts from an SSE stream,
+//  6. yields text-delta stream parts.
 //
 // No real network access is performed — every request hits 127.0.0.1.
-// Mirrors the Kotlin StructuredE2ETest.kt.
-
 package aimux
 
 import (
@@ -155,6 +153,18 @@ func waitForChannelFull[T any](t *testing.T, channel chan T) {
 	}
 }
 
+// chatModel is a model that speaks the OpenAI chat-completions wire format at
+// baseURL: the registry's OpenAI-compatible preset (the OpenAI constructors
+// talk to the Responses API).
+func chatModel(t *testing.T, baseURL string) *Model {
+	t.Helper()
+	m, err := ProviderWithBase("deepseek", "sk-test-fake-key", "gpt-4o", baseURL)
+	if err != nil {
+		t.Fatalf("ProviderWithBase: %v", err)
+	}
+	return m
+}
+
 // ── Canned OpenAI responses ────────────────────────────────────────────────────
 
 // plainOpenAIResponse is a plain OpenAI text response (no tool calls).
@@ -194,7 +204,7 @@ func weatherToolsOpts(toolChoice string) string {
 			{
 				"type": "function",
 				"name": "get_weather",
-				"input_schema": map[string]any{
+				"inputSchema": map[string]any{
 					"type": "object",
 					"properties": map[string]any{
 						"location": map[string]any{"type": "string"},
@@ -204,7 +214,7 @@ func weatherToolsOpts(toolChoice string) string {
 		},
 	}
 	if toolChoice != "" {
-		opts["tool_choice"] = toolChoice
+		opts["toolChoice"] = toolChoice
 	}
 	b, _ := json.Marshal(opts)
 	return string(b)
@@ -217,7 +227,7 @@ func TestE2E_GenerateTextParsesText(t *testing.T) {
 	defer srv.Close()
 	srv.SetResponse(plainOpenAIResponse)
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	result, err := m.GenerateText(`"What is Rust?"`, "")
@@ -243,7 +253,7 @@ func TestE2E_GenerateTextParsesToolCalls(t *testing.T) {
 	defer srv.Close()
 	srv.SetResponse(toolCallOpenAIResponse)
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	result, err := m.GenerateText(`"What is the weather in Tokyo?"`, weatherToolsOpts(""))
@@ -273,21 +283,23 @@ func TestE2E_GenerateTextParsesToolCalls(t *testing.T) {
 		t.Errorf("expected location Tokyo, got %v", input["location"])
 	}
 
-	// raw.content should contain a "ToolCall" variant.
+	// raw.content should contain a "tool-call" part.
 	var rawWrap struct {
 		Raw struct {
-			Content []map[string]json.RawMessage `json:"content"`
+			Content []struct {
+				Type string `json:"type"`
+			} `json:"content"`
 		} `json:"raw"`
 	}
 	json.Unmarshal([]byte(result), &rawWrap)
-	hasToolCallVariant := false
+	hasToolCallPart := false
 	for _, part := range rawWrap.Raw.Content {
-		if _, ok := part["ToolCall"]; ok {
-			hasToolCallVariant = true
+		if part.Type == "tool-call" {
+			hasToolCallPart = true
 		}
 	}
-	if !hasToolCallVariant {
-		t.Error("expected raw.content to contain a ToolCall variant")
+	if !hasToolCallPart {
+		t.Error("expected raw.content to contain a tool-call part")
 	}
 }
 
@@ -296,7 +308,7 @@ func TestE2E_MultiRoleMessagesReachProvider(t *testing.T) {
 	defer srv.Close()
 	srv.SetResponse(plainOpenAIResponse)
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	msgs := []ModelMessage{
@@ -343,7 +355,7 @@ func TestE2E_ToolChoiceReachesProvider(t *testing.T) {
 	defer srv.Close()
 	srv.SetResponse(toolCallOpenAIResponse)
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	_, err := m.GenerateText(`"What is the weather in Tokyo?"`, weatherToolsOpts("required"))
@@ -370,7 +382,7 @@ func TestE2E_StreamTextParsesToolCallStreamParts(t *testing.T) {
 	sseResponse := buildToolCallSSE()
 	srv.SetResponse(sseResponse)
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	stream := m.StreamText(`"What is the weather in Tokyo?"`, weatherToolsOpts(""))
@@ -386,17 +398,16 @@ func TestE2E_StreamTextParsesToolCallStreamParts(t *testing.T) {
 		t.Fatal("expected at least one stream part")
 	}
 
-	// The first part should be a StreamStart or ToolCall stream part.
-	// We just verify we got JSON objects with known tags.
+	// We just verify we got JSON objects tagged by "type".
 	for _, p := range parts {
 		sp, err := ParseStreamPart(p)
 		if err != nil {
 			t.Errorf("failed to parse stream part: %v", err)
 			continue
 		}
-		// Tags we expect: StreamStart, TextDelta, ToolCall, Finish, etc.
-		if sp.Tag == "" {
-			t.Error("expected non-empty stream part tag")
+		// Types we expect: stream-start, text-delta, tool-call, finish, etc.
+		if sp.Type == "" {
+			t.Error("expected non-empty stream part type")
 		}
 	}
 }
@@ -410,7 +421,7 @@ func TestE2E_StreamTextYieldsTextDeltas(t *testing.T) {
 	sseResponse := buildTextDeltaSSE()
 	srv.SetResponse(sseResponse)
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	stream := m.StreamText(`"Say hello"`, "")
@@ -423,12 +434,12 @@ func TestE2E_StreamTextYieldsTextDeltas(t *testing.T) {
 			t.Errorf("failed to parse: %v", err)
 			continue
 		}
-		switch sp.Tag {
-		case "TextDelta":
+		switch sp.Type {
+		case "text-delta":
 			var td TextDeltaPayload
-			json.Unmarshal(sp.Payload, &td)
+			json.Unmarshal(sp.Raw, &td)
 			textBuilder.WriteString(td.Delta)
-		case "Finish":
+		case "finish":
 			gotFinish = true
 		}
 	}
@@ -441,7 +452,7 @@ func TestE2E_StreamTextYieldsTextDeltas(t *testing.T) {
 		t.Errorf("expected 'Hello world', got %q", text)
 	}
 	if !gotFinish {
-		t.Error("expected a Finish part")
+		t.Error("expected a finish part")
 	}
 }
 
@@ -464,7 +475,7 @@ func TestStreamTextContextCancelsInFlightRequest(t *testing.T) {
 	}))
 	defer srv.Close()
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 	ctx, cancel := context.WithCancel(context.Background())
 	stream := m.StreamTextContext(ctx, `"wait"`, "")
@@ -501,7 +512,7 @@ func TestStreamCancelUnblocksFullPartsChannel(t *testing.T) {
 	srv, allSent, requestDone := newBackpressureSSEServer(800)
 	defer srv.Close()
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 	stream := m.StreamText(`"fill the channel"`, "")
 	defer stream.Cancel()
@@ -567,8 +578,7 @@ func buildToolCallSSE() string {
 
 // ── Tool-call full round-trip ─────────────────────────────────────────────────
 
-// TestE2E_ToolCallFullRoundTrip mirrors Kotlin's `tool-call full round-trip`
-// scenario: the model requests a tool call, the client sends the tool result
+// TestE2E_ToolCallFullRoundTrip: the model requests a tool call, the client sends the tool result
 // back as a multi-part tool message, and the model produces a final answer.
 // This verifies ModelMessage multi-part content reaches the provider correctly.
 func TestE2E_ToolCallFullRoundTrip(t *testing.T) {
@@ -578,7 +588,7 @@ func TestE2E_ToolCallFullRoundTrip(t *testing.T) {
 	// Queue: first call returns a tool_call, second call returns the final text.
 	srv.setResponses(toolCallOpenAIResponse, plainOpenAIResponse)
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	// 1. First call: model requests get_weather tool call.
@@ -596,15 +606,15 @@ func TestE2E_ToolCallFullRoundTrip(t *testing.T) {
 	}
 
 	// 2. Build a multi-part tool-result message using ModelMessage's Content
-	//    (any) field — this is the path that was previously unsupported.
+	//    (any) field.
 	toolResultMsg := ModelMessage{
 		Role: RoleTool,
 		Content: []map[string]any{
 			{
-				"type":         "tool_result",
-				"tool_call_id": call.ToolCallID,
-				"tool_name":    call.ToolName,
-				"result":       map[string]any{"temperature": "20C"},
+				"type":       "tool-result",
+				"toolCallId": call.ToolCallID,
+				"toolName":   call.ToolName,
+				"output":     map[string]any{"type": "json", "value": map[string]any{"temperature": "20C"}},
 			},
 		},
 	}
@@ -626,7 +636,7 @@ func TestE2E_ToolCallFullRoundTrip(t *testing.T) {
 	}
 
 	// Verify the second request carried the tool-result message correctly.
-	// The engine converts multi-part tool_result content into OpenAI's wire
+	// The engine converts a tool-result part into OpenAI's wire
 	// format: {"role":"tool","tool_call_id":"...","content":"<stringified result>"}.
 	reqBody := srv.LastRequestBody()
 	var body struct {
diff --git a/bindings/go/error.go b/bindings/go/error.go
index 048e7b441..ac33a330f 100644
--- a/bindings/go/error.go
+++ b/bindings/go/error.go
@@ -47,7 +47,7 @@ const (
 	CodeLoadAPIKey Code = 18
 	// CodeLoadSetting: a required provider setting was not passed and its
 	// fallback environment variable is unset (LoadSettingError).
-	CodeLoadSetting Code = 19
+	CodeLoadSetting       Code = 19
 	CodeNoOutputGenerated Code = 20
 )
 
diff --git a/bindings/go/error_test.go b/bindings/go/error_test.go
index 5a9632d75..a12f88e63 100644
--- a/bindings/go/error_test.go
+++ b/bindings/go/error_test.go
@@ -160,7 +160,7 @@ func TestRetryableIsNotDerivedFromStatus(t *testing.T) {
 // Malformed raw JSON is caught by the binding before the C call: a plain
 // error naming the parameter, not an *Error.
 func TestInvalidJSONNamesParameter(t *testing.T) {
-	err := RegisterProviders("{not json")
+	err := InitProxy("{not json")
 	if err == nil {
 		t.Fatal("expected error for malformed config")
 	}
@@ -417,7 +417,7 @@ func TestMarshalledOptionsRejectRawPassThrough(t *testing.T) {
 		// invalid UTF-8 becomes U+FFFD, NUL becomes a \u0000 escape. No error here.
 		{"plain string field is sanitized, not rejected", func() error {
 			s, e := marshalJSON("opts", &SpeechCallOptions{Text: "hi\xffthere\x00"})
-			if e == nil && !strings.Contains(s, `"hi\ufffdthere\u0000"`) {
+			if e == nil && !(strings.Contains(s, `"hi\ufffdthere\u0000"`) || strings.Contains(s, "\"hi\uFFFDthere\\u0000\"")) {
 				t.Errorf("expected U+FFFD / \\u0000 coercion, got %s", s)
 			}
 			return e
@@ -460,13 +460,13 @@ func TestEmbedRejectsRawPassThroughOpts(t *testing.T) {
 	if _, err := m.Embed([]string{"hi"}, &EmbeddingCallOptions{
 		ProviderOptions: jsonObj(`{"a":"\ud800"}`),
 	}); err == nil || !strings.Contains(err.Error(), "opts: invalid JSON") {
-		t.Fatalf("Embed with a lone surrogate in provider_options: %v", err)
+		t.Fatalf("Embed with a lone surrogate in providerOptions: %v", err)
 	}
 }
 
 func TestCodeFromCRejectsOutOfRange(t *testing.T) {
-	// 4 is retired; 14 is Retry; 15..17 are tool-call errors.
-	for _, bad := range []int{0, 4, 20, 999} {
+	// 4 is retired; 14 is Retry; 15..17 are tool-call errors; 20 is the last.
+	for _, bad := range []int{0, 4, 21, 999} {
 		if _, ok := codeFromC(bad); ok {
 			t.Fatalf("%d is not an AiMuxError variant", bad)
 		}
diff --git a/bindings/go/examples/generate/main.go b/bindings/go/examples/generate/main.go
index 4dcf69335..010d7ff9e 100644
--- a/bindings/go/examples/generate/main.go
+++ b/bindings/go/examples/generate/main.go
@@ -41,9 +41,9 @@ func main() {
 		log.Fatalf("stream failed: %v", err)
 	}
 	for part := range stream.Parts() {
-		if part.Tag == "TextDelta" {
+		if part.Type == "text-delta" {
 			var td aimux.TextDeltaPayload
-			json.Unmarshal(part.Payload, &td)
+			json.Unmarshal(part.Raw, &td)
 			fmt.Print(td.Delta)
 		}
 	}
diff --git a/bindings/go/multimodal.go b/bindings/go/multimodal.go
index 05417a26d..189ca5a22 100644
--- a/bindings/go/multimodal.go
+++ b/bindings/go/multimodal.go
@@ -779,19 +779,20 @@ type TranscriptionSession struct {
 }
 
 // InputAudioFormat is the input audio format for streaming transcription
-// (RFC-0028): e.g. {"format_type": "audio/pcm", "rate": 24000}.
+// (RFC-0028): e.g. {"formatType": "audio/pcm", "rate": 24000}.
 type InputAudioFormat struct {
-	FormatType string  `json:"format_type"`
+	FormatType string  `json:"formatType"`
 	Rate       *uint32 `json:"rate,omitempty"`
 }
 
 // TranscriptionSessionOpts are the optional session options (RFC-0028):
-// input audio format, provider options, headers, include_raw_chunks.
+// input audio format, provider options, headers, includeRawChunks, timeout.
 type TranscriptionSessionOpts struct {
-	InputAudioFormat *InputAudioFormat          `json:"input_audio_format,omitempty"`
-	ProviderOptions  map[string]json.RawMessage `json:"provider_options,omitempty"`
+	InputAudioFormat *InputAudioFormat          `json:"inputAudioFormat,omitempty"`
+	ProviderOptions  map[string]json.RawMessage `json:"providerOptions,omitempty"`
 	Headers          map[string]string          `json:"headers,omitempty"`
-	IncludeRawChunks *bool                      `json:"include_raw_chunks,omitempty"`
+	IncludeRawChunks *bool                      `json:"includeRawChunks,omitempty"`
+	Timeout          *TimeoutConfiguration      `json:"timeout,omitempty"`
 }
 
 // ErrTranscriptionEnded is returned by NextPart when the stream ended
diff --git a/bindings/go/multimodal_test.go b/bindings/go/multimodal_test.go
index b695ab57b..4eb66f42e 100644
--- a/bindings/go/multimodal_test.go
+++ b/bindings/go/multimodal_test.go
@@ -134,7 +134,7 @@ func TestE2E_Embedding(t *testing.T) {
 	if len(result.Embeddings[0]) != 3 {
 		t.Errorf("expected 3 dimensions, got %d", len(result.Embeddings[0]))
 	}
-	if result.Usage == nil || result.Usage.Tokens == nil || *result.Usage.Tokens != 5 {
+	if result.Usage == nil || result.Usage.Tokens != 5 {
 		t.Error("usage tokens mismatch")
 	}
 }
@@ -143,15 +143,20 @@ func TestE2E_Embedding(t *testing.T) {
 
 func TestE2E_SpeechResultParsing(t *testing.T) {
 	result, err := ParseSpeechResult(`{
-		"audio": {"Base64": "aGVsbG8="},
+		"audio": "aGVsbG8=",
 		"warnings": [],
-		"response": {"id": "resp-1"}
+		"response": {"modelId": "tts-1"}
 	}`)
 	if err != nil {
 		t.Fatalf("parse failed: %v", err)
 	}
-	if result.Audio.Base64 == nil || *result.Audio.Base64 != "aGVsbG8=" {
-		t.Error("audio base64 mismatch")
+	if string(result.Audio) != "hello" {
+		t.Errorf("audio mismatch: %q", result.Audio)
+	}
+	// The byte-array form decodes to the same bytes.
+	result, err = ParseSpeechResult(`{"audio": [104,105], "warnings": [], "response": {}}`)
+	if err != nil || string(result.Audio) != "hi" {
+		t.Errorf("binary audio: %q, %v", result.Audio, err)
 	}
 }
 
@@ -159,18 +164,18 @@ func TestE2E_SpeechResultParsing(t *testing.T) {
 
 func TestE2E_ImageResultParsing(t *testing.T) {
 	result, err := ParseImageResult(`{
-		"images": {"Base64": ["aW1hZ2Ux"]},
+		"images": ["aW1hZ2Ux"],
 		"warnings": [],
-		"response": {"id": "resp-1"}
+		"response": {"modelId": "dall-e-3"}
 	}`)
 	if err != nil {
 		t.Fatalf("parse failed: %v", err)
 	}
-	if len(result.Images.Base64) != 1 {
-		t.Fatalf("expected 1 image, got %d", len(result.Images.Base64))
+	if len(result.Images) != 1 {
+		t.Fatalf("expected 1 image, got %d", len(result.Images))
 	}
-	if result.Images.Base64[0] != "aW1hZ2Ux" {
-		t.Error("image base64 mismatch")
+	if string(result.Images[0]) != "image1" {
+		t.Error("image mismatch")
 	}
 }
 
@@ -179,7 +184,7 @@ func TestE2E_ImageResultParsing(t *testing.T) {
 func TestE2E_TranscriptionResultParsing(t *testing.T) {
 	result, err := ParseTranscriptionResult(`{
 		"text": "Hello world",
-		"segments": [{"text": "Hello", "start": 0.0, "end": 1.0}],
+		"segments": [{"text": "Hello", "startSecond": 0.0, "endSecond": 1.0}],
 		"language": "en",
 		"warnings": [],
 		"response": {"id": "resp-1"}
@@ -206,8 +211,8 @@ func TestE2E_TranscriptionResultParsing(t *testing.T) {
 func TestE2E_RerankingResultParsing(t *testing.T) {
 	result, err := ParseRerankingResult(`{
 		"ranking": [
-			{"index": 1, "relevance_score": 0.95},
-			{"index": 0, "relevance_score": 0.30}
+			{"index": 1, "relevanceScore": 0.95},
+			{"index": 0, "relevanceScore": 0.30}
 		],
 		"warnings": []
 	}`)
@@ -229,9 +234,9 @@ func TestE2E_RerankingResultParsing(t *testing.T) {
 
 func TestE2E_VideoResultParsing(t *testing.T) {
 	result, err := ParseVideoResult(`{
-		"videos": [{"Url": {"url": "https://example.com/video.mp4", "media_type": "video/mp4"}}],
+		"videos": [{"type": "url", "url": "https://example.com/video.mp4", "mediaType": "video/mp4"}],
 		"warnings": [],
-		"response": {"id": "resp-1"}
+		"response": {"modelId": "veo"}
 	}`)
 	if err != nil {
 		t.Fatalf("parse failed: %v", err)
@@ -239,7 +244,7 @@ func TestE2E_VideoResultParsing(t *testing.T) {
 	if len(result.Videos) != 1 {
 		t.Fatalf("expected 1 video, got %d", len(result.Videos))
 	}
-	if result.Videos[0].Url == nil || result.Videos[0].Url.URL != "https://example.com/video.mp4" {
+	if result.Videos[0].Type != "url" || result.Videos[0].URL != "https://example.com/video.mp4" {
 		t.Error("video URL mismatch")
 	}
 }
@@ -266,7 +271,7 @@ func TestVideoPollOptionsWireFormat(t *testing.T) {
 	if !ok {
 		t.Fatalf("poll missing from wire format: %s", encoded)
 	}
-	if poll["interval_ms"] != float64(intervalMS) || poll["timeout_ms"] != float64(timeoutMS) {
+	if poll["intervalMs"] != float64(intervalMS) || poll["timeoutMs"] != float64(timeoutMS) {
 		t.Fatalf("unexpected poll wire format: %s", encoded)
 	}
 }
@@ -299,8 +304,8 @@ func TestE2E_SearchResultParsing(t *testing.T) {
 
 func TestE2E_UploadFileResultParsing(t *testing.T) {
 	result, err := ParseUploadFileResult(`{
-		"provider_reference": {"openai": "file-abc123"},
-		"media_type": "application/pdf",
+		"providerReference": {"openai": "file-abc123"},
+		"mediaType": "application/pdf",
 		"filename": "doc.pdf",
 		"warnings": []
 	}`)
@@ -319,7 +324,7 @@ func TestE2E_UploadFileResultParsing(t *testing.T) {
 
 func TestE2E_RerankWithOptions(t *testing.T) {
 	// Cohere reranking uses its own endpoint; we verify options marshaling.
-	docs := `{"Text":{"values":["doc1","doc2","doc3"]}}`
+	docs := `{"type":"text","values":["doc1","doc2","doc3"]}`
 	opts := &RerankingCallOptions{
 		Documents: json.RawMessage(docs),
 		Query:     "which is most relevant?",
@@ -336,8 +341,8 @@ func TestE2E_RerankWithOptions(t *testing.T) {
 	if !strings.Contains(s, `"query"`) {
 		t.Errorf("expected query field: %s", s)
 	}
-	if !strings.Contains(s, `"top_n"`) {
-		t.Errorf("expected top_n field: %s", s)
+	if !strings.Contains(s, `"topN"`) {
+		t.Errorf("expected topN field: %s", s)
 	}
 }
 
diff --git a/bindings/go/multimodal_types.go b/bindings/go/multimodal_types.go
index 3eb2eabf7..7e78abfbe 100644
--- a/bindings/go/multimodal_types.go
+++ b/bindings/go/multimodal_types.go
@@ -1,13 +1,41 @@
 // Typed multimodal data structures mirroring the aimux-core wire format
-// (same shapes as the ts-rs generated .ts types in bindings/node/src/types/).
-//
-// Field names use JSON tags matching the wire format's snake_case. These types
-// are intentionally lenient on decode (unknown keys ignored, every field
-// optional) so future engine additions don't break existing clients.
+// (same shapes as the ts-rs generated .ts types in bindings/node/src/types/):
+// camelCase field names, unions tagged by "type". These types are
+// intentionally lenient on decode (unknown keys ignored, every field optional)
+// so future engine additions don't break existing clients.
 
 package aimux
 
-import "encoding/json"
+import (
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+)
+
+// Bytes is binary data on the wire: a base64 string or an array of byte
+// values. It decodes either form; it encodes as a base64 string.
+type Bytes []byte
+
+func (b *Bytes) UnmarshalJSON(data []byte) error {
+	if len(data) > 0 && data[0] == '[' {
+		var values []uint8
+		if err := json.Unmarshal(data, &values); err != nil {
+			return err
+		}
+		*b = values
+		return nil
+	}
+	var text string
+	if err := json.Unmarshal(data, &text); err != nil {
+		return fmt.Errorf("aimux: bytes must be a base64 string or an array of numbers: %w", err)
+	}
+	decoded, err := base64.StdEncoding.DecodeString(text)
+	if err != nil {
+		return fmt.Errorf("aimux: invalid base64: %w", err)
+	}
+	*b = decoded
+	return nil
+}
 
 // ── Shared types ────────────────────────────────────────────────────────────
 
@@ -18,7 +46,7 @@ type Warning = json.RawMessage
 
 // EmbeddingUsage is token usage for an embedding call (input tokens only).
 type EmbeddingUsage struct {
-	Tokens *uint32 `json:"tokens,omitempty"`
+	Tokens uint32 `json:"tokens"`
 }
 
 // EmbeddingResponse is provider response metadata for embeddings.
@@ -31,28 +59,25 @@ type EmbeddingResponse struct {
 type EmbeddingResult struct {
 	Embeddings       [][]float32        `json:"embeddings"`
 	Usage            *EmbeddingUsage    `json:"usage,omitempty"`
-	ProviderMetadata json.RawMessage    `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage    `json:"providerMetadata,omitempty"`
 	Response         *EmbeddingResponse `json:"response,omitempty"`
 	Warnings         []Warning          `json:"warnings,omitempty"`
 }
 
 // EmbeddingCallOptions is the options for an embedding call.
 type EmbeddingCallOptions struct {
-	Values          []string              `json:"values,omitempty"`
-	MaxRetries      *uint32               `json:"max_retries,omitempty"`
+	Values          []string              `json:"values"`
+	MaxRetries      *uint32               `json:"maxRetries,omitempty"`
 	Timeout         *TimeoutConfiguration `json:"timeout,omitempty"`
-	ProviderOptions jsonObj               `json:"provider_options"`
+	ProviderOptions jsonObj               `json:"providerOptions"`
 	Headers         map[string]string     `json:"headers,omitempty"`
 }
 
 // ── Speech (TTS) ──────────────────────────────────────────────────────────────
 
-// AudioData is generated audio: base64 string or raw binary bytes.
-// Wire format: {"Base64": "string"} | {"Binary": [numbers]}
-type AudioData struct {
-	Base64 *string `json:"Base64,omitempty"`
-	Binary []uint8 `json:"Binary,omitempty"`
-}
+// AudioData is generated audio: a base64 string or an array of bytes on the
+// wire, decoded to bytes.
+type AudioData = Bytes
 
 // SpeechRequest is request metadata for speech generation.
 type SpeechRequest struct {
@@ -64,7 +89,7 @@ type SpeechRequest struct {
 // SpeechResponse is provider response metadata for speech.
 type SpeechResponse struct {
 	Timestamp *string           `json:"timestamp,omitempty"`
-	ModelID   *string           `json:"model_id,omitempty"`
+	ModelID   *string           `json:"modelId,omitempty"`
 	Headers   map[string]string `json:"headers,omitempty"`
 	Body      any               `json:"body,omitempty"`
 }
@@ -75,42 +100,40 @@ type SpeechResult struct {
 	Warnings         []Warning       `json:"warnings,omitempty"`
 	Request          *SpeechRequest  `json:"request,omitempty"`
 	Response         SpeechResponse  `json:"response"`
-	ProviderMetadata json.RawMessage `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage `json:"providerMetadata,omitempty"`
 }
 
 // SpeechCallOptions is the options for speech generation.
 type SpeechCallOptions struct {
 	Text            string                `json:"text"`
 	Voice           *string               `json:"voice,omitempty"`
-	OutputFormat    *string               `json:"output_format,omitempty"`
+	OutputFormat    *string               `json:"outputFormat,omitempty"`
 	Instructions    *string               `json:"instructions,omitempty"`
 	Speed           *float64              `json:"speed,omitempty"`
 	Language        *string               `json:"language,omitempty"`
-	MaxRetries      *uint32               `json:"max_retries,omitempty"`
+	MaxRetries      *uint32               `json:"maxRetries,omitempty"`
 	Timeout         *TimeoutConfiguration `json:"timeout,omitempty"`
-	ProviderOptions jsonObj               `json:"provider_options"`
+	ProviderOptions jsonObj               `json:"providerOptions"`
 	Headers         map[string]string     `json:"headers,omitempty"`
 }
 
 // ── Image ─────────────────────────────────────────────────────────────────────
 
-// ImageOutputs is the generated images: all base64 or all binary.
-type ImageOutputs struct {
-	Base64 []string  `json:"Base64,omitempty"`
-	Binary [][]uint8 `json:"Binary,omitempty"`
-}
+// ImageOutputs is the generated images: base64 strings or byte arrays on the
+// wire, decoded to bytes.
+type ImageOutputs = []Bytes
 
 // ImageUsage is token usage for image generation (if reported).
 type ImageUsage struct {
-	InputTokens  *uint32 `json:"input_tokens,omitempty"`
-	OutputTokens *uint32 `json:"output_tokens,omitempty"`
-	TotalTokens  *uint32 `json:"total_tokens,omitempty"`
+	InputTokens  *uint32 `json:"inputTokens,omitempty"`
+	OutputTokens *uint32 `json:"outputTokens,omitempty"`
+	TotalTokens  *uint32 `json:"totalTokens,omitempty"`
 }
 
 // ImageResponse is provider response metadata for images.
 type ImageResponse struct {
 	Timestamp *string           `json:"timestamp,omitempty"`
-	ModelID   *string           `json:"model_id,omitempty"`
+	ModelID   *string           `json:"modelId,omitempty"`
 	Headers   map[string]string `json:"headers,omitempty"`
 }
 
@@ -118,7 +141,7 @@ type ImageResponse struct {
 type ImageResult struct {
 	Images           ImageOutputs    `json:"images"`
 	Warnings         []Warning       `json:"warnings,omitempty"`
-	ProviderMetadata json.RawMessage `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage `json:"providerMetadata,omitempty"`
 	Response         ImageResponse   `json:"response"`
 	Usage            *ImageUsage     `json:"usage,omitempty"`
 }
@@ -126,15 +149,15 @@ type ImageResult struct {
 // ImageCallOptions is the options for image generation.
 type ImageCallOptions struct {
 	Prompt          *string               `json:"prompt,omitempty"`
-	N               *int                  `json:"n,omitempty"`
-	Size            *string               `json:"size,omitempty"`
-	AspectRatio     *string               `json:"aspect_ratio,omitempty"`
+	N               int                   `json:"n"`
+	Size            *string               `json:"size,omitempty"`        // "WxH", e.g. "1024x1024"
+	AspectRatio     *string               `json:"aspectRatio,omitempty"` // "W:H", e.g. "16:9"
 	Seed            *uint64               `json:"seed,omitempty"`
 	Files           []json.RawMessage     `json:"files,omitempty"`
 	Mask            json.RawMessage       `json:"mask,omitempty"`
-	MaxRetries      *uint32               `json:"max_retries,omitempty"`
+	MaxRetries      *uint32               `json:"maxRetries,omitempty"`
 	Timeout         *TimeoutConfiguration `json:"timeout,omitempty"`
-	ProviderOptions jsonObj               `json:"provider_options"`
+	ProviderOptions jsonObj               `json:"providerOptions"`
 	Headers         map[string]string     `json:"headers,omitempty"`
 }
 
@@ -145,8 +168,8 @@ type TranscriptionSegment struct {
 	// StartSecond/EndSecond are required f64 in the core (not Option), so they
 	// are values rather than pointers and must always be marshalled.
 	Text        string  `json:"text"`
-	StartSecond float64 `json:"start_second"`
-	EndSecond   float64 `json:"end_second"`
+	StartSecond float64 `json:"startSecond"`
+	EndSecond   float64 `json:"endSecond"`
 }
 
 // TranscriptionRequest is request metadata for transcription.
@@ -159,7 +182,7 @@ type TranscriptionRequest struct {
 // TranscriptionResponse is provider response metadata for transcription.
 type TranscriptionResponse struct {
 	Timestamp *string           `json:"timestamp,omitempty"`
-	ModelID   *string           `json:"model_id,omitempty"`
+	ModelID   *string           `json:"modelId,omitempty"`
 	Headers   map[string]string `json:"headers,omitempty"`
 	Body      any               `json:"body,omitempty"`
 }
@@ -169,20 +192,20 @@ type TranscriptionResult struct {
 	Text              string                 `json:"text"`
 	Segments          []TranscriptionSegment `json:"segments,omitempty"`
 	Language          *string                `json:"language,omitempty"`
-	DurationInSeconds *float64               `json:"duration_in_seconds,omitempty"`
+	DurationInSeconds *float64               `json:"durationInSeconds,omitempty"`
 	Warnings          []Warning              `json:"warnings,omitempty"`
 	Request           *TranscriptionRequest  `json:"request,omitempty"`
 	Response          TranscriptionResponse  `json:"response"`
-	ProviderMetadata  json.RawMessage        `json:"provider_metadata,omitempty"`
+	ProviderMetadata  json.RawMessage        `json:"providerMetadata,omitempty"`
 }
 
 // TranscriptionCallOptions is the options for transcription.
 type TranscriptionCallOptions struct {
-	Audio           json.RawMessage       `json:"audio"`
-	MediaType       string                `json:"media_type"`
-	MaxRetries      *uint32               `json:"max_retries,omitempty"`
+	Audio           Bytes                 `json:"audio"`
+	MediaType       string                `json:"mediaType"`
+	MaxRetries      *uint32               `json:"maxRetries,omitempty"`
 	Timeout         *TimeoutConfiguration `json:"timeout,omitempty"`
-	ProviderOptions jsonObj               `json:"provider_options"`
+	ProviderOptions jsonObj               `json:"providerOptions"`
 	Headers         map[string]string     `json:"headers,omitempty"`
 }
 
@@ -191,14 +214,14 @@ type TranscriptionCallOptions struct {
 // RerankingRank is a single reranked entry.
 type RerankingRank struct {
 	Index          int     `json:"index"`
-	RelevanceScore float64 `json:"relevance_score"`
+	RelevanceScore float64 `json:"relevanceScore"`
 }
 
 // RerankingResponse is provider response metadata for reranking.
 type RerankingResponse struct {
 	ID        *string           `json:"id,omitempty"`
 	Timestamp *string           `json:"timestamp,omitempty"`
-	ModelID   *string           `json:"model_id,omitempty"`
+	ModelID   *string           `json:"modelId,omitempty"`
 	Headers   map[string]string `json:"headers,omitempty"`
 	Body      any               `json:"body,omitempty"`
 }
@@ -206,7 +229,7 @@ type RerankingResponse struct {
 // RerankingResult is the result of a reranking call.
 type RerankingResult struct {
 	Ranking          []RerankingRank    `json:"ranking"`
-	ProviderMetadata json.RawMessage    `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage    `json:"providerMetadata,omitempty"`
 	Warnings         []Warning          `json:"warnings,omitempty"`
 	Response         *RerankingResponse `json:"response,omitempty"`
 }
@@ -215,45 +238,28 @@ type RerankingResult struct {
 type RerankingCallOptions struct {
 	Documents       json.RawMessage       `json:"documents"`
 	Query           string                `json:"query"`
-	TopN            *int                  `json:"top_n,omitempty"`
-	MaxRetries      *uint32               `json:"max_retries,omitempty"`
+	TopN            *int                  `json:"topN,omitempty"`
+	MaxRetries      *uint32               `json:"maxRetries,omitempty"`
 	Timeout         *TimeoutConfiguration `json:"timeout,omitempty"`
-	ProviderOptions jsonObj               `json:"provider_options"`
+	ProviderOptions jsonObj               `json:"providerOptions"`
 	Headers         map[string]string     `json:"headers,omitempty"`
 }
 
 // ── Video ───────────────────────────────────────────────────────────────────
 
-// VideoData is generated video: URL, base64, or binary.
-// Wire format: {"Url": {"url":"...", "media_type":"..."}} | {"Base64": {...}} | {"Binary": {...}}
+// VideoData is generated video, tagged by Type: "url" (URL), "base64" or
+// "binary" (Data, decoded to bytes).
 type VideoData struct {
-	Url    *VideoUrlData    `json:"Url,omitempty"`
-	Base64 *VideoBase64Data `json:"Base64,omitempty"`
-	Binary *VideoBinaryData `json:"Binary,omitempty"`
-}
-
-// VideoUrlData is the URL variant of VideoData.
-type VideoUrlData struct {
-	URL       string `json:"url"`
-	MediaType string `json:"media_type,omitempty"`
-}
-
-// VideoBase64Data is the base64 variant of VideoData.
-type VideoBase64Data struct {
-	Data      string `json:"data"`
-	MediaType string `json:"media_type,omitempty"`
-}
-
-// VideoBinaryData is the binary variant of VideoData.
-type VideoBinaryData struct {
-	Data      []uint8 `json:"data"`
-	MediaType string  `json:"media_type,omitempty"`
+	Type      string `json:"type"`
+	URL       string `json:"url,omitempty"`
+	Data      Bytes  `json:"data,omitempty"`
+	MediaType string `json:"mediaType"`
 }
 
 // VideoResponse is provider response metadata for video.
 type VideoResponse struct {
 	Timestamp *string           `json:"timestamp,omitempty"`
-	ModelID   *string           `json:"model_id,omitempty"`
+	ModelID   *string           `json:"modelId,omitempty"`
 	Headers   map[string]string `json:"headers,omitempty"`
 }
 
@@ -261,27 +267,33 @@ type VideoResponse struct {
 type VideoResult struct {
 	Videos           []VideoData     `json:"videos"`
 	Warnings         []Warning       `json:"warnings,omitempty"`
-	ProviderMetadata json.RawMessage `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage `json:"providerMetadata,omitempty"`
 	Response         VideoResponse   `json:"response"`
 }
 
 // VideoPollOptions overrides Core's pacing for an asynchronous video job.
 type VideoPollOptions struct {
-	IntervalMS *uint64 `json:"interval_ms,omitempty"`
-	TimeoutMS  *uint64 `json:"timeout_ms,omitempty"`
+	IntervalMS *uint64 `json:"intervalMs,omitempty"`
+	TimeoutMS  *uint64 `json:"timeoutMs,omitempty"`
 }
 
 // VideoCallOptions is the options for video generation.
 type VideoCallOptions struct {
 	Prompt          *string               `json:"prompt,omitempty"`
 	N               *int                  `json:"n,omitempty"`
-	AspectRatio     *string               `json:"aspect_ratio,omitempty"`
-	Resolution      *string               `json:"resolution,omitempty"`
+	AspectRatio     *string               `json:"aspectRatio,omitempty"` // "W:H", e.g. "16:9"
+	Resolution      *string               `json:"resolution,omitempty"`  // "WxH", e.g. "1280x720"
+	Duration        *float64              `json:"duration,omitempty"`
+	Fps             *float64              `json:"fps,omitempty"`
 	Seed            *uint64               `json:"seed,omitempty"`
-	MaxRetries      *uint32               `json:"max_retries,omitempty"`
+	Image           json.RawMessage       `json:"image,omitempty"`
+	FrameImages     []json.RawMessage     `json:"frameImages,omitempty"`
+	InputReferences []json.RawMessage     `json:"inputReferences,omitempty"`
+	GenerateAudio   *bool                 `json:"generateAudio,omitempty"`
+	MaxRetries      *uint32               `json:"maxRetries,omitempty"`
 	Poll            *VideoPollOptions     `json:"poll,omitempty"`
 	Timeout         *TimeoutConfiguration `json:"timeout,omitempty"`
-	ProviderOptions jsonObj               `json:"provider_options"`
+	ProviderOptions jsonObj               `json:"providerOptions"`
 	Headers         map[string]string     `json:"headers,omitempty"`
 }
 
@@ -292,7 +304,7 @@ type SearchResultItem struct {
 	Title      *string  `json:"title,omitempty"`
 	URL        *string  `json:"url,omitempty"`
 	Content    *string  `json:"content,omitempty"`
-	RawContent *string  `json:"raw_content,omitempty"`
+	RawContent *string  `json:"rawContent,omitempty"`
 	Score      *float64 `json:"score,omitempty"`
 }
 
@@ -306,7 +318,7 @@ type SearchResponse struct {
 type SearchResult struct {
 	Results          []SearchResultItem `json:"results"`
 	Answer           *string            `json:"answer,omitempty"`
-	ProviderMetadata json.RawMessage    `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage    `json:"providerMetadata,omitempty"`
 	Warnings         []Warning          `json:"warnings,omitempty"`
 	Response         *SearchResponse    `json:"response,omitempty"`
 }
@@ -314,14 +326,14 @@ type SearchResult struct {
 // SearchCallOptions is the options for a search call.
 type SearchCallOptions struct {
 	Query             string                `json:"query"`
-	MaxResults        *int                  `json:"max_results,omitempty"`
-	IncludeRawContent *bool                 `json:"include_raw_content,omitempty"`
-	TimeRange         *string               `json:"time_range,omitempty"`
-	IncludeDomains    []string              `json:"include_domains,omitempty"`
-	ExcludeDomains    []string              `json:"exclude_domains,omitempty"`
-	MaxRetries        *uint32               `json:"max_retries,omitempty"`
+	MaxResults        *int                  `json:"maxResults,omitempty"`
+	IncludeRawContent *bool                 `json:"includeRawContent,omitempty"`
+	TimeRange         *string               `json:"timeRange,omitempty"`
+	IncludeDomains    []string              `json:"includeDomains,omitempty"`
+	ExcludeDomains    []string              `json:"excludeDomains,omitempty"`
+	MaxRetries        *uint32               `json:"maxRetries,omitempty"`
 	Timeout           *TimeoutConfiguration `json:"timeout,omitempty"`
-	ProviderOptions   jsonObj               `json:"provider_options"`
+	ProviderOptions   jsonObj               `json:"providerOptions"`
 	Headers           map[string]string     `json:"headers,omitempty"`
 }
 
@@ -329,19 +341,20 @@ type SearchCallOptions struct {
 
 // UploadFileResult is the result of a file upload.
 type UploadFileResult struct {
-	ProviderReference map[string]string `json:"provider_reference"`
-	MediaType         *string           `json:"media_type,omitempty"`
+	ProviderReference map[string]string `json:"providerReference"`
+	MediaType         *string           `json:"mediaType,omitempty"`
 	Filename          *string           `json:"filename,omitempty"`
-	ProviderMetadata  json.RawMessage   `json:"provider_metadata,omitempty"`
+	ProviderMetadata  json.RawMessage   `json:"providerMetadata,omitempty"`
 	Warnings          []Warning         `json:"warnings,omitempty"`
 }
 
 // UploadFileCallOptions is the options for a file upload.
 type UploadFileCallOptions struct {
+	// Data is {"type":"data","data":<base64 or bytes>} or {"type":"text","text":...}.
 	Data            json.RawMessage `json:"data"`
-	MediaType       string          `json:"media_type"`
+	MediaType       string          `json:"mediaType"`
 	Filename        *string         `json:"filename,omitempty"`
-	ProviderOptions jsonObj         `json:"provider_options"`
+	ProviderOptions jsonObj         `json:"providerOptions"`
 }
 
 // jsonObj is a json.RawMessage that serializes as {} when nil (instead of null).
diff --git a/bindings/go/multimodal_withbase_test.go b/bindings/go/multimodal_withbase_test.go
index 3c0c713f4..c7eacb1b2 100644
--- a/bindings/go/multimodal_withbase_test.go
+++ b/bindings/go/multimodal_withbase_test.go
@@ -166,8 +166,7 @@ func TestE2E_SpeechViaMock(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	// Binary audio comes back as AudioData::Binary
-	if len(r.Audio.Binary) == 0 {
+	if len(r.Audio) == 0 {
 		t.Error("empty audio")
 	}
 }
@@ -185,12 +184,12 @@ func TestE2E_ImageViaMock(t *testing.T) {
 
 	prompt := "otter"
 	n := 1
-	r, err := ParseImageResult(must(m.Generate(&ImageCallOptions{Prompt: &prompt, N: &n})))
+	r, err := ParseImageResult(must(m.Generate(&ImageCallOptions{Prompt: &prompt, N: n})))
 	if err != nil {
 		t.Fatal(err)
 	}
-	if len(r.Images.Base64) != 1 {
-		t.Fatalf("got %d images", len(r.Images.Base64))
+	if len(r.Images) != 1 {
+		t.Fatalf("got %d images", len(r.Images))
 	}
 }
 
@@ -269,7 +268,7 @@ func TestE2E_RerankingViaMock(t *testing.T) {
 	topN := 2
 	r, err := ParseRerankingResult(must(m.Rerank(&RerankingCallOptions{
 		Query:     "which?",
-		Documents: json.RawMessage(`{"Text":{"values":["doc1","doc2"]}}`),
+		Documents: json.RawMessage(`{"type":"text","values":["doc1","doc2"]}`),
 		TopN:      &topN,
 	})))
 	if err != nil {
@@ -290,11 +289,11 @@ func TestE2E_VideoViaMock(t *testing.T) {
 	}
 	defer m.Close()
 
-	r, err := ParseVideoResult(`{"videos":[{"Url":{"url":"https://example.com/v.mp4","media_type":"video/mp4"}}]}`)
+	r, err := ParseVideoResult(`{"videos":[{"type":"url","url":"https://example.com/v.mp4","mediaType":"video/mp4"}],"warnings":[],"response":{}}`)
 	if err != nil {
 		t.Fatal(err)
 	}
-	if len(r.Videos) != 1 || r.Videos[0].Url == nil || r.Videos[0].Url.URL != "https://example.com/v.mp4" {
+	if len(r.Videos) != 1 || r.Videos[0].Type != "url" || r.Videos[0].URL != "https://example.com/v.mp4" {
 		t.Fatal("video result parse mismatch")
 	}
 }
diff --git a/bindings/go/repair.go b/bindings/go/repair.go
index 017eb7925..62a7f2c62 100644
--- a/bindings/go/repair.go
+++ b/bindings/go/repair.go
@@ -24,12 +24,12 @@ import (
 // argument *text*, not a parsed object. It is what a repair function reads
 // and what it returns.
 type RawToolCall struct {
-	ToolCallID       string          `json:"tool_call_id"`
-	ToolName         string          `json:"tool_name"`
+	ToolCallID       string          `json:"toolCallId"`
+	ToolName         string          `json:"toolName"`
 	Input            string          `json:"input"`
-	ProviderExecuted *bool           `json:"provider_executed,omitempty"`
+	ProviderExecuted *bool           `json:"providerExecuted,omitempty"`
 	Dynamic          *bool           `json:"dynamic,omitempty"`
-	ProviderMetadata json.RawMessage `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage `json:"providerMetadata,omitempty"`
 }
 
 // ToolCallRepairContext is the argument handed to a RepairToolCallFunc — the
@@ -39,13 +39,13 @@ type RawToolCall struct {
 // engine from the prompt and options of the call being repaired.
 type ToolCallRepairContext struct {
 	// ToolCall is the failed call, with its raw argument text.
-	ToolCall RawToolCall `json:"tool_call"`
-	// Error is the serialized AiMuxError the call carries (e.g.
-	// {"InvalidToolInput": {...}} or {"NoSuchTool": {...}}).
+	ToolCall RawToolCall `json:"toolCall"`
+	// Error is the serialized AiMuxError (AI SDK error JSON, tagged by "name" the call carries (e.g.
+	// {"name":"AI_InvalidToolInputError",...} or {"name":"AI_NoSuchToolError",...}).
 	Error json.RawMessage `json:"error"`
 	// InputSchema is the JSON schema of the named tool, or an empty-object
 	// schema when the name does not resolve to a tool.
-	InputSchema  json.RawMessage `json:"input_schema"`
+	InputSchema  json.RawMessage `json:"inputSchema"`
 	Tools        []Tool          `json:"tools"`
 	Messages     []ModelMessage  `json:"messages"`
 	Instructions *string         `json:"instructions"`
@@ -106,7 +106,7 @@ func applyToolCallRepairJSON(toolCallJSON, optsJSON, replyJSON string) (string,
 }
 
 // applyToolCallRepairToResultJSON patches a whole result document: both
-// tool_calls and the matching response_messages tool-call part.
+// toolCalls and the matching responseMessages tool-call part.
 func applyToolCallRepairToResultJSON(resultJSON, optsJSON, toolCallID, replyJSON string) (string, error) {
 	cResult := C.CString(resultJSON)
 	defer C.free(unsafe.Pointer(cResult))
@@ -148,7 +148,7 @@ func repairReply(repair RepairToolCallFunc, contextJSON string) (string, error)
 	}
 	reply, err := json.Marshal(struct {
 		Type     string       `json:"type"`
-		ToolCall *RawToolCall `json:"tool_call"`
+		ToolCall *RawToolCall `json:"toolCall"`
 	}{"repaired", replacement})
 	if err != nil {
 		return "", fmt.Errorf("aimux: failed to encode repaired tool call: %w", err)
@@ -158,14 +158,14 @@ func repairReply(repair RepairToolCallFunc, contextJSON string) (string, error)
 
 // invalidToolCalls lists the invalid tool calls of a result document, in
 // order, as (id, JSON) pairs. A GenerateObjectResult carries the text result
-// under `raw`; a GenerateTextResult's own `raw` has no tool_calls, so the
-// same lookup covers both — and StreamTextResultAggregated, whose tool_calls
+// under `raw`; a GenerateTextResult's own `raw` has no toolCalls, so the
+// same lookup covers both — and StreamTextResultAggregated, whose toolCalls
 // are top-level.
 func invalidToolCalls(resultJSON string) [][2]string {
 	var doc struct {
-		ToolCalls []json.RawMessage `json:"tool_calls"`
+		ToolCalls []json.RawMessage `json:"toolCalls"`
 		Raw       struct {
-			ToolCalls []json.RawMessage `json:"tool_calls"`
+			ToolCalls []json.RawMessage `json:"toolCalls"`
 		} `json:"raw"`
 	}
 	if json.Unmarshal([]byte(resultJSON), &doc) != nil {
@@ -178,7 +178,7 @@ func invalidToolCalls(resultJSON string) [][2]string {
 	var invalid [][2]string
 	for _, call := range calls {
 		var head struct {
-			ToolCallID string `json:"tool_call_id"`
+			ToolCallID string `json:"toolCallId"`
 			Invalid    *bool  `json:"invalid"`
 		}
 		if json.Unmarshal(call, &head) != nil || head.Invalid == nil || !*head.Invalid {
@@ -217,7 +217,7 @@ func repairResultJSON(resultJSON, promptJSON, optsJSON string, repair RepairTool
 }
 
 // repairStreamToolCall is the streaming counterpart: it resolves one
-// {"ToolCall": …} stream part payload, returning the payload to deliver.
+// {"type":"tool-call",…} stream part, returning the payload to deliver.
 // A valid call, or one made without a tool set, passes through untouched.
 // Tool-input delta parts never reach here — they are forwarded immediately,
 // as the AI SDK does.
diff --git a/bindings/go/roundtrip_fields_test.go b/bindings/go/roundtrip_fields_test.go
index 61bdd1294..8fe7b808e 100644
--- a/bindings/go/roundtrip_fields_test.go
+++ b/bindings/go/roundtrip_fields_test.go
@@ -4,7 +4,7 @@
 // Go prompts are raw JSON, so the binding's job is to pass them through the
 // FFI into the Rust engine unchanged. This test proves the full pipe for the
 // extended-thinking input round-trip: an assistant reasoning part carrying a
-// `signature` is echoed as an Anthropic thinking block on the wire when
+// providerOptions.anthropic.signature is echoed as an Anthropic thinking block on the wire when
 // thinking is enabled.
 
 package aimux
@@ -30,12 +30,12 @@ func TestE2E_AnthropicEchoesReasoningSignature(t *testing.T) {
 	prompt := `[
 		{"role":"user","content":[{"type":"text","text":"hi"}]},
 		{"role":"assistant","content":[
-			{"type":"reasoning","text":"pondering","signature":"sig-go-1"},
+			{"type":"reasoning","text":"pondering","providerOptions":{"anthropic":{"signature":"sig-go-1"}}},
 			{"type":"text","text":"answer"}
 		]},
 		{"role":"user","content":[{"type":"text","text":"go on"}]}
 	]`
-	opts := `{"provider_options":{"anthropic":{"thinking":{"type":"enabled","budgetTokens":1024}}}}`
+	opts := `{"providerOptions":{"anthropic":{"thinking":{"type":"enabled","budgetTokens":1024}}}}`
 
 	if _, err := m.GenerateText(prompt, opts); err != nil {
 		t.Fatalf("GenerateText: %v", err)
diff --git a/bindings/go/roundtrip_test.go b/bindings/go/roundtrip_test.go
index 360e830d2..71e5bdf3a 100644
--- a/bindings/go/roundtrip_test.go
+++ b/bindings/go/roundtrip_test.go
@@ -1,9 +1,9 @@
 // Round-trip (de)serialization tests for the typed wire-format types.
 //
-// Mirrors Kotlin TypedModelRoundTripTest.kt: build a typed value, marshal to
-// JSON, unmarshal back, assert equality. These pin down wire-format field
-// names and tag spellings the typed wrappers rely on — without them, a rename
-// in aimux-core could silently break Go clients.
+// Build a typed value, marshal to JSON, unmarshal back, assert equality. These
+// pin down wire-format field names and tag spellings the typed wrappers rely
+// on (camelCase, the AI SDK JSON) — without them, a rename in aimux-core could
+// silently break Go clients.
 //
 // Also includes regression tests for the two bugs found in code review:
 //   - ToolChoiceTool JSON injection (now uses json.Marshal)
@@ -122,9 +122,9 @@ func TestToolCallRoundTrip(t *testing.T) {
 	if err != nil {
 		t.Fatalf("marshal failed: %v", err)
 	}
-	// Verify wire field names match Kotlin (snake_case).
+	// Verify wire field names (camelCase).
 	s := string(b)
-	for _, want := range []string{`"tool_call_id"`, `"tool_name"`, `"input"`, `"provider_executed"`, `"dynamic"`, `"provider_metadata"`} {
+	for _, want := range []string{`"toolCallId"`, `"toolName"`, `"input"`, `"providerExecuted"`, `"dynamic"`, `"providerMetadata"`} {
 		if !contains(s, want) {
 			t.Errorf("expected %s in wire JSON, got %s", want, s)
 		}
@@ -134,19 +134,19 @@ func TestToolCallRoundTrip(t *testing.T) {
 		t.Fatalf("unmarshal failed: %v", err)
 	}
 	if decoded.ToolCallID != original.ToolCallID {
-		t.Errorf("tool_call_id mismatch: got %s, want %s", decoded.ToolCallID, original.ToolCallID)
+		t.Errorf("toolCallId mismatch: got %s, want %s", decoded.ToolCallID, original.ToolCallID)
 	}
 	if decoded.ToolName != original.ToolName {
-		t.Errorf("tool_name mismatch: got %s, want %s", decoded.ToolName, original.ToolName)
+		t.Errorf("toolName mismatch: got %s, want %s", decoded.ToolName, original.ToolName)
 	}
 	if decoded.ProviderExecuted == nil || *decoded.ProviderExecuted != true {
-		t.Error("provider_executed did not round-trip")
+		t.Error("providerExecuted did not round-trip")
 	}
 	if decoded.Dynamic == nil || *decoded.Dynamic != false {
 		t.Error("dynamic did not round-trip")
 	}
 	if string(decoded.ProviderMetadata) != string(original.ProviderMetadata) {
-		t.Errorf("provider_metadata mismatch: got %s, want %s", decoded.ProviderMetadata, original.ProviderMetadata)
+		t.Errorf("providerMetadata mismatch: got %s, want %s", decoded.ProviderMetadata, original.ProviderMetadata)
 	}
 }
 
@@ -179,9 +179,10 @@ func TestModelMessageMultiPartRoundTrip(t *testing.T) {
 		Role: RoleTool,
 		Content: []map[string]any{
 			{
-				"type":         "tool_result",
-				"tool_call_id": "call_1",
-				"result":       map[string]any{"temp": "20"},
+				"type":       "tool-result",
+				"toolCallId": "call_1",
+				"toolName":   "get_weather",
+				"output":     map[string]any{"type": "json", "value": map[string]any{"temp": "20"}},
 			},
 		},
 	}
@@ -203,14 +204,14 @@ func TestModelMessageMultiPartRoundTrip(t *testing.T) {
 	if len(raw.Content) != 1 {
 		t.Fatalf("expected 1 content part, got %d", len(raw.Content))
 	}
-	// Verify the tool_result structure.
+	// Verify the tool-result structure.
 	var part map[string]any
 	json.Unmarshal(raw.Content[0], &part)
-	if part["type"] != "tool_result" {
-		t.Errorf("expected type 'tool_result', got %v", part["type"])
+	if part["type"] != "tool-result" {
+		t.Errorf("expected type 'tool-result', got %v", part["type"])
 	}
-	if part["tool_call_id"] != "call_1" {
-		t.Errorf("expected tool_call_id 'call_1', got %v", part["tool_call_id"])
+	if part["toolCallId"] != "call_1" {
+		t.Errorf("expected toolCallId 'call_1', got %v", part["toolCallId"])
 	}
 }
 
@@ -239,11 +240,11 @@ func TestGenerateTextOptionsRoundTrip(t *testing.T) {
 	if err != nil {
 		t.Fatalf("marshal failed: %v", err)
 	}
-	// Verify snake_case wire fields present.
+	// Verify camelCase wire fields present.
 	s := string(b)
 	for _, want := range []string{
-		`"temperature"`, `"top_k"`, `"reasoning"`, `"instructions"`,
-		`"tool_choice"`, `"tools"`, `"input_schema"`,
+		`"temperature"`, `"topK"`, `"reasoning"`, `"instructions"`,
+		`"toolChoice"`, `"tools"`, `"inputSchema"`,
 	} {
 		if !contains(s, want) {
 			t.Errorf("expected %s in wire JSON, got %s", want, s)
@@ -275,22 +276,22 @@ func TestGenerateTextOptionsDefaultOmitsFields(t *testing.T) {
 	}
 }
 
-// RFC-0016 M2: include_raw_chunks round-trips through the typed options.
+// includeRawChunks round-trips through the typed options.
 func TestGenerateTextOptionsIncludeRawChunksRoundTrip(t *testing.T) {
 	original := GenerateTextOptions{IncludeRawChunks: boolPtr(true)}
 	b, err := json.Marshal(original)
 	if err != nil {
 		t.Fatalf("marshal: %v", err)
 	}
-	if !strings.Contains(string(b), `"include_raw_chunks":true`) {
-		t.Errorf("expected include_raw_chunks:true in %s", b)
+	if !strings.Contains(string(b), `"includeRawChunks":true`) {
+		t.Errorf("expected includeRawChunks:true in %s", b)
 	}
 	var back GenerateTextOptions
 	if err := json.Unmarshal(b, &back); err != nil {
 		t.Fatalf("unmarshal: %v", err)
 	}
 	if back.IncludeRawChunks == nil || *back.IncludeRawChunks != true {
-		t.Errorf("round-trip lost include_raw_chunks: %#v", back.IncludeRawChunks)
+		t.Errorf("round-trip lost includeRawChunks: %#v", back.IncludeRawChunks)
 	}
 }
 
@@ -318,11 +319,11 @@ func TestStreamPartParseRoundTrip(t *testing.T) {
 	cases := []struct {
 		name string
 		json string
-		tag  string
+		typ  string
 	}{
-		{"TextDelta", `{"TextDelta":{"id":"tx1","delta":"Hello"}}`, "TextDelta"},
-		{"Finish", `{"Finish":{"finish_reason":{"unified":{"stop":null}}}}`, "Finish"},
-		{"StreamStart", `{"StreamStart":{"warnings":[]}}`, "StreamStart"},
+		{"text-delta", `{"type":"text-delta","id":"tx1","delta":"Hello"}`, "text-delta"},
+		{"finish", `{"type":"finish","finishReason":{"unified":"stop"},"usage":{"inputTokens":{},"outputTokens":{}}}`, "finish"},
+		{"stream-start", `{"type":"stream-start","warnings":[]}`, "stream-start"},
 	}
 	for _, c := range cases {
 		t.Run(c.name, func(t *testing.T) {
@@ -330,34 +331,40 @@ func TestStreamPartParseRoundTrip(t *testing.T) {
 			if err != nil {
 				t.Fatalf("parse failed: %v", err)
 			}
-			if sp.Tag != c.tag {
-				t.Errorf("tag mismatch: got %s, want %s", sp.Tag, c.tag)
+			if sp.Type != c.typ {
+				t.Errorf("type mismatch: got %s, want %s", sp.Type, c.typ)
 			}
-			if len(sp.Payload) == 0 {
-				t.Error("expected non-empty payload")
+			if string(sp.Raw) != c.json {
+				t.Errorf("raw mismatch: got %s", sp.Raw)
 			}
 		})
 	}
 }
 
 func TestParseTextDeltaPayload(t *testing.T) {
-	sp, _ := ParseStreamPart(`{"TextDelta":{"id":"tx1","delta":"Hello"}}`)
+	sp, _ := ParseStreamPart(`{"type":"text-delta","id":"tx1","delta":"Hello"}`)
 	var td TextDeltaPayload
-	json.Unmarshal(sp.Payload, &td)
+	json.Unmarshal(sp.Raw, &td)
 	if td.ID != "tx1" || td.Delta != "Hello" {
 		t.Errorf("got %+v", td)
 	}
 }
 
-// TestParseStreamPartRejectsInvalid validates the externally-tagged union
-// enforcement: zero keys or multiple keys must produce an error.
+func TestStreamPartSource(t *testing.T) {
+	sp, _ := ParseStreamPart(`{"type":"source","sourceType":"url","id":"s1","url":"https://example.com"}`)
+	src, err := sp.Source()
+	if err != nil || src.URL == nil || src.URL.URL != "https://example.com" {
+		t.Fatalf("source: %+v, %v", src, err)
+	}
+}
+
+// TestParseStreamPartRejectsInvalid: a part without a "type" is an error.
 func TestParseStreamPartRejectsInvalid(t *testing.T) {
 	cases := []struct {
 		name string
 		json string
 	}{
 		{"empty_object", `{}`},
-		{"multiple_keys", `{"TextDelta":{"delta":"a"},"Finish":{}}`},
 		{"not_object", `[]`},
 		{"invalid_json", `{`},
 	}
diff --git a/bindings/go/trace_test.go b/bindings/go/trace_test.go
index 6e0a05fa6..1e9a26a29 100644
--- a/bindings/go/trace_test.go
+++ b/bindings/go/trace_test.go
@@ -1,7 +1,7 @@
 // trace_test.go — RFC-0015 wire-contract tests (pure JSON, no FFI).
 //
 // Verifies Go can parse the probe wire types exactly as Rust serializes them
-// (snake_case fields, hex fingerprints, optional verdicts).
+// (camelCase fields, hex fingerprints, optional verdicts).
 
 package aimux
 
@@ -13,34 +13,34 @@ import (
 const traceRecordJSON = `{
   "provider": "openai",
   "model": "gpt-4o",
-  "request_id": "req-1",
-  "session_id": "sess-1",
-  "call_id": "trace-1",
-  "sent_at_unix_ms": 1785900000000,
-  "ttft_ms": 42,
+  "requestId": "req-1",
+  "sessionId": "sess-1",
+  "callId": "trace-1",
+  "sentAtUnixMs": 1785900000000,
+  "ttftMs": 42,
   "fingerprint": {
-    "body_hash": "0123456789abcdef0123456789abcdef",
-    "len_bytes": 10240,
-    "block_size": 4096,
-    "block_hashes": ["aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb"],
-    "token_estimate": 2560
+    "bodyHash": "0123456789abcdef0123456789abcdef",
+    "lenBytes": 10240,
+    "blockSize": 4096,
+    "blockHashes": ["aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb"],
+    "tokenEstimate": 2560
   },
   "usage": {
-    "input_total": 2560,
-    "input_no_cache": 1536,
-    "cache_read": 1024,
-    "cache_write": 0,
-    "output_total": 10,
-    "output_text": 10
+    "inputTotal": 2560,
+    "inputNoCache": 1536,
+    "cacheRead": 1024,
+    "cacheWrite": 0,
+    "outputTotal": 10,
+    "outputText": 10
   },
-  "request_cache_hints": {"requested_write": true},
+  "requestCacheHints": {"requestedWrite": true},
   "verdict": {
     "kind": "Trusted",
     "confidence": "High",
     "violated": [],
-    "expected_max": 1024,
+    "expectedMax": 1024,
     "claimed": 1024,
-    "lcp_bytes": 4096,
+    "lcpBytes": 4096,
     "notes": []
   }
 }`
@@ -54,19 +54,19 @@ func TestTraceRecordParses(t *testing.T) {
 		t.Errorf("provider/model mismatch: %s / %s", rec.Provider, rec.Model)
 	}
 	if rec.SessionID == nil || *rec.SessionID != "sess-1" {
-		t.Errorf("session_id mismatch: %v", rec.SessionID)
+		t.Errorf("sessionId mismatch: %v", rec.SessionID)
 	}
 	if rec.Fingerprint.BodyHash != "0123456789abcdef0123456789abcdef" {
-		t.Errorf("body_hash mismatch: %s", rec.Fingerprint.BodyHash)
+		t.Errorf("bodyHash mismatch: %s", rec.Fingerprint.BodyHash)
 	}
 	if len(rec.Fingerprint.BlockHashes) != 2 {
 		t.Errorf("expected 2 block hashes, got %d", len(rec.Fingerprint.BlockHashes))
 	}
 	if rec.Usage.CacheRead == nil || *rec.Usage.CacheRead != 1024 {
-		t.Errorf("cache_read mismatch: %v", rec.Usage.CacheRead)
+		t.Errorf("cacheRead mismatch: %v", rec.Usage.CacheRead)
 	}
 	if rec.RequestCacheHints == nil || !rec.RequestCacheHints.RequestedWrite {
-		t.Error("request_cache_hints mismatch")
+		t.Error("requestCacheHints mismatch")
 	}
 	if rec.Verdict == nil {
 		t.Fatal("verdict must be present")
@@ -85,11 +85,11 @@ func TestTraceRecordParses(t *testing.T) {
 func TestTraceStatsAndChainParse(t *testing.T) {
 	statsJSON := `{
 	  "provider": "openai", "model": "gpt-4o",
-	  "requests": 3, "input_tokens_total": 7680,
-	  "claimed_cache_read_total": 2048, "claimed_cache_write_total": 0,
-	  "reported_hit_rate": 0.26666666666666666,
-	  "client_upper_bound_hit_rate": 0.4,
-	  "verdict_counts": {"Trusted": 2, "SuspectOverclaim": 1},
+	  "requests": 3, "inputTokensTotal": 7680,
+	  "claimedCacheReadTotal": 2048, "claimedCacheWriteTotal": 0,
+	  "reportedHitRate": 0.26666666666666666,
+	  "clientUpperBoundHitRate": 0.4,
+	  "verdictCounts": {"Trusted": 2, "SuspectOverclaim": 1},
 	  "errors": 0
 	}`
 	var stats TraceStats
@@ -100,16 +100,16 @@ func TestTraceStatsAndChainParse(t *testing.T) {
 		t.Errorf("stats mismatch: %+v", stats)
 	}
 	if stats.VerdictCounts["Trusted"] != 2 {
-		t.Errorf("verdict_counts mismatch: %v", stats.VerdictCounts)
+		t.Errorf("verdictCounts mismatch: %v", stats.VerdictCounts)
 	}
 
 	chainJSON := `{
-	  "session_id": "sess-1",
-	  "record_ids": ["trace-1", "trace-2"],
-	  "prefix_stability": 0.8,
+	  "sessionId": "sess-1",
+	  "recordIds": ["trace-1", "trace-2"],
+	  "prefixStability": 0.8,
 	  "breaks": [{
-	    "at_record_id": "trace-2", "prev_record_id": "trace-1",
-	    "lcp_bytes": 1024, "expected_break": false, "kind": "Unknown"
+	    "atRecordId": "trace-2", "prevRecordId": "trace-1",
+	    "lcpBytes": 1024, "expectedBreak": false, "kind": "Unknown"
 	  }]
 	}`
 	var chain SessionChainView
diff --git a/bindings/go/typed.go b/bindings/go/typed.go
index d0d3f6d33..8298b31a6 100644
--- a/bindings/go/typed.go
+++ b/bindings/go/typed.go
@@ -71,7 +71,7 @@ func (m *Model) Generate(prompt any, opts *GenerateTextOptions) (*GenerateTextRe
 // GenerateObj is the typed object-generation method (M12, RFC-0016).
 //
 // Same signature as Generate; returns a typed *GenerateObjectResult. Pass
-// response_format: { "Json": { ... } } via opts for schema control; the
+// ResponseFormat {"type":"json","schema":{...}} via opts for schema control; the
 // engine applies JSON repair before parsing.
 func (m *Model) GenerateObj(prompt any, opts *GenerateTextOptions) (*GenerateObjectResult, error) {
 	promptJSON, err := marshalPrompt(prompt)
@@ -144,9 +144,9 @@ func (s *TypedStream) Cancel() { s.raw.Cancel() }
 //
 //	stream, err := model.Stream("Write a haiku", nil)
 //	for part := range stream.Parts() {
-//	    if part.Tag == "TextDelta" {
+//	    if part.Type == "text-delta" {
 //	        var td TextDeltaPayload
-//	        json.Unmarshal(part.Payload, &td)
+//	        json.Unmarshal(part.Raw, &td)
 //	        fmt.Print(td.Delta)
 //	    }
 //	}
@@ -198,14 +198,14 @@ func (m *Model) StreamContext(
 			}
 			// Only a complete tool call can be repaired; input deltas are
 			// forwarded untouched, as the AI SDK does (RFC-0035 §4).
-			if sp.Tag == "ToolCall" && opts != nil && opts.RepairToolCall != nil {
-				payload, err := repairStreamToolCall(sp.Payload, promptJSON, optsJSON, opts.RepairToolCall)
+			if sp.Type == "tool-call" && opts != nil && opts.RepairToolCall != nil {
+				raw, err := repairStreamToolCall(sp.Raw, promptJSON, optsJSON, opts.RepairToolCall)
 				if err != nil {
 					ts.err = err
 					rawStream.cancel(err)
 					return
 				}
-				sp.Payload = payload
+				sp.Raw = raw
 			}
 			select {
 			case ts.parts <- sp:
diff --git a/bindings/go/typed_test.go b/bindings/go/typed_test.go
index 2852791cb..2d08f5807 100644
--- a/bindings/go/typed_test.go
+++ b/bindings/go/typed_test.go
@@ -18,7 +18,7 @@ func TestTypedGenerateString(t *testing.T) {
 	defer srv.Close()
 	srv.SetResponse(plainOpenAIResponse)
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	// Typed call with plain string prompt.
@@ -36,7 +36,7 @@ func TestTypedGenerateMessages(t *testing.T) {
 	defer srv.Close()
 	srv.SetResponse(plainOpenAIResponse)
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	msgs := []ModelMessage{
@@ -65,7 +65,7 @@ func TestTypedGenerateWithOptions(t *testing.T) {
 	defer srv.Close()
 	srv.SetResponse(toolCallOpenAIResponse)
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	temp := 0.7
@@ -111,7 +111,7 @@ func TestTypedStreamTextDeltas(t *testing.T) {
 	srv.SetContentType("text/event-stream")
 	srv.SetResponse(buildTextDeltaSSE())
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	stream, err := m.Stream("Say hello", nil)
@@ -122,12 +122,12 @@ func TestTypedStreamTextDeltas(t *testing.T) {
 	var textBuilder string
 	var gotFinish bool
 	for part := range stream.Parts() {
-		switch part.Tag {
-		case "TextDelta":
+		switch part.Type {
+		case "text-delta":
 			var td TextDeltaPayload
-			json.Unmarshal(part.Payload, &td)
+			json.Unmarshal(part.Raw, &td)
 			textBuilder += td.Delta
-		case "Finish":
+		case "finish":
 			gotFinish = true
 		}
 	}
@@ -148,7 +148,7 @@ func TestTypedStreamMessages(t *testing.T) {
 	srv.SetContentType("text/event-stream")
 	srv.SetResponse(buildTextDeltaSSE())
 
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	msgs := []ModelMessage{
@@ -207,7 +207,7 @@ func TestTypedStreamContextAlreadyCanceled(t *testing.T) {
 func TestTypedStreamCancelUnblocksFullPartsChannel(t *testing.T) {
 	srv, allSent, requestDone := newBackpressureSSEServer(1200)
 	defer srv.Close()
-	m := OpenAIWithBase("sk-test-fake-key", "gpt-4o", srv.URL)
+	m := chatModel(t, srv.URL)
 	defer m.Close()
 
 	stream, err := m.Stream("fill the typed channel", nil)
diff --git a/bindings/go/types.go b/bindings/go/types.go
index 0b919cf6d..77cf2c6ee 100644
--- a/bindings/go/types.go
+++ b/bindings/go/types.go
@@ -1,12 +1,13 @@
-// Typed data structures mirroring the aimux-core wire format (same shapes as
-// the Kotlin Types.kt and the generated TypeScript .d.ts types).
+// Typed data structures mirroring the aimux-core wire format: the JSON the
+// Rust engine reads and writes, which is the AI SDK JSON (camelCase field
+// names, unions tagged by a "type" key, absent rather than null for optional
+// fields). The generated TypeScript types in bindings/node/src/types are the
+// exact shape.
 //
-// Field names use JSON tags matching the wire format's snake_case. The raw JSON
-// boundary is handled by Model.GenerateText / Model.StreamText — this layer
-// only provides typed parsing so callers don't manually dig through JSON.
-//
-// These types are intentionally lenient on decode (unknown keys ignored, every
-// field optional) so future engine additions don't break existing clients.
+// The raw JSON boundary is handled by Model.GenerateText / Model.StreamText —
+// this layer only provides typed parsing so callers don't manually dig through
+// JSON. Decoding is lenient (unknown keys ignored) so future engine additions
+// don't break existing clients.
 
 package aimux
 
@@ -38,7 +39,6 @@ const (
 )
 
 // ReasoningEffort controls how much reasoning effort the model spends.
-// Mirrors Kotlin ReasoningEffort (Types.kt:72-81).
 type ReasoningEffort string
 
 const (
@@ -87,27 +87,27 @@ type URLSource struct {
 	ID               string          `json:"id"`
 	URL              string          `json:"url"`
 	Title            *string         `json:"title,omitempty"`
-	ProviderMetadata json.RawMessage `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage `json:"providerMetadata,omitempty"`
 }
 
 type DocumentSource struct {
 	ID               string          `json:"id"`
-	MediaType        string          `json:"media_type"`
+	MediaType        string          `json:"mediaType"`
 	Title            string          `json:"title"`
 	Filename         *string         `json:"filename,omitempty"`
-	ProviderMetadata json.RawMessage `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage `json:"providerMetadata,omitempty"`
 }
 
 func (s Source) MarshalJSON() ([]byte, error) {
 	if s.URL != nil && s.Document == nil {
 		return json.Marshal(struct {
-			SourceType string `json:"source_type"`
+			SourceType string `json:"sourceType"`
 			*URLSource
 		}{"url", s.URL})
 	}
 	if s.Document != nil && s.URL == nil {
 		return json.Marshal(struct {
-			SourceType string `json:"source_type"`
+			SourceType string `json:"sourceType"`
 			*DocumentSource
 		}{"document", s.Document})
 	}
@@ -120,17 +120,17 @@ func (s *Source) UnmarshalJSON(data []byte) error {
 		return err
 	}
 	var sourceType string
-	if err := json.Unmarshal(fields["source_type"], &sourceType); err != nil {
-		return fmt.Errorf("aimux: invalid source_type: %w", err)
+	if err := json.Unmarshal(fields["sourceType"], &sourceType); err != nil {
+		return fmt.Errorf("aimux: invalid sourceType: %w", err)
 	}
 	required := []string{"id"}
 	switch sourceType {
 	case "url":
 		required = append(required, "url")
 	case "document":
-		required = append(required, "media_type", "title")
+		required = append(required, "mediaType", "title")
 	default:
-		return fmt.Errorf("aimux: unknown source_type %q", sourceType)
+		return fmt.Errorf("aimux: unknown sourceType %q", sourceType)
 	}
 	for _, name := range required {
 		var value *string
@@ -157,9 +157,9 @@ func (s *Source) UnmarshalJSON(data []byte) error {
 // InputTokenUsage is input token usage detail with cache breakdown.
 type InputTokenUsage struct {
 	Total      *uint32 `json:"total,omitempty"`
-	NoCache    *uint32 `json:"no_cache,omitempty"`
-	CacheRead  *uint32 `json:"cache_read,omitempty"`
-	CacheWrite *uint32 `json:"cache_write,omitempty"`
+	NoCache    *uint32 `json:"noCache,omitempty"`
+	CacheRead  *uint32 `json:"cacheRead,omitempty"`
+	CacheWrite *uint32 `json:"cacheWrite,omitempty"`
 }
 
 // OutputTokenUsage is output token usage detail.
@@ -171,8 +171,8 @@ type OutputTokenUsage struct {
 
 // Usage is token usage statistics.
 type Usage struct {
-	InputTokens  InputTokenUsage            `json:"input_tokens,omitempty"`
-	OutputTokens OutputTokenUsage           `json:"output_tokens,omitempty"`
+	InputTokens  InputTokenUsage            `json:"inputTokens,omitempty"`
+	OutputTokens OutputTokenUsage           `json:"outputTokens,omitempty"`
 	Raw          map[string]json.RawMessage `json:"raw,omitempty"`
 }
 
@@ -183,123 +183,120 @@ type FinishReason struct {
 }
 
 // ToolCall represents a tool call requested by the model.
-// Mirrors Kotlin ToolCall (Types.kt:157-164).
 type ToolCall struct {
-	ToolCallID       string          `json:"tool_call_id"`
-	ToolName         string          `json:"tool_name"`
+	ToolCallID       string          `json:"toolCallId"`
+	ToolName         string          `json:"toolName"`
 	Input            json.RawMessage `json:"input,omitempty"`
-	ProviderExecuted *bool           `json:"provider_executed,omitempty"`
+	ProviderExecuted *bool           `json:"providerExecuted,omitempty"`
 	Dynamic          *bool           `json:"dynamic,omitempty"`
 	// ProviderMetadata carries provider-specific data associated with this call.
-	ProviderMetadata json.RawMessage `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage `json:"providerMetadata,omitempty"`
 	// Invalid is set by Core when the tool call stays invalid after optional repair.
 	Invalid *bool `json:"invalid,omitempty"`
 	// Error is the typed lookup, parse, schema, or repair failure for an invalid call.
 	Error json.RawMessage `json:"error,omitempty"`
 }
 
-// ContentPart is a single content part in the raw response.
-// The wire format uses internally-tagged unions; we keep it as raw JSON
-// for forward compatibility.
+// ContentPart is a single content part, in a prompt message or a result: an
+// object tagged by "type" ("text", "tool-call", "tool-result", "source", ...).
+// It is kept as raw JSON for forward compatibility. A tool-result part is
+// {"type":"tool-result","toolCallId":...,"toolName":...,"output":{...}} where
+// output is {"type":"text"|"json"|"error-text"|"error-json","value":...},
+// {"type":"execution-denied","reason":...} or {"type":"content","value":[...]}.
 type ContentPart = json.RawMessage
 
-// ResponseMetadata describes the provider HTTP response.
-// Mirrors Kotlin ResponseMetadata (Types.kt:141-146).
+// ResponseMetadata describes the provider response.
 type ResponseMetadata struct {
 	ID        *string `json:"id,omitempty"`
 	Timestamp *string `json:"timestamp,omitempty"`
-	ModelID   *string `json:"model_id,omitempty"`
+	ModelID   *string `json:"modelId,omitempty"`
 }
 
-// GenerateResponseMetadata is response metadata with optional HTTP headers and body.
-type GenerateResponseMetadata struct {
+// ResponseInfo is response metadata with optional HTTP headers and body.
+type ResponseInfo struct {
 	ResponseMetadata
 	Headers map[string]string `json:"headers,omitempty"`
 	Body    json.RawMessage   `json:"body,omitempty"`
 }
 
-// GenerateRequest is provider request information with an optional HTTP body.
-type GenerateRequest struct {
+// RequestInfo is provider request information with an optional HTTP body.
+type RequestInfo struct {
 	Body json.RawMessage `json:"body,omitempty"`
 }
 
 // GenerateResult is the raw provider result.
-// Mirrors Kotlin GenerateResult (Types.kt:853-871).
 type GenerateResult struct {
-	Content          []ContentPart             `json:"content,omitempty"`
-	FinishReason     FinishReason              `json:"finish_reason,omitempty"`
-	Usage            Usage                     `json:"usage,omitempty"`
-	Warnings         []json.RawMessage         `json:"warnings,omitempty"`
-	ProviderMetadata json.RawMessage           `json:"provider_metadata,omitempty"`
-	Response         *GenerateResponseMetadata `json:"response,omitempty"`
-	Request          *GenerateRequest          `json:"request,omitempty"`
+	Content          []ContentPart     `json:"content,omitempty"`
+	FinishReason     FinishReason      `json:"finishReason,omitempty"`
+	Usage            Usage             `json:"usage,omitempty"`
+	Warnings         []json.RawMessage `json:"warnings,omitempty"`
+	ProviderMetadata json.RawMessage   `json:"providerMetadata,omitempty"`
+	Response         *ResponseInfo     `json:"response,omitempty"`
+	Request          *RequestInfo      `json:"request,omitempty"`
 }
 
 // GenerateTextResult is the typed result of a GenerateText call.
-// Mirrors Kotlin GenerateTextResult (Types.kt:878-886).
 type GenerateTextResult struct {
 	Content          []ContentPart     `json:"content,omitempty"`
 	Text             string            `json:"text"`
-	ToolCalls        []ToolCall        `json:"tool_calls,omitempty"`
-	FinishReason     FinishReason      `json:"finish_reason,omitempty"`
+	ToolCalls        []ToolCall        `json:"toolCalls,omitempty"`
+	FinishReason     FinishReason      `json:"finishReason,omitempty"`
 	Usage            Usage             `json:"usage,omitempty"`
 	Warnings         []json.RawMessage `json:"warnings,omitempty"`
 	Raw              GenerateResult    `json:"raw"`
 	Reasoning        []json.RawMessage `json:"reasoning,omitempty"`
-	ReasoningText    string            `json:"reasoning_text,omitempty"`
+	ReasoningText    string            `json:"reasoningText,omitempty"`
 	Sources          []Source          `json:"sources,omitempty"`
 	Files            []json.RawMessage `json:"files,omitempty"`
-	ResponseMessages []ModelMessage    `json:"response_messages,omitempty"`
-	// RawFinishReason is the raw provider-specific finish reason string (M12).
-	RawFinishReason *string `json:"raw_finish_reason,omitempty"`
+	ResponseMessages []ModelMessage    `json:"responseMessages,omitempty"`
+	// RawFinishReason is the raw provider-specific finish reason string.
+	RawFinishReason *string `json:"rawFinishReason,omitempty"`
 	// ProviderMetadata is provider-specific metadata (e.g. Anthropic cache info).
-	// Mirrored from raw.provider_metadata for top-level convenience. Weak type
-	// (json.RawMessage) — same strategy as GenerateResult.ProviderMetadata.
-	ProviderMetadata json.RawMessage `json:"provider_metadata,omitempty"`
-	// Response is the response metadata (id, timestamp, model_id), mirrored from
-	// raw.response for top-level convenience.
-	Response ResponseMetadata `json:"response,omitempty"`
+	// Mirrored from raw.providerMetadata for top-level convenience.
+	ProviderMetadata json.RawMessage `json:"providerMetadata,omitempty"`
+	// Request is the request information (body) from the provider.
+	Request RequestInfo `json:"request"`
+	// Response is the response metadata, headers and body from the provider.
+	Response ResponseInfo `json:"response"`
 	// TotalUsage is total token usage across all steps. In single-step mode
 	// (aimux's default), equals Usage. Provided for AI SDK parity.
-	TotalUsage Usage `json:"total_usage,omitempty"`
+	TotalUsage Usage `json:"totalUsage,omitempty"`
 }
 
-// GenerateObjectResult is the typed result of a GenerateObject call (M12).
-// Mirrors Rust GenerateObjectResult. `object` is an arbitrary JSON value
-// (weak type — json.RawMessage) and `raw` is the full GenerateTextResult.
+// GenerateObjectResult is the typed result of a GenerateObject call.
+// `object` is an arbitrary JSON value and `raw` is the full GenerateTextResult.
 type GenerateObjectResult struct {
 	Object           json.RawMessage    `json:"object,omitempty"`
-	FinishReason     FinishReason       `json:"finish_reason,omitempty"`
-	RawFinishReason  *string            `json:"raw_finish_reason,omitempty"`
+	FinishReason     FinishReason       `json:"finishReason,omitempty"`
+	RawFinishReason  *string            `json:"rawFinishReason,omitempty"`
 	Usage            Usage              `json:"usage,omitempty"`
 	Warnings         []json.RawMessage  `json:"warnings,omitempty"`
 	Reasoning        *string            `json:"reasoning,omitempty"`
-	ProviderMetadata json.RawMessage    `json:"provider_metadata,omitempty"`
+	ProviderMetadata json.RawMessage    `json:"providerMetadata,omitempty"`
 	Response         ResponseMetadata   `json:"response,omitempty"`
 	Raw              GenerateTextResult `json:"raw"`
 }
 
-// StreamTextResultAggregated is the aggregated result of stream_text (M11).
-// Mirrors Rust StreamTextResultAggregated. reasoning/sources/files use weak
-// types (json.RawMessage) — same strategy as GenerateTextResult.
+// StreamTextResultAggregated is the aggregated result of a consumed stream.
 type StreamTextResultAggregated struct {
 	Content          []ContentPart     `json:"content,omitempty"`
 	Text             string            `json:"text"`
 	Reasoning        []json.RawMessage `json:"reasoning,omitempty"`
-	ReasoningText    string            `json:"reasoning_text,omitempty"`
-	ToolCalls        []ToolCall        `json:"tool_calls,omitempty"`
+	ReasoningText    string            `json:"reasoningText,omitempty"`
+	ToolCalls        []ToolCall        `json:"toolCalls,omitempty"`
 	Sources          []Source          `json:"sources,omitempty"`
 	Files            []json.RawMessage `json:"files,omitempty"`
-	FinishReason     FinishReason      `json:"finish_reason,omitempty"`
-	RawFinishReason  *string           `json:"raw_finish_reason,omitempty"`
+	FinishReason     FinishReason      `json:"finishReason,omitempty"`
+	RawFinishReason  *string           `json:"rawFinishReason,omitempty"`
 	Usage            Usage             `json:"usage,omitempty"`
-	TotalUsage       Usage             `json:"total_usage,omitempty"`
+	TotalUsage       Usage             `json:"totalUsage,omitempty"`
 	Warnings         []json.RawMessage `json:"warnings,omitempty"`
-	ProviderMetadata json.RawMessage   `json:"provider_metadata,omitempty"`
-	// Response is the response metadata (id, timestamp, model_id) if emitted by
-	// the stream. Nullable (streaming may have none).
-	Response         *ResponseMetadata `json:"response,omitempty"`
-	ResponseMessages []ModelMessage    `json:"response_messages,omitempty"`
+	ProviderMetadata json.RawMessage   `json:"providerMetadata,omitempty"`
+	// Request is the request information from the provider.
+	Request RequestInfo `json:"request"`
+	// Response is the response metadata and headers from the last step.
+	Response         ResponseInfo   `json:"response"`
+	ResponseMessages []ModelMessage `json:"responseMessages,omitempty"`
 }
 
 // ParseGenerateTextResult parses the JSON string returned by Model.GenerateText
@@ -313,7 +310,7 @@ func ParseGenerateTextResult(jsonStr string) (*GenerateTextResult, error) {
 }
 
 // ParseGenerateObjectResult parses the JSON string returned by
-// Model.GenerateObject into a typed GenerateObjectResult (M12).
+// Model.GenerateObject into a typed GenerateObjectResult.
 func ParseGenerateObjectResult(jsonStr string) (*GenerateObjectResult, error) {
 	var r GenerateObjectResult
 	if err := json.Unmarshal([]byte(jsonStr), &r); err != nil {
@@ -323,7 +320,7 @@ func ParseGenerateObjectResult(jsonStr string) (*GenerateObjectResult, error) {
 }
 
 // ParseStreamTextResultAggregated parses the JSON string returned by
-// Model.ConsumeStreamText into a typed StreamTextResultAggregated (M11).
+// Model.ConsumeStreamText into a typed StreamTextResultAggregated.
 func ParseStreamTextResultAggregated(jsonStr string) (*StreamTextResultAggregated, error) {
 	var r StreamTextResultAggregated
 	if err := json.Unmarshal([]byte(jsonStr), &r); err != nil {
@@ -336,8 +333,7 @@ func ParseStreamTextResultAggregated(jsonStr string) (*StreamTextResultAggregate
 
 // ModelMessage is a single message in a conversation.
 // Content is `any` so it can be a plain string (common case) or a slice of
-// typed ContentParts (multi-part content, e.g. tool results).
-// Mirrors Kotlin ModelMessage (Types.kt:523-548).
+// ContentParts (multi-part content, e.g. tool results).
 type ModelMessage struct {
 	Role    Role `json:"role"`
 	Content any  `json:"content"`
@@ -360,37 +356,35 @@ func MarshalMessages(msgs []ModelMessage) (string, error) {
 // ── GenerateTextOptions (for typed options building) ──────────────────────────
 
 // GenerateTextOptions is the typed options for text generation.
-// All fields are optional (pointer types) to match the engine's schema.
-// Mirrors Kotlin GenerateTextOptions (Types.kt:561-579).
-// TimeoutConfiguration sets per-call timeouts (RFC-0016 H3).
+// TimeoutConfiguration sets per-call timeouts.
 type TimeoutConfiguration struct {
-	TotalMs      *uint64 `json:"total_ms,omitempty"`
-	StepMs       *uint64 `json:"step_ms,omitempty"`
-	FirstChunkMs *uint64 `json:"first_chunk_ms,omitempty"`
-	ChunkMs      *uint64 `json:"chunk_ms,omitempty"`
+	TotalMs      *uint64 `json:"totalMs,omitempty"`
+	StepMs       *uint64 `json:"stepMs,omitempty"`
+	FirstChunkMs *uint64 `json:"firstChunkMs,omitempty"`
+	ChunkMs      *uint64 `json:"chunkMs,omitempty"`
 }
 
-// GenerateTextOptions mirrors the shared wire options.
+// GenerateTextOptions mirrors the shared wire options. All fields are optional.
 type GenerateTextOptions struct {
-	MaxOutputTokens  *uint32               `json:"max_output_tokens,omitempty"`
+	MaxOutputTokens  *uint32               `json:"maxOutputTokens,omitempty"`
 	Temperature      *float64              `json:"temperature,omitempty"`
-	StopSequences    []string              `json:"stop_sequences,omitempty"`
-	TopP             *float64              `json:"top_p,omitempty"`
-	TopK             *float64              `json:"top_k,omitempty"`
-	PresencePenalty  *float64              `json:"presence_penalty,omitempty"`
-	FrequencyPenalty *float64              `json:"frequency_penalty,omitempty"`
-	ResponseFormat   json.RawMessage       `json:"response_format,omitempty"`
+	StopSequences    []string              `json:"stopSequences,omitempty"`
+	TopP             *float64              `json:"topP,omitempty"`
+	TopK             *float64              `json:"topK,omitempty"`
+	PresencePenalty  *float64              `json:"presencePenalty,omitempty"`
+	FrequencyPenalty *float64              `json:"frequencyPenalty,omitempty"`
+	ResponseFormat   json.RawMessage       `json:"responseFormat,omitempty"`
 	Seed             *uint64               `json:"seed,omitempty"`
 	Tools            []Tool                `json:"tools,omitempty"`
-	ToolChoice       ToolChoice            `json:"tool_choice,omitempty"`
+	ToolChoice       ToolChoice            `json:"toolChoice,omitempty"`
 	Headers          map[string]string     `json:"headers,omitempty"`
-	ProviderOptions  json.RawMessage       `json:"provider_options,omitempty"`
+	ProviderOptions  json.RawMessage       `json:"providerOptions,omitempty"`
 	Reasoning        *ReasoningEffort      `json:"reasoning,omitempty"`
 	Instructions     *string               `json:"instructions,omitempty"`
-	MaxRetries       *uint32               `json:"max_retries,omitempty"`
+	MaxRetries       *uint32               `json:"maxRetries,omitempty"`
 	Timeout          *TimeoutConfiguration `json:"timeout,omitempty"`
-	IncludeRawChunks *bool                 `json:"include_raw_chunks,omitempty"`
-	SessionID        *string               `json:"session_id,omitempty"`
+	IncludeRawChunks *bool                 `json:"includeRawChunks,omitempty"`
+	SessionID        *string               `json:"sessionId,omitempty"`
 	// RepairToolCall repairs tool calls the model got wrong (RFC-0035). It
 	// runs in Go, after generation, on every call the engine marked invalid —
 	// never serialized into the options sent to the engine.
@@ -406,16 +400,19 @@ type FunctionToolInputExample struct {
 	Input map[string]json.RawMessage `json:"input"`
 }
 
-// Tool is a function tool definition (the "function" variant).
-// Mirrors Kotlin Tool.Function (Types.kt:209-230).
+// Tool is a tool definition: a function tool (Type "function": Name,
+// InputSchema, ...) or a provider-defined tool (Type "provider": ID, Name, Args).
 type Tool struct {
-	Type            string                     `json:"type"` // always "function"
+	Type            string                     `json:"type"` // "function" or "provider"
 	Name            string                     `json:"name"`
 	Description     *string                    `json:"description,omitempty"`
-	InputSchema     json.RawMessage            `json:"input_schema,omitempty"`
+	InputSchema     json.RawMessage            `json:"inputSchema,omitempty"`
 	Strict          *bool                      `json:"strict,omitempty"`
-	ProviderOptions map[string]json.RawMessage `json:"provider_options,omitempty"`
-	InputExamples   []FunctionToolInputExample `json:"input_examples,omitempty"`
+	ProviderOptions map[string]json.RawMessage `json:"providerOptions,omitempty"`
+	InputExamples   []FunctionToolInputExample `json:"inputExamples,omitempty"`
+	// ID and Args are the provider tool's id ("<provider>.<tool>") and arguments.
+	ID   string                     `json:"id,omitempty"`
+	Args map[string]json.RawMessage `json:"args,omitempty"`
 }
 
 // MarshalOptions serializes GenerateTextOptions to JSON. Returns "" for nil opts.
@@ -432,45 +429,43 @@ func MarshalOptions(opts *GenerateTextOptions) (string, error) {
 
 // ── StreamPart parsing ─────────────────────────────────────────────────────
 
-// StreamPart is a parsed stream part. The wire format uses externally-tagged
-// JSON (e.g. {"TextDelta":{"delta":"..."}}). We decode the tag and payload.
+// StreamPart is a parsed stream part. The wire format is an object tagged by
+// its "type" key (e.g. {"type":"text-delta","id":"...","delta":"..."}).
 type StreamPart struct {
-	Tag     string          `json:"-"`
-	Payload json.RawMessage `json:"-"`
+	// Type is the part's "type" (e.g. "text-delta", "tool-call", "finish").
+	Type string `json:"-"`
+	// Raw is the whole part object, "type" key included. Decode it into the
+	// struct of your choice, e.g. TextDeltaPayload.
+	Raw json.RawMessage `json:"-"`
 }
 
 // ParseStreamPart parses a StreamPart JSON string.
 func ParseStreamPart(jsonStr string) (*StreamPart, error) {
-	var raw map[string]json.RawMessage
-	if err := json.Unmarshal([]byte(jsonStr), &raw); err != nil {
-		return nil, fmt.Errorf("aimux: failed to parse StreamPart: %w", err)
-	}
-	// Externally-tagged: the map must have exactly one key.
-	if len(raw) == 0 {
-		return nil, fmt.Errorf("aimux: StreamPart has no variant tag: %s", jsonStr)
+	var head struct {
+		Type string `json:"type"`
 	}
-	if len(raw) > 1 {
-		return nil, fmt.Errorf("aimux: StreamPart has multiple variant tags (%d), expected 1: %s", len(raw), jsonStr)
+	if err := json.Unmarshal([]byte(jsonStr), &head); err != nil {
+		return nil, fmt.Errorf("aimux: failed to parse StreamPart: %w", err)
 	}
-	for tag, payload := range raw {
-		return &StreamPart{Tag: tag, Payload: payload}, nil
+	if head.Type == "" {
+		return nil, fmt.Errorf("aimux: StreamPart has no type: %s", jsonStr)
 	}
-	return &StreamPart{}, nil
+	return &StreamPart{Type: head.Type, Raw: json.RawMessage(jsonStr)}, nil
 }
 
-// Source decodes the payload of a Source stream part.
+// Source decodes a "source" stream part.
 func (p *StreamPart) Source() (*Source, error) {
-	if p.Tag != "Source" {
-		return nil, fmt.Errorf("aimux: expected Source stream part, got %s", p.Tag)
+	if p.Type != "source" {
+		return nil, fmt.Errorf("aimux: expected source stream part, got %s", p.Type)
 	}
 	var source Source
-	if err := json.Unmarshal(p.Payload, &source); err != nil {
+	if err := json.Unmarshal(p.Raw, &source); err != nil {
 		return nil, err
 	}
 	return &source, nil
 }
 
-// TextDeltaPayload is the payload of a {"TextDelta":{...}} stream part.
+// TextDeltaPayload is a {"type":"text-delta",...} stream part.
 type TextDeltaPayload struct {
 	ID    string `json:"id,omitempty"`
 	Delta string `json:"delta"`
@@ -482,8 +477,8 @@ type TextDeltaPayload struct {
 type TraceFilter struct {
 	Provider    *string `json:"provider,omitempty"`
 	Model       *string `json:"model,omitempty"`
-	SessionID   *string `json:"session_id,omitempty"`
-	SinceUnixMs *int64  `json:"since_unix_ms,omitempty"`
+	SessionID   *string `json:"sessionId,omitempty"`
+	SinceUnixMs *int64  `json:"sinceUnixMs,omitempty"`
 }
 
 // TraceStats is one (provider, model) aggregation group.
@@ -491,14 +486,14 @@ type TraceStats struct {
 	Provider                string            `json:"provider"`
 	Model                   string            `json:"model"`
 	Requests                uint64            `json:"requests"`
-	InputTokensTotal        uint64            `json:"input_tokens_total"`
-	ClaimedCacheReadTotal   uint64            `json:"claimed_cache_read_total"`
-	ClaimedCacheWriteTotal  uint64            `json:"claimed_cache_write_total"`
-	ReportedHitRate         *float64          `json:"reported_hit_rate,omitempty"`
-	ClientUpperBoundHitRate *float64          `json:"client_upper_bound_hit_rate,omitempty"`
-	VerdictCounts           map[string]uint64 `json:"verdict_counts"`
-	TTFTP50Ms               *uint64           `json:"ttft_p50_ms,omitempty"`
-	TTFTP95Ms               *uint64           `json:"ttft_p95_ms,omitempty"`
+	InputTokensTotal        uint64            `json:"inputTokensTotal"`
+	ClaimedCacheReadTotal   uint64            `json:"claimedCacheReadTotal"`
+	ClaimedCacheWriteTotal  uint64            `json:"claimedCacheWriteTotal"`
+	ReportedHitRate         *float64          `json:"reportedHitRate,omitempty"`
+	ClientUpperBoundHitRate *float64          `json:"clientUpperBoundHitRate,omitempty"`
+	VerdictCounts           map[string]uint64 `json:"verdictCounts"`
+	TTFTP50Ms               *uint64           `json:"ttftP50Ms,omitempty"`
+	TTFTP95Ms               *uint64           `json:"ttftP95Ms,omitempty"`
 	Errors                  uint64            `json:"errors"`
 }
 
@@ -506,60 +501,60 @@ type TraceStats struct {
 type TraceRecord struct {
 	Provider             string             `json:"provider"`
 	Model                string             `json:"model"`
-	RequestID            *string            `json:"request_id,omitempty"`
-	SessionID            *string            `json:"session_id,omitempty"`
-	TraceID              string             `json:"call_id"`
-	SentAtUnixMs         int64              `json:"sent_at_unix_ms"`
-	LCPTokenUpper        *uint64            `json:"lcp_token_upper,omitempty"`
-	TTFTMs               *uint64            `json:"ttft_ms,omitempty"`
+	RequestID            *string            `json:"requestId,omitempty"`
+	SessionID            *string            `json:"sessionId,omitempty"`
+	TraceID              string             `json:"callId"`
+	SentAtUnixMs         int64              `json:"sentAtUnixMs"`
+	LCPTokenUpper        *uint64            `json:"lcpTokenUpper,omitempty"`
+	TTFTMs               *uint64            `json:"ttftMs,omitempty"`
 	Fingerprint          Fingerprint        `json:"fingerprint"`
 	Usage                UsageSnapshot      `json:"usage"`
-	ResponseCacheHeaders map[string]string  `json:"response_cache_headers,omitempty"`
-	RequestCacheHints    *RequestCacheHints `json:"request_cache_hints,omitempty"`
+	ResponseCacheHeaders map[string]string  `json:"responseCacheHeaders,omitempty"`
+	RequestCacheHints    *RequestCacheHints `json:"requestCacheHints,omitempty"`
 	Verdict              json.RawMessage    `json:"verdict,omitempty"`
 	Error                *string            `json:"error,omitempty"`
 }
 
 // RequestCacheHints is the best-effort request-side cache hint snapshot.
 type RequestCacheHints struct {
-	RequestedWrite bool `json:"requested_write"`
+	RequestedWrite bool `json:"requestedWrite"`
 }
 
 // Fingerprint is the block-hash chain of a denoised request body (hex).
 type Fingerprint struct {
-	BodyHash      string   `json:"body_hash"`
-	LenBytes      uint64   `json:"len_bytes"`
-	BlockSize     uint64   `json:"block_size"`
-	BlockHashes   []string `json:"block_hashes"`
-	TokenEstimate uint64   `json:"token_estimate"`
+	BodyHash      string   `json:"bodyHash"`
+	LenBytes      uint64   `json:"lenBytes"`
+	BlockSize     uint64   `json:"blockSize"`
+	BlockHashes   []string `json:"blockHashes"`
+	TokenEstimate uint64   `json:"tokenEstimate"`
 }
 
 // UsageSnapshot is the 7-field flat usage snapshot + raw passthrough.
 type UsageSnapshot struct {
-	InputTotal      *uint64         `json:"input_total,omitempty"`
-	InputNoCache    *uint64         `json:"input_no_cache,omitempty"`
-	CacheRead       *uint64         `json:"cache_read,omitempty"`
-	CacheWrite      *uint64         `json:"cache_write,omitempty"`
-	OutputTotal     *uint64         `json:"output_total,omitempty"`
-	OutputText      *uint64         `json:"output_text,omitempty"`
-	OutputReasoning *uint64         `json:"output_reasoning,omitempty"`
+	InputTotal      *uint64         `json:"inputTotal,omitempty"`
+	InputNoCache    *uint64         `json:"inputNoCache,omitempty"`
+	CacheRead       *uint64         `json:"cacheRead,omitempty"`
+	CacheWrite      *uint64         `json:"cacheWrite,omitempty"`
+	OutputTotal     *uint64         `json:"outputTotal,omitempty"`
+	OutputText      *uint64         `json:"outputText,omitempty"`
+	OutputReasoning *uint64         `json:"outputReasoning,omitempty"`
 	Raw             json.RawMessage `json:"raw,omitempty"`
 }
 
 // SessionChainView is the per-session ordered chain view.
 type SessionChainView struct {
-	SessionID       string        `json:"session_id"`
-	RecordIDs       []string      `json:"record_ids"`
-	PrefixStability float64       `json:"prefix_stability"`
+	SessionID       string        `json:"sessionId"`
+	RecordIDs       []string      `json:"recordIds"`
+	PrefixStability float64       `json:"prefixStability"`
 	Breaks          []PrefixBreak `json:"breaks"`
 }
 
 // PrefixBreak marks a prefix break between consecutive session records.
 type PrefixBreak struct {
-	AtRecordID    string `json:"at_record_id"`
-	PrevRecordID  string `json:"prev_record_id"`
-	LCPBytes      uint64 `json:"lcp_bytes"`
-	ExpectedBreak bool   `json:"expected_break"`
+	AtRecordID    string `json:"atRecordId"`
+	PrevRecordID  string `json:"prevRecordId"`
+	LCPBytes      uint64 `json:"lcpBytes"`
+	ExpectedBreak bool   `json:"expectedBreak"`
 	Kind          string `json:"kind"`
 }
 
diff --git a/bindings/java/build.gradle.kts b/bindings/java/build.gradle.kts
index c4fb8a511..7335bba18 100644
--- a/bindings/java/build.gradle.kts
+++ b/bindings/java/build.gradle.kts
@@ -36,8 +36,12 @@ dependencies {
 
 tasks.test {
     useJUnitPlatform()
-    // Point JNA at the freshly built native library (cargo build --release -p aimux-ffi).
-    systemProperty("jna.library.path", projectDir.resolve("../../target/release").canonicalPath)
+    // Point JNA at the freshly built native library (cargo build --release -p aimux-ffi),
+    // or at AIMUX_FFI_LIB_DIR when set (e.g. a debug build or a shared target dir).
+    systemProperty(
+        "jna.library.path",
+        System.getenv("AIMUX_FFI_LIB_DIR") ?: projectDir.resolve("../../target/release").canonicalPath
+    )
     testLogging {
         events("passed", "skipped", "failed")
     }
diff --git a/bindings/java/src/main/java/ai/arcships/aimux/Aimux.java b/bindings/java/src/main/java/ai/arcships/aimux/Aimux.java
index 6525e1c34..e6f00ed42 100644
--- a/bindings/java/src/main/java/ai/arcships/aimux/Aimux.java
+++ b/bindings/java/src/main/java/ai/arcships/aimux/Aimux.java
@@ -130,26 +130,6 @@ public static void recordingTryFlush() {
         }
     }
 
-    /**
-     * Register external OpenAI-compatible providers from a JSON config string
-     * (RFC-0020).
-     *
-     * <p>{@code configJson} is {@code { "providers": [ { "name", "base_url", ... } ] }}.
-     * Entries override same-named built-ins or add new ones. Like
-     * {@link #initRecording}, this mutates process-global registry state.
-     *
-     * @param configJson Provider registry config JSON.
-     * @throws AimuxException if the registry rejects the config
-     *         ({@link AimuxException.InvalidArgumentError}).
-     */
-    public static void registerProviders(String configJson) {
-        AimuxResult.requireJsonNonNull(configJson, "configJson");
-        Pointer e = AimuxFFI.INSTANCE.aimux_register_providers(configJson);
-        if (e != null) {
-            throw AimuxResult.expectAimuxError(e, "registerProviders");
-        }
-    }
-
     /**
      * Set the global proxy configuration (M6, RFC-0016). Must be called before
      * the first {@code generateText} / {@code streamText} call; a no-op if the
diff --git a/bindings/java/src/main/java/ai/arcships/aimux/AimuxFFI.java b/bindings/java/src/main/java/ai/arcships/aimux/AimuxFFI.java
index 80a192905..9d4a83d1e 100644
--- a/bindings/java/src/main/java/ai/arcships/aimux/AimuxFFI.java
+++ b/bindings/java/src/main/java/ai/arcships/aimux/AimuxFFI.java
@@ -379,9 +379,6 @@ Pointer aimux_file_upload(long handle, String dataBase64,
     /** Create a mock replay model from recorded JSONL. Writes the handle to {@code outHandle}; returns the error or null. */
     Pointer aimux_mock_replay_new(String recordingsJsonl, LongByReference outHandle);
 
-    /** Register external OpenAI-compatible providers from JSON config (RFC-0020). Returns the error or null. */
-    Pointer aimux_register_providers(String configJson);
-
     /** Set the global proxy configuration (M6, RFC-0016). Returns the error or null. */
     Pointer aimux_init_proxy(String configJson);
 
diff --git a/bindings/java/src/main/java/ai/arcships/aimux/Model.java b/bindings/java/src/main/java/ai/arcships/aimux/Model.java
index 4755e3a2f..f28ba5d23 100644
--- a/bindings/java/src/main/java/ai/arcships/aimux/Model.java
+++ b/bindings/java/src/main/java/ai/arcships/aimux/Model.java
@@ -368,9 +368,9 @@ public static Model azureWithBase(String apiKey, String baseUrl, String deployme
      *                   from the registry entry.
      * @param modelId    Model id.
      * @param configJson Optional JSON object of ProviderOptions
-     *                   ({@code {"base_url": "...", "headers": {...}, "organization": "...",
+     *                   ({@code {"baseUrl": "...", "headers": {...}, "organization": "...",
      *                   "project": "...", "params": {...}}}); {@code null} for defaults.
-     *                   {@code max_retries} (a per-call option) and {@code body_overrides}
+     *                   {@code maxRetries} (a per-call option) and {@code bodyOverrides}
      *                   (removed) are rejected as {@link AimuxException.InvalidArgumentError}.
      * @return A new {@link Model}.
      * @throws AimuxException if the provider could not be constructed
@@ -469,7 +469,7 @@ public static Model mockReplay(String recordingsJsonl) {
      * @param models     child models (must be non-empty; closed models throw).
      * @param configJson optional config: {@code {"router": "rule"|"weighted",
      *                   "weights": [...], "fallback": "on_error"|"none",
-     *                   "provider_name", "model_id"}} — all optional.
+     *                   "providerName", "modelId"}} — all optional.
      * @return a new RouterModel wrapping the children.
      */
     public static Model router(java.util.List<Model> models, String configJson) {
@@ -493,9 +493,9 @@ public static Model router(java.util.List<Model> models, String configJson) {
      *
      * @param references reference models (may be null/empty — runs aggregator only).
      * @param aggregator the aggregator model (must be non-null and open).
-     * @param configJson optional MoaConfig: {@code {"provider_name", "model_id",
-     *                   "aggregator_instructions", "strip_reference_tools",
-     *                   "fail_mode": "best_effort"|"fail_fast"}}.
+     * @param configJson optional MoaConfig: {@code {"providerName", "modelId",
+     *                   "aggregatorInstructions", "stripReferenceTools",
+     *                   "failMode": "best_effort"|"fail_fast"}}.
      * @return a new MoaModel.
      */
     public static Model moa(java.util.List<Model> references, Model aggregator, String configJson) {
diff --git a/bindings/java/src/main/java/ai/arcships/aimux/MultimodalTypes.java b/bindings/java/src/main/java/ai/arcships/aimux/MultimodalTypes.java
index 34fc9de76..4c6c52ad9 100644
--- a/bindings/java/src/main/java/ai/arcships/aimux/MultimodalTypes.java
+++ b/bindings/java/src/main/java/ai/arcships/aimux/MultimodalTypes.java
@@ -31,8 +31,8 @@
  *
  * <p>Port of the Kotlin binding's {@code MultimodalTypes.kt}. Same shapes as the
  * ts-rs generated {@code .ts} types in {@code bindings/node/src/types/}. Field
- * names are camelCase in Java and map to the wire format's snake_case via
- * {@link JsonProperty}.
+ * names are camelCase in Java and on the wire (the AI SDK JSON); {@link JsonProperty}
+ * pins them.
  *
  * <p>These types are intentionally lenient on decode (unknown keys ignored,
  * every field has a default) so future engine additions don't break existing
@@ -44,12 +44,11 @@
  *     .readValue(jsonStr, MultimodalTypes.EmbeddingResult.class);
  * }</pre>
  *
- * <p>The {@code Base64}/{@code Binary}/{@code Url} unions ({@link AudioData},
- * {@link ImageOutputs}, {@link VideoData}) are serde-style externally-tagged
- * enums on the wire ({@code {"Base64": ...}} / {@code {"Binary": ...}} /
- * {@code {"Url": ...}}), so each has a custom serializer — the same pattern as
- * {@link Types.FileBytes} in {@code Types.java} and
- * {@code StreamPartSerializer} in the Kotlin binding.
+ * <p>The byte unions ({@link AudioData}, {@link ImageOutputs}, {@link VideoFileData})
+ * are untagged on the wire (a base64 string or an int array), and
+ * {@link VideoData} / {@link VideoFile} are tagged on {@code type}, so each has a
+ * custom serializer — the same pattern as {@link Types.FileBytes} in
+ * {@code Types.java}.
  */
 public final class MultimodalTypes {
     private MultimodalTypes() {}
@@ -187,7 +186,7 @@ public boolean equals(Object o) {
     public static class EmbeddingResult {
         @JsonProperty("embeddings") private List<List<Float>> embeddings = new ArrayList<>();
         @JsonProperty("usage") private EmbeddingUsage usage;
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
         @JsonProperty("response") private EmbeddingResponse response;
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
 
@@ -252,9 +251,9 @@ public boolean equals(Object o) {
     /** Options for an embedding call. */
     public static class EmbeddingCallOptions {
         @JsonProperty("values") private List<String> values = new ArrayList<>();
-        @JsonProperty("provider_options") private JsonNode providerOptions;
+        @JsonProperty("providerOptions") private JsonNode providerOptions;
         @JsonProperty("headers") private Map<String, String> headers;
-        @JsonProperty("max_retries") private Integer maxRetries;
+        @JsonProperty("maxRetries") private Integer maxRetries;
         @JsonProperty("timeout") private Types.TimeoutConfiguration timeout;
 
         @JsonCreator
@@ -325,8 +324,7 @@ public boolean equals(Object o) {
     /**
      * Generated audio: a base64 string or raw binary bytes.
      *
-     * <p>Wire format (serde externally-tagged): {@code {"Base64": "..."}} |
-     * {@code {"Binary": [n,...]}}.
+     * <p>Wire format (untagged): a base64 string | {@code [n,...]}.
      */
     public abstract static class AudioData {
         private AudioData() {}
@@ -382,15 +380,14 @@ public boolean equals(Object o) {
         }
     }
 
-    /** Custom (de)serializer for {@link AudioData}: {@code {"Base64": "..."}} | {@code {"Binary": [n,...]}}. */
+    /** Untagged (de)serializer for {@link AudioData}: a base64 string or an int array. */
     public static class AudioDataSerializer extends JsonSerializer<AudioData> {
         @Override
         public void serialize(AudioData value, JsonGenerator gen, SerializerProvider serializers) throws IOException {
-            gen.writeStartObject();
             if (value instanceof AudioData.Base64) {
-                gen.writeStringField("Base64", ((AudioData.Base64) value).getValue());
+                gen.writeString(((AudioData.Base64) value).getValue());
             } else if (value instanceof AudioData.Binary) {
-                gen.writeArrayFieldStart("Binary");
+                gen.writeStartArray();
                 for (Integer b : ((AudioData.Binary) value).getValue()) {
                     gen.writeNumber(b);
                 }
@@ -398,7 +395,6 @@ public void serialize(AudioData value, JsonGenerator gen, SerializerProvider ser
             } else {
                 throw new IOException("Unknown AudioData: " + value);
             }
-            gen.writeEndObject();
         }
     }
 
@@ -407,26 +403,17 @@ public static class AudioDataDeserializer extends JsonDeserializer<AudioData> {
         @Override
         public AudioData deserialize(JsonParser p, DeserializationContext ctxt) throws IOException {
             JsonNode node = p.getCodec().readTree(p);
-            if (!node.isObject() || node.size() != 1) {
-                throw new IOException("AudioData must be a single-key externally-tagged object, got: " + node);
+            if (node.isTextual()) {
+                return new AudioData.Base64(node.asText());
             }
-            String tag = node.fieldNames().next();
-            JsonNode inner = node.get(tag);
-            switch (tag) {
-                case "Base64":
-                    return new AudioData.Base64(inner.asText());
-                case "Binary": {
-                    List<Integer> bytes = new ArrayList<>();
-                    if (inner.isArray()) {
-                        for (JsonNode item : inner) {
-                            bytes.add(item.asInt());
-                        }
-                    }
-                    return new AudioData.Binary(bytes);
+            if (node.isArray()) {
+                List<Integer> bytes = new ArrayList<>();
+                for (JsonNode item : node) {
+                    bytes.add(item.asInt());
                 }
-                default:
-                    throw new IOException("Unknown AudioData tag: '" + tag + "'");
+                return new AudioData.Binary(bytes);
             }
+            throw new IOException("AudioData must be a base64 string or an int array, got: " + node);
         }
     }
 
@@ -476,7 +463,7 @@ public boolean equals(Object o) {
     /** Provider response metadata for speech. */
     public static class SpeechResponse {
         @JsonProperty("timestamp") private String timestamp;
-        @JsonProperty("model_id") private String modelId;
+        @JsonProperty("modelId") private String modelId;
         @JsonProperty("headers") private Map<String, String> headers;
         @JsonProperty("body") private JsonNode body;
 
@@ -541,7 +528,7 @@ public static class SpeechResult {
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
         @JsonProperty("request") private SpeechRequest request;
         @JsonProperty("response") private SpeechResponse response = new SpeechResponse();
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
         @JsonCreator
         SpeechResult() {}
@@ -608,13 +595,13 @@ public boolean equals(Object o) {
     public static class SpeechCallOptions {
         @JsonProperty("text") private String text = "";
         @JsonProperty("voice") private String voice;
-        @JsonProperty("output_format") private String outputFormat;
+        @JsonProperty("outputFormat") private String outputFormat;
         @JsonProperty("instructions") private String instructions;
         @JsonProperty("speed") private Double speed;
         @JsonProperty("language") private String language;
-        @JsonProperty("provider_options") private JsonNode providerOptions;
+        @JsonProperty("providerOptions") private JsonNode providerOptions;
         @JsonProperty("headers") private Map<String, String> headers;
-        @JsonProperty("max_retries") private Integer maxRetries;
+        @JsonProperty("maxRetries") private Integer maxRetries;
         @JsonProperty("timeout") private Types.TimeoutConfiguration timeout;
 
         @JsonCreator
@@ -715,8 +702,7 @@ public int hashCode() {
     /**
      * Generated images: all base64 strings or all binary byte arrays.
      *
-     * <p>Wire format (serde externally-tagged):
-     * {@code {"Base64": ["...", ...]}} | {@code {"Binary": [[n,...], ...]}}.
+     * <p>Wire format (untagged): {@code ["...", ...]} | {@code [[n,...], ...]}.
      */
     public abstract static class ImageOutputs {
         private ImageOutputs() {}
@@ -772,19 +758,16 @@ public boolean equals(Object o) {
         }
     }
 
-    /** Custom (de)serializer for {@link ImageOutputs}: {@code {"Base64": [...]}} | {@code {"Binary": [[...],...]}}. */
+    /** Untagged (de)serializer for {@link ImageOutputs}: an array of base64 strings or of int arrays. */
     public static class ImageOutputsSerializer extends JsonSerializer<ImageOutputs> {
         @Override
         public void serialize(ImageOutputs value, JsonGenerator gen, SerializerProvider serializers) throws IOException {
-            gen.writeStartObject();
+            gen.writeStartArray();
             if (value instanceof ImageOutputs.Base64) {
-                gen.writeArrayFieldStart("Base64");
                 for (String s : ((ImageOutputs.Base64) value).getValue()) {
                     gen.writeString(s);
                 }
-                gen.writeEndArray();
             } else if (value instanceof ImageOutputs.Binary) {
-                gen.writeArrayFieldStart("Binary");
                 for (List<Integer> row : ((ImageOutputs.Binary) value).getValue()) {
                     gen.writeStartArray();
                     for (Integer b : row) {
@@ -792,60 +775,45 @@ public void serialize(ImageOutputs value, JsonGenerator gen, SerializerProvider
                     }
                     gen.writeEndArray();
                 }
-                gen.writeEndArray();
             } else {
                 throw new IOException("Unknown ImageOutputs: " + value);
             }
-            gen.writeEndObject();
+            gen.writeEndArray();
         }
     }
 
-    /** Custom (de)serializer for {@link ImageOutputs}. */
+    /** Custom (de)serializer for {@link ImageOutputs}; an empty array decodes as {@link ImageOutputs.Base64}. */
     public static class ImageOutputsDeserializer extends JsonDeserializer<ImageOutputs> {
         @Override
         public ImageOutputs deserialize(JsonParser p, DeserializationContext ctxt) throws IOException {
             JsonNode node = p.getCodec().readTree(p);
-            if (!node.isObject() || node.size() != 1) {
-                throw new IOException("ImageOutputs must be a single-key externally-tagged object, got: " + node);
+            if (!node.isArray()) {
+                throw new IOException("ImageOutputs must be an array, got: " + node);
             }
-            String tag = node.fieldNames().next();
-            JsonNode inner = node.get(tag);
-            switch (tag) {
-                case "Base64": {
-                    List<String> values = new ArrayList<>();
-                    if (inner.isArray()) {
-                        for (JsonNode item : inner) {
-                            values.add(item.asText());
-                        }
-                    }
-                    return new ImageOutputs.Base64(values);
-                }
-                case "Binary": {
-                    List<List<Integer>> rows = new ArrayList<>();
-                    if (inner.isArray()) {
-                        for (JsonNode rowNode : inner) {
-                            List<Integer> row = new ArrayList<>();
-                            if (rowNode.isArray()) {
-                                for (JsonNode item : rowNode) {
-                                    row.add(item.asInt());
-                                }
-                            }
-                            rows.add(row);
-                        }
+            if (node.size() > 0 && node.get(0).isArray()) {
+                List<List<Integer>> rows = new ArrayList<>();
+                for (JsonNode rowNode : node) {
+                    List<Integer> row = new ArrayList<>();
+                    for (JsonNode item : rowNode) {
+                        row.add(item.asInt());
                     }
-                    return new ImageOutputs.Binary(rows);
+                    rows.add(row);
                 }
-                default:
-                    throw new IOException("Unknown ImageOutputs tag: '" + tag + "'");
+                return new ImageOutputs.Binary(rows);
             }
+            List<String> values = new ArrayList<>();
+            for (JsonNode item : node) {
+                values.add(item.asText());
+            }
+            return new ImageOutputs.Base64(values);
         }
     }
 
     /** Token usage for image generation (if reported). */
     public static class ImageUsage {
-        @JsonProperty("input_tokens") private Long inputTokens;
-        @JsonProperty("output_tokens") private Long outputTokens;
-        @JsonProperty("total_tokens") private Long totalTokens;
+        @JsonProperty("inputTokens") private Long inputTokens;
+        @JsonProperty("outputTokens") private Long outputTokens;
+        @JsonProperty("totalTokens") private Long totalTokens;
 
         @JsonCreator
         ImageUsage() {}
@@ -898,7 +866,7 @@ public boolean equals(Object o) {
     /** Provider response metadata for images. */
     public static class ImageResponse {
         @JsonProperty("timestamp") private String timestamp;
-        @JsonProperty("model_id") private String modelId;
+        @JsonProperty("modelId") private String modelId;
         @JsonProperty("headers") private Map<String, String> headers;
 
         @JsonCreator
@@ -953,7 +921,7 @@ public boolean equals(Object o) {
     public static class ImageResult {
         @JsonProperty("images") private ImageOutputs images = new ImageOutputs.Base64(new ArrayList<String>());
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
         @JsonProperty("response") private ImageResponse response = new ImageResponse();
         @JsonProperty("usage") private ImageUsage usage;
 
@@ -1023,13 +991,13 @@ public static class ImageCallOptions {
         @JsonProperty("prompt") private String prompt;
         @JsonProperty("n") private Integer n;
         @JsonProperty("size") private String size;
-        @JsonProperty("aspect_ratio") private String aspectRatio;
+        @JsonProperty("aspectRatio") private String aspectRatio;
         @JsonProperty("seed") private Long seed;
         @JsonProperty("files") private List<JsonNode> files = new ArrayList<>();
         @JsonProperty("mask") private JsonNode mask;
-        @JsonProperty("provider_options") private JsonNode providerOptions;
+        @JsonProperty("providerOptions") private JsonNode providerOptions;
         @JsonProperty("headers") private Map<String, String> headers;
-        @JsonProperty("max_retries") private Integer maxRetries;
+        @JsonProperty("maxRetries") private Integer maxRetries;
         @JsonProperty("timeout") private Types.TimeoutConfiguration timeout;
 
         @JsonCreator
@@ -1138,9 +1106,9 @@ public static class TranscriptionSegment {
         @JsonProperty("text") private String text = "";
         // Core declares these as required f64 (TranscriptionSegment.ts), so they
         // are primitives: NON_NULL inclusion would drop boxed nulls and the core
-        // would reject the payload with `missing field start_second`.
-        @JsonProperty("start_second") private double startSecond;
-        @JsonProperty("end_second") private double endSecond;
+        // would reject the payload with `missing field startSecond`.
+        @JsonProperty("startSecond") private double startSecond;
+        @JsonProperty("endSecond") private double endSecond;
 
         @JsonCreator
         TranscriptionSegment() {}
@@ -1236,7 +1204,7 @@ public boolean equals(Object o) {
     /** Provider response metadata for transcription. */
     public static class TranscriptionResponse {
         @JsonProperty("timestamp") private String timestamp;
-        @JsonProperty("model_id") private String modelId;
+        @JsonProperty("modelId") private String modelId;
         @JsonProperty("headers") private Map<String, String> headers;
         @JsonProperty("body") private JsonNode body;
 
@@ -1300,11 +1268,11 @@ public static class TranscriptionResult {
         @JsonProperty("text") private String text = "";
         @JsonProperty("segments") private List<TranscriptionSegment> segments = new ArrayList<>();
         @JsonProperty("language") private String language;
-        @JsonProperty("duration_in_seconds") private Double durationInSeconds;
+        @JsonProperty("durationInSeconds") private Double durationInSeconds;
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
         @JsonProperty("request") private TranscriptionRequest request;
         @JsonProperty("response") private TranscriptionResponse response = new TranscriptionResponse();
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
         @JsonCreator
         TranscriptionResult() {}
@@ -1391,10 +1359,10 @@ public int hashCode() {
     /** Options for transcription. */
     public static class TranscriptionCallOptions {
         @JsonProperty("audio") private JsonNode audio = JsonNodeFactory.instance.objectNode();
-        @JsonProperty("media_type") private String mediaType = "";
-        @JsonProperty("provider_options") private JsonNode providerOptions;
+        @JsonProperty("mediaType") private String mediaType = "";
+        @JsonProperty("providerOptions") private JsonNode providerOptions;
         @JsonProperty("headers") private Map<String, String> headers;
-        @JsonProperty("max_retries") private Integer maxRetries;
+        @JsonProperty("maxRetries") private Integer maxRetries;
         @JsonProperty("timeout") private Types.TimeoutConfiguration timeout;
 
         @JsonCreator
@@ -1471,7 +1439,7 @@ public boolean equals(Object o) {
     /** A single reranked entry. */
     public static class RerankingRank {
         @JsonProperty("index") private int index;
-        @JsonProperty("relevance_score") private double relevanceScore;
+        @JsonProperty("relevanceScore") private double relevanceScore;
 
         @JsonCreator
         RerankingRank() {}
@@ -1520,7 +1488,7 @@ public boolean equals(Object o) {
     public static class RerankingResponse {
         @JsonProperty("id") private String id;
         @JsonProperty("timestamp") private String timestamp;
-        @JsonProperty("model_id") private String modelId;
+        @JsonProperty("modelId") private String modelId;
         @JsonProperty("headers") private Map<String, String> headers;
         @JsonProperty("body") private JsonNode body;
 
@@ -1587,7 +1555,7 @@ public boolean equals(Object o) {
     /** Result of a reranking call. */
     public static class RerankingResult {
         @JsonProperty("ranking") private List<RerankingRank> ranking = new ArrayList<>();
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
         @JsonProperty("response") private RerankingResponse response;
 
@@ -1651,10 +1619,10 @@ public boolean equals(Object o) {
     public static class RerankingCallOptions {
         @JsonProperty("documents") private JsonNode documents = JsonNodeFactory.instance.arrayNode();
         @JsonProperty("query") private String query = "";
-        @JsonProperty("top_n") private Integer topN;
-        @JsonProperty("provider_options") private JsonNode providerOptions;
+        @JsonProperty("topN") private Integer topN;
+        @JsonProperty("providerOptions") private JsonNode providerOptions;
         @JsonProperty("headers") private Map<String, String> headers;
-        @JsonProperty("max_retries") private Integer maxRetries;
+        @JsonProperty("maxRetries") private Integer maxRetries;
         @JsonProperty("timeout") private Types.TimeoutConfiguration timeout;
 
         @JsonCreator
@@ -1736,7 +1704,7 @@ public boolean equals(Object o) {
     /** The URL variant of {@link VideoData}. */
     public static class VideoUrlData {
         @JsonProperty("url") private String url = "";
-        @JsonProperty("media_type") private String mediaType = "";
+        @JsonProperty("mediaType") private String mediaType = "";
 
         @JsonCreator
         VideoUrlData() {}
@@ -1784,7 +1752,7 @@ public boolean equals(Object o) {
     /** The base64 variant of {@link VideoData}. */
     public static class VideoBase64Data {
         @JsonProperty("data") private String data = "";
-        @JsonProperty("media_type") private String mediaType = "";
+        @JsonProperty("mediaType") private String mediaType = "";
 
         @JsonCreator
         VideoBase64Data() {}
@@ -1832,7 +1800,7 @@ public boolean equals(Object o) {
     /** The binary variant of {@link VideoData}. */
     public static class VideoBinaryData {
         @JsonProperty("data") private List<Integer> data = new ArrayList<>();
-        @JsonProperty("media_type") private String mediaType = "";
+        @JsonProperty("mediaType") private String mediaType = "";
 
         @JsonCreator
         VideoBinaryData() {}
@@ -1880,8 +1848,8 @@ public boolean equals(Object o) {
     /**
      * Generated video: a URL, base64 string, or raw binary bytes.
      *
-     * <p>Wire format (serde externally-tagged):
-     * {@code {"Url": {...}}} | {@code {"Base64": {...}}} | {@code {"Binary": {...}}}.
+     * <p>Wire format (tagged on {@code type}): {@code {"type":"url",...}} |
+     * {@code {"type":"base64",...}} | {@code {"type":"binary",...}}.
      */
     public abstract static class VideoData {
         private VideoData() {}
@@ -1962,24 +1930,24 @@ public boolean equals(Object o) {
         }
     }
 
-    /** Custom (de)serializer for {@link VideoData}: {@code {"Url": {...}}} | {@code {"Base64": {...}}} | {@code {"Binary": {...}}}. */
+    /** Custom (de)serializer for {@link VideoData} — internally tagged on `"type"` with the payload inlined. */
     public static class VideoDataSerializer extends JsonSerializer<VideoData> {
         @Override
         public void serialize(VideoData value, JsonGenerator gen, SerializerProvider serializers) throws IOException {
-            gen.writeStartObject();
+            ObjectNode out = JsonNodeFactory.instance.objectNode();
             if (value instanceof VideoData.Url) {
-                gen.writeFieldName("Url");
-                gen.writeTree(AimuxJson.MAPPER.valueToTree(((VideoData.Url) value).getValue()));
+                out.put("type", "url");
+                out.setAll((ObjectNode) AimuxJson.MAPPER.valueToTree(((VideoData.Url) value).getValue()));
             } else if (value instanceof VideoData.Base64) {
-                gen.writeFieldName("Base64");
-                gen.writeTree(AimuxJson.MAPPER.valueToTree(((VideoData.Base64) value).getValue()));
+                out.put("type", "base64");
+                out.setAll((ObjectNode) AimuxJson.MAPPER.valueToTree(((VideoData.Base64) value).getValue()));
             } else if (value instanceof VideoData.Binary) {
-                gen.writeFieldName("Binary");
-                gen.writeTree(AimuxJson.MAPPER.valueToTree(((VideoData.Binary) value).getValue()));
+                out.put("type", "binary");
+                out.setAll((ObjectNode) AimuxJson.MAPPER.valueToTree(((VideoData.Binary) value).getValue()));
             } else {
                 throw new IOException("Unknown VideoData: " + value);
             }
-            gen.writeEndObject();
+            gen.writeTree(out);
         }
     }
 
@@ -1988,23 +1956,16 @@ public static class VideoDataDeserializer extends JsonDeserializer<VideoData> {
         @Override
         public VideoData deserialize(JsonParser p, DeserializationContext ctxt) throws IOException {
             JsonNode node = p.getCodec().readTree(p);
-            if (!node.isObject() || node.size() != 1) {
-                throw new IOException("VideoData must be a single-key externally-tagged object, got: " + node);
-            }
-            String tag = node.fieldNames().next();
-            JsonNode inner = node.get(tag);
-            if (!inner.isObject()) {
-                inner = JsonNodeFactory.instance.objectNode();
-            }
+            String tag = node.path("type").asText();
             switch (tag) {
-                case "Url":
-                    return new VideoData.Url(AimuxJson.MAPPER.treeToValue(inner, VideoUrlData.class));
-                case "Base64":
-                    return new VideoData.Base64(AimuxJson.MAPPER.treeToValue(inner, VideoBase64Data.class));
-                case "Binary":
-                    return new VideoData.Binary(AimuxJson.MAPPER.treeToValue(inner, VideoBinaryData.class));
+                case "url":
+                    return new VideoData.Url(AimuxJson.MAPPER.treeToValue(node, VideoUrlData.class));
+                case "base64":
+                    return new VideoData.Base64(AimuxJson.MAPPER.treeToValue(node, VideoBase64Data.class));
+                case "binary":
+                    return new VideoData.Binary(AimuxJson.MAPPER.treeToValue(node, VideoBinaryData.class));
                 default:
-                    throw new IOException("Unknown VideoData tag: '" + tag + "'");
+                    throw new IOException("Unknown VideoData type: '" + tag + "'");
             }
         }
     }
@@ -2012,7 +1973,7 @@ public VideoData deserialize(JsonParser p, DeserializationContext ctxt) throws I
     /** Provider response metadata for video. */
     public static class VideoResponse {
         @JsonProperty("timestamp") private String timestamp;
-        @JsonProperty("model_id") private String modelId;
+        @JsonProperty("modelId") private String modelId;
         @JsonProperty("headers") private Map<String, String> headers;
 
         @JsonCreator
@@ -2067,7 +2028,7 @@ public boolean equals(Object o) {
     public static class VideoResult {
         @JsonProperty("videos") private List<VideoData> videos = new ArrayList<>();
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
         @JsonProperty("response") private VideoResponse response = new VideoResponse();
 
         @JsonCreator
@@ -2129,8 +2090,7 @@ public boolean equals(Object o) {
     /**
      * File payload for a {@link VideoFile}: a base64 string or raw binary bytes.
      *
-     * <p>Wire format (serde externally-tagged): {@code {"Base64": "..."}} |
-     * {@code {"Binary": [n,...]}}.
+     * <p>Wire format (untagged): a base64 string | {@code [n,...]}.
      */
     public abstract static class VideoFileData {
         private VideoFileData() {}
@@ -2186,15 +2146,14 @@ public boolean equals(Object o) {
         }
     }
 
-    /** Custom (de)serializer for {@link VideoFileData}: {@code {"Base64": "..."}} | {@code {"Binary": [n,...]}}. */
+    /** Untagged (de)serializer for {@link VideoFileData}: a base64 string or an int array. */
     public static class VideoFileDataSerializer extends JsonSerializer<VideoFileData> {
         @Override
         public void serialize(VideoFileData value, JsonGenerator gen, SerializerProvider serializers) throws IOException {
-            gen.writeStartObject();
             if (value instanceof VideoFileData.Base64) {
-                gen.writeStringField("Base64", ((VideoFileData.Base64) value).getValue());
+                gen.writeString(((VideoFileData.Base64) value).getValue());
             } else if (value instanceof VideoFileData.Binary) {
-                gen.writeArrayFieldStart("Binary");
+                gen.writeStartArray();
                 for (Integer b : ((VideoFileData.Binary) value).getValue()) {
                     gen.writeNumber(b);
                 }
@@ -2202,7 +2161,6 @@ public void serialize(VideoFileData value, JsonGenerator gen, SerializerProvider
             } else {
                 throw new IOException("Unknown VideoFileData: " + value);
             }
-            gen.writeEndObject();
         }
     }
 
@@ -2211,32 +2169,23 @@ public static class VideoFileDataDeserializer extends JsonDeserializer<VideoFile
         @Override
         public VideoFileData deserialize(JsonParser p, DeserializationContext ctxt) throws IOException {
             JsonNode node = p.getCodec().readTree(p);
-            if (!node.isObject() || node.size() != 1) {
-                throw new IOException("VideoFileData must be a single-key externally-tagged object, got: " + node);
+            if (node.isTextual()) {
+                return new VideoFileData.Base64(node.asText());
             }
-            String tag = node.fieldNames().next();
-            JsonNode inner = node.get(tag);
-            switch (tag) {
-                case "Base64":
-                    return new VideoFileData.Base64(inner.asText());
-                case "Binary": {
-                    List<Integer> bytes = new ArrayList<>();
-                    if (inner.isArray()) {
-                        for (JsonNode item : inner) {
-                            bytes.add(item.asInt());
-                        }
-                    }
-                    return new VideoFileData.Binary(bytes);
+            if (node.isArray()) {
+                List<Integer> bytes = new ArrayList<>();
+                for (JsonNode item : node) {
+                    bytes.add(item.asInt());
                 }
-                default:
-                    throw new IOException("Unknown VideoFileData tag: '" + tag + "'");
+                return new VideoFileData.Binary(bytes);
             }
+            throw new IOException("VideoFileData must be a base64 string or an int array, got: " + node);
         }
     }
 
     /** The {@code File} variant payload of {@link VideoFile}. */
     public static class VideoFileFileData {
-        @JsonProperty("media_type") private String mediaType = "";
+        @JsonProperty("mediaType") private String mediaType = "";
         @JsonProperty("data") private VideoFileData data = new VideoFileData.Base64("");
 
         @JsonCreator
@@ -2285,7 +2234,7 @@ public boolean equals(Object o) {
     /** The {@code Url} variant payload of {@link VideoFile}. */
     public static class VideoFileUrlData {
         @JsonProperty("url") private String url = "";
-        @JsonProperty("media_type") private String mediaType;
+        @JsonProperty("mediaType") private String mediaType;
 
         @JsonCreator
         VideoFileUrlData() {}
@@ -2333,8 +2282,8 @@ public boolean equals(Object o) {
     /**
      * A video or image file used for video editing or image-to-video generation.
      *
-     * <p>Wire format (serde externally-tagged): {@code {"File": {...}}} |
-     * {@code {"Url": {...}}}.
+     * <p>Wire format (tagged on {@code type}): {@code {"type":"file",...}} |
+     * {@code {"type":"url",...}}.
      */
     public abstract static class VideoFile {
         private VideoFile() {}
@@ -2390,21 +2339,21 @@ public boolean equals(Object o) {
         }
     }
 
-    /** Custom (de)serializer for {@link VideoFile}: {@code {"File": {...}}} | {@code {"Url": {...}}}. */
+    /** Custom (de)serializer for {@link VideoFile} — internally tagged on `"type"` with the payload inlined. */
     public static class VideoFileSerializer extends JsonSerializer<VideoFile> {
         @Override
         public void serialize(VideoFile value, JsonGenerator gen, SerializerProvider serializers) throws IOException {
-            gen.writeStartObject();
+            ObjectNode out = JsonNodeFactory.instance.objectNode();
             if (value instanceof VideoFile.File) {
-                gen.writeFieldName("File");
-                gen.writeTree(AimuxJson.MAPPER.valueToTree(((VideoFile.File) value).getValue()));
+                out.put("type", "file");
+                out.setAll((ObjectNode) AimuxJson.MAPPER.valueToTree(((VideoFile.File) value).getValue()));
             } else if (value instanceof VideoFile.Url) {
-                gen.writeFieldName("Url");
-                gen.writeTree(AimuxJson.MAPPER.valueToTree(((VideoFile.Url) value).getValue()));
+                out.put("type", "url");
+                out.setAll((ObjectNode) AimuxJson.MAPPER.valueToTree(((VideoFile.Url) value).getValue()));
             } else {
                 throw new IOException("Unknown VideoFile: " + value);
             }
-            gen.writeEndObject();
+            gen.writeTree(out);
         }
     }
 
@@ -2413,35 +2362,28 @@ public static class VideoFileDeserializer extends JsonDeserializer<VideoFile> {
         @Override
         public VideoFile deserialize(JsonParser p, DeserializationContext ctxt) throws IOException {
             JsonNode node = p.getCodec().readTree(p);
-            if (!node.isObject() || node.size() != 1) {
-                throw new IOException("VideoFile must be a single-key externally-tagged object, got: " + node);
-            }
-            String tag = node.fieldNames().next();
-            JsonNode inner = node.get(tag);
-            if (!inner.isObject()) {
-                inner = JsonNodeFactory.instance.objectNode();
-            }
+            String tag = node.path("type").asText();
             switch (tag) {
-                case "File":
-                    return new VideoFile.File(AimuxJson.MAPPER.treeToValue(inner, VideoFileFileData.class));
-                case "Url":
-                    return new VideoFile.Url(AimuxJson.MAPPER.treeToValue(inner, VideoFileUrlData.class));
+                case "file":
+                    return new VideoFile.File(AimuxJson.MAPPER.treeToValue(node, VideoFileFileData.class));
+                case "url":
+                    return new VideoFile.Url(AimuxJson.MAPPER.treeToValue(node, VideoFileUrlData.class));
                 default:
-                    throw new IOException("Unknown VideoFile tag: '" + tag + "'");
+                    throw new IOException("Unknown VideoFile type: '" + tag + "'");
             }
         }
     }
 
     /** The role a frame image plays in video generation. */
     public enum VideoFrameType {
-        @JsonProperty("FirstFrame") FIRST_FRAME,
-        @JsonProperty("LastFrame") LAST_FRAME,
+        @JsonProperty("first_frame") FIRST_FRAME,
+        @JsonProperty("last_frame") LAST_FRAME,
     }
 
     /** A role-tagged image input for image-to-video and first-last-frame generation. */
     public static class VideoFrameImage {
         @JsonProperty("image") private VideoFile image = new VideoFile.File(new VideoFileFileData());
-        @JsonProperty("frame_type") private VideoFrameType frameType = VideoFrameType.FIRST_FRAME;
+        @JsonProperty("frameType") private VideoFrameType frameType = VideoFrameType.FIRST_FRAME;
 
         @JsonCreator
         VideoFrameImage() {}
@@ -2488,8 +2430,8 @@ public boolean equals(Object o) {
 
     /** Per-call pacing overrides for the Core-owned video status poll loop. */
     public static class VideoPollOptions {
-        @JsonProperty("interval_ms") private Long intervalMs;
-        @JsonProperty("timeout_ms") private Long timeoutMs;
+        @JsonProperty("intervalMs") private Long intervalMs;
+        @JsonProperty("timeoutMs") private Long timeoutMs;
 
         @JsonCreator
         VideoPollOptions() {}
@@ -2538,18 +2480,18 @@ public String toString() {
     public static class VideoCallOptions {
         @JsonProperty("prompt") private String prompt;
         @JsonProperty("n") private Integer n;
-        @JsonProperty("aspect_ratio") private String aspectRatio;
+        @JsonProperty("aspectRatio") private String aspectRatio;
         @JsonProperty("resolution") private String resolution;
         @JsonProperty("duration") private Long duration;
         @JsonProperty("fps") private Long fps;
         @JsonProperty("seed") private Long seed;
         @JsonProperty("image") private VideoFile image;
-        @JsonProperty("frame_images") private List<VideoFrameImage> frameImages;
-        @JsonProperty("input_references") private List<VideoFile> inputReferences;
-        @JsonProperty("generate_audio") private Boolean generateAudio;
-        @JsonProperty("provider_options") private JsonNode providerOptions;
+        @JsonProperty("frameImages") private List<VideoFrameImage> frameImages;
+        @JsonProperty("inputReferences") private List<VideoFile> inputReferences;
+        @JsonProperty("generateAudio") private Boolean generateAudio;
+        @JsonProperty("providerOptions") private JsonNode providerOptions;
         @JsonProperty("headers") private Map<String, String> headers;
-        @JsonProperty("max_retries") private Integer maxRetries;
+        @JsonProperty("maxRetries") private Integer maxRetries;
         @JsonProperty("poll") private VideoPollOptions poll;
         @JsonProperty("timeout") private Types.TimeoutConfiguration timeout;
 
@@ -2690,7 +2632,7 @@ public static class SearchResultItem {
         @JsonProperty("title") private String title = "";
         @JsonProperty("url") private String url = "";
         @JsonProperty("content") private String content = "";
-        @JsonProperty("raw_content") private String rawContent;
+        @JsonProperty("rawContent") private String rawContent;
         @JsonProperty("score") private Double score;
 
         @JsonCreator
@@ -2806,7 +2748,7 @@ public boolean equals(Object o) {
     public static class SearchResult {
         @JsonProperty("results") private List<SearchResultItem> results = new ArrayList<>();
         @JsonProperty("answer") private String answer;
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
         @JsonProperty("response") private SearchResponse response;
 
@@ -2874,14 +2816,14 @@ public boolean equals(Object o) {
     /** Options for a search call. */
     public static class SearchCallOptions {
         @JsonProperty("query") private String query = "";
-        @JsonProperty("max_results") private Integer maxResults;
-        @JsonProperty("include_raw_content") private Boolean includeRawContent;
-        @JsonProperty("time_range") private String timeRange;
-        @JsonProperty("include_domains") private List<String> includeDomains = new ArrayList<>();
-        @JsonProperty("exclude_domains") private List<String> excludeDomains = new ArrayList<>();
-        @JsonProperty("provider_options") private JsonNode providerOptions;
+        @JsonProperty("maxResults") private Integer maxResults;
+        @JsonProperty("includeRawContent") private Boolean includeRawContent;
+        @JsonProperty("timeRange") private String timeRange;
+        @JsonProperty("includeDomains") private List<String> includeDomains = new ArrayList<>();
+        @JsonProperty("excludeDomains") private List<String> excludeDomains = new ArrayList<>();
+        @JsonProperty("providerOptions") private JsonNode providerOptions;
         @JsonProperty("headers") private Map<String, String> headers;
-        @JsonProperty("max_retries") private Integer maxRetries;
+        @JsonProperty("maxRetries") private Integer maxRetries;
         @JsonProperty("timeout") private Types.TimeoutConfiguration timeout;
 
         @JsonCreator
@@ -2984,10 +2926,10 @@ public int hashCode() {
 
     /** Result of a file upload. */
     public static class UploadFileResult {
-        @JsonProperty("provider_reference") private Map<String, String> providerReference = new HashMap<>();
-        @JsonProperty("media_type") private String mediaType;
+        @JsonProperty("providerReference") private Map<String, String> providerReference = new HashMap<>();
+        @JsonProperty("mediaType") private String mediaType;
         @JsonProperty("filename") private String filename;
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
 
         @JsonCreator
@@ -3056,9 +2998,9 @@ public int hashCode() {
     /** Options for a file upload. */
     public static class UploadFileCallOptions {
         @JsonProperty("data") private JsonNode data = JsonNodeFactory.instance.objectNode();
-        @JsonProperty("media_type") private String mediaType = "";
+        @JsonProperty("mediaType") private String mediaType = "";
         @JsonProperty("filename") private String filename;
-        @JsonProperty("provider_options") private JsonNode providerOptions;
+        @JsonProperty("providerOptions") private JsonNode providerOptions;
         @JsonProperty("headers") private Map<String, String> headers;
 
         @JsonCreator
diff --git a/bindings/java/src/main/java/ai/arcships/aimux/ToolCallRepairs.java b/bindings/java/src/main/java/ai/arcships/aimux/ToolCallRepairs.java
index 959caa3a8..fc560a916 100644
--- a/bindings/java/src/main/java/ai/arcships/aimux/ToolCallRepairs.java
+++ b/bindings/java/src/main/java/ai/arcships/aimux/ToolCallRepairs.java
@@ -44,7 +44,7 @@ static String apply(String toolCallJson, String optsJson, String replyJson) {
         return AimuxResult.extractString(e, out, "apply_tool_call_repair");
     }
 
-    /** Patch a whole result document (tool_calls and response_messages). */
+    /** Patch a whole result document (toolCalls and responseMessages). */
     static String applyToResult(String resultJson, String optsJson, String toolCallId, String replyJson) {
         PointerByReference out = new PointerByReference();
         Pointer e = AimuxFFI.INSTANCE.aimux_apply_tool_call_repair_to_result(
@@ -82,7 +82,7 @@ static String repairResult(String resultJson, String promptJson, String optsJson
                 continue; // no tool set — never repaired (AI SDK rule)
             }
             patched = applyToResult(patched, optsJson,
-                call.path("tool_call_id").asText(), reply(hook, contextJson));
+                call.path("toolCallId").asText(), reply(hook, contextJson));
         }
         return patched;
     }
@@ -102,8 +102,8 @@ static String repairStreamPart(String partJson, String promptJson, String optsJs
             return partJson;
         }
         JsonNode part = parse(partJson);
-        JsonNode call = part.path("ToolCall");
-        if (!call.isObject() || !call.path("invalid").asBoolean(false)) {
+        JsonNode call = part;
+        if (!"tool-call".equals(part.path("type").asText()) || !call.path("invalid").asBoolean(false)) {
             return partJson;
         }
         String contextJson = context(call.toString(), promptJson, optsJson);
@@ -112,7 +112,8 @@ static String repairStreamPart(String partJson, String promptJson, String optsJs
         }
         String repaired = apply(call.toString(), optsJson, reply(offCallbackThread(hook, model), contextJson));
         ObjectNode out = Types.AimuxJson.MAPPER.createObjectNode();
-        out.set("ToolCall", parse(repaired));
+        out.put("type", "tool-call");
+        out.setAll((ObjectNode) parse(repaired));
         return out.toString();
     }
 
@@ -180,7 +181,7 @@ private static String reply(ToolCallRepair hook, String contextJson) {
             replacement = hook.repair(context);
         } catch (Exception e) {
             // The host's own failure is data, not a thrown error: it comes back
-            // as ToolCallRepair { original_error, cause }.
+            // as ToolCallRepair { originalError, cause }.
             ObjectNode failed = Types.AimuxJson.MAPPER.createObjectNode();
             failed.put("type", "failed");
             failed.put("message", e.getMessage() == null ? e.toString() : e.getMessage());
@@ -191,14 +192,14 @@ private static String reply(ToolCallRepair hook, String contextJson) {
         }
         ObjectNode repaired = Types.AimuxJson.MAPPER.createObjectNode();
         repaired.put("type", "repaired");
-        repaired.set("tool_call", Types.AimuxJson.MAPPER.valueToTree(replacement));
+        repaired.set("toolCall", Types.AimuxJson.MAPPER.valueToTree(replacement));
         return repaired.toString();
     }
 
-    /** {@code tool_calls}, from a text result or from the {@code raw} of an object result. */
+    /** {@code toolCalls}, from a text result or from the {@code raw} of an object result. */
     private static JsonNode toolCalls(JsonNode result) {
-        JsonNode calls = result.path("tool_calls");
-        return calls.isArray() ? calls : result.path("raw").path("tool_calls");
+        JsonNode calls = result.path("toolCalls");
+        return calls.isArray() ? calls : result.path("raw").path("toolCalls");
     }
 
     private static boolean isNull(String json) {
diff --git a/bindings/java/src/main/java/ai/arcships/aimux/Types.java b/bindings/java/src/main/java/ai/arcships/aimux/Types.java
index ca6eb20e9..4082f820e 100644
--- a/bindings/java/src/main/java/ai/arcships/aimux/Types.java
+++ b/bindings/java/src/main/java/ai/arcships/aimux/Types.java
@@ -39,8 +39,8 @@
  * `bindings/kotlin` (which in turn mirrors the ts-rs output in
  * `bindings/node/src/types`).
  *
- * Field names are camelCase in Java and mapped to the wire format's snake_case
- * via {@link JsonProperty}. The raw JSON boundary is handled by
+ * Field names are camelCase in Java and on the wire (the AI SDK JSON), so
+ * {@link JsonProperty} only pins them. The raw JSON boundary is handled by
  * {@link TypedModel} — callers of this layer never parse JSON by hand.
  *
  * These types are intentionally lenient on decode (unknown keys ignored, every
@@ -69,7 +69,7 @@ private AimuxJson() {}
         public static final ObjectMapper MAPPER = createMapper();
 
         /**
-         * Inner mapper used by the externally-tagged polymorphic serializers
+         * Inner mapper used by the `type`-tagged polymorphic serializers
          * ({@link ContentPartSerializer}, {@link StreamPartSerializer},
          * {@link GenerateContentSerializer}) to serialize the concrete variant
          * as a plain POJO. It deliberately does NOT register those four
@@ -109,7 +109,7 @@ private static ObjectMapper configure(ObjectMapper m, boolean polymorphic) {
             module.addDeserializer(FileData.class, new FileDataDeserializer());
             module.addSerializer(GeneratedFileData.class, new GeneratedFileDataSerializer());
             module.addDeserializer(GeneratedFileData.class, new GeneratedFileDataDeserializer());
-            // Externally-tagged polymorphic serializers — only on the public
+            // `type`-tagged polymorphic serializers — only on the public
             // mapper. They use INNER_MAPPER internally to avoid self-recursion.
             if (polymorphic) {
                 module.addSerializer(MessageContent.class, new MessageContentSerializer());
@@ -162,9 +162,9 @@ public enum ReasoningEffort {
 
     public static class InputTokenUsage {
         @JsonProperty("total") private Long total;
-        @JsonProperty("no_cache") private Long noCache;
-        @JsonProperty("cache_read") private Long cacheRead;
-        @JsonProperty("cache_write") private Long cacheWrite;
+        @JsonProperty("noCache") private Long noCache;
+        @JsonProperty("cacheRead") private Long cacheRead;
+        @JsonProperty("cacheWrite") private Long cacheWrite;
         @JsonCreator InputTokenUsage() {}
         public InputTokenUsage(Long total, Long noCache, Long cacheRead, Long cacheWrite) {
             this.total = total;
@@ -228,12 +228,12 @@ public static class Builder {
     /**
      * Token usage statistics.
      *
-     * Mirrors `Usage.ts`: `{ input_tokens: InputTokenUsage, output_tokens: OutputTokenUsage,
-     * raw?: JSONObject | null }`.
+     * Mirrors `Usage.ts`: `{ inputTokens: InputTokenUsage, outputTokens: OutputTokenUsage,
+     * raw?: JSONObject }`.
      */
     public static class Usage {
-        @JsonProperty("input_tokens") private InputTokenUsage inputTokens = new InputTokenUsage();
-        @JsonProperty("output_tokens") private OutputTokenUsage outputTokens = new OutputTokenUsage();
+        @JsonProperty("inputTokens") private InputTokenUsage inputTokens = new InputTokenUsage();
+        @JsonProperty("outputTokens") private OutputTokenUsage outputTokens = new OutputTokenUsage();
         @JsonProperty("raw") private ObjectNode raw;
 
         @JsonCreator
@@ -334,12 +334,12 @@ public int hashCode() {
      * Metadata about the API response.
      *
      * Mirrors `ResponseMetadata.ts`: `{ id: string | null, timestamp: string | null,
-     * model_id: string | null }`.
+     * modelId: string | null }`.
      */
     public static class ResponseMetadata {
         @JsonProperty("id") private String id;
         @JsonProperty("timestamp") private String timestamp;
-        @JsonProperty("model_id") private String modelId;
+        @JsonProperty("modelId") private String modelId;
 
         @JsonCreator
         ResponseMetadata() {}
@@ -387,20 +387,20 @@ public int hashCode() {
     /**
      * A tool call requested by the model.
      *
-     * Mirrors `ToolCall.ts`: `{ tool_call_id, tool_name, input: JsonValue,
-     * provider_executed?: bool | null, dynamic?: bool | null,
-     * provider_metadata?: JsonValue | null }`.
+     * Mirrors `ToolCall.ts`: `{ toolCallId, toolName, input: JsonValue,
+     * providerExecuted?: bool, dynamic?: bool, providerMetadata?: JsonValue,
+     * invalid?: bool, error?: AiMuxError }`.
      *
      * `input` is a {@link JsonNode} because it is usually an arbitrary JSON object
      * (the tool arguments) whose shape is tool-specific.
      */
     public static class ToolCall {
-        @JsonProperty("tool_call_id") private String toolCallId = "";
-        @JsonProperty("tool_name") private String toolName = "";
+        @JsonProperty("toolCallId") private String toolCallId = "";
+        @JsonProperty("toolName") private String toolName = "";
         @JsonProperty("input") private JsonNode input = emptyObject();
-        @JsonProperty("provider_executed") private Boolean providerExecuted;
+        @JsonProperty("providerExecuted") private Boolean providerExecuted;
         @JsonProperty("dynamic") private Boolean dynamic;
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
         @JsonProperty("invalid") private Boolean invalid;
         @JsonProperty("error") private JsonNode error;
 
@@ -507,15 +507,15 @@ public boolean equals(Object o) {
     /**
      * A user-defined function tool.
      *
-     * Mirrors `FunctionTool.ts`. `input_schema` is a {@link JsonNode} (a JSON Schema).
+     * Mirrors `FunctionTool.ts`. `inputSchema` is a {@link JsonNode} (a JSON Schema).
      */
     public static class FunctionTool {
         @JsonProperty("name") private String name = "";
         @JsonProperty("description") private String description;
-        @JsonProperty("input_schema") private JsonNode inputSchema = emptyObject();
+        @JsonProperty("inputSchema") private JsonNode inputSchema = emptyObject();
         @JsonProperty("strict") private Boolean strict;
-        @JsonProperty("provider_options") private Map<String, JsonNode> providerOptions;
-        @JsonProperty("input_examples") private List<FunctionToolInputExample> inputExamples;
+        @JsonProperty("providerOptions") private Map<String, JsonNode> providerOptions;
+        @JsonProperty("inputExamples") private List<FunctionToolInputExample> inputExamples;
 
         @JsonCreator
         FunctionTool() {}
@@ -649,10 +649,10 @@ private Tool() {}
         public static class Function extends Tool {
             @JsonProperty("name") private String name = "";
             @JsonProperty("description") private String description;
-            @JsonProperty("input_schema") private JsonNode inputSchema = emptyObject();
+            @JsonProperty("inputSchema") private JsonNode inputSchema = emptyObject();
             @JsonProperty("strict") private Boolean strict;
-            @JsonProperty("provider_options") private Map<String, JsonNode> providerOptions;
-            @JsonProperty("input_examples") private List<FunctionToolInputExample> inputExamples;
+            @JsonProperty("providerOptions") private Map<String, JsonNode> providerOptions;
+            @JsonProperty("inputExamples") private List<FunctionToolInputExample> inputExamples;
 
             @JsonCreator
             Function() {}
@@ -916,6 +916,67 @@ public ToolChoice deserialize(JsonParser p, DeserializationContext ctxt) throws
         }
     }
 
+    // ─────────────────────────────────────────────────────────────────────────────
+    // ToolResultOutput (the `output` of a prompt-side tool-result part).
+    // ─────────────────────────────────────────────────────────────────────────────
+
+    /**
+     * Output of a tool call in a prompt message.
+     *
+     * Mirrors `ToolResultOutput.ts`, tagged on `type`: `text` / `error-text`
+     * carry a string `value`, `json` / `error-json` a JSON `value`, `content` an
+     * array of `{type:"text"|"file"|"custom",...}` items, and `execution-denied`
+     * an optional `reason`. `providerOptions` is absent on `content`.
+     */
+    public static class ToolResultOutput {
+        @JsonProperty("type") private String type = "text";
+        @JsonProperty("value") private JsonNode value;
+        @JsonProperty("reason") private String reason;
+        @JsonProperty("providerOptions") private JsonNode providerOptions;
+
+        @JsonCreator
+        ToolResultOutput() {}
+
+        private ToolResultOutput(String type, JsonNode value, String reason) {
+            this.type = type;
+            this.value = value;
+            this.reason = reason;
+        }
+
+        public static ToolResultOutput text(String value) { return new ToolResultOutput("text", TextNode.valueOf(value), null); }
+        public static ToolResultOutput json(JsonNode value) { return new ToolResultOutput("json", value, null); }
+        public static ToolResultOutput executionDenied(String reason) { return new ToolResultOutput("execution-denied", null, reason); }
+        public static ToolResultOutput errorText(String value) { return new ToolResultOutput("error-text", TextNode.valueOf(value), null); }
+        public static ToolResultOutput errorJson(JsonNode value) { return new ToolResultOutput("error-json", value, null); }
+        /** `items` is a JSON array of `text` / `file` / `custom` content items. */
+        public static ToolResultOutput content(ArrayNode items) { return new ToolResultOutput("content", items, null); }
+
+        /** One of `text`, `json`, `execution-denied`, `error-text`, `error-json`, `content`. */
+        public String getType() { return type; }
+        public JsonNode getValue() { return value; }
+        public String getReason() { return reason; }
+        public JsonNode getProviderOptions() { return providerOptions; }
+
+        /** Returns a copy carrying the given provider options (ignored by `content`). */
+        public ToolResultOutput withProviderOptions(JsonNode options) {
+            ToolResultOutput copy = new ToolResultOutput(type, value, reason);
+            copy.providerOptions = options;
+            return copy;
+        }
+
+        @Override
+        public boolean equals(Object o) {
+            if (this == o) return true;
+            if (!(o instanceof ToolResultOutput)) return false;
+            ToolResultOutput that = (ToolResultOutput) o;
+            return Objects.equals(type, that.type) && Objects.equals(value, that.value)
+                && Objects.equals(reason, that.reason) && Objects.equals(providerOptions, that.providerOptions);
+        }
+
+        @Override
+        public int hashCode() { return Objects.hash(type, value, reason, providerOptions); }
+    }
+
     // ─────────────────────────────────────────────────────────────────────────────
     // ContentPart (multi-part message content).
     //
@@ -930,7 +991,7 @@ public ToolChoice deserialize(JsonParser p, DeserializationContext ctxt) throws
      * Mirrors `ContentPart.ts` (internally tagged on `type`). Shared between
      * {@link ModelMessage} (user-facing) and the provider-facing prompt. Every
      * field has a default and unknown keys are ignored on decode, so future part
-     * additions do not break existing clients. `provider_options` is a
+     * additions do not break existing clients. `providerOptions` is a
      * {@link JsonNode} because it is an opaque `Record<string, JSONObject>`.
      */
     public abstract static class ContentPart {
@@ -938,7 +999,7 @@ private ContentPart() {}
 
         public static class Text extends ContentPart {
             @JsonProperty("text") private String text = "";
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
 
             @JsonCreator
             Text() {}
@@ -977,8 +1038,8 @@ public boolean equals(Object o) {
 
         public static class Image extends ContentPart {
             @JsonProperty("image") private List<Integer> image = new ArrayList<>();
-            @JsonProperty("media_type") private String mediaType = "";
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("mediaType") private String mediaType = "";
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
 
             @JsonCreator
             Image() {}
@@ -1023,9 +1084,9 @@ public boolean equals(Object o) {
 
         public static class File extends ContentPart {
             @JsonProperty("data") private List<Integer> data = new ArrayList<>();
-            @JsonProperty("media_type") private String mediaType = "";
+            @JsonProperty("mediaType") private String mediaType = "";
             @JsonProperty("filename") private String filename;
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
 
             @JsonCreator
             File() {}
@@ -1075,9 +1136,9 @@ public boolean equals(Object o) {
 
         public static class FileBase64 extends ContentPart {
             @JsonProperty("data") private String data = "";
-            @JsonProperty("media_type") private String mediaType = "";
+            @JsonProperty("mediaType") private String mediaType = "";
             @JsonProperty("filename") private String filename;
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
 
             @JsonCreator
             FileBase64() {}
@@ -1127,8 +1188,8 @@ public boolean equals(Object o) {
 
         public static class FileUrl extends ContentPart {
             @JsonProperty("url") private String url = "";
-            @JsonProperty("media_type") private String mediaType = "";
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("mediaType") private String mediaType = "";
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
 
             @JsonCreator
             FileUrl() {}
@@ -1172,10 +1233,10 @@ public boolean equals(Object o) {
         }
 
         public static class FileReference extends ContentPart {
-            @JsonProperty("media_type") private String mediaType = "";
+            @JsonProperty("mediaType") private String mediaType = "";
             @JsonProperty("reference") private JsonNode reference = emptyObject();
             @JsonProperty("filename") private String filename;
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
 
             @JsonCreator
             FileReference() {}
@@ -1228,7 +1289,7 @@ public boolean equals(Object o) {
         public static class Reasoning extends ContentPart {
             @JsonProperty("text") private String text = "";
             @JsonProperty("signature") private String signature;
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
 
             @JsonCreator
             Reasoning() {}
@@ -1272,11 +1333,11 @@ public boolean equals(Object o) {
         }
 
         public static class ToolCall extends ContentPart {
-            @JsonProperty("tool_call_id") private String toolCallId = "";
-            @JsonProperty("tool_name") private String toolName = "";
+            @JsonProperty("toolCallId") private String toolCallId = "";
+            @JsonProperty("toolName") private String toolName = "";
             @JsonProperty("input") private JsonNode input = emptyObject();
-            @JsonProperty("provider_executed") private Boolean providerExecuted;
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("providerExecuted") private Boolean providerExecuted;
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
 
             @JsonCreator
             ToolCall() {}
@@ -1335,58 +1396,40 @@ public int hashCode() {
         }
 
         public static class ToolResult extends ContentPart {
-            @JsonProperty("tool_call_id") private String toolCallId = "";
-            @JsonProperty("result") private JsonNode result = emptyObject();
-            @JsonProperty("tool_name") private String toolName;
-            @JsonProperty("is_error") private Boolean isError;
-            @JsonProperty("preliminary") private Boolean preliminary;
-            @JsonProperty("dynamic") private Boolean dynamic;
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("toolCallId") private String toolCallId = "";
+            @JsonProperty("toolName") private String toolName = "";
+            @JsonProperty("output") private ToolResultOutput output = ToolResultOutput.text("");
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
 
             @JsonCreator
             ToolResult() {}
 
-            private ToolResult(String toolCallId, JsonNode result, String toolName, Boolean isError,
-                               Boolean preliminary, Boolean dynamic, JsonNode providerOptions) {
+            private ToolResult(String toolCallId, String toolName, ToolResultOutput output, JsonNode providerOptions) {
                 this.toolCallId = toolCallId;
-                this.result = result;
                 this.toolName = toolName;
-                this.isError = isError;
-                this.preliminary = preliminary;
-                this.dynamic = dynamic;
+                this.output = output;
                 this.providerOptions = providerOptions;
             }
 
             public String getToolCallId() { return toolCallId; }
-            public JsonNode getResult() { return result; }
             public String getToolName() { return toolName; }
-            public Boolean getIsError() { return isError; }
-            public Boolean getPreliminary() { return preliminary; }
-            public Boolean getDynamic() { return dynamic; }
+            public ToolResultOutput getOutput() { return output; }
             public JsonNode getProviderOptions() { return providerOptions; }
 
             public static Builder builder() { return new Builder(); }
 
             public static class Builder {
                 private String toolCallId = "";
-                private JsonNode result = emptyObject();
-                private String toolName;
-                private Boolean isError;
-                private Boolean preliminary;
-                private Boolean dynamic;
+                private String toolName = "";
+                private ToolResultOutput output = ToolResultOutput.text("");
                 private JsonNode providerOptions;
 
                 public Builder toolCallId(String v) { this.toolCallId = v; return this; }
-                public Builder result(JsonNode v) { this.result = v; return this; }
                 public Builder toolName(String v) { this.toolName = v; return this; }
-                public Builder isError(Boolean v) { this.isError = v; return this; }
-                public Builder preliminary(Boolean v) { this.preliminary = v; return this; }
-                public Builder dynamic(Boolean v) { this.dynamic = v; return this; }
+                public Builder output(ToolResultOutput v) { this.output = v; return this; }
                 public Builder providerOptions(JsonNode v) { this.providerOptions = v; return this; }
 
-                public ToolResult build() {
-                    return new ToolResult(toolCallId, result, toolName, isError, preliminary, dynamic, providerOptions);
-                }
+                public ToolResult build() { return new ToolResult(toolCallId, toolName, output, providerOptions); }
             }
 
             @Override
@@ -1395,22 +1438,17 @@ public boolean equals(Object o) {
                 if (!(o instanceof ToolResult)) return false;
                 ToolResult that = (ToolResult) o;
                 return Objects.equals(toolCallId, that.toolCallId)
-                    && Objects.equals(result, that.result)
                     && Objects.equals(toolName, that.toolName)
-                    && Objects.equals(isError, that.isError)
-                    && Objects.equals(preliminary, that.preliminary)
-                    && Objects.equals(dynamic, that.dynamic)
+                    && Objects.equals(output, that.output)
                     && Objects.equals(providerOptions, that.providerOptions);
             }
 
             @Override
-            public int hashCode() {
-                return Objects.hash(toolCallId, result, toolName, isError, preliminary, dynamic, providerOptions);
-            }
+            public int hashCode() { return Objects.hash(toolCallId, toolName, output, providerOptions); }
         }
         public static class Custom extends ContentPart {
             @JsonProperty("kind") private String kind;
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
             @JsonCreator Custom() {}
             public Custom(String kind, JsonNode providerOptions) {
                 this.kind = kind;
@@ -1435,8 +1473,8 @@ public static class Builder {
         }
         public static class ReasoningFile extends ContentPart {
             @JsonProperty("data") private GeneratedFileData data;
-            @JsonProperty("media_type") private String mediaType;
-            @JsonProperty("provider_options") private JsonNode providerOptions;
+            @JsonProperty("mediaType") private String mediaType;
+            @JsonProperty("providerOptions") private JsonNode providerOptions;
             @JsonCreator ReasoningFile() {}
             public ReasoningFile(GeneratedFileData data, String mediaType, JsonNode providerOptions) {
                 this.data = data;
@@ -1464,12 +1502,12 @@ public static class Builder {
             @Override public int hashCode() { return Objects.hash(data, mediaType, providerOptions); }
         }
         public static class ToolApprovalRequest extends ContentPart {
-            @JsonProperty("approval_id") private String approvalId;
-            @JsonProperty("tool_call_id") private String toolCallId;
+            @JsonProperty("approvalId") private String approvalId;
+            @JsonProperty("toolCallId") private String toolCallId;
             @JsonProperty("reason") private String reason;
-            @JsonProperty("is_automatic") private Boolean isAutomatic;
+            @JsonProperty("isAutomatic") private Boolean isAutomatic;
             @JsonProperty("signature") private String signature;
-            @JsonProperty("input_schema_input") private JsonNode inputSchemaInput;
+            @JsonProperty("inputSchemaInput") private JsonNode inputSchemaInput;
             @JsonCreator ToolApprovalRequest() {}
             public ToolApprovalRequest(String approvalId, String toolCallId, String reason, Boolean isAutomatic, String signature, JsonNode inputSchemaInput) {
                 this.approvalId = approvalId;
@@ -1526,15 +1564,15 @@ private static String variantTag(ContentPart value) {
             if (value instanceof ContentPart.Text) return "text";
             if (value instanceof ContentPart.Image) return "image";
             if (value instanceof ContentPart.File) return "file";
-            if (value instanceof ContentPart.FileBase64) return "file_base64";
-            if (value instanceof ContentPart.FileUrl) return "file_url";
-            if (value instanceof ContentPart.FileReference) return "file_reference";
+            if (value instanceof ContentPart.FileBase64) return "file-base64";
+            if (value instanceof ContentPart.FileUrl) return "file-url";
+            if (value instanceof ContentPart.FileReference) return "file-reference";
             if (value instanceof ContentPart.Reasoning) return "reasoning";
-            if (value instanceof ContentPart.ToolCall) return "tool_call";
-            if (value instanceof ContentPart.ToolResult) return "tool_result";
+            if (value instanceof ContentPart.ToolCall) return "tool-call";
+            if (value instanceof ContentPart.ToolResult) return "tool-result";
             if (value instanceof ContentPart.Custom) return "custom";
-            if (value instanceof ContentPart.ReasoningFile) return "reasoning_file";
-            if (value instanceof ContentPart.ToolApprovalRequest) return "tool_approval_request";
+            if (value instanceof ContentPart.ReasoningFile) return "reasoning-file";
+            if (value instanceof ContentPart.ToolApprovalRequest) return "tool-approval-request";
             throw new IllegalArgumentException("Unknown ContentPart: " + value);
         }
     }
@@ -1552,15 +1590,15 @@ public ContentPart deserialize(JsonParser p, DeserializationContext ctxt) throws
                 case "text": return AimuxJson.MAPPER.treeToValue(node, ContentPart.Text.class);
                 case "image": return AimuxJson.MAPPER.treeToValue(node, ContentPart.Image.class);
                 case "file": return AimuxJson.MAPPER.treeToValue(node, ContentPart.File.class);
-                case "file_base64": return AimuxJson.MAPPER.treeToValue(node, ContentPart.FileBase64.class);
-                case "file_url": return AimuxJson.MAPPER.treeToValue(node, ContentPart.FileUrl.class);
-                case "file_reference": return AimuxJson.MAPPER.treeToValue(node, ContentPart.FileReference.class);
+                case "file-base64": return AimuxJson.MAPPER.treeToValue(node, ContentPart.FileBase64.class);
+                case "file-url": return AimuxJson.MAPPER.treeToValue(node, ContentPart.FileUrl.class);
+                case "file-reference": return AimuxJson.MAPPER.treeToValue(node, ContentPart.FileReference.class);
                 case "reasoning": return AimuxJson.MAPPER.treeToValue(node, ContentPart.Reasoning.class);
-                case "tool_call": return AimuxJson.MAPPER.treeToValue(node, ContentPart.ToolCall.class);
-                case "tool_result": return AimuxJson.MAPPER.treeToValue(node, ContentPart.ToolResult.class);
+                case "tool-call": return AimuxJson.MAPPER.treeToValue(node, ContentPart.ToolCall.class);
+                case "tool-result": return AimuxJson.MAPPER.treeToValue(node, ContentPart.ToolResult.class);
                 case "custom": return AimuxJson.MAPPER.treeToValue(node, ContentPart.Custom.class);
-                case "reasoning_file": return AimuxJson.MAPPER.treeToValue(node, ContentPart.ReasoningFile.class);
-                case "tool_approval_request": return AimuxJson.MAPPER.treeToValue(node, ContentPart.ToolApprovalRequest.class);
+                case "reasoning-file": return AimuxJson.MAPPER.treeToValue(node, ContentPart.ReasoningFile.class);
+                case "tool-approval-request": return AimuxJson.MAPPER.treeToValue(node, ContentPart.ToolApprovalRequest.class);
                 default:
                     throw new IOException("Unknown ContentPart type: '" + type + "'");
             }
@@ -1757,10 +1795,10 @@ public int hashCode() {
      * backoff and the whole streamed response.
      */
     public static class TimeoutConfiguration {
-        @JsonProperty("total_ms") private Long totalMs;
-        @JsonProperty("step_ms") private Long stepMs;
-        @JsonProperty("first_chunk_ms") private Long firstChunkMs;
-        @JsonProperty("chunk_ms") private Long chunkMs;
+        @JsonProperty("totalMs") private Long totalMs;
+        @JsonProperty("stepMs") private Long stepMs;
+        @JsonProperty("firstChunkMs") private Long firstChunkMs;
+        @JsonProperty("chunkMs") private Long chunkMs;
 
         @JsonCreator
         TimeoutConfiguration() {}
@@ -1827,12 +1865,12 @@ public int hashCode() {
      * inspects and what it returns as the replacement.
      */
     public static class RawToolCall {
-        @JsonProperty("tool_call_id") private String toolCallId = "";
-        @JsonProperty("tool_name") private String toolName = "";
+        @JsonProperty("toolCallId") private String toolCallId = "";
+        @JsonProperty("toolName") private String toolName = "";
         @JsonProperty("input") private String input = "";
-        @JsonProperty("provider_executed") private Boolean providerExecuted;
+        @JsonProperty("providerExecuted") private Boolean providerExecuted;
         @JsonProperty("dynamic") private Boolean dynamic;
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
         @JsonCreator
         RawToolCall() {}
@@ -1886,9 +1924,9 @@ public RawToolCall build() {
      * {@link #getMessages()} is the conversation the model actually saw.
      */
     public static class ToolCallRepairContext {
-        @JsonProperty("tool_call") private RawToolCall toolCall;
+        @JsonProperty("toolCall") private RawToolCall toolCall;
         @JsonProperty("error") private JsonNode error;
-        @JsonProperty("input_schema") private JsonNode inputSchema;
+        @JsonProperty("inputSchema") private JsonNode inputSchema;
         @JsonProperty("tools") private List<Tool> tools;
         @JsonProperty("messages") private List<ModelMessage> messages;
         @JsonProperty("instructions") private String instructions;
@@ -1907,25 +1945,25 @@ public static class ToolCallRepairContext {
     }
 
     public static class GenerateTextOptions {
-        @JsonProperty("max_output_tokens") private Long maxOutputTokens;
+        @JsonProperty("maxOutputTokens") private Long maxOutputTokens;
         @JsonProperty("temperature") private Double temperature;
-        @JsonProperty("stop_sequences") private List<String> stopSequences;
-        @JsonProperty("top_p") private Double topP;
-        @JsonProperty("top_k") private Double topK;
-        @JsonProperty("presence_penalty") private Double presencePenalty;
-        @JsonProperty("frequency_penalty") private Double frequencyPenalty;
-        @JsonProperty("response_format") private JsonNode responseFormat;
+        @JsonProperty("stopSequences") private List<String> stopSequences;
+        @JsonProperty("topP") private Double topP;
+        @JsonProperty("topK") private Double topK;
+        @JsonProperty("presencePenalty") private Double presencePenalty;
+        @JsonProperty("frequencyPenalty") private Double frequencyPenalty;
+        @JsonProperty("responseFormat") private JsonNode responseFormat;
         @JsonProperty("seed") private Long seed;
         @JsonProperty("tools") private List<Tool> tools;
-        @JsonProperty("tool_choice") private ToolChoice toolChoice;
+        @JsonProperty("toolChoice") private ToolChoice toolChoice;
         @JsonProperty("headers") private Map<String, String> headers;
-        @JsonProperty("provider_options") private Map<String, JsonNode> providerOptions;
+        @JsonProperty("providerOptions") private Map<String, JsonNode> providerOptions;
         @JsonProperty("reasoning") private ReasoningEffort reasoning;
         @JsonProperty("instructions") private String instructions;
-        @JsonProperty("max_retries") private Long maxRetries;
-        @JsonProperty("include_raw_chunks") private Boolean includeRawChunks;
+        @JsonProperty("maxRetries") private Long maxRetries;
+        @JsonProperty("includeRawChunks") private Boolean includeRawChunks;
         @JsonProperty("timeout") private TimeoutConfiguration timeout;
-        @JsonProperty("session_id") private String sessionId;
+        @JsonProperty("sessionId") private String sessionId;
         // Host-side only (RFC-0035): a function cannot cross the C ABI, and the
         // native layer must never see this key.
         @JsonIgnore private ToolCallRepair repairToolCall;
@@ -2099,15 +2137,15 @@ public int hashCode() {
     // ─────────────────────────────────────────────────────────────────────────────
     // File bytes / file data (shared V4 file types).
     //
-    // `FileBytes` and `GeneratedFileData` are externally-tagged enums
-    // (`{"Binary": [...]}`, `{"Data": {"data": ...}}`, ...). Each is modeled with a
-    // custom serializer that dispatches on the single tag key.
+    // `FileBytes` is untagged (an int array or a base64 string); `FileData` and
+    // `GeneratedFileData` are tagged on `type` (`{"type":"data","data":...}`, ...).
+    // Each is modeled with a custom serializer.
     // ─────────────────────────────────────────────────────────────────────────────
 
     /**
      * Either raw bytes or a base64-encoded string.
      *
-     * Mirrors `FileBytes.ts`: `{"Binary": Array<number>} | {"Base64": string}`.
+     * Mirrors `FileBytes.ts`: `Array<number> | string`.
      * The binary payload is a list of byte ints (0–255).
      */
     public abstract static class FileBytes {
@@ -2158,59 +2196,48 @@ public boolean equals(Object o) {
         }
     }
 
-    /** Custom (de)serializer for {@link FileBytes} — externally tagged. */
+    /** Serializes {@link FileBytes} untagged: a JSON int array (binary) or a string (base64). */
     public static class FileBytesSerializer extends JsonSerializer<FileBytes> {
         @Override
         public void serialize(FileBytes value, JsonGenerator gen, SerializerProvider serializers) throws IOException {
-            gen.writeStartObject();
             if (value instanceof FileBytes.Binary) {
-                gen.writeArrayFieldStart("Binary");
+                gen.writeStartArray();
                 for (Integer b : ((FileBytes.Binary) value).getData()) {
                     gen.writeNumber(b);
                 }
                 gen.writeEndArray();
             } else if (value instanceof FileBytes.Base64) {
-                gen.writeStringField("Base64", ((FileBytes.Base64) value).getData());
+                gen.writeString(((FileBytes.Base64) value).getData());
             } else {
                 throw new IOException("Unknown FileBytes: " + value);
             }
-            gen.writeEndObject();
         }
     }
 
-    /** Custom (de)serializer for {@link FileBytes}. */
+    /** Deserializes {@link FileBytes}: a JSON array is binary, a JSON string is base64. */
     public static class FileBytesDeserializer extends JsonDeserializer<FileBytes> {
         @Override
         public FileBytes deserialize(JsonParser p, DeserializationContext ctxt) throws IOException {
             JsonNode node = p.getCodec().readTree(p);
-            if (!node.isObject() || node.size() != 1) {
-                throw new IOException("FileBytes must be a single-key externally-tagged object, got: " + node);
-            }
-            String tag = node.fieldNames().next();
-            JsonNode inner = node.get(tag);
-            switch (tag) {
-                case "Binary": {
-                    List<Integer> data = new ArrayList<>();
-                    if (inner.isArray()) {
-                        for (JsonNode item : inner) {
-                            data.add(item.asInt());
-                        }
-                    }
-                    return new FileBytes.Binary(data);
+            if (node.isArray()) {
+                List<Integer> data = new ArrayList<>();
+                for (JsonNode item : node) {
+                    data.add(item.asInt());
                 }
-                case "Base64":
-                    return new FileBytes.Base64(inner.asText());
-                default:
-                    throw new IOException("Unknown FileBytes tag: '" + tag + "'");
+                return new FileBytes.Binary(data);
             }
+            if (node.isTextual()) {
+                return new FileBytes.Base64(node.asText());
+            }
+            throw new IOException("FileBytes must be an int array or a base64 string, got: " + node);
         }
     }
 
     /**
      * File data as a tagged discriminated union.
      *
-     * Mirrors `FileData.ts`: `{"Data": {"data": FileBytes}} | {"Url": {"url": ...}}
-     * | {"Reference": {"reference": {...}}} | {"Text": {"text": ...}}`.
+     * Mirrors `FileData.ts`: `{"type":"data","data":FileBytes} | {"type":"url","url":...}
+     * | {"type":"reference","reference":{...}} | {"type":"text","text":...}`.
      */
     public abstract static class FileData {
         private FileData() {}
@@ -2238,7 +2265,7 @@ public boolean equals(Object o) {
 
         public static class Url extends FileData {
             @JsonProperty("url") private String url;
-            @JsonProperty("original_url") private String originalUrl;
+            @JsonProperty("originalUrl") private String originalUrl;
             @JsonCreator Url() {}
             public Url(String url, String originalUrl) {
                 this.url = url;
@@ -2304,30 +2331,26 @@ public boolean equals(Object o) {
         }
     }
 
-    /** Custom (de)serializer for {@link FileData} — externally tagged. */
+    /** Custom (de)serializer for {@link FileData} — internally tagged on `"type"`. */
     public static class FileDataSerializer extends JsonSerializer<FileData> {
         @Override
         public void serialize(FileData value, JsonGenerator gen, SerializerProvider serializers) throws IOException {
             gen.writeStartObject();
             if (value instanceof FileData.Data) {
-                gen.writeObjectFieldStart("Data");
+                gen.writeStringField("type", "data");
                 gen.writeFieldName("data");
                 gen.writeTree(AimuxJson.MAPPER.valueToTree(((FileData.Data) value).getData()));
-                gen.writeEndObject();
             } else if (value instanceof FileData.Url) {
-                gen.writeObjectFieldStart("Url");
+                gen.writeStringField("type", "url");
                 gen.writeStringField("url", ((FileData.Url) value).getUrl());
                 if (((FileData.Url) value).getOriginalUrl() != null)
-                    gen.writeStringField("original_url", ((FileData.Url) value).getOriginalUrl());
-                gen.writeEndObject();
+                    gen.writeStringField("originalUrl", ((FileData.Url) value).getOriginalUrl());
             } else if (value instanceof FileData.Reference) {
-                gen.writeObjectFieldStart("Reference");
+                gen.writeStringField("type", "reference");
                 gen.writeObjectField("reference", ((FileData.Reference) value).getReference());
-                gen.writeEndObject();
             } else if (value instanceof FileData.Text) {
-                gen.writeObjectFieldStart("Text");
+                gen.writeStringField("type", "text");
                 gen.writeStringField("text", ((FileData.Text) value).getText());
-                gen.writeEndObject();
             } else {
                 throw new IOException("Unknown FileData: " + value);
             }
@@ -2340,23 +2363,18 @@ public static class FileDataDeserializer extends JsonDeserializer<FileData> {
         @Override
         public FileData deserialize(JsonParser p, DeserializationContext ctxt) throws IOException {
             JsonNode node = p.getCodec().readTree(p);
-            if (!node.isObject() || node.size() != 1) {
-                throw new IOException("FileData must be a single-key externally-tagged object, got: " + node);
-            }
-            String tag = node.fieldNames().next();
-            JsonNode inner = node.get(tag);
-            JsonNode innerObj = inner.isObject() ? inner : AimuxJson.MAPPER.createObjectNode();
+            String tag = node.path("type").asText();
             switch (tag) {
-                case "Data":
-                    return new FileData.Data(AimuxJson.MAPPER.treeToValue(innerObj.get("data"), FileBytes.class));
-                case "Url":
-                    return new FileData.Url(innerObj.path("url").asText(), innerObj.hasNonNull("original_url") ? innerObj.get("original_url").asText() : null);
-                case "Reference":
-                    return new FileData.Reference(innerObj.get("reference") == null ? emptyObject() : innerObj.get("reference"));
-                case "Text":
-                    return new FileData.Text(innerObj.path("text").asText());
+                case "data":
+                    return new FileData.Data(AimuxJson.MAPPER.treeToValue(node.get("data"), FileBytes.class));
+                case "url":
+                    return new FileData.Url(node.path("url").asText(), node.hasNonNull("originalUrl") ? node.get("originalUrl").asText() : null);
+                case "reference":
+                    return new FileData.Reference(node.get("reference") == null ? emptyObject() : node.get("reference"));
+                case "text":
+                    return new FileData.Text(node.path("text").asText());
                 default:
-                    throw new IOException("Unknown FileData tag: '" + tag + "'");
+                    throw new IOException("Unknown FileData type: '" + tag + "'");
             }
         }
     }
@@ -2364,7 +2382,7 @@ public FileData deserialize(JsonParser p, DeserializationContext ctxt) throws IO
     /**
      * File data as a tagged discriminated union.
      *
-     * Mirrors `GeneratedFileData.ts`: `{"Data": {"data": FileBytes}} | {"Url": {"url": ...}}
+     * Mirrors `GeneratedFileData.ts`: `{"type":"data","data":FileBytes} | {"type":"url","url":...}`.
      * Generated files contain only data or a URL.
      */
     public abstract static class GeneratedFileData {
@@ -2393,7 +2411,7 @@ public boolean equals(Object o) {
 
         public static class Url extends GeneratedFileData {
             @JsonProperty("url") private String url;
-            @JsonProperty("original_url") private String originalUrl;
+            @JsonProperty("originalUrl") private String originalUrl;
             @JsonCreator Url() {}
             public Url(String url, String originalUrl) {
                 this.url = url;
@@ -2419,22 +2437,20 @@ public static class Builder {
 
     }
 
-    /** Custom (de)serializer for {@link GeneratedFileData} — externally tagged. */
+    /** Custom (de)serializer for {@link GeneratedFileData} — internally tagged on `"type"`. */
     public static class GeneratedFileDataSerializer extends JsonSerializer<GeneratedFileData> {
         @Override
         public void serialize(GeneratedFileData value, JsonGenerator gen, SerializerProvider serializers) throws IOException {
             gen.writeStartObject();
             if (value instanceof GeneratedFileData.Data) {
-                gen.writeObjectFieldStart("Data");
+                gen.writeStringField("type", "data");
                 gen.writeFieldName("data");
                 gen.writeTree(AimuxJson.MAPPER.valueToTree(((GeneratedFileData.Data) value).getData()));
-                gen.writeEndObject();
             } else if (value instanceof GeneratedFileData.Url) {
-                gen.writeObjectFieldStart("Url");
+                gen.writeStringField("type", "url");
                 gen.writeStringField("url", ((GeneratedFileData.Url) value).getUrl());
                 if (((GeneratedFileData.Url) value).getOriginalUrl() != null)
-                    gen.writeStringField("original_url", ((GeneratedFileData.Url) value).getOriginalUrl());
-                gen.writeEndObject();
+                    gen.writeStringField("originalUrl", ((GeneratedFileData.Url) value).getOriginalUrl());
             } else {
                 throw new IOException("Unknown GeneratedFileData: " + value);
             }
@@ -2447,19 +2463,14 @@ public static class GeneratedFileDataDeserializer extends JsonDeserializer<Gener
         @Override
         public GeneratedFileData deserialize(JsonParser p, DeserializationContext ctxt) throws IOException {
             JsonNode node = p.getCodec().readTree(p);
-            if (!node.isObject() || node.size() != 1) {
-                throw new IOException("GeneratedFileData must be a single-key externally-tagged object, got: " + node);
-            }
-            String tag = node.fieldNames().next();
-            JsonNode inner = node.get(tag);
-            JsonNode innerObj = inner.isObject() ? inner : AimuxJson.MAPPER.createObjectNode();
+            String tag = node.path("type").asText();
             switch (tag) {
-                case "Data":
-                    return new GeneratedFileData.Data(AimuxJson.MAPPER.treeToValue(innerObj.get("data"), FileBytes.class));
-                case "Url":
-                    return new GeneratedFileData.Url(innerObj.path("url").asText(), innerObj.hasNonNull("original_url") ? innerObj.get("original_url").asText() : null);
+                case "data":
+                    return new GeneratedFileData.Data(AimuxJson.MAPPER.treeToValue(node.get("data"), FileBytes.class));
+                case "url":
+                    return new GeneratedFileData.Url(node.path("url").asText(), node.hasNonNull("originalUrl") ? node.get("originalUrl").asText() : null);
                 default:
-                    throw new IOException("Unknown GeneratedFileData tag: '" + tag + "'");
+                    throw new IOException("Unknown GeneratedFileData type: '" + tag + "'");
             }
         }
     }
@@ -2467,20 +2478,20 @@ public GeneratedFileData deserialize(JsonParser p, DeserializationContext ctxt)
     // ─────────────────────────────────────────────────────────────────────────────
     // GenerateContent / GenerateResult (the `raw` field of GenerateTextResult).
     //
-    // `GenerateContent` is externally tagged (`{"Text": {...}}`, `{"ToolCall":
-    // {...}}`, ...). Unrecognized variants fall back to GenerateContent.Unknown
+    // `GenerateContent` is internally tagged on `type` (`{"type":"text",...}`,
+    // `{"type":"tool-call",...}`, ...). Unrecognized variants fall back to GenerateContent.Unknown
     // for forward compatibility (mirroring StreamPart).
     // ─────────────────────────────────────────────────────────────────────────────
 
     /** A URL or document used as a source for the response. */
-    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.PROPERTY, property = "source_type")
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.PROPERTY, property = "sourceType")
     @JsonSubTypes({
         @JsonSubTypes.Type(value = Source.Url.class, name = "url"),
         @JsonSubTypes.Type(value = Source.Document.class, name = "document")
     })
     public abstract static class Source {
         @JsonProperty("id") private final String id;
-        @JsonProperty("provider_metadata") private final JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private final JsonNode providerMetadata;
 
         protected Source(String id, JsonNode providerMetadata) {
             this.id = Objects.requireNonNull(id, "id");
@@ -2497,7 +2508,7 @@ public static final class Url extends Source {
             @JsonCreator
             public Url(@JsonProperty("id") String id, @JsonProperty("url") String url,
                        @JsonProperty("title") String title,
-                       @JsonProperty("provider_metadata") JsonNode providerMetadata) {
+                       @JsonProperty("providerMetadata") JsonNode providerMetadata) {
                 super(id, providerMetadata);
                 this.url = Objects.requireNonNull(url, "url");
                 this.title = title;
@@ -2517,16 +2528,16 @@ public Url(@JsonProperty("id") String id, @JsonProperty("url") String url,
         }
 
         public static final class Document extends Source {
-            @JsonProperty("media_type") private final String mediaType;
+            @JsonProperty("mediaType") private final String mediaType;
             @JsonProperty("title") private final String title;
             @JsonProperty("filename") private final String filename;
 
             @JsonCreator
-            public Document(@JsonProperty("id") String id, @JsonProperty("media_type") String mediaType,
+            public Document(@JsonProperty("id") String id, @JsonProperty("mediaType") String mediaType,
                             @JsonProperty("title") String title, @JsonProperty("filename") String filename,
-                            @JsonProperty("provider_metadata") JsonNode providerMetadata) {
+                            @JsonProperty("providerMetadata") JsonNode providerMetadata) {
                 super(id, providerMetadata);
-                this.mediaType = Objects.requireNonNull(mediaType, "media_type");
+                this.mediaType = Objects.requireNonNull(mediaType, "mediaType");
                 this.title = Objects.requireNonNull(title, "title");
                 this.filename = filename;
             }
@@ -2551,7 +2562,7 @@ public Document(@JsonProperty("id") String id, @JsonProperty("media_type") Strin
     /**
      * A content item in the generation result.
      *
-     * Mirrors `GenerateContent.ts` (externally tagged). `provider_metadata` is a
+     * Mirrors `GenerateContent.ts` (tagged on `type`). `providerMetadata` is a
      * {@link JsonNode} (`ProviderMetadata = serde_json::Value`). Every field has a
      * default; the `File` variant has no `filename` (matching Rust).
      */
@@ -2560,7 +2571,7 @@ private GenerateContent() {}
 
         public static class Text extends GenerateContent {
             @JsonProperty("text") private String text = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             Text() {}
@@ -2598,12 +2609,12 @@ public boolean equals(Object o) {
         }
 
         public static class ToolCall extends GenerateContent {
-            @JsonProperty("tool_call_id") private String toolCallId = "";
-            @JsonProperty("tool_name") private String toolName = "";
+            @JsonProperty("toolCallId") private String toolCallId = "";
+            @JsonProperty("toolName") private String toolName = "";
             @JsonProperty("input") private JsonNode input = emptyObject();
-            @JsonProperty("provider_executed") private Boolean providerExecuted;
+            @JsonProperty("providerExecuted") private Boolean providerExecuted;
             @JsonProperty("dynamic") private Boolean dynamic;
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             ToolCall() {}
@@ -2682,7 +2693,7 @@ public static class Source extends GenerateContent {
 
         public static class Reasoning extends GenerateContent {
             @JsonProperty("text") private String text = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             Reasoning() {}
@@ -2721,8 +2732,8 @@ public boolean equals(Object o) {
 
         public static class File extends GenerateContent {
             @JsonProperty("data") private GeneratedFileData data = new GeneratedFileData.Data(new FileBytes.Base64(""));
-            @JsonProperty("media_type") private String mediaType = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("mediaType") private String mediaType = "";
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             File() {}
@@ -2766,13 +2777,13 @@ public boolean equals(Object o) {
         }
 
         public static class ToolResult extends GenerateContent {
-            @JsonProperty("tool_call_id") private String toolCallId = "";
-            @JsonProperty("tool_name") private String toolName = "";
+            @JsonProperty("toolCallId") private String toolCallId = "";
+            @JsonProperty("toolName") private String toolName = "";
             @JsonProperty("result") private JsonNode result = emptyObject();
-            @JsonProperty("is_error") private Boolean isError;
+            @JsonProperty("isError") private Boolean isError;
             @JsonProperty("preliminary") private Boolean preliminary;
             @JsonProperty("dynamic") private Boolean dynamic;
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             ToolResult() {}
@@ -2842,7 +2853,7 @@ public int hashCode() {
 
         public static class Custom extends GenerateContent {
             @JsonProperty("kind") private String kind;
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
             @JsonCreator Custom() {}
             public Custom(String kind, JsonNode providerMetadata) {
                 this.kind = kind;
@@ -2867,8 +2878,8 @@ public static class Builder {
         }
         public static class ReasoningFile extends GenerateContent {
             @JsonProperty("data") private GeneratedFileData data;
-            @JsonProperty("media_type") private String mediaType;
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("mediaType") private String mediaType;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
             @JsonCreator ReasoningFile() {}
             public ReasoningFile(GeneratedFileData data, String mediaType, JsonNode providerMetadata) {
                 this.data = data;
@@ -2896,9 +2907,9 @@ public static class Builder {
             @Override public int hashCode() { return Objects.hash(data, mediaType, providerMetadata); }
         }
         public static class ToolApprovalRequest extends GenerateContent {
-            @JsonProperty("approval_id") private String approvalId;
-            @JsonProperty("tool_call_id") private String toolCallId;
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("approvalId") private String approvalId;
+            @JsonProperty("toolCallId") private String toolCallId;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
             @JsonCreator ToolApprovalRequest() {}
             public ToolApprovalRequest(String approvalId, String toolCallId, JsonNode providerMetadata) {
                 this.approvalId = approvalId;
@@ -2954,36 +2965,34 @@ public boolean equals(Object o) {
         }
     }
 
-    /** The externally-tagged variant name for this {@link GenerateContent} (e.g. "Text", "ToolCall"). */
+    /** The wire `type` tag of this {@link GenerateContent} (e.g. "text", "tool-call"). */
     public static String variantTag(GenerateContent value) {
-        if (value instanceof GenerateContent.Text) return "Text";
-        if (value instanceof GenerateContent.ToolCall) return "ToolCall";
-        if (value instanceof GenerateContent.Source) return "Source";
-        if (value instanceof GenerateContent.Reasoning) return "Reasoning";
-        if (value instanceof GenerateContent.File) return "File";
-        if (value instanceof GenerateContent.ToolResult) return "ToolResult";
-        if (value instanceof GenerateContent.Custom) return "Custom";
-        if (value instanceof GenerateContent.ReasoningFile) return "ReasoningFile";
-        if (value instanceof GenerateContent.ToolApprovalRequest) return "ToolApprovalRequest";
+        if (value instanceof GenerateContent.Text) return "text";
+        if (value instanceof GenerateContent.ToolCall) return "tool-call";
+        if (value instanceof GenerateContent.Source) return "source";
+        if (value instanceof GenerateContent.Reasoning) return "reasoning";
+        if (value instanceof GenerateContent.File) return "file";
+        if (value instanceof GenerateContent.ToolResult) return "tool-result";
+        if (value instanceof GenerateContent.Custom) return "custom";
+        if (value instanceof GenerateContent.ReasoningFile) return "reasoning-file";
+        if (value instanceof GenerateContent.ToolApprovalRequest) return "tool-approval-request";
         if (value instanceof GenerateContent.Unknown) return ((GenerateContent.Unknown) value).getTag();
         throw new IllegalArgumentException("Unknown GenerateContent: " + value);
     }
 
-    /** Custom (de)serializer for {@link GenerateContent} — externally tagged. */
+    /** Custom (de)serializer for {@link GenerateContent} — internally tagged on `"type"`. */
     public static class GenerateContentSerializer extends JsonSerializer<GenerateContent> {
         @Override
         public void serialize(GenerateContent value, JsonGenerator gen, SerializerProvider serializers) throws IOException {
-            String tag = variantTag(value);
             if (value instanceof GenerateContent.Unknown) {
-                ObjectNode out = JsonNodeFactory.instance.objectNode();
-                out.set(tag, ((GenerateContent.Unknown) value).getData());
-                gen.writeTree(out);
+                gen.writeTree(((GenerateContent.Unknown) value).getData());
                 return;
             }
             ObjectNode node = (ObjectNode) AimuxJson.INNER_MAPPER.valueToTree(
                 value instanceof GenerateContent.Source ? ((GenerateContent.Source) value).getSource() : value);
             ObjectNode out = JsonNodeFactory.instance.objectNode();
-            out.set(tag, node);
+            out.put("type", variantTag(value));
+            out.setAll(node);
             gen.writeTree(out);
         }
     }
@@ -2993,23 +3002,21 @@ public static class GenerateContentDeserializer extends JsonDeserializer<Generat
         @Override
         public GenerateContent deserialize(JsonParser p, DeserializationContext ctxt) throws IOException {
             JsonNode node = p.getCodec().readTree(p);
-            if (!node.isObject() || node.size() != 1) {
-                throw new IOException("GenerateContent must be a single-key externally-tagged object, got: " + node);
+            if (!node.isObject()) {
+                throw new IOException("GenerateContent must be a JSON object, got: " + node);
             }
-            String tag = node.fieldNames().next();
-            JsonNode inner = node.get(tag);
-            JsonNode innerObj = inner.isObject() ? inner : AimuxJson.MAPPER.createObjectNode();
+            String tag = node.path("type").asText();
             switch (tag) {
-                case "Text": return AimuxJson.MAPPER.treeToValue(innerObj, GenerateContent.Text.class);
-                case "ToolCall": return AimuxJson.MAPPER.treeToValue(innerObj, GenerateContent.ToolCall.class);
-                case "Source": return new GenerateContent.Source(AimuxJson.MAPPER.treeToValue(innerObj, Types.Source.class));
-                case "Reasoning": return AimuxJson.MAPPER.treeToValue(innerObj, GenerateContent.Reasoning.class);
-                case "File": return AimuxJson.MAPPER.treeToValue(innerObj, GenerateContent.File.class);
-                case "ToolResult": return AimuxJson.MAPPER.treeToValue(innerObj, GenerateContent.ToolResult.class);
-                case "Custom": return AimuxJson.MAPPER.treeToValue(innerObj, GenerateContent.Custom.class);
-                case "ReasoningFile": return AimuxJson.MAPPER.treeToValue(innerObj, GenerateContent.ReasoningFile.class);
-                case "ToolApprovalRequest": return AimuxJson.MAPPER.treeToValue(innerObj, GenerateContent.ToolApprovalRequest.class);
-                default: return new GenerateContent.Unknown(tag, inner);
+                case "text": return AimuxJson.MAPPER.treeToValue(node, GenerateContent.Text.class);
+                case "tool-call": return AimuxJson.MAPPER.treeToValue(node, GenerateContent.ToolCall.class);
+                case "source": return new GenerateContent.Source(AimuxJson.MAPPER.treeToValue(node, Types.Source.class));
+                case "reasoning": return AimuxJson.MAPPER.treeToValue(node, GenerateContent.Reasoning.class);
+                case "file": return AimuxJson.MAPPER.treeToValue(node, GenerateContent.File.class);
+                case "tool-result": return AimuxJson.MAPPER.treeToValue(node, GenerateContent.ToolResult.class);
+                case "custom": return AimuxJson.MAPPER.treeToValue(node, GenerateContent.Custom.class);
+                case "reasoning-file": return AimuxJson.MAPPER.treeToValue(node, GenerateContent.ReasoningFile.class);
+                case "tool-approval-request": return AimuxJson.MAPPER.treeToValue(node, GenerateContent.ToolApprovalRequest.class);
+                default: return new GenerateContent.Unknown(tag, node);
             }
         }
     }
@@ -3018,7 +3025,7 @@ public GenerateContent deserialize(JsonParser p, DeserializationContext ctxt) th
     public static class ResponseInfo {
         @JsonProperty("id") private String id;
         @JsonProperty("timestamp") private String timestamp;
-        @JsonProperty("model_id") private String modelId;
+        @JsonProperty("modelId") private String modelId;
         @JsonProperty("headers") private Map<String, String> headers;
         @JsonProperty("body") private JsonNode body;
 
@@ -3116,10 +3123,10 @@ public boolean equals(Object o) {
      */
     public static class GenerateResult {
         @JsonProperty("content") private List<GenerateContent> content = new ArrayList<>();
-        @JsonProperty("finish_reason") private FinishReason finishReason = new FinishReason();
+        @JsonProperty("finishReason") private FinishReason finishReason = new FinishReason();
         @JsonProperty("usage") private Usage usage = new Usage();
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
         @JsonProperty("response") private ResponseInfo response;
         @JsonProperty("request") private RequestInfo request;
 
@@ -3146,7 +3153,7 @@ private GenerateResult(List<GenerateContent> content, FinishReason finishReason,
         public ResponseInfo getResponse() { return response; }
         public RequestInfo getRequest() { return request; }
 
-        /** Names of the variant tags present in {@link #getContent()} (e.g. "Text", "ToolCall"). */
+        /** Wire `type` tags present in {@link #getContent()} (e.g. "text", "tool-call"). */
         public List<String> getContentVariantTags() {
             List<String> tags = new ArrayList<>();
             for (GenerateContent item : content) {
@@ -3155,7 +3162,7 @@ public List<String> getContentVariantTags() {
             return tags;
         }
 
-        /** `true` if any content item carries the given externally-tagged variant. */
+        /** `true` if any content item carries the given `type` tag. */
         public boolean hasContentVariant(String tag) {
             for (GenerateContent item : content) {
                 if (variantTag(item).equals(tag)) return true;
@@ -3216,20 +3223,21 @@ public int hashCode() {
     public static class GenerateTextResult {
         @JsonProperty("content") private List<JsonNode> content = new ArrayList<>();
         @JsonProperty("text") private String text = "";
-        @JsonProperty("tool_calls") private List<ToolCall> toolCalls = new ArrayList<>();
-        @JsonProperty("finish_reason") private FinishReason finishReason = new FinishReason();
+        @JsonProperty("toolCalls") private List<ToolCall> toolCalls = new ArrayList<>();
+        @JsonProperty("finishReason") private FinishReason finishReason = new FinishReason();
         @JsonProperty("usage") private Usage usage = new Usage();
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
         @JsonProperty("raw") private GenerateResult raw = new GenerateResult();
         @JsonProperty("reasoning") private List<JsonNode> reasoning = new ArrayList<>();
-        @JsonProperty("reasoning_text") private String reasoningText = "";
+        @JsonProperty("reasoningText") private String reasoningText = "";
         @JsonProperty("sources") private List<Source> sources = new ArrayList<>();
         @JsonProperty("files") private List<JsonNode> files = new ArrayList<>();
-        @JsonProperty("response_messages") private List<ModelMessage> responseMessages = new ArrayList<>();
-        @JsonProperty("raw_finish_reason") private String rawFinishReason;
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
-        @JsonProperty("response") private ResponseMetadata response = new ResponseMetadata();
-        @JsonProperty("total_usage") private Usage totalUsage = new Usage();
+        @JsonProperty("responseMessages") private List<ModelMessage> responseMessages = new ArrayList<>();
+        @JsonProperty("rawFinishReason") private String rawFinishReason;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
+        @JsonProperty("request") private RequestInfo request = new RequestInfo();
+        @JsonProperty("response") private ResponseInfo response = new ResponseInfo();
+        @JsonProperty("totalUsage") private Usage totalUsage = new Usage();
 
         @JsonCreator
         GenerateTextResult() {}
@@ -3239,7 +3247,7 @@ private GenerateTextResult(String text, List<ToolCall> toolCalls, FinishReason f
                                    List<JsonNode> reasoning, String reasoningText,
                                    List<Source> sources, List<JsonNode> files,
                                    List<ModelMessage> responseMessages, String rawFinishReason,
-                                   JsonNode providerMetadata, ResponseMetadata response, Usage totalUsage) {
+                                   JsonNode providerMetadata, ResponseInfo response, Usage totalUsage) {
             this.text = text;
             this.toolCalls = toolCalls;
             this.finishReason = finishReason;
@@ -3271,7 +3279,8 @@ private GenerateTextResult(String text, List<ToolCall> toolCalls, FinishReason f
         public List<ModelMessage> getResponseMessages() { return responseMessages; }
         public String getRawFinishReason() { return rawFinishReason; }
         public JsonNode getProviderMetadata() { return providerMetadata; }
-        public ResponseMetadata getResponse() { return response; }
+        public RequestInfo getRequest() { return request; }
+        public ResponseInfo getResponse() { return response; }
         public Usage getTotalUsage() { return totalUsage; }
 
         public static Builder builder() { return new Builder(); }
@@ -3290,7 +3299,7 @@ public static class Builder {
             private List<ModelMessage> responseMessages = new ArrayList<>();
             private String rawFinishReason;
             private JsonNode providerMetadata;
-            private ResponseMetadata response = new ResponseMetadata();
+            private ResponseInfo response = new ResponseInfo();
             private Usage totalUsage = new Usage();
 
             public Builder text(String v) { this.text = v; return this; }
@@ -3306,7 +3315,7 @@ public static class Builder {
             public Builder responseMessages(List<ModelMessage> v) { this.responseMessages = v; return this; }
             public Builder rawFinishReason(String v) { this.rawFinishReason = v; return this; }
             public Builder providerMetadata(JsonNode v) { this.providerMetadata = v; return this; }
-            public Builder response(ResponseMetadata v) { this.response = v; return this; }
+            public Builder response(ResponseInfo v) { this.response = v; return this; }
             public Builder totalUsage(Usage v) { this.totalUsage = v; return this; }
 
             public GenerateTextResult build() {
@@ -3356,12 +3365,12 @@ public int hashCode() {
      */
     public static class GenerateObjectResult {
         @JsonProperty("object") private JsonNode object = emptyObject();
-        @JsonProperty("finish_reason") private FinishReason finishReason = new FinishReason();
-        @JsonProperty("raw_finish_reason") private String rawFinishReason;
+        @JsonProperty("finishReason") private FinishReason finishReason = new FinishReason();
+        @JsonProperty("rawFinishReason") private String rawFinishReason;
         @JsonProperty("usage") private Usage usage = new Usage();
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
         @JsonProperty("reasoning") private String reasoning;
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
         @JsonProperty("response") private ResponseMetadata response = new ResponseMetadata();
         @JsonProperty("raw") private GenerateTextResult raw = new GenerateTextResult();
 
@@ -3457,18 +3466,19 @@ public static class StreamTextResultAggregated {
         @JsonProperty("content") private List<JsonNode> content = new ArrayList<>();
         @JsonProperty("text") private String text = "";
         @JsonProperty("reasoning") private List<JsonNode> reasoning = new ArrayList<>();
-        @JsonProperty("reasoning_text") private String reasoningText = "";
-        @JsonProperty("tool_calls") private List<ToolCall> toolCalls = new ArrayList<>();
+        @JsonProperty("reasoningText") private String reasoningText = "";
+        @JsonProperty("toolCalls") private List<ToolCall> toolCalls = new ArrayList<>();
         @JsonProperty("sources") private List<Source> sources = new ArrayList<>();
         @JsonProperty("files") private List<JsonNode> files = new ArrayList<>();
-        @JsonProperty("finish_reason") private FinishReason finishReason = new FinishReason();
-        @JsonProperty("raw_finish_reason") private String rawFinishReason;
+        @JsonProperty("finishReason") private FinishReason finishReason = new FinishReason();
+        @JsonProperty("rawFinishReason") private String rawFinishReason;
         @JsonProperty("usage") private Usage usage = new Usage();
-        @JsonProperty("total_usage") private Usage totalUsage = new Usage();
+        @JsonProperty("totalUsage") private Usage totalUsage = new Usage();
         @JsonProperty("warnings") private List<JsonNode> warnings = new ArrayList<>();
-        @JsonProperty("provider_metadata") private JsonNode providerMetadata;
-        @JsonProperty("response") private ResponseMetadata response;
-        @JsonProperty("response_messages") private List<ModelMessage> responseMessages = new ArrayList<>();
+        @JsonProperty("providerMetadata") private JsonNode providerMetadata;
+        @JsonProperty("request") private RequestInfo request = new RequestInfo();
+        @JsonProperty("response") private ResponseInfo response;
+        @JsonProperty("responseMessages") private List<ModelMessage> responseMessages = new ArrayList<>();
 
         @JsonCreator
         StreamTextResultAggregated() {}
@@ -3477,7 +3487,7 @@ private StreamTextResultAggregated(String text, List<JsonNode> reasoning, String
                                            List<ToolCall> toolCalls, List<Source> sources, List<JsonNode> files,
                                            FinishReason finishReason, String rawFinishReason, Usage usage,
                                            Usage totalUsage, List<JsonNode> warnings, JsonNode providerMetadata,
-                                           ResponseMetadata response, List<ModelMessage> responseMessages) {
+                                           ResponseInfo response, List<ModelMessage> responseMessages) {
             this.text = text;
             this.reasoning = reasoning;
             this.reasoningText = reasoningText;
@@ -3507,7 +3517,8 @@ private StreamTextResultAggregated(String text, List<JsonNode> reasoning, String
         public Usage getTotalUsage() { return totalUsage; }
         public List<JsonNode> getWarnings() { return warnings; }
         public JsonNode getProviderMetadata() { return providerMetadata; }
-        public ResponseMetadata getResponse() { return response; }
+        public RequestInfo getRequest() { return request; }
+        public ResponseInfo getResponse() { return response; }
         public List<ModelMessage> getResponseMessages() { return responseMessages; }
 
         public static Builder builder() { return new Builder(); }
@@ -3525,7 +3536,7 @@ public static class Builder {
             private Usage totalUsage = new Usage();
             private List<JsonNode> warnings = new ArrayList<>();
             private JsonNode providerMetadata;
-            private ResponseMetadata response;
+            private ResponseInfo response;
             private List<ModelMessage> responseMessages = new ArrayList<>();
 
             public Builder text(String v) { this.text = v; return this; }
@@ -3540,7 +3551,7 @@ public static class Builder {
             public Builder totalUsage(Usage v) { this.totalUsage = v; return this; }
             public Builder warnings(List<JsonNode> v) { this.warnings = v; return this; }
             public Builder providerMetadata(JsonNode v) { this.providerMetadata = v; return this; }
-            public Builder response(ResponseMetadata v) { this.response = v; return this; }
+            public Builder response(ResponseInfo v) { this.response = v; return this; }
             public Builder responseMessages(List<ModelMessage> v) { this.responseMessages = v; return this; }
 
             public StreamTextResultAggregated build() {
@@ -3582,22 +3593,23 @@ public int hashCode() {
     // ─────────────────────────────────────────────────────────────────────────────
     // StreamPart (the streaming chunk type).
     //
-    // `StreamPart` is externally tagged (`{"TextDelta": {...}}`, `{"ToolCall":
-    // {...}}`, ...). Unrecognized variants fall back to StreamPart.Unknown for
+    // `StreamPart` is tagged on `type` (`{"type":"text-delta",...}`,
+    // `{"type":"tool-call",...}`, ...). It mirrors the TextStreamPart that
+    // `aimux_stream_text` yields. Unrecognized variants fall back to StreamPart.Unknown for
     // forward compatibility.
     // ─────────────────────────────────────────────────────────────────────────────
 
     /**
      * A single streaming chunk.
      *
-     * Mirrors `StreamPart.ts` (externally tagged).
+     * Mirrors `TextStreamPart.ts` (tagged on `type`).
      */
     public abstract static class StreamPart {
         private StreamPart() {}
 
         public static class TextStart extends StreamPart {
             @JsonProperty("id") private String id = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             TextStart() {}
@@ -3637,7 +3649,7 @@ public boolean equals(Object o) {
         public static class TextDelta extends StreamPart {
             @JsonProperty("id") private String id = "";
             @JsonProperty("delta") private String delta = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             TextDelta() {}
@@ -3682,7 +3694,7 @@ public boolean equals(Object o) {
 
         public static class TextEnd extends StreamPart {
             @JsonProperty("id") private String id = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             TextEnd() {}
@@ -3751,9 +3763,9 @@ public boolean equals(Object o) {
         }
 
         public static class Finish extends StreamPart {
-            @JsonProperty("finish_reason") private FinishReason finishReason = new FinishReason();
+            @JsonProperty("finishReason") private FinishReason finishReason = new FinishReason();
             @JsonProperty("usage") private Usage usage = new Usage();
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             Finish() {}
@@ -3796,13 +3808,43 @@ public boolean equals(Object o) {
             public int hashCode() { return Objects.hash(finishReason, usage, providerMetadata); }
         }
 
+        /** End of one step, carrying the step's response (`finish-step`). */
+        public static class FinishStep extends StreamPart {
+            @JsonProperty("finishReason") private FinishReason finishReason = new FinishReason();
+            @JsonProperty("usage") private Usage usage = new Usage();
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
+            @JsonProperty("response") private ResponseInfo response = new ResponseInfo();
+
+            @JsonCreator
+            FinishStep() {}
+
+            public FinishReason getFinishReason() { return finishReason; }
+            public Usage getUsage() { return usage; }
+            public JsonNode getProviderMetadata() { return providerMetadata; }
+            public ResponseInfo getResponse() { return response; }
+
+            @Override
+            public boolean equals(Object o) {
+                if (this == o) return true;
+                if (!(o instanceof FinishStep)) return false;
+                FinishStep that = (FinishStep) o;
+                return Objects.equals(finishReason, that.finishReason)
+                    && Objects.equals(usage, that.usage)
+                    && Objects.equals(providerMetadata, that.providerMetadata)
+                    && Objects.equals(response, that.response);
+            }
+
+            @Override
+            public int hashCode() { return Objects.hash(finishReason, usage, providerMetadata, response); }
+        }
+
         public static class ToolInputStart extends StreamPart {
             @JsonProperty("id") private String id = "";
-            @JsonProperty("tool_name") private String toolName = "";
-            @JsonProperty("provider_executed") private Boolean providerExecuted;
+            @JsonProperty("toolName") private String toolName = "";
+            @JsonProperty("providerExecuted") private Boolean providerExecuted;
             @JsonProperty("dynamic") private Boolean dynamic;
             @JsonProperty("title") private String title;
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             ToolInputStart() {}
@@ -3868,7 +3910,7 @@ public int hashCode() {
         public static class ToolInputDelta extends StreamPart {
             @JsonProperty("id") private String id = "";
             @JsonProperty("delta") private String delta = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             ToolInputDelta() {}
@@ -3913,7 +3955,7 @@ public boolean equals(Object o) {
 
         public static class ToolInputEnd extends StreamPart {
             @JsonProperty("id") private String id = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             ToolInputEnd() {}
@@ -3951,12 +3993,12 @@ public boolean equals(Object o) {
         }
 
         public static class ToolCall extends StreamPart {
-            @JsonProperty("tool_call_id") private String toolCallId = "";
-            @JsonProperty("tool_name") private String toolName = "";
+            @JsonProperty("toolCallId") private String toolCallId = "";
+            @JsonProperty("toolName") private String toolName = "";
             @JsonProperty("input") private JsonNode input = emptyObject();
-            @JsonProperty("provider_executed") private Boolean providerExecuted;
+            @JsonProperty("providerExecuted") private Boolean providerExecuted;
             @JsonProperty("dynamic") private Boolean dynamic;
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
             @JsonProperty("invalid") private Boolean invalid;
             @JsonProperty("error") private JsonNode error;
 
@@ -4037,13 +4079,13 @@ public int hashCode() {
         }
 
         public static class ToolResult extends StreamPart {
-            @JsonProperty("tool_call_id") private String toolCallId = "";
-            @JsonProperty("tool_name") private String toolName = "";
+            @JsonProperty("toolCallId") private String toolCallId = "";
+            @JsonProperty("toolName") private String toolName = "";
             @JsonProperty("result") private JsonNode result = emptyObject();
-            @JsonProperty("is_error") private Boolean isError;
+            @JsonProperty("isError") private Boolean isError;
             @JsonProperty("preliminary") private Boolean preliminary;
             @JsonProperty("dynamic") private Boolean dynamic;
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             ToolResult() {}
@@ -4114,8 +4156,8 @@ public int hashCode() {
         /** A file generated by the model (e.g. an image or document). */
         public static class File extends StreamPart {
             @JsonProperty("data") private GeneratedFileData data = new GeneratedFileData.Data(new FileBytes.Base64(""));
-            @JsonProperty("media_type") private String mediaType = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("mediaType") private String mediaType = "";
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             File() {}
@@ -4160,7 +4202,7 @@ public boolean equals(Object o) {
 
         public static class ReasoningStart extends StreamPart {
             @JsonProperty("id") private String id = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             ReasoningStart() {}
@@ -4200,7 +4242,7 @@ public boolean equals(Object o) {
         public static class ReasoningDelta extends StreamPart {
             @JsonProperty("id") private String id = "";
             @JsonProperty("delta") private String delta = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             ReasoningDelta() {}
@@ -4245,7 +4287,7 @@ public boolean equals(Object o) {
 
         public static class ReasoningEnd extends StreamPart {
             @JsonProperty("id") private String id = "";
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
 
             @JsonCreator
             ReasoningEnd() {}
@@ -4285,7 +4327,7 @@ public boolean equals(Object o) {
         public static class ResponseMetadata extends StreamPart {
             @JsonProperty("id") private String id;
             @JsonProperty("timestamp") private String timestamp;
-            @JsonProperty("model_id") private String modelId;
+            @JsonProperty("modelId") private String modelId;
 
             @JsonCreator
             ResponseMetadata() {}
@@ -4341,7 +4383,7 @@ public static class Source extends StreamPart {
         }
 
         public static class Raw extends StreamPart {
-            @JsonProperty("raw_value") private JsonNode rawValue = emptyObject();
+            @JsonProperty("rawValue") private JsonNode rawValue = emptyObject();
 
             @JsonCreator
             Raw() {}
@@ -4404,7 +4446,7 @@ public boolean equals(Object o) {
 
         public static class Custom extends StreamPart {
             @JsonProperty("kind") private String kind;
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
             @JsonCreator Custom() {}
             public Custom(String kind, JsonNode providerMetadata) {
                 this.kind = kind;
@@ -4429,7 +4471,7 @@ public static class Builder {
         }
         public static class ReasoningFile extends StreamPart {
             @JsonProperty("file") private GenerateContent.File file;
-            @JsonProperty("provider_metadata") private JsonNode providerMetadata;
+            @JsonProperty("providerMetadata") private JsonNode providerMetadata;
             @JsonCreator ReasoningFile() {}
             public ReasoningFile(GenerateContent.File file, JsonNode providerMetadata) {
                 this.file = file;
@@ -4453,10 +4495,10 @@ public static class Builder {
             @Override public int hashCode() { return Objects.hash(file, providerMetadata); }
         }
         public static class ToolApprovalRequest extends StreamPart {
-            @JsonProperty("approval_id") private String approvalId;
-            @JsonProperty("tool_call") private Types.ToolCall toolCall;
+            @JsonProperty("approvalId") private String approvalId;
+            @JsonProperty("toolCall") private Types.ToolCall toolCall;
             @JsonProperty("reason") private String reason;
-            @JsonProperty("is_automatic") private Boolean isAutomatic;
+            @JsonProperty("isAutomatic") private Boolean isAutomatic;
             @JsonProperty("signature") private String signature;
             @JsonCreator ToolApprovalRequest() {}
             public ToolApprovalRequest(String approvalId, Types.ToolCall toolCall, String reason, Boolean isAutomatic, String signature) {
@@ -4521,48 +4563,47 @@ public boolean equals(Object o) {
         }
     }
 
-    /** The externally-tagged variant name for this {@link StreamPart} (e.g. "TextDelta", "ToolCall"). */
+    /** The wire `type` tag of this {@link StreamPart} (e.g. "text-delta", "tool-call"). */
     public static String variantTag(StreamPart value) {
-        if (value instanceof StreamPart.TextStart) return "TextStart";
-        if (value instanceof StreamPart.TextDelta) return "TextDelta";
-        if (value instanceof StreamPart.TextEnd) return "TextEnd";
-        if (value instanceof StreamPart.StreamStart) return "StreamStart";
-        if (value instanceof StreamPart.Finish) return "Finish";
-        if (value instanceof StreamPart.ToolInputStart) return "ToolInputStart";
-        if (value instanceof StreamPart.ToolInputDelta) return "ToolInputDelta";
-        if (value instanceof StreamPart.ToolInputEnd) return "ToolInputEnd";
-        if (value instanceof StreamPart.ToolCall) return "ToolCall";
-        if (value instanceof StreamPart.ToolResult) return "ToolResult";
-        if (value instanceof StreamPart.File) return "File";
-        if (value instanceof StreamPart.ReasoningStart) return "ReasoningStart";
-        if (value instanceof StreamPart.ReasoningDelta) return "ReasoningDelta";
-        if (value instanceof StreamPart.ReasoningEnd) return "ReasoningEnd";
-        if (value instanceof StreamPart.ResponseMetadata) return "ResponseMetadata";
-        if (value instanceof StreamPart.Source) return "Source";
-        if (value instanceof StreamPart.Raw) return "Raw";
-        if (value instanceof StreamPart.Error) return "Error";
-        if (value instanceof StreamPart.Custom) return "Custom";
-        if (value instanceof StreamPart.ReasoningFile) return "ReasoningFile";
-        if (value instanceof StreamPart.ToolApprovalRequest) return "ToolApprovalRequest";
+        if (value instanceof StreamPart.TextStart) return "text-start";
+        if (value instanceof StreamPart.TextDelta) return "text-delta";
+        if (value instanceof StreamPart.TextEnd) return "text-end";
+        if (value instanceof StreamPart.StreamStart) return "stream-start";
+        if (value instanceof StreamPart.Finish) return "finish";
+        if (value instanceof StreamPart.FinishStep) return "finish-step";
+        if (value instanceof StreamPart.ToolInputStart) return "tool-input-start";
+        if (value instanceof StreamPart.ToolInputDelta) return "tool-input-delta";
+        if (value instanceof StreamPart.ToolInputEnd) return "tool-input-end";
+        if (value instanceof StreamPart.ToolCall) return "tool-call";
+        if (value instanceof StreamPart.ToolResult) return "tool-result";
+        if (value instanceof StreamPart.File) return "file";
+        if (value instanceof StreamPart.ReasoningStart) return "reasoning-start";
+        if (value instanceof StreamPart.ReasoningDelta) return "reasoning-delta";
+        if (value instanceof StreamPart.ReasoningEnd) return "reasoning-end";
+        if (value instanceof StreamPart.ResponseMetadata) return "response-metadata";
+        if (value instanceof StreamPart.Source) return "source";
+        if (value instanceof StreamPart.Raw) return "raw";
+        if (value instanceof StreamPart.Error) return "error";
+        if (value instanceof StreamPart.Custom) return "custom";
+        if (value instanceof StreamPart.ReasoningFile) return "reasoning-file";
+        if (value instanceof StreamPart.ToolApprovalRequest) return "tool-approval-request";
         if (value instanceof StreamPart.Unknown) return ((StreamPart.Unknown) value).getTag();
         throw new IllegalArgumentException("Unknown StreamPart: " + value);
     }
 
-    /** Custom (de)serializer for {@link StreamPart} — externally tagged. */
+    /** Custom (de)serializer for {@link StreamPart} — internally tagged on `"type"`. */
     public static class StreamPartSerializer extends JsonSerializer<StreamPart> {
         @Override
         public void serialize(StreamPart value, JsonGenerator gen, SerializerProvider serializers) throws IOException {
-            String tag = variantTag(value);
             if (value instanceof StreamPart.Unknown) {
-                ObjectNode out = JsonNodeFactory.instance.objectNode();
-                out.set(tag, ((StreamPart.Unknown) value).getData());
-                gen.writeTree(out);
+                gen.writeTree(((StreamPart.Unknown) value).getData());
                 return;
             }
             ObjectNode node = (ObjectNode) AimuxJson.INNER_MAPPER.valueToTree(
                 value instanceof StreamPart.Source ? ((StreamPart.Source) value).getSource() : value);
             ObjectNode out = JsonNodeFactory.instance.objectNode();
-            out.set(tag, node);
+            out.put("type", variantTag(value));
+            out.setAll(node);
             gen.writeTree(out);
         }
     }
@@ -4572,35 +4613,34 @@ public static class StreamPartDeserializer extends JsonDeserializer<StreamPart>
         @Override
         public StreamPart deserialize(JsonParser p, DeserializationContext ctxt) throws IOException {
             JsonNode node = p.getCodec().readTree(p);
-            if (!node.isObject() || node.size() != 1) {
-                throw new IOException("StreamPart must be a single-key externally-tagged object, got: " + node);
+            if (!node.isObject()) {
+                throw new IOException("StreamPart must be a JSON object, got: " + node);
             }
-            String tag = node.fieldNames().next();
-            JsonNode inner = node.get(tag);
-            JsonNode innerObj = inner.isObject() ? inner : AimuxJson.MAPPER.createObjectNode();
+            String tag = node.path("type").asText();
             switch (tag) {
-                case "TextStart": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.TextStart.class);
-                case "TextDelta": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.TextDelta.class);
-                case "TextEnd": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.TextEnd.class);
-                case "StreamStart": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.StreamStart.class);
-                case "Finish": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.Finish.class);
-                case "ToolInputStart": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ToolInputStart.class);
-                case "ToolInputDelta": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ToolInputDelta.class);
-                case "ToolInputEnd": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ToolInputEnd.class);
-                case "ToolCall": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ToolCall.class);
-                case "ToolResult": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ToolResult.class);
-                case "File": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.File.class);
-                case "ReasoningStart": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ReasoningStart.class);
-                case "ReasoningDelta": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ReasoningDelta.class);
-                case "ReasoningEnd": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ReasoningEnd.class);
-                case "ResponseMetadata": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ResponseMetadata.class);
-                case "Source": return new StreamPart.Source(AimuxJson.MAPPER.treeToValue(innerObj, Types.Source.class));
-                case "Raw": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.Raw.class);
-                case "Error": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.Error.class);
-                case "Custom": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.Custom.class);
-                case "ReasoningFile": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ReasoningFile.class);
-                case "ToolApprovalRequest": return AimuxJson.MAPPER.treeToValue(innerObj, StreamPart.ToolApprovalRequest.class);
-                default: return new StreamPart.Unknown(tag, inner);
+                case "text-start": return AimuxJson.MAPPER.treeToValue(node, StreamPart.TextStart.class);
+                case "text-delta": return AimuxJson.MAPPER.treeToValue(node, StreamPart.TextDelta.class);
+                case "text-end": return AimuxJson.MAPPER.treeToValue(node, StreamPart.TextEnd.class);
+                case "stream-start": return AimuxJson.MAPPER.treeToValue(node, StreamPart.StreamStart.class);
+                case "finish": return AimuxJson.MAPPER.treeToValue(node, StreamPart.Finish.class);
+                case "finish-step": return AimuxJson.MAPPER.treeToValue(node, StreamPart.FinishStep.class);
+                case "tool-input-start": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ToolInputStart.class);
+                case "tool-input-delta": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ToolInputDelta.class);
+                case "tool-input-end": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ToolInputEnd.class);
+                case "tool-call": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ToolCall.class);
+                case "tool-result": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ToolResult.class);
+                case "file": return AimuxJson.MAPPER.treeToValue(node, StreamPart.File.class);
+                case "reasoning-start": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ReasoningStart.class);
+                case "reasoning-delta": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ReasoningDelta.class);
+                case "reasoning-end": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ReasoningEnd.class);
+                case "response-metadata": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ResponseMetadata.class);
+                case "source": return new StreamPart.Source(AimuxJson.MAPPER.treeToValue(node, Types.Source.class));
+                case "raw": return AimuxJson.MAPPER.treeToValue(node, StreamPart.Raw.class);
+                case "error": return AimuxJson.MAPPER.treeToValue(node, StreamPart.Error.class);
+                case "custom": return AimuxJson.MAPPER.treeToValue(node, StreamPart.Custom.class);
+                case "reasoning-file": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ReasoningFile.class);
+                case "tool-approval-request": return AimuxJson.MAPPER.treeToValue(node, StreamPart.ToolApprovalRequest.class);
+                default: return new StreamPart.Unknown(tag, node);
             }
         }
     }
@@ -4613,7 +4653,8 @@ private static ObjectNode emptyObject() {
     // OpenAI Chat Completions output (RFC-0026).
     //
     // Mirrors `aimux-core::openai_output`. Field names are camelCase in Java and
-    // mapped to the wire's snake_case via @JsonProperty. The `type` field is JSON
+    // mapped to the OpenAI wire's snake_case via @JsonProperty (this vendor shape is
+    // not part of the AI SDK JSON). The `type` field is JSON
     // `"type"` (Rust `#[serde(rename = "type")]`) → `toolType`. Arbitrary-JSON
     // fields (`logprobs`, `annotations`) are JsonNode.
     // ─────────────────────────────────────────────────────────────────────────────
diff --git a/bindings/java/src/test/java/ai/arcships/aimux/AimuxExceptionTest.java b/bindings/java/src/test/java/ai/arcships/aimux/AimuxExceptionTest.java
index b3d493d49..dd4f5ba41 100644
--- a/bindings/java/src/test/java/ai/arcships/aimux/AimuxExceptionTest.java
+++ b/bindings/java/src/test/java/ai/arcships/aimux/AimuxExceptionTest.java
@@ -88,7 +88,7 @@ void deadHandleIsBindingInvariant() {
 
     @Test
     void ffiInvalidWireJsonIsBindingInvariant() {
-        Pointer eh = AimuxFFI.INSTANCE.aimux_register_providers("{not json");
+        Pointer eh = AimuxFFI.INSTANCE.aimux_init_proxy("{not json");
         assertThat(eh).isNotNull();
         assertThat(AimuxFFI.INSTANCE.aimux_error_code(eh)).isEqualTo(202);
         RuntimeException e = AimuxResult.expectAimuxError(eh, "ctx");
@@ -112,19 +112,19 @@ void expectRecordingErrorRejectsAnAiMuxErrorCode() {
 
     @Test
     void trailingGarbageJsonIsIllegalArgument() {
-        assertThatThrownBy(() -> Aimux.registerProviders("{} garbage"))
+        assertThatThrownBy(() -> Aimux.initProxy("{} garbage"))
             .isInstanceOf(IllegalArgumentException.class)
             .hasMessageContaining("configJson");
-        assertThatThrownBy(() -> Aimux.registerProviders(""))
+        assertThatThrownBy(() -> Aimux.initProxy(""))
             .isInstanceOf(IllegalArgumentException.class)
             .hasMessageContaining("configJson: invalid JSON: empty");
-        assertThatThrownBy(() -> Aimux.registerProviders(null))
+        assertThatThrownBy(() -> Aimux.initProxy(null))
             .isInstanceOf(NullPointerException.class);
     }
 
     @Test
     void malformedConfigJsonIsIllegalArgument() {
-        assertThatThrownBy(() -> Aimux.registerProviders("{not json"))
+        assertThatThrownBy(() -> Aimux.initProxy("{not json"))
             .isInstanceOf(IllegalArgumentException.class)
             .isNotInstanceOf(AimuxException.class)
             .hasMessageContaining("configJson");
@@ -248,12 +248,13 @@ void loadErrorsHaveTheirOwnCodesNamesAndNoEnvVarLocally() {
 
     @Test
     void codesOutsideTheRustEnumAreRejected() {
-        // Tool errors occupy 15..17 and 18 / 19 are the missing key / setting;
-        // retired code 4 and values beyond 19 are header/library mismatches.
+        // Tool errors occupy 15..17, 18 / 19 are the missing key / setting and 20
+        // is NoOutputGenerated; retired code 4 and values beyond 20 are
+        // header/library mismatches.
         assertThat(AimuxException.of(15, "")).isInstanceOf(AimuxException.NoSuchToolError.class);
         assertThatThrownBy(() -> AimuxException.of(4, ""))
             .isInstanceOf(IllegalStateException.class);
-        assertThatThrownBy(() -> AimuxException.of(20, ""))
+        assertThatThrownBy(() -> AimuxException.of(21, ""))
             .isInstanceOf(IllegalStateException.class);
         assertThatThrownBy(() -> AimuxException.of(999, "m"))
             .isInstanceOf(IllegalStateException.class);
diff --git a/bindings/java/src/test/java/ai/arcships/aimux/MockProviderServer.java b/bindings/java/src/test/java/ai/arcships/aimux/MockProviderServer.java
index 599e35663..576058d44 100644
--- a/bindings/java/src/test/java/ai/arcships/aimux/MockProviderServer.java
+++ b/bindings/java/src/test/java/ai/arcships/aimux/MockProviderServer.java
@@ -105,6 +105,14 @@ public String baseUrl() {
         return "http://127.0.0.1:" + port();
     }
 
+    /**
+     * Provider config for an OpenAI-compatible chat-completions model pointed
+     * at this server (the OpenAI factory itself speaks the Responses API).
+     */
+    public String chatConfig() {
+        return "{\"baseUrl\":\"" + baseUrl() + "\"}";
+    }
+
     public String lastRequestBody() {
         return lastRequestBody;
     }
diff --git a/bindings/java/src/test/java/ai/arcships/aimux/ModelTest.java b/bindings/java/src/test/java/ai/arcships/aimux/ModelTest.java
index 0edddeec5..9ce6409ba 100644
--- a/bindings/java/src/test/java/ai/arcships/aimux/ModelTest.java
+++ b/bindings/java/src/test/java/ai/arcships/aimux/ModelTest.java
@@ -65,7 +65,7 @@ void providerCreatesModelInstanceFromRegistry() {
     @Test
     void generateTextReturnsJsonString() {
         try (MockProviderServer server = new MockProviderServer();
-             Model model = Model.openaiWithBase("sk-test-fake-key", "gpt-4o-mini", server.baseUrl())) {
+             Model model = Model.provider("deepseek", "sk-test-fake-key", "gpt-4o-mini", server.chatConfig())) {
             server.setResponseBody(plainOpenAiResponse());
 
             String resultJson = model.generateText("\"Hello\"");
@@ -105,7 +105,7 @@ void closeThenCallThrows() {
     @Test
     void streamTextInvokesOnDone() {
         try (MockProviderServer server = new MockProviderServer();
-             Model model = Model.openaiWithBase("sk-test-fake-key", "gpt-4o-mini", server.baseUrl())) {
+             Model model = Model.provider("deepseek", "sk-test-fake-key", "gpt-4o-mini", server.chatConfig())) {
             server.setContentType("text/event-stream");
             server.setResponseBody(openAiSseResponse());
 
@@ -146,7 +146,7 @@ void streamTextInvokesOnDone() {
     @Test
     void streamTextStreamIteratesParts() {
         try (MockProviderServer server = new MockProviderServer();
-             Model model = Model.openaiWithBase("sk-test-fake-key", "gpt-4o-mini", server.baseUrl())) {
+             Model model = Model.provider("deepseek", "sk-test-fake-key", "gpt-4o-mini", server.chatConfig())) {
             server.setContentType("text/event-stream");
             server.setResponseBody(openAiSseResponse());
 
@@ -184,7 +184,7 @@ void streamTextStreamReturnsAStream() {
     @Test
     void concurrentCloseAndGenerateDoesNotCrash() throws InterruptedException {
         try (MockProviderServer server = new MockProviderServer();
-             Model model = Model.openaiWithBase("sk-test-fake-key", "gpt-4o-mini", server.baseUrl())) {
+             Model model = Model.provider("deepseek", "sk-test-fake-key", "gpt-4o-mini", server.chatConfig())) {
             server.setResponseBody(plainOpenAiResponse());
 
             final int iterations = 300;
diff --git a/bindings/java/src/test/java/ai/arcships/aimux/MultimodalCoverageTest.java b/bindings/java/src/test/java/ai/arcships/aimux/MultimodalCoverageTest.java
index b5e01c8c3..ba1e75df1 100644
--- a/bindings/java/src/test/java/ai/arcships/aimux/MultimodalCoverageTest.java
+++ b/bindings/java/src/test/java/ai/arcships/aimux/MultimodalCoverageTest.java
@@ -158,8 +158,8 @@ void searchParsesAnswerField() {
         try (SearchModel model = SearchModel.tavilyWithBase("sk-test", server.baseUrl())) {
             JSONObject opts = new JSONObject()
                 .put("query", "What is Rust?")
-                .put("max_results", 5)
-                .put("provider_options", new JSONObject());
+                .put("maxResults", 5)
+                .put("providerOptions", new JSONObject());
             JSONObject result = new JSONObject(model.search(opts.toString()));
 
             assertThat(result.getJSONArray("results").length()).isEqualTo(1);
@@ -176,7 +176,7 @@ void deepseekGenerateTextViaMock() {
                 + "\"choices\":[{\"index\":0,\"message\":{\"role\":\"assistant\",\"content\":\"Hi\"},\"finish_reason\":\"stop\"}],"
                 + "\"usage\":{\"prompt_tokens\":1,\"completion_tokens\":1,\"total_tokens\":2}}");
 
-        try (Model model = Model.openaiWithBase("sk-test", "deepseek-chat", server.baseUrl())) {
+        try (Model model = Model.provider("deepseek", "sk-test-fake-key", "deepseek-chat", server.chatConfig())) {
             String result = model.generateText("\"Say hi\"");
             assertThat(new JSONObject(result).getString("text")).isEqualTo("Hi");
         }
diff --git a/bindings/java/src/test/java/ai/arcships/aimux/MultimodalE2ETest.java b/bindings/java/src/test/java/ai/arcships/aimux/MultimodalE2ETest.java
index 7ccfdb3de..7680fb087 100644
--- a/bindings/java/src/test/java/ai/arcships/aimux/MultimodalE2ETest.java
+++ b/bindings/java/src/test/java/ai/arcships/aimux/MultimodalE2ETest.java
@@ -76,15 +76,13 @@ void speechGenerateReturnsBinaryAudio() {
             JSONObject opts = new JSONObject()
                 .put("text", "Hi")
                 .put("voice", "alloy")
-                .put("output_format", "mp3")
-                .put("provider_options", new JSONObject());
+                .put("outputFormat", "mp3")
+                .put("providerOptions", new JSONObject());
 
             JSONObject result = new JSONObject(model.generate(opts.toString()));
 
-            // Wire format: {"audio":{"Binary":[<bytes>]}, ...}
-            JSONObject audio = result.getJSONObject("audio");
-            assertThat(audio.has("Binary")).isTrue();
-            assertThat(audio.getJSONArray("Binary").length()).isGreaterThan(0);
+            // Wire format: {"audio":[<bytes>], ...} (untagged: int array = binary)
+            assertThat(result.getJSONArray("audio").length()).isGreaterThan(0);
         }
     }
 
@@ -99,12 +97,12 @@ void imageGenerateParsesBase64Images() {
             JSONObject opts = new JSONObject()
                 .put("prompt", "otter")
                 .put("n", 1)
-                .put("provider_options", new JSONObject());
+                .put("providerOptions", new JSONObject());
 
             JSONObject result = new JSONObject(model.generate(opts.toString()));
 
-            // Wire format: {"images":{"Base64":["aW1hZ2Ux"]}, ...}
-            JSONArray base64 = result.getJSONObject("images").getJSONArray("Base64");
+            // Wire format: {"images":["aW1hZ2Ux"], ...} (untagged: strings = base64)
+            JSONArray base64 = result.getJSONArray("images");
             assertThat(base64.length()).isEqualTo(1);
             assertThat(base64.getString(0)).isEqualTo("aW1hZ2Ux");
         }
@@ -137,18 +135,18 @@ void rerankParsesRanking() {
                 "sk-test", "rerank-v3.0", server.baseUrl())) {
             JSONObject opts = new JSONObject()
                 .put("query", "which?")
-                .put("documents", new JSONObject("{\"Text\":{\"values\":[\"doc1\",\"doc2\"]}}"))
-                .put("top_n", 2)
-                .put("provider_options", new JSONObject());
+                .put("documents", new JSONObject("{\"type\":\"text\",\"values\":[\"doc1\",\"doc2\"]}"))
+                .put("topN", 2)
+                .put("providerOptions", new JSONObject());
 
             JSONObject result = new JSONObject(model.rerank(opts.toString()));
 
-            // Wire format: {"ranking":[{"index":1,"relevance_score":0.95}, ...], ...}
+            // Wire format: {"ranking":[{"index":1,"relevanceScore":0.95}, ...], ...}
             JSONArray ranking = result.getJSONArray("ranking");
             assertThat(ranking.length()).isEqualTo(2);
             JSONObject first = ranking.getJSONObject(0);
             assertThat(first.getInt("index")).isEqualTo(1);
-            assertThat(first.getDouble("relevance_score")).isEqualTo(0.95);
+            assertThat(first.getDouble("relevanceScore")).isEqualTo(0.95);
         }
     }
 
@@ -165,8 +163,8 @@ void searchParsesResults() {
                 "sk-test", server.baseUrl())) {
             JSONObject opts = new JSONObject()
                 .put("query", "What is Rust?")
-                .put("max_results", 5)
-                .put("provider_options", new JSONObject());
+                .put("maxResults", 5)
+                .put("providerOptions", new JSONObject());
 
             JSONObject result = new JSONObject(model.search(opts.toString()));
 
@@ -189,8 +187,8 @@ void uploadFileParsesProviderReference() {
         try (Files files = Files.openaiWithBase("sk-test", server.baseUrl())) {
             JSONObject result = new JSONObject(files.uploadFile("dGVzdA==", "application/pdf"));
 
-            // Wire format: {"provider_reference":{"openai":"file-abc"}, ...}
-            assertThat(result.getJSONObject("provider_reference").getString("openai"))
+            // Wire format: {"providerReference":{"openai":"file-abc"}, ...}
+            assertThat(result.getJSONObject("providerReference").getString("openai"))
                 .isEqualTo("file-abc");
         }
     }
@@ -211,13 +209,13 @@ void videoConstructionAndResultParsing() {
             assertThat(model).isNotNull();
         }
 
-        // Wire format: {"videos":[{"Url":{"url":"...","media_type":"..."}}], ...}
+        // Wire format: {"videos":[{"type":"url","url":"...","mediaType":"..."}], ...}
         JSONObject parsed = new JSONObject(
-            "{\"videos\":[{\"Url\":{\"url\":\"https://example.com/v.mp4\","
-                + "\"media_type\":\"video/mp4\"}}]}");
+            "{\"videos\":[{\"type\":\"url\",\"url\":\"https://example.com/v.mp4\","
+                + "\"mediaType\":\"video/mp4\"}]}");
         JSONArray videos = parsed.getJSONArray("videos");
         assertThat(videos.length()).isEqualTo(1);
-        assertThat(videos.getJSONObject(0).getJSONObject("Url").getString("url"))
+        assertThat(videos.getJSONObject(0).getString("url"))
             .isEqualTo("https://example.com/v.mp4");
     }
 }
diff --git a/bindings/java/src/test/java/ai/arcships/aimux/MultimodalTypesTest.java b/bindings/java/src/test/java/ai/arcships/aimux/MultimodalTypesTest.java
index 99f709af1..19b4799ea 100644
--- a/bindings/java/src/test/java/ai/arcships/aimux/MultimodalTypesTest.java
+++ b/bindings/java/src/test/java/ai/arcships/aimux/MultimodalTypesTest.java
@@ -26,11 +26,10 @@
  *
  * <p>Special attention is paid to the three hand-written
  * {@code JsonSerializer}/{@code JsonDeserializer} pairs —
- * {@link MultimodalTypes.AudioData} (Base64/Binary),
- * {@link MultimodalTypes.ImageOutputs} (Base64/Binary) and
- * {@link MultimodalTypes.VideoData} (Url/Base64/Binary) — which implement the
- * serde-style externally-tagged enum wire format ({@code {"Base64": ...}}) and
- * are the most bug-prone part of the file. Dedicated tests below exercise every
+ * {@link MultimodalTypes.AudioData} (Base64/Binary, untagged),
+ * {@link MultimodalTypes.ImageOutputs} (Base64/Binary, untagged) and
+ * {@link MultimodalTypes.VideoData} (Url/Base64/Binary, tagged on {@code type})
+ * — which are the most bug-prone part of the file. Dedicated tests below exercise every
  * variant of each union.
  *
  * <p>These are pure serialization tests: no native library is loaded — the
@@ -79,7 +78,7 @@ void embeddingResultRoundTrip() throws Exception {
     @Test
     void speechResultRoundTrip() throws Exception {
         // Same shape as Go ParseSpeechResult — exercises AudioData.Base64.
-        String json = "{\"audio\":{\"Base64\":\"aGVsbG8=\"},"
+        String json = "{\"audio\":\"aGVsbG8=\","
             + "\"warnings\":[],\"response\":{\"id\":\"resp-1\"}}";
 
         MultimodalTypes.SpeechResult result =
@@ -95,9 +94,9 @@ void speechResultRoundTrip() throws Exception {
             M.readValue(out, MultimodalTypes.SpeechResult.class);
         assertThat(rt).isEqualTo(result);
 
-        // The custom AudioData serializer re-emits the {"Base64": "..."} tag.
+        // The custom AudioData serializer re-emits the bare base64 string.
         assertThat(M.readTree(out)).as("wire form: %s", out).isEqualTo(M.readTree(
-            "{\"audio\":{\"Base64\":\"aGVsbG8=\"},\"warnings\":[],\"response\":{}}"));
+            "{\"audio\":\"aGVsbG8=\",\"warnings\":[],\"response\":{}}"));
     }
 
     // ── Image ──────────────────────────────────────────────────────────────
@@ -105,7 +104,7 @@ void speechResultRoundTrip() throws Exception {
     @Test
     void imageResultRoundTrip() throws Exception {
         // Same shape as Go ParseImageResult — exercises ImageOutputs.Base64.
-        String json = "{\"images\":{\"Base64\":[\"aW1hZ2Ux\"]},"
+        String json = "{\"images\":[\"aW1hZ2Ux\"],"
             + "\"warnings\":[],\"response\":{\"id\":\"resp-1\"}}";
 
         MultimodalTypes.ImageResult result =
@@ -123,9 +122,9 @@ void imageResultRoundTrip() throws Exception {
             M.readValue(out, MultimodalTypes.ImageResult.class);
         assertThat(rt).isEqualTo(result);
 
-        // The custom ImageOutputs serializer re-emits the {"Base64": [...]} tag.
+        // The custom ImageOutputs serializer re-emits the bare string array.
         assertThat(M.readTree(out)).as("wire form: %s", out).isEqualTo(M.readTree(
-            "{\"images\":{\"Base64\":[\"aW1hZ2Ux\"]},\"warnings\":[],\"response\":{}}"));
+            "{\"images\":[\"aW1hZ2Ux\"],\"warnings\":[],\"response\":{}}"));
     }
 
     // ── Transcription (STT) ────────────────────────────────────────────────
@@ -133,9 +132,9 @@ void imageResultRoundTrip() throws Exception {
     @Test
     void transcriptionResultRoundTrip() throws Exception {
         // Same shape as Go ParseTranscriptionResult.
-        // Timing keys are `start_second`/`end_second` — see TranscriptionSegment.ts.
+        // Timing keys are `startSecond`/`endSecond` — see TranscriptionSegment.ts.
         String json = "{\"text\":\"Hello world\","
-            + "\"segments\":[{\"text\":\"Hello\",\"start_second\":0.0,\"end_second\":1.0}],"
+            + "\"segments\":[{\"text\":\"Hello\",\"startSecond\":0.0,\"endSecond\":1.0}],"
             + "\"language\":\"en\",\"warnings\":[],\"response\":{\"id\":\"resp-1\"}}";
 
         MultimodalTypes.TranscriptionResult result =
@@ -161,11 +160,11 @@ void transcriptionResultRoundTrip() throws Exception {
         assertThat(rt.getText()).isEqualTo("Hello world");
         assertThat(rt.getLanguage()).isEqualTo("en");
 
-        // Wire form retains start_second:0.0 and the empty `warnings` list
+        // Wire form retains startSecond:0.0 and the empty `warnings` list
         // (NON_NULL suppresses only null, not zero values or empty collections).
         assertThat(M.readTree(out)).as("wire form: %s", out).isEqualTo(M.readTree(
             "{\"text\":\"Hello world\",\"segments\":"
-                + "[{\"text\":\"Hello\",\"start_second\":0.0,\"end_second\":1.0}],"
+                + "[{\"text\":\"Hello\",\"startSecond\":0.0,\"endSecond\":1.0}],"
                 + "\"language\":\"en\",\"warnings\":[],\"response\":{}}"));
     }
 
@@ -175,8 +174,8 @@ void transcriptionResultRoundTrip() throws Exception {
     void rerankingResultRoundTrip() throws Exception {
         // Same shape as Go ParseRerankingResult.
         String json = "{\"ranking\":["
-            + "{\"index\":1,\"relevance_score\":0.95},"
-            + "{\"index\":0,\"relevance_score\":0.30}"
+            + "{\"index\":1,\"relevanceScore\":0.95},"
+            + "{\"index\":0,\"relevanceScore\":0.30}"
             + "],\"warnings\":[]}";
 
         MultimodalTypes.RerankingResult result =
@@ -198,8 +197,8 @@ void rerankingResultRoundTrip() throws Exception {
         // int is never null, so it is not suppressed) and the empty `warnings`
         // list; it round-trips losslessly either way.
         assertThat(M.readTree(out)).as("wire form: %s", out).isEqualTo(M.readTree(
-            "{\"ranking\":[{\"index\":1,\"relevance_score\":0.95},"
-            + "{\"index\":0,\"relevance_score\":0.3}],\"warnings\":[]}"));
+            "{\"ranking\":[{\"index\":1,\"relevanceScore\":0.95},"
+            + "{\"index\":0,\"relevanceScore\":0.3}],\"warnings\":[]}"));
     }
 
     // ── Video ──────────────────────────────────────────────────────────────
@@ -207,8 +206,8 @@ void rerankingResultRoundTrip() throws Exception {
     @Test
     void videoResultRoundTrip() throws Exception {
         // Same shape as Go ParseVideoResult — exercises VideoData.Url.
-        String json = "{\"videos\":[{\"Url\":"
-            + "{\"url\":\"https://example.com/video.mp4\",\"media_type\":\"video/mp4\"}}],"
+        String json = "{\"videos\":[{\"type\":\"url\","
+            + "\"url\":\"https://example.com/video.mp4\",\"mediaType\":\"video/mp4\"}],"
             + "\"warnings\":[],\"response\":{\"id\":\"resp-1\"}}";
 
         MultimodalTypes.VideoResult result =
@@ -227,10 +226,10 @@ void videoResultRoundTrip() throws Exception {
             M.readValue(out, MultimodalTypes.VideoResult.class);
         assertThat(rt).isEqualTo(result);
 
-        // The custom VideoData serializer re-emits the {"Url": {...}} tag.
+        // The custom VideoData serializer re-emits the {"type":"url",...} tag.
         assertThat(M.readTree(out)).as("wire form: %s", out).isEqualTo(M.readTree(
-            "{\"videos\":[{\"Url\":{\"url\":\"https://example.com/video.mp4\","
-            + "\"media_type\":\"video/mp4\"}}],\"warnings\":[],\"response\":{}}"));
+            "{\"videos\":[{\"type\":\"url\",\"url\":\"https://example.com/video.mp4\","
+            + "\"mediaType\":\"video/mp4\"}],\"warnings\":[],\"response\":{}}"));
     }
 
     @Test
@@ -244,8 +243,8 @@ void videoPollOptionsRoundTrip() throws Exception {
             .build();
 
         String json = M.writeValueAsString(options);
-        assertThat(M.readTree(json).path("poll").path("interval_ms").asLong()).isEqualTo(1_000L);
-        assertThat(M.readTree(json).path("poll").path("timeout_ms").asLong()).isEqualTo(120_000L);
+        assertThat(M.readTree(json).path("poll").path("intervalMs").asLong()).isEqualTo(1_000L);
+        assertThat(M.readTree(json).path("poll").path("timeoutMs").asLong()).isEqualTo(120_000L);
 
         MultimodalTypes.VideoCallOptions decoded =
             M.readValue(json, MultimodalTypes.VideoCallOptions.class);
@@ -301,13 +300,13 @@ void searchResultRoundTrip() throws Exception {
     @Test
     void uploadFileResultRoundTrip() throws Exception {
         // Same shape as Go ParseUploadFileResult.
-        String json = "{\"provider_reference\":{\"openai\":\"file-abc123\"},"
-            + "\"media_type\":\"application/pdf\",\"filename\":\"doc.pdf\",\"warnings\":[]}";
+        String json = "{\"providerReference\":{\"openai\":\"file-abc123\"},"
+            + "\"mediaType\":\"application/pdf\",\"filename\":\"doc.pdf\",\"warnings\":[]}";
 
         MultimodalTypes.UploadFileResult result =
             M.readValue(json, MultimodalTypes.UploadFileResult.class);
 
-        // Decode (mirrors Go: provider_reference.openai, media_type).
+        // Decode (mirrors Go: providerReference.openai, mediaType).
         assertThat(result.getProviderReference()).containsEntry("openai", "file-abc123");
         assertThat(result.getMediaType()).isEqualTo("application/pdf");
         assertThat(result.getFilename()).isEqualTo("doc.pdf");
@@ -318,8 +317,8 @@ void uploadFileResultRoundTrip() throws Exception {
         assertThat(rt).isEqualTo(result);
 
         assertThat(M.readTree(out)).as("wire form: %s", out).isEqualTo(M.readTree(
-            "{\"provider_reference\":{\"openai\":\"file-abc123\"},"
-            + "\"media_type\":\"application/pdf\",\"filename\":\"doc.pdf\",\"warnings\":[]}"));
+            "{\"providerReference\":{\"openai\":\"file-abc123\"},"
+            + "\"mediaType\":\"application/pdf\",\"filename\":\"doc.pdf\",\"warnings\":[]}"));
     }
 
     // ── Custom serializer focus: the three sealed unions ───────────────────
@@ -331,7 +330,7 @@ void audioDataSerializerAllVariants() throws Exception {
         // Base64 variant.
         MultimodalTypes.AudioData base64 = new MultimodalTypes.AudioData.Base64("aGVsbG8=");
         String base64Json = M.writeValueAsString(base64);
-        assertThat(base64Json).isEqualTo("{\"Base64\":\"aGVsbG8=\"}");
+        assertThat(base64Json).isEqualTo("\"aGVsbG8=\"");
         MultimodalTypes.AudioData base64Back =
             M.readValue(base64Json, MultimodalTypes.AudioData.class);
         assertThat(base64Back).isInstanceOf(MultimodalTypes.AudioData.Base64.class);
@@ -342,14 +341,14 @@ void audioDataSerializerAllVariants() throws Exception {
         MultimodalTypes.AudioData binary =
             new MultimodalTypes.AudioData.Binary(Arrays.asList(72, 101, 108, 108, 111));
         String binaryJson = M.writeValueAsString(binary);
-        assertThat(binaryJson).isEqualTo("{\"Binary\":[72,101,108,108,111]}");
+        assertThat(binaryJson).isEqualTo("[72,101,108,108,111]");
         MultimodalTypes.AudioData binaryBack =
             M.readValue(binaryJson, MultimodalTypes.AudioData.class);
         assertThat(binaryBack).isInstanceOf(MultimodalTypes.AudioData.Binary.class);
         assertThat(((MultimodalTypes.AudioData.Binary) binaryBack).getValue())
             .isEqualTo(Arrays.asList(72, 101, 108, 108, 111));
 
-        // Unknown tag → the deserializer rejects it.
+        // Neither a string nor an array → the deserializer rejects it.
         assertThatThrownBy(() -> M.readValue("{\"Quux\":\"x\"}", MultimodalTypes.AudioData.class))
             .isInstanceOf(IOException.class);
     }
@@ -360,7 +359,7 @@ void imageOutputsSerializerAllVariants() throws Exception {
         MultimodalTypes.ImageOutputs base64 =
             new MultimodalTypes.ImageOutputs.Base64(Arrays.asList("aW1hZ2Ux", "aW1hZ2Uy"));
         String base64Json = M.writeValueAsString(base64);
-        assertThat(base64Json).isEqualTo("{\"Base64\":[\"aW1hZ2Ux\",\"aW1hZ2Uy\"]}");
+        assertThat(base64Json).isEqualTo("[\"aW1hZ2Ux\",\"aW1hZ2Uy\"]");
         MultimodalTypes.ImageOutputs base64Back =
             M.readValue(base64Json, MultimodalTypes.ImageOutputs.class);
         assertThat(base64Back).isInstanceOf(MultimodalTypes.ImageOutputs.Base64.class);
@@ -371,7 +370,7 @@ void imageOutputsSerializerAllVariants() throws Exception {
         MultimodalTypes.ImageOutputs binary = new MultimodalTypes.ImageOutputs.Binary(
             Arrays.asList(Arrays.asList(1, 2, 3), Arrays.asList(4, 5, 6)));
         String binaryJson = M.writeValueAsString(binary);
-        assertThat(binaryJson).isEqualTo("{\"Binary\":[[1,2,3],[4,5,6]]}");
+        assertThat(binaryJson).isEqualTo("[[1,2,3],[4,5,6]]");
         MultimodalTypes.ImageOutputs binaryBack =
             M.readValue(binaryJson, MultimodalTypes.ImageOutputs.class);
         assertThat(binaryBack).isInstanceOf(MultimodalTypes.ImageOutputs.Binary.class);
@@ -389,7 +388,7 @@ void videoDataSerializerAllVariants() throws Exception {
             MultimodalTypes.VideoUrlData.of("https://example.com/v.mp4", "video/mp4"));
         String urlJson = M.writeValueAsString(url);
         assertThat(urlJson).isEqualTo(
-            "{\"Url\":{\"url\":\"https://example.com/v.mp4\",\"media_type\":\"video/mp4\"}}");
+            "{\"type\":\"url\",\"url\":\"https://example.com/v.mp4\",\"mediaType\":\"video/mp4\"}");
         MultimodalTypes.VideoData urlBack = M.readValue(urlJson, MultimodalTypes.VideoData.class);
         assertThat(urlBack).isInstanceOf(MultimodalTypes.VideoData.Url.class);
         assertThat(((MultimodalTypes.VideoData.Url) urlBack).getValue().getUrl())
@@ -400,7 +399,7 @@ void videoDataSerializerAllVariants() throws Exception {
             MultimodalTypes.VideoBase64Data.of("AAAA", "video/mp4"));
         String base64Json = M.writeValueAsString(base64);
         assertThat(base64Json).isEqualTo(
-            "{\"Base64\":{\"data\":\"AAAA\",\"media_type\":\"video/mp4\"}}");
+            "{\"type\":\"base64\",\"data\":\"AAAA\",\"mediaType\":\"video/mp4\"}");
         MultimodalTypes.VideoData base64Back =
             M.readValue(base64Json, MultimodalTypes.VideoData.class);
         assertThat(base64Back).isInstanceOf(MultimodalTypes.VideoData.Base64.class);
@@ -412,14 +411,14 @@ void videoDataSerializerAllVariants() throws Exception {
             MultimodalTypes.VideoBinaryData.of(Arrays.asList(1, 2, 3), "video/mp4"));
         String binaryJson = M.writeValueAsString(binary);
         assertThat(binaryJson).isEqualTo(
-            "{\"Binary\":{\"data\":[1,2,3],\"media_type\":\"video/mp4\"}}");
+            "{\"type\":\"binary\",\"data\":[1,2,3],\"mediaType\":\"video/mp4\"}");
         MultimodalTypes.VideoData binaryBack =
             M.readValue(binaryJson, MultimodalTypes.VideoData.class);
         assertThat(binaryBack).isInstanceOf(MultimodalTypes.VideoData.Binary.class);
         assertThat(((MultimodalTypes.VideoData.Binary) binaryBack).getValue().getData())
             .isEqualTo(Arrays.asList(1, 2, 3));
 
-        assertThatThrownBy(() -> M.readValue("{\"Quux\":{}}", MultimodalTypes.VideoData.class))
+        assertThatThrownBy(() -> M.readValue("{\"type\":\"quux\"}", MultimodalTypes.VideoData.class))
             .isInstanceOf(IOException.class);
     }
 }
diff --git a/bindings/java/src/test/java/ai/arcships/aimux/StructuredE2ETest.java b/bindings/java/src/test/java/ai/arcships/aimux/StructuredE2ETest.java
index ddd322506..a8c8a40ac 100644
--- a/bindings/java/src/test/java/ai/arcships/aimux/StructuredE2ETest.java
+++ b/bindings/java/src/test/java/ai/arcships/aimux/StructuredE2ETest.java
@@ -20,14 +20,14 @@
  * the Kotlin binding's StructuredE2ETest.kt).
  *
  * <p>These spin up a local mock HTTP server speaking the OpenAI
- * chat-completions wire format, point an OpenAI {@code Model} at it via the
- * base-url constructor, and assert that the binding correctly:
+ * chat-completions wire format, point an OpenAI-compatible {@code Model} at it
+ * through the provider registry, and assert that the binding correctly:
  * <ol>
- *   <li>parses tool_calls out of a provider response,</li>
+ *   <li>parses toolCalls out of a provider response,</li>
  *   <li>forwards multi-role messages,</li>
- *   <li>forwards {@code tool_choice},</li>
+ *   <li>forwards {@code toolChoice},</li>
  *   <li>parses tool-call stream parts from an SSE stream,</li>
- *   <li>runs a full tool-call round-trip (tool_call + tool_result + final answer).</li>
+ *   <li>runs a full tool-call round-trip (tool-call + tool-result + final answer).</li>
  * </ol>
  *
  * No real network access is performed — every request hits 127.0.0.1.
@@ -160,12 +160,12 @@ private static String weatherToolsOpts(String includeToolChoice) {
             new JSONObject()
                 .put("type", "function")
                 .put("name", "get_weather")
-                .put("input_schema", new JSONObject()
+                .put("inputSchema", new JSONObject()
                     .put("type", "object")
                     .put("properties", new JSONObject()
                         .put("location", new JSONObject().put("type", "string"))))));
         if (includeToolChoice != null) {
-            opts.put("tool_choice", includeToolChoice);
+            opts.put("toolChoice", includeToolChoice);
         }
         return opts.toString();
     }
@@ -180,23 +180,23 @@ private static String weatherToolsOpts() {
     void generateTextParsesToolCalls() {
         server.setResponseBody(toolCallOpenAiResponse());
 
-        try (Model model = Model.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+        try (Model model = Model.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             String resultJson = model.generateText("\"What is the weather in Tokyo?\"", weatherToolsOpts());
             JSONObject result = new JSONObject(resultJson);
 
-            // tool_calls[0].tool_name == "get_weather"
-            JSONArray toolCalls = result.getJSONArray("tool_calls");
+            // toolCalls[0].toolName == "get_weather"
+            JSONArray toolCalls = result.getJSONArray("toolCalls");
             assertThat(toolCalls.length()).isGreaterThan(0);
             JSONObject firstCall = toolCalls.getJSONObject(0);
-            assertThat(firstCall.getString("tool_name")).isEqualTo("get_weather");
-            assertThat(firstCall.getString("tool_call_id")).isEqualTo("call_abc");
+            assertThat(firstCall.getString("toolName")).isEqualTo("get_weather");
+            assertThat(firstCall.getString("toolCallId")).isEqualTo("call_abc");
             assertThat(firstCall.getJSONObject("input").getString("location")).isEqualTo("Tokyo");
 
-            // raw.content contains a "ToolCall" variant.
+            // raw.content contains a "tool-call" variant.
             JSONArray rawContent = result.getJSONObject("raw").getJSONArray("content");
             boolean hasToolCallVariant = false;
             for (int i = 0; i < rawContent.length(); i++) {
-                if (rawContent.getJSONObject(i).has("ToolCall")) {
+                if ("tool-call".equals(rawContent.getJSONObject(i).optString("type"))) {
                     hasToolCallVariant = true;
                     break;
                 }
@@ -215,7 +215,7 @@ void multiRoleMessagesReachProvider() {
             .put(new JSONObject().put("role", "system").put("content", "You are a helpful assistant."))
             .put(new JSONObject().put("role", "user").put("content", "What is Rust?"));
 
-        try (Model model = Model.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+        try (Model model = Model.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             String resultJson = model.generateText(prompt.toString());
             JSONObject result = new JSONObject(resultJson);
             assertThat(result.getString("text"))
@@ -235,7 +235,7 @@ void multiRoleMessagesReachProvider() {
     void toolChoiceReachesProvider() {
         server.setResponseBody(toolCallOpenAiResponse());
 
-        try (Model model = Model.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+        try (Model model = Model.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             model.generateText("\"What is the weather in Tokyo?\"", weatherToolsOpts("required"));
 
             JSONObject reqBody = new JSONObject(server.lastRequestBody());
@@ -251,7 +251,7 @@ void streamTextParsesToolCallStreamParts() {
         server.setContentType("text/event-stream");
         server.setResponseBody(toolCallSseResponse());
 
-        try (Model model = Model.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+        try (Model model = Model.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             // Run the blocking FFI stream call on a separate thread so the JNA
             // callbacks (which attach to the calling thread) don't deadlock with
             // tokio's block_on. The main thread collects parts via a queue.
@@ -286,24 +286,24 @@ void streamTextParsesToolCallStreamParts() {
             }
 
             assertThat(collected).isNotEmpty();
-            // Stream must surface tool-call activity: either a complete ToolCall
-            // part or ToolInputDelta deltas (or both).
+            // Stream must surface tool-call activity: either a complete tool-call
+            // part or tool-input-delta deltas (or both).
             boolean hasToolPart = false;
             for (String part : collected) {
-                if (part.contains("\"ToolCall\"") || part.contains("\"ToolInputDelta\"")) {
+                if (part.contains("\"type\":\"tool-call\"") || part.contains("\"type\":\"tool-input-delta\"")) {
                     hasToolPart = true;
                     break;
                 }
             }
             assertThat(hasToolPart)
-                .as("expected a ToolCall or ToolInputDelta stream part; got: %s", collected)
+                .as("expected a tool-call or tool-input-delta stream part; got: %s", collected)
                 .isTrue();
 
             // The finish part should also be present, carrying the tool-calls
             // finish reason.
             boolean hasFinish = false;
             for (String part : collected) {
-                if (part.contains("\"Finish\"")) {
+                if (part.contains("\"type\":\"finish\"")) {
                     hasFinish = true;
                     break;
                 }
@@ -320,43 +320,45 @@ void toolCallFullRoundTrip() {
         // second (FIFO queue).
         server.setResponses(toolCallOpenAiResponse(), finalTextResponse());
 
-        try (Model model = Model.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+        try (Model model = Model.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             // 1st call: provider asks to call get_weather.
             String firstResultJson =
                 model.generateText("\"What's the weather in Tokyo?\"", weatherToolsOpts());
             JSONObject firstResult = new JSONObject(firstResultJson);
-            JSONArray toolCalls = firstResult.getJSONArray("tool_calls");
+            JSONArray toolCalls = firstResult.getJSONArray("toolCalls");
             assertThat(toolCalls.length()).isGreaterThan(0);
             JSONObject firstCall = toolCalls.getJSONObject(0);
-            assertThat(firstCall.getString("tool_name")).isEqualTo("get_weather");
-            assertThat(firstCall.getString("tool_call_id")).isEqualTo("call_abc");
+            assertThat(firstCall.getString("toolName")).isEqualTo("get_weather");
+            assertThat(firstCall.getString("toolCallId")).isEqualTo("call_abc");
             assertThat(firstCall.getJSONObject("input").getString("location"))
                 .isEqualTo("Tokyo");
 
             // 2nd call: echo the full conversation back, including the
-            // assistant's tool_call and the ToolResult we synthesised.
-            // Input uses the engine's ContentPart variants (tool_call /
-            // tool_result); the engine converts these to the OpenAI wire
-            // format on the outbound request. NOTE: the tool_result part uses
-            // the `result` field (never `output`) — the engine's
-            // ContentPart::ToolResult requires it (aimux-core/content.rs).
+            // assistant's tool-call and the tool-result we synthesised.
+            // Input uses the AI SDK ContentPart variants (tool-call /
+            // tool-result); the engine converts these to the OpenAI wire
+            // format on the outbound request. The tool-result part carries
+            // `output` (a ToolResultOutput union), never `result`.
             JSONArray messages = new JSONArray()
                 .put(new JSONObject().put("role", "user").put("content", "What's the weather in Tokyo?"))
                 .put(new JSONObject().put("role", "assistant")
                     .put("content", new JSONArray().put(
                         new JSONObject()
-                            .put("type", "tool_call")
-                            .put("tool_call_id", "call_abc")
-                            .put("tool_name", "get_weather")
+                            .put("type", "tool-call")
+                            .put("toolCallId", "call_abc")
+                            .put("toolName", "get_weather")
                             .put("input", new JSONObject().put("location", "Tokyo")))))
                 .put(new JSONObject().put("role", "tool")
                     .put("content", new JSONArray().put(
                         new JSONObject()
-                            .put("type", "tool_result")
-                            .put("tool_call_id", "call_abc")
-                            .put("result", new JSONObject()
-                                .put("temperature", 22)
-                                .put("condition", "sunny")))));
+                            .put("type", "tool-result")
+                            .put("toolCallId", "call_abc")
+                            .put("toolName", "get_weather")
+                            .put("output", new JSONObject()
+                                .put("type", "json")
+                                .put("value", new JSONObject()
+                                    .put("temperature", 22)
+                                    .put("condition", "sunny"))))));
 
             String secondResultJson = model.generateText(messages.toString(), weatherToolsOpts());
             JSONObject secondResult = new JSONObject(secondResultJson);
diff --git a/bindings/java/src/test/java/ai/arcships/aimux/TypedModelTest.java b/bindings/java/src/test/java/ai/arcships/aimux/TypedModelTest.java
index 6ffbb0bc4..5decbd6f3 100644
--- a/bindings/java/src/test/java/ai/arcships/aimux/TypedModelTest.java
+++ b/bindings/java/src/test/java/ai/arcships/aimux/TypedModelTest.java
@@ -167,7 +167,7 @@ void topLevelToolCallProviderMetadataRoundTrips() throws Exception {
         String json = Types.AimuxJson.MAPPER.writeValueAsString(original);
         Types.ToolCall decoded = Types.AimuxJson.MAPPER.readValue(json, Types.ToolCall.class);
 
-        assertThat(json).contains("\"provider_metadata\"");
+        assertThat(json).contains("\"providerMetadata\"");
         assertThat(decoded.getProviderMetadata()).isEqualTo(metadata);
         assertThat(decoded).isEqualTo(original);
     }
@@ -177,7 +177,7 @@ void generateTextReturnsTypedGenerateTextResultWithTextAndRawContent() {
         server.setResponseBody(plainOpenAiResponse());
 
         try (TypedModel model =
-                 TypedModel.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+                 TypedModel.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             Types.GenerateTextResult result = model.generateText("What is Rust?");
 
             // No manual parsing: the wrapper returned a typed object.
@@ -204,7 +204,7 @@ void generateTextParsesToolCallsIntoTypedObjects() {
             .build();
 
         try (TypedModel model =
-                 TypedModel.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+                 TypedModel.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             Types.GenerateTextResult result =
                 model.generateText("What is the weather in Tokyo?", options);
 
@@ -216,8 +216,8 @@ void generateTextParsesToolCallsIntoTypedObjects() {
             // input is a JsonNode (tool arguments); inspect it directly.
             assertThat(call.getInput().get("location").asText()).isEqualTo("Tokyo");
 
-            // .raw.content carries the ToolCall variant tag.
-            assertThat(result.getRaw().hasContentVariant("ToolCall")).isTrue();
+            // .raw.content carries the tool-call variant tag.
+            assertThat(result.getRaw().hasContentVariant("tool-call")).isTrue();
         }
     }
 
@@ -231,11 +231,11 @@ void typedToolsAndToolChoiceReachProvider() {
             .build();
 
         try (TypedModel model =
-                 TypedModel.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+                 TypedModel.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             model.generateText("What is the weather in Tokyo?", options);
 
             // The serialized options crossed the JSON boundary as the engine's
-            // snake_case shape with only the fields the caller set.
+            // camelCase shape with only the fields the caller set.
             JSONObject reqBody = new JSONObject(server.lastRequestBody());
             assertThat(reqBody.getString("tool_choice")).isEqualTo("required");
             assertThat(reqBody.has("tools")).isTrue();
@@ -260,7 +260,7 @@ void multiRoleModelMessageListReachesProvider() {
             Types.ModelMessage.text(Types.Role.USER, "What is Rust?"));
 
         try (TypedModel model =
-                 TypedModel.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+                 TypedModel.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             Types.GenerateTextResult result = model.generateText(messages);
 
             assertThat(result.getText()).isEqualTo("Rust is a systems programming language.");
@@ -289,7 +289,7 @@ void streamTextStreamDecodesTypedParts() {
             .build();
 
         try (TypedModel model =
-                 TypedModel.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+                 TypedModel.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             List<Types.StreamPart> parts = assertTimeout(Duration.ofSeconds(30),
                 () -> model.streamTextStream("What is the weather in Tokyo?", options)
                     .collect(Collectors.toList()));
@@ -326,7 +326,7 @@ void streamTextCallbackDeliversTypedParts() {
         java.util.concurrent.atomic.AtomicBoolean done = new java.util.concurrent.atomic.AtomicBoolean(false);
 
         try (TypedModel model =
-                 TypedModel.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+                 TypedModel.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             assertTimeout(Duration.ofSeconds(30), () ->
                 model.streamText("What is the weather in Tokyo?", options,
                     collected::add,           // onPart — typed StreamPart
@@ -346,7 +346,7 @@ void streamTextCallbackDeliversTypedParts() {
     void ofModelDoesNotOwnHandle() {
         server.setResponseBody(plainOpenAiResponse());
 
-        Model raw = Model.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl());
+        Model raw = Model.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig());
         // of() wraps without owning — closing the TypedModel must NOT close the raw Model.
         TypedModel typed = TypedModel.of(raw);
         typed.close();
@@ -370,7 +370,7 @@ void streamTextStreamFromMessageList() {
             .build();
 
         try (TypedModel model =
-                 TypedModel.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+                 TypedModel.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             List<Types.StreamPart> parts = assertTimeout(Duration.ofSeconds(30),
                 () -> model.streamTextStream(messages, options).collect(Collectors.toList()));
 
@@ -386,7 +386,7 @@ void generateTextInvalidPromptThrowsAimuxException() {
         // The typed layer JSON-encodes the prompt, so any string is valid wire
         // JSON; the failure is the mock provider's response → AimuxException.
         try (TypedModel model =
-                 TypedModel.openaiWithBase("sk-test-fake-key", "gpt-4o", server.baseUrl())) {
+                 TypedModel.provider("deepseek", "sk-test-fake-key", "gpt-4o", server.chatConfig())) {
             assertThatThrownBy(() -> model.generateText("not-valid-json"))
                 .isInstanceOf(AimuxException.class);
         }
diff --git a/bindings/kotlin/build.gradle.kts b/bindings/kotlin/build.gradle.kts
index 003e4e0a6..e2fa32782 100644
--- a/bindings/kotlin/build.gradle.kts
+++ b/bindings/kotlin/build.gradle.kts
@@ -33,7 +33,12 @@ kotlin {
 
 tasks.test {
     useJUnitPlatform()
-    systemProperty("jna.library.path", rootProject.projectDir.resolve("../../target/release").canonicalPath)
+    // AIMUX_FFI_LIB_DIR points at the directory holding libaimux_ffi (e.g. a
+    // custom CARGO_TARGET_DIR/debug); the default is the in-tree release build.
+    systemProperty(
+        "jna.library.path",
+        System.getenv("AIMUX_FFI_LIB_DIR") ?: rootProject.projectDir.resolve("../../target/release").canonicalPath,
+    )
     // Gradle prints only "there were failing tests" by default; the names live
     // in an HTML report that CI logs never surface. Without this, a red build
     // tells you nothing about which test broke.
diff --git a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Errors.kt b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Errors.kt
index 5de65650c..ec4012c92 100644
--- a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Errors.kt
+++ b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Errors.kt
@@ -522,7 +522,7 @@ class ToolCallRepairError(
     cause: Throwable? = null,
     retryable: Boolean = false,
     /**
-     * The original lookup/parse/validation error as externally-tagged wire
+     * The original lookup/parse/validation error as name-tagged wire
      * JSON (the same encoding as [ToolCall.error]); null when synthesized
      * locally.
      */
diff --git a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Model.kt b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Model.kt
index 5c9561869..5f5fd6acb 100644
--- a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Model.kt
+++ b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Model.kt
@@ -221,9 +221,6 @@ internal interface AimuxFFI : Library {
     fun aimux_recording_try_flush(): Pointer?
     fun aimux_mock_replay_new(recordingsJsonl: String, outHandle: LongByReference): Pointer?
 
-    // [AiMuxError] Register external providers from JSON config (RFC-0020).
-    fun aimux_register_providers(configJson: String): Pointer?
-
     // [AiMuxError] Set the global proxy configuration (M6, RFC-0016).
     fun aimux_init_proxy(configJson: String): Pointer?
 
@@ -534,7 +531,7 @@ class Model internal constructor(handle: Long) : Closeable {
      * Generate a structured JSON object (M12, RFC-0016).
      *
      * Same signature as [generateText]; returns a JSON-serialized
-     * `GenerateObjectResult`. Pass `response_format: { "Json": { ... } }` via
+     * `GenerateObjectResult`. Pass `responseFormat: { "type": "json", "schema": { ... } }` via
      * [optsJson] for schema control; aimux-core applies JSON repair before
      * parsing.
      *
@@ -581,7 +578,7 @@ class Model internal constructor(handle: Long) : Closeable {
      * Stream text from the model.
      *
      * Blocks the calling thread until the stream completes. Recoverable
-     * frame errors (a malformed SSE frame) arrive as `StreamPart::Error`
+     * frame errors (a malformed SSE frame) arrive as `error` stream parts
      * data parts and the stream continues; only transport/Core failures
      * throw [AimuxException] (on_done is not invoked on failure), after any
      * parts already delivered to [onPart].
@@ -893,7 +890,9 @@ class Model internal constructor(handle: Long) : Closeable {
          * @param apiKey     API key, or null to read the provider's env var from
          *                   the registry entry.
          * @param modelId    Model id.
-         * @param configJson Optional JSON object of ProviderOptions; null for defaults.
+         * @param configJson Optional JSON object of settings (`baseUrl`, `headers`,
+         *                   `organization`, `project`, `params`); any other key is
+         *                   rejected as an invalid argument. Null for defaults.
          */
         fun provider(name: String, apiKey: String? = null, modelId: String, configJson: String? = null): Model {
             requireJson("configJson", configJson)
@@ -943,7 +942,7 @@ class Model internal constructor(handle: Long) : Closeable {
          * @param models     child models (must be non-empty).
          * @param configJson optional config: `{"router": "rule"|"weighted",
          *                   "weights": [...], "fallback": "on_error"|"none",
-         *                   "provider_name", "model_id"}` — all optional.
+         *                   "providerName", "modelId"}` — all optional.
          * @throws IllegalArgumentException if `configJson` is malformed JSON.
          * @throws AimuxException on AiMuxError.
          */
@@ -982,22 +981,6 @@ class Model internal constructor(handle: Long) : Closeable {
             })
         }
 
-        /**
-         * Register external OpenAI-compatible providers from a JSON config string
-         * (RFC-0020).
-         *
-         * `configJson` is `{ "providers": [ { "name", "base_url", ... } ] }`.
-         * Entries override same-named built-ins or add new ones. Like
-         * `initRecording`, this mutates process-global registry state.
-         *
-         * @throws IllegalArgumentException if `configJson` is blank or malformed JSON.
-         * @throws AimuxException on AiMuxError (bad schema / unknown protocol).
-         */
-        fun registerProviders(configJson: String) {
-            requireJsonRequired("configJson", configJson)
-            FFI.lib.aimux_register_providers(configJson)?.let { throw expectAimuxError(it, "registerProviders") }
-        }
-
         /**
          * Set the global proxy configuration (M6, RFC-0016). Must be called
          * before the first `generateText` / `streamText` call; a no-op if the
diff --git a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Multimodal.kt b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Multimodal.kt
index eeaa2fae1..ef60b6ba9 100644
--- a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Multimodal.kt
+++ b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Multimodal.kt
@@ -234,8 +234,8 @@ class TranscriptionModel private constructor(handle: Long) : Closeable {
      * Start a streaming transcription session (RFC-0028) on this model.
      * Requires a model that supports streaming (realtime models).
      *
-     * @param optsJson    optional session options JSON (`input_audio_format` /
-     *                    `provider_options` / `headers` / `include_raw_chunks`).
+     * @param optsJson    optional session options JSON (`inputAudioFormat` /
+     *                    `providerOptions` / `headers` / `includeRawChunks`).
      * @param abortHandle abort handle (from `Model.abortSignalNew()`), or 0.
      * @return a new live session.
      */
diff --git a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/MultimodalTypes.kt b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/MultimodalTypes.kt
index 3f547296f..8fb8fdedf 100644
--- a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/MultimodalTypes.kt
+++ b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/MultimodalTypes.kt
@@ -1,31 +1,34 @@
 /**
  * aimux — typed multimodal data structures mirroring the aimux-core wire format.
  *
- * Same shapes as the ts-rs generated `.ts` types in `bindings/node/src/types/` and
- * the Go structs in `bindings/go/multimodal_types.go`. Field names use camelCase
- * in Kotlin and map to the wire format's snake_case via [kotlinx.serialization.SerialName].
+ * Same shapes as the ts-rs generated `.ts` types in `bindings/node/src/types/`:
+ * camelCase field names (Kotlin property names equal the wire names), unions
+ * tagged on `type`, optional fields absent rather than null.
  *
- * These types are intentionally lenient on decode (unknown keys ignored, every
- * field has a default) so future engine additions don't break existing clients.
- * The serialization config lives in [AimuxJson]. Decode a JSON string returned
- * by a [Multimodal][aimux] model with, for example:
+ * Decode is lenient (unknown keys ignored, every result field has a default)
+ * so future engine additions don't break existing clients. The serialization
+ * config lives in [AimuxJson]. Decode a JSON string returned by a
+ * [Multimodal][aimux] model with, for example:
  *
  * ```kotlin
  * val result: EmbeddingResult = AimuxJson.decodeFromString(EmbeddingResult.serializer(), jsonStr)
  * ```
  *
- * The `Base64`/`Binary`/`Url` unions ([AudioData], [ImageOutputs], [VideoData])
- * are serde-style externally-tagged enums on the wire
- * (`{"Base64": ...}` / `{"Binary": ...}` / `{"Url": ...}`), so each has a custom
- * serializer (see [StreamPartSerializer] in `Types.kt` for the same pattern).
+ * [AudioData] and [ImageOutputs] are untagged unions on the wire (a string is
+ * base64, an array is binary), so each has a small custom serializer; the
+ * `Array<number> | string` input fields reuse [FileBytes].
  */
 
 package ai.arcships.aimux
 
+import kotlinx.serialization.EncodeDefault
+import kotlinx.serialization.ExperimentalSerializationApi
 import kotlinx.serialization.KSerializer
 import kotlinx.serialization.SerialName
 import kotlinx.serialization.Serializable
 import kotlinx.serialization.SerializationException
+import kotlinx.serialization.descriptors.PrimitiveKind
+import kotlinx.serialization.descriptors.PrimitiveSerialDescriptor
 import kotlinx.serialization.descriptors.SerialDescriptor
 import kotlinx.serialization.descriptors.buildClassSerialDescriptor
 import kotlinx.serialization.encoding.Decoder
@@ -34,18 +37,45 @@ import kotlinx.serialization.json.JsonArray
 import kotlinx.serialization.json.JsonDecoder
 import kotlinx.serialization.json.JsonElement
 import kotlinx.serialization.json.JsonEncoder
-import kotlinx.serialization.json.JsonObject
 import kotlinx.serialization.json.JsonPrimitive
 import kotlinx.serialization.json.jsonArray
-import kotlinx.serialization.json.jsonObject
 import kotlinx.serialization.json.jsonPrimitive
 
 // ─────────────────────────────────────────────────────────────────────────────
 // Shared types.
 // ─────────────────────────────────────────────────────────────────────────────
 
-/** A provider warning (arbitrary JSON). Mirrors Go's `Warning = json.RawMessage`. */
-typealias Warning = JsonElement
+/** A pixel size. Mirrors `Size.ts`: serialized as the string `"WxH"` (e.g. `"1024x1024"`). */
+@Serializable(with = SizeSerializer::class)
+data class Size(val width: Int, val height: Int)
+
+/** An aspect ratio. Mirrors `AspectRatio.ts`: serialized as the string `"W:H"` (e.g. `"16:9"`). */
+@Serializable(with = AspectRatioSerializer::class)
+data class AspectRatio(val width: Int, val height: Int)
+
+private fun parsePair(kind: String, sep: Char, text: String): Pair<Int, Int> {
+    val parts = text.split(sep)
+    val w = parts.getOrNull(0)?.toIntOrNull()
+    val h = parts.getOrNull(1)?.toIntOrNull()
+    if (parts.size != 2 || w == null || h == null) {
+        throw SerializationException("$kind must look like \"W${sep}H\", got: \"$text\"")
+    }
+    return w to h
+}
+
+object SizeSerializer : KSerializer<Size> {
+    override val descriptor: SerialDescriptor = PrimitiveSerialDescriptor("aimux.Size", PrimitiveKind.STRING)
+    override fun deserialize(decoder: Decoder): Size =
+        parsePair("Size", 'x', decoder.decodeString()).let { Size(it.first, it.second) }
+    override fun serialize(encoder: Encoder, value: Size) = encoder.encodeString("${value.width}x${value.height}")
+}
+
+object AspectRatioSerializer : KSerializer<AspectRatio> {
+    override val descriptor: SerialDescriptor = PrimitiveSerialDescriptor("aimux.AspectRatio", PrimitiveKind.STRING)
+    override fun deserialize(decoder: Decoder): AspectRatio =
+        parsePair("AspectRatio", ':', decoder.decodeString()).let { AspectRatio(it.first, it.second) }
+    override fun serialize(encoder: Encoder, value: AspectRatio) = encoder.encodeString("${value.width}:${value.height}")
+}
 
 // ─────────────────────────────────────────────────────────────────────────────
 // Embedding
@@ -69,7 +99,7 @@ data class EmbeddingResponse(
 data class EmbeddingResult(
     val embeddings: List<List<Float>> = emptyList(),
     val usage: EmbeddingUsage? = null,
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    val providerMetadata: ProviderMetadata? = null,
     val response: EmbeddingResponse? = null,
     val warnings: List<Warning> = emptyList(),
 )
@@ -78,9 +108,9 @@ data class EmbeddingResult(
 @Serializable
 data class EmbeddingCallOptions(
     val values: List<String> = emptyList(),
-    @SerialName("max_retries") val maxRetries: Long? = null,
+    val maxRetries: Long? = null,
     val timeout: TimeoutConfiguration? = null,
-    @SerialName("provider_options") val providerOptions: JsonElement? = null,
+    val providerOptions: ProviderOptions? = null,
     val headers: Map<String, String>? = null,
 )
 
@@ -91,7 +121,7 @@ data class EmbeddingCallOptions(
 /**
  * Generated audio: a base64 string or raw binary bytes.
  *
- * Wire format (serde externally-tagged): `{"Base64": "..."}` | `{"Binary": [n,...]}`.
+ * Wire format (untagged, `AudioData.ts`): `"..."` (base64) | `[n,...]` (binary).
  */
 @Serializable(with = AudioDataSerializer::class)
 sealed interface AudioData {
@@ -112,7 +142,7 @@ data class SpeechRequest(
 @Serializable
 data class SpeechResponse(
     val timestamp: String? = null,
-    @SerialName("model_id") val modelId: String? = null,
+    val modelId: String? = null,
     val headers: Map<String, String>? = null,
     val body: JsonElement? = null,
 )
@@ -124,7 +154,7 @@ data class SpeechResult(
     val warnings: List<Warning> = emptyList(),
     val request: SpeechRequest? = null,
     val response: SpeechResponse = SpeechResponse(),
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    val providerMetadata: ProviderMetadata? = null,
 )
 
 /** Options for speech generation. */
@@ -132,13 +162,13 @@ data class SpeechResult(
 data class SpeechCallOptions(
     val text: String = "",
     val voice: String? = null,
-    @SerialName("output_format") val outputFormat: String? = null,
+    val outputFormat: String? = null,
     val instructions: String? = null,
     val speed: Double? = null,
     val language: String? = null,
-    @SerialName("max_retries") val maxRetries: Long? = null,
+    val maxRetries: Long? = null,
     val timeout: TimeoutConfiguration? = null,
-    @SerialName("provider_options") val providerOptions: JsonElement? = null,
+    val providerOptions: ProviderOptions? = null,
     val headers: Map<String, String>? = null,
 )
 
@@ -149,8 +179,7 @@ data class SpeechCallOptions(
 /**
  * Generated images: all base64 strings or all binary byte arrays.
  *
- * Wire format (serde externally-tagged):
- * `{"Base64": ["...", ...]}` | `{"Binary": [[n,...], ...]}`.
+ * Wire format (untagged, `ImageOutputs.ts`): `["...", ...]` | `[[n,...], ...]`.
  */
 @Serializable(with = ImageOutputsSerializer::class)
 sealed interface ImageOutputs {
@@ -164,16 +193,16 @@ sealed interface ImageOutputs {
 /** Token usage for image generation (if reported). */
 @Serializable
 data class ImageUsage(
-    @SerialName("input_tokens") val inputTokens: Long? = null,
-    @SerialName("output_tokens") val outputTokens: Long? = null,
-    @SerialName("total_tokens") val totalTokens: Long? = null,
+    val inputTokens: Long? = null,
+    val outputTokens: Long? = null,
+    val totalTokens: Long? = null,
 )
 
 /** Provider response metadata for images. */
 @Serializable
 data class ImageResponse(
     val timestamp: String? = null,
-    @SerialName("model_id") val modelId: String? = null,
+    val modelId: String? = null,
     val headers: Map<String, String>? = null,
 )
 
@@ -182,24 +211,40 @@ data class ImageResponse(
 data class ImageResult(
     val images: ImageOutputs = ImageOutputs.Base64(emptyList()),
     val warnings: List<Warning> = emptyList(),
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    val providerMetadata: ProviderMetadata? = null,
     val response: ImageResponse = ImageResponse(),
     val usage: ImageUsage? = null,
 )
 
-/** Options for image generation. */
+/** An input image (edit source or mask), tagged on `type`. Mirrors `ImageFile.ts`. */
+@Serializable
+sealed interface ImageFile {
+    @Serializable
+    @SerialName("file")
+    data class File(val mediaType: String, val data: FileBytes) : ImageFile
+
+    @Serializable
+    @SerialName("url")
+    data class Url(val url: String) : ImageFile
+}
+
+/**
+ * Options for image generation. `n` and `providerOptions` are required on the
+ * wire, so they are always encoded (defaults: 1 image, no options).
+ */
+@OptIn(ExperimentalSerializationApi::class)
 @Serializable
 data class ImageCallOptions(
     val prompt: String? = null,
-    val n: Int? = null,
-    val size: String? = null,
-    @SerialName("aspect_ratio") val aspectRatio: String? = null,
+    @EncodeDefault val n: Int = 1,
+    val size: Size? = null,
+    val aspectRatio: AspectRatio? = null,
     val seed: Long? = null,
-    val files: List<JsonElement> = emptyList(),
-    val mask: JsonElement? = null,
-    @SerialName("max_retries") val maxRetries: Long? = null,
+    val files: List<ImageFile>? = null,
+    val mask: ImageFile? = null,
+    @EncodeDefault val providerOptions: ProviderOptions = emptyMap(),
+    val maxRetries: Long? = null,
     val timeout: TimeoutConfiguration? = null,
-    @SerialName("provider_options") val providerOptions: JsonElement? = null,
     val headers: Map<String, String>? = null,
 )
 
@@ -210,26 +255,22 @@ data class ImageCallOptions(
 /** A transcript segment with timing. */
 @Serializable
 data class TranscriptionSegment(
-    // Required f64 in the core (no Option) — see TranscriptionSegment.ts — so
-    // non-nullable here; the defaults follow the file's lenient-decode contract.
-    // Decode-only in practice: segments arrive from the core, which always emits
-    // all three fields, and nothing here sends them back.
     val text: String = "",
-    @SerialName("start_second") val startSecond: Double = 0.0,
-    @SerialName("end_second") val endSecond: Double = 0.0,
+    val startSecond: Double = 0.0,
+    val endSecond: Double = 0.0,
 )
 
 /** Request metadata for transcription. */
 @Serializable
 data class TranscriptionRequest(
-    @SerialName("body") val body: String? = null,
+    val body: String? = null,
 )
 
 /** Provider response metadata for transcription. */
 @Serializable
 data class TranscriptionResponse(
     val timestamp: String? = null,
-    @SerialName("model_id") val modelId: String? = null,
+    val modelId: String? = null,
     val headers: Map<String, String>? = null,
     val body: JsonElement? = null,
 )
@@ -240,21 +281,21 @@ data class TranscriptionResult(
     val text: String = "",
     val segments: List<TranscriptionSegment> = emptyList(),
     val language: String? = null,
-    @SerialName("duration_in_seconds") val durationInSeconds: Double? = null,
+    val durationInSeconds: Double? = null,
     val warnings: List<Warning> = emptyList(),
     val request: TranscriptionRequest? = null,
     val response: TranscriptionResponse = TranscriptionResponse(),
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    val providerMetadata: ProviderMetadata? = null,
 )
 
-/** Options for transcription. */
+/** Options for transcription. `audio` is raw bytes or a base64 string. */
 @Serializable
 data class TranscriptionCallOptions(
-    val audio: JsonElement = JsonObject(emptyMap()),
-    @SerialName("media_type") val mediaType: String = "",
-    @SerialName("max_retries") val maxRetries: Long? = null,
+    val audio: FileBytes = FileBytes.Base64(""),
+    val mediaType: String = "",
+    val maxRetries: Long? = null,
     val timeout: TimeoutConfiguration? = null,
-    @SerialName("provider_options") val providerOptions: JsonElement? = null,
+    val providerOptions: ProviderOptions? = null,
     val headers: Map<String, String>? = null,
 )
 
@@ -266,7 +307,7 @@ data class TranscriptionCallOptions(
 @Serializable
 data class RerankingRank(
     val index: Int = 0,
-    @SerialName("relevance_score") val relevanceScore: Double = 0.0,
+    val relevanceScore: Double = 0.0,
 )
 
 /** Provider response metadata for reranking. */
@@ -274,7 +315,7 @@ data class RerankingRank(
 data class RerankingResponse(
     val id: String? = null,
     val timestamp: String? = null,
-    @SerialName("model_id") val modelId: String? = null,
+    val modelId: String? = null,
     val headers: Map<String, String>? = null,
     val body: JsonElement? = null,
 )
@@ -283,20 +324,32 @@ data class RerankingResponse(
 @Serializable
 data class RerankingResult(
     val ranking: List<RerankingRank> = emptyList(),
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    val warnings: List<Warning> = emptyList(),
+    val providerMetadata: ProviderMetadata? = null,
+    val warnings: List<Warning>? = null,
     val response: RerankingResponse? = null,
 )
 
+/** The documents to rerank, tagged on `type`. Mirrors `RerankingDocuments.ts`. */
+@Serializable
+sealed interface RerankingDocuments {
+    @Serializable
+    @SerialName("text")
+    data class Text(val values: List<String>) : RerankingDocuments
+
+    @Serializable
+    @SerialName("object")
+    data class Objects(val values: List<JsonElement>) : RerankingDocuments
+}
+
 /** Options for reranking. */
 @Serializable
 data class RerankingCallOptions(
-    val documents: JsonElement = JsonArray(emptyList()),
+    val documents: RerankingDocuments,
     val query: String = "",
-    @SerialName("top_n") val topN: Int? = null,
-    @SerialName("max_retries") val maxRetries: Long? = null,
+    val topN: Int? = null,
+    val maxRetries: Long? = null,
     val timeout: TimeoutConfiguration? = null,
-    @SerialName("provider_options") val providerOptions: JsonElement? = null,
+    val providerOptions: ProviderOptions? = null,
     val headers: Map<String, String>? = null,
 )
 
@@ -304,50 +357,54 @@ data class RerankingCallOptions(
 // Video
 // ─────────────────────────────────────────────────────────────────────────────
 
-/** The URL variant of [VideoData]. */
-@Serializable
-data class VideoUrlData(
-    val url: String = "",
-    @SerialName("media_type") val mediaType: String = "",
-)
-
-/** The base64 variant of [VideoData]. */
-@Serializable
-data class VideoBase64Data(
-    val data: String = "",
-    @SerialName("media_type") val mediaType: String = "",
-)
-
-/** The binary variant of [VideoData]. */
+/** Generated video, tagged on `type`. Mirrors `VideoData.ts`. */
 @Serializable
-data class VideoBinaryData(
-    val data: List<Int> = emptyList(),
-    @SerialName("media_type") val mediaType: String = "",
-)
-
-/**
- * Generated video: a URL, base64 string, or raw binary bytes.
- *
- * Wire format (serde externally-tagged):
- * `{"Url": {...}}` | `{"Base64": {...}}` | `{"Binary": {...}}`.
- */
-@Serializable(with = VideoDataSerializer::class)
 sealed interface VideoData {
+    val mediaType: String
+
     /** A URL pointing at the generated video. */
-    data class Url(val value: VideoUrlData) : VideoData
+    @Serializable
+    @SerialName("url")
+    data class Url(val url: String, override val mediaType: String) : VideoData
 
     /** Base64-encoded video. */
-    data class Base64(val value: VideoBase64Data) : VideoData
+    @Serializable
+    @SerialName("base64")
+    data class Base64(val data: String, override val mediaType: String) : VideoData
 
     /** Raw binary video bytes (each element is a 0–255 byte value). */
-    data class Binary(val value: VideoBinaryData) : VideoData
+    @Serializable
+    @SerialName("binary")
+    data class Binary(val data: List<Int>, override val mediaType: String) : VideoData
+}
+
+/** An input video/image, tagged on `type`. Mirrors `VideoFile.ts`. */
+@Serializable
+sealed interface VideoFile {
+    @Serializable
+    @SerialName("file")
+    data class File(val mediaType: String, val data: FileBytes) : VideoFile
+
+    @Serializable
+    @SerialName("url")
+    data class Url(val url: String, val mediaType: String? = null) : VideoFile
 }
 
+@Serializable
+enum class VideoFrameType {
+    @SerialName("first_frame") FIRST_FRAME,
+    @SerialName("last_frame") LAST_FRAME,
+}
+
+/** An image pinned to a frame of the video. Mirrors `VideoFrameImage.ts`. */
+@Serializable
+data class VideoFrameImage(val image: VideoFile, val frameType: VideoFrameType)
+
 /** Provider response metadata for video. */
 @Serializable
 data class VideoResponse(
     val timestamp: String? = null,
-    @SerialName("model_id") val modelId: String? = null,
+    val modelId: String? = null,
     val headers: Map<String, String>? = null,
 )
 
@@ -356,29 +413,36 @@ data class VideoResponse(
 data class VideoResult(
     val videos: List<VideoData> = emptyList(),
     val warnings: List<Warning> = emptyList(),
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    val providerMetadata: ProviderMetadata? = null,
     val response: VideoResponse = VideoResponse(),
 )
 
 /** Per-call pacing overrides for the Core-owned video status poll loop. */
 @Serializable
 data class VideoPollOptions(
-    @SerialName("interval_ms") val intervalMs: Long? = null,
-    @SerialName("timeout_ms") val timeoutMs: Long? = null,
+    val intervalMs: Long? = null,
+    val timeoutMs: Long? = null,
 )
 
-/** Options for video generation. */
+/** Options for video generation. `n` and `providerOptions` are always encoded. */
+@OptIn(ExperimentalSerializationApi::class)
 @Serializable
 data class VideoCallOptions(
     val prompt: String? = null,
-    val n: Int? = null,
-    @SerialName("aspect_ratio") val aspectRatio: String? = null,
-    val resolution: String? = null,
+    @EncodeDefault val n: Int = 1,
+    val aspectRatio: AspectRatio? = null,
+    val resolution: Size? = null,
+    val duration: Double? = null,
+    val fps: Double? = null,
     val seed: Long? = null,
-    @SerialName("max_retries") val maxRetries: Long? = null,
+    val image: VideoFile? = null,
+    val frameImages: List<VideoFrameImage>? = null,
+    val inputReferences: List<VideoFile>? = null,
+    val generateAudio: Boolean? = null,
+    @EncodeDefault val providerOptions: ProviderOptions = emptyMap(),
+    val maxRetries: Long? = null,
     val poll: VideoPollOptions? = null,
     val timeout: TimeoutConfiguration? = null,
-    @SerialName("provider_options") val providerOptions: JsonElement? = null,
     val headers: Map<String, String>? = null,
 )
 
@@ -392,8 +456,9 @@ data class SearchResultItem(
     val title: String? = null,
     val url: String? = null,
     val content: String? = null,
-    @SerialName("raw_content") val rawContent: String? = null,
+    val rawContent: String? = null,
     val score: Double? = null,
+    val providerMetadata: ProviderMetadata? = null,
 )
 
 /** Provider response metadata for search. */
@@ -408,7 +473,7 @@ data class SearchResponse(
 data class SearchResult(
     val results: List<SearchResultItem> = emptyList(),
     val answer: String? = null,
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    val providerMetadata: ProviderMetadata? = null,
     val warnings: List<Warning> = emptyList(),
     val response: SearchResponse? = null,
 )
@@ -417,14 +482,14 @@ data class SearchResult(
 @Serializable
 data class SearchCallOptions(
     val query: String = "",
-    @SerialName("max_results") val maxResults: Int? = null,
-    @SerialName("include_raw_content") val includeRawContent: Boolean? = null,
-    @SerialName("time_range") val timeRange: String? = null,
-    @SerialName("include_domains") val includeDomains: List<String> = emptyList(),
-    @SerialName("exclude_domains") val excludeDomains: List<String> = emptyList(),
-    @SerialName("max_retries") val maxRetries: Long? = null,
+    val maxResults: Int? = null,
+    val includeRawContent: Boolean? = null,
+    val timeRange: String? = null,
+    val includeDomains: List<String>? = null,
+    val excludeDomains: List<String>? = null,
+    val maxRetries: Long? = null,
     val timeout: TimeoutConfiguration? = null,
-    @SerialName("provider_options") val providerOptions: JsonElement? = null,
+    val providerOptions: ProviderOptions? = null,
     val headers: Map<String, String>? = null,
 )
 
@@ -435,114 +500,88 @@ data class SearchCallOptions(
 /** Result of a file upload. */
 @Serializable
 data class UploadFileResult(
-    @SerialName("provider_reference") val providerReference: Map<String, String> = emptyMap(),
-    @SerialName("media_type") val mediaType: String? = null,
+    val providerReference: Map<String, String> = emptyMap(),
+    val mediaType: String? = null,
     val filename: String? = null,
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    val providerMetadata: ProviderMetadata? = null,
     val warnings: List<Warning> = emptyList(),
 )
 
+/** Upload payload, tagged on `type`. Mirrors `UploadFileData.ts`. */
+@Serializable
+sealed interface UploadFileData {
+    @Serializable
+    @SerialName("data")
+    data class Data(val data: FileBytes) : UploadFileData
+
+    @Serializable
+    @SerialName("text")
+    data class Text(val text: String) : UploadFileData
+}
+
 /** Options for a file upload. */
 @Serializable
 data class UploadFileCallOptions(
-    val data: JsonElement = JsonObject(emptyMap()),
-    @SerialName("media_type") val mediaType: String = "",
+    val data: UploadFileData = UploadFileData.Data(FileBytes.Base64("")),
+    val mediaType: String = "",
     val filename: String? = null,
-    @SerialName("provider_options") val providerOptions: JsonElement? = null,
+    val providerOptions: ProviderOptions? = null,
 )
 
 // ─────────────────────────────────────────────────────────────────────────────
-// Custom serializers for the serde-style externally-tagged unions.
-// (Same approach as [StreamPartSerializer] in `Types.kt`.)
+// Custom serializers for the untagged unions.
 // ─────────────────────────────────────────────────────────────────────────────
 
-/** (De)serializer for [AudioData]: `{"Base64": "..."}` | `{"Binary": [n,...]}`. */
+/** (De)serializer for [AudioData]: `"..."` | `[n,...]`. */
 object AudioDataSerializer : KSerializer<AudioData> {
     override val descriptor: SerialDescriptor = buildClassSerialDescriptor("aimux.AudioData")
 
     override fun deserialize(decoder: Decoder): AudioData {
         val json = decoder as? JsonDecoder
             ?: throw SerializationException("AudioData can only be decoded from JSON")
-        val obj = json.decodeJsonElement().jsonObject
-        require(obj.size == 1) { "AudioData must be a single-key object, got: $obj" }
-        val (tag, inner) = obj.entries.single()
-        return when (tag) {
-            "Base64" -> AudioData.Base64(inner.jsonPrimitive.content)
-            "Binary" -> AudioData.Binary(inner.jsonArray.map { it.jsonPrimitive.content.toInt() })
-            else -> throw SerializationException("Unknown AudioData variant: '$tag'")
+        return when (val el = json.decodeJsonElement()) {
+            is JsonArray -> AudioData.Binary(el.map { it.jsonPrimitive.content.toInt() })
+            is JsonPrimitive -> AudioData.Base64(el.content)
+            else -> throw SerializationException("AudioData must be a base64 string or a byte array, got: $el")
         }
     }
 
     override fun serialize(encoder: Encoder, value: AudioData) {
         val json = encoder as? JsonEncoder
             ?: throw SerializationException("AudioData can only be encoded to JSON")
-        val (tag, inner) = when (value) {
-            is AudioData.Base64 -> "Base64" to JsonPrimitive(value.value)
-            is AudioData.Binary -> "Binary" to JsonArray(value.value.map { JsonPrimitive(it) })
-        }
-        json.encodeJsonElement(JsonObject(mapOf(tag to inner)))
+        json.encodeJsonElement(
+            when (value) {
+                is AudioData.Base64 -> JsonPrimitive(value.value)
+                is AudioData.Binary -> JsonArray(value.value.map { JsonPrimitive(it) })
+            }
+        )
     }
 }
 
-/** (De)serializer for [ImageOutputs]: `{"Base64": [...]}` | `{"Binary": [[...],...]}`. */
+/** (De)serializer for [ImageOutputs]: `["...",...]` | `[[n,...],...]`. */
 object ImageOutputsSerializer : KSerializer<ImageOutputs> {
     override val descriptor: SerialDescriptor = buildClassSerialDescriptor("aimux.ImageOutputs")
 
     override fun deserialize(decoder: Decoder): ImageOutputs {
         val json = decoder as? JsonDecoder
             ?: throw SerializationException("ImageOutputs can only be decoded from JSON")
-        val obj = json.decodeJsonElement().jsonObject
-        require(obj.size == 1) { "ImageOutputs must be a single-key object, got: $obj" }
-        val (tag, inner) = obj.entries.single()
-        return when (tag) {
-            "Base64" -> ImageOutputs.Base64(inner.jsonArray.map { it.jsonPrimitive.content })
-            "Binary" -> ImageOutputs.Binary(
-                inner.jsonArray.map { row -> row.jsonArray.map { it.jsonPrimitive.content.toInt() } },
-            )
-            else -> throw SerializationException("Unknown ImageOutputs variant: '$tag'")
+        val items = json.decodeJsonElement().jsonArray
+        // An empty list carries no evidence either way; treat it as base64.
+        return if (items.firstOrNull() is JsonArray) {
+            ImageOutputs.Binary(items.map { row -> row.jsonArray.map { it.jsonPrimitive.content.toInt() } })
+        } else {
+            ImageOutputs.Base64(items.map { it.jsonPrimitive.content })
         }
     }
 
     override fun serialize(encoder: Encoder, value: ImageOutputs) {
         val json = encoder as? JsonEncoder
             ?: throw SerializationException("ImageOutputs can only be encoded to JSON")
-        val (tag, inner) = when (value) {
-            is ImageOutputs.Base64 -> "Base64" to JsonArray(value.value.map { JsonPrimitive(it) })
-            is ImageOutputs.Binary -> "Binary" to JsonArray(value.value.map { row -> JsonArray(row.map { JsonPrimitive(it) }) })
-        }
-        json.encodeJsonElement(JsonObject(mapOf(tag to inner)))
-    }
-}
-
-/** (De)serializer for [VideoData]: `{"Url": {...}}` | `{"Base64": {...}}` | `{"Binary": {...}}`. */
-object VideoDataSerializer : KSerializer<VideoData> {
-    override val descriptor: SerialDescriptor = buildClassSerialDescriptor("aimux.VideoData")
-
-    override fun deserialize(decoder: Decoder): VideoData {
-        val json = decoder as? JsonDecoder
-            ?: throw SerializationException("VideoData can only be decoded from JSON")
-        val obj = json.decodeJsonElement().jsonObject
-        require(obj.size == 1) { "VideoData must be a single-key object, got: $obj" }
-        val (tag, inner) = obj.entries.single()
-        val innerObj = inner as? JsonObject ?: JsonObject(emptyMap())
-        val ctx = json.json
-        return when (tag) {
-            "Url" -> VideoData.Url(ctx.decodeFromJsonElement(VideoUrlData.serializer(), innerObj))
-            "Base64" -> VideoData.Base64(ctx.decodeFromJsonElement(VideoBase64Data.serializer(), innerObj))
-            "Binary" -> VideoData.Binary(ctx.decodeFromJsonElement(VideoBinaryData.serializer(), innerObj))
-            else -> throw SerializationException("Unknown VideoData variant: '$tag'")
-        }
-    }
-
-    override fun serialize(encoder: Encoder, value: VideoData) {
-        val json = encoder as? JsonEncoder
-            ?: throw SerializationException("VideoData can only be encoded to JSON")
-        val ctx = json.json
-        val (tag, inner) = when (value) {
-            is VideoData.Url -> "Url" to ctx.encodeToJsonElement(VideoUrlData.serializer(), value.value)
-            is VideoData.Base64 -> "Base64" to ctx.encodeToJsonElement(VideoBase64Data.serializer(), value.value)
-            is VideoData.Binary -> "Binary" to ctx.encodeToJsonElement(VideoBinaryData.serializer(), value.value)
-        }
-        json.encodeJsonElement(JsonObject(mapOf(tag to inner)))
+        json.encodeJsonElement(
+            when (value) {
+                is ImageOutputs.Base64 -> JsonArray(value.value.map { JsonPrimitive(it) })
+                is ImageOutputs.Binary -> JsonArray(value.value.map { row -> JsonArray(row.map { JsonPrimitive(it) }) })
+            }
+        )
     }
 }
diff --git a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/ToolCallRepair.kt b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/ToolCallRepair.kt
index dd913f937..3cae4e4c8 100644
--- a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/ToolCallRepair.kt
+++ b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/ToolCallRepair.kt
@@ -12,7 +12,6 @@
 
 package ai.arcships.aimux
 
-import kotlinx.serialization.SerialName
 import kotlinx.serialization.Serializable
 import kotlinx.serialization.json.JsonArray
 import kotlinx.serialization.json.JsonElement
@@ -40,19 +39,19 @@ import java.util.concurrent.FutureTask
  */
 @Serializable
 data class RawToolCall(
-    @SerialName("tool_call_id") val toolCallId: String,
-    @SerialName("tool_name") val toolName: String,
+    val toolCallId: String,
+    val toolName: String,
     val input: String,
-    @SerialName("provider_executed") val providerExecuted: Boolean? = null,
+    val providerExecuted: Boolean? = null,
     val dynamic: Boolean? = null,
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    val providerMetadata: ProviderMetadata? = null,
 )
 
 /**
  * Everything a repair function is given about one invalid tool call.
  *
  * Mirrors the AI SDK `repairToolCall` argument. [error] is the serialized
- * `AiMuxError` that made the call invalid (e.g. `{"InvalidToolInput": …}`),
+ * `AiMuxError` that made the call invalid (e.g. `{"name":"AI_InvalidToolInputError",…}`),
  * carried verbatim from the call rather than re-derived. [inputSchema] is the
  * JSON Schema of the named function tool, or the AI SDK's empty-object schema
  * when the name resolves to no function tool. [messages] and [instructions]
@@ -60,9 +59,9 @@ data class RawToolCall(
  */
 @Serializable
 data class ToolCallRepairContext(
-    @SerialName("tool_call") val toolCall: RawToolCall,
+    val toolCall: RawToolCall,
     val error: JsonElement,
-    @SerialName("input_schema") val inputSchema: JsonElement,
+    val inputSchema: JsonElement,
     val tools: List<Tool> = emptyList(),
     val messages: List<ModelMessage> = emptyList(),
     val instructions: String? = null,
@@ -88,7 +87,7 @@ typealias RepairToolCall = (ToolCallRepairContext) -> RawToolCall?
 /**
  * Build the repair argument for one invalid tool call.
  *
- * @param toolCallJson One `tool_calls` entry with `"invalid": true`.
+ * @param toolCallJson One `toolCalls` entry with `"invalid": true`.
  * @param promptJson The prompt the call was generated with (the same string
  *   passed to `generate_text` / `stream_text`).
  * @param optsJson The options the call was generated with (same rule).
@@ -108,7 +107,7 @@ fun toolCallRepairContext(toolCallJson: String, promptJson: String, optsJson: St
  * Resolve one invalid tool call against a repair reply, returning the final
  * `ToolCall` JSON — valid, or invalid carrying a nested `ToolCallRepair` error.
  *
- * @param replyJson `{"type":"repaired","tool_call":…}` | `{"type":"unchanged"}`
+ * @param replyJson `{"type":"repaired","toolCall":…}` | `{"type":"unchanged"}`
  *   | `{"type":"failed","message":…}`.
  * @throws InvalidArgumentError when the options carry no tools, or the call is
  *   not an invalid one.
@@ -126,7 +125,7 @@ fun applyToolCallRepair(toolCallJson: String, optsJson: String?, replyJson: Stri
  * Apply a repair reply to a whole serialized result (`GenerateTextResult`,
  * `GenerateObjectResult` via its nested `raw`, or `StreamTextResultAggregated`).
  *
- * Rewrites both `tool_calls` and the matching `response_messages` tool-call
+ * Rewrites both `toolCalls` and the matching `responseMessages` tool-call
  * part — the transcript must agree with the repaired call, or the next turn
  * sends the model the unrepaired arguments again.
  *
@@ -167,7 +166,7 @@ private fun repairReply(repair: RepairToolCall, contextJson: String): String {
         val replacement = repair(context) ?: return REPLY_UNCHANGED
         buildJsonObject {
             put("type", "repaired")
-            put("tool_call", AimuxJson.encodeToJsonElement(RawToolCall.serializer(), replacement))
+            put("toolCall", AimuxJson.encodeToJsonElement(RawToolCall.serializer(), replacement))
         }.toString()
     } catch (e: Exception) {
         // A host exception has no typed counterpart in core; only its message
@@ -197,8 +196,8 @@ internal fun repairToolCallsInResult(
 ): String {
     val root = AimuxJson.parseToJsonElement(resultJson) as? JsonObject ?: return resultJson
     // GenerateObjectResult nests the text result under `raw`.
-    val target = if (root.containsKey("tool_calls")) root else root["raw"] as? JsonObject ?: return resultJson
-    val invalid = (target["tool_calls"] as? JsonArray)
+    val target = if (root.containsKey("toolCalls")) root else root["raw"] as? JsonObject ?: return resultJson
+    val invalid = (target["toolCalls"] as? JsonArray)
         ?.mapNotNull { it as? JsonObject }
         ?.filter(::isInvalid)
         ?: return resultJson
@@ -207,7 +206,7 @@ internal fun repairToolCallsInResult(
     for (call in invalid) {
         val contextJson = toolCallRepairContext(call.toString(), promptJson, optsJson)
         if (AimuxJson.parseToJsonElement(contextJson) is JsonNull) continue
-        val id = call["tool_call_id"]?.jsonPrimitive?.content ?: continue
+        val id = call["toolCallId"]?.jsonPrimitive?.content ?: continue
         patched = applyToolCallRepairToResult(patched, optsJson, id, repairReply(repair, contextJson))
     }
     return patched
@@ -250,7 +249,7 @@ internal fun offCallbackThread(model: Model, repair: RepairToolCall): RepairTool
 }
 
 /**
- * Repair an invalid `ToolCall` stream part, returning the replacement part
+ * Repair an invalid `tool-call` stream part, returning the replacement part
  * JSON. Any other part (tool-input deltas included) is returned untouched, so
  * deltas still reach the caller immediately — AI SDK behaviour.
  */
@@ -260,13 +259,12 @@ internal fun repairToolCallStreamPart(
     optsJson: String?,
     repair: RepairToolCall,
 ): String {
-    // StreamPart is externally tagged: {"ToolCall": {…}}.
+    // A tool-call part is the ToolCall itself plus `"type": "tool-call"`.
     val part = AimuxJson.parseToJsonElement(partJson) as? JsonObject ?: return partJson
-    val call = part["ToolCall"] as? JsonObject ?: return partJson
-    if (!isInvalid(call)) return partJson
+    if ((part["type"] as? JsonPrimitive)?.content != "tool-call" || !isInvalid(part)) return partJson
 
-    val contextJson = toolCallRepairContext(call.toString(), promptJson, optsJson)
+    val contextJson = toolCallRepairContext(part.toString(), promptJson, optsJson)
     if (AimuxJson.parseToJsonElement(contextJson) is JsonNull) return partJson
-    val repaired = applyToolCallRepair(call.toString(), optsJson, repairReply(repair, contextJson))
-    return JsonObject(part + ("ToolCall" to AimuxJson.parseToJsonElement(repaired))).toString()
+    val repaired = applyToolCallRepair(part.toString(), optsJson, repairReply(repair, contextJson))
+    return JsonObject(mapOf("type" to JsonPrimitive("tool-call")) + (AimuxJson.parseToJsonElement(repaired) as JsonObject)).toString()
 }
diff --git a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/TypedModel.kt b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/TypedModel.kt
index 8d5f2d3ed..6b2fac6db 100644
--- a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/TypedModel.kt
+++ b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/TypedModel.kt
@@ -265,7 +265,7 @@ class TypedModel(private val raw: Model, private val ownsModel: Boolean = false)
                     rawPartJson
                 }
                 try {
-                    onPart(AimuxJson.decodeFromString(StreamPartSerializer, partJson))
+                    onPart(AimuxJson.decodeFromString(StreamPart.serializer(), partJson))
                 } catch (error: Exception) {
                     onError(
                         "failed to decode StreamPart: ${error.message ?: error::class.simpleName}",
diff --git a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Types.kt b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Types.kt
index 1d0e0b3ff..fb4d77b2b 100644
--- a/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Types.kt
+++ b/bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Types.kt
@@ -2,13 +2,15 @@
  * aimux — typed data classes mirroring the ts-rs output in `bindings/node/src/types`
  * (the generated TypeScript types).
  *
- * Field names use camelCase in Kotlin and are mapped to the wire format's
- * snake_case via [kotlinx.serialization.SerialName]. The raw JSON boundary is
- * handled by [TypedModel] — callers of this layer never parse JSON by hand.
+ * The wire JSON is the AI SDK's: camelCase field names, unions tagged with a
+ * `type` key (a few use `role`, `name` or `sourceType`), optional fields
+ * absent rather than null. Kotlin property names equal the wire names, so no
+ * [kotlinx.serialization.SerialName] is needed on fields. The raw JSON boundary
+ * is handled by [TypedModel] — callers of this layer never parse JSON by hand.
  *
- * These types intentionally lenient on decode (unknown keys ignored, every
- * field has a default) so that future provider/engine additions do not break
- * existing clients. The serialization config lives in [AimuxJson].
+ * Decode is lenient (unknown keys ignored, every result field has a default)
+ * so engine additions do not break existing clients. The serialization config
+ * lives in [AimuxJson].
  */
 
 package ai.arcships.aimux
@@ -31,17 +33,18 @@ import kotlinx.serialization.json.JsonElement
 import kotlinx.serialization.json.JsonEncoder
 import kotlinx.serialization.json.JsonObject
 import kotlinx.serialization.json.JsonPrimitive
-import kotlinx.serialization.json.jsonObject
 import kotlinx.serialization.json.jsonPrimitive
 
 // ─────────────────────────────────────────────────────────────────────────────
 // Shared Json instance.
 //
 //  - ignoreUnknownKeys  : tolerate forward-compatible fields from the engine.
-//  - explicitNulls=false: omit null fields when encoding (so GenerateTextOptions
-//                         only carries fields the caller actually set, matching
-//                         the engine's optional-everything schema).
+//  - explicitNulls=false: omit null fields when encoding (optional fields are
+//                         absent on the wire, never null).
 //  - encodeDefaults=false: do not encode default values (keeps payloads small).
+//
+// Unions tagged with `type` use kotlinx's default sealed-class polymorphism
+// (`classDiscriminator` stays "type").
 // ─────────────────────────────────────────────────────────────────────────────
 
 val AimuxJson: Json = Json {
@@ -50,6 +53,12 @@ val AimuxJson: Json = Json {
     encodeDefaults = false
 }
 
+/** Provider-specific options / metadata: `{ providerName: { key: value } }`. */
+typealias ProviderOptions = Map<String, Map<String, JsonElement>>
+
+/** Provider-specific metadata on a result (same shape as [ProviderOptions]). */
+typealias ProviderMetadata = Map<String, Map<String, JsonElement>>
+
 // ─────────────────────────────────────────────────────────────────────────────
 // Enums (simple string enums on the wire).
 // ─────────────────────────────────────────────────────────────────────────────
@@ -90,9 +99,9 @@ enum class ReasoningEffort {
 @Serializable
 data class InputTokenUsage(
     val total: Long? = null,
-    @SerialName("no_cache") val noCache: Long? = null,
-    @SerialName("cache_read") val cacheRead: Long? = null,
-    @SerialName("cache_write") val cacheWrite: Long? = null,
+    val noCache: Long? = null,
+    val cacheRead: Long? = null,
+    val cacheWrite: Long? = null,
 )
 
 @Serializable
@@ -102,16 +111,11 @@ data class OutputTokenUsage(
     val reasoning: Long? = null,
 )
 
-/**
- * Token usage statistics.
- *
- * Mirrors `Usage.ts`: `{ input_tokens: InputTokenUsage, output_tokens: OutputTokenUsage,
- * raw?: JsonValue | null }`.
- */
+/** Token usage statistics. Mirrors `Usage.ts`. */
 @Serializable
 data class Usage(
-    @SerialName("input_tokens") val inputTokens: InputTokenUsage = InputTokenUsage(),
-    @SerialName("output_tokens") val outputTokens: OutputTokenUsage = OutputTokenUsage(),
+    val inputTokens: InputTokenUsage = InputTokenUsage(),
+    val outputTokens: OutputTokenUsage = OutputTokenUsage(),
     val raw: JsonObject? = null,
 ) {
     companion object {
@@ -121,35 +125,27 @@ data class Usage(
     }
 }
 
-/**
- * Unified finish reason.
- *
- * Mirrors `FinishReason.ts`: `{ unified: FinishReasonUnified, raw: string | null }`.
- */
+/** Unified finish reason. Mirrors `FinishReason.ts`: `raw` is absent when the provider sent none. */
 @Serializable
 data class FinishReason(
     val unified: FinishReasonUnified = FinishReasonUnified.OTHER,
     val raw: String? = null,
 )
 
-/**
- * Metadata about the API response.
- *
- * Mirrors `ResponseMetadata.ts`: `{ id: string | null, timestamp: string | null,
- * model_id: string | null }`.
- */
+/** Mirrors `ResponseMetadata.ts`. */
 @Serializable
 data class ResponseMetadata(
     val id: String? = null,
     val timestamp: String? = null,
-    @SerialName("model_id") val modelId: String? = null,
+    val modelId: String? = null,
 )
 
+/** Mirrors `ResponseInfo.ts`: [ResponseMetadata] plus the HTTP headers and body. */
 @Serializable
 data class ResponseInfo(
     val id: String? = null,
     val timestamp: String? = null,
-    @SerialName("model_id") val modelId: String? = null,
+    val modelId: String? = null,
     val headers: Map<String, String>? = null,
     val body: JsonElement? = null,
 )
@@ -158,26 +154,21 @@ data class ResponseInfo(
 data class RequestInfo(val body: JsonElement? = null)
 
 /**
- * A tool call requested by the model.
- *
- * Mirrors `ToolCall.ts`: `{ tool_call_id, tool_name, input: JsonValue,
- * provider_executed?: bool | null, dynamic?: bool | null,
- * provider_metadata?: JsonValue | null }`.
+ * A tool call requested by the model. Mirrors `ToolCall.ts`.
  *
- * `input` is a [JsonElement] because it is usually an arbitrary JSON object
- * (the tool arguments) whose shape is tool-specific.
- *
- * `invalid` is set by Core when the tool call stays invalid after optional
- * repair; `error` is the typed lookup, parse, schema, or repair failure.
+ * `input` is the parsed argument value (usually an object). `invalid` is set
+ * by Core when the call stays invalid after optional repair; `error` is the
+ * `name`-tagged `AiMuxError` JSON of the lookup, parse, schema, or repair
+ * failure (`{"name":"AI_NoSuchToolError",...}`).
  */
 @Serializable
 data class ToolCall(
-    @SerialName("tool_call_id") val toolCallId: String,
-    @SerialName("tool_name") val toolName: String,
+    val toolCallId: String,
+    val toolName: String,
     val input: JsonElement = JsonObject(emptyMap()),
-    @SerialName("provider_executed") val providerExecuted: Boolean? = null,
-    @SerialName("dynamic") val dynamic: Boolean? = null,
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    val providerExecuted: Boolean? = null,
+    val dynamic: Boolean? = null,
+    val providerMetadata: ProviderMetadata? = null,
     val invalid: Boolean? = null,
     val error: JsonElement? = null,
 )
@@ -185,34 +176,24 @@ data class ToolCall(
 // ─────────────────────────────────────────────────────────────────────────────
 // Tools (input side of GenerateTextOptions).
 //
-// `Tool` is an internally-tagged union on `type` (`"function" | "provider"`),
-// which matches kotlinx.serialization's default sealed-class polymorphism with
-// the `"type"` discriminator — no custom serializer needed.
+// `Tool` is an internally-tagged union on `type` (`"function" | "provider"`).
 // ─────────────────────────────────────────────────────────────────────────────
 
 @Serializable
 data class FunctionToolInputExample(val input: JsonObject)
 
-/**
- * A user-defined function tool.
- *
- * Mirrors `FunctionTool.ts`. `input_schema` is a [JsonElement] (a JSON Schema).
- */
+/** A user-defined function tool. Mirrors `FunctionTool.ts`; `inputSchema` is a JSON Schema. */
 @Serializable
 data class FunctionTool(
     val name: String,
     val description: String? = null,
-    @SerialName("input_schema") val inputSchema: JsonElement,
+    val inputSchema: JsonElement,
     val strict: Boolean? = null,
-    @SerialName("provider_options") val providerOptions: Map<String, JsonElement>? = null,
-    @SerialName("input_examples") val inputExamples: List<FunctionToolInputExample>? = null,
+    val providerOptions: ProviderOptions? = null,
+    val inputExamples: List<FunctionToolInputExample>? = null,
 )
 
-/**
- * A provider-defined tool (e.g. `anthropic.web_search_20250305`).
- *
- * Mirrors `ProviderTool.ts`.
- */
+/** A provider-defined tool (e.g. `anthropic.web_search_20250305`). Mirrors `ProviderTool.ts`. */
 @Serializable
 data class ProviderTool(
     val id: String,
@@ -221,10 +202,8 @@ data class ProviderTool(
 )
 
 /**
- * A tool that can be either a function tool or a provider tool.
- *
- * Mirrors `Tool.ts`. Serialized as `{"type":"function", ...}` /
- * `{"type":"provider", ...}` (internal `"type"` discriminator).
+ * A function tool or a provider tool. Mirrors `Tool.ts`; serialized as
+ * `{"type":"function", ...}` / `{"type":"provider", ...}`.
  */
 @Serializable
 sealed interface Tool {
@@ -233,10 +212,10 @@ sealed interface Tool {
     data class Function(
         val name: String,
         val description: String? = null,
-        @SerialName("input_schema") val inputSchema: JsonElement,
+        val inputSchema: JsonElement,
         val strict: Boolean? = null,
-        @SerialName("provider_options") val providerOptions: Map<String, JsonElement>? = null,
-        @SerialName("input_examples") val inputExamples: List<FunctionToolInputExample>? = null,
+        val providerOptions: ProviderOptions? = null,
+        val inputExamples: List<FunctionToolInputExample>? = null,
     ) : Tool {
         companion object {
             /** Convenience constructor from a [FunctionTool]. */
@@ -268,8 +247,8 @@ sealed interface Tool {
  * How the model should choose tools.
  *
  * Mirrors `ToolChoice.ts`: `"auto" | "none" | "required" | { type: "tool",
- * toolName: "..." }`. This is a mixed tagged/untagged shape (bare strings plus a
- * tagged object), so a custom serializer handles the two forms.
+ * toolName: "..." }`. This is a mixed untagged/tagged shape (bare strings plus
+ * a tagged object), so a custom serializer handles the two forms.
  */
 @Serializable(with = ToolChoiceSerializer::class)
 sealed interface ToolChoice {
@@ -329,170 +308,254 @@ object ToolChoiceSerializer : KSerializer<ToolChoice> {
 }
 
 // ─────────────────────────────────────────────────────────────────────────────
-// ContentPart (multi-part message content).
-//
-// `ContentPart` is an internally-tagged enum (`{"type": "text", "text": ...}`,
-// `{"type": "image", "image": [...], ...}`). A custom serializer dispatches on
-// the `"type"` key and re-injects it on encode.
+// File bytes / file data.
 // ─────────────────────────────────────────────────────────────────────────────
 
 /**
- * A part of a multi-part message.
+ * Raw bytes or a base64 string. Mirrors `FileBytes.ts`: `Array<number> | string`
+ * (untagged — a JSON array is binary, a JSON string is base64). Also the shape
+ * of every `Array<number> | string` field (audio input, image / video file data).
+ */
+@Serializable(with = FileBytesSerializer::class)
+sealed interface FileBytes {
+
+    /** Raw binary bytes (a JSON array of 0–255 ints on the wire). */
+    data class Binary(val data: List<Int> = emptyList()) : FileBytes
+
+    /** A base64-encoded string. */
+    data class Base64(val data: String = "") : FileBytes
+}
+
+object FileBytesSerializer : KSerializer<FileBytes> {
+    override val descriptor: SerialDescriptor = buildClassSerialDescriptor("aimux.FileBytes")
+
+    override fun deserialize(decoder: Decoder): FileBytes {
+        val json = decoder as? JsonDecoder
+            ?: throw SerializationException("FileBytes can only be decoded from JSON")
+        return when (val el = json.decodeJsonElement()) {
+            is JsonArray -> FileBytes.Binary(el.map { it.jsonPrimitive.content.toInt() })
+            is JsonPrimitive -> FileBytes.Base64(el.content)
+            else -> throw SerializationException("FileBytes must be a byte array or a base64 string, got: $el")
+        }
+    }
+
+    override fun serialize(encoder: Encoder, value: FileBytes) {
+        val json = encoder as? JsonEncoder
+            ?: throw SerializationException("FileBytes can only be encoded to JSON")
+        json.encodeJsonElement(
+            when (value) {
+                is FileBytes.Binary -> JsonArray(value.data.map { JsonPrimitive(it) })
+                is FileBytes.Base64 -> JsonPrimitive(value.data)
+            }
+        )
+    }
+}
+
+/** File data, tagged on `type`. Mirrors `FileData.ts`. */
+@Serializable
+sealed interface FileData {
+
+    @Serializable
+    @SerialName("data")
+    data class Data(val data: FileBytes = FileBytes.Base64("")) : FileData
+
+    @Serializable
+    @SerialName("url")
+    data class Url(val url: String = "", val originalUrl: String? = null) : FileData
+
+    @Serializable
+    @SerialName("reference")
+    data class Reference(val reference: Map<String, String> = emptyMap()) : FileData
+
+    @Serializable
+    @SerialName("text")
+    data class Text(val text: String = "") : FileData
+}
+
+/** Generated file data, tagged on `type`. Mirrors `GeneratedFileData.ts`. */
+@Serializable
+sealed interface GeneratedFileData {
+
+    @Serializable
+    @SerialName("data")
+    data class Data(val data: FileBytes = FileBytes.Base64("")) : GeneratedFileData
+
+    @Serializable
+    @SerialName("url")
+    data class Url(val url: String = "", val originalUrl: String? = null) : GeneratedFileData
+}
+
+/** A file produced by the model. Mirrors `GeneratedFile.ts`. */
+@Serializable
+data class GeneratedFile(
+    val data: GeneratedFileData = GeneratedFileData.Data(),
+    val mediaType: String = "",
+    val providerMetadata: ProviderMetadata? = null,
+)
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Tool results inside a prompt (ContentPart `tool-result`).
+// ─────────────────────────────────────────────────────────────────────────────
+
+/** One element of a `content` tool output. Mirrors `ToolResultContent.ts`. */
+@Serializable
+sealed interface ToolResultContent {
+    @Serializable
+    @SerialName("text")
+    data class Text(val text: String = "", val providerOptions: ProviderOptions? = null) : ToolResultContent
+
+    @Serializable
+    @SerialName("file")
+    data class File(
+        val data: FileData = FileData.Data(),
+        val mediaType: String = "",
+        val filename: String? = null,
+        val providerOptions: ProviderOptions? = null,
+    ) : ToolResultContent
+
+    @Serializable
+    @SerialName("custom")
+    data class Custom(val providerOptions: ProviderOptions? = null) : ToolResultContent
+}
+
+/** The output of a tool call, tagged on `type`. Mirrors `ToolResultOutput.ts`. */
+@Serializable
+sealed interface ToolResultOutput {
+    @Serializable
+    @SerialName("text")
+    data class Text(val value: String, val providerOptions: ProviderOptions? = null) : ToolResultOutput
+
+    @Serializable
+    @SerialName("json")
+    data class JsonValue(val value: JsonElement, val providerOptions: ProviderOptions? = null) : ToolResultOutput
+
+    @Serializable
+    @SerialName("execution-denied")
+    data class ExecutionDenied(val reason: String? = null, val providerOptions: ProviderOptions? = null) : ToolResultOutput
+
+    @Serializable
+    @SerialName("error-text")
+    data class ErrorText(val value: String, val providerOptions: ProviderOptions? = null) : ToolResultOutput
+
+    @Serializable
+    @SerialName("error-json")
+    data class ErrorJson(val value: JsonElement, val providerOptions: ProviderOptions? = null) : ToolResultOutput
+
+    @Serializable
+    @SerialName("content")
+    data class Content(val value: List<ToolResultContent>) : ToolResultOutput
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// ContentPart (multi-part message content), tagged on `type`.
+// ─────────────────────────────────────────────────────────────────────────────
+
+/**
+ * A part of a multi-part message. Mirrors `ContentPart.ts`.
  *
- * Mirrors `ContentPart.ts` (internally tagged on `type`). Shared between
- * [ModelMessage] (user-facing) and the provider-facing prompt. Every field has
- * a default and unknown keys are ignored on decode (see [AimuxJson]), so future
- * part additions do not break existing clients. `provider_options` is a
- * [JsonElement] because it is an opaque `Record<string, JSONObject>`.
+ * Shared by [ModelMessage] (user-facing) and the provider-facing prompt. A
+ * tool result carries a [ToolResultOutput] (`output`), not a bare value.
  */
-@Serializable(with = ContentPartSerializer::class)
+@Serializable
 sealed interface ContentPart {
     @Serializable
-    data class Custom(val kind: String, @SerialName("provider_options") val providerOptions: JsonElement? = null) : ContentPart
+    @SerialName("text")
+    data class Text(val text: String = "", val providerOptions: ProviderOptions? = null) : ContentPart
 
     @Serializable
-    data class ReasoningFile(val data: GeneratedFileData, @SerialName("media_type") val mediaType: String, @SerialName("provider_options") val providerOptions: JsonElement? = null) : ContentPart
+    @SerialName("custom")
+    data class Custom(val kind: String, val providerOptions: ProviderOptions? = null) : ContentPart
 
     @Serializable
-    data class ToolApprovalRequest(@SerialName("approval_id") val approvalId: String, @SerialName("tool_call_id") val toolCallId: String, val reason: String? = null, @SerialName("is_automatic") val isAutomatic: Boolean? = null, val signature: String? = null, @SerialName("input_schema_input") val inputSchemaInput: JsonElement? = null) : ContentPart
-
+    @SerialName("reasoning-file")
+    data class ReasoningFile(
+        val data: GeneratedFileData,
+        val mediaType: String,
+        val providerOptions: ProviderOptions? = null,
+    ) : ContentPart
 
     @Serializable
-    data class Text(
-        val text: String = "",
-        @SerialName("provider_options") val providerOptions: JsonElement? = null,
+    @SerialName("tool-approval-request")
+    data class ToolApprovalRequest(
+        val approvalId: String,
+        val toolCallId: String,
+        val reason: String? = null,
+        val isAutomatic: Boolean? = null,
+        val signature: String? = null,
+        val inputSchemaInput: JsonElement? = null,
     ) : ContentPart
 
     @Serializable
+    @SerialName("image")
     data class Image(
         val image: List<Int> = emptyList(),
-        @SerialName("media_type") val mediaType: String = "",
-        @SerialName("provider_options") val providerOptions: JsonElement? = null,
+        val mediaType: String = "",
+        val providerOptions: ProviderOptions? = null,
     ) : ContentPart
 
     @Serializable
+    @SerialName("file")
     data class File(
         val data: List<Int> = emptyList(),
-        @SerialName("media_type") val mediaType: String = "",
+        val mediaType: String = "",
         val filename: String? = null,
-        @SerialName("provider_options") val providerOptions: JsonElement? = null,
+        val providerOptions: ProviderOptions? = null,
     ) : ContentPart
 
     @Serializable
+    @SerialName("file-base64")
     data class FileBase64(
         val data: String = "",
-        @SerialName("media_type") val mediaType: String = "",
+        val mediaType: String = "",
         val filename: String? = null,
-        @SerialName("provider_options") val providerOptions: JsonElement? = null,
+        val providerOptions: ProviderOptions? = null,
     ) : ContentPart
 
     @Serializable
+    @SerialName("file-url")
     data class FileUrl(
         val url: String = "",
-        @SerialName("media_type") val mediaType: String = "",
-        @SerialName("provider_options") val providerOptions: JsonElement? = null,
+        val mediaType: String = "",
+        val providerOptions: ProviderOptions? = null,
     ) : ContentPart
 
     @Serializable
+    @SerialName("file-reference")
     data class FileReference(
-        @SerialName("media_type") val mediaType: String = "",
+        val mediaType: String = "",
         val reference: JsonElement = JsonObject(emptyMap()),
         val filename: String? = null,
-        @SerialName("provider_options") val providerOptions: JsonElement? = null,
+        val providerOptions: ProviderOptions? = null,
     ) : ContentPart
 
     @Serializable
+    @SerialName("reasoning")
     data class Reasoning(
         val text: String = "",
         val signature: String? = null,
-        @SerialName("provider_options") val providerOptions: JsonElement? = null,
+        val providerOptions: ProviderOptions? = null,
     ) : ContentPart
 
     @Serializable
+    @SerialName("tool-call")
     data class ToolCall(
-        @SerialName("tool_call_id") val toolCallId: String = "",
-        @SerialName("tool_name") val toolName: String = "",
+        val toolCallId: String = "",
+        val toolName: String = "",
         val input: JsonElement = JsonObject(emptyMap()),
-        @SerialName("provider_options") val providerOptions: JsonElement? = null,
-        @SerialName("provider_executed") val providerExecuted: Boolean? = null,
+        val providerExecuted: Boolean? = null,
+        val providerOptions: ProviderOptions? = null,
     ) : ContentPart
 
     @Serializable
+    @SerialName("tool-result")
     data class ToolResult(
-        @SerialName("tool_call_id") val toolCallId: String = "",
-        val result: JsonElement = JsonObject(emptyMap()),
-        @SerialName("tool_name") val toolName: String? = null,
-        @SerialName("is_error") val isError: Boolean? = null,
-        @SerialName("preliminary") val preliminary: Boolean? = null,
-        @SerialName("dynamic") val dynamic: Boolean? = null,
-        @SerialName("provider_options") val providerOptions: JsonElement? = null,
+        val toolCallId: String,
+        val toolName: String,
+        val output: ToolResultOutput,
+        val providerOptions: ProviderOptions? = null,
     ) : ContentPart
 }
 
-/**
- * Custom (de)serializer for [ContentPart].
- *
- * The wire format is internally tagged: each part is a JSON object whose
- * `"type"` key selects the variant. On decode the `"type"` key is read and the
- * matching variant's generated serializer decodes the object (the `"type"` key
- * is ignored thanks to `ignoreUnknownKeys`). On encode the variant is encoded
- * and the `"type"` key is re-injected.
- */
-object ContentPartSerializer : KSerializer<ContentPart> {
-    override val descriptor: SerialDescriptor =
-        buildClassSerialDescriptor("aimux.ContentPart")
-
-    override fun deserialize(decoder: Decoder): ContentPart {
-        val json = decoder as? JsonDecoder
-            ?: throw SerializationException("ContentPart can only be decoded from JSON")
-        val element = json.decodeJsonElement()
-        val obj = element.jsonObject
-        val type = obj["type"]?.jsonPrimitive?.content
-            ?: throw SerializationException("ContentPart is missing the 'type' discriminator: $element")
-        val ctx = json.json
-        return when (type) {
-            "custom" -> ctx.decodeFromJsonElement(ContentPart.Custom.serializer(), obj)
-            "reasoning_file" -> ctx.decodeFromJsonElement(ContentPart.ReasoningFile.serializer(), obj)
-            "tool_approval_request" -> ctx.decodeFromJsonElement(ContentPart.ToolApprovalRequest.serializer(), obj)
-            "text" -> ctx.decodeFromJsonElement(ContentPart.Text.serializer(), obj)
-            "image" -> ctx.decodeFromJsonElement(ContentPart.Image.serializer(), obj)
-            "file" -> ctx.decodeFromJsonElement(ContentPart.File.serializer(), obj)
-            "file_base64" -> ctx.decodeFromJsonElement(ContentPart.FileBase64.serializer(), obj)
-            "file_url" -> ctx.decodeFromJsonElement(ContentPart.FileUrl.serializer(), obj)
-            "file_reference" -> ctx.decodeFromJsonElement(ContentPart.FileReference.serializer(), obj)
-            "reasoning" -> ctx.decodeFromJsonElement(ContentPart.Reasoning.serializer(), obj)
-            "tool_call" -> ctx.decodeFromJsonElement(ContentPart.ToolCall.serializer(), obj)
-            "tool_result" -> ctx.decodeFromJsonElement(ContentPart.ToolResult.serializer(), obj)
-            else -> throw SerializationException("Unknown ContentPart type: '$type'")
-        }
-    }
-
-    override fun serialize(encoder: Encoder, value: ContentPart) {
-        val json = encoder as? JsonEncoder
-            ?: throw SerializationException("ContentPart can only be encoded to JSON")
-        val ctx = json.json
-        val (typeTag, inner) = when (value) {
-            is ContentPart.Custom -> "custom" to ctx.encodeToJsonElement(ContentPart.Custom.serializer(), value)
-            is ContentPart.ReasoningFile -> "reasoning_file" to ctx.encodeToJsonElement(ContentPart.ReasoningFile.serializer(), value)
-            is ContentPart.ToolApprovalRequest -> "tool_approval_request" to ctx.encodeToJsonElement(ContentPart.ToolApprovalRequest.serializer(), value)
-            is ContentPart.Text -> "text" to ctx.encodeToJsonElement(ContentPart.Text.serializer(), value)
-            is ContentPart.Image -> "image" to ctx.encodeToJsonElement(ContentPart.Image.serializer(), value)
-            is ContentPart.File -> "file" to ctx.encodeToJsonElement(ContentPart.File.serializer(), value)
-            is ContentPart.FileBase64 -> "file_base64" to ctx.encodeToJsonElement(ContentPart.FileBase64.serializer(), value)
-            is ContentPart.FileUrl -> "file_url" to ctx.encodeToJsonElement(ContentPart.FileUrl.serializer(), value)
-            is ContentPart.FileReference -> "file_reference" to ctx.encodeToJsonElement(ContentPart.FileReference.serializer(), value)
-            is ContentPart.Reasoning -> "reasoning" to ctx.encodeToJsonElement(ContentPart.Reasoning.serializer(), value)
-            is ContentPart.ToolCall -> "tool_call" to ctx.encodeToJsonElement(ContentPart.ToolCall.serializer(), value)
-            is ContentPart.ToolResult -> "tool_result" to ctx.encodeToJsonElement(ContentPart.ToolResult.serializer(), value)
-        }
-        val merged = JsonObject(buildMap {
-            put("type", JsonPrimitive(typeTag))
-            putAll(inner as? JsonObject ?: JsonObject(emptyMap()))
-        })
-        json.encodeJsonElement(merged)
-    }
-}
-
 // ─────────────────────────────────────────────────────────────────────────────
 // ModelMessage (prompt side).
 // ─────────────────────────────────────────────────────────────────────────────
@@ -500,9 +563,8 @@ object ContentPartSerializer : KSerializer<ContentPart> {
 /**
  * Message body: either a simple string or multi-part content.
  *
- * Mirrors `MessageContent.ts`: `string | Array<ContentPart>`. Modeled as a
- * sealed union with a custom (untagged) serializer: a JSON string decodes to
- * [MessageContent.Text], a JSON array decodes to [MessageContent.Parts].
+ * Mirrors `MessageContent.ts`: `string | Array<ContentPart>` (untagged): a JSON
+ * string decodes to [MessageContent.Text], a JSON array to [MessageContent.Parts].
  */
 @Serializable(with = MessageContentSerializer::class)
 sealed interface MessageContent {
@@ -514,9 +576,6 @@ sealed interface MessageContent {
     data class Parts(val parts: List<ContentPart> = emptyList()) : MessageContent
 }
 
-/**
- * Custom (de)serializer for the untagged [MessageContent] union.
- */
 object MessageContentSerializer : KSerializer<MessageContent> {
     override val descriptor: SerialDescriptor =
         buildClassSerialDescriptor("aimux.MessageContent")
@@ -552,10 +611,8 @@ object MessageContentSerializer : KSerializer<MessageContent> {
 }
 
 /**
- * A single user-facing chat message.
- *
- * Mirrors `ModelMessage.ts`: `{ role: Role, content: MessageContent }` where
- * [MessageContent] is a string-or-parts union. Use [contentString] /
+ * A single user-facing chat message. Mirrors `ModelMessage.ts`:
+ * `{ role: Role, content: MessageContent }`. Use [contentString] /
  * [contentParts] for ergonomic access, or the companion factories to build one.
  */
 @Serializable
@@ -590,24 +647,39 @@ data class ModelMessage(
 // ─────────────────────────────────────────────────────────────────────────────
 
 /**
- * Per-call timeout configuration.
- *
- * Mirrors `TimeoutConfiguration.ts`. All values are milliseconds; `null`
- * disables the corresponding limit. A `total` timeout also covers retry
- * backoff and the whole streamed response.
+ * Per-call timeout configuration. Mirrors `TimeoutConfiguration.ts`. All values
+ * are milliseconds; `null` disables the corresponding limit. A `total` timeout
+ * also covers retry backoff and the whole streamed response.
  */
 @Serializable
 data class TimeoutConfiguration(
     /** Overall timeout for the entire call (including retries and, for streaming, the whole stream), in milliseconds. */
-    @SerialName("total_ms") val totalMs: Long? = null,
+    val totalMs: Long? = null,
     /** Timeout for one model operation attempt, in milliseconds. */
-    @SerialName("step_ms") val stepMs: Long? = null,
+    val stepMs: Long? = null,
     /** Timeout waiting for the first stream chunk (streaming only). */
-    @SerialName("first_chunk_ms") val firstChunkMs: Long? = null,
+    val firstChunkMs: Long? = null,
     /** Maximum idle time between consecutive stream chunks (streaming only). */
-    @SerialName("chunk_ms") val chunkMs: Long? = null,
+    val chunkMs: Long? = null,
 )
 
+/** Response format, tagged on `type`. Mirrors `ResponseFormat.ts`. */
+@Serializable
+sealed interface ResponseFormat {
+    @Serializable
+    @SerialName("text")
+    data object Text : ResponseFormat
+
+    /** JSON output; `schema` is a JSON Schema for structured output. */
+    @Serializable
+    @SerialName("json")
+    data class Json(
+        val schema: JsonElement? = null,
+        val name: String? = null,
+        val description: String? = null,
+    ) : ResponseFormat
+}
+
 /**
  * User-facing options for `generate_text` / `stream_text`.
  *
@@ -617,30 +689,29 @@ data class TimeoutConfiguration(
  */
 @Serializable
 data class GenerateTextOptions(
-    @SerialName("max_output_tokens") val maxOutputTokens: Long? = null,
+    val maxOutputTokens: Long? = null,
     val temperature: Double? = null,
-    @SerialName("stop_sequences") val stopSequences: List<String>? = null,
-    @SerialName("top_p") val topP: Double? = null,
-    @SerialName("top_k") val topK: Double? = null,
-    @SerialName("presence_penalty") val presencePenalty: Double? = null,
-    @SerialName("frequency_penalty") val frequencyPenalty: Double? = null,
-    /** Response format. Untyped ([JsonElement]) — use a `ResponseFormat` JSON object if needed. */
-    @SerialName("response_format") val responseFormat: JsonElement? = null,
+    val stopSequences: List<String>? = null,
+    val topP: Double? = null,
+    val topK: Double? = null,
+    val presencePenalty: Double? = null,
+    val frequencyPenalty: Double? = null,
+    val responseFormat: ResponseFormat? = null,
     val seed: Long? = null,
     val tools: List<Tool>? = null,
-    @SerialName("tool_choice") val toolChoice: ToolChoice? = null,
+    val toolChoice: ToolChoice? = null,
     val headers: Map<String, String>? = null,
-    @SerialName("provider_options") val providerOptions: Map<String, JsonElement>? = null,
+    val providerOptions: ProviderOptions? = null,
     val reasoning: ReasoningEffort? = null,
     val instructions: String? = null,
     /** Per-call retry count (0 = disable retries). */
-    @SerialName("max_retries") val maxRetries: Long? = null,
-    /** Emit raw provider stream chunks as `StreamPart.Raw` (debugging; OpenAI-compatible family only). */
-    @SerialName("include_raw_chunks") val includeRawChunks: Boolean? = null,
+    val maxRetries: Long? = null,
+    /** Emit raw provider stream chunks as `raw` stream parts (debugging; OpenAI-compatible family only). */
+    val includeRawChunks: Boolean? = null,
     /** Per-call timeout configuration (overall / first chunk / inter-chunk idle, in ms). */
-    @SerialName("timeout") val timeout: TimeoutConfiguration? = null,
+    val timeout: TimeoutConfiguration? = null,
     /** Session identifier (RFC-0024): groups consecutive calls into a session. */
-    @SerialName("session_id") val sessionId: String? = null,
+    val sessionId: String? = null,
     /**
      * One repair attempt per invalid tool call (RFC-0035), run on the JVM after
      * the call returns. Mirrors AI SDK `repairToolCall`.
@@ -670,205 +741,16 @@ data class GenerateTextOptions(
 )
 
 // ─────────────────────────────────────────────────────────────────────────────
-// File bytes / file data (shared V4 file types).
-//
-// `FileBytes` and `FileData` are externally-tagged enums
-// (`{"Binary": [...]}`, `{"Data": {"data": ...}}`, ...). Each is modeled with a
-// custom serializer that dispatches on the single tag key.
-// ─────────────────────────────────────────────────────────────────────────────
-
-/**
- * Either raw bytes or a base64-encoded string.
- *
- * Mirrors `FileBytes.ts`: `{"Binary": Array<number>} | {"Base64": string}`.
- * The binary payload is a list of byte ints (serde's default `Vec<u8>` encoding).
- */
-@Serializable(with = FileBytesSerializer::class)
-sealed interface FileBytes {
-
-    /** Raw binary bytes (a JSON array of 0–255 ints on the wire). */
-    data class Binary(val data: List<Int> = emptyList()) : FileBytes
-
-    /** A base64-encoded string. */
-    data class Base64(val data: String = "") : FileBytes
-}
-
-/**
- * Custom (de)serializer for [FileBytes].
- *
- * Externally tagged: `{"Binary": [..]}` / `{"Base64": "..."}`. The inner value
- * is the raw array/string (not an object), so it is read/written directly.
- */
-object FileBytesSerializer : KSerializer<FileBytes> {
-    override val descriptor: SerialDescriptor =
-        buildClassSerialDescriptor("aimux.FileBytes")
-
-    override fun deserialize(decoder: Decoder): FileBytes {
-        val json = decoder as? JsonDecoder
-            ?: throw SerializationException("FileBytes can only be decoded from JSON")
-        val element = json.decodeJsonElement()
-        val obj = element.jsonObject
-        require(obj.size == 1) {
-            "FileBytes must be a single-key externally-tagged object, got: $element"
-        }
-        val (tag, inner) = obj.entries.single()
-        return when (tag) {
-            "Binary" -> {
-                val arr = (inner as? JsonArray) ?: JsonArray(emptyList())
-                FileBytes.Binary(arr.map { it.jsonPrimitive.content.toInt() })
-            }
-            "Base64" -> FileBytes.Base64(inner.jsonPrimitive.content)
-            else -> throw SerializationException("Unknown FileBytes tag: '$tag'")
-        }
-    }
-
-    override fun serialize(encoder: Encoder, value: FileBytes) {
-        val json = encoder as? JsonEncoder
-            ?: throw SerializationException("FileBytes can only be encoded to JSON")
-        val (tag, inner) = when (value) {
-            is FileBytes.Binary -> "Binary" to JsonArray(value.data.map { JsonPrimitive(it) })
-            is FileBytes.Base64 -> "Base64" to JsonPrimitive(value.data)
-        }
-        json.encodeJsonElement(JsonObject(mapOf(tag to inner)))
-    }
-}
-
-/**
- * File data as a tagged discriminated union.
- *
- * Mirrors `FileData.ts`: `{"Data": {"data": FileBytes}} | {"Url": {"url": ...}}
- * | {"Reference": {"reference": {...}}} | {"Text": {"text": ...}}`.
- */
-@Serializable(with = FileDataSerializer::class)
-sealed interface FileData {
-
-    @Serializable
-    data class Data(val data: FileBytes = FileBytes.Base64("")) : FileData
-
-    @Serializable
-    data class Url(val url: String = "", @SerialName("original_url") val originalUrl: String? = null) : FileData
-
-    @Serializable
-    data class Reference(val reference: JsonElement = JsonObject(emptyMap())) : FileData
-
-    @Serializable
-    data class Text(val text: String = "") : FileData
-}
-
-/**
- * Custom (de)serializer for [FileData].
- *
- * Externally tagged: each variant is a single-key object whose value is the
- * variant's inner object; delegation goes to the matching variant's generated
- * serializer.
- */
-object FileDataSerializer : KSerializer<FileData> {
-    override val descriptor: SerialDescriptor =
-        buildClassSerialDescriptor("aimux.FileData")
-
-    override fun deserialize(decoder: Decoder): FileData {
-        val json = decoder as? JsonDecoder
-            ?: throw SerializationException("FileData can only be decoded from JSON")
-        val element = json.decodeJsonElement()
-        val obj = element.jsonObject
-        require(obj.size == 1) {
-            "FileData must be a single-key externally-tagged object, got: $element"
-        }
-        val (tag, inner) = obj.entries.single()
-        val innerObj = inner as? JsonObject ?: JsonObject(emptyMap())
-        val ctx = json.json
-        return when (tag) {
-            "Data" -> ctx.decodeFromJsonElement(FileData.Data.serializer(), innerObj)
-            "Url" -> ctx.decodeFromJsonElement(FileData.Url.serializer(), innerObj)
-            "Reference" -> ctx.decodeFromJsonElement(FileData.Reference.serializer(), innerObj)
-            "Text" -> ctx.decodeFromJsonElement(FileData.Text.serializer(), innerObj)
-            else -> throw SerializationException("Unknown FileData tag: '$tag'")
-        }
-    }
-
-    override fun serialize(encoder: Encoder, value: FileData) {
-        val json = encoder as? JsonEncoder
-            ?: throw SerializationException("FileData can only be encoded to JSON")
-        val ctx = json.json
-        val (tag, inner) = when (value) {
-            is FileData.Data -> "Data" to ctx.encodeToJsonElement(FileData.Data.serializer(), value)
-            is FileData.Url -> "Url" to ctx.encodeToJsonElement(FileData.Url.serializer(), value)
-            is FileData.Reference -> "Reference" to ctx.encodeToJsonElement(FileData.Reference.serializer(), value)
-            is FileData.Text -> "Text" to ctx.encodeToJsonElement(FileData.Text.serializer(), value)
-        }
-        json.encodeJsonElement(JsonObject(mapOf(tag to inner)))
-    }
-}
-
-@Serializable(with = GeneratedFileDataSerializer::class)
-sealed interface GeneratedFileData {
-
-    @Serializable
-    data class Data(val data: FileBytes = FileBytes.Base64("")) : GeneratedFileData
-
-    @Serializable
-    data class Url(val url: String = "", @SerialName("original_url") val originalUrl: String? = null) : GeneratedFileData
-
-
-}
-
-/**
- * Custom (de)serializer for [GeneratedFileData].
- *
- * Externally tagged: each variant is a single-key object whose value is the
- * variant's inner object; delegation goes to the matching variant's generated
- * serializer.
- */
-object GeneratedFileDataSerializer : KSerializer<GeneratedFileData> {
-    override val descriptor: SerialDescriptor =
-        buildClassSerialDescriptor("aimux.GeneratedFileData")
-
-    override fun deserialize(decoder: Decoder): GeneratedFileData {
-        val json = decoder as? JsonDecoder
-            ?: throw SerializationException("GeneratedFileData can only be decoded from JSON")
-        val element = json.decodeJsonElement()
-        val obj = element.jsonObject
-        require(obj.size == 1) {
-            "GeneratedFileData must be a single-key externally-tagged object, got: $element"
-        }
-        val (tag, inner) = obj.entries.single()
-        val innerObj = inner as? JsonObject ?: JsonObject(emptyMap())
-        val ctx = json.json
-        return when (tag) {
-            "Data" -> ctx.decodeFromJsonElement(GeneratedFileData.Data.serializer(), innerObj)
-            "Url" -> ctx.decodeFromJsonElement(GeneratedFileData.Url.serializer(), innerObj)
-            else -> throw SerializationException("Unknown GeneratedFileData tag: '$tag'")
-        }
-    }
-
-    override fun serialize(encoder: Encoder, value: GeneratedFileData) {
-        val json = encoder as? JsonEncoder
-            ?: throw SerializationException("GeneratedFileData can only be encoded to JSON")
-        val ctx = json.json
-        val (tag, inner) = when (value) {
-            is GeneratedFileData.Data -> "Data" to ctx.encodeToJsonElement(GeneratedFileData.Data.serializer(), value)
-            is GeneratedFileData.Url -> "Url" to ctx.encodeToJsonElement(GeneratedFileData.Url.serializer(), value)
-        }
-        json.encodeJsonElement(JsonObject(mapOf(tag to inner)))
-    }
-}
-
-// ─────────────────────────────────────────────────────────────────────────────
-// GenerateContent / GenerateResult (the `raw` field of GenerateTextResult).
-//
-// `GenerateContent` is an externally-tagged enum
-// (`{"Text": {...}}`, `{"ToolCall": {...}}`, ...). It is modeled as a sealed
-// interface with a custom serializer; unrecognized variants fall back to
-// [GenerateContent.Unknown] for forward compatibility (mirroring [StreamPart]).
+// Results.
 // ─────────────────────────────────────────────────────────────────────────────
 
-/** A URL or document used as a source for the response. */
+/** A URL or document used as a source for the response. Mirrors `Source.ts` (tagged on `sourceType`). */
 @OptIn(ExperimentalSerializationApi::class)
 @Serializable
-@JsonClassDiscriminator("source_type")
+@JsonClassDiscriminator("sourceType")
 sealed interface Source {
     val id: String
-    val providerMetadata: JsonElement?
+    val providerMetadata: ProviderMetadata?
 
     @Serializable
     @SerialName("url")
@@ -876,549 +758,368 @@ sealed interface Source {
         override val id: String,
         val url: String,
         val title: String? = null,
-        @SerialName("provider_metadata") override val providerMetadata: JsonElement? = null,
+        override val providerMetadata: ProviderMetadata? = null,
     ) : Source
 
     @Serializable
     @SerialName("document")
     data class Document(
         override val id: String,
-        @SerialName("media_type") val mediaType: String,
+        val mediaType: String,
         val title: String,
         val filename: String? = null,
-        @SerialName("provider_metadata") override val providerMetadata: JsonElement? = null,
+        override val providerMetadata: ProviderMetadata? = null,
     ) : Source
 }
 
+/** A provider warning, tagged on `type`. Mirrors `Warning.ts`. */
+@Serializable
+sealed interface Warning {
+    @Serializable
+    @SerialName("unsupported")
+    data class Unsupported(val feature: String, val details: String? = null) : Warning
+
+    @Serializable
+    @SerialName("compatibility")
+    data class Compatibility(val feature: String, val details: String? = null) : Warning
+
+    @Serializable
+    @SerialName("deprecated")
+    data class Deprecated(val setting: String, val message: String) : Warning
+
+    @Serializable
+    @SerialName("other")
+    data class Other(val message: String) : Warning
+}
+
 /**
- * A content item in the generation result.
- *
- * Mirrors `GenerateContent.ts` (externally tagged). `provider_metadata` is a
- * [JsonElement] (`ProviderMetadata = serde_json::Value`). Every field has a
- * default; the `File` variant has no `filename` (matching Rust).
+ * A content item of the provider-level result (`GenerateResult.content`),
+ * tagged on `type`. Mirrors `GenerateContent<RawToolCall, RawToolApprovalRequest,
+ * GeneratedFile>`: a tool call's `input` is the raw argument text, and an
+ * approval request carries just the ids.
  */
-@Serializable(with = GenerateContentSerializer::class)
+@Serializable
 sealed interface GenerateContent {
     @Serializable
-    data class Custom(
-        val kind: String,
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    ) : GenerateContent
+    @SerialName("text")
+    data class Text(val text: String = "", val providerMetadata: ProviderMetadata? = null) : GenerateContent
 
     @Serializable
-    data class ReasoningFile(
-        val data: GeneratedFileData,
-        @SerialName("media_type") val mediaType: String,
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    @SerialName("tool-call")
+    data class ToolCall(
+        val toolCallId: String = "",
+        val toolName: String = "",
+        val input: String = "",
+        val providerExecuted: Boolean? = null,
+        val dynamic: Boolean? = null,
+        val providerMetadata: ProviderMetadata? = null,
     ) : GenerateContent
 
     @Serializable
-    data class ToolApprovalRequest(
-        @SerialName("approval_id") val approvalId: String,
-        @SerialName("tool_call_id") val toolCallId: String,
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    @SerialName("source")
+    data class Source(
+        val sourceType: String,
+        val id: String,
+        val url: String? = null,
+        val mediaType: String? = null,
+        val title: String? = null,
+        val filename: String? = null,
+        val providerMetadata: ProviderMetadata? = null,
     ) : GenerateContent
 
+    @Serializable
+    @SerialName("reasoning")
+    data class Reasoning(val text: String = "", val providerMetadata: ProviderMetadata? = null) : GenerateContent
 
     @Serializable
-    data class Text(
-        val text: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    @SerialName("file")
+    data class File(
+        val data: GeneratedFileData = GeneratedFileData.Data(),
+        val mediaType: String = "",
+        val providerMetadata: ProviderMetadata? = null,
     ) : GenerateContent
 
     @Serializable
-    data class ToolCall(
-        @SerialName("tool_call_id") val toolCallId: String = "",
-        @SerialName("tool_name") val toolName: String = "",
-        val input: JsonElement = JsonObject(emptyMap()),
-        @SerialName("provider_executed") val providerExecuted: Boolean? = null,
-        @SerialName("dynamic") val dynamic: Boolean? = null,
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    @SerialName("reasoning-file")
+    data class ReasoningFile(
+        val data: GeneratedFileData = GeneratedFileData.Data(),
+        val mediaType: String = "",
+        val providerMetadata: ProviderMetadata? = null,
     ) : GenerateContent
 
-    data class Source(val source: ai.arcships.aimux.Source) : GenerateContent
-
     @Serializable
-    data class Reasoning(
-        val text: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    ) : GenerateContent
+    @SerialName("custom")
+    data class Custom(val kind: String, val providerMetadata: ProviderMetadata? = null) : GenerateContent
 
     @Serializable
-    data class File(
-        val data: GeneratedFileData = GeneratedFileData.Data(),
-        @SerialName("media_type") val mediaType: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    @SerialName("tool-approval-request")
+    data class ToolApprovalRequest(
+        val approvalId: String,
+        val toolCallId: String,
+        val providerMetadata: ProviderMetadata? = null,
     ) : GenerateContent
 
     @Serializable
+    @SerialName("tool-result")
     data class ToolResult(
-        @SerialName("tool_call_id") val toolCallId: String = "",
-        @SerialName("tool_name") val toolName: String = "",
+        val toolCallId: String = "",
+        val toolName: String = "",
         val result: JsonElement = JsonObject(emptyMap()),
-        @SerialName("is_error") val isError: Boolean? = null,
-        @SerialName("preliminary") val preliminary: Boolean? = null,
-        @SerialName("dynamic") val dynamic: Boolean? = null,
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+        val isError: Boolean? = null,
+        val preliminary: Boolean? = null,
+        val dynamic: Boolean? = null,
+        val providerMetadata: ProviderMetadata? = null,
     ) : GenerateContent
-
-    /** Fallback for variants introduced after this wrapper was written. */
-    data class Unknown(val tag: String, val data: JsonElement) : GenerateContent
 }
 
-/** The externally-tagged variant name for this [GenerateContent] (e.g. "Text", "ToolCall"). */
-val GenerateContent.variantTag: String
-    get() = when (this) {
-        is GenerateContent.Custom -> "Custom"
-        is GenerateContent.ReasoningFile -> "ReasoningFile"
-        is GenerateContent.ToolApprovalRequest -> "ToolApprovalRequest"
-        is GenerateContent.Text -> "Text"
-        is GenerateContent.ToolCall -> "ToolCall"
-        is GenerateContent.Source -> "Source"
-        is GenerateContent.Reasoning -> "Reasoning"
-        is GenerateContent.File -> "File"
-        is GenerateContent.ToolResult -> "ToolResult"
-        is GenerateContent.Unknown -> tag
-    }
-
-/**
- * Custom (de)serializer for [GenerateContent].
- *
- * The wire format is externally tagged: each item is a single-key JSON object
- * `{"<VariantName>": { ...inner... }}`. This serializer reads the tag, then
- * delegates to the matching variant's generated serializer. Unknown tags become
- * [GenerateContent.Unknown].
- */
-object GenerateContentSerializer : KSerializer<GenerateContent> {
-    override val descriptor: SerialDescriptor =
-        buildClassSerialDescriptor("aimux.GenerateContent")
-
-    override fun deserialize(decoder: Decoder): GenerateContent {
-        val json = decoder as? JsonDecoder
-            ?: throw SerializationException("GenerateContent can only be decoded from JSON")
-        val element = json.decodeJsonElement()
-        val obj = element.jsonObject
-        require(obj.size == 1) {
-            "GenerateContent must be a single-key externally-tagged object, got: $element"
-        }
-        val (tag, inner) = obj.entries.single()
-        val innerObj = inner as? JsonObject ?: JsonObject(emptyMap())
-        val ctx = json.json
-        return when (tag) {
-            "Custom" -> ctx.decodeFromJsonElement(GenerateContent.Custom.serializer(), innerObj)
-            "ReasoningFile" -> ctx.decodeFromJsonElement(GenerateContent.ReasoningFile.serializer(), innerObj)
-            "ToolApprovalRequest" -> ctx.decodeFromJsonElement(GenerateContent.ToolApprovalRequest.serializer(), innerObj)
-            "Text" -> ctx.decodeFromJsonElement(GenerateContent.Text.serializer(), innerObj)
-            "ToolCall" -> ctx.decodeFromJsonElement(GenerateContent.ToolCall.serializer(), innerObj)
-            "Source" -> GenerateContent.Source(ctx.decodeFromJsonElement(Source.serializer(), innerObj))
-            "Reasoning" -> ctx.decodeFromJsonElement(GenerateContent.Reasoning.serializer(), innerObj)
-            "File" -> ctx.decodeFromJsonElement(GenerateContent.File.serializer(), innerObj)
-            "ToolResult" -> ctx.decodeFromJsonElement(GenerateContent.ToolResult.serializer(), innerObj)
-            else -> GenerateContent.Unknown(tag, inner)
-        }
-    }
-
-    override fun serialize(encoder: Encoder, value: GenerateContent) {
-        val json = encoder as? JsonEncoder
-            ?: throw SerializationException("GenerateContent can only be encoded to JSON")
-        val ctx = json.json
-        val (tag, inner) = when (value) {
-            is GenerateContent.Custom -> "Custom" to ctx.encodeToJsonElement(GenerateContent.Custom.serializer(), value)
-            is GenerateContent.ReasoningFile -> "ReasoningFile" to ctx.encodeToJsonElement(GenerateContent.ReasoningFile.serializer(), value)
-            is GenerateContent.ToolApprovalRequest -> "ToolApprovalRequest" to ctx.encodeToJsonElement(GenerateContent.ToolApprovalRequest.serializer(), value)
-            is GenerateContent.Text -> "Text" to ctx.encodeToJsonElement(GenerateContent.Text.serializer(), value)
-            is GenerateContent.ToolCall -> "ToolCall" to ctx.encodeToJsonElement(GenerateContent.ToolCall.serializer(), value)
-            is GenerateContent.Source -> "Source" to ctx.encodeToJsonElement(Source.serializer(), value.source)
-            is GenerateContent.Reasoning -> "Reasoning" to ctx.encodeToJsonElement(GenerateContent.Reasoning.serializer(), value)
-            is GenerateContent.File -> "File" to ctx.encodeToJsonElement(GenerateContent.File.serializer(), value)
-            is GenerateContent.ToolResult -> "ToolResult" to ctx.encodeToJsonElement(GenerateContent.ToolResult.serializer(), value)
-            is GenerateContent.Unknown -> value.tag to value.data
-        }
-        json.encodeJsonElement(JsonObject(mapOf(tag to inner)))
-    }
-}
-
-/**
- * Result of `LanguageModel::do_generate` (non-streaming) — the raw provider
- * result surfaced via `GenerateTextResult.raw`.
- *
- * Mirrors `GenerateResult.ts`. [content] holds typed [GenerateContent] items.
- */
+/** Result of `LanguageModel::do_generate`, surfaced as `GenerateTextResult.raw`. Mirrors `GenerateResult.ts`. */
 @Serializable
 data class GenerateResult(
     val content: List<GenerateContent> = emptyList(),
-    @SerialName("finish_reason") val finishReason: FinishReason = FinishReason(),
+    val finishReason: FinishReason = FinishReason(),
     val usage: Usage = Usage(),
-    val warnings: List<JsonElement> = emptyList(),
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    val response: ResponseInfo? = null,
+    val warnings: List<Warning> = emptyList(),
+    val providerMetadata: ProviderMetadata? = null,
     val request: RequestInfo? = null,
-) {
-    /** Names of the variant tags present in [content] (e.g. "Text", "ToolCall"). */
-    val contentVariantTags: List<String>
-        get() = content.map { it.variantTag }
-
-    /** `true` if any content item carries the given externally-tagged variant. */
-    fun hasContentVariant(tag: String): Boolean =
-        content.any { it.variantTag == tag }
-}
+    val response: ResponseInfo? = null,
+)
 
 /**
- * Result of `generate_text` (user-facing).
+ * Result of `generate_text` (user-facing). Mirrors `GenerateTextResult.ts`.
  *
- * Mirrors `GenerateTextResult.ts`.
+ * `content`, `reasoning` and the tool-call items are kept as raw [JsonElement]s
+ * (their element types are generic over the call layer).
  */
 @Serializable
 data class GenerateTextResult(
+    val content: List<JsonElement>? = null,
     val text: String = "",
-    val content: List<JsonElement> = emptyList(),
-    @SerialName("tool_calls") val toolCalls: List<ToolCall> = emptyList(),
-    @SerialName("finish_reason") val finishReason: FinishReason = FinishReason(),
+    val toolCalls: List<ToolCall> = emptyList(),
+    val finishReason: FinishReason = FinishReason(),
     val usage: Usage = Usage(),
-    val warnings: List<JsonElement> = emptyList(),
+    val warnings: List<Warning> = emptyList(),
     val raw: GenerateResult = GenerateResult(),
-    // M7: top-level aggregation fields
     val reasoning: List<JsonElement> = emptyList(),
-    @SerialName("reasoning_text") val reasoningText: String = "",
+    val reasoningText: String = "",
     val sources: List<Source> = emptyList(),
-    val files: List<JsonElement> = emptyList(),
-    @SerialName("response_messages") val responseMessages: List<ModelMessage> = emptyList(),
-    // M12: raw provider-specific finish reason string.
-    @SerialName("raw_finish_reason") val rawFinishReason: String? = null,
-    // Provider-specific metadata (e.g. Anthropic cache info). Mirrored from
-    // raw.provider_metadata for top-level convenience. Weak type (JsonElement?).
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    // Response metadata (id, timestamp, model_id). Mirrored from raw.response.
-    val response: ResponseMetadata = ResponseMetadata(),
-    // Total token usage across all steps. In single-step mode (aimux's
-    // default), equals usage. Provided for AI SDK parity.
-    @SerialName("total_usage") val totalUsage: Usage = Usage(),
+    val files: List<GeneratedFile> = emptyList(),
+    val responseMessages: List<ModelMessage> = emptyList(),
+    val rawFinishReason: String? = null,
+    val providerMetadata: ProviderMetadata? = null,
+    val request: RequestInfo = RequestInfo(),
+    val response: ResponseInfo = ResponseInfo(),
+    /** Total token usage across all steps (equals [usage] in single-step mode). */
+    val totalUsage: Usage = Usage(),
 )
 
 /**
- * Result of `generate_object` (user-facing, M12). The parsed JSON object plus
- * convenience fields from the underlying `generate_text` call.
- *
- * Mirrors `GenerateObjectResult.ts`. `object` is a [JsonElement] (arbitrary
- * JSON value, weak type).
+ * Result of `generate_object` (user-facing). The parsed JSON object plus
+ * convenience fields from the underlying `generate_text` call. Mirrors
+ * `GenerateObjectResult.ts`; `object` is an arbitrary JSON value.
  */
 @Serializable
 data class GenerateObjectResult(
-    // `object` is an arbitrary JSON value — weak type (JsonElement).
     val `object`: JsonElement,
-    @SerialName("finish_reason") val finishReason: FinishReason = FinishReason(),
-    @SerialName("raw_finish_reason") val rawFinishReason: String? = null,
+    val finishReason: FinishReason = FinishReason(),
+    val rawFinishReason: String? = null,
     val usage: Usage = Usage(),
-    val warnings: List<JsonElement> = emptyList(),
-    // Concatenated reasoning text (if the model produced reasoning/thinking).
+    val warnings: List<Warning> = emptyList(),
+    /** Concatenated reasoning text, if the model produced reasoning. */
     val reasoning: String? = null,
-    // Provider-specific metadata (e.g. Anthropic cache info). Weak type.
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    // Response metadata (id, timestamp, model_id).
+    val providerMetadata: ProviderMetadata? = null,
     val response: ResponseMetadata = ResponseMetadata(),
     val raw: GenerateTextResult = GenerateTextResult(),
 )
 
 /**
- * Aggregated result of `stream_text().consume()` (M11). Mirrors
- * `GenerateTextResult`'s user-facing fields (without `raw`, since streaming
- * has no `GenerateResult` equivalent).
- *
- * Mirrors `StreamTextResultAggregated.ts`. reasoning/sources/files use weak
- * types (JsonElement) — same strategy as [GenerateTextResult].
+ * Aggregated result of `stream_text().consume()`. Mirrors
+ * `StreamTextResultAggregated.ts` (`GenerateTextResult` without `raw`).
  */
 @Serializable
 data class StreamTextResultAggregated(
+    val content: List<JsonElement>? = null,
     val text: String = "",
-    val content: List<JsonElement> = emptyList(),
-    // reasoning/sources/files use weak types (JsonElement).
     val reasoning: List<JsonElement> = emptyList(),
-    @SerialName("reasoning_text") val reasoningText: String = "",
-    @SerialName("tool_calls") val toolCalls: List<ToolCall> = emptyList(),
+    val reasoningText: String = "",
+    val toolCalls: List<ToolCall> = emptyList(),
     val sources: List<Source> = emptyList(),
-    val files: List<JsonElement> = emptyList(),
-    @SerialName("finish_reason") val finishReason: FinishReason = FinishReason(),
-    @SerialName("raw_finish_reason") val rawFinishReason: String? = null,
+    val files: List<GeneratedFile> = emptyList(),
+    val finishReason: FinishReason = FinishReason(),
+    val rawFinishReason: String? = null,
     val usage: Usage = Usage(),
-    // Total token usage across all steps. In single-step mode (aimux's
-    // default), equals usage. Provided for AI SDK parity.
-    @SerialName("total_usage") val totalUsage: Usage = Usage(),
-    val warnings: List<JsonElement> = emptyList(),
-    // Provider-specific metadata from the Finish chunk. Weak type.
-    @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    // Response metadata (id, timestamp, model_id) if emitted by the stream.
-    val response: ResponseMetadata? = null,
-    @SerialName("response_messages") val responseMessages: List<ModelMessage> = emptyList(),
+    val totalUsage: Usage = Usage(),
+    val warnings: List<Warning> = emptyList(),
+    val providerMetadata: ProviderMetadata? = null,
+    val request: RequestInfo = RequestInfo(),
+    val response: ResponseInfo = ResponseInfo(),
+    val responseMessages: List<ModelMessage> = emptyList(),
 )
 
 // ─────────────────────────────────────────────────────────────────────────────
-// StreamPart (the streaming chunk type).
+// StreamPart (the streaming chunk type), tagged on `type`.
 //
-// `StreamPart` is an externally-tagged enum
-// (`{"TextDelta": {...}}`, `{"ToolCall": {...}}`, ...). It is modeled as a
-// sealed class with a custom deserializer that dispatches on the single tag
-// key; unrecognized variants fall back to [StreamPart.Unknown] for forward
-// compatibility.
+// Mirrors `TextStreamPart.ts`, the part type `aimux_stream_text` emits. An
+// unknown `type` fails decoding ([TypedModel] reports it through `onError`).
 // ─────────────────────────────────────────────────────────────────────────────
 
-@Serializable(with = StreamPartSerializer::class)
+@Serializable
 sealed interface StreamPart {
     @Serializable
-    data class Custom(
-        val kind: String,
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    ) : StreamPart
-
-    @Serializable
-    data class ReasoningFile(
-        val file: GenerateContent.File,
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    ) : StreamPart
-
-    @Serializable
-    data class ToolApprovalRequest(
-        @SerialName("approval_id") val approvalId: String,
-        @SerialName("tool_call") val toolCall: ai.arcships.aimux.ToolCall,
-        val reason: String? = null,
-        @SerialName("is_automatic") val isAutomatic: Boolean? = null,
-        val signature: String? = null,
-    ) : StreamPart
-
-
-    @Serializable
-    data class TextStart(
-        val id: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    ) : StreamPart
+    @SerialName("text-start")
+    data class TextStart(val id: String = "", val providerMetadata: ProviderMetadata? = null) : StreamPart
 
     @Serializable
+    @SerialName("text-delta")
     data class TextDelta(
         val id: String = "",
         val delta: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+        val providerMetadata: ProviderMetadata? = null,
     ) : StreamPart
 
     @Serializable
-    data class TextEnd(
-        val id: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    ) : StreamPart
+    @SerialName("text-end")
+    data class TextEnd(val id: String = "", val providerMetadata: ProviderMetadata? = null) : StreamPart
 
     @Serializable
-    data class StreamStart(val warnings: List<JsonElement> = emptyList()) : StreamPart
+    @SerialName("stream-start")
+    data class StreamStart(val warnings: List<Warning> = emptyList()) : StreamPart
 
     @Serializable
+    @SerialName("finish")
     data class Finish(
-        @SerialName("finish_reason") val finishReason: FinishReason = FinishReason(),
+        val finishReason: FinishReason = FinishReason(),
+        val usage: Usage = Usage(),
+        val providerMetadata: ProviderMetadata? = null,
+    ) : StreamPart
+
+    @Serializable
+    @SerialName("finish-step")
+    data class FinishStep(
+        val finishReason: FinishReason = FinishReason(),
         val usage: Usage = Usage(),
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+        val providerMetadata: ProviderMetadata? = null,
+        val response: ResponseInfo = ResponseInfo(),
     ) : StreamPart
 
+    /** `error` is the `name`-tagged `AiMuxError` JSON. */
+    @Serializable
+    @SerialName("error")
+    data class Error(val error: JsonElement = JsonObject(emptyMap())) : StreamPart
+
     @Serializable
+    @SerialName("tool-input-start")
     data class ToolInputStart(
         val id: String = "",
-        @SerialName("tool_name") val toolName: String = "",
-        @SerialName("provider_executed") val providerExecuted: Boolean? = null,
-        @SerialName("dynamic") val dynamic: Boolean? = null,
+        val toolName: String = "",
+        val providerExecuted: Boolean? = null,
+        val dynamic: Boolean? = null,
         val title: String? = null,
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+        val providerMetadata: ProviderMetadata? = null,
     ) : StreamPart
 
     @Serializable
+    @SerialName("tool-input-delta")
     data class ToolInputDelta(
         val id: String = "",
         val delta: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+        val providerMetadata: ProviderMetadata? = null,
     ) : StreamPart
 
     @Serializable
-    data class ToolInputEnd(
-        val id: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    ) : StreamPart
+    @SerialName("tool-input-end")
+    data class ToolInputEnd(val id: String = "", val providerMetadata: ProviderMetadata? = null) : StreamPart
 
     /**
      * `invalid` is set by Core when the tool call stays invalid after optional
      * repair; `error` is the typed lookup, parse, schema, or repair failure.
      */
     @Serializable
+    @SerialName("tool-call")
     data class ToolCall(
-        @SerialName("tool_call_id") val toolCallId: String = "",
-        @SerialName("tool_name") val toolName: String = "",
+        val toolCallId: String = "",
+        val toolName: String = "",
         val input: JsonElement = JsonObject(emptyMap()),
-        @SerialName("provider_executed") val providerExecuted: Boolean? = null,
-        @SerialName("dynamic") val dynamic: Boolean? = null,
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+        val providerExecuted: Boolean? = null,
+        val dynamic: Boolean? = null,
+        val providerMetadata: ProviderMetadata? = null,
         val invalid: Boolean? = null,
         val error: JsonElement? = null,
     ) : StreamPart
 
     @Serializable
+    @SerialName("tool-result")
     data class ToolResult(
-        @SerialName("tool_call_id") val toolCallId: String = "",
-        @SerialName("tool_name") val toolName: String = "",
+        val toolCallId: String = "",
+        val toolName: String = "",
         val result: JsonElement = JsonObject(emptyMap()),
-        @SerialName("is_error") val isError: Boolean? = null,
-        @SerialName("preliminary") val preliminary: Boolean? = null,
-        @SerialName("dynamic") val dynamic: Boolean? = null,
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+        val isError: Boolean? = null,
+        val preliminary: Boolean? = null,
+        val dynamic: Boolean? = null,
+        val providerMetadata: ProviderMetadata? = null,
     ) : StreamPart
 
-    // ── P2: file ──
     /** A file generated by the model (e.g. an image or document). */
     @Serializable
+    @SerialName("file")
     data class File(
-        val data: JsonElement = JsonObject(emptyMap()),
-        @SerialName("media_type") val mediaType: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+        val data: GeneratedFileData = GeneratedFileData.Data(),
+        val mediaType: String = "",
+        val providerMetadata: ProviderMetadata? = null,
     ) : StreamPart
 
     @Serializable
-    data class ReasoningStart(
-        val id: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    @SerialName("reasoning-file")
+    data class ReasoningFile(
+        val file: GeneratedFile = GeneratedFile(),
+        val providerMetadata: ProviderMetadata? = null,
     ) : StreamPart
 
     @Serializable
-    data class ReasoningDelta(
-        val id: String = "",
-        val delta: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
-    ) : StreamPart
+    @SerialName("custom")
+    data class Custom(val kind: String, val providerMetadata: ProviderMetadata? = null) : StreamPart
 
     @Serializable
-    data class ReasoningEnd(
-        val id: String = "",
-        @SerialName("provider_metadata") val providerMetadata: JsonElement? = null,
+    @SerialName("tool-approval-request")
+    data class ToolApprovalRequest(
+        val approvalId: String,
+        val toolCall: ai.arcships.aimux.ToolCall,
+        val reason: String? = null,
+        val isAutomatic: Boolean? = null,
+        val signature: String? = null,
     ) : StreamPart
 
     @Serializable
-    data class ResponseMetadata(
-        val id: String? = null,
-        val timestamp: String? = null,
-        @SerialName("model_id") val modelId: String? = null,
-    ) : StreamPart
-
-    data class Source(val source: ai.arcships.aimux.Source) : StreamPart
+    @SerialName("reasoning-start")
+    data class ReasoningStart(val id: String = "", val providerMetadata: ProviderMetadata? = null) : StreamPart
 
     @Serializable
-    data class Raw(
-        @SerialName("raw_value") val rawValue: JsonElement = JsonObject(emptyMap()),
+    @SerialName("reasoning-delta")
+    data class ReasoningDelta(
+        val id: String = "",
+        val delta: String = "",
+        val providerMetadata: ProviderMetadata? = null,
     ) : StreamPart
 
     @Serializable
-    data class Error(val error: JsonElement = JsonObject(emptyMap())) : StreamPart
-
-    /** Fallback for variants introduced after this wrapper was written. */
-    data class Unknown(val tag: String, val data: JsonElement) : StreamPart
-}
-
-/** The externally-tagged variant name for this [StreamPart] (e.g. "TextDelta", "ToolCall"). */
-val StreamPart.variantTag: String
-    get() = when (this) {
-        is StreamPart.Custom -> "Custom"
-        is StreamPart.ReasoningFile -> "ReasoningFile"
-        is StreamPart.ToolApprovalRequest -> "ToolApprovalRequest"
-        is StreamPart.TextStart -> "TextStart"
-        is StreamPart.TextDelta -> "TextDelta"
-        is StreamPart.TextEnd -> "TextEnd"
-        is StreamPart.StreamStart -> "StreamStart"
-        is StreamPart.Finish -> "Finish"
-        is StreamPart.ToolInputStart -> "ToolInputStart"
-        is StreamPart.ToolInputDelta -> "ToolInputDelta"
-        is StreamPart.ToolInputEnd -> "ToolInputEnd"
-        is StreamPart.ToolCall -> "ToolCall"
-        is StreamPart.ToolResult -> "ToolResult"
-        is StreamPart.File -> "File"
-        is StreamPart.ReasoningStart -> "ReasoningStart"
-        is StreamPart.ReasoningDelta -> "ReasoningDelta"
-        is StreamPart.ReasoningEnd -> "ReasoningEnd"
-        is StreamPart.ResponseMetadata -> "ResponseMetadata"
-        is StreamPart.Source -> "Source"
-        is StreamPart.Raw -> "Raw"
-        is StreamPart.Error -> "Error"
-        is StreamPart.Unknown -> tag
-    }
-
-/**
- * Custom (de)serializer for [StreamPart].
- *
- * The wire format is externally tagged: each part is a single-key JSON object
- * `{"<VariantName>": { ...inner... }}`. This serializer reads the tag, then
- * delegates to the matching variant's generated serializer. Unknown tags become
- * [StreamPart.Unknown].
- */
-object StreamPartSerializer : KSerializer<StreamPart> {
-    override val descriptor: SerialDescriptor =
-        buildClassSerialDescriptor("aimux.StreamPart")
+    @SerialName("reasoning-end")
+    data class ReasoningEnd(val id: String = "", val providerMetadata: ProviderMetadata? = null) : StreamPart
 
-    override fun deserialize(decoder: Decoder): StreamPart {
-        val json = decoder as? JsonDecoder
-            ?: throw SerializationException("StreamPart can only be decoded from JSON")
-        val element = json.decodeJsonElement()
-        val obj = element.jsonObject
-        require(obj.size == 1) {
-            "StreamPart must be a single-key externally-tagged object, got: $element"
-        }
-        val (tag, inner) = obj.entries.single()
-        val innerObj = inner as? JsonObject ?: JsonObject(emptyMap())
-        val ctx = json.json
-        return when (tag) {
-            "Custom" -> ctx.decodeFromJsonElement(StreamPart.Custom.serializer(), innerObj)
-            "ReasoningFile" -> ctx.decodeFromJsonElement(StreamPart.ReasoningFile.serializer(), innerObj)
-            "ToolApprovalRequest" -> ctx.decodeFromJsonElement(StreamPart.ToolApprovalRequest.serializer(), innerObj)
-            "TextStart" -> ctx.decodeFromJsonElement(StreamPart.TextStart.serializer(), innerObj)
-            "TextDelta" -> ctx.decodeFromJsonElement(StreamPart.TextDelta.serializer(), innerObj)
-            "TextEnd" -> ctx.decodeFromJsonElement(StreamPart.TextEnd.serializer(), innerObj)
-            "StreamStart" -> ctx.decodeFromJsonElement(StreamPart.StreamStart.serializer(), innerObj)
-            "Finish" -> ctx.decodeFromJsonElement(StreamPart.Finish.serializer(), innerObj)
-            "ToolInputStart" -> ctx.decodeFromJsonElement(StreamPart.ToolInputStart.serializer(), innerObj)
-            "ToolInputDelta" -> ctx.decodeFromJsonElement(StreamPart.ToolInputDelta.serializer(), innerObj)
-            "ToolInputEnd" -> ctx.decodeFromJsonElement(StreamPart.ToolInputEnd.serializer(), innerObj)
-            "ToolCall" -> ctx.decodeFromJsonElement(StreamPart.ToolCall.serializer(), innerObj)
-            "ToolResult" -> ctx.decodeFromJsonElement(StreamPart.ToolResult.serializer(), innerObj)
-            "File" -> ctx.decodeFromJsonElement(StreamPart.File.serializer(), innerObj)
-            "ReasoningStart" -> ctx.decodeFromJsonElement(StreamPart.ReasoningStart.serializer(), innerObj)
-            "ReasoningDelta" -> ctx.decodeFromJsonElement(StreamPart.ReasoningDelta.serializer(), innerObj)
-            "ReasoningEnd" -> ctx.decodeFromJsonElement(StreamPart.ReasoningEnd.serializer(), innerObj)
-            "ResponseMetadata" -> ctx.decodeFromJsonElement(StreamPart.ResponseMetadata.serializer(), innerObj)
-            "Source" -> StreamPart.Source(ctx.decodeFromJsonElement(Source.serializer(), innerObj))
-            "Raw" -> ctx.decodeFromJsonElement(StreamPart.Raw.serializer(), innerObj)
-            "Error" -> ctx.decodeFromJsonElement(StreamPart.Error.serializer(), innerObj)
-            else -> StreamPart.Unknown(tag, inner)
-        }
-    }
+    /** A source cited by the response (flat: `type: "source"` plus the [Source] fields). */
+    @Serializable
+    @SerialName("source")
+    data class Source(
+        val sourceType: String,
+        val id: String,
+        val url: String? = null,
+        val mediaType: String? = null,
+        val title: String? = null,
+        val filename: String? = null,
+        val providerMetadata: ProviderMetadata? = null,
+    ) : StreamPart
 
-    override fun serialize(encoder: Encoder, value: StreamPart) {
-        val json = encoder as? JsonEncoder
-            ?: throw SerializationException("StreamPart can only be encoded to JSON")
-        val ctx = json.json
-        val (tag, inner) = when (value) {
-            is StreamPart.Custom -> "Custom" to ctx.encodeToJsonElement(StreamPart.Custom.serializer(), value)
-            is StreamPart.ReasoningFile -> "ReasoningFile" to ctx.encodeToJsonElement(StreamPart.ReasoningFile.serializer(), value)
-            is StreamPart.ToolApprovalRequest -> "ToolApprovalRequest" to ctx.encodeToJsonElement(StreamPart.ToolApprovalRequest.serializer(), value)
-            is StreamPart.TextStart -> "TextStart" to ctx.encodeToJsonElement(StreamPart.TextStart.serializer(), value)
-            is StreamPart.TextDelta -> "TextDelta" to ctx.encodeToJsonElement(StreamPart.TextDelta.serializer(), value)
-            is StreamPart.TextEnd -> "TextEnd" to ctx.encodeToJsonElement(StreamPart.TextEnd.serializer(), value)
-            is StreamPart.StreamStart -> "StreamStart" to ctx.encodeToJsonElement(StreamPart.StreamStart.serializer(), value)
-            is StreamPart.Finish -> "Finish" to ctx.encodeToJsonElement(StreamPart.Finish.serializer(), value)
-            is StreamPart.ToolInputStart -> "ToolInputStart" to ctx.encodeToJsonElement(StreamPart.ToolInputStart.serializer(), value)
-            is StreamPart.ToolInputDelta -> "ToolInputDelta" to ctx.encodeToJsonElement(StreamPart.ToolInputDelta.serializer(), value)
-            is StreamPart.ToolInputEnd -> "ToolInputEnd" to ctx.encodeToJsonElement(StreamPart.ToolInputEnd.serializer(), value)
-            is StreamPart.ToolCall -> "ToolCall" to ctx.encodeToJsonElement(StreamPart.ToolCall.serializer(), value)
-            is StreamPart.ToolResult -> "ToolResult" to ctx.encodeToJsonElement(StreamPart.ToolResult.serializer(), value)
-            is StreamPart.File -> "File" to ctx.encodeToJsonElement(StreamPart.File.serializer(), value)
-            is StreamPart.ReasoningStart -> "ReasoningStart" to ctx.encodeToJsonElement(StreamPart.ReasoningStart.serializer(), value)
-            is StreamPart.ReasoningDelta -> "ReasoningDelta" to ctx.encodeToJsonElement(StreamPart.ReasoningDelta.serializer(), value)
-            is StreamPart.ReasoningEnd -> "ReasoningEnd" to ctx.encodeToJsonElement(StreamPart.ReasoningEnd.serializer(), value)
-            is StreamPart.ResponseMetadata -> "ResponseMetadata" to ctx.encodeToJsonElement(StreamPart.ResponseMetadata.serializer(), value)
-            is StreamPart.Source -> "Source" to ctx.encodeToJsonElement(Source.serializer(), value.source)
-            is StreamPart.Raw -> "Raw" to ctx.encodeToJsonElement(StreamPart.Raw.serializer(), value)
-            is StreamPart.Error -> "Error" to ctx.encodeToJsonElement(StreamPart.Error.serializer(), value)
-            is StreamPart.Unknown -> value.tag to value.data
-        }
-        json.encodeJsonElement(JsonObject(mapOf(tag to inner)))
-    }
+    @Serializable
+    @SerialName("raw")
+    data class Raw(val rawValue: JsonElement = JsonObject(emptyMap())) : StreamPart
 }
 
 // ─────────────────────────────────────────────────────────────────────────────
diff --git a/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/ErrorsTest.kt b/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/ErrorsTest.kt
index 0263ae286..2d34d74e5 100644
--- a/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/ErrorsTest.kt
+++ b/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/ErrorsTest.kt
@@ -47,22 +47,6 @@ class ErrorsTest {
             .hasMessageContaining("api_key")
     }
 
-    /** Malformed JSON text that reaches C → invariant (the public API rejects it before C). */
-    @Test
-    fun `malformed wire JSON is a C ABI failure naming the parameter`() {
-        val e = FFI.lib.aimux_register_providers("{not json")
-        assertThat(e).isNotNull
-        assertThat(FFI.lib.aimux_error_code(e)).isEqualTo(202)
-        assertThatThrownBy { throw expectAimuxError(e!!) }
-            .isInstanceOf(IllegalStateException::class.java)
-            .isNotInstanceOf(AimuxException::class.java)
-            .hasMessageContaining("config_json")
-        // The public API never lets it reach C.
-        assertThatThrownBy { Model.registerProviders("{not json") }
-            .isInstanceOf(IllegalArgumentException::class.java)
-            .hasMessageContaining("configJson")
-    }
-
     /** A C-ABI-only utility decodes through expectFfiError. */
     @Test
     fun `expectFfiError reads the message and frees`() {
@@ -237,8 +221,8 @@ class ErrorsTest {
     fun `createByCode rejects codes outside the enum`() {
         assertThatThrownBy { AimuxException.createByCode(999, "?") }
             .isInstanceOf(IllegalStateException::class.java)
-        // 20 is the first unassigned value and is rejected.
-        assertThatThrownBy { AimuxException.createByCode(20, "?") }
+        // 21 is the first unassigned value and is rejected.
+        assertThatThrownBy { AimuxException.createByCode(21, "?") }
             .isInstanceOf(IllegalStateException::class.java)
         assertThatThrownBy { AimuxException.createByCode(AIMUX_OK, "?") }
             .isInstanceOf(IllegalStateException::class.java)
diff --git a/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/MultimodalE2ETest.kt b/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/MultimodalE2ETest.kt
index ea99780d6..24423630b 100644
--- a/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/MultimodalE2ETest.kt
+++ b/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/MultimodalE2ETest.kt
@@ -78,19 +78,16 @@ class MultimodalE2ETest {
         server.responseBody = "SGVsbG8gd29ybGQ="
 
         SpeechModel.openai("sk-test", "tts-1", server.baseUrl).use { model ->
-            val opts = JSONObject().apply {
-                put("text", "Hi")
-                put("voice", "alloy")
-                put("output_format", "mp3")
-                put("provider_options", JSONObject())
-            }.toString()
-
-            val result = JSONObject(model.generate(opts))
-
-            // Wire format: {"audio":{"Binary":[<bytes>]}, ...}
-            val audio = result.getJSONObject("audio")
-            assertThat(audio.has("Binary")).isTrue()
-            assertThat(audio.getJSONArray("Binary").length()).isGreaterThan(0)
+            val opts = AimuxJson.encodeToString(
+                SpeechCallOptions.serializer(),
+                SpeechCallOptions(text = "Hi", voice = "alloy", outputFormat = "mp3"),
+            )
+
+            val result = AimuxJson.decodeFromString(SpeechResult.serializer(), model.generate(opts))
+
+            // Wire format: {"audio":[<bytes>], ...} (untagged: an array is binary).
+            val audio = result.audio as AudioData.Binary
+            assertThat(audio.value).isNotEmpty()
         }
     }
 
@@ -101,18 +98,20 @@ class MultimodalE2ETest {
         server.responseBody = """{"data":[{"b64_json":"aW1hZ2Ux"}]}"""
 
         ImageModel.openai("sk-test", "dall-e-3", server.baseUrl).use { model ->
-            val opts = JSONObject().apply {
-                put("prompt", "otter")
-                put("n", 1)
-                put("provider_options", JSONObject())
-            }.toString()
-
-            val result = JSONObject(model.generate(opts))
-
-            // Wire format: {"images":{"Base64":["aW1hZ2Ux"]}, ...}
-            val base64 = result.getJSONObject("images").getJSONArray("Base64")
-            assertThat(base64.length()).isEqualTo(1)
-            assertThat(base64.getString(0)).isEqualTo("aW1hZ2Ux")
+            // `n` and `providerOptions` are required on the wire and always encoded.
+            val opts = AimuxJson.encodeToString(
+                ImageCallOptions.serializer(),
+                ImageCallOptions(prompt = "otter", size = Size(1024, 1024), aspectRatio = AspectRatio(16, 9)),
+            )
+            assertThat(JSONObject(opts).getString("size")).isEqualTo("1024x1024")
+            assertThat(JSONObject(opts).getString("aspectRatio")).isEqualTo("16:9")
+            assertThat(JSONObject(opts).has("n")).isTrue()
+            assertThat(JSONObject(opts).has("providerOptions")).isTrue()
+
+            val result = AimuxJson.decodeFromString(ImageResult.serializer(), model.generate(opts))
+
+            // Wire format: {"images":["aW1hZ2Ux"], ...} (untagged: strings are base64).
+            assertThat(result.images).isEqualTo(ImageOutputs.Base64(listOf("aW1hZ2Ux")))
         }
     }
 
@@ -138,21 +137,21 @@ class MultimodalE2ETest {
             """{"results":[{"index":1,"relevance_score":0.95},{"index":0,"relevance_score":0.3}]}"""
 
         RerankingModel.cohere("sk-test", "rerank-v3.0", server.baseUrl).use { model ->
-            val opts = JSONObject().apply {
-                put("query", "which?")
-                put("documents", JSONObject("""{"Text":{"values":["doc1","doc2"]}}"""))
-                put("top_n", 2)
-                put("provider_options", JSONObject())
-            }.toString()
-
-            val result = JSONObject(model.rerank(opts))
-
-            // Wire format: {"ranking":[{"index":1,"relevance_score":0.95}, ...], ...}
-            val ranking = result.getJSONArray("ranking")
-            assertThat(ranking.length()).isEqualTo(2)
-            val first = ranking.getJSONObject(0)
-            assertThat(first.getInt("index")).isEqualTo(1)
-            assertThat(first.getDouble("relevance_score")).isEqualTo(0.95)
+            val opts = AimuxJson.encodeToString(
+                RerankingCallOptions.serializer(),
+                RerankingCallOptions(
+                    documents = RerankingDocuments.Text(listOf("doc1", "doc2")),
+                    query = "which?",
+                    topN = 2,
+                ),
+            )
+            assertThat(JSONObject(opts).getJSONObject("documents").getString("type")).isEqualTo("text")
+
+            val result = AimuxJson.decodeFromString(RerankingResult.serializer(), model.rerank(opts))
+
+            // Wire format: {"ranking":[{"index":1,"relevanceScore":0.95}, ...], ...}
+            assertThat(result.ranking).hasSize(2)
+            assertThat(result.ranking[0]).isEqualTo(RerankingRank(index = 1, relevanceScore = 0.95))
         }
     }
 
@@ -164,11 +163,10 @@ class MultimodalE2ETest {
             """{"results":[{"title":"Rust","url":"https://rust-lang.org","content":"Rust is..."}],"answer":"Rust is a systems language."}"""
 
         SearchModel.tavily("sk-test", server.baseUrl).use { model ->
-            val opts = JSONObject().apply {
-                put("query", "What is Rust?")
-                put("max_results", 5)
-                put("provider_options", JSONObject())
-            }.toString()
+            val opts = AimuxJson.encodeToString(
+                SearchCallOptions.serializer(),
+                SearchCallOptions(query = "What is Rust?", maxResults = 5),
+            )
 
             val result = JSONObject(model.search(opts))
 
@@ -189,8 +187,8 @@ class MultimodalE2ETest {
         Files.openai("sk-test", server.baseUrl).use { model ->
             val result = JSONObject(model.uploadFile("dGVzdA==", "application/pdf"))
 
-            // Wire format: {"provider_reference":{"openai":"file-abc"}, ...}
-            assertThat(result.getJSONObject("provider_reference").getString("openai"))
+            // Wire format: {"providerReference":{"openai":"file-abc"}, ...}
+            assertThat(result.getJSONObject("providerReference").getString("openai"))
                 .isEqualTo("file-abc")
         }
     }
@@ -209,13 +207,11 @@ class MultimodalE2ETest {
             assertThat(model).isNotNull()
         }
 
-        // Wire format: {"videos":[{"Url":{"url":"...","media_type":"..."}}], ...}
-        val parsed = JSONObject(
-            """{"videos":[{"Url":{"url":"https://example.com/v.mp4","media_type":"video/mp4"}}]}""",
+        // Wire format: {"videos":[{"type":"url","url":"...","mediaType":"..."}], ...}
+        val parsed = AimuxJson.decodeFromString(
+            VideoResult.serializer(),
+            """{"videos":[{"type":"url","url":"https://example.com/v.mp4","mediaType":"video/mp4"}]}""",
         )
-        val videos = parsed.getJSONArray("videos")
-        assertThat(videos.length()).isEqualTo(1)
-        assertThat(videos.getJSONObject(0).getJSONObject("Url").getString("url"))
-            .isEqualTo("https://example.com/v.mp4")
+        assertThat(parsed.videos).containsExactly(VideoData.Url("https://example.com/v.mp4", "video/mp4"))
     }
 }
diff --git a/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/StructuredE2ETest.kt b/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/StructuredE2ETest.kt
index 7cf05c114..680302a4b 100644
--- a/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/StructuredE2ETest.kt
+++ b/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/StructuredE2ETest.kt
@@ -15,16 +15,24 @@ import java.nio.charset.StandardCharsets
 // Structured end-to-end tests for the Kotlin/JVM binding.
 //
 // These spin up a local mock HTTP server speaking the OpenAI chat-completions
-// wire format, point an OpenAI `Model` at it via the `base_url` constructor,
+// wire format, point an OpenAI `Model` at it via the `baseUrl` constructor,
 // and assert that the binding correctly:
-//   1. parses tool_calls out of a provider response,
+//   1. parses toolCalls out of a provider response,
 //   2. forwards multi-role messages,
-//   3. forwards `tool_choice`,
+//   3. forwards `toolChoice`,
 //   4. parses tool-call stream parts from an SSE stream.
 //
 // No real network access is performed — every request hits 127.0.0.1.
 // ─────────────────────────────────────────────────────────────────────────────
 
+/**
+ * A chat-completions model pointed at [baseUrl]. `Model.openai` speaks the
+ * Responses API, so these tests reach the OpenAI-compatible chat wire through
+ * a registry provider instead.
+ */
+internal fun chatCompletionsModel(baseUrl: String): Model =
+    Model.provider("deepseek", "sk-test-fake-key", "gpt-4o", """{"baseUrl":"$baseUrl"}""")
+
 /// A tiny mock HTTP server that records the last request body and replays a
 /// preset response. Listens on a random loopback port.
 internal class MockProviderServer {
@@ -206,7 +214,7 @@ class StructuredE2ETest {
                         put("type", "function")
                         put("name", "get_weather")
                         put(
-                            "input_schema",
+                            "inputSchema",
                             JSONObject().apply {
                                 put("type", "object")
                                 put(
@@ -221,7 +229,7 @@ class StructuredE2ETest {
                 ),
             )
             if (includeToolChoice != null) {
-                put("tool_choice", includeToolChoice)
+                put("toolChoice", includeToolChoice)
             }
         }.toString()
 
@@ -231,22 +239,22 @@ class StructuredE2ETest {
     fun `generateText parses tool_calls`() {
         server.responseBody = toolCallOpenAiResponse
 
-        Model.openai("sk-test-fake-key", "gpt-4o", server.baseUrl).use { model ->
+        chatCompletionsModel(server.baseUrl).use { model ->
             val resultJson = model.generateText("\"What is the weather in Tokyo?\"", weatherToolsOpts())
             val result = JSONObject(resultJson)
 
-            // tool_calls[0].tool_name == "get_weather"
-            val toolCalls = result.getJSONArray("tool_calls")
+            // toolCalls[0].toolName == "get_weather"
+            val toolCalls = result.getJSONArray("toolCalls")
             assertThat(toolCalls.length()).isGreaterThan(0)
             val firstCall = toolCalls.getJSONObject(0)
-            assertThat(firstCall.getString("tool_name")).isEqualTo("get_weather")
-            assertThat(firstCall.getString("tool_call_id")).isEqualTo("call_abc")
+            assertThat(firstCall.getString("toolName")).isEqualTo("get_weather")
+            assertThat(firstCall.getString("toolCallId")).isEqualTo("call_abc")
             assertThat(firstCall.getJSONObject("input").getString("location")).isEqualTo("Tokyo")
 
-            // raw.content contains a "ToolCall" variant.
+            // raw.content contains a `tool-call` item.
             val rawContent = result.getJSONObject("raw").getJSONArray("content")
             val hasToolCallVariant = (0 until rawContent.length()).any {
-                rawContent.getJSONObject(it).has("ToolCall")
+                rawContent.getJSONObject(it).getString("type") == "tool-call"
             }
             assertThat(hasToolCallVariant).isTrue()
         }
@@ -269,7 +277,7 @@ class StructuredE2ETest {
             })
         }.toString()
 
-        Model.openai("sk-test-fake-key", "gpt-4o", server.baseUrl).use { model ->
+        chatCompletionsModel(server.baseUrl).use { model ->
             val resultJson = model.generateText(promptJson)
             val result = JSONObject(resultJson)
             assertThat(result.getString("text"))
@@ -289,7 +297,7 @@ class StructuredE2ETest {
     fun `tool_choice reaches provider`() {
         server.responseBody = toolCallOpenAiResponse
 
-        Model.openai("sk-test-fake-key", "gpt-4o", server.baseUrl).use { model ->
+        chatCompletionsModel(server.baseUrl).use { model ->
             model.generateText("\"What is the weather in Tokyo?\"", weatherToolsOpts(includeToolChoice = "required"))
 
             val reqBody = JSONObject(server.lastRequestBody)
@@ -404,7 +412,7 @@ class StructuredE2ETest {
             append("data: [DONE]\n\n")
         }
 
-        Model.openai("sk-test-fake-key", "gpt-4o", server.baseUrl).use { model ->
+        chatCompletionsModel(server.baseUrl).use { model ->
             // Run the blocking FFI stream call on a separate thread so the JNA
             // callbacks (which attach to the calling thread) don't deadlock with
             // tokio's block_on. The main thread collects parts via a queue.
@@ -428,18 +436,18 @@ class StructuredE2ETest {
             streamThread.join(5000)
 
             assertThat(collected).isNotEmpty()
-            // Stream must surface tool-call activity: either a complete ToolCall
-            // part or ToolInputDelta deltas (or both).
+            // Stream must surface tool-call activity: either a complete tool-call
+            // part or tool-input-delta deltas (or both).
             val hasToolPart = collected.any {
-                it.contains("\"ToolCall\"") || it.contains("\"ToolInputDelta\"")
+                it.contains("\"type\":\"tool-call\"") || it.contains("\"type\":\"tool-input-delta\"")
             }
             assertThat(hasToolPart)
-                .`as`("expected a ToolCall or ToolInputDelta stream part; got: $collected")
+                .`as`("expected a tool-call or tool-input-delta stream part; got: $collected")
                 .isTrue()
 
             // The finish part should also be present, carrying the tool-calls
             // finish reason.
-            val hasFinish = collected.any { it.contains("\"Finish\"") }
+            val hasFinish = collected.any { it.contains("\"type\":\"finish\"") }
             assertThat(hasFinish).isTrue()
         }
     }
@@ -480,23 +488,23 @@ class StructuredE2ETest {
 
         server.setResponses(toolCallOpenAiResponse, finalTextResponse)
 
-        Model.openai("sk-test-fake-key", "gpt-4o", server.baseUrl).use { model ->
+        chatCompletionsModel(server.baseUrl).use { model ->
             // 1st call: provider asks to call get_weather.
             val firstResultJson =
                 model.generateText("\"What's the weather in Tokyo?\"", weatherToolsOpts())
             val firstResult = JSONObject(firstResultJson)
-            val toolCalls = firstResult.getJSONArray("tool_calls")
+            val toolCalls = firstResult.getJSONArray("toolCalls")
             assertThat(toolCalls.length()).isGreaterThan(0)
             val firstCall = toolCalls.getJSONObject(0)
-            assertThat(firstCall.getString("tool_name")).isEqualTo("get_weather")
-            assertThat(firstCall.getString("tool_call_id")).isEqualTo("call_abc")
+            assertThat(firstCall.getString("toolName")).isEqualTo("get_weather")
+            assertThat(firstCall.getString("toolCallId")).isEqualTo("call_abc")
             assertThat(firstCall.getJSONObject("input").getString("location"))
                 .isEqualTo("Tokyo")
 
             // 2nd call: echo the full conversation back, including the
-            // assistant's tool_call and the ToolResult we synthesised.
-            // Input uses the engine's ContentPart variants (tool_call /
-            // tool_result); the engine converts these to the OpenAI wire
+            // assistant's tool call and the tool result we synthesised.
+            // Input uses the engine's ContentPart variants (tool-call /
+            // tool-result); the engine converts these to the OpenAI wire
             // format on the outbound request.
             val messages = JSONArray().apply {
                 put(JSONObject().apply {
@@ -509,9 +517,9 @@ class StructuredE2ETest {
                         "content",
                         JSONArray().put(
                             JSONObject().apply {
-                                put("type", "tool_call")
-                                put("tool_call_id", "call_abc")
-                                put("tool_name", "get_weather")
+                                put("type", "tool-call")
+                                put("toolCallId", "call_abc")
+                                put("toolName", "get_weather")
                                 put("input", JSONObject().put("location", "Tokyo"))
                             },
                         ),
@@ -523,11 +531,15 @@ class StructuredE2ETest {
                         "content",
                         JSONArray().put(
                             JSONObject().apply {
-                                put("type", "tool_result")
-                                put("tool_call_id", "call_abc")
+                                put("type", "tool-result")
+                                put("toolCallId", "call_abc")
+                                put("toolName", "get_weather")
                                 put(
                                     "output",
-                                    JSONObject().put("temperature", 22).put("condition", "sunny"),
+                                    JSONObject().put("type", "json").put(
+                                        "value",
+                                        JSONObject().put("temperature", 22).put("condition", "sunny"),
+                                    ),
                                 )
                             },
                         ),
diff --git a/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/TypedModelTest.kt b/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/TypedModelTest.kt
index 83704cc46..28357c75a 100644
--- a/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/TypedModelTest.kt
+++ b/bindings/kotlin/src/test/kotlin/ai/arcships/aimux/TypedModelTest.kt
@@ -18,7 +18,7 @@ import org.junit.jupiter.api.Test
 // (OpenAI chat-completions wire format) and assert that the typed wrapper:
 //   1. returns a [GenerateTextResult] object (no manual JSON parsing),
 //   2. surfaces `.text`, `.toolCalls[0].toolName`, and `.raw.content`,
-//   3. forwards typed `tools` / `tool_choice` onto the provider request,
+//   3. forwards typed `tools` / `toolChoice` onto the provider request,
 //   4. forwards multi-role [ModelMessage] conversations.
 //
 // No real network access — every request hits 127.0.0.1.
@@ -140,7 +140,7 @@ class TypedModelTest {
     fun `generateText returns a typed GenerateTextResult with text and raw content`() {
         server.responseBody = plainOpenAiResponse
 
-        TypedModel.openai("sk-test-fake-key", "gpt-4o", server.baseUrl).use { model ->
+        TypedModel(chatCompletionsModel(server.baseUrl), ownsModel = true).use { model ->
             val result = model.generateText("What is Rust?")
 
             // No manual parsing: the wrapper returned a typed object.
@@ -166,7 +166,7 @@ class TypedModelTest {
             toolChoice = ToolChoice.AUTO,
         )
 
-        TypedModel.openai("sk-test-fake-key", "gpt-4o", server.baseUrl).use { model ->
+        TypedModel(chatCompletionsModel(server.baseUrl), ownsModel = true).use { model ->
             val result = model.generateText("What is the weather in Tokyo?", options)
 
             // .toolCalls[0].toolName / .toolCallId / .input — all typed.
@@ -178,8 +178,8 @@ class TypedModelTest {
             assertThat(call.input.jsonObject["location"]!!.jsonPrimitive.content)
                 .isEqualTo("Tokyo")
 
-            // .raw.content carries the ToolCall variant tag.
-            assertThat(result.raw.hasContentVariant("ToolCall")).isTrue()
+            // .raw.content carries the tool-call item (input is the raw argument text).
+            assertThat(result.raw.content.any { it is GenerateContent.ToolCall }).isTrue()
         }
     }
 
@@ -192,11 +192,11 @@ class TypedModelTest {
             toolChoice = ToolChoice.REQUIRED,
         )
 
-        TypedModel.openai("sk-test-fake-key", "gpt-4o", server.baseUrl).use { model ->
+        TypedModel(chatCompletionsModel(server.baseUrl), ownsModel = true).use { model ->
             model.generateText("What is the weather in Tokyo?", options)
 
-            // The serialized options crossed the JSON boundary as the engine's
-            // snake_case shape with only the fields the caller set.
+            // The serialized options crossed the JSON boundary in the engine's
+            // camelCase shape with only the fields the caller set.
             val reqBody = JSONObject(server.lastRequestBody)
             assertThat(reqBody.getString("tool_choice")).isEqualTo("required")
             assertThat(reqBody.has("tools")).isTrue()
@@ -221,7 +221,7 @@ class TypedModelTest {
             ModelMessage.text(Role.USER, "What is Rust?"),
         )
 
-        TypedModel.openai("sk-test-fake-key", "gpt-4o", server.baseUrl).use { model ->
+        TypedModel(chatCompletionsModel(server.baseUrl), ownsModel = true).use { model ->
             val result = model.generateText(messages)
 
             assertThat(result.text).isEqualTo("Rust is a systems programming language.")
@@ -248,11 +248,9 @@ class TypedModelTest {
 // asserts the round-tripped value equals the original. They also pin down the
 // wire-format quirks the typed wrappers rely on:
 //   - [GenerateContent.File] carries no `filename` (unlike [ContentPart.File]),
-//   - [GenerateContent.ToolResult] / [ContentPart.ToolResult] use the `result`
-//     field (never `output`) and carry `tool_name` / `is_error` / `preliminary`
-//     / `dynamic`,
-//   - [GenerateContent.Unknown] preserves an unrecognized external tag (forward
-//     compatibility).
+//   - [GenerateContent.ToolResult] keeps the executed-tool `result` shape, while
+//     [ContentPart.ToolResult] carries a [ToolResultOutput] under `output`,
+//   - every union is tagged on `type` with the AI SDK's kebab-case names.
 // ─────────────────────────────────────────────────────────────────────────────
 class TypedModelRoundTripTest {
 
@@ -262,17 +260,15 @@ class TypedModelRoundTripTest {
             toolCallId = "call_1",
             toolName = "get_weather",
             input = JsonObject(mapOf("city" to JsonPrimitive("Paris"))),
-            providerMetadata = JsonObject(
-                mapOf("openai" to JsonObject(mapOf("itemId" to JsonPrimitive("item_1")))),
-            ),
+            providerMetadata = mapOf("openai" to mapOf("itemId" to JsonPrimitive("item_1"))),
         )
         val json = AimuxJson.encodeToString(ToolCall.serializer(), original)
-        assertThat(json).contains("\"provider_metadata\"")
+        assertThat(json).contains("\"providerMetadata\"")
         val decoded = AimuxJson.decodeFromString(ToolCall.serializer(), json)
         assertThat(decoded).isEqualTo(original)
     }
 
-    // ── GenerateContent (externally tagged) ──────────────────────────────
+    // ── GenerateContent (tagged on `type`) ──────────────────────────────
 
     @Test
     fun `GenerateContent Text round-trips`() {
@@ -287,12 +283,13 @@ class TypedModelRoundTripTest {
         val original = GenerateContent.ToolCall(
             toolCallId = "call_1",
             toolName = "get_weather",
-            input = JsonObject(mapOf("location" to JsonPrimitive("Tokyo"))),
+            input = """{"location":"Tokyo"}""",
             providerExecuted = false,
             dynamic = true,
             providerMetadata = null,
         )
         val json = AimuxJson.encodeToString(GenerateContent.serializer(), original)
+        assertThat(json).contains("\"type\":\"tool-call\"")
         val decoded = AimuxJson.decodeFromString(GenerateContent.serializer(), json)
         assertThat(decoded).isEqualTo(original)
     }
@@ -325,23 +322,7 @@ class TypedModelRoundTripTest {
         assertThat(decoded).isEqualTo(original)
     }
 
-    @Test
-    fun `GenerateContent Unknown preserves an unrecognized external tag`() {
-        val payload = JsonObject(mapOf("future" to JsonPrimitive("data")))
-        val original = GenerateContent.Unknown(tag = "FutureVariant", data = payload)
-        val json = AimuxJson.encodeToString(GenerateContent.serializer(), original)
-        val decoded = AimuxJson.decodeFromString(GenerateContent.serializer(), json)
-
-        // The whole value round-trips intact.
-        assertThat(decoded).isEqualTo(original)
-        // The unrecognized variant survives as Unknown and keeps its tag.
-        assertThat(decoded).isInstanceOf(GenerateContent.Unknown::class.java)
-        assertThat((decoded as GenerateContent.Unknown).tag).isEqualTo("FutureVariant")
-        assertThat(decoded.variantTag).isEqualTo("FutureVariant")
-        assertThat(json).contains("\"FutureVariant\"")
-    }
-
-    // ── ContentPart (internally tagged on `type`) ───────────────────────
+    // ── ContentPart (tagged on `type`) ───────────────────────
 
     @Test
     fun `ContentPart Text round-trips`() {
@@ -364,36 +345,30 @@ class TypedModelRoundTripTest {
         val json = AimuxJson.encodeToString(ContentPart.serializer(), original)
         val decoded = AimuxJson.decodeFromString(ContentPart.serializer(), json)
         assertThat(decoded).isEqualTo(original)
-        assertThat(json).contains("\"tool_call\"")
-        assertThat(json).contains("\"provider_executed\":true")
+        assertThat(json).contains("\"type\":\"tool-call\"")
+        assertThat(json).contains("\"providerExecuted\":true")
     }
 
     @Test
-    fun `ContentPart ToolResult round-trips with result not output`() {
+    fun `ContentPart ToolResult round-trips with a typed output`() {
         val original = ContentPart.ToolResult(
             toolCallId = "call_1",
-            result = JsonObject(mapOf("temp" to JsonPrimitive("20"))),
             toolName = "get_weather",
-            isError = true,
-            preliminary = true,
-            dynamic = false,
+            output = ToolResultOutput.JsonValue(JsonObject(mapOf("temp" to JsonPrimitive("20")))),
             providerOptions = null,
         )
         val json = AimuxJson.encodeToString(ContentPart.serializer(), original)
 
-        // `result` field, never `output`; carries the full ToolResult field set.
-        assertThat(json).contains("\"result\"").doesNotContain("\"output\"")
-        assertThat(json).contains("\"tool_name\"")
-        assertThat(json).contains("\"is_error\"")
-        assertThat(json).contains("\"preliminary\"")
-        assertThat(json).contains("\"dynamic\"")
-        assertThat(json).contains("\"tool_result\"")
+        // `output` is a tagged union; the old result / isError fields are gone.
+        assertThat(json).contains("\"type\":\"tool-result\"")
+        assertThat(json).contains("\"output\":{\"type\":\"json\",\"value\":{\"temp\":\"20\"}}")
+        assertThat(json).doesNotContain("\"result\"").doesNotContain("\"isError\"")
 
         val decoded = AimuxJson.decodeFromString(ContentPart.serializer(), json)
         assertThat(decoded).isEqualTo(original)
     }
 
-    // ── FileBytes / FileData (externally tagged) ────────────────────────
+    // ── FileBytes (untagged) / FileData (tagged on `type`) ────────────────────────
 
     @Test
     fun `FileBytes Binary round-trips`() {
@@ -401,7 +376,7 @@ class TypedModelRoundTripTest {
         val json = AimuxJson.encodeToString(FileBytes.serializer(), original)
         val decoded = AimuxJson.decodeFromString(FileBytes.serializer(), json)
         assertThat(decoded).isEqualTo(original)
-        assertThat(json).contains("\"Binary\"")
+        assertThat(json).isEqualTo("[1,2,3,255]")
     }
 
     @Test
@@ -410,7 +385,7 @@ class TypedModelRoundTripTest {
         val json = AimuxJson.encodeToString(FileBytes.serializer(), original)
         val decoded = AimuxJson.decodeFromString(FileBytes.serializer(), json)
         assertThat(decoded).isEqualTo(original)
-        assertThat(json).contains("\"Base64\"")
+        assertThat(json).isEqualTo("\"aGVsbG8=\"")
     }
 
     @Test
@@ -419,7 +394,7 @@ class TypedModelRoundTripTest {
         val json = AimuxJson.encodeToString(FileData.serializer(), original)
         val decoded = AimuxJson.decodeFromString(FileData.serializer(), json)
         assertThat(decoded).isEqualTo(original)
-        assertThat(json).contains("\"Data\"")
+        assertThat(json).contains("\"type\":\"data\"")
     }
 
     @Test
@@ -428,7 +403,7 @@ class TypedModelRoundTripTest {
         val json = AimuxJson.encodeToString(FileData.serializer(), original)
         val decoded = AimuxJson.decodeFromString(FileData.serializer(), json)
         assertThat(decoded).isEqualTo(original)
-        assertThat(json).contains("\"Url\"")
+        assertThat(json).contains("\"type\":\"url\"")
     }
 
     // ── GenerateResult (integration: mixed content variants) ───────────
@@ -441,7 +416,7 @@ class TypedModelRoundTripTest {
                 GenerateContent.ToolCall(
                     toolCallId = "call_1",
                     toolName = "get_weather",
-                    input = JsonObject(mapOf("location" to JsonPrimitive("Tokyo"))),
+                    input = """{"location":"Tokyo"}""",
                 ),
                 GenerateContent.ToolResult(
                     toolCallId = "call_1",
@@ -459,11 +434,9 @@ class TypedModelRoundTripTest {
         // The whole structure round-trips intact.
         assertThat(decoded).isEqualTo(original)
 
-        // Variant order and tags are preserved.
-        assertThat(decoded.contentVariantTags)
+        // Variant order is preserved.
+        assertThat(decoded.content.map { it::class.simpleName })
             .containsExactly("Text", "ToolCall", "ToolResult")
-        assertThat(decoded.hasContentVariant("ToolCall")).isTrue()
-        assertThat(decoded.hasContentVariant("Reasoning")).isFalse()
 
         // Spot-check the decoded content variants by type and field.
         assertThat(decoded.content[0]).isInstanceOf(GenerateContent.Text::class.java)
@@ -476,8 +449,9 @@ class TypedModelRoundTripTest {
         assertThat((decoded.content[2] as GenerateContent.ToolResult).result.toString())
             .isEqualTo("""{"temp":"20"}""")
 
-        // The external tags are present on the wire, and there is no `output`.
-        assertThat(json).contains("\"Text\"").contains("\"ToolCall\"").contains("\"ToolResult\"")
+        // The tags are present on the wire, and there is no `output`.
+        assertThat(json).contains("\"type\":\"text\"").contains("\"type\":\"tool-call\"")
+            .contains("\"type\":\"tool-result\"")
         assertThat(json).contains("\"result\"").doesNotContain("\"output\"")
     }
 }
diff --git a/bindings/node/__test__/cassette.test.ts b/bindings/node/__test__/cassette.test.ts
index bdfb52160..81dd3f395 100644
--- a/bindings/node/__test__/cassette.test.ts
+++ b/bindings/node/__test__/cassette.test.ts
@@ -27,7 +27,7 @@ test('cassette: OpenAI generate', async (t) => {
     t.is(typeof result.text, 'string', 'should have text field')
     t.true(result.text.length > 0, 'text should be non-empty')
     t.truthy(result.usage, 'should have usage')
-    t.truthy(result.finish_reason, 'should have finish_reason')
+    t.truthy(result.finishReason, 'should have finishReason')
   } finally {
     await srv.stop()
   }
@@ -46,9 +46,9 @@ test('cassette: OpenAI stream', async (t) => {
     }
     t.true(parts.length > 0, 'should receive stream parts')
 
-    const types = parts.map((p) => Object.keys(p)[0])
-    t.true(types.includes('StreamStart'), 'should have StreamStart')
-    t.true(types.includes('Finish'), 'should have Finish')
+    const types = parts.map((p) => p.type)
+    t.true(types.includes('stream-start'), 'should have stream-start')
+    t.true(types.includes('finish'), 'should have finish')
   } finally {
     await srv.stop()
   }
diff --git a/bindings/node/__test__/decision.test.ts b/bindings/node/__test__/decision.test.ts
index c742519de..21da4f085 100644
--- a/bindings/node/__test__/decision.test.ts
+++ b/bindings/node/__test__/decision.test.ts
@@ -57,8 +57,8 @@ test('missing or null usage counts preserve valid decision answers', async t =>
     usage = counts
     const result = await decide(model, options)
     t.is(Object.keys(result.answers).length, 3)
-    t.is(result.usage?.input_tokens.total, counts.input_tokens ?? null)
-    t.is(result.usage?.output_tokens.total, counts.output_tokens ?? null)
+    t.is(result.usage?.inputTokens.total ?? null, counts.input_tokens ?? null)
+    t.is(result.usage?.outputTokens.total ?? null, counts.output_tokens ?? null)
     t.deepEqual(result.usage?.raw, counts)
   }
 })
diff --git a/bindings/node/__test__/e2e.test.ts b/bindings/node/__test__/e2e.test.ts
index cdebb8ae1..56ad216a9 100644
--- a/bindings/node/__test__/e2e.test.ts
+++ b/bindings/node/__test__/e2e.test.ts
@@ -82,7 +82,7 @@ test('e2e: OpenAI generateText via mock server', async (t) => {
 
     t.is(r.text, 'Rust is a systems programming language.')
     t.truthy(r.usage)
-    t.truthy(r.finish_reason)
+    t.truthy(r.finishReason)
   } finally {
     await closeServer(server)
   }
@@ -105,8 +105,8 @@ test('e2e: OpenAI streamText via mock server', async (t) => {
     t.true(parts.length > 0)
 
     const text = parts
-      .filter((p) => p.TextDelta)
-      .map((p) => p.TextDelta.delta)
+      .filter((p) => p.type === 'text-delta')
+      .map((p) => p.delta)
       .join('')
 
     t.is(text, 'Hello world')
@@ -149,8 +149,8 @@ test('e2e: Anthropic streamText via mock server', async (t) => {
     t.true(parts.length > 0)
 
     const text = parts
-      .filter((p) => p.TextDelta)
-      .map((p) => p.TextDelta.delta)
+      .filter((p) => p.type === 'text-delta')
+      .map((p) => p.delta)
       .join('')
 
     t.is(text, 'Hello from Claude')
@@ -159,7 +159,7 @@ test('e2e: Anthropic streamText via mock server', async (t) => {
   }
 })
 
-test('e2e: OpenAI generateText with options (max_tokens, temperature)', async (t) => {
+test('e2e: OpenAI generateText with options (maxOutputTokens, temperature)', async (t) => {
   let receivedBody: any = null
   const { server, url } = await startMockServer((req, res) => {
     let body = ''
@@ -173,7 +173,7 @@ test('e2e: OpenAI generateText with options (max_tokens, temperature)', async (t
 
   try {
     const model = await deepseek('test-key', 'gpt-4o', url)
-    const opts = JSON.stringify({ max_output_tokens: 100, temperature: 0.5 })
+    const opts = JSON.stringify({ maxOutputTokens: 100, temperature: 0.5 })
     await model.generateText(JSON.stringify('Hello'), opts)
 
     // Verify the options were passed through to the HTTP request
@@ -204,7 +204,7 @@ const OPENAI_TOOL_CALL_RESPONSE = JSON.stringify({
   usage: { prompt_tokens: 20, completion_tokens: 10, total_tokens: 30 },
 })
 
-test('e2e: OpenAI generateText parses tool_calls (structured content)', async (t) => {
+test('e2e: OpenAI generateText parses toolCalls (structured content)', async (t) => {
   const { server, url } = await startMockServer((req, res) => {
     res.writeHead(200, { 'content-type': 'application/json' })
     res.end(OPENAI_TOOL_CALL_RESPONSE)
@@ -217,7 +217,7 @@ test('e2e: OpenAI generateText parses tool_calls (structured content)', async (t
         type: 'function',
         name: 'get_weather',
         description: 'Get weather for a location',
-        input_schema: {
+        inputSchema: {
           type: 'object',
           properties: { location: { type: 'string' } },
           required: ['location'],
@@ -226,22 +226,22 @@ test('e2e: OpenAI generateText parses tool_calls (structured content)', async (t
     })
     const r = JSON.parse(await model.generateText(JSON.stringify("What's the weather in Tokyo?"), opts))
 
-    // Convenience field: tool_calls extracted
-    t.is(r.tool_calls.length, 1)
-    t.is(r.tool_calls[0].tool_name, 'get_weather')
-    t.is(r.tool_calls[0].tool_call_id, 'call_abc')
-    t.deepEqual(r.tool_calls[0].input, { location: 'Tokyo' })
+    // Convenience field: toolCalls extracted
+    t.is(r.toolCalls.length, 1)
+    t.is(r.toolCalls[0].toolName, 'get_weather')
+    t.is(r.toolCalls[0].toolCallId, 'call_abc')
+    t.deepEqual(r.toolCalls[0].input, { location: 'Tokyo' })
 
-    // Structured content: raw.content contains the ToolCall variant
+    // Structured content: raw.content contains the tool-call variant
     t.truthy(r.raw)
     t.true(Array.isArray(r.raw.content))
-    const tc = r.raw.content.find((c: any) => c.ToolCall)
-    t.truthy(tc, 'raw.content contains a ToolCall variant')
-    t.is(tc.ToolCall.tool_name, 'get_weather')
-    t.is(tc.ToolCall.tool_call_id, 'call_abc')
+    const tc = r.raw.content.find((c: any) => c.type === 'tool-call')
+    t.truthy(tc, 'raw.content contains a tool-call variant')
+    t.is(tc.toolName, 'get_weather')
+    t.is(tc.toolCallId, 'call_abc')
     // raw content keeps the provider's argument text; the parsed object
     // lives on the top-level toolCalls.
-    t.is(tc.ToolCall.input, '{"location":"Tokyo"}')
+    t.is(tc.input, '{"location":"Tokyo"}')
   } finally {
     await closeServer(server)
   }
@@ -284,7 +284,7 @@ test('e2e: OpenAI generateText with multi-role messages (system + user)', async
 
 // ── ToolChoice ──────────────────────────────────────────────────────────────
 
-test('e2e: OpenAI generateText with tool_choice: required', async (t) => {
+test('e2e: OpenAI generateText with toolChoice: required', async (t) => {
   let receivedBody: any = null
   const { server, url } = await startMockServer((req, res) => {
     let body = ''
@@ -302,13 +302,13 @@ test('e2e: OpenAI generateText with tool_choice: required', async (t) => {
       tools: [{
         type: 'function',
         name: 'get_weather',
-        input_schema: { type: 'object', properties: { location: { type: 'string' } } },
+        inputSchema: { type: 'object', properties: { location: { type: 'string' } } },
       }],
-      tool_choice: 'required',
+      toolChoice: 'required',
     })
     await model.generateText(JSON.stringify('Hello'), opts)
 
-    // tool_choice reaches the provider request body as "required"
+    // toolChoice reaches the provider request body as "required"
     t.is(receivedBody.tool_choice, 'required')
   } finally {
     await closeServer(server)
@@ -336,7 +336,7 @@ test('e2e: OpenAI streamText parses tool-call stream parts', async (t) => {
       tools: [{
         type: 'function',
         name: 'get_weather',
-        input_schema: { type: 'object', properties: { location: { type: 'string' } } },
+        inputSchema: { type: 'object', properties: { location: { type: 'string' } } },
       }],
     })
     const parts: any[] = []
@@ -344,16 +344,16 @@ test('e2e: OpenAI streamText parses tool-call stream parts', async (t) => {
       parts.push(JSON.parse(json))
     }
 
-    // The stream must contain a ToolCall or ToolInputDelta part (not just TextDelta/Finish)
-    const hasToolPart = parts.some(
-      (p) => p.ToolCall || p.ToolInputDelta || p.ToolInputStart,
+    // The stream must contain a tool-call or tool-input part (not just text-delta/finish)
+    const hasToolPart = parts.some((p) =>
+      ['tool-call', 'tool-input-delta', 'tool-input-start'].includes(p.type),
     )
     t.true(hasToolPart, 'stream contained a tool-related StreamPart')
 
-    // A complete ToolCall part should carry the parsed tool name + input
-    const toolCall = parts.find((p) => p.ToolCall)
+    // A complete tool-call part should carry the parsed tool name + input
+    const toolCall = parts.find((p) => p.type === 'tool-call')
     if (toolCall) {
-      t.is(toolCall.ToolCall.tool_name, 'get_weather')
+      t.is(toolCall.toolName, 'get_weather')
     }
   } finally {
     await closeServer(server)
@@ -362,7 +362,7 @@ test('e2e: OpenAI streamText parses tool-call stream parts', async (t) => {
 
 // ── Tool-call full round-trip ───────────────────────────────────────────────
 
-test('e2e: OpenAI tool-call full round-trip (ToolCall → ToolResult → final text)', async (t) => {
+test('e2e: OpenAI tool-call full round-trip (tool-call → tool-result → final text)', async (t) => {
   // Two calls: the mock returns tool_calls on the first, final text on the second.
   let callCount = 0
   let secondRequestBody: any = null
@@ -385,7 +385,7 @@ test('e2e: OpenAI tool-call full round-trip (ToolCall → ToolResult → final t
         type: 'function',
         name: 'get_weather',
         description: 'Get weather for a location',
-        input_schema: {
+        inputSchema: {
           type: 'object',
           properties: { location: { type: 'string' } },
           required: ['location'],
@@ -395,30 +395,31 @@ test('e2e: OpenAI tool-call full round-trip (ToolCall → ToolResult → final t
 
     // Step 1: first call — model requests a tool call.
     const r1 = JSON.parse(await model.generateText(JSON.stringify("What's the weather in Tokyo?"), opts))
-    t.is(r1.tool_calls[0].tool_name, 'get_weather')
-    t.is(r1.tool_calls[0].tool_call_id, 'call_abc')
+    t.is(r1.toolCalls[0].toolName, 'get_weather')
+    t.is(r1.toolCalls[0].toolCallId, 'call_abc')
 
     // Step 2: user "executes" the tool, then builds the full conversation:
     //   user → assistant(tool_call) → tool(result)
-    // Input uses the engine's ContentPart variants (tool_call / tool_result);
+    // Input uses the engine's ContentPart variants (tool-call / tool-result);
     // the engine converts these to the OpenAI wire format on the outbound request.
     const messages = [
       { role: 'user', content: "What's the weather in Tokyo?" },
       {
         role: 'assistant',
         content: [{
-          type: 'tool_call',
-          tool_call_id: 'call_abc',
-          tool_name: 'get_weather',
+          type: 'tool-call',
+          toolCallId: 'call_abc',
+          toolName: 'get_weather',
           input: { location: 'Tokyo' },
         }],
       },
       {
         role: 'tool',
         content: [{
-          type: 'tool_result',
-          tool_call_id: 'call_abc',
-          result: { temperature: 22, condition: 'sunny' },
+          type: 'tool-result',
+          toolCallId: 'call_abc',
+          toolName: 'get_weather',
+          output: { type: 'json', value: { temperature: 22, condition: 'sunny' } },
         }],
       },
     ]
diff --git a/bindings/node/__test__/error.test.ts b/bindings/node/__test__/error.test.ts
index 12eab5d7a..a8ab022e6 100644
--- a/bindings/node/__test__/error.test.ts
+++ b/bindings/node/__test__/error.test.ts
@@ -66,7 +66,7 @@ test('stream error path yields a typed AimuxError', async (t) => {
     baseUrl: 'http://127.0.0.1:1',
   })
   const err = (await t.throwsAsync(async () => {
-    for await (const _part of streamText(model, 'hello', { max_retries: 0 })) {
+    for await (const _part of streamText(model, 'hello', { maxRetries: 0 })) {
       // never yields
     }
   })) as AimuxError
@@ -119,7 +119,7 @@ test('APICallError carries every field the response produced', async (t) => {
   try {
     const model = await native.openai('test-key', 'gpt-4o', { baseUrl: url })
     const err = (await t.throwsAsync(() =>
-      model.generateText(JSON.stringify('hi'), JSON.stringify({ max_retries: 0 })),
+      model.generateText(JSON.stringify('hi'), JSON.stringify({ maxRetries: 0 })),
     )) as APICallError
     t.true(err instanceof APICallError)
     t.true(err instanceof native.APICallError)
@@ -156,7 +156,7 @@ test('exhausted retries throw RetryError with the per-attempt history', async (t
   try {
     const model = await native.openai('test-key', 'gpt-4o', { baseUrl: url })
     const err = (await t.throwsAsync(() =>
-      model.generateText(JSON.stringify('hi'), JSON.stringify({ max_retries: 1 })),
+      model.generateText(JSON.stringify('hi'), JSON.stringify({ maxRetries: 1 })),
     )) as RetryError
     t.true(err instanceof RetryError)
     t.true(err instanceof AimuxError)
@@ -239,7 +239,7 @@ test.serial('a missing API key is a LoadAPIKeyError at call time carrying the en
   try {
     const model = await provider('groq', null, 'some-model')
     const err = (await t.throwsAsync(() =>
-      model.generateText(JSON.stringify('hi'), JSON.stringify({ max_retries: 0 })),
+      model.generateText(JSON.stringify('hi'), JSON.stringify({ maxRetries: 0 })),
     )) as LoadAPIKeyError
     t.true(err instanceof LoadAPIKeyError)
     t.true(err instanceof AimuxError)
@@ -269,7 +269,7 @@ test.serial('a missing provider setting is a LoadSettingError at call time', asy
   try {
     const model = await native.azure('test-key', '', 'some-deployment')
     const err = (await t.throwsAsync(() =>
-      model.generateText(JSON.stringify('hi'), JSON.stringify({ max_retries: 0 })),
+      model.generateText(JSON.stringify('hi'), JSON.stringify({ maxRetries: 0 })),
     )) as LoadSettingError
     t.true(err instanceof LoadSettingError)
     t.true(err instanceof AimuxError)
diff --git a/bindings/node/__test__/index.test.ts b/bindings/node/__test__/index.test.ts
index 029db1701..fbb4a7d97 100644
--- a/bindings/node/__test__/index.test.ts
+++ b/bindings/node/__test__/index.test.ts
@@ -20,7 +20,7 @@ test('native module loads and exports functions', (t) => {
 })
 
 test('video poll options are exported from the public type barrel', (t) => {
-  const poll: VideoPollOptions = { interval_ms: 1_000, timeout_ms: 120_000 }
+  const poll: VideoPollOptions = { intervalMs: 1_000, timeoutMs: 120_000 }
   const options: Pick<VideoCallOptions, 'poll'> = { poll }
   t.deepEqual(options.poll, poll)
 })
@@ -41,16 +41,20 @@ test('provider() creates a registry model instance (RFC-0017 phase 4)', async (t
   t.is(typeof model.streamText, 'function')
 })
 
-test('provider() with undefined apiKey reads the registry env var', async (t) => {
-  // ABACUS_API_KEY is unset in CI. As in the AI SDK, the key is loaded when a
-  // request is made, so construction succeeds and the call fails with a clear
-  // error (missing env key), proving the env-var path is wired.
-  delete process.env.ABACUS_API_KEY
-  const model = await provider('abacus', undefined, 'm')
-  await t.throwsAsync(
-    () => model.generateText(JSON.stringify('hi')),
-    { message: /api key/i },
-  )
+test.serial('provider() with undefined apiKey reads the registry env var', async (t) => {
+  // The key is resolved at call time: with the env var unset the call must
+  // fail with a clear error (missing env key), proving the env-var path is wired.
+  const saved = process.env.GROQ_API_KEY
+  delete process.env.GROQ_API_KEY
+  try {
+    const model = await provider('groq', undefined, 'm')
+    await t.throwsAsync(
+      () => model.generateText(JSON.stringify('hi'), JSON.stringify({ maxRetries: 0 })),
+      { message: /api key/i },
+    )
+  } finally {
+    if (saved !== undefined) process.env.GROQ_API_KEY = saved
+  }
 })
 
 test('provider() rejects unknown provider names with the available list', async (t) => {
diff --git a/bindings/node/__test__/multimodal_e2e.test.ts b/bindings/node/__test__/multimodal_e2e.test.ts
index ba780f4ed..85d51d58f 100644
--- a/bindings/node/__test__/multimodal_e2e.test.ts
+++ b/bindings/node/__test__/multimodal_e2e.test.ts
@@ -80,7 +80,7 @@ const FILES_RESPONSE = JSON.stringify({
 // fetch result). A single-response mock can't cover the full flow, so the video
 // test only exercises construction + result-JSON parsing (see test below).
 const VIDEO_RESULT_JSON = JSON.stringify({
-  videos: [{ Url: { url: 'https://example.com/v.mp4', media_type: 'video/mp4' } }],
+  videos: [{ type: 'url', url: 'https://example.com/v.mp4', mediaType: 'video/mp4' }],
 })
 
 // ── Tests ───────────────────────────────────────────────────────────────────
@@ -114,11 +114,11 @@ test('e2e: OpenAI speech (TTS) via mock server returns binary audio', async (t)
   try {
     const speaker = await openaiSpeech('test-key', 'tts-1', url)
     const r = JSON.parse(
-      await speaker.generate(JSON.stringify({ text: 'Hi', voice: 'alloy', output_format: 'mp3' })),
+      await speaker.generate(JSON.stringify({ text: 'Hi', voice: 'alloy', outputFormat: 'mp3' })),
     )
 
-    t.truthy(r.audio.Binary, 'audio comes back as AudioData::Binary')
-    t.true(r.audio.Binary.length > 0, 'binary audio is non-empty')
+    t.true(Array.isArray(r.audio), 'audio comes back as AudioData::Binary (a byte array)')
+    t.true(r.audio.length > 0, 'binary audio is non-empty')
   } finally {
     await closeServer(server)
   }
@@ -132,14 +132,14 @@ test('e2e: OpenAI image generation via mock server', async (t) => {
 
   try {
     const imager = await openaiImage('test-key', 'dall-e-3', url)
-    // ImageCallOptions has non-Option required fields `n` and `provider_options`,
+    // ImageCallOptions has non-Option required fields `n` and `providerOptions`,
     // so both must be present in the opts JSON for serde deserialization.
     const r = JSON.parse(
-      await imager.generate(JSON.stringify({ prompt: 'otter', n: 1, provider_options: {} })),
+      await imager.generate(JSON.stringify({ prompt: 'otter', n: 1, providerOptions: {} })),
     )
 
-    t.is(r.images.Base64.length, 1)
-    t.is(r.images.Base64[0], 'aW1hZ2Ux')
+    t.is(r.images.length, 1)
+    t.is(r.images[0], 'aW1hZ2Ux')
   } finally {
     await closeServer(server)
   }
@@ -169,14 +169,14 @@ test('e2e: Cohere reranking via mock server', async (t) => {
 
   try {
     const reranker = await cohereReranking('test-key', 'rerank-v3.0', url)
-    // `docs_json` deserializes into the externally-tagged `RerankingDocuments`
-    // enum, so it must be `{"Text":{"values":[...]}}` (not a bare array).
+    // `docs_json` deserializes into the `type`-tagged `RerankingDocuments`
+    // enum, so it must be `{"type":"text","values":[...]}` (not a bare array).
     const r = JSON.parse(
-      await reranker.rerank('which?', JSON.stringify({ Text: { values: ['doc1', 'doc2'] } })),
+      await reranker.rerank('which?', JSON.stringify({ type: 'text', values: ['doc1', 'doc2'] })),
     )
 
     t.is(r.ranking.length, 2)
-    t.is(r.ranking[0].relevance_score, 0.95)
+    t.is(r.ranking[0].relevanceScore, 0.95)
   } finally {
     await closeServer(server)
   }
@@ -210,7 +210,7 @@ test('e2e: OpenAI files upload via mock server', async (t) => {
     const files = await openaiFiles('test-key', url)
     const r = JSON.parse(await files.uploadFile('dGVzdA==', 'application/pdf'))
 
-    t.is(r.provider_reference.openai, 'file-abc')
+    t.is(r.providerReference.openai, 'file-abc')
   } finally {
     await closeServer(server)
   }
@@ -225,6 +225,6 @@ test('e2e: Google video construction + result parsing', async (t) => {
   t.truthy(video, 'googleVideo factory constructs without throwing')
 
   const r = JSON.parse(VIDEO_RESULT_JSON)
-  t.is(r.videos[0].Url.url, 'https://example.com/v.mp4')
-  t.is(r.videos[0].Url.media_type, 'video/mp4')
+  t.is(r.videos[0].url, 'https://example.com/v.mp4')
+  t.is(r.videos[0].mediaType, 'video/mp4')
 })
diff --git a/bindings/node/__test__/reasoning_replay.test.ts b/bindings/node/__test__/reasoning_replay.test.ts
index b9b86dff3..2f64b8abb 100644
--- a/bindings/node/__test__/reasoning_replay.test.ts
+++ b/bindings/node/__test__/reasoning_replay.test.ts
@@ -2,15 +2,14 @@
 // reported against aimux 0.1.1 when driving OpenAI-compatible thinking models
 // (e.g. DeepSeek `deepseek-v4-flash`) in multi-turn tool-call conversations.
 //
-// Issue 1: a `tool`-role message carrying `ContentPart[]` with a `tool_result`
-//   part built from the legacy `output` field was rejected by `ModelPrompt`
-//   deserialization ("data did not match any variant of untagged enum
-//   ModelPrompt"). The Rust `ContentPart::ToolResult.result` field now accepts
-//   `output` as a serde alias, so both shapes round-trip.
+// Issue 1: a `tool`-role message carrying `ContentPart[]` with a `tool-result`
+//   part (the AI SDK `ToolResultPart` shape, with a structured `output`) was
+//   rejected by `ModelPrompt` deserialization ("data did not match any variant
+//   of untagged enum ModelPrompt").
 //
 // Issue 2: thinking models require prior assistant `reasoning_content` to be
 //   replayed on later turns. The OpenAI message converter now lifts
-//   `ContentPart::Reasoning` parts to a top-level `reasoning_content` string.
+//   `reasoning` content parts to a top-level `reasoning_content` string.
 //
 // These tests use a mock HTTP server (no real API) and assert the outbound
 // request body shape that the Rust engine produces.
@@ -52,9 +51,9 @@ const CHAT_RESPONSE = JSON.stringify({
   usage: { prompt_tokens: 10, completion_tokens: 1, total_tokens: 11 },
 })
 
-// ── Issue 1: tool role ContentPart[] with the legacy `output` field ─────────
+// ── Issue 1: tool role ContentPart[] with a structured `output` ─────────────
 
-test('issue 1: tool_result with legacy `output` field is accepted and carries tool_call_id', async (t) => {
+test('issue 1: tool-result with a structured `output` is accepted and carries tool_call_id', async (t) => {
   let requestBody: any = null
   const { server, url } = await startMockServer(async (req, res) => {
     requestBody = await readBody(req)
@@ -65,25 +64,25 @@ test('issue 1: tool_result with legacy `output` field is accepted and carries to
   try {
     const model = await deepseek('test-key', 'deepseek-v4-flash', url)
 
-    // The exact shape from the user's failing case: a tool message whose
-    // ContentPart uses the legacy `output` field (Vercel AI SDK / 0.1.0 TS).
+    // A tool message whose ContentPart carries the AI SDK `ToolResultPart` shape.
     const messages = [
       { role: 'user', content: 'write a file' },
       {
         role: 'assistant',
         content: [{
-          type: 'tool_call',
-          tool_call_id: 'tc1',
-          tool_name: 'write_file',
+          type: 'tool-call',
+          toolCallId: 'tc1',
+          toolName: 'write_file',
           input: { path: '/tmp/test.txt' },
         }],
       },
       {
         role: 'tool',
         content: [{
-          type: 'tool_result',
-          tool_call_id: 'tc1',
-          output: 'Successfully wrote to /tmp/test.txt',
+          type: 'tool-result',
+          toolCallId: 'tc1',
+          toolName: 'write_file',
+          output: { type: 'text', value: 'Successfully wrote to /tmp/test.txt' },
         }],
       },
     ]
@@ -105,7 +104,7 @@ test('issue 1: tool_result with legacy `output` field is accepted and carries to
 
 // ── Issue 2: reasoning_content replay on the request side ───────────────────
 
-test('issue 2: assistant reasoning + tool_call lifts reasoning to reasoning_content', async (t) => {
+test('issue 2: assistant reasoning + tool-call lifts reasoning to reasoning_content', async (t) => {
   let requestBody: any = null
   const { server, url } = await startMockServer(async (req, res) => {
     requestBody = await readBody(req)
@@ -117,7 +116,7 @@ test('issue 2: assistant reasoning + tool_call lifts reasoning to reasoning_cont
     const model = await deepseek('test-key', 'deepseek-v4-flash', url)
 
     // Reproduces the user's table row: assistant
-    // [{reasoning},{tool_call}] + tool result. The first turn produced a
+    // [{reasoning},{tool-call}] + tool result. The first turn produced a
     // reasoning_content + tool_call; the second turn must replay both.
     // A tool message carries tool_result parts only (AI SDK `ToolModelMessage`).
     const messages = [
@@ -125,17 +124,17 @@ test('issue 2: assistant reasoning + tool_call lifts reasoning to reasoning_cont
       {
         role: 'assistant',
         content: [
-          { type: 'reasoning', text: 'I need to inspect files before answering.', signature: null },
-          { type: 'tool_call', tool_call_id: 'call_1', tool_name: 'read_file', input: { path: 'README.md' } },
+          { type: 'reasoning', text: 'I need to inspect files before answering.' },
+          { type: 'tool-call', toolCallId: 'call_1', toolName: 'read_file', input: { path: 'README.md' } },
         ],
       },
       {
         role: 'tool',
         content: [{
-          type: 'tool_result',
-          tool_call_id: 'call_1',
-          tool_name: 'read_file',
-          output: 'contents of README.md',
+          type: 'tool-result',
+          toolCallId: 'call_1',
+          toolName: 'read_file',
+          output: { type: 'text', value: 'contents of README.md' },
         }],
       },
     ]
@@ -173,7 +172,7 @@ test('issue 2: assistant reasoning + text (no tool calls) emits reasoning_conten
       {
         role: 'assistant',
         content: [
-          { type: 'reasoning', text: '2 plus 2 equals 4.', signature: null },
+          { type: 'reasoning', text: '2 plus 2 equals 4.' },
           { type: 'text', text: '4' },
         ],
       },
@@ -208,9 +207,9 @@ test('issue 2: assistant without reasoning omits reasoning_content', async (t) =
       {
         role: 'assistant',
         content: [{
-          type: 'tool_call',
-          tool_call_id: 'call_1',
-          tool_name: 'write_file',
+          type: 'tool-call',
+          toolCallId: 'call_1',
+          toolName: 'write_file',
           input: { path: '/tmp/x' },
         }],
       },
diff --git a/bindings/node/__test__/roundtrip-fields.test.ts b/bindings/node/__test__/roundtrip-fields.test.ts
index 66b8924cd..d66f98cc6 100644
--- a/bindings/node/__test__/roundtrip-fields.test.ts
+++ b/bindings/node/__test__/roundtrip-fields.test.ts
@@ -6,7 +6,7 @@
 //     `signature` is echoed as a thinking block on the wire (input
 //     round-trip), and a thinking-block response surfaces `signature` on
 //     the result (response visibility, #131).
-//   - google: a tool_result's `toolName` reaches the request as
+//   - google: a tool-result's `toolName` reaches the request as
 //     functionResponse.name (#127).
 
 import test from 'ava'
@@ -69,14 +69,14 @@ test('roundtrip: anthropic echoes prompt reasoning signature and surfaces respon
       {
         role: 'assistant',
         content: [
-          { type: 'reasoning', text: 'pondering', provider_options: { anthropic: { signature: 'sig-prompt-1' } } },
+          { type: 'reasoning', text: 'pondering', providerOptions: { anthropic: { signature: 'sig-prompt-1' } } },
           { type: 'text', text: 'answer' },
         ],
       },
       { role: 'user', content: [{ type: 'text', text: 'go on' }] },
     ]
     const opts = JSON.stringify({
-      provider_options: {
+      providerOptions: {
         anthropic: { thinking: { type: 'enabled', budgetTokens: 1024 } },
       },
     })
@@ -120,12 +120,12 @@ test('roundtrip: google toolName reaches functionResponse.name', async (t) => {
       {
         role: 'assistant',
         content: [
-          { type: 'tool_call', tool_call_id: 'call-1', tool_name: 'weather', input: { location: 'SF' } },
+          { type: 'tool-call', toolCallId: 'call-1', toolName: 'weather', input: { location: 'SF' } },
         ],
       },
       {
         role: 'tool',
-        content: [{ type: 'tool_result', tool_call_id: 'call-1', tool_name: 'weather', result: { temp: 70 } }],
+        content: [{ type: 'tool-result', toolCallId: 'call-1', toolName: 'weather', output: { type: 'json', value: { temp: 70 } } }],
       },
     ]
     await model.generateText(JSON.stringify(prompt))
diff --git a/bindings/node/__test__/session.test.ts b/bindings/node/__test__/session.test.ts
index 1cc0b8a45..b01304eab 100644
--- a/bindings/node/__test__/session.test.ts
+++ b/bindings/node/__test__/session.test.ts
@@ -57,29 +57,29 @@ test.serial('sessionId groups calls and query API returns typed results', async
   const model = await deepseek('sk-test-fake-key', 'gpt-4o-mini', { baseUrl: url })
 
   // Two calls in the same session.
-  await generateText(model, 'first', { session_id: 'sess-1' })
-  await generateText(model, 'second', { session_id: 'sess-1' })
+  await generateText(model, 'first', { sessionId: 'sess-1' })
+  await generateText(model, 'second', { sessionId: 'sess-1' })
 
   const calls: SessionCall[] = getSessionCalls('sess-1')
   t.is(calls.length, 2)
   t.is(calls[0].step, 0)
   t.is(calls[1].step, 1)
-  t.truthy(calls[0].call_id)
-  t.truthy(calls[0].recorded_at)
-  t.not(calls[0].call_id, calls[1].call_id)
+  t.truthy(calls[0].callId)
+  t.truthy(calls[0].recordedAt)
+  t.not(calls[0].callId, calls[1].callId)
 
   // A call without a sessionId (inference off) is not grouped.
   await generateText(model, 'third')
   const sessions: SessionView[] = getSessions()
   t.is(sessions.length, 1)
-  t.is(sessions[0].session_id, 'sess-1')
+  t.is(sessions[0].sessionId, 'sess-1')
   t.is(sessions[0].source, 'Explicit')
 
   // Unknown session → empty.
   t.deepEqual(getSessionCalls('nope'), [])
 
   // Separate explicit session.
-  await generateText(model, 'other', { session_id: 'sess-2' })
+  await generateText(model, 'other', { sessionId: 'sess-2' })
   t.is(getSessions().length, 2)
   t.is(getSessionCalls('sess-2').length, 1)
 })
@@ -97,7 +97,7 @@ test.serial('opt-in inferer groups prefix continuations into auto sessions', asy
   await generateText(model, [{ role: 'user', content: 'u1' }, { role: 'assistant', content: 'a1' }, { role: 'user', content: 'u2' }])
 
   const sessions = getSessions()
-  const autos = sessions.filter((s) => s.session_id.startsWith('auto-'))
+  const autos = sessions.filter((s) => s.sessionId.startsWith('auto-'))
   t.is(autos.length, 1, 'prefix continuation stays in one auto session')
   t.is(autos[0].source, 'Inferred')
   t.is(autos[0].calls.length, 2)
diff --git a/bindings/node/__test__/trace.test.ts b/bindings/node/__test__/trace.test.ts
index f6c2e7ac2..67b356f68 100644
--- a/bindings/node/__test__/trace.test.ts
+++ b/bindings/node/__test__/trace.test.ts
@@ -55,7 +55,7 @@ test('trace() records calls and query API returns typed results', async (t) => {
 
   // > 4 KiB user message so block-aligned prefixes actually match.
   const big = 'x'.repeat(5000)
-  await generateText(traced, [{ role: 'user', content: big }], { session_id: 'sess-1' })
+  await generateText(traced, [{ role: 'user', content: big }], { sessionId: 'sess-1' })
   await generateText(
     traced,
     [
@@ -63,7 +63,7 @@ test('trace() records calls and query API returns typed results', async (t) => {
       { role: 'assistant', content: 'a1' },
       { role: 'user', content: 'u2' },
     ],
-    { session_id: 'sess-1' },
+    { sessionId: 'sess-1' },
   )
 
   // Untraced models reject the query API.
@@ -74,28 +74,26 @@ test('trace() records calls and query API returns typed results', async (t) => {
   t.is(stats.length, 1)
   t.is(stats[0].provider, 'deepseek.chat')
   t.is(stats[0].requests, 2)
-  t.truthy(stats[0].reported_hit_rate)
-  t.truthy(stats[0].client_upper_bound_hit_rate)
-  t.truthy(stats[0].verdict_counts)
+  t.truthy(stats[0].verdictCounts)
 
   const chain = JSON.parse(traced.traceSessionChain('sess-1')) as any
-  t.is(chain.record_ids.length, 2)
-  t.truthy(chain.prefix_stability)
+  t.is(chain.recordIds.length, 2)
+  t.truthy(chain.prefixStability)
 
   const jsonl = traced.traceExportJsonl()
   const lines = jsonl.trim().split('\n')
   t.is(lines.length, 2, 'one TraceRecord per line')
   const first = JSON.parse(lines[0])
-  t.truthy(first.fingerprint.body_hash)
-  t.is(first.session_id, 'sess-1')
+  t.truthy(first.fingerprint.bodyHash)
+  t.is(first.sessionId, 'sess-1')
 
   traced.traceClear()
   t.is(traced.traceExportJsonl().trim(), '')
 })
 
-test('default streams carry no Raw parts (RFC-0016 M2 contract)', async (t) => {
+test('default streams carry no raw parts (RFC-0016 M2 contract)', async (t) => {
   // The RFC-0015 stream meta part was removed in favor of the M2 contract:
-  // Raw parts appear only when include_raw_chunks is set. Probing itself is
+  // raw parts appear only when includeRawChunks is set. Probing itself is
   // unaffected (TraceLayer records internally; request bodies remain
   // available on the non-streaming path — RFC-0023 runtime recording is the
   // planned (draft) mechanism for future streaming-side access).
@@ -110,10 +108,10 @@ test('default streams carry no Raw parts (RFC-0016 M2 contract)', async (t) => {
   }
 
   t.truthy(parts.length >= 1)
-  t.truthy(parts[0].includes('StreamStart'), `part[0] must be StreamStart: ${parts[0]}`)
+  t.is(JSON.parse(parts[0]).type, 'stream-start', `part[0] must be stream-start: ${parts[0]}`)
   t.true(
-    parts.every((p) => !p.includes('"Raw"')),
-    'no Raw parts by default',
+    parts.every((p) => JSON.parse(p).type !== 'raw'),
+    'no raw parts by default',
   )
 })
 
@@ -124,5 +122,5 @@ test('non-traced model still generates normally', async (t) => {
   const model = await deepseek('sk-test-fake-key', 'gpt-4o-mini', { baseUrl: url })
   const result: GenerateTextResult = await generateText(model, 'hello')
   t.truthy(result.text)
-  t.is(result.usage.input_tokens.total, 200)
+  t.is(result.usage.inputTokens.total, 200)
 })
diff --git a/bindings/node/__test__/transcription-session.test.ts b/bindings/node/__test__/transcription-session.test.ts
index ab3ea6630..9462b96fd 100644
--- a/bindings/node/__test__/transcription-session.test.ts
+++ b/bindings/node/__test__/transcription-session.test.ts
@@ -130,7 +130,7 @@ test.serial('nextPart timeout preserves the live session', async (t) => {
   const session = await startTranscriptionSession(model, null)
   try {
     const first = JSON.parse((await session.nextPart(3000)) as string)
-    t.truthy('StreamStart' in first)
+    t.is(first.type, 'stream-start')
 
     const err = (await t.throwsAsync(session.nextPart(250))) as AimuxError
     // Timeout is a spent budget, not a provider-call transport failure.
@@ -140,7 +140,7 @@ test.serial('nextPart timeout preserves the live session', async (t) => {
 
     // The session stayed live across the timeout.
     const late = JSON.parse((await session.nextPart(5000)) as string)
-    t.truthy('TranscriptDelta' in late)
+    t.is(late.type, 'transcript-delta')
   } finally {
     session.close()
   }
@@ -159,7 +159,7 @@ test.serial('server error event rejects with a typed APICallError', async (t) =>
   const session = await startTranscriptionSession(model, null)
   try {
     const first = JSON.parse((await session.nextPart(5000)) as string)
-    t.truthy('StreamStart' in first)
+    t.is(first.type, 'stream-start')
 
     const err = (await t.throwsAsync(session.nextPart(5000))) as AimuxError
     t.true(err instanceof APICallError)
diff --git a/bindings/node/__test__/wrapper.test.ts b/bindings/node/__test__/wrapper.test.ts
index b80b30c10..1212f4d36 100644
--- a/bindings/node/__test__/wrapper.test.ts
+++ b/bindings/node/__test__/wrapper.test.ts
@@ -2,8 +2,8 @@
 //
 // Verifies that `generateText` / `streamText` (from src/index.ts) erase the
 // JSON-string boundary: inputs are typed objects and outputs are parsed into
-// typed objects (`.text`, `.tool_calls`, `.raw.content`, stream `TextDelta` /
-// `ToolCall`). Uses the same mock-HTTP-server pattern as e2e.test.ts — no real
+// typed objects (`.text`, `.toolCalls`, `.raw.content`, stream `text-delta` /
+// `tool-call`). Uses the same mock-HTTP-server pattern as e2e.test.ts — no real
 // API calls.
 
 import test from 'ava'
@@ -14,7 +14,7 @@ import { createServer, type Server } from 'node:http'
 import { deepseek, anthropic, generateText, streamText } from '../src/index.ts'
 import type {
   GenerateTextResult,
-  StreamPart,
+  TextStreamPart,
   ModelMessage,
   Tool,
   ToolChoice,
@@ -123,12 +123,12 @@ const OPENAI_STREAM_TOOL_BODY = [
   'data: [DONE]\n\n',
 ].join('')
 
-// A typed tool definition + tool_choice, reused across tests.
+// A typed tool definition + toolChoice, reused across tests.
 const weatherTool: Tool = {
   type: 'function',
   name: 'get_weather',
   description: 'Get weather for a location',
-  input_schema: {
+  inputSchema: {
     type: 'object',
     properties: { location: { type: 'string' } },
     required: ['location'],
@@ -137,7 +137,7 @@ const weatherTool: Tool = {
 
 // ── Tests ───────────────────────────────────────────────────────────────────
 
-test('wrapper: generateText returns a typed object (.text / .usage / .finish_reason / .raw.content)', async (t) => {
+test('wrapper: generateText returns a typed object (.text / .usage / .finishReason / .raw.content)', async (t) => {
   const { server, url } = await startMockServer((req, res) => {
     res.writeHead(200, { 'content-type': 'application/json' })
     res.end(OPENAI_CHAT_RESPONSE)
@@ -151,20 +151,20 @@ test('wrapper: generateText returns a typed object (.text / .usage / .finish_rea
     // Top-level convenience fields are typed and present.
     t.is(result.text, 'Rust is a systems programming language.')
     t.truthy(result.usage, 'usage is present')
-    t.truthy(result.finish_reason, 'finish_reason is present')
+    t.truthy(result.finishReason, 'finishReason is present')
     t.true(Array.isArray(result.warnings))
 
     // raw is the provider-facing GenerateResult; .content is a typed array.
     t.truthy(result.raw, 'raw provider result is present')
     t.true(Array.isArray(result.raw.content), 'raw.content is a typed array')
-    const textPart = result.raw.content.find((c) => 'Text' in c)
-    t.truthy(textPart, 'raw.content contains a Text content part')
+    const textPart = result.raw.content.find((c) => c.type === 'text')
+    t.truthy(textPart, 'raw.content contains a text content part')
   } finally {
     await closeServer(server)
   }
 })
 
-test('wrapper: generateText parses tool_calls + raw.content ToolCall', async (t) => {
+test('wrapper: generateText parses toolCalls + raw.content tool-call', async (t) => {
   const { server, url } = await startMockServer((req, res) => {
     res.writeHead(200, { 'content-type': 'application/json' })
     res.end(OPENAI_TOOL_CALL_RESPONSE)
@@ -177,27 +177,27 @@ test('wrapper: generateText parses tool_calls + raw.content ToolCall', async (t)
       tools: [weatherTool],
     })
 
-    // Convenience field: tool_calls extracted and typed.
-    t.is(result.tool_calls.length, 1)
-    t.is(result.tool_calls[0].tool_name, 'get_weather')
-    t.is(result.tool_calls[0].tool_call_id, 'call_abc')
-    t.deepEqual(result.tool_calls[0].input, { location: 'Tokyo' })
-
-    // Structured content: raw.content carries the ToolCall variant.
-    const tc = result.raw.content.find((c) => 'ToolCall' in c)
-    t.truthy(tc, 'raw.content contains a ToolCall content part')
-    if (tc && 'ToolCall' in tc) {
-      t.is(tc.ToolCall.tool_name, 'get_weather')
-      t.is(tc.ToolCall.tool_call_id, 'call_abc')
+    // Convenience field: toolCalls extracted and typed.
+    t.is(result.toolCalls.length, 1)
+    t.is(result.toolCalls[0].toolName, 'get_weather')
+    t.is(result.toolCalls[0].toolCallId, 'call_abc')
+    t.deepEqual(result.toolCalls[0].input, { location: 'Tokyo' })
+
+    // Structured content: raw.content carries the tool-call variant.
+    const tc = result.raw.content.find((c) => c.type === 'tool-call')
+    t.truthy(tc, 'raw.content contains a tool-call content part')
+    if (tc && tc.type === 'tool-call') {
+      t.is(tc.toolName, 'get_weather')
+      t.is(tc.toolCallId, 'call_abc')
       // raw content keeps the provider's argument text (see e2e.test.ts).
-      t.is(tc.ToolCall.input, '{"location":"Tokyo"}')
+      t.is(tc.input, '{"location":"Tokyo"}')
     }
   } finally {
     await closeServer(server)
   }
 })
 
-test('wrapper: generateText passes tool_choice through to the provider', async (t) => {
+test('wrapper: generateText passes toolChoice through to the provider', async (t) => {
   const cap = capturingHandler(OPENAI_TOOL_CALL_RESPONSE)
   const { server, url } = await startMockServer(cap.handler)
 
@@ -206,7 +206,7 @@ test('wrapper: generateText passes tool_choice through to the provider', async (
     const toolChoice: ToolChoice = 'required'
     await generateText(model, 'Hello', {
       tools: [weatherTool],
-      tool_choice: toolChoice,
+      toolChoice,
     })
 
     const received = JSON.parse(cap.body)
@@ -260,7 +260,7 @@ test('wrapper: generateText works across providers (Anthropic mock)', async (t)
   }
 })
 
-test('wrapper: streamText yields typed TextDelta parts', async (t) => {
+test('wrapper: streamText yields typed text-delta parts', async (t) => {
   const { server, url } = await startMockServer((req, res) => {
     res.writeHead(200, { 'content-type': 'text/event-stream' })
     res.end(OPENAI_STREAM_BODY)
@@ -268,7 +268,7 @@ test('wrapper: streamText yields typed TextDelta parts', async (t) => {
 
   try {
     const model = await deepseek('test-key', 'gpt-4o', url)
-    const parts: StreamPart[] = []
+    const parts: TextStreamPart[] = []
     // streamText is an async generator of typed StreamParts — no JSON.parse.
     for await (const part of streamText(model, 'Say hello')) {
       parts.push(part)
@@ -276,18 +276,18 @@ test('wrapper: streamText yields typed TextDelta parts', async (t) => {
 
     t.true(parts.length > 0)
 
-    // Narrow the tagged union by key presence (typed, not `any`).
+    // Narrow the tagged union by its `type` tag (typed, not `any`).
     const deltas = parts
-      .filter((p): p is Extract<StreamPart, { TextDelta: unknown }> => 'TextDelta' in p)
-      .map((p) => p.TextDelta.delta)
+      .filter((p): p is Extract<TextStreamPart, { type: 'text-delta' }> => p.type === 'text-delta')
+      .map((p) => p.delta)
     t.is(deltas.join(''), 'Hello world')
   } finally {
     await closeServer(server)
   }
 })
 
-test('wrapper: streamText emits Raw parts when include_raw_chunks set', async (t) => {
-  // RFC-0016 M2: include_raw_chunks surfaces one Raw part per JSON SSE event
+test('wrapper: streamText emits raw parts when includeRawChunks set', async (t) => {
+  // RFC-0016 M2: includeRawChunks surfaces one raw part per JSON SSE event
   // (parsed payload, before the parsed parts; [DONE] excluded).
   const { server, url } = await startMockServer((req, res) => {
     res.writeHead(200, { 'content-type': 'text/event-stream' })
@@ -296,32 +296,32 @@ test('wrapper: streamText emits Raw parts when include_raw_chunks set', async (t
 
   try {
     const model = await deepseek('test-key', 'gpt-4o', url)
-    const parts: StreamPart[] = []
+    const parts: TextStreamPart[] = []
     for await (const part of streamText(model, 'Say hello', {
-      include_raw_chunks: true,
+      includeRawChunks: true,
     })) {
       parts.push(part)
     }
 
     const rawParts = parts.filter(
-      (p): p is Extract<StreamPart, { Raw: unknown }> => 'Raw' in p,
+      (p): p is Extract<TextStreamPart, { type: 'raw' }> => p.type === 'raw',
     )
     // OPENAI_STREAM_BODY has 3 JSON events (2 content + 1 usage chunk);
     // the [DONE] sentinel emits no Raw.
     t.is(rawParts.length, 3)
-    const firstRaw = rawParts[0].Raw as { raw_value: any }
-    t.is(firstRaw.raw_value.choices[0].delta.content, 'Hello')
+    const firstRaw = rawParts[0].rawValue as any
+    t.is(firstRaw.choices[0].delta.content, 'Hello')
 
-    // Raw for the first event precedes its TextDelta.
+    // The raw part for the first event precedes its text-delta.
     const firstRawIdx = parts.indexOf(rawParts[0])
-    const firstTextIdx = parts.findIndex((p) => 'TextDelta' in p)
-    t.true(firstRawIdx < firstTextIdx, 'Raw must precede the parsed parts')
+    const firstTextIdx = parts.findIndex((p) => p.type === 'text-delta')
+    t.true(firstRawIdx < firstTextIdx, 'raw must precede the parsed parts')
   } finally {
     await closeServer(server)
   }
 })
 
-test('wrapper: no Raw parts by default', async (t) => {
+test('wrapper: no raw parts by default', async (t) => {
   const { server, url } = await startMockServer((req, res) => {
     res.writeHead(200, { 'content-type': 'text/event-stream' })
     res.end(OPENAI_STREAM_BODY)
@@ -329,17 +329,17 @@ test('wrapper: no Raw parts by default', async (t) => {
 
   try {
     const model = await deepseek('test-key', 'gpt-4o', url)
-    const parts: StreamPart[] = []
+    const parts: TextStreamPart[] = []
     for await (const part of streamText(model, 'Say hello')) {
       parts.push(part)
     }
-    t.true(parts.every((p) => !('Raw' in p)), 'no Raw parts expected by default')
+    t.true(parts.every((p) => p.type !== 'raw'), 'no raw parts expected by default')
   } finally {
     await closeServer(server)
   }
 })
 
-test('wrapper: streamText yields typed ToolCall parts', async (t) => {
+test('wrapper: streamText yields typed tool-call parts', async (t) => {
   const { server, url } = await startMockServer((req, res) => {
     res.writeHead(200, { 'content-type': 'text/event-stream' })
     res.end(OPENAI_STREAM_TOOL_BODY)
@@ -347,7 +347,7 @@ test('wrapper: streamText yields typed ToolCall parts', async (t) => {
 
   try {
     const model = await deepseek('test-key', 'gpt-4o', url)
-    const parts: StreamPart[] = []
+    const parts: TextStreamPart[] = []
     for await (const part of streamText(model, "What's the weather?", {
       tools: [weatherTool],
     })) {
@@ -356,15 +356,15 @@ test('wrapper: streamText yields typed ToolCall parts', async (t) => {
 
     // The stream must contain a tool-related part.
     t.true(
-      parts.some((p) => 'ToolCall' in p || 'ToolInputStart' in p || 'ToolInputDelta' in p),
+      parts.some((p) => ['tool-call', 'tool-input-start', 'tool-input-delta'].includes(p.type)),
       'stream contained a tool-related StreamPart',
     )
 
-    // A complete ToolCall part carries the parsed tool name + input (typed).
-    const toolCall = parts.find((p): p is Extract<StreamPart, { ToolCall: unknown }> => 'ToolCall' in p)
+    // A complete tool-call part carries the parsed tool name + input (typed).
+    const toolCall = parts.find((p): p is Extract<TextStreamPart, { type: 'tool-call' }> => p.type === 'tool-call')
     if (toolCall) {
-      t.is(toolCall.ToolCall.tool_name, 'get_weather')
-      t.deepEqual(toolCall.ToolCall.input, { location: 'Tokyo' })
+      t.is(toolCall.toolName, 'get_weather')
+      t.deepEqual(toolCall.input, { location: 'Tokyo' })
     }
   } finally {
     await closeServer(server)
diff --git a/bindings/node/examples/generate-text.ts b/bindings/node/examples/generate-text.ts
index 11588b93f..a35b95938 100644
--- a/bindings/node/examples/generate-text.ts
+++ b/bindings/node/examples/generate-text.ts
@@ -16,7 +16,7 @@ async function main() {
   const result = await generateText(model, 'Explain Rust ownership in one sentence.')
   console.log('Text:', result.text)
   console.log('Usage:', result.usage)
-  console.log('Finish reason:', result.finish_reason)
+  console.log('Finish reason:', result.finishReason)
 }
 
 main().catch(console.error)
diff --git a/bindings/node/examples/stream-text.ts b/bindings/node/examples/stream-text.ts
index a2c1b9e74..65b465cd5 100644
--- a/bindings/node/examples/stream-text.ts
+++ b/bindings/node/examples/stream-text.ts
@@ -3,7 +3,6 @@
 // Run: OPENAI_API_KEY=sk-... node --experimental-strip-types examples/stream-text.ts
 
 import { openai, streamText } from '../src/index.ts'
-import type { StreamPart } from '../src/index.ts'
 
 async function main() {
   const apiKey = process.env.OPENAI_API_KEY
@@ -16,12 +15,10 @@ async function main() {
 
   console.log('Streaming:\n')
   for await (const part of streamText(model, 'Write a haiku about Rust.')) {
-    // StreamPart is a tagged union — check the type field
-    const obj = part as Record<string, unknown>
-    const key = Object.keys(obj)[0]
-    if (key === 'TextDelta') {
-      process.stdout.write((obj[key] as { delta: string }).delta)
-    } else if (key === 'Finish') {
+    // TextStreamPart is a tagged union — check the type field
+    if (part.type === 'text-delta') {
+      process.stdout.write(part.delta)
+    } else if (part.type === 'finish') {
       console.log('\n\n[done]')
     }
   }
diff --git a/bindings/node/index.d.ts b/bindings/node/index.d.ts
index c311db1b0..35d4d8947 100644
--- a/bindings/node/index.d.ts
+++ b/bindings/node/index.d.ts
@@ -112,7 +112,7 @@ export declare class Model {
    * Generate a structured JSON object from the model (M12, RFC-0016).
    *
    * Same signature as [`Model::generate_text`]; returns a JSON-serialized
-   * `GenerateObjectResult`. Pass `response_format: { "Json": { ... } }`
+   * `GenerateObjectResult`. Pass `responseFormat: { "type": "json", ... }`
    * via `options` for schema control; the function applies JSON repair
    * before parsing.
    */
@@ -298,7 +298,7 @@ export declare function anthropicAws(apiKey: string, region: string, modelId: st
  * Resolve one invalid tool call against a host's repair reply.
  *
  * `optsJson` is the same string the call was generated with; the tool set
- * comes from it. `replyJson` is `{"type":"repaired","tool_call":{…}}`,
+ * comes from it. `replyJson` is `{"type":"repaired","toolCall":{…}}`,
  * `{"type":"unchanged"}`, or `{"type":"failed","message":"…"}`. Returns the
  * resulting `ToolCall` JSON — valid, or invalid carrying a nested
  * `ToolCallRepairError`.
@@ -311,8 +311,8 @@ export declare function applyToolCallRepair(toolCallJson: string, optsJson: stri
 
 /**
  * Apply a repair reply to a serialized `GenerateTextResult` or
- * `GenerateObjectResult`, rewriting both `tool_calls` and the matching
- * `response_messages` tool-call part. `optsJson` is the same string the call
+ * `GenerateObjectResult`, rewriting both `toolCalls` and the matching
+ * `responseMessages` tool-call part. `optsJson` is the same string the call
  * was generated with.
  *
  * The OpenAI-shaped result has no equivalent: it carries no `invalid` /
@@ -564,14 +564,14 @@ export declare function tavilySearch(apiKey: string, baseUrl?: string | undefine
 /**
  * Build the repair argument for one invalid tool call.
  *
- * `toolCallJson` is a `GenerateTextResult.tool_calls` entry with
+ * `toolCallJson` is a `GenerateTextResult.toolCalls` entry with
  * `invalid: true`. `prompt` and `optsJson` are **the same two strings the
  * call was generated with** (`generateText` / `streamText`); messages,
  * instructions, and the tool set are derived from them here, so no caller
  * repeats that derivation.
  *
- * Returns `{tool_call, error, input_schema, tools, messages, instructions}`
- * — the AI SDK `repairToolCall` argument, with `tool_call.input` the
+ * Returns `{toolCall, error, inputSchema, tools, messages, instructions}`
+ * — the AI SDK `repairToolCall` argument, with `toolCall.input` the
  * provider's raw argument text — or the JSON literal `"null"` when the
  * options carried no tools: as in the AI SDK, a call made without a tool set
  * is never repaired, and the caller skips it. `"null"` is a success.
diff --git a/bindings/node/src/index.ts b/bindings/node/src/index.ts
index db64f3a0a..a4bc898e3 100644
--- a/bindings/node/src/index.ts
+++ b/bindings/node/src/index.ts
@@ -246,11 +246,11 @@ export type {
 
 /** The argument passed to a {@link RepairToolCall} function. */
 export type ToolCallRepairContext = {
-  tool_call: RawToolCall
+  toolCall: RawToolCall
   /** The lookup / parse / validation failure that made the call invalid. */
   error: AiMuxError
   /** JSON Schema of the called tool; the empty-object schema when unknown. */
-  input_schema: JsonValue
+  inputSchema: JsonValue
   tools: Tool[]
   messages: ModelMessage[]
   instructions: string | null
@@ -316,7 +316,7 @@ async function repairReplyFor(
   let reply: ToolCallRepairReply
   try {
     const replacement = await repair(context)
-    reply = replacement ? { type: 'repaired', tool_call: replacement } : { type: 'unchanged' }
+    reply = replacement ? { type: 'repaired', toolCall: replacement } : { type: 'unchanged' }
   } catch (e) {
     reply = { type: 'failed', message: errorMessage(e) }
   }
@@ -325,7 +325,7 @@ async function repairReplyFor(
 
 /**
  * Repair every invalid call in a serialized result, before it is decoded.
- * `tool_calls` sits at the top level of a generate-text result and under `raw`
+ * `toolCalls` sits at the top level of a generate-text result and under `raw`
  * in a generate-object one.
  */
 async function repairResultJson(
@@ -335,17 +335,17 @@ async function repairResultJson(
   repair: RepairToolCall,
 ): Promise<string> {
   const result = JSON.parse(resultJson) as {
-    tool_calls?: ToolCall[]
-    raw?: { tool_calls?: ToolCall[] }
+    toolCalls?: ToolCall[]
+    raw?: { toolCalls?: ToolCall[] }
   }
-  const calls = result.tool_calls ?? result.raw?.tool_calls ?? []
+  const calls = result.toolCalls ?? result.raw?.toolCalls ?? []
 
   let patched = resultJson
   for (const call of calls) {
     if (call.invalid !== true) continue
     const reply = await repairReplyFor(call, promptJson, optsJson, repair)
     if (reply === null) continue
-    patched = applyToolCallRepairToResult(patched, optsJson, call.tool_call_id, reply)
+    patched = applyToolCallRepairToResult(patched, optsJson, call.toolCallId, reply)
   }
   return patched
 }
@@ -355,8 +355,8 @@ async function repairResultJson(
  *
  * @param model   - A raw model instance from `openai()`, `anthropic()`, etc.
  * @param prompt  - A plain string or an array of typed chat messages.
- * @param options - Optional typed generation options (tools, tool_choice,
- *                  temperature, response_format, …). Invalid tool calls arrive
+ * @param options - Optional typed generation options (tools, toolChoice,
+ *                  temperature, responseFormat, …). Invalid tool calls arrive
  *                  with `invalid`/`error` set on the tool call; pass
  *                  {@link RepairToolCall | `repairToolCall`} to fix them up
  *                  before the result is decoded.
@@ -395,7 +395,7 @@ export async function generateText(
  *
  * @param model   - A raw model instance from `openai()`, `anthropic()`, etc.
  * @param prompt  - A plain string or an array of typed chat messages.
- * @param options - Optional typed generation options (tools, tool_choice, …).
+ * @param options - Optional typed generation options (tools, toolChoice, …).
  * @param signal  - Optional `AbortSignal`; aborting it cancels the stream.
  *
  * Internally drives the raw `model.streamText(JSON.stringify(prompt), …)`
@@ -407,7 +407,7 @@ export async function generateText(
  * import { openai, streamText } from 'aimux'
  * const model = await openai(apiKey, 'gpt-4o')
  * for await (const part of streamText(model, 'Write a haiku about Rust.')) {
- *   if ('TextDelta' in part) process.stdout.write(part.TextDelta.delta)
+ *   if (part.type === 'text-delta') process.stdout.write(part.delta)
  * }
  * ```
  */
@@ -424,15 +424,14 @@ export async function* streamText(
   const gen = await model.streamText(promptJson, optsJson, bridge)
   for await (const json of gen) {
     const part = JSON.parse(json) as TextStreamPart
-    // Only the settled ToolCall part is repairable; ToolInputDelta parts are
+    // Only the settled tool-call part is repairable; tool-input-delta parts are
     // the provider's raw text and pass through immediately (AI SDK behaviour).
-    if (repair && 'ToolCall' in part && part.ToolCall.invalid === true) {
-      const reply = await repairReplyFor(part.ToolCall, promptJson, optsJson, repair)
+    if (repair && part.type === 'tool-call' && part.invalid === true) {
+      const reply = await repairReplyFor(part, promptJson, optsJson, repair)
       if (reply !== null) {
         yield {
-          ToolCall: JSON.parse(
-            applyToolCallRepair(JSON.stringify(part.ToolCall), optsJson, reply),
-          ) as Extract<TextStreamPart, { ToolCall: unknown }>['ToolCall'],
+          type: 'tool-call',
+          ...(JSON.parse(applyToolCallRepair(JSON.stringify(part), optsJson, reply)) as ToolCall),
         }
         continue
       }
diff --git a/bindings/node/src/lib.rs b/bindings/node/src/lib.rs
index 786af95cd..711e9bf31 100644
--- a/bindings/node/src/lib.rs
+++ b/bindings/node/src/lib.rs
@@ -250,7 +250,7 @@ impl Model {
     /// Generate a structured JSON object from the model (M12, RFC-0016).
     ///
     /// Same signature as [`Model::generate_text`]; returns a JSON-serialized
-    /// `GenerateObjectResult`. Pass `response_format: { "Json": { ... } }`
+    /// `GenerateObjectResult`. Pass `responseFormat: { "type": "json", ... }`
     /// via `options` for schema control; the function applies JSON repair
     /// before parsing.
     #[napi(ts_return_type = "Promise<string>")]
@@ -1621,14 +1621,14 @@ fn parse_opts(json: Option<&str>) -> MResult<GenerateTextOptions> {
 
 /// Build the repair argument for one invalid tool call.
 ///
-/// `toolCallJson` is a `GenerateTextResult.tool_calls` entry with
+/// `toolCallJson` is a `GenerateTextResult.toolCalls` entry with
 /// `invalid: true`. `prompt` and `optsJson` are **the same two strings the
 /// call was generated with** (`generateText` / `streamText`); messages,
 /// instructions, and the tool set are derived from them here, so no caller
 /// repeats that derivation.
 ///
-/// Returns `{tool_call, error, input_schema, tools, messages, instructions}`
-/// — the AI SDK `repairToolCall` argument, with `tool_call.input` the
+/// Returns `{toolCall, error, inputSchema, tools, messages, instructions}`
+/// — the AI SDK `repairToolCall` argument, with `toolCall.input` the
 /// provider's raw argument text — or the JSON literal `"null"` when the
 /// options carried no tools: as in the AI SDK, a call made without a tool set
 /// is never repaired, and the caller skips it. `"null"` is a success.
@@ -1660,7 +1660,7 @@ pub fn tool_call_repair_context(
 /// Resolve one invalid tool call against a host's repair reply.
 ///
 /// `optsJson` is the same string the call was generated with; the tool set
-/// comes from it. `replyJson` is `{"type":"repaired","tool_call":{…}}`,
+/// comes from it. `replyJson` is `{"type":"repaired","toolCall":{…}}`,
 /// `{"type":"unchanged"}`, or `{"type":"failed","message":"…"}`. Returns the
 /// resulting `ToolCall` JSON — valid, or invalid carrying a nested
 /// `ToolCallRepairError`.
@@ -1689,8 +1689,8 @@ pub fn apply_tool_call_repair(
 }
 
 /// Apply a repair reply to a serialized `GenerateTextResult` or
-/// `GenerateObjectResult`, rewriting both `tool_calls` and the matching
-/// `response_messages` tool-call part. `optsJson` is the same string the call
+/// `GenerateObjectResult`, rewriting both `toolCalls` and the matching
+/// `responseMessages` tool-call part. `optsJson` is the same string the call
 /// was generated with.
 ///
 /// The OpenAI-shaped result has no equivalent: it carries no `invalid` /
diff --git a/bindings/node/src/types/AiMuxError.ts b/bindings/node/src/types/AiMuxError.ts
index 0a44a0a3e..87cd2b938 100644
--- a/bindings/node/src/types/AiMuxError.ts
+++ b/bindings/node/src/types/AiMuxError.ts
@@ -9,30 +9,8 @@ import type { RetryError } from "./RetryError";
  * failure came from. `ApiCallError` is boxed only to keep the Rust enum
  * compact; serde and every binding still observe the same object shape.
  */
-export type AiMuxError = { "ApiCall": ApiCallError } | { "Retry": RetryError } | { "JsonParse": string } | { "NoOutputGenerated": string } | { "InvalidResponseData": string } | { "ToolCallNotFoundForApproval": { tool_call_id: string, approval_id: string, } } | { "NoSuchTool": { tool_name: string, available_tools?: Array<string> | null,
+export type AiMuxError = { "name": "AI_APICallError" } & ApiCallError | { "name": "AI_RetryError" } & RetryError | { "name": "AI_JSONParseError", message: string, } | { "name": "AI_NoOutputGeneratedError", message: string, } | { "name": "AI_InvalidResponseDataError", message: string, } | { "name": "AI_ToolCallNotFoundForApprovalError", toolCallId: string, approvalId: string, } | { "name": "AI_NoSuchToolError", toolName: string, availableTools?: Array<string>, toolInput?: string, } | { "name": "AI_InvalidToolInputError", toolName: string, toolInput: string, cause: string, } | { "name": "AI_ToolCallRepairError", originalError: AiMuxError, cause: AiMuxError, } | { "name": "AI_InvalidArgumentError", message: string, } | { "name": "AI_InvalidPromptError", message: string, } | { "name": "AI_LoadAPIKeyError", envVar: string, description: string, } | { "name": "AI_LoadSettingError", envVar: string, 
 /**
- * Original argument text, when supplied by the provider. Absent in
- * older serialized errors and errors constructed without a call.
+ * `name` is the union's tag key, so the setting's name is `settingName`.
  */
-tool_input?: string | null, } } | { "InvalidToolInput": { tool_name: string, tool_input: string, cause: string, } } | { "ToolCallRepair": { original_error: AiMuxError, cause: AiMuxError, } } | { "InvalidArgument": string } | { "InvalidPrompt": string } | { "LoadApiKey": {
-/**
- * The environment variable consulted as the fallback.
- */
-env_var: string,
-/**
- * What the key is for, e.g. `"OpenAI"`.
- */
-description: string, } } | { "LoadSetting": {
-/**
- * The environment variable consulted as the fallback.
- */
-env_var: string,
-/**
- * The setting's parameter name, e.g. `"region"`.
- */
-name: string, } } | { "TokenExpired": string } | { "UnsupportedFunctionality": string } | { "NoSuchModel": { model_id: string,
-/**
- * What kind of model was requested (`"languageModel"`,
- * `"imageModel"`, …), the AI SDK's `modelType`.
- */
-model_type: string, } } | { "NoSuchProvider": { provider_id: string, model_id: string, model_type: string, available_providers: Array<string>, } } | { "Timeout": string } | { "Aborted": string } | { "Other": string };
+settingName: string, } | { "name": "AI_TokenExpiredError", message: string, } | { "name": "AI_UnsupportedFunctionalityError", message: string, } | { "name": "AI_NoSuchModelError", modelId: string, modelType: string, } | { "name": "AI_NoSuchProviderError", providerId: string, modelId: string, modelType: string, availableProviders: Array<string>, } | { "name": "TimeoutError", message: string, } | { "name": "AbortError", message: string, } | { "name": "AI_Error", message: string, };
diff --git a/bindings/node/src/types/ApiCallError.ts b/bindings/node/src/types/ApiCallError.ts
index fa26fb2d0..6dd28f598 100644
--- a/bindings/node/src/types/ApiCallError.ts
+++ b/bindings/node/src/types/ApiCallError.ts
@@ -31,20 +31,20 @@ url: string,
 /**
  * Sanitized request values used to create the API request.
  */
-request_body_values: JsonValue, 
+requestBodyValues: JsonValue, 
 /**
  * HTTP status of the response, when it came from one
  * (`APICallError.statusCode`). Always the *observed* status: the HTTP
  * layer fills it for every response-derived error; errors built without
  * an HTTP exchange leave it `None`.
  */
-status_code: number | null, 
+statusCode?: number, 
 /**
  * Provider's machine-readable code (OpenAI's `code`/`type`,
  * e.g. `"rate_limit_exceeded"` vs `"insufficient_quota"`). Never an HTTP
  * status. Our normalized take on `APICallError.data`.
  */
-provider_code: string | null, 
+providerCode?: string, 
 /**
  * Human-readable failure text. Provider text is verbatim when available;
  * locally detected transport/parse failures include their source detail.
@@ -56,19 +56,19 @@ message: string,
  * lossless evidence when `message`/`provider_code` are extractions.
  * `None` when the error did not come from a response body.
  */
-response_body: string | null, 
+responseBody?: string, 
 /**
  * Sanitized response headers.
  */
-response_headers: { [key in string]: string } | null, 
+responseHeaders?: { [key in string]: string }, 
 /**
  * Parsed provider error data.
  */
-data: JsonValue | null, 
+data?: JsonValue, 
 /**
  * Whether retrying can help (`APICallError.isRetryable`) — stored at
  * construction, exactly like the AI SDK: the response path computes it
  * from the status (408/409/429 or 5xx), and the transport path (no response, so
  * no status to compute from) sets it to `true` explicitly.
  */
-is_retryable: boolean, };
+isRetryable: boolean, };
diff --git a/bindings/node/src/types/AspectRatio.ts b/bindings/node/src/types/AspectRatio.ts
index c466b9b65..fe8b3f70c 100644
--- a/bindings/node/src/types/AspectRatio.ts
+++ b/bindings/node/src/types/AspectRatio.ts
@@ -7,4 +7,4 @@
  * `aspectRatio` options. Aligned with the TS template-literal type
  * `` `${number}:${number}` ``.
  */
-export type AspectRatio = { width: number, height: number, };
+export type AspectRatio = `${number}:${number}`;
diff --git a/bindings/node/src/types/AudioChunk.ts b/bindings/node/src/types/AudioChunk.ts
index 062104028..785ad95b9 100644
--- a/bindings/node/src/types/AudioChunk.ts
+++ b/bindings/node/src/types/AudioChunk.ts
@@ -3,4 +3,4 @@
 /**
  * A chunk of audio in a streaming transcription request.
  */
-export type AudioChunk = { "Binary": Array<number> } | { "Base64": string };
+export type AudioChunk = Array<number> | string;
diff --git a/bindings/node/src/types/AudioData.ts b/bindings/node/src/types/AudioData.ts
index b2ae98665..3f9a7e85f 100644
--- a/bindings/node/src/types/AudioData.ts
+++ b/bindings/node/src/types/AudioData.ts
@@ -6,4 +6,4 @@
  * The TS result type is `string | Uint8Array`. Providers should return data
  * without unnecessary conversion.
  */
-export type AudioData = { "Base64": string } | { "Binary": Array<number> };
+export type AudioData = string | Array<number>;
diff --git a/bindings/node/src/types/AudioInput.ts b/bindings/node/src/types/AudioInput.ts
index e513ffda3..b786aec38 100644
--- a/bindings/node/src/types/AudioInput.ts
+++ b/bindings/node/src/types/AudioInput.ts
@@ -3,4 +3,4 @@
 /**
  * Audio input: raw bytes or a base64-encoded string.
  */
-export type AudioInput = { "Binary": Array<number> } | { "Base64": string };
+export type AudioInput = Array<number> | string;
diff --git a/bindings/node/src/types/CallOptions.ts b/bindings/node/src/types/CallOptions.ts
index 80ffacc3c..a384768c1 100644
--- a/bindings/node/src/types/CallOptions.ts
+++ b/bindings/node/src/types/CallOptions.ts
@@ -22,70 +22,70 @@ prompt: Array<LanguageModelMessage>,
 /**
  * Maximum tokens to generate.
  */
-max_output_tokens: number | null, 
+maxOutputTokens?: number, 
 /**
  * Sampling temperature.
  */
-temperature: number | null, 
+temperature?: number, 
 /**
  * Stop sequences.
  */
-stop_sequences: Array<string> | null, 
+stopSequences?: Array<string>, 
 /**
  * Nucleus sampling `top_p`.
  */
-top_p: number | null, 
+topP?: number, 
 /**
  * Top-k sampling.
  */
-top_k: number | null, 
+topK?: number, 
 /**
  * Presence penalty.
  */
-presence_penalty: number | null, 
+presencePenalty?: number, 
 /**
  * Frequency penalty.
  */
-frequency_penalty: number | null, 
+frequencyPenalty?: number, 
 /**
  * Response format (text or JSON).
  */
-response_format: ResponseFormat | null, 
+responseFormat?: ResponseFormat, 
 /**
  * Seed for reproducibility.
  */
-seed: number | null, 
+seed?: number, 
 /**
  * Tools available to the model (function tools and/or provider-defined tools).
  */
-tools: Array<Tool> | null, 
+tools?: Array<Tool>, 
 /**
  * How the model should choose tools.
  */
-tool_choice: ToolChoice | null, 
+toolChoice?: ToolChoice, 
 /**
  * Extra HTTP headers.
  */
-headers: { [key in string]: string } | null, 
+headers?: { [key in string]: string }, 
 /**
  * Provider-specific options (keyed by provider name).
  */
-provider_options: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Top-level reasoning effort. Maps to OpenAI `reasoning_effort` and
  * Anthropic `thinking` config.
  */
-reasoning: ReasoningEffort | null, 
+reasoning?: ReasoningEffort, 
 /**
  * Per-call retry count override. `None` uses the provider's configured
  * Core operation retry. `Some(0)` disables retries.
  */
-max_retries: number | null, 
+maxRetries?: number, 
 /**
  * Per-call timeout configuration (total / first-chunk / chunk idle).
  * `None` = no timeouts (provider defaults still apply at the HTTP layer).
  */
-timeout: TimeoutConfiguration | null, 
+timeout?: TimeoutConfiguration, 
 /**
  * Session identifier, for grouping consecutive calls into a session
  * (observability, see RFC-0024). Explicit values take precedence; when
@@ -94,7 +94,7 @@ timeout: TimeoutConfiguration | null,
  * local grouping, while session headers in `headers` are for upstream
  * routing — both may share an id value but travel different paths.
  */
-session_id: string | null, 
+sessionId?: string, 
 /**
  * Emit raw provider stream chunks as `StreamPart::Raw` (debugging aid).
  * When `Some(true)`, streaming providers yield one `Raw` part per JSON
@@ -105,4 +105,4 @@ session_id: string | null,
  * openai-compatible registry providers); other provider families ignore
  * it for now (RFC-0016 M2).
  */
-include_raw_chunks: boolean | null, };
+includeRawChunks?: boolean, };
diff --git a/bindings/node/src/types/ContentPart.ts b/bindings/node/src/types/ContentPart.ts
index fd965cc5e..3c517067a 100644
--- a/bindings/node/src/types/ContentPart.ts
+++ b/bindings/node/src/types/ContentPart.ts
@@ -1,5 +1,6 @@
 // This file was generated by [ts-rs](https://github.com/Aleph-Alpha/ts-rs). Do not edit this file manually.
 import type { GeneratedFileData } from "./GeneratedFileData";
+import type { ToolResultOutput } from "./ToolResultOutput";
 import type { JsonValue } from "./serde_json/JsonValue";
 
 /**
@@ -7,61 +8,34 @@ import type { JsonValue } from "./serde_json/JsonValue";
  *
  * Shared between `ModelMessage` (user-facing) and `LanguageModelPrompt` (provider-facing).
  */
-export type ContentPart = { "type": "text", text: string, 
+export type ContentPart = { "type": "text", text: string,
 /**
  * Provider-specific options for this part (e.g. `openai.promptCacheBreakpoint`).
  */
-provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "custom", kind: string, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "reasoning_file", data: GeneratedFileData, media_type: string, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "tool_approval_request", approval_id: string, tool_call_id: string, reason?: string | null, is_automatic?: boolean | null, signature?: string | null, input_schema_input?: JsonValue | null, } | { "type": "image", image: Array<number>, media_type: string, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "file", data: Array<number>, media_type: string, filename?: string | null, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "file_base64", data: string, media_type: string, filename?: string | null, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "file_url", url: string, media_type: string, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "file_reference", media_type: string, reference: JsonValue, filename?: string | null, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "reasoning", text: string, signature: string | null, 
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "custom", kind: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "reasoning-file", data: GeneratedFileData, mediaType: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-approval-request", approvalId: string, toolCallId: string, reason?: string, isAutomatic?: boolean, signature?: string, inputSchemaInput?: JsonValue, } | { "type": "image", image: Array<number>, mediaType: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "file", data: Array<number>, mediaType: string, filename?: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "file-base64", data: string, mediaType: string, filename?: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "file-url", url: string, mediaType: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "file-reference", mediaType: string, reference: JsonValue, filename?: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "reasoning", text: string, signature?: string,
 /**
  * Provider-specific options for this part (e.g.
  * `anthropic.signature`, `anthropic.redactedData`,
  * `anthropic.cacheControl`).
  */
-provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "tool_call", tool_call_id: string, tool_name: string, 
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-call", toolCallId: string, toolName: string,
 /**
  * Arguments as a JSON value (usually an object).
  */
-input: JsonValue, 
+input: JsonValue,
 /**
  * Whether the tool call is executed by the provider rather than by
  * the client. This is part of the prompt contract because providers
  * need it to replay server tool calls on a later turn.
  */
-provider_executed?: boolean | null, 
+providerExecuted?: boolean,
 /**
  * Provider-specific options for this part (e.g.
  * `anthropic.cacheControl`).
  */
-provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "tool_result", tool_call_id: string, 
-/**
- * The tool's output (usually a JSON value or plain text).
- *
- * Accepts the legacy `output` field name (used by the Vercel AI SDK
- * and the 0.1.0 TypeScript bindings) as an alias during deserialization,
- * so multi-part tool messages constructed either way round-trip
- * correctly. Serialization always emits `result`.
- */
-result: JsonValue, 
-/**
- * The name of the tool that produced this result (optional on the
- * user-input side; providers that need it can look it up from the
- * preceding tool call).
- */
-tool_name?: string | null, 
-/**
- * Whether the result is an error or error message.
- */
-is_error?: boolean | null, 
-/**
- * Whether the result is preliminary (replaces prior, e.g. image previews).
- */
-preliminary?: boolean | null, 
-/**
- * Whether the tool is dynamic (defined at runtime, e.g. MCP tools).
- */
-dynamic?: boolean | null, 
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-result", toolCallId: string, toolName: string, output: ToolResultOutput,
 /**
  * Provider-specific options for this part (e.g.
  * `anthropic.cacheControl`).
  */
-provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, };
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/CustomPart.ts b/bindings/node/src/types/CustomPart.ts
index 39f66564d..2ea5ec97f 100644
--- a/bindings/node/src/types/CustomPart.ts
+++ b/bindings/node/src/types/CustomPart.ts
@@ -4,4 +4,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * Provider-specific content identified by its provider and kind.
  */
-export type CustomPart = { kind: string, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type CustomPart = { kind: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/EmbeddingCallOptions.ts b/bindings/node/src/types/EmbeddingCallOptions.ts
index 0d16b0ea1..33cfcf038 100644
--- a/bindings/node/src/types/EmbeddingCallOptions.ts
+++ b/bindings/node/src/types/EmbeddingCallOptions.ts
@@ -15,16 +15,16 @@ values: Array<string>,
 /**
  * Per-call retry override. `None` uses the model default.
  */
-max_retries: number | null, 
+maxRetries?: number, 
 /**
  * Per-call operation timeout.
  */
-timeout: TimeoutConfiguration | null, 
+timeout?: TimeoutConfiguration, 
 /**
  * Additional provider-specific options, keyed by provider name.
  */
-provider_options: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Additional HTTP headers to send with the request.
  */
-headers: { [key in string]: string } | null, };
+headers?: { [key in string]: string }, };
diff --git a/bindings/node/src/types/EmbeddingResponse.ts b/bindings/node/src/types/EmbeddingResponse.ts
index c758f7b32..72bc97787 100644
--- a/bindings/node/src/types/EmbeddingResponse.ts
+++ b/bindings/node/src/types/EmbeddingResponse.ts
@@ -8,8 +8,8 @@ export type EmbeddingResponse = {
 /**
  * Response headers.
  */
-headers: { [key in string]: string } | null, 
+headers?: { [key in string]: string }, 
 /**
  * The response body (opaque JSON).
  */
-body: JsonValue | null, };
+body?: JsonValue, };
diff --git a/bindings/node/src/types/EmbeddingResult.ts b/bindings/node/src/types/EmbeddingResult.ts
index fe4fc025a..c550a2dda 100644
--- a/bindings/node/src/types/EmbeddingResult.ts
+++ b/bindings/node/src/types/EmbeddingResult.ts
@@ -17,15 +17,15 @@ embeddings: Array<Array<number>>,
 /**
  * Token usage (input tokens only).
  */
-usage: EmbeddingUsage | null, 
+usage?: EmbeddingUsage, 
 /**
  * Additional provider-specific metadata.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Optional response information for debugging.
  */
-response: EmbeddingResponse | null, 
+response?: EmbeddingResponse, 
 /**
  * Warnings for the call, e.g. unsupported settings.
  */
diff --git a/bindings/node/src/types/EvaluationAnswer.ts b/bindings/node/src/types/EvaluationAnswer.ts
index 13aad7bc3..6c783e68e 100644
--- a/bindings/node/src/types/EvaluationAnswer.ts
+++ b/bindings/node/src/types/EvaluationAnswer.ts
@@ -12,7 +12,7 @@ choice: string,
 /**
  * Complete distribution over the question's options, when available.
  */
-probabilities?: { [key in string]: number } | null, } | { "type": "score", 
+probabilities?: { [key in string]: number }, } | { "type": "score", 
 /**
  * Fractional position in `[0, levels - 1]`.
  */
@@ -21,7 +21,7 @@ score: number,
  * Complete distribution keyed by zero-based level indices as
  * strings; when supplied, `score` is its probability-weighted mean.
  */
-probabilities?: { [key in string]: number } | null, } | { "type": "boolean", 
+probabilities?: { [key in string]: number }, } | { "type": "boolean", 
 /**
  * Model-estimated P(true), in `[0, 1]`. Not confidence in either
  * outcome.
diff --git a/bindings/node/src/types/EvaluationCallOptions.ts b/bindings/node/src/types/EvaluationCallOptions.ts
index 702e85cf1..671bdb60e 100644
--- a/bindings/node/src/types/EvaluationCallOptions.ts
+++ b/bindings/node/src/types/EvaluationCallOptions.ts
@@ -15,4 +15,4 @@ state: EvaluationInput,
 /**
  * Questions by id; insertion order is the question order.
  */
-questions: { [key in string]: EvaluationQuestion }, headers: { [key in string]: string } | null, provider_options: { [key in string]: { [key in string]: JsonValue } } | null, };
+questions: { [key in string]: EvaluationQuestion }, headers?: { [key in string]: string }, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/EvaluationQuestion.ts b/bindings/node/src/types/EvaluationQuestion.ts
index a0f49741b..1f6ae6136 100644
--- a/bindings/node/src/types/EvaluationQuestion.ts
+++ b/bindings/node/src/types/EvaluationQuestion.ts
@@ -14,4 +14,4 @@ criteria: { [key in string]: EvaluationInput | null }, } | { "type": "score", in
 /**
  * At least two ordered levels, indexed from zero.
  */
-criteria: Array<EvaluationInput | null>, } | { "type": "boolean", instructions: EvaluationInput, criteria?: BooleanCriteria | null, };
+criteria: Array<EvaluationInput | null>, } | { "type": "boolean", instructions: EvaluationInput, criteria?: BooleanCriteria, };
diff --git a/bindings/node/src/types/EvaluationResult.ts b/bindings/node/src/types/EvaluationResult.ts
index 63bb20c33..646491258 100644
--- a/bindings/node/src/types/EvaluationResult.ts
+++ b/bindings/node/src/types/EvaluationResult.ts
@@ -13,4 +13,4 @@ export type EvaluationResult = {
 /**
  * Exactly one answer per question, under the original question ids.
  */
-answers: { [key in string]: EvaluationAnswer }, rounding: EvaluationRounding | null, usage: EvaluationUsage | null, warnings: Array<Warning>, provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, response: ResponseInfo | null, };
+answers: { [key in string]: EvaluationAnswer }, rounding?: EvaluationRounding, usage?: EvaluationUsage, warnings: Array<Warning>, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, response?: ResponseInfo, };
diff --git a/bindings/node/src/types/EvaluationRounding.ts b/bindings/node/src/types/EvaluationRounding.ts
index e9d848269..127dd78ec 100644
--- a/bindings/node/src/types/EvaluationRounding.ts
+++ b/bindings/node/src/types/EvaluationRounding.ts
@@ -3,4 +3,4 @@
 /**
  * Decimal places the provider rounded its output to; omit for full precision.
  */
-export type EvaluationRounding = { probability_decimals: number | null, score_decimals: number | null, };
+export type EvaluationRounding = { probabilityDecimals?: number, scoreDecimals?: number, };
diff --git a/bindings/node/src/types/EvaluationUsage.ts b/bindings/node/src/types/EvaluationUsage.ts
index 467679edb..bf034ac58 100644
--- a/bindings/node/src/types/EvaluationUsage.ts
+++ b/bindings/node/src/types/EvaluationUsage.ts
@@ -3,4 +3,4 @@
 /**
  * Token usage of an evaluation.
  */
-export type EvaluationUsage = { input_tokens: number | null, output_tokens: number | null, };
+export type EvaluationUsage = { inputTokens?: number, outputTokens?: number, };
diff --git a/bindings/node/src/types/FileBytes.ts b/bindings/node/src/types/FileBytes.ts
index b06229434..c17813377 100644
--- a/bindings/node/src/types/FileBytes.ts
+++ b/bindings/node/src/types/FileBytes.ts
@@ -7,4 +7,4 @@
  * the upstream API returns base64, return [`FileBytes::Base64`]; if it
  * returns binary, return [`FileBytes::Binary`].
  */
-export type FileBytes = { "Binary": Array<number> } | { "Base64": string };
+export type FileBytes = Array<number> | string;
diff --git a/bindings/node/src/types/FileData.ts b/bindings/node/src/types/FileData.ts
index cec9eccbb..2410acc1e 100644
--- a/bindings/node/src/types/FileData.ts
+++ b/bindings/node/src/types/FileData.ts
@@ -6,4 +6,4 @@ import type { FileBytes } from "./FileBytes";
  *
  * Aligned with V4 `SharedV4FileData`.
  */
-export type FileData = { "Data": { data: FileBytes, } } | { "Url": { url: string, original_url?: string | null, } } | { "Reference": { reference: { [key in string]: string }, } } | { "Text": { text: string, } };
+export type FileData = { "type": "data", data: FileBytes, } | { "type": "url", url: string, originalUrl?: string, } | { "type": "reference", reference: { [key in string]: string }, } | { "type": "text", text: string, };
diff --git a/bindings/node/src/types/FilePart.ts b/bindings/node/src/types/FilePart.ts
index e412d1c0a..89a95b17b 100644
--- a/bindings/node/src/types/FilePart.ts
+++ b/bindings/node/src/types/FilePart.ts
@@ -5,4 +5,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * File content part (the only file part; `data` is a tagged `FileData`).
  */
-export type FilePart = { data: FileData, media_type: string, filename?: string | null, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type FilePart = { data: FileData, mediaType: string, filename?: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/Fingerprint.ts b/bindings/node/src/types/Fingerprint.ts
index 0849548a0..6198bad6a 100644
--- a/bindings/node/src/types/Fingerprint.ts
+++ b/bindings/node/src/types/Fingerprint.ts
@@ -8,13 +8,13 @@ export type Fingerprint = {
 /**
  * 128-bit hash of the whole denoised body (hex).
  */
-body_hash: string, len_bytes: bigint, block_size: bigint, 
+bodyHash: string, lenBytes: bigint, blockSize: bigint, 
 /**
  * Block-hash chain (hex), one entry per block.
  */
-block_hashes: Array<string>, 
+blockHashes: Array<string>, 
 /**
  * Token estimate — byte proxy `len_bytes / 4` when no tokenizer is
  * attached (RFC-0015 §4.1: W capped at one block in this mode).
  */
-token_estimate: bigint, };
+tokenEstimate: bigint, };
diff --git a/bindings/node/src/types/FinishReason.ts b/bindings/node/src/types/FinishReason.ts
index 3e557eb11..8a73936e6 100644
--- a/bindings/node/src/types/FinishReason.ts
+++ b/bindings/node/src/types/FinishReason.ts
@@ -12,4 +12,4 @@ unified: FinishReasonUnified,
 /**
  * The raw provider-specific reason string.
  */
-raw: string | null, };
+raw?: string, };
diff --git a/bindings/node/src/types/FunctionTool.ts b/bindings/node/src/types/FunctionTool.ts
index c43b94a6a..b369875b0 100644
--- a/bindings/node/src/types/FunctionTool.ts
+++ b/bindings/node/src/types/FunctionTool.ts
@@ -13,23 +13,23 @@ name: string,
 /**
  * Human-readable description.
  */
-description: string | null, 
+description?: string, 
 /**
  * JSON Schema describing the tool's parameters.
  */
-input_schema: JsonValue, 
+inputSchema: JsonValue, 
 /**
  * Whether the tool supports strict schema enforcement (OpenAI strict mode).
  */
-strict: boolean | null, 
+strict?: boolean, 
 /**
  * Provider-specific options, keyed by provider name
  * (e.g. `{"anthropic": {"eagerInputStreaming": true}}`). Aligned with the
  * V4 `providerOptions` field on function tools.
  */
-provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Example inputs for the tool (V4 `inputExamples`), used by some providers
  * to emit `input_examples` in the request body.
  */
-input_examples?: Array<FunctionToolInputExample> | null, };
+inputExamples?: Array<FunctionToolInputExample>, };
diff --git a/bindings/node/src/types/GenerateContent.ts b/bindings/node/src/types/GenerateContent.ts
index 456e003b9..2d0de8e94 100644
--- a/bindings/node/src/types/GenerateContent.ts
+++ b/bindings/node/src/types/GenerateContent.ts
@@ -10,4 +10,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * A content item in the generation result.
  */
-export type GenerateContent<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFile> = { "Text": { text: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolCall": C } | { "Source": Source } | { "Reasoning": ReasoningOutput } | { "File": GeneratedFile } | { "ReasoningFile": R } | { "Custom": { kind: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolApprovalRequest": A } | { "ToolResult": ToolResult };
+export type GenerateContent<C = RawToolCall, A = RawToolApprovalRequest, R = GeneratedFile> = { "type": "text", text: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-call" } & C | { "type": "source" } & Source | { "type": "reasoning" } & ReasoningOutput | { "type": "file" } & GeneratedFile | { "type": "reasoning-file" } & R | { "type": "custom", kind: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-approval-request" } & A | { "type": "tool-result" } & ToolResult;
diff --git a/bindings/node/src/types/GenerateObjectResult.ts b/bindings/node/src/types/GenerateObjectResult.ts
index 89e5eb445..b47ba2960 100644
--- a/bindings/node/src/types/GenerateObjectResult.ts
+++ b/bindings/node/src/types/GenerateObjectResult.ts
@@ -18,11 +18,11 @@ object: JsonValue,
 /**
  * Why generation stopped.
  */
-finish_reason: FinishReason, 
+finishReason: FinishReason, 
 /**
  * Raw provider-specific finish reason string.
  */
-raw_finish_reason: string | null, 
+rawFinishReason?: string, 
 /**
  * Token usage.
  */
@@ -34,11 +34,11 @@ warnings: Array<Warning>,
 /**
  * Concatenated reasoning text (if the model produced reasoning/thinking).
  */
-reasoning: string | null, 
+reasoning?: string, 
 /**
  * Provider-specific metadata (e.g. Anthropic cache info).
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Response metadata (id, timestamp, model_id).
  */
diff --git a/bindings/node/src/types/GenerateResult.ts b/bindings/node/src/types/GenerateResult.ts
index 70e2ef62c..979381eab 100644
--- a/bindings/node/src/types/GenerateResult.ts
+++ b/bindings/node/src/types/GenerateResult.ts
@@ -21,7 +21,7 @@ content: Array<GenerateContent<RawToolCall, RawToolApprovalRequest, GeneratedFil
 /**
  * Why generation stopped.
  */
-finish_reason: FinishReason, 
+finishReason: FinishReason, 
 /**
  * Token usage.
  */
@@ -33,12 +33,12 @@ warnings: Array<Warning>,
 /**
  * Provider-specific metadata.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Optional request information for telemetry and debugging.
  */
-request: RequestInfo | null, 
+request?: RequestInfo, 
 /**
  * Optional response metadata, headers, and body.
  */
-response: ResponseInfo | null, };
+response?: ResponseInfo, };
diff --git a/bindings/node/src/types/GenerateTextOptions.ts b/bindings/node/src/types/GenerateTextOptions.ts
index 4bbbcfc83..d80dc0be7 100644
--- a/bindings/node/src/types/GenerateTextOptions.ts
+++ b/bindings/node/src/types/GenerateTextOptions.ts
@@ -12,31 +12,31 @@ import type { JsonValue } from "./serde_json/JsonValue";
  * Unlike `CallOptions` (provider-facing), this does not include `prompt`
  * (passed separately) and defaults are more ergonomic.
  */
-export type GenerateTextOptions = { max_output_tokens: number | null, temperature: number | null, stop_sequences: Array<string> | null, top_p: number | null, top_k: number | null, presence_penalty: number | null, frequency_penalty: number | null, response_format: ResponseFormat | null, seed: number | null, tools: Array<Tool> | null, tool_choice: ToolChoice | null, headers: { [key in string]: string } | null, provider_options: { [key in string]: { [key in string]: JsonValue } } | null, 
+export type GenerateTextOptions = { maxOutputTokens?: number, temperature?: number, stopSequences?: Array<string>, topP?: number, topK?: number, presencePenalty?: number, frequencyPenalty?: number, responseFormat?: ResponseFormat, seed?: number, tools?: Array<Tool>, toolChoice?: ToolChoice, headers?: { [key in string]: string }, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Top-level reasoning effort.
  */
-reasoning: ReasoningEffort | null, 
+reasoning?: ReasoningEffort, 
 /**
  * System instructions prepended to the prompt.
  */
-instructions: string | null, 
+instructions?: string, 
 /**
  * Per-call retry count override. `None` = provider default, `Some(0)` = disable.
  */
-max_retries: number | null, 
+maxRetries?: number, 
 /**
  * Per-call timeout configuration (total / first-chunk / chunk idle).
  */
-timeout: TimeoutConfiguration | null, 
+timeout?: TimeoutConfiguration, 
 /**
  * Session identifier, for grouping consecutive calls into a session
  * (observability, see RFC-0024). Orthogonal to RFC-0019 session-affinity
  * headers. When `None` and the optional session inferer is enabled, one
  * may be inferred from prompt-prefix continuation.
  */
-session_id: string | null, 
+sessionId?: string, 
 /**
  * Emit raw provider stream chunks as `StreamPart::Raw` (debugging aid).
  */
-include_raw_chunks: boolean | null, };
+includeRawChunks?: boolean, };
diff --git a/bindings/node/src/types/GenerateTextResult.ts b/bindings/node/src/types/GenerateTextResult.ts
index 98b9ac920..4fde225a8 100644
--- a/bindings/node/src/types/GenerateTextResult.ts
+++ b/bindings/node/src/types/GenerateTextResult.ts
@@ -18,76 +18,76 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * Result of `generate_text` (user-facing).
  */
-export type GenerateTextResult = {
+export type GenerateTextResult = { 
 /**
  * Ordered generated content with parsed tool calls.
  */
-content?: Array<GenerateContent<ToolCall, ToolApprovalRequestOutput, ReasoningFileOutput>>,
+content?: Array<GenerateContent<ToolCall, ToolApprovalRequestOutput, ReasoningFileOutput>>, 
 /**
  * The generated text (concatenated from all text content parts).
  */
-text: string,
+text: string, 
 /**
  * Tool calls requested by the model.
  */
-tool_calls: Array<ToolCall>,
+toolCalls: Array<ToolCall>, 
 /**
  * Why generation stopped.
  */
-finish_reason: FinishReason,
+finishReason: FinishReason, 
 /**
  * Token usage.
  */
-usage: Usage,
+usage: Usage, 
 /**
  * Warnings from the provider.
  */
-warnings: Array<Warning>,
+warnings: Array<Warning>, 
 /**
  * Raw provider result (for advanced use).
  */
-raw: GenerateResult,
+raw: GenerateResult, 
 /**
  * Reasoning / thinking segments from the model.
  */
-reasoning: Array<ReasoningPart>,
+reasoning: Array<ReasoningPart>, 
 /**
  * Concatenated reasoning text (convenience for `reasoning.iter().map(text).join("")`).
  */
-reasoning_text: string,
+reasoningText: string, 
 /**
  * Sources / citations (search-preview models).
  */
-sources: Array<Source>,
+sources: Array<Source>, 
 /**
  * Files generated by the model.
  */
-files: Array<GeneratedFile>,
+files: Array<GeneratedFile>, 
 /**
  * Assistant messages ready to append to the prompt for the next turn
  * (solves the multi-turn "manually build assistant message" footgun).
  */
-response_messages: Array<ModelMessage>,
+responseMessages: Array<ModelMessage>, 
 /**
  * The raw provider-specific finish reason string (e.g. "stop",
  * "end_turn", "safety"). Useful when `finish_reason.unified` is `Other`.
  */
-raw_finish_reason: string | null,
+rawFinishReason?: string, 
 /**
  * Provider-specific metadata (e.g. Anthropic cache info). Mirrored from
  * `raw.provider_metadata` for top-level convenience.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null,
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Request information from the provider.
  */
-request: RequestInfo,
+request: RequestInfo, 
 /**
  * Response metadata, headers, and body from the provider.
  */
-response: ResponseInfo,
+response: ResponseInfo, 
 /**
  * Total token usage across all steps. In single-step mode (aimux's
  * default), `total_usage` equals `usage`. Provided for AI SDK parity.
  */
-total_usage: Usage, };
+totalUsage: Usage, };
diff --git a/bindings/node/src/types/GeneratedFile.ts b/bindings/node/src/types/GeneratedFile.ts
index 48333ce42..35493ce51 100644
--- a/bindings/node/src/types/GeneratedFile.ts
+++ b/bindings/node/src/types/GeneratedFile.ts
@@ -8,4 +8,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
  * or URL. The payload of both `GenerateContent::File` and
  * `StreamPart::File`, and the element of the text results' `files`.
  */
-export type GeneratedFile = { data: GeneratedFileData, media_type: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type GeneratedFile = { data: GeneratedFileData, mediaType: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/GeneratedFileData.ts b/bindings/node/src/types/GeneratedFileData.ts
index f8b500f58..91d743a20 100644
--- a/bindings/node/src/types/GeneratedFileData.ts
+++ b/bindings/node/src/types/GeneratedFileData.ts
@@ -4,4 +4,4 @@ import type { FileBytes } from "./FileBytes";
 /**
  * Data or a URL returned for a generated file.
  */
-export type GeneratedFileData = { "Data": { data: FileBytes, } } | { "Url": { url: string, original_url?: string | null, } };
+export type GeneratedFileData = { "type": "data", data: FileBytes, } | { "type": "url", url: string, originalUrl?: string, };
diff --git a/bindings/node/src/types/HttpExchange.ts b/bindings/node/src/types/HttpExchange.ts
index 336f21769..7e0f6916f 100644
--- a/bindings/node/src/types/HttpExchange.ts
+++ b/bindings/node/src/types/HttpExchange.ts
@@ -11,7 +11,7 @@ export type HttpExchange = {
  * Composite step this exchange belongs to (e.g. `router[0]:openai/gpt-4o`
  * or `moa.ref[1]:...`). `None` for a plain, non-composite operation.
  */
-step?: string | null, 
+step?: string, 
 /**
  * Core operation attempt, starting at 1. Attempt numbers are unique
  * across the whole call, including composite child steps, so
@@ -21,11 +21,11 @@ attempt: number,
 /**
  * HTTP exchange within the operation attempt, starting at 1.
  */
-exchange_index: number, request: HttpRecord, 
+exchangeIndex: number, request: HttpRecord, 
 /**
  * None = 请求失败未获响应。
  */
-response: ResponseRecord | null, timing: TimingRecord, error: string | null, 
+response?: ResponseRecord, timing: TimingRecord, error?: string, 
 /**
  * 流式:该 exchange 是否已终结(收到 response 补全)。非流式恒 true。
  */
diff --git a/bindings/node/src/types/HttpRecord.ts b/bindings/node/src/types/HttpRecord.ts
index 90ae88a53..55db75602 100644
--- a/bindings/node/src/types/HttpRecord.ts
+++ b/bindings/node/src/types/HttpRecord.ts
@@ -8,4 +8,4 @@ headers: Array<[string, string]>,
 /**
  * 明文(脱敏后);None = 无 body。
  */
-body: string | null, };
+body?: string, };
diff --git a/bindings/node/src/types/ImageCallOptions.ts b/bindings/node/src/types/ImageCallOptions.ts
index 1b490a4cb..b900b647e 100644
--- a/bindings/node/src/types/ImageCallOptions.ts
+++ b/bindings/node/src/types/ImageCallOptions.ts
@@ -15,7 +15,7 @@ export type ImageCallOptions = {
  * Prompt for the image generation. `None` for operations (e.g. upscaling)
  * that do not require a prompt.
  */
-prompt: string | null, 
+prompt?: string, 
 /**
  * Number of images to generate.
  */
@@ -24,37 +24,37 @@ n: number,
  * Size of the images, in `{width}x{height}` format.
  * `None` uses the provider's default size.
  */
-size: Size | null, 
+size?: Size, 
 /**
  * Aspect ratio of the images, in `{width}:{height}` format.
  * `None` uses the provider's default aspect ratio.
  */
-aspect_ratio: AspectRatio | null, 
+aspectRatio?: AspectRatio, 
 /**
  * Seed for deterministic generation. `None` uses the provider's default.
  */
-seed: number | null, 
+seed?: number, 
 /**
  * Images for image editing or variation generation.
  */
-files: Array<ImageFile> | null, 
+files?: Array<ImageFile>, 
 /**
  * Mask image for inpainting operations.
  */
-mask: ImageFile | null, 
+mask?: ImageFile, 
 /**
  * Additional provider-specific options, keyed by provider name.
  */
-provider_options: { [key in string]: { [key in string]: JsonValue } }, 
+providerOptions: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Per-call retry override. `None` uses the model default.
  */
-max_retries: number | null, 
+maxRetries?: number, 
 /**
  * Per-call operation timeout.
  */
-timeout: TimeoutConfiguration | null, 
+timeout?: TimeoutConfiguration, 
 /**
  * Additional HTTP headers to send with the request.
  */
-headers: { [key in string]: string } | null, };
+headers?: { [key in string]: string }, };
diff --git a/bindings/node/src/types/ImageFile.ts b/bindings/node/src/types/ImageFile.ts
index 545e7cef5..64aef8164 100644
--- a/bindings/node/src/types/ImageFile.ts
+++ b/bindings/node/src/types/ImageFile.ts
@@ -6,16 +6,16 @@ import type { ImageFileData } from "./ImageFileData";
  *
  * Aligned with V4 `ImageModelV4File`.
  */
-export type ImageFile = { "File": { 
+export type ImageFile = { "type": "file", 
 /**
  * IANA media type, e.g. `"image/png"`.
  */
-media_type: string, 
+mediaType: string, 
 /**
  * File data as base64 string or binary bytes.
  */
-data: ImageFileData, } } | { "Url": { 
+data: ImageFileData, } | { "type": "url", 
 /**
  * The URL of the image file.
  */
-url: string, } };
+url: string, };
diff --git a/bindings/node/src/types/ImageFileData.ts b/bindings/node/src/types/ImageFileData.ts
index 96be9636f..46fceec6a 100644
--- a/bindings/node/src/types/ImageFileData.ts
+++ b/bindings/node/src/types/ImageFileData.ts
@@ -3,4 +3,4 @@
 /**
  * File payload for an [`ImageFile::File`]: base64 string or raw bytes.
  */
-export type ImageFileData = { "Base64": string } | { "Binary": Array<number> };
+export type ImageFileData = string | Array<number>;
diff --git a/bindings/node/src/types/ImageOutputs.ts b/bindings/node/src/types/ImageOutputs.ts
index 38be2705b..9ff238058 100644
--- a/bindings/node/src/types/ImageOutputs.ts
+++ b/bindings/node/src/types/ImageOutputs.ts
@@ -9,4 +9,4 @@
  * upstream API returns base64, return [`ImageOutputs::Base64`]; if binary,
  * return [`ImageOutputs::Binary`].
  */
-export type ImageOutputs = { "Base64": Array<string> } | { "Binary": Array<Array<number>> };
+export type ImageOutputs = Array<string> | Array<Array<number>>;
diff --git a/bindings/node/src/types/ImageResponse.ts b/bindings/node/src/types/ImageResponse.ts
index 0f5f33dd0..3e7e3660e 100644
--- a/bindings/node/src/types/ImageResponse.ts
+++ b/bindings/node/src/types/ImageResponse.ts
@@ -7,12 +7,12 @@ export type ImageResponse = {
 /**
  * Timestamp for the start of the generated response (ISO 8601 string).
  */
-timestamp: string | null, 
+timestamp?: string, 
 /**
  * The ID of the model that was used to generate the response.
  */
-model_id: string | null, 
+modelId?: string, 
 /**
  * Response headers.
  */
-headers: { [key in string]: string } | null, };
+headers?: { [key in string]: string }, };
diff --git a/bindings/node/src/types/ImageResult.ts b/bindings/node/src/types/ImageResult.ts
index 4276c580f..cdf420366 100644
--- a/bindings/node/src/types/ImageResult.ts
+++ b/bindings/node/src/types/ImageResult.ts
@@ -22,7 +22,7 @@ warnings: Array<Warning>,
 /**
  * Additional provider-specific metadata, keyed by provider name.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Response information for telemetry and debugging.
  */
@@ -30,4 +30,4 @@ response: ImageResponse,
 /**
  * Optional token usage for the call (if the provider reports it).
  */
-usage: ImageUsage | null, };
+usage?: ImageUsage, };
diff --git a/bindings/node/src/types/ImageUsage.ts b/bindings/node/src/types/ImageUsage.ts
index 86fb55df9..51e882868 100644
--- a/bindings/node/src/types/ImageUsage.ts
+++ b/bindings/node/src/types/ImageUsage.ts
@@ -9,12 +9,12 @@ export type ImageUsage = {
 /**
  * Number of input (prompt) tokens used.
  */
-input_tokens: number | null, 
+inputTokens?: number, 
 /**
  * Number of output tokens used, if reported.
  */
-output_tokens: number | null, 
+outputTokens?: number, 
 /**
  * Total tokens as reported by the provider.
  */
-total_tokens: number | null, };
+totalTokens?: number, };
diff --git a/bindings/node/src/types/InputAudioFormat.ts b/bindings/node/src/types/InputAudioFormat.ts
index c45946104..20d045790 100644
--- a/bindings/node/src/types/InputAudioFormat.ts
+++ b/bindings/node/src/types/InputAudioFormat.ts
@@ -7,8 +7,8 @@ export type InputAudioFormat = {
 /**
  * Audio format type, e.g. `"audio/pcm"`, `"audio/pcmu"`, `"audio/pcma"`.
  */
-format_type: string, 
+formatType: string, 
 /**
  * Sample rate in Hz. Only applicable for formats that require a rate.
  */
-rate: number | null, };
+rate?: number, };
diff --git a/bindings/node/src/types/InputRecord.ts b/bindings/node/src/types/InputRecord.ts
index 61cb96552..51348b0b6 100644
--- a/bindings/node/src/types/InputRecord.ts
+++ b/bindings/node/src/types/InputRecord.ts
@@ -12,7 +12,7 @@ export type InputRecord = {
  * Decision contract used for validation and deterministic offline replay.
  * Connection settings remain outside the recording.
  */
-decision_capabilities?: DecisionCapabilities | null, operation?: RecordingOperation,
+decisionCapabilities?: DecisionCapabilities | null, operation?: RecordingOperation,
 /**
  * 完整 prompt(消息数组,含 ContentPart::Image 等多模态)。
  */
diff --git a/bindings/node/src/types/InputTokenUsage.ts b/bindings/node/src/types/InputTokenUsage.ts
index 658af2dab..10d3c62eb 100644
--- a/bindings/node/src/types/InputTokenUsage.ts
+++ b/bindings/node/src/types/InputTokenUsage.ts
@@ -7,16 +7,16 @@ export type InputTokenUsage = {
 /**
  * Total tokens.
  */
-total: number | null, 
+total?: number, 
 /**
  * Tokens that were not served from cache.
  */
-no_cache: number | null, 
+noCache?: number, 
 /**
  * Tokens served from cache (read).
  */
-cache_read: number | null, 
+cacheRead?: number, 
 /**
  * Tokens written to cache.
  */
-cache_write: number | null, };
+cacheWrite?: number, };
diff --git a/bindings/node/src/types/LanguageModelMessage.ts b/bindings/node/src/types/LanguageModelMessage.ts
index a7052a024..4efb3d34d 100644
--- a/bindings/node/src/types/LanguageModelMessage.ts
+++ b/bindings/node/src/types/LanguageModelMessage.ts
@@ -10,4 +10,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
  * `provider_options` is message-level (e.g. `anthropic.cacheControl`),
  * mirroring `LanguageModelV4Message.providerOptions`.
  */
-export type LanguageModelMessage = { "role": "system", content: string, provider_options: { [key in string]: { [key in string]: JsonValue } } | null, } | { "role": "user", content: Array<UserPart>, provider_options: { [key in string]: { [key in string]: JsonValue } } | null, } | { "role": "assistant", content: Array<AssistantPart>, provider_options: { [key in string]: { [key in string]: JsonValue } } | null, } | { "role": "tool", content: Array<ToolPart>, provider_options: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type LanguageModelMessage = { "role": "system", content: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "role": "user", content: Array<UserPart>, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "role": "assistant", content: Array<AssistantPart>, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "role": "tool", content: Array<ToolPart>, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/LanguageModelReasoningPart.ts b/bindings/node/src/types/LanguageModelReasoningPart.ts
index 4fb52fa16..de94608a6 100644
--- a/bindings/node/src/types/LanguageModelReasoningPart.ts
+++ b/bindings/node/src/types/LanguageModelReasoningPart.ts
@@ -4,4 +4,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * Reasoning / thinking content part.
  */
-export type LanguageModelReasoningPart = { text: string, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type LanguageModelReasoningPart = { text: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/ModelCapabilities.ts b/bindings/node/src/types/ModelCapabilities.ts
index e38d3622b..c5926ecb4 100644
--- a/bindings/node/src/types/ModelCapabilities.ts
+++ b/bindings/node/src/types/ModelCapabilities.ts
@@ -7,11 +7,11 @@ export type ModelCapabilities = {
 /**
  * Supports function/tool calling.
  */
-tool_call: boolean, 
+toolCall: boolean, 
 /**
  * Supports structured / JSON-schema output.
  */
-structured_output: boolean, 
+structuredOutput: boolean, 
 /**
  * Accepts a `temperature` parameter.
  */
diff --git a/bindings/node/src/types/ModelCost.ts b/bindings/node/src/types/ModelCost.ts
index 39e510c55..434d43609 100644
--- a/bindings/node/src/types/ModelCost.ts
+++ b/bindings/node/src/types/ModelCost.ts
@@ -4,4 +4,4 @@
  * Pricing (advisory, read-only metadata). Units follow the source catalog
  * (anya2a uses USD per 1M tokens).
  */
-export type ModelCost = { input?: number | null, output?: number | null, cache_read?: number | null, cache_write?: number | null, };
+export type ModelCost = { input?: number, output?: number, cacheRead?: number, cacheWrite?: number, };
diff --git a/bindings/node/src/types/ModelLimits.ts b/bindings/node/src/types/ModelLimits.ts
index 8835a50e6..9688b1e8c 100644
--- a/bindings/node/src/types/ModelLimits.ts
+++ b/bindings/node/src/types/ModelLimits.ts
@@ -7,12 +7,12 @@ export type ModelLimits = {
 /**
  * Maximum total context window (input + output), in tokens.
  */
-context?: number | null, 
+context?: number, 
 /**
  * Maximum output tokens.
  */
-output?: number | null, 
+output?: number, 
 /**
  * Maximum input tokens (when distinct from `context`).
  */
-input?: number | null, };
+input?: number, };
diff --git a/bindings/node/src/types/ModelSpec.ts b/bindings/node/src/types/ModelSpec.ts
index 69ead413e..6c435f5a1 100644
--- a/bindings/node/src/types/ModelSpec.ts
+++ b/bindings/node/src/types/ModelSpec.ts
@@ -14,12 +14,12 @@ import type { JsonValue } from "./serde_json/JsonValue";
  * All fields are advisory — aimux never auto-applies them in the request path.
  * The caller reads them to decide how to configure `GenerateTextOptions`.
  */
-export type ModelSpec = { display_name?: string | null, type: ModelType, limits: ModelLimits, modalities: ModelModalities, capabilities: ModelCapabilities, reasoning?: ReasoningSpec | null, cost?: ModelCost | null, source: CatalogueSource, 
+export type ModelSpec = { displayName?: string, type: ModelType, limits: ModelLimits, modalities: ModelModalities, capabilities: ModelCapabilities, reasoning?: ReasoningSpec, cost?: ModelCost, source: CatalogueSource, 
 /**
  * Provider name (normalized to aimux registry spelling).
  */
-provider?: string | null, 
+provider?: string, 
 /**
  * Verbatim source entry, preserved for forward compatibility / debugging.
  */
-raw?: JsonValue | null, };
+raw?: JsonValue, };
diff --git a/bindings/node/src/types/OutcomeRecord.ts b/bindings/node/src/types/OutcomeRecord.ts
index 38777b88a..4f5c94f36 100644
--- a/bindings/node/src/types/OutcomeRecord.ts
+++ b/bindings/node/src/types/OutcomeRecord.ts
@@ -9,13 +9,13 @@ export type OutcomeRecord = {
 /**
  * Normalized decision result, independent of a provider's wire format.
  */
-decision_result?: JsonValue | null, status: OutcomeStatus, finish_reason: string | null, error: string | null,
+decisionResult?: JsonValue | null, status: OutcomeStatus, finishReason?: string, error?: string,
 /**
  * Lossless structured domain error. In particular, `RetryError.errors`
  * keeps the complete attempt history rather than only its display text.
  */
-error_value: JsonValue | null,
+errorValue?: JsonValue,
 /**
  * 序列化的 Usage。
  */
-usage: JsonValue | null, };
+usage?: JsonValue, };
diff --git a/bindings/node/src/types/OutputTokenUsage.ts b/bindings/node/src/types/OutputTokenUsage.ts
index 47528d6d8..38a53cdb0 100644
--- a/bindings/node/src/types/OutputTokenUsage.ts
+++ b/bindings/node/src/types/OutputTokenUsage.ts
@@ -3,12 +3,12 @@
 /**
  * Output token usage detail.
  */
-export type OutputTokenUsage = { total: number | null, 
+export type OutputTokenUsage = { total?: number, 
 /**
  * Text (non-reasoning) tokens (output side).
  */
-text: number | null, 
+text?: number, 
 /**
  * Reasoning tokens (output side).
  */
-reasoning: number | null, };
+reasoning?: number, };
diff --git a/bindings/node/src/types/PrefixBreak.ts b/bindings/node/src/types/PrefixBreak.ts
index 78aa70bdb..dadebc493 100644
--- a/bindings/node/src/types/PrefixBreak.ts
+++ b/bindings/node/src/types/PrefixBreak.ts
@@ -4,4 +4,4 @@ import type { BreakKind } from "./BreakKind";
 /**
  * A prefix break between two consecutive records of a session.
  */
-export type PrefixBreak = { at_record_id: string, prev_record_id: string, lcp_bytes: bigint, expected_break: boolean, kind: BreakKind, };
+export type PrefixBreak = { atRecordId: string, prevRecordId: string, lcpBytes: bigint, expectedBreak: boolean, kind: BreakKind, };
diff --git a/bindings/node/src/types/ProviderAuditSpec.ts b/bindings/node/src/types/ProviderAuditSpec.ts
index 6f60f4e92..a6189f4d6 100644
--- a/bindings/node/src/types/ProviderAuditSpec.ts
+++ b/bindings/node/src/types/ProviderAuditSpec.ts
@@ -13,7 +13,7 @@ family: ProviderFamily,
 /**
  * Token quantization granularity; `None` = no quantization.
  */
-gran: bigint | null, 
+gran?: bigint, 
 /**
  * Minimum cache threshold (tokens): below it `claimed` must be 0.
  */
@@ -21,7 +21,7 @@ threshold: bigint,
 /**
  * TTL_idle (ms), conservative upper bound.
  */
-ttl_ms: bigint, 
+ttlMs: bigint, 
 /**
  * gpt-5.6+ semantics (no quantization, implicit-breakpoint whitelist,
  * write/read equality).
@@ -31,4 +31,4 @@ model56plus: boolean,
  * Claimed-hit ceiling: above this multiple of the client upper bound
  * the verdict is overclaim (W) even in shared mode. `None` = use U+τ.
  */
-shared_ceiling_mult: number | null, };
+sharedCeilingMult?: number, };
diff --git a/bindings/node/src/types/ProviderRecord.ts b/bindings/node/src/types/ProviderRecord.ts
index a59cfc113..a7a5071dc 100644
--- a/bindings/node/src/types/ProviderRecord.ts
+++ b/bindings/node/src/types/ProviderRecord.ts
@@ -10,7 +10,7 @@ export type ProviderRecord = {
 /**
  * provider 名(registry 名),回放重建的查找键,如 "openai"、"deepseek"。
  */
-provider_id: string, 
+providerId: string, 
 /**
  * `model.provider()`,如 "openai.chat"。
  */
@@ -18,4 +18,4 @@ provider: string,
 /**
  * `model.model_id()`,如 "gpt-4o"。
  */
-model_id: string, };
+modelId: string, };
diff --git a/bindings/node/src/types/RawToolApprovalRequest.ts b/bindings/node/src/types/RawToolApprovalRequest.ts
index 5b3dd37db..fc5e8e66b 100644
--- a/bindings/node/src/types/RawToolApprovalRequest.ts
+++ b/bindings/node/src/types/RawToolApprovalRequest.ts
@@ -4,4 +4,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * Provider approval request for a provider-executed call.
  */
-export type RawToolApprovalRequest = { approval_id: string, tool_call_id: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type RawToolApprovalRequest = { approvalId: string, toolCallId: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/RawToolCall.ts b/bindings/node/src/types/RawToolCall.ts
index c44ead3c6..a924f9231 100644
--- a/bindings/node/src/types/RawToolCall.ts
+++ b/bindings/node/src/types/RawToolCall.ts
@@ -7,4 +7,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
  * The wire shape matches [`ToolCall`] field for field, except that `input`
  * is the provider's raw argument *text* rather than a parsed value.
  */
-export type RawToolCall = { tool_call_id: string, tool_name: string, input: string, provider_executed?: boolean | null, dynamic?: boolean | null, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type RawToolCall = { toolCallId: string, toolName: string, input: string, providerExecuted?: boolean, dynamic?: boolean, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/ReasoningFileOutput.ts b/bindings/node/src/types/ReasoningFileOutput.ts
index 57eaa2c0b..59f585e2c 100644
--- a/bindings/node/src/types/ReasoningFileOutput.ts
+++ b/bindings/node/src/types/ReasoningFileOutput.ts
@@ -2,4 +2,4 @@
 import type { GeneratedFile } from "./GeneratedFile";
 import type { JsonValue } from "./serde_json/JsonValue";
 
-export type ReasoningFileOutput = { file: GeneratedFile, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type ReasoningFileOutput = { file: GeneratedFile, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/ReasoningFilePart.ts b/bindings/node/src/types/ReasoningFilePart.ts
index b8c1e4aab..9b70a21f3 100644
--- a/bindings/node/src/types/ReasoningFilePart.ts
+++ b/bindings/node/src/types/ReasoningFilePart.ts
@@ -5,4 +5,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * A file generated as part of reasoning.
  */
-export type ReasoningFilePart = { data: GeneratedFileData, media_type: string, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type ReasoningFilePart = { data: GeneratedFileData, mediaType: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/ReasoningOutput.ts b/bindings/node/src/types/ReasoningOutput.ts
index 90a4f59b8..60837bf23 100644
--- a/bindings/node/src/types/ReasoningOutput.ts
+++ b/bindings/node/src/types/ReasoningOutput.ts
@@ -8,4 +8,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
  * (`providerMetadata.anthropic.signature`). The payload of
  * `GenerateContent::Reasoning` and the element of the text results' `reasoning`.
  */
-export type ReasoningOutput = { text: string, provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type ReasoningOutput = { text: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/ReasoningPart.ts b/bindings/node/src/types/ReasoningPart.ts
index 37a7c5873..70292cde6 100644
--- a/bindings/node/src/types/ReasoningPart.ts
+++ b/bindings/node/src/types/ReasoningPart.ts
@@ -5,4 +5,4 @@ import type { ReasoningOutput } from "./ReasoningOutput";
 /**
  * Text or a generated file produced during reasoning.
  */
-export type ReasoningPart = ReasoningOutput | ReasoningFileOutput;
+export type ReasoningPart = { "type": "reasoning" } & ReasoningOutput | { "type": "reasoning-file" } & ReasoningFileOutput;
diff --git a/bindings/node/src/types/ReasoningSpec.ts b/bindings/node/src/types/ReasoningSpec.ts
index 2ce4266a0..03f35424a 100644
--- a/bindings/node/src/types/ReasoningSpec.ts
+++ b/bindings/node/src/types/ReasoningSpec.ts
@@ -13,23 +13,23 @@ supported: boolean,
 /**
  * Whether reasoning should be on by default.
  */
-default_enabled: boolean, 
+defaultEnabled: boolean, 
 /**
  * Primary control mode.
  */
-mode: ReasoningMode | null, 
+mode?: ReasoningMode, 
 /**
  * Default effort level (when `mode` is Effort/Mixed).
  */
-effort_default?: string | null, 
+effortDefault?: string, 
 /**
  * Supported effort levels.
  */
-effort_options?: Array<string>, 
+effortOptions?: Array<string>, 
 /**
  * Minimum token budget (when `mode` is Budget/Mixed).
  */
-budget_min?: number | null, 
+budgetMin?: number, 
 /**
  * Interleaved reasoning (thinking blocks between tool calls).
  */
@@ -37,4 +37,4 @@ interleaved: boolean,
 /**
  * Output visibility.
  */
-visibility: ReasoningVisibility | null, };
+visibility?: ReasoningVisibility, };
diff --git a/bindings/node/src/types/Recording.ts b/bindings/node/src/types/Recording.ts
index 8ef090200..28c75efe9 100644
--- a/bindings/node/src/types/Recording.ts
+++ b/bindings/node/src/types/Recording.ts
@@ -16,11 +16,11 @@ schema: number,
 /**
  * 暴露逻辑调用 ID,关联三层。
  */
-call_id: string, 
+callId: string, 
 /**
  * ISO 8601 时间戳。
  */
-recorded_at: string, 
+recordedAt: string, 
 /**
  * ① 输入侧:调用参数(prompt + options)。
  */
@@ -45,14 +45,14 @@ complete: boolean,
  * 传输层封闭信号:层 B 声明"不再有 exchange"(P1:层 A 收尾自动发;
  * P2:由层 B 发送)。false = 仍可能来 exchange(记录暂不可定稿)。
  */
-transport_closed: boolean, 
+transportClosed: boolean, 
 /**
  * 会话归组(RFC-0024 P3):所在会话 id。None = 未归组(无 session_id 且
  * 推断关闭)。由 `Recorder::record_session` 填充,写入 InputRecord 之前
  * 或之后均可(writer 端按 call_id 合并)。
  */
-session_id?: string | null, 
+sessionId?: string, 
 /**
  * 会话内步号(0 起,由 SessionStore 分配)。与 `session_id` 同生命周期。
  */
-step?: number | null, };
+step?: number, };
diff --git a/bindings/node/src/types/RequestCacheHints.ts b/bindings/node/src/types/RequestCacheHints.ts
index 9e3ff369d..7f90b919c 100644
--- a/bindings/node/src/types/RequestCacheHints.ts
+++ b/bindings/node/src/types/RequestCacheHints.ts
@@ -7,4 +7,4 @@ export type RequestCacheHints = {
 /**
  * Whether any message requested `cache_control` (write) this call.
  */
-requested_write: boolean, };
+requestedWrite: boolean, };
diff --git a/bindings/node/src/types/RequestInfo.ts b/bindings/node/src/types/RequestInfo.ts
index 098d3be65..ce8a9ae4e 100644
--- a/bindings/node/src/types/RequestInfo.ts
+++ b/bindings/node/src/types/RequestInfo.ts
@@ -10,4 +10,4 @@ export type RequestInfo = {
 /**
  * The request body that was sent (opaque JSON).
  */
-body: JsonValue | null, };
+body?: JsonValue, };
diff --git a/bindings/node/src/types/RerankingCallOptions.ts b/bindings/node/src/types/RerankingCallOptions.ts
index 3de208995..57bfbf122 100644
--- a/bindings/node/src/types/RerankingCallOptions.ts
+++ b/bindings/node/src/types/RerankingCallOptions.ts
@@ -20,20 +20,20 @@ query: string,
 /**
  * Optional limit: return only the top `n` documents.
  */
-top_n: number | null, 
+topN?: number, 
 /**
  * Per-call retry override. `None` uses the model default.
  */
-max_retries: number | null, 
+maxRetries?: number, 
 /**
  * Per-call operation timeout.
  */
-timeout: TimeoutConfiguration | null, 
+timeout?: TimeoutConfiguration, 
 /**
  * Additional provider-specific options, keyed by provider name.
  */
-provider_options: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Additional HTTP headers to send with the request.
  */
-headers: { [key in string]: string } | null, };
+headers?: { [key in string]: string }, };
diff --git a/bindings/node/src/types/RerankingDocuments.ts b/bindings/node/src/types/RerankingDocuments.ts
index 2fd5f51f2..3e5941d58 100644
--- a/bindings/node/src/types/RerankingDocuments.ts
+++ b/bindings/node/src/types/RerankingDocuments.ts
@@ -6,4 +6,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
  *
  * Aligned with V4 `RerankingModelV4CallOptions.documents`.
  */
-export type RerankingDocuments = { "Text": { values: Array<string>, } } | { "Object": { values: Array<JsonValue>, } };
+export type RerankingDocuments = { "type": "text", values: Array<string>, } | { "type": "object", values: Array<JsonValue>, };
diff --git a/bindings/node/src/types/RerankingRank.ts b/bindings/node/src/types/RerankingRank.ts
index 986d3b2e0..6d236bd16 100644
--- a/bindings/node/src/types/RerankingRank.ts
+++ b/bindings/node/src/types/RerankingRank.ts
@@ -11,4 +11,4 @@ index: number,
 /**
  * The relevance score of the document after reranking.
  */
-relevance_score: number, };
+relevanceScore: number, };
diff --git a/bindings/node/src/types/RerankingResponse.ts b/bindings/node/src/types/RerankingResponse.ts
index 83a869529..179b39959 100644
--- a/bindings/node/src/types/RerankingResponse.ts
+++ b/bindings/node/src/types/RerankingResponse.ts
@@ -8,20 +8,20 @@ export type RerankingResponse = {
 /**
  * ID for the generated response, if the provider sends one.
  */
-id: string | null, 
+id?: string, 
 /**
  * Timestamp for the start of the generated response (ISO 8601 string).
  */
-timestamp: string | null, 
+timestamp?: string, 
 /**
  * The ID of the model that was used to generate the response.
  */
-model_id: string | null, 
+modelId?: string, 
 /**
  * Response headers.
  */
-headers: { [key in string]: string } | null, 
+headers?: { [key in string]: string }, 
 /**
  * Response body (opaque JSON).
  */
-body: JsonValue | null, };
+body?: JsonValue, };
diff --git a/bindings/node/src/types/RerankingResult.ts b/bindings/node/src/types/RerankingResult.ts
index 4d339296c..6534cba8e 100644
--- a/bindings/node/src/types/RerankingResult.ts
+++ b/bindings/node/src/types/RerankingResult.ts
@@ -19,12 +19,12 @@ ranking: Array<RerankingRank>,
 /**
  * Additional provider-specific metadata, keyed by provider name.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Warnings for the call, e.g. unsupported settings.
  */
-warnings: Array<Warning> | null, 
+warnings?: Array<Warning>, 
 /**
  * Optional response information for debugging.
  */
-response: RerankingResponse | null, };
+response?: RerankingResponse, };
diff --git a/bindings/node/src/types/ResponseFormat.ts b/bindings/node/src/types/ResponseFormat.ts
index 88f2405e1..19760bc4f 100644
--- a/bindings/node/src/types/ResponseFormat.ts
+++ b/bindings/node/src/types/ResponseFormat.ts
@@ -4,4 +4,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * How the model should format its response.
  */
-export type ResponseFormat = "Text" | { "Json": { schema: JsonValue | null, name: string | null, description: string | null, } };
+export type ResponseFormat = { "type": "text" } | { "type": "json", schema?: JsonValue, name?: string, description?: string, };
diff --git a/bindings/node/src/types/ResponseInfo.ts b/bindings/node/src/types/ResponseInfo.ts
index 52103ad78..c832686c2 100644
--- a/bindings/node/src/types/ResponseInfo.ts
+++ b/bindings/node/src/types/ResponseInfo.ts
@@ -14,20 +14,20 @@ export type ResponseInfo = {
 /**
  * ID of the generated response, when supplied by the provider.
  */
-id?: string | null, 
+id?: string, 
 /**
  * Timestamp for the start of the generated response (ISO 8601 string).
  */
-timestamp: string | null, 
+timestamp?: string, 
 /**
  * The ID of the model that was used to generate the response.
  */
-model_id: string | null, 
+modelId?: string, 
 /**
  * Response headers.
  */
-headers: { [key in string]: string } | null, 
+headers?: { [key in string]: string }, 
 /**
  * The response body (opaque JSON).
  */
-body: JsonValue | null, };
+body?: JsonValue, };
diff --git a/bindings/node/src/types/ResponseMetadata.ts b/bindings/node/src/types/ResponseMetadata.ts
index 4c2084e8c..7ead4483e 100644
--- a/bindings/node/src/types/ResponseMetadata.ts
+++ b/bindings/node/src/types/ResponseMetadata.ts
@@ -3,4 +3,4 @@
 /**
  * Metadata about the API response.
  */
-export type ResponseMetadata = { id: string | null, timestamp: string | null, model_id: string | null, };
+export type ResponseMetadata = { id?: string, timestamp?: string, modelId?: string, };
diff --git a/bindings/node/src/types/ResponseRecord.ts b/bindings/node/src/types/ResponseRecord.ts
index 28abf8d33..0edbedf7e 100644
--- a/bindings/node/src/types/ResponseRecord.ts
+++ b/bindings/node/src/types/ResponseRecord.ts
@@ -4,8 +4,8 @@ export type ResponseRecord = { status: number, headers: Array<[string, string]>,
 /**
  * 非流式:完整 JSON;流式:原始 SSE 拼接文本(上限截断)。
  */
-body: string | null, stream_chunks: number | null, 
+body?: string, streamChunks?: number, 
 /**
  * 首字节延迟(流式)。
  */
-ttfb_ms?: number | null, };
+ttfbMs?: number, };
diff --git a/bindings/node/src/types/RuntimeModel.ts b/bindings/node/src/types/RuntimeModel.ts
index f12ea8802..5a131cbac 100644
--- a/bindings/node/src/types/RuntimeModel.ts
+++ b/bindings/node/src/types/RuntimeModel.ts
@@ -16,8 +16,8 @@ id: string,
 /**
  * `owned_by` field from OpenAI-compatible `/models` responses.
  */
-owned_by?: string | null, 
+ownedBy?: string, 
 /**
  * `created` unix timestamp from OpenAI-compatible `/models` responses.
  */
-created?: number | null, };
+created?: number, };
diff --git a/bindings/node/src/types/SearchCallOptions.ts b/bindings/node/src/types/SearchCallOptions.ts
index 1658b5e28..d4c69eb05 100644
--- a/bindings/node/src/types/SearchCallOptions.ts
+++ b/bindings/node/src/types/SearchCallOptions.ts
@@ -13,40 +13,40 @@ query: string,
 /**
  * Maximum number of results to return.
  */
-max_results: number | null, 
+maxResults?: number, 
 /**
  * Whether to include raw page content in results.
  *
  * Provider support varies; providers that cannot honor this should
  * issue a warning rather than erroring.
  */
-include_raw_content: boolean | null, 
+includeRawContent?: boolean, 
 /**
  * Optional time range filter (e.g. `"day"`, `"week"`, `"month"`,
  * `"year"`). Provider support varies.
  */
-time_range: string | null, 
+timeRange?: string, 
 /**
  * Optional list of domains to include in results.
  */
-include_domains: Array<string> | null, 
+includeDomains?: Array<string>, 
 /**
  * Optional list of domains to exclude from results.
  */
-exclude_domains: Array<string> | null, 
+excludeDomains?: Array<string>, 
 /**
  * Per-call retry override. `None` uses the model default.
  */
-max_retries: number | null, 
+maxRetries?: number, 
 /**
  * Per-call operation timeout.
  */
-timeout: TimeoutConfiguration | null, 
+timeout?: TimeoutConfiguration, 
 /**
  * Additional provider-specific options, keyed by provider name.
  */
-provider_options: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Additional HTTP headers to send with the request.
  */
-headers: { [key in string]: string } | null, };
+headers?: { [key in string]: string }, };
diff --git a/bindings/node/src/types/SearchResponse.ts b/bindings/node/src/types/SearchResponse.ts
index 287e4d97b..ced52e84b 100644
--- a/bindings/node/src/types/SearchResponse.ts
+++ b/bindings/node/src/types/SearchResponse.ts
@@ -8,8 +8,8 @@ export type SearchResponse = {
 /**
  * Response headers.
  */
-headers: { [key in string]: string } | null, 
+headers?: { [key in string]: string }, 
 /**
  * The response body (opaque JSON).
  */
-body: JsonValue | null, };
+body?: JsonValue, };
diff --git a/bindings/node/src/types/SearchResult.ts b/bindings/node/src/types/SearchResult.ts
index ee3f6a49b..b009ede13 100644
--- a/bindings/node/src/types/SearchResult.ts
+++ b/bindings/node/src/types/SearchResult.ts
@@ -16,11 +16,11 @@ results: Array<SearchResultItem>,
  * An optional direct answer / summary (some providers return
  * an AI-generated answer alongside results).
  */
-answer: string | null, 
+answer?: string, 
 /**
  * Additional provider-specific metadata.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Warnings for the call.
  */
@@ -28,4 +28,4 @@ warnings: Array<Warning>,
 /**
  * Optional response information for debugging.
  */
-response: SearchResponse | null, };
+response?: SearchResponse, };
diff --git a/bindings/node/src/types/SearchResultItem.ts b/bindings/node/src/types/SearchResultItem.ts
index 0c8b8d67a..7a4b9fb14 100644
--- a/bindings/node/src/types/SearchResultItem.ts
+++ b/bindings/node/src/types/SearchResultItem.ts
@@ -8,25 +8,25 @@ export type SearchResultItem = {
 /**
  * The title of the result (e.g. page title).
  */
-title: string | null, 
+title?: string, 
 /**
  * The URL of the result.
  */
-url: string | null, 
+url?: string, 
 /**
  * A snippet/summary of the result content.
  */
-content: string | null, 
+content?: string, 
 /**
  * Raw page content (when `include_raw_content` is requested and
  * supported by the provider).
  */
-raw_content: string | null, 
+rawContent?: string, 
 /**
  * A relevance score (0.0–1.0) if the provider returns one.
  */
-score: number | null, 
+score?: number, 
 /**
  * Provider-specific metadata for this result.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, };
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/SessionCall.ts b/bindings/node/src/types/SessionCall.ts
index 695d5a603..79c38fd77 100644
--- a/bindings/node/src/types/SessionCall.ts
+++ b/bindings/node/src/types/SessionCall.ts
@@ -10,7 +10,7 @@ export type SessionCall = {
  * Recording / TraceRecord `call_id` (RFC-0015/0023): once those land,
  * this field carries their value.
  */
-call_id: string, 
+callId: string, 
 /**
  * Step within the session (0-based).
  */
@@ -18,4 +18,4 @@ step: number,
 /**
  * When the call was recorded (RFC 3339 UTC, millisecond precision).
  */
-recorded_at: string, };
+recordedAt: string, };
diff --git a/bindings/node/src/types/SessionChainView.ts b/bindings/node/src/types/SessionChainView.ts
index 3212b456a..3d416fd18 100644
--- a/bindings/node/src/types/SessionChainView.ts
+++ b/bindings/node/src/types/SessionChainView.ts
@@ -4,4 +4,4 @@ import type { PrefixBreak } from "./PrefixBreak";
 /**
  * Session chain view (append-only order + prefix stability, RFC-0015 §5.3).
  */
-export type SessionChainView = { session_id: string, record_ids: Array<string>, prefix_stability: number, breaks: Array<PrefixBreak>, };
+export type SessionChainView = { sessionId: string, recordIds: Array<string>, prefixStability: number, breaks: Array<PrefixBreak>, };
diff --git a/bindings/node/src/types/SessionStepStat.ts b/bindings/node/src/types/SessionStepStat.ts
index 1eeb3bf02..a75b0652f 100644
--- a/bindings/node/src/types/SessionStepStat.ts
+++ b/bindings/node/src/types/SessionStepStat.ts
@@ -12,17 +12,17 @@ step: number,
 /**
  * The call's id (association key into Recording / replay, RFC-0023).
  */
-call_id: string, 
+callId: string, 
 /**
  * Reported hit rate for this step (`cache_read / input_total`); `None`
  * when the call carried no input/cache-read usage.
  */
-hit_rate?: number | null, 
+hitRate?: number, 
 /**
  * Audit verdict (present only when an auditor is attached).
  */
-verdict?: Verdict | null, 
+verdict?: Verdict, 
 /**
  * Error string for failed calls (failures are part of the session).
  */
-error?: string | null, };
+error?: string, };
diff --git a/bindings/node/src/types/SessionView.ts b/bindings/node/src/types/SessionView.ts
index 60a153778..709ebfef1 100644
--- a/bindings/node/src/types/SessionView.ts
+++ b/bindings/node/src/types/SessionView.ts
@@ -5,7 +5,7 @@ import type { SessionSource } from "./SessionSource";
 /**
  * Aggregated view of one session (session_id → ordered calls).
  */
-export type SessionView = { session_id: string, 
+export type SessionView = { sessionId: string, 
 /**
  * Where the session id came from (first call's source).
  */
diff --git a/bindings/node/src/types/Size.ts b/bindings/node/src/types/Size.ts
index 960627114..32711645a 100644
--- a/bindings/node/src/types/Size.ts
+++ b/bindings/node/src/types/Size.ts
@@ -7,4 +7,4 @@
  * `resolution` options. Aligned with the TS template-literal type
  * `` `${number}x${number}` ``.
  */
-export type Size = { width: number, height: number, };
+export type Size = `${number}x${number}`;
diff --git a/bindings/node/src/types/Source.ts b/bindings/node/src/types/Source.ts
index f9239ad68..49c10d7d3 100644
--- a/bindings/node/src/types/Source.ts
+++ b/bindings/node/src/types/Source.ts
@@ -6,4 +6,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
  * payload of both `GenerateContent::Source` and `StreamPart::Source`, and the
  * element of the text results' `sources`.
  */
-export type Source = { "source_type": "url", id: string, url: string, title?: string | null, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "source_type": "document", id: string, media_type: string, title: string, filename?: string | null, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type Source = { "sourceType": "url", id: string, url: string, title?: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "sourceType": "document", id: string, mediaType: string, title: string, filename?: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/SpeechCallOptions.ts b/bindings/node/src/types/SpeechCallOptions.ts
index d9424d439..9f947c252 100644
--- a/bindings/node/src/types/SpeechCallOptions.ts
+++ b/bindings/node/src/types/SpeechCallOptions.ts
@@ -16,38 +16,38 @@ text: string,
  * The voice to use (provider-specific voice ID, name, or other
  * identifier).
  */
-voice: string | null, 
+voice?: string, 
 /**
  * The desired output format for the audio, e.g. `"mp3"`, `"wav"`.
  */
-output_format: string | null, 
+outputFormat?: string, 
 /**
  * Instructions for the speech generation, e.g.
  * `"Speak in a slow and steady tone"`.
  */
-instructions: string | null, 
+instructions?: string, 
 /**
  * The speed of the speech generation.
  */
-speed: number | null, 
+speed?: number, 
 /**
  * The language for speech generation (ISO 639-1 code, e.g. `"en"`, or
  * `"auto"` for automatic detection). Provider support varies.
  */
-language: string | null, 
+language?: string, 
 /**
  * Additional provider-specific options, keyed by provider name.
  */
-provider_options: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Per-call retry override. `None` uses the model default.
  */
-max_retries: number | null, 
+maxRetries?: number, 
 /**
  * Per-call operation timeout.
  */
-timeout: TimeoutConfiguration | null, 
+timeout?: TimeoutConfiguration, 
 /**
  * Additional HTTP headers to send with the request.
  */
-headers: { [key in string]: string } | null, };
+headers?: { [key in string]: string }, };
diff --git a/bindings/node/src/types/SpeechRequest.ts b/bindings/node/src/types/SpeechRequest.ts
index 1d5fbe382..6131173a4 100644
--- a/bindings/node/src/types/SpeechRequest.ts
+++ b/bindings/node/src/types/SpeechRequest.ts
@@ -8,4 +8,4 @@ export type SpeechRequest = {
 /**
  * Response body (HTTP providers only).
  */
-body: JsonValue | null, };
+body?: JsonValue, };
diff --git a/bindings/node/src/types/SpeechResponse.ts b/bindings/node/src/types/SpeechResponse.ts
index 9d83fe59d..0c447bbe0 100644
--- a/bindings/node/src/types/SpeechResponse.ts
+++ b/bindings/node/src/types/SpeechResponse.ts
@@ -8,16 +8,16 @@ export type SpeechResponse = {
 /**
  * Timestamp for the start of the generated response (ISO 8601 string).
  */
-timestamp: string | null, 
+timestamp?: string, 
 /**
  * The ID of the model that was used to generate the response.
  */
-model_id: string | null, 
+modelId?: string, 
 /**
  * Response headers.
  */
-headers: { [key in string]: string } | null, 
+headers?: { [key in string]: string }, 
 /**
  * Response body (opaque JSON).
  */
-body: JsonValue | null, };
+body?: JsonValue, };
diff --git a/bindings/node/src/types/SpeechResult.ts b/bindings/node/src/types/SpeechResult.ts
index 4dd161ffe..5fe572a8e 100644
--- a/bindings/node/src/types/SpeechResult.ts
+++ b/bindings/node/src/types/SpeechResult.ts
@@ -22,7 +22,7 @@ warnings: Array<Warning>,
 /**
  * Optional request information for telemetry and debugging.
  */
-request: SpeechRequest | null, 
+request?: SpeechRequest, 
 /**
  * Response information for telemetry and debugging.
  */
@@ -30,4 +30,4 @@ response: SpeechResponse,
 /**
  * Additional provider-specific metadata, keyed by provider name.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, };
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/StreamPart.ts b/bindings/node/src/types/StreamPart.ts
index 03c3f862f..2f96fba3b 100644
--- a/bindings/node/src/types/StreamPart.ts
+++ b/bindings/node/src/types/StreamPart.ts
@@ -18,28 +18,28 @@ import type { JsonValue } from "./serde_json/JsonValue";
  * [`StreamPart`] (raw argument text); `stream_text` emits [`TextStreamPart`]
  * (parsed and validated [`ToolCall`]).
  */
-export type StreamPart = { "TextStart": { id: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "TextDelta": { id: string, delta: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "TextEnd": { id: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "StreamStart": { warnings: Array<Warning>, } } | { "Finish": { finish_reason: FinishReason, usage: Usage, provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "Error": { error: AiMuxError, } } | { "ToolInputStart": { id: string, tool_name: string, 
+export type StreamPart = { "type": "text-start", id: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "text-delta", id: string, delta: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "text-end", id: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "stream-start", warnings: Array<Warning>, } | { "type": "finish", finishReason: FinishReason, usage: Usage, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "error", error: AiMuxError, } | { "type": "tool-input-start", id: string, toolName: string, 
 /**
  * Whether the tool call will be executed by the provider.
  */
-provider_executed?: boolean | null, 
+providerExecuted?: boolean, 
 /**
  * Whether the tool is dynamic (defined at runtime, e.g. MCP tools).
  */
-dynamic?: boolean | null, 
+dynamic?: boolean, 
 /**
  * Optional title for the tool call.
  */
-title?: string | null, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolInputDelta": { id: string, delta: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolInputEnd": { id: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolCall": RawToolCall } | { "ToolResult": ToolResult } | { "File": GeneratedFile } | { "ReasoningFile": GeneratedFile } | { "Custom": { kind: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolApprovalRequest": RawToolApprovalRequest } | { "ReasoningStart": { id: string,
+title?: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-input-delta", id: string, delta: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-input-end", id: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-call" } & RawToolCall | { "type": "tool-result" } & ToolResult | { "type": "file" } & GeneratedFile | { "type": "reasoning-file" } & GeneratedFile | { "type": "custom", kind: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-approval-request" } & RawToolApprovalRequest | { "type": "reasoning-start", id: string, 
 /**
  * Provider-specific metadata (e.g. xAI `itemId`).
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ReasoningDelta": { id: string, delta: string, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "reasoning-delta", id: string, delta: string, 
 /**
  * Provider-specific metadata (e.g. xAI `itemId`).
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ReasoningEnd": { id: string, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "reasoning-end", id: string, 
 /**
  * Provider-specific metadata (e.g. xAI `itemId`, `reasoningEncryptedContent`).
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ResponseMetadata": ResponseMetadata } | { "Source": Source } | { "Raw": { raw_value: JsonValue, } };
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "response-metadata" } & ResponseMetadata | { "type": "source" } & Source | { "type": "raw", rawValue: JsonValue, };
diff --git a/bindings/node/src/types/StreamResponseInfo.ts b/bindings/node/src/types/StreamResponseInfo.ts
index 8426efa47..3afa51caf 100644
--- a/bindings/node/src/types/StreamResponseInfo.ts
+++ b/bindings/node/src/types/StreamResponseInfo.ts
@@ -3,4 +3,4 @@
 /**
  * Response information returned when starting a language-model stream.
  */
-export type StreamResponseInfo = { headers: { [key in string]: string } | null, };
+export type StreamResponseInfo = { headers?: { [key in string]: string }, };
diff --git a/bindings/node/src/types/StreamTextResultAggregated.ts b/bindings/node/src/types/StreamTextResultAggregated.ts
index bf8b91600..7192e5db3 100644
--- a/bindings/node/src/types/StreamTextResultAggregated.ts
+++ b/bindings/node/src/types/StreamTextResultAggregated.ts
@@ -19,69 +19,69 @@ import type { JsonValue } from "./serde_json/JsonValue";
  * `GenerateTextResult`'s user-facing fields (without `raw`, since streaming
  * has no `GenerateResult` equivalent).
  */
-export type StreamTextResultAggregated = {
+export type StreamTextResultAggregated = { 
 /**
  * Ordered content from the last generation step.
  */
-content?: Array<GenerateContent<ToolCall, ToolApprovalRequestOutput, ReasoningFileOutput>>,
+content?: Array<GenerateContent<ToolCall, ToolApprovalRequestOutput, ReasoningFileOutput>>, 
 /**
  * The generated text (concatenated `TextDelta`).
  */
-text: string,
+text: string, 
 /**
  * Reasoning / thinking segments.
  */
-reasoning: Array<ReasoningPart>,
+reasoning: Array<ReasoningPart>, 
 /**
  * Concatenated reasoning text.
  */
-reasoning_text: string,
+reasoningText: string, 
 /**
  * Tool calls requested by the model.
  */
-tool_calls: Array<ToolCall>,
+toolCalls: Array<ToolCall>, 
 /**
  * Sources / citations.
  */
-sources: Array<Source>,
+sources: Array<Source>, 
 /**
  * Files generated by the model.
  */
-files: Array<GeneratedFile>,
+files: Array<GeneratedFile>, 
 /**
  * Why generation stopped.
  */
-finish_reason: FinishReason,
+finishReason: FinishReason, 
 /**
  * Raw provider-specific finish reason string.
  */
-raw_finish_reason: string | null,
+rawFinishReason?: string, 
 /**
  * Token usage.
  */
-usage: Usage,
+usage: Usage, 
 /**
  * Total token usage across all steps. In single-step mode (aimux's
  * default), `total_usage` equals `usage`. Provided for AI SDK parity.
  */
-total_usage: Usage,
+totalUsage: Usage, 
 /**
  * Warnings from the provider.
  */
-warnings: Array<Warning>,
+warnings: Array<Warning>, 
 /**
  * Provider-specific metadata from the Finish chunk.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null,
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Request information from the provider.
  */
-request: RequestInfo,
+request: RequestInfo, 
 /**
  * Response metadata and headers from the last step.
  */
-response: ResponseInfo,
+response: ResponseInfo, 
 /**
  * Assistant messages ready to append for the next turn.
  */
-response_messages: Array<ModelMessage>, };
+responseMessages: Array<ModelMessage>, };
diff --git a/bindings/node/src/types/TextContent.ts b/bindings/node/src/types/TextContent.ts
index 09ad96b47..2db52e39b 100644
--- a/bindings/node/src/types/TextContent.ts
+++ b/bindings/node/src/types/TextContent.ts
@@ -10,4 +10,4 @@ import type { ToolCall } from "./ToolCall";
 import type { ToolResult } from "./ToolResult";
 import type { JsonValue } from "./serde_json/JsonValue";
 
-export type TextContent = { "Text": { text: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolCall": ToolCall } | { "Source": Source } | { "Reasoning": ReasoningOutput } | { "File": GeneratedFile } | { "ReasoningFile": ReasoningFileOutput } | { "Custom": { kind: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolApprovalRequest": ToolApprovalRequestOutput } | { "ToolResult": ToolResult };
+export type TextContent = { "type": "text", text: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-call" } & ToolCall | { "type": "source" } & Source | { "type": "reasoning" } & ReasoningOutput | { "type": "file" } & GeneratedFile | { "type": "reasoning-file" } & ReasoningFileOutput | { "type": "custom", kind: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-approval-request" } & ToolApprovalRequestOutput | { "type": "tool-result" } & ToolResult;
diff --git a/bindings/node/src/types/TextPart.ts b/bindings/node/src/types/TextPart.ts
index 2e01804cf..9cc05f537 100644
--- a/bindings/node/src/types/TextPart.ts
+++ b/bindings/node/src/types/TextPart.ts
@@ -4,4 +4,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * Text content part.
  */
-export type TextPart = { text: string, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type TextPart = { text: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/TextStreamPart.ts b/bindings/node/src/types/TextStreamPart.ts
index 7457a4b09..1f2537033 100644
--- a/bindings/node/src/types/TextStreamPart.ts
+++ b/bindings/node/src/types/TextStreamPart.ts
@@ -15,28 +15,28 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * A public call-layer stream chunk, with parsed tool calls.
  */
-export type TextStreamPart = { "TextStart": { id: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "TextDelta": { id: string, delta: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "TextEnd": { id: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "StreamStart": { warnings: Array<Warning>, } } | { "Finish": { finish_reason: FinishReason, usage: Usage, provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "FinishStep": { finish_reason: FinishReason, usage: Usage, provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, response: ResponseInfo, } } | { "Error": { error: AiMuxError, } } | { "ToolInputStart": { id: string, tool_name: string,
+export type TextStreamPart = { "type": "text-start", id: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "text-delta", id: string, delta: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "text-end", id: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "stream-start", warnings: Array<Warning>, } | { "type": "finish", finishReason: FinishReason, usage: Usage, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "finish-step", finishReason: FinishReason, usage: Usage, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, response: ResponseInfo, } | { "type": "error", error: AiMuxError, } | { "type": "tool-input-start", id: string, toolName: string, 
 /**
  * Whether the tool call will be executed by the provider.
  */
-provider_executed?: boolean | null,
+providerExecuted?: boolean, 
 /**
  * Whether the tool is dynamic (defined at runtime, e.g. MCP tools).
  */
-dynamic?: boolean | null,
+dynamic?: boolean, 
 /**
  * Optional title for the tool call.
  */
-title?: string | null, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolInputDelta": { id: string, delta: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolInputEnd": { id: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolCall": ToolCall } | { "ToolResult": ToolResult } | { "File": GeneratedFile } | { "ReasoningFile": ReasoningFileOutput } | { "Custom": { kind: string, provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ToolApprovalRequest": ToolApprovalRequestOutput } | { "ReasoningStart": { id: string,
+title?: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-input-delta", id: string, delta: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-input-end", id: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-call" } & ToolCall | { "type": "tool-result" } & ToolResult | { "type": "file" } & GeneratedFile | { "type": "reasoning-file" } & ReasoningFileOutput | { "type": "custom", kind: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "tool-approval-request" } & ToolApprovalRequestOutput | { "type": "reasoning-start", id: string, 
 /**
  * Provider-specific metadata (e.g. xAI `itemId`).
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ReasoningDelta": { id: string, delta: string,
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "reasoning-delta", id: string, delta: string, 
 /**
  * Provider-specific metadata (e.g. xAI `itemId`).
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ReasoningEnd": { id: string,
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "reasoning-end", id: string, 
 /**
  * Provider-specific metadata (e.g. xAI `itemId`, `reasoningEncryptedContent`).
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "Source": Source } | { "Raw": { raw_value: JsonValue, } };
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "source" } & Source | { "type": "raw", rawValue: JsonValue, };
diff --git a/bindings/node/src/types/TimeoutConfiguration.ts b/bindings/node/src/types/TimeoutConfiguration.ts
index d547f0a0f..086865d32 100644
--- a/bindings/node/src/types/TimeoutConfiguration.ts
+++ b/bindings/node/src/types/TimeoutConfiguration.ts
@@ -13,12 +13,12 @@ export type TimeoutConfiguration = {
  * Overall timeout for the entire call (including retries and, for
  * streaming, the whole stream), in milliseconds.
  */
-total_ms: number | null, 
+totalMs?: number, 
 /**
  * Timeout for one generation step, including that step's attempts and
  * retry backoff, in milliseconds. Aimux currently has one step.
  */
-step_ms: number | null, 
+stepMs?: number, 
 /**
  * Timeout waiting for the first stream chunk (streaming only).
  *
@@ -26,8 +26,8 @@ step_ms: number | null,
  * and any retries before the first semantic output: it is the
  * user-perceived time-to-first-output budget, not a per-attempt timer.
  */
-first_chunk_ms: number | null, 
+firstChunkMs?: number, 
 /**
  * Maximum idle time between consecutive stream chunks (streaming only).
  */
-chunk_ms: number | null, };
+chunkMs?: number, };
diff --git a/bindings/node/src/types/TimingRecord.ts b/bindings/node/src/types/TimingRecord.ts
index b24ae6478..cac3c5e8a 100644
--- a/bindings/node/src/types/TimingRecord.ts
+++ b/bindings/node/src/types/TimingRecord.ts
@@ -4,4 +4,4 @@ export type TimingRecord = {
 /**
  * ms 级延迟,TS 用 number(远小于 2^53,避免 bigint)。
  */
-latency_ms: number, ttfb_ms: number | null, };
+latencyMs: number, ttfbMs?: number, };
diff --git a/bindings/node/src/types/ToolApprovalRequestOutput.ts b/bindings/node/src/types/ToolApprovalRequestOutput.ts
index 8e5d70b34..9b4f34fc1 100644
--- a/bindings/node/src/types/ToolApprovalRequestOutput.ts
+++ b/bindings/node/src/types/ToolApprovalRequestOutput.ts
@@ -4,4 +4,4 @@ import type { ToolCall } from "./ToolCall";
 /**
  * Approval request exposed by text generation.
  */
-export type ToolApprovalRequestOutput = { approval_id: string, tool_call: ToolCall, reason?: string | null, is_automatic?: boolean | null, signature?: string | null, };
+export type ToolApprovalRequestOutput = { approvalId: string, toolCall: ToolCall, reason?: string, isAutomatic?: boolean, signature?: string, };
diff --git a/bindings/node/src/types/ToolApprovalResponsePart.ts b/bindings/node/src/types/ToolApprovalResponsePart.ts
index 12163b325..d624ac2f3 100644
--- a/bindings/node/src/types/ToolApprovalResponsePart.ts
+++ b/bindings/node/src/types/ToolApprovalResponsePart.ts
@@ -4,4 +4,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * The user's decision for a provider-executed tool approval request.
  */
-export type ToolApprovalResponsePart = { approval_id: string, approved: boolean, reason?: string | null, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type ToolApprovalResponsePart = { approvalId: string, approved: boolean, reason?: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/ToolCall.ts b/bindings/node/src/types/ToolCall.ts
index 46500aa89..442cf0c80 100644
--- a/bindings/node/src/types/ToolCall.ts
+++ b/bindings/node/src/types/ToolCall.ts
@@ -9,11 +9,11 @@ export type ToolCall = {
 /**
  * Provider-assigned call id.
  */
-tool_call_id: string, 
+toolCallId: string, 
 /**
  * Tool name.
  */
-tool_name: string, 
+toolName: string, 
 /**
  * Parsed arguments, or the original string when `invalid` is true and the
  * provider input was not valid JSON.
@@ -22,21 +22,21 @@ input: JsonValue,
 /**
  * Whether the tool call will be executed by the provider.
  */
-provider_executed?: boolean | null, 
+providerExecuted?: boolean, 
 /**
  * Whether the tool is dynamic (defined at runtime, e.g. MCP tools).
  */
-dynamic?: boolean | null, 
+dynamic?: boolean, 
 /**
  * Additional provider-specific metadata associated with this call.
  */
-provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Set when lookup, JSON parsing, or schema validation still failed after
  * the optional repair attempt.
  */
-invalid?: boolean | null, 
+invalid?: boolean, 
 /**
  * Typed failure associated with an invalid tool call.
  */
-error?: AiMuxError | null, };
+error?: AiMuxError, };
diff --git a/bindings/node/src/types/ToolCallPart.ts b/bindings/node/src/types/ToolCallPart.ts
index bac25885b..ca46e2b8b 100644
--- a/bindings/node/src/types/ToolCallPart.ts
+++ b/bindings/node/src/types/ToolCallPart.ts
@@ -4,4 +4,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * Tool call content part.
  */
-export type ToolCallPart = { tool_call_id: string, tool_name: string, input: JsonValue, provider_executed?: boolean | null, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type ToolCallPart = { toolCallId: string, toolName: string, input: JsonValue, providerExecuted?: boolean, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/ToolCallRepairReply.ts b/bindings/node/src/types/ToolCallRepairReply.ts
index 6de422082..4208158c3 100644
--- a/bindings/node/src/types/ToolCallRepairReply.ts
+++ b/bindings/node/src/types/ToolCallRepairReply.ts
@@ -7,7 +7,7 @@ import type { RawToolCall } from "./RawToolCall";
  * The bindings cannot carry a [`ToolCallRepair`] closure, so a host instead
  * reads the invalid call off the result, repairs it in its own language, and
  * hands the answer back as this value. Wire shape:
- * `{"type":"repaired","tool_call":{…}}`, `{"type":"unchanged"}`, or
+ * `{"type":"repaired","toolCall":{…}}`, `{"type":"unchanged"}`, or
  * `{"type":"failed","message":"…"}`.
  */
-export type ToolCallRepairReply = { "type": "repaired", tool_call: RawToolCall, } | { "type": "unchanged" } | { "type": "failed", message: string, };
+export type ToolCallRepairReply = { "type": "repaired", toolCall: RawToolCall, } | { "type": "unchanged" } | { "type": "failed", message: string, };
diff --git a/bindings/node/src/types/ToolResult.ts b/bindings/node/src/types/ToolResult.ts
index 544c2e653..6f9839746 100644
--- a/bindings/node/src/types/ToolResult.ts
+++ b/bindings/node/src/types/ToolResult.ts
@@ -9,7 +9,7 @@ export type ToolResult = {
 /**
  * Must match the corresponding call's `tool_call_id`.
  */
-tool_call_id: string, tool_name: string, 
+toolCallId: string, toolName: string, 
 /**
  * The tool's output (usually a JSON-serializable value or plain text).
  */
@@ -17,16 +17,16 @@ result: JsonValue,
 /**
  * Whether the result is an error or error message.
  */
-is_error?: boolean | null, 
+isError?: boolean, 
 /**
  * Whether the result is preliminary (replaces prior, e.g. image previews).
  */
-preliminary?: boolean | null, 
+preliminary?: boolean, 
 /**
  * Whether the tool is dynamic (defined at runtime, e.g. MCP tools).
  */
-dynamic?: boolean | null, 
+dynamic?: boolean, 
 /**
  * Additional provider-specific metadata for the tool result.
  */
-provider_metadata?: { [key in string]: { [key in string]: JsonValue } } | null, };
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/ToolResultContent.ts b/bindings/node/src/types/ToolResultContent.ts
index 3511c045d..742dc5312 100644
--- a/bindings/node/src/types/ToolResultContent.ts
+++ b/bindings/node/src/types/ToolResultContent.ts
@@ -6,4 +6,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * Content within a tool result.
  */
-export type ToolResultContent = { "type": "text" } & TextPart | { "type": "file" } & FilePart | { "type": "custom", provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type ToolResultContent = { "type": "text" } & TextPart | { "type": "file" } & FilePart | { "type": "custom", providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/ToolResultOutput.ts b/bindings/node/src/types/ToolResultOutput.ts
index 2855f2918..f0d7c4672 100644
--- a/bindings/node/src/types/ToolResultOutput.ts
+++ b/bindings/node/src/types/ToolResultOutput.ts
@@ -5,4 +5,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * The provider-facing result of a tool call.
  */
-export type ToolResultOutput = { "type": "text", value: string, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "json", value: JsonValue, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "execution-denied", reason?: string | null, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "error-text", value: string, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "error-json", value: JsonValue, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, } | { "type": "content", value: Array<ToolResultContent>, };
+export type ToolResultOutput = { "type": "text", value: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "json", value: JsonValue, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "execution-denied", reason?: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "error-text", value: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "error-json", value: JsonValue, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "content", value: Array<ToolResultContent>, };
diff --git a/bindings/node/src/types/ToolResultPart.ts b/bindings/node/src/types/ToolResultPart.ts
index 6a479932d..36d87abb1 100644
--- a/bindings/node/src/types/ToolResultPart.ts
+++ b/bindings/node/src/types/ToolResultPart.ts
@@ -5,4 +5,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * Tool result content part.
  */
-export type ToolResultPart = { tool_call_id: string, tool_name: string, output: ToolResultOutput, provider_options?: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type ToolResultPart = { toolCallId: string, toolName: string, output: ToolResultOutput, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/TraceFilter.ts b/bindings/node/src/types/TraceFilter.ts
index 314c21083..59d2b8ec3 100644
--- a/bindings/node/src/types/TraceFilter.ts
+++ b/bindings/node/src/types/TraceFilter.ts
@@ -3,4 +3,4 @@
 /**
  * Query filter (RFC-0015 §5.3).
  */
-export type TraceFilter = { provider: string | null, model: string | null, session_id: string | null, since_unix_ms: bigint | null, };
+export type TraceFilter = { provider?: string, model?: string, sessionId?: string, sinceUnixMs?: bigint, };
diff --git a/bindings/node/src/types/TraceRecord.ts b/bindings/node/src/types/TraceRecord.ts
index fe903fb3f..cfd91ae37 100644
--- a/bindings/node/src/types/TraceRecord.ts
+++ b/bindings/node/src/types/TraceRecord.ts
@@ -7,41 +7,41 @@ import type { Verdict } from "./Verdict";
 /**
  * One probed call. Plaintext never persists — fingerprints only.
  */
-export type TraceRecord = { provider: string, model: string, request_id: string | null, 
+export type TraceRecord = { provider: string, model: string, requestId?: string, 
 /**
  * From `CallOptions.session_id` (RFC-0024) — explicit first, inference
  * fallback; may also be the layer's default session.
  */
-session_id: string | null, 
+sessionId?: string, 
 /**
  * Call-level unique id (association key for Recording / replay,
  * RFC-0023).
  */
-call_id: string, 
+callId: string, 
 /**
  * When the call was sent (epoch ms).
  */
-sent_at_unix_ms: number, 
+sentAtUnixMs: number, 
 /**
  * Client-side LCP token upper bound (block upper bound, byte-proxy
  * len/4). `None` when no history matched. Consumed by `TraceStats`
  * aggregation.
  */
-lcp_token_upper?: bigint | null, 
+lcpTokenUpper?: bigint, 
 /**
  * Time to first streamed token (ms). Non-streaming: `None`.
  */
-ttft_ms: number | null, fingerprint: Fingerprint, usage: UsageSnapshot, 
+ttftMs?: number, fingerprint: Fingerprint, usage: UsageSnapshot, 
 /**
  * Response-side cache headers (e.g. `x-openrouter-cache-status`).
  */
-response_cache_headers?: { [key in string]: string } | null, request_cache_hints?: RequestCacheHints | null, 
+responseCacheHeaders?: { [key in string]: string }, requestCacheHints?: RequestCacheHints, 
 /**
  * Audit verdict (`None` when no auditor is attached — the default).
  */
-verdict?: Verdict | null, 
+verdict?: Verdict, 
 /**
  * Error string when the call failed (still recorded: failures are part
  * of the session).
  */
-error?: string | null, };
+error?: string, };
diff --git a/bindings/node/src/types/TraceStats.ts b/bindings/node/src/types/TraceStats.ts
index 2f836deac..d182f8fd5 100644
--- a/bindings/node/src/types/TraceStats.ts
+++ b/bindings/node/src/types/TraceStats.ts
@@ -4,4 +4,4 @@
  * Aggregated statistics (RFC-0015 §5.2 — the two hit rates are reported
  * side by side, never merged into a single number).
  */
-export type TraceStats = { provider: string, model: string, requests: bigint, input_tokens_total: bigint, claimed_cache_read_total: bigint, claimed_cache_write_total: bigint, reported_hit_rate: number | null, client_upper_bound_hit_rate: number | null, verdict_counts: { [key in string]: bigint }, ttft_p50_ms: bigint | null, ttft_p95_ms: bigint | null, errors: bigint, };
+export type TraceStats = { provider: string, model: string, requests: bigint, inputTokensTotal: bigint, claimedCacheReadTotal: bigint, claimedCacheWriteTotal: bigint, reportedHitRate?: number, clientUpperBoundHitRate?: number, verdictCounts: { [key in string]: bigint }, ttftP50Ms?: bigint, ttftP95Ms?: bigint, errors: bigint, };
diff --git a/bindings/node/src/types/TranscriptionCallOptions.ts b/bindings/node/src/types/TranscriptionCallOptions.ts
index e5e7debdc..bb28dcfcd 100644
--- a/bindings/node/src/types/TranscriptionCallOptions.ts
+++ b/bindings/node/src/types/TranscriptionCallOptions.ts
@@ -16,20 +16,20 @@ audio: AudioInput,
 /**
  * The IANA media type of the audio data, e.g. `"audio/mp3"`.
  */
-media_type: string, 
+mediaType: string, 
 /**
  * Additional provider-specific options, keyed by provider name.
  */
-provider_options: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Per-call retry override. `None` uses the model default.
  */
-max_retries: number | null, 
+maxRetries?: number, 
 /**
  * Per-call operation timeout.
  */
-timeout: TimeoutConfiguration | null, 
+timeout?: TimeoutConfiguration, 
 /**
  * Additional HTTP headers to send with the request.
  */
-headers: { [key in string]: string } | null, };
+headers?: { [key in string]: string }, };
diff --git a/bindings/node/src/types/TranscriptionRequest.ts b/bindings/node/src/types/TranscriptionRequest.ts
index d1dac6010..d20adefc3 100644
--- a/bindings/node/src/types/TranscriptionRequest.ts
+++ b/bindings/node/src/types/TranscriptionRequest.ts
@@ -7,4 +7,4 @@ export type TranscriptionRequest = {
 /**
  * Raw request HTTP body that was sent (JSON stringified).
  */
-body: string | null, };
+body?: string, };
diff --git a/bindings/node/src/types/TranscriptionResponse.ts b/bindings/node/src/types/TranscriptionResponse.ts
index cdf3331b2..7508f88b6 100644
--- a/bindings/node/src/types/TranscriptionResponse.ts
+++ b/bindings/node/src/types/TranscriptionResponse.ts
@@ -8,16 +8,16 @@ export type TranscriptionResponse = {
 /**
  * Timestamp for the start of the generated response (ISO 8601 string).
  */
-timestamp: string | null, 
+timestamp?: string, 
 /**
  * The ID of the model that was used to generate the response.
  */
-model_id: string | null, 
+modelId?: string, 
 /**
  * Response headers.
  */
-headers: { [key in string]: string } | null, 
+headers?: { [key in string]: string }, 
 /**
  * Response body (opaque JSON).
  */
-body: JsonValue | null, };
+body?: JsonValue, };
diff --git a/bindings/node/src/types/TranscriptionResult.ts b/bindings/node/src/types/TranscriptionResult.ts
index 2c5356fc1..8f2cd7f08 100644
--- a/bindings/node/src/types/TranscriptionResult.ts
+++ b/bindings/node/src/types/TranscriptionResult.ts
@@ -22,11 +22,11 @@ segments: Array<TranscriptionSegment>,
 /**
  * The detected language (ISO 639-1 code, e.g. `"en"`), if detected.
  */
-language: string | null, 
+language?: string, 
 /**
  * The total duration of the audio in seconds, if determined.
  */
-duration_in_seconds: number | null, 
+durationInSeconds?: number, 
 /**
  * Warnings for the call, e.g. unsupported settings.
  */
@@ -34,7 +34,7 @@ warnings: Array<Warning>,
 /**
  * Optional request information for telemetry and debugging.
  */
-request: TranscriptionRequest | null, 
+request?: TranscriptionRequest, 
 /**
  * Response information for telemetry and debugging.
  */
@@ -42,4 +42,4 @@ response: TranscriptionResponse,
 /**
  * Additional provider-specific metadata, keyed by provider name.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, };
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/TranscriptionSegment.ts b/bindings/node/src/types/TranscriptionSegment.ts
index dcb5e02a2..1165793e7 100644
--- a/bindings/node/src/types/TranscriptionSegment.ts
+++ b/bindings/node/src/types/TranscriptionSegment.ts
@@ -11,8 +11,8 @@ text: string,
 /**
  * The start time of this segment in seconds.
  */
-start_second: number, 
+startSecond: number, 
 /**
  * The end time of this segment in seconds.
  */
-end_second: number, };
+endSecond: number, };
diff --git a/bindings/node/src/types/TranscriptionStreamPart.ts b/bindings/node/src/types/TranscriptionStreamPart.ts
index ecfa20e24..eb9e00737 100644
--- a/bindings/node/src/types/TranscriptionStreamPart.ts
+++ b/bindings/node/src/types/TranscriptionStreamPart.ts
@@ -8,4 +8,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
  *
  * Aligned with V4 `TranscriptionModelV4StreamPart`.
  */
-export type TranscriptionStreamPart = { "StreamStart": { warnings: Array<Warning>, } } | { "TranscriptDelta": { id: string | null, delta: string, provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "TranscriptPartial": { id: string | null, text: string, start_second: number | null, duration_in_seconds: number | null, channel_index: number | null, provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "TranscriptFinal": { id: string | null, text: string, start_second: number | null, end_second: number | null, channel_index: number | null, provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "ResponseMetadata": { timestamp: string | null, model_id: string | null, headers: { [key in string]: string } | null, body: JsonValue | null, } } | { "Finish": { text: string, segments: Array<TranscriptionSegment>, language: string | null, duration_in_seconds: number | null, provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, } } | { "Raw": { raw_value: JsonValue, } } | { "Error": { error: JsonValue, } };
+export type TranscriptionStreamPart = { "type": "stream-start", warnings: Array<Warning>, } | { "type": "transcript-delta", id?: string, delta: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "transcript-partial", id?: string, text: string, startSecond?: number, durationInSeconds?: number, channelIndex?: number, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "transcript-final", id?: string, text: string, startSecond?: number, endSecond?: number, channelIndex?: number, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "response-metadata", timestamp?: string, modelId?: string, headers?: { [key in string]: string }, body?: JsonValue, } | { "type": "finish", text: string, segments: Array<TranscriptionSegment>, language?: string, durationInSeconds?: number, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, } | { "type": "raw", rawValue: JsonValue, } | { "type": "error", error: JsonValue, };
diff --git a/bindings/node/src/types/UploadFileCallOptions.ts b/bindings/node/src/types/UploadFileCallOptions.ts
index b39d69c19..bbcae7f87 100644
--- a/bindings/node/src/types/UploadFileCallOptions.ts
+++ b/bindings/node/src/types/UploadFileCallOptions.ts
@@ -15,12 +15,12 @@ data: UploadFileData,
 /**
  * The IANA media type of the file, e.g. `"application/pdf"`.
  */
-media_type: string, 
+mediaType: string, 
 /**
  * The filename of the file.
  */
-filename: string | null, 
+filename?: string, 
 /**
  * Additional provider-specific options, keyed by provider name.
  */
-provider_options: { [key in string]: { [key in string]: JsonValue } } | null, };
+providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/UploadFileData.ts b/bindings/node/src/types/UploadFileData.ts
index 6b50bff3d..de6a48ad3 100644
--- a/bindings/node/src/types/UploadFileData.ts
+++ b/bindings/node/src/types/UploadFileData.ts
@@ -8,4 +8,4 @@ import type { FileBytes } from "./FileBytes";
  * `SharedV4FileData` (URLs and provider references are not valid uploads), so
  * this is a dedicated two-variant enum rather than the full [`crate::shared::FileData`].
  */
-export type UploadFileData = { "Data": { data: FileBytes, } } | { "Text": { text: string, } };
+export type UploadFileData = { "type": "data", data: FileBytes, } | { "type": "text", text: string, };
diff --git a/bindings/node/src/types/UploadFileResult.ts b/bindings/node/src/types/UploadFileResult.ts
index 981c04b59..b671ec9ef 100644
--- a/bindings/node/src/types/UploadFileResult.ts
+++ b/bindings/node/src/types/UploadFileResult.ts
@@ -12,20 +12,20 @@ export type UploadFileResult = {
  * A provider reference mapping provider names to provider-specific file
  * identifiers.
  */
-provider_reference: { [key in string]: string }, 
+providerReference: { [key in string]: string }, 
 /**
  * The IANA media type of the uploaded file, if available from the
  * provider.
  */
-media_type: string | null, 
+mediaType?: string, 
 /**
  * The filename of the uploaded file, if available from the provider.
  */
-filename: string | null, 
+filename?: string, 
 /**
  * Additional provider-specific metadata, keyed by provider name.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Warnings from the provider.
  */
diff --git a/bindings/node/src/types/UploadSkillCallOptions.ts b/bindings/node/src/types/UploadSkillCallOptions.ts
index 7d6e3aa38..7ffbc3259 100644
--- a/bindings/node/src/types/UploadSkillCallOptions.ts
+++ b/bindings/node/src/types/UploadSkillCallOptions.ts
@@ -2,4 +2,4 @@
 import type { SkillFile } from "./SkillFile";
 import type { JsonValue } from "./serde_json/JsonValue";
 
-export type UploadSkillCallOptions = { files: Array<SkillFile>, display_title: string | null, provider_options: { [key in string]: { [key in string]: JsonValue } } | null, };
+export type UploadSkillCallOptions = { files: Array<SkillFile>, displayTitle?: string, providerOptions?: { [key in string]: { [key in string]: JsonValue } }, };
diff --git a/bindings/node/src/types/UploadSkillResult.ts b/bindings/node/src/types/UploadSkillResult.ts
index 17a61f981..01340ea2c 100644
--- a/bindings/node/src/types/UploadSkillResult.ts
+++ b/bindings/node/src/types/UploadSkillResult.ts
@@ -2,4 +2,4 @@
 import type { Warning } from "./Warning";
 import type { JsonValue } from "./serde_json/JsonValue";
 
-export type UploadSkillResult = { provider_reference: { [key in string]: string }, display_title: string | null, name: string | null, description: string | null, latest_version: string | null, provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, warnings: Array<Warning>, };
+export type UploadSkillResult = { providerReference: { [key in string]: string }, displayTitle?: string, name?: string, description?: string, latestVersion?: string, providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, warnings: Array<Warning>, };
diff --git a/bindings/node/src/types/Usage.ts b/bindings/node/src/types/Usage.ts
index 385ac2af6..bcde45b69 100644
--- a/bindings/node/src/types/Usage.ts
+++ b/bindings/node/src/types/Usage.ts
@@ -5,8 +5,8 @@ import type { OutputTokenUsage } from "./OutputTokenUsage";
 /**
  * Token usage statistics.
  */
-export type Usage = { input_tokens: InputTokenUsage, output_tokens: OutputTokenUsage, 
+export type Usage = { inputTokens: InputTokenUsage, outputTokens: OutputTokenUsage, 
 /**
  * Raw usage information from the provider (opaque, provider-specific).
  */
-raw?: Record<string, unknown> | null, };
+raw?: Record<string, unknown>, };
diff --git a/bindings/node/src/types/UsageSnapshot.ts b/bindings/node/src/types/UsageSnapshot.ts
index 2db90e806..368e98740 100644
--- a/bindings/node/src/types/UsageSnapshot.ts
+++ b/bindings/node/src/types/UsageSnapshot.ts
@@ -4,9 +4,9 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * Token usage snapshot (7 flat fields + raw passthrough).
  */
-export type UsageSnapshot = { input_total: bigint | null, input_no_cache: bigint | null, cache_read: bigint | null, cache_write: bigint | null, output_total: bigint | null, output_text: bigint | null, output_reasoning: bigint | null, 
+export type UsageSnapshot = { inputTotal?: bigint, inputNoCache?: bigint, cacheRead?: bigint, cacheWrite?: bigint, outputTotal?: bigint, outputText?: bigint, outputReasoning?: bigint, 
 /**
  * Raw provider usage payload (opaque passthrough; numbers only — no
  * prompt text).
  */
-raw?: JsonValue | null, };
+raw?: JsonValue, };
diff --git a/bindings/node/src/types/Verdict.ts b/bindings/node/src/types/Verdict.ts
index 14b282efa..f48c0dedf 100644
--- a/bindings/node/src/types/Verdict.ts
+++ b/bindings/node/src/types/Verdict.ts
@@ -13,8 +13,8 @@ violated: Array<string>,
 /**
  * Expected upper bound U (tokens).
  */
-expected_max: bigint, claimed: bigint, 
+expectedMax: bigint, claimed: bigint, 
 /**
  * Client-side LCP (bytes).
  */
-lcp_bytes: bigint, notes: Array<string>, };
+lcpBytes: bigint, notes: Array<string>, };
diff --git a/bindings/node/src/types/VideoCallOptions.ts b/bindings/node/src/types/VideoCallOptions.ts
index 9f3791316..63235c733 100644
--- a/bindings/node/src/types/VideoCallOptions.ts
+++ b/bindings/node/src/types/VideoCallOptions.ts
@@ -16,7 +16,7 @@ export type VideoCallOptions = {
 /**
  * Text prompt for the video generation. `None` when not required.
  */
-prompt: string | null, 
+prompt?: string, 
 /**
  * Number of videos to generate. Default `1`; most models only support
  * `n = 1` due to computational cost.
@@ -25,57 +25,57 @@ n: number,
 /**
  * Aspect ratio, in `{width}:{height}` format (e.g. `"16:9"`).
  */
-aspect_ratio: AspectRatio | null, 
+aspectRatio?: AspectRatio, 
 /**
  * Resolution, in `{width}x{height}` format (e.g. `"1280x720"`).
  */
-resolution: Size | null, 
+resolution?: Size, 
 /**
  * Duration of the video in seconds. Typically 3–10 seconds.
  */
-duration: number | null, 
+duration?: number, 
 /**
  * Frames per second. Common values: 24, 30, 60.
  */
-fps: number | null, 
+fps?: number, 
 /**
  * Seed for deterministic generation. `None` uses a random seed.
  */
-seed: number | null, 
+seed?: number, 
 /**
  * Input image for image-to-video generation (the starting frame).
  */
-image: VideoFile | null, 
+image?: VideoFile, 
 /**
  * Role-tagged image inputs for first-last-frame generation.
  */
-frame_images: Array<VideoFrameImage> | null, 
+frameImages?: Array<VideoFrameImage>, 
 /**
  * Reference inputs for reference-to-video generation (images or videos).
  */
-input_references: Array<VideoFile> | null, 
+inputReferences?: Array<VideoFile>, 
 /**
  * Whether the model should generate audio alongside the video.
  */
-generate_audio: boolean | null, 
+generateAudio?: boolean, 
 /**
  * Additional provider-specific options, keyed by provider name.
  */
-provider_options: { [key in string]: { [key in string]: JsonValue } }, 
+providerOptions: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Per-call retry override. `None` uses the model default.
  */
-max_retries: number | null, 
+maxRetries?: number, 
 /**
  * Per-call poll pacing override for the start/status flow. Unset fields
  * fall back to the model's [`VideoModel::poll_config`].
  */
-poll: VideoPollOptions | null, 
+poll?: VideoPollOptions, 
 /**
  * Per-call operation timeout.
  */
-timeout: TimeoutConfiguration | null, 
+timeout?: TimeoutConfiguration, 
 /**
  * Additional HTTP headers to send with the request.
  */
-headers: { [key in string]: string } | null, };
+headers?: { [key in string]: string }, };
diff --git a/bindings/node/src/types/VideoData.ts b/bindings/node/src/types/VideoData.ts
index 443d5449b..5af500ef5 100644
--- a/bindings/node/src/types/VideoData.ts
+++ b/bindings/node/src/types/VideoData.ts
@@ -6,4 +6,4 @@
  * Aligned with V4 `VideoModelV4VideoData`. Most providers return URLs due to
  * large file sizes.
  */
-export type VideoData = { "Url": { url: string, media_type: string, } } | { "Base64": { data: string, media_type: string, } } | { "Binary": { data: Array<number>, media_type: string, } };
+export type VideoData = { "type": "url", url: string, mediaType: string, } | { "type": "base64", data: string, mediaType: string, } | { "type": "binary", data: Array<number>, mediaType: string, };
diff --git a/bindings/node/src/types/VideoFile.ts b/bindings/node/src/types/VideoFile.ts
index 9d32e3648..b74cd2ef2 100644
--- a/bindings/node/src/types/VideoFile.ts
+++ b/bindings/node/src/types/VideoFile.ts
@@ -6,15 +6,15 @@ import type { VideoFileData } from "./VideoFileData";
  *
  * Aligned with V4 `VideoModelV4File`.
  */
-export type VideoFile = { "File": { 
+export type VideoFile = { "type": "file", 
 /**
  * IANA media type, e.g. `"video/mp4"` or `"image/png"`.
  */
-media_type: string, 
+mediaType: string, 
 /**
  * File data as a base64 string or binary bytes.
  */
-data: VideoFileData, } } | { "Url": { 
+data: VideoFileData, } | { "type": "url", 
 /**
  * The URL of the video or image file.
  */
@@ -22,4 +22,4 @@ url: string,
 /**
  * The media type of the referenced file, when known.
  */
-media_type: string | null, } };
+mediaType?: string, };
diff --git a/bindings/node/src/types/VideoFileData.ts b/bindings/node/src/types/VideoFileData.ts
index d7b1d6a6b..dec5fdba2 100644
--- a/bindings/node/src/types/VideoFileData.ts
+++ b/bindings/node/src/types/VideoFileData.ts
@@ -3,4 +3,4 @@
 /**
  * File payload for a [`VideoFile::File`]: base64 string or raw bytes.
  */
-export type VideoFileData = { "Base64": string } | { "Binary": Array<number> };
+export type VideoFileData = string | Array<number>;
diff --git a/bindings/node/src/types/VideoFrameImage.ts b/bindings/node/src/types/VideoFrameImage.ts
index de49fcb3d..070d96b90 100644
--- a/bindings/node/src/types/VideoFrameImage.ts
+++ b/bindings/node/src/types/VideoFrameImage.ts
@@ -15,4 +15,4 @@ image: VideoFile,
 /**
  * Which frame this image represents.
  */
-frame_type: VideoFrameType, };
+frameType: VideoFrameType, };
diff --git a/bindings/node/src/types/VideoFrameType.ts b/bindings/node/src/types/VideoFrameType.ts
index 4279f6dfe..fb10478ee 100644
--- a/bindings/node/src/types/VideoFrameType.ts
+++ b/bindings/node/src/types/VideoFrameType.ts
@@ -5,4 +5,4 @@
  *
  * Aligned with V4 `VideoModelV4FrameType`.
  */
-export type VideoFrameType = "FirstFrame" | "LastFrame";
+export type VideoFrameType = "first_frame" | "last_frame";
diff --git a/bindings/node/src/types/VideoPollOptions.ts b/bindings/node/src/types/VideoPollOptions.ts
index 5950d03ed..5fbf42814 100644
--- a/bindings/node/src/types/VideoPollOptions.ts
+++ b/bindings/node/src/types/VideoPollOptions.ts
@@ -7,8 +7,8 @@ export type VideoPollOptions = {
 /**
  * Delay between consecutive status checks, in milliseconds.
  */
-interval_ms: number | null, 
+intervalMs?: number, 
 /**
  * Maximum total time to wait for completion, in milliseconds.
  */
-timeout_ms: number | null, };
+timeoutMs?: number, };
diff --git a/bindings/node/src/types/VideoResponse.ts b/bindings/node/src/types/VideoResponse.ts
index 9e7fff9cb..e7721fdab 100644
--- a/bindings/node/src/types/VideoResponse.ts
+++ b/bindings/node/src/types/VideoResponse.ts
@@ -7,12 +7,12 @@ export type VideoResponse = {
 /**
  * Timestamp for the start of the generated response (ISO 8601 string).
  */
-timestamp: string | null, 
+timestamp?: string, 
 /**
  * The ID of the model that was used to generate the response.
  */
-model_id: string | null, 
+modelId?: string, 
 /**
  * Response headers.
  */
-headers: { [key in string]: string } | null, };
+headers?: { [key in string]: string }, };
diff --git a/bindings/node/src/types/VideoResult.ts b/bindings/node/src/types/VideoResult.ts
index 8ab4fec8f..df032c3e0 100644
--- a/bindings/node/src/types/VideoResult.ts
+++ b/bindings/node/src/types/VideoResult.ts
@@ -21,7 +21,7 @@ warnings: Array<Warning>,
 /**
  * Additional provider-specific metadata, keyed by provider name.
  */
-provider_metadata: { [key in string]: { [key in string]: JsonValue } } | null, 
+providerMetadata?: { [key in string]: { [key in string]: JsonValue } }, 
 /**
  * Response information for telemetry and debugging.
  */
diff --git a/bindings/node/src/types/Warning.ts b/bindings/node/src/types/Warning.ts
index 14edc447c..23e1093ae 100644
--- a/bindings/node/src/types/Warning.ts
+++ b/bindings/node/src/types/Warning.ts
@@ -3,4 +3,4 @@
 /**
  * A warning issued by the provider (e.g. unsupported parameter).
  */
-export type Warning = { "Unsupported": { feature: string, details: string | null, } } | { "Compatibility": { feature: string, details: string | null, } } | { "Deprecated": { setting: string, message: string, } } | { "Other": { message: string, } };
+export type Warning = { "type": "unsupported", feature: string, details?: string, } | { "type": "compatibility", feature: string, details?: string, } | { "type": "deprecated", setting: string, message: string, } | { "type": "other", message: string, };
diff --git a/bindings/python/python/aimux/__init__.py b/bindings/python/python/aimux/__init__.py
index 1709a57ad..0e0494967 100644
--- a/bindings/python/python/aimux/__init__.py
+++ b/bindings/python/python/aimux/__init__.py
@@ -270,12 +270,12 @@ def _prompt_to_json(prompt: Union[str, List[Dict[str, Any]]]) -> str:
     return json.dumps({"prompt": prompt})
 
 
-#: The ``options["repair_tool_call"]`` function (RFC-0035), the equivalent of
+#: The ``options["repairToolCall"]`` function (RFC-0035), the equivalent of
 #: the AI SDK ``repairToolCall``.
 #:
 #: It is called once per tool call that came back ``invalid``, with the repair
-#: context ``{tool_call, error, input_schema, tools, messages, instructions}``
-#: — ``tool_call["input"]`` being the provider's raw argument text. Return a
+#: context ``{toolCall, error, inputSchema, tools, messages, instructions}``
+#: — ``toolCall["input"]`` being the provider's raw argument text. Return a
 #: replacement raw tool call (same shape, ``input`` a string) to re-validate,
 #: or ``None`` to leave the call invalid. Raising marks the repair failed: the
 #: call keeps its original error with the exception message as the cause.
@@ -291,16 +291,16 @@ def _prompt_to_json(prompt: Union[str, List[Dict[str, Any]]]) -> str:
 def _opts_to_json(options: Optional[Dict[str, Any]]) -> Optional[str]:
     """Convert options dict to JSON string.
 
-    ``repair_tool_call`` is a Python callable and stays on this side of the
+    ``repairToolCall`` is a Python callable and stays on this side of the
     boundary; everything else is passed through to the native layer.
     """
     if options is None:
         return None
-    return json.dumps({k: v for k, v in options.items() if k != "repair_tool_call"})
+    return json.dumps({k: v for k, v in options.items() if k != "repairToolCall"})
 
 
 def _repair_fn(options: Optional[Dict[str, Any]]) -> Optional[_repair.RepairAdapter]:
-    fn = options.get("repair_tool_call") if options else None
+    fn = options.get("repairToolCall") if options else None
     if fn is None:
         return None
     return lambda context: _repair.run_hook(fn, context)
@@ -319,7 +319,7 @@ def generate_text(
         options: Optional generation options.
 
     Returns:
-        Dict with keys: text, tool_calls, finish_reason, usage, warnings, raw.
+        Dict with keys: text, toolCalls, finishReason, usage, warnings, raw.
     """
     prompt_json = _prompt_to_json(prompt)
     opts_json = _opts_to_json(options)
@@ -338,17 +338,17 @@ def generate_object(
     """Generate a structured JSON object from the model (M12, RFC-0016).
 
     Same signature as ``generate_text``. Pass
-    ``response_format: {"Json": {...}}`` via options for schema control; the
+    ``responseFormat: {"type": "json", ...}`` via options for schema control; the
     function applies JSON repair before parsing.
 
     Args:
         model: A model instance from openai(), anthropic(), etc.
         prompt: A string or a list of message dicts.
-        options: Optional generation options (incl. response_format).
+        options: Optional generation options (incl. responseFormat).
 
     Returns:
-        Dict with keys: object, finish_reason, raw_finish_reason, usage,
-        warnings, reasoning, provider_metadata, response, raw.
+        Dict with keys: object, finishReason, rawFinishReason, usage,
+        warnings, reasoning, providerMetadata, response, raw.
     """
     prompt_json = _prompt_to_json(prompt)
     opts_json = _opts_to_json(options)
@@ -392,12 +392,12 @@ def stream_text(
     prompt: Union[str, List[Dict[str, Any]]],
     options: Optional[Dict[str, Any]] = None,
 ):
-    """Stream text from a model. Yields StreamPart dicts.
+    """Stream text from a model. Yields StreamPart dicts (tagged by ``type``).
 
     Usage:
         for part in stream_text(model, "Write a haiku about Rust."):
-            if "TextDelta" in part:
-                print(part["TextDelta"]["delta"], end="")
+            if part["type"] == "text-delta":
+                print(part["delta"], end="")
     """
     prompt_json = _prompt_to_json(prompt)
     opts_json = _opts_to_json(options)
@@ -448,7 +448,7 @@ def stream_text_as_openai(
     ``chat.completion.chunk`` object. Works with any provider.
 
     Stream options (include_usage, include_reasoning) are read from
-    options["provider_options"]["openai"]["stream_options"] (both default True).
+    options["providerOptions"]["openai"]["stream_options"] (both default True).
 
     Usage:
         for chunk in stream_text_as_openai(model, "Write a haiku."):
diff --git a/bindings/python/python/aimux/_repair.py b/bindings/python/python/aimux/_repair.py
index b8ed800a4..7ec2b7bdd 100644
--- a/bindings/python/python/aimux/_repair.py
+++ b/bindings/python/python/aimux/_repair.py
@@ -50,7 +50,7 @@ def _reply(repair: RepairAdapter, context: Dict[str, Any]) -> str:
         return json.dumps({"type": "failed", "message": str(exc)})
     if replacement is None:
         return json.dumps({"type": "unchanged"})
-    return json.dumps({"type": "repaired", "tool_call": replacement})
+    return json.dumps({"type": "repaired", "toolCall": replacement})
 
 
 def _context(call: Dict[str, Any], prompt_json: str, opts_json: Optional[str]):
@@ -66,14 +66,14 @@ def repair_result(
 ) -> str:
     """Repair every invalid tool call of a serialized result, in order.
 
-    Handles ``generate_text`` / ``generate_object`` (``raw.tool_calls``) and the
+    Handles ``generate_text`` / ``generate_object`` (``raw.toolCalls``) and the
     aggregated stream result alike; one repair attempt per call, as in the
     AI SDK. Returns the patched result JSON.
     """
     result = json.loads(result_json)
-    calls = result.get("tool_calls")
+    calls = result.get("toolCalls")
     if calls is None:
-        calls = result.get("raw", {}).get("tool_calls") or []
+        calls = result.get("raw", {}).get("toolCalls") or []
     for call in calls:
         if not call.get("invalid"):
             continue
@@ -81,7 +81,7 @@ def repair_result(
         if context is None:
             continue
         result_json = apply_tool_call_repair_to_result(
-            result_json, opts_json, call["tool_call_id"], _reply(repair, context)
+            result_json, opts_json, call["toolCallId"], _reply(repair, context)
         )
     return result_json
 
@@ -110,14 +110,13 @@ def repair_stream_part(
     opts_json: Optional[str],
     repair: RepairAdapter,
 ) -> str:
-    """Replace an invalid ``ToolCall`` stream part with its repaired form.
+    """Replace an invalid ``tool-call`` stream part with its repaired form.
 
     Every other part — tool-input deltas included — is forwarded untouched, so
     repair never delays the stream.
     """
-    part = json.loads(part_json)
-    call = part.get("ToolCall") if isinstance(part, dict) else None
-    if call is None or not call.get("invalid"):
+    call = json.loads(part_json)
+    if call.get("type") != "tool-call" or not call.get("invalid"):
         return part_json
     context = _context(call, prompt_json, opts_json)
     if context is None:
@@ -125,4 +124,4 @@ def repair_stream_part(
     repaired = apply_tool_call_repair(
         json.dumps(call), opts_json, _reply(repair, context)
     )
-    return json.dumps({"ToolCall": json.loads(repaired)})
+    return json.dumps({"type": "tool-call", **json.loads(repaired)})
diff --git a/bindings/python/python/aimux/wrapper.py b/bindings/python/python/aimux/wrapper.py
index e2e42d3d9..d5ddcc289 100644
--- a/bindings/python/python/aimux/wrapper.py
+++ b/bindings/python/python/aimux/wrapper.py
@@ -12,9 +12,11 @@
 
 The Pydantic models mirror the Rust core types in ``aimux-core/src/`` (see
 ``bindings/node/src/types/*.ts`` for the field structure generated by ts-rs).
-Field names are ``snake_case`` to match the Rust serde wire format; the few
-camelCase exceptions required by the wire format (``toolName``) are handled
-internally by custom serializers.
+The wire format is the AI SDK's: JSON keys are camelCase and unions are
+discriminated by ``type`` (kebab-case values). Python attributes stay
+``snake_case``; every model maps them to the camelCase JSON keys through an
+alias generator, so both ``tool_call_id`` and ``toolCallId`` validate and
+serialization uses the camelCase form.
 """
 
 from __future__ import annotations
@@ -22,7 +24,8 @@
 import json
 from typing import Annotated, Any, Callable, Dict, Iterator, List, Literal, Optional, Union
 
-from pydantic import BaseModel, ConfigDict, Field, RootModel, model_serializer, model_validator
+from pydantic import BaseModel, ConfigDict, Field, TypeAdapter
+from pydantic.alias_generators import to_camel
 
 from . import _repair
 from .aimux import Model
@@ -91,6 +94,16 @@
 ]
 
 
+class _Wire(BaseModel):
+    """Base of every aimux model: ``snake_case`` attributes, camelCase JSON keys."""
+
+    model_config = ConfigDict(alias_generator=to_camel, populate_by_name=True)
+
+
+class _Meta(_Wire):
+    provider_metadata: Optional[Dict[str, Any]] = None
+
+
 # ─────────────────────────────────────────────────────────────────────────────
 # String-literal type aliases (mirror serde rename rules)
 # ─────────────────────────────────────────────────────────────────────────────
@@ -108,10 +121,10 @@
     "provider-default", "none", "minimal", "low", "medium", "high", "xhigh"
 ]
 
-#: Opaque provider error value (Rust ``AiMuxError``). A mixed newtype/struct
-#: externally-tagged enum, so it is left as an arbitrary JSON value rather than
-#: a discriminated union. Named ``AiMuxErrorValue`` to avoid confusion with the
-#: raised ``AimuxError`` exception.
+#: Opaque provider error value (Rust ``AiMuxError``): an object discriminated
+#: by ``name`` (``"AI_APICallError"`` …), left as an arbitrary JSON value.
+#: Named ``AiMuxErrorValue`` to avoid confusion with the raised ``AimuxError``
+#: exception.
 AiMuxErrorValue = Any
 
 
@@ -119,20 +132,20 @@
 # Simple struct models
 # ─────────────────────────────────────────────────────────────────────────────
 
-class InputTokenUsage(BaseModel):
+class InputTokenUsage(_Wire):
     total: Optional[int] = None
     no_cache: Optional[int] = None
     cache_read: Optional[int] = None
     cache_write: Optional[int] = None
 
 
-class OutputTokenUsage(BaseModel):
+class OutputTokenUsage(_Wire):
     total: Optional[int] = None
     text: Optional[int] = None
     reasoning: Optional[int] = None
 
 
-class Usage(BaseModel):
+class Usage(_Wire):
     """Token usage statistics (Rust ``Usage``)."""
 
     input_tokens: InputTokenUsage = Field(default_factory=InputTokenUsage)
@@ -140,14 +153,14 @@ class Usage(BaseModel):
     raw: Optional[Dict[str, Any]] = None
 
 
-class FinishReason(BaseModel):
+class FinishReason(_Wire):
     """Why generation stopped (Rust ``FinishReason``)."""
 
     unified: FinishReasonUnified
     raw: Optional[str] = None
 
 
-class ResponseMetadata(BaseModel):
+class ResponseMetadata(_Wire):
     """Metadata about the API response (Rust ``ResponseMetadata``)."""
 
     id: Optional[str] = None
@@ -156,11 +169,13 @@ class ResponseMetadata(BaseModel):
 
 
 class GenerateResponseMetadata(ResponseMetadata):
+    """Response metadata with headers and body (Rust ``ResponseInfo``)."""
+
     headers: Optional[Dict[str, str]] = None
     body: Optional[Any] = None
 
 
-class ToolCall(BaseModel):
+class ToolCall(_Meta):
     """A tool call requested by the model (Rust ``tool::ToolCall``)."""
 
     tool_call_id: str
@@ -168,12 +183,11 @@ class ToolCall(BaseModel):
     input: Any
     provider_executed: Optional[bool] = None
     dynamic: Optional[bool] = None
-    provider_metadata: Optional[Any] = None
     invalid: Optional[bool] = None
     error: Optional[AiMuxErrorValue] = None
 
 
-class RawToolCall(BaseModel):
+class RawToolCall(_Meta):
     """A tool call as the provider emitted it (Rust ``RawToolCall``).
 
     Unlike :class:`ToolCall`, ``input`` is the **raw argument text** — the
@@ -186,256 +200,114 @@ class RawToolCall(BaseModel):
     input: str
     provider_executed: Optional[bool] = None
     dynamic: Optional[bool] = None
-    provider_metadata: Optional[Any] = None
 
 
 # ─────────────────────────────────────────────────────────────────────────────
-# External-tag enums (Rust serde default for enums)
-#
-# Rust serializes these as ``{"Variant": {fields}}``. Pydantic discriminated
-# unions use an *internal* tag, so each enum is modelled as a ``RootModel``
-# over a discriminated union with a ``type`` field, plus a before-validator
-# that converts the external tag into the internal discriminator and a wrap
-# serializer that converts it back.
+# Unions: internally tagged by ``type`` (kebab-case), as in the AI SDK.
 # ─────────────────────────────────────────────────────────────────────────────
 
-def _external_tag_before(data: Any) -> Any:
-    """Convert ``{"Variant": {payload}}`` → ``{"type": "Variant", **payload}``.
-
-    Passes other shapes through unchanged (so already-internal-tagged data and
-    scalars still validate).
-    """
-    if isinstance(data, dict) and len(data) == 1:
-        (tag, payload), = data.items()
-        if isinstance(payload, dict):
-            return {"type": tag, **payload}
-    return data
-
-
-def _external_tag_serializer(out: Any) -> Any:
-    """Convert ``{"type": "Variant", **fields}`` → ``{"Variant": {fields}}``."""
-    if isinstance(out, dict) and "type" in out:
-        tag = out.pop("type")
-        return {tag: out}
-    return out
-
-
 # ── Warning ──────────────────────────────────────────────────────────────────
 
-class _WarningUnsupported(BaseModel):
-    type: Literal["Unsupported"]
+class _WarningUnsupported(_Wire):
+    type: Literal["unsupported"]
     feature: str
     details: Optional[str] = None
 
 
-class _WarningCompatibility(BaseModel):
-    type: Literal["Compatibility"]
+class _WarningCompatibility(_Wire):
+    type: Literal["compatibility"]
     feature: str
     details: Optional[str] = None
 
 
-class _WarningDeprecated(BaseModel):
-    type: Literal["Deprecated"]
+class _WarningDeprecated(_Wire):
+    type: Literal["deprecated"]
     setting: str
     message: str
 
 
-class _WarningOther(BaseModel):
-    type: Literal["Other"]
+class _WarningOther(_Wire):
+    type: Literal["other"]
     message: str
 
 
-_WarningUnion = Annotated[
-    Union[
-        _WarningUnsupported,
-        _WarningCompatibility,
-        _WarningDeprecated,
-        _WarningOther,
-    ],
+#: A provider warning (Rust ``Warning``).
+Warning = Annotated[
+    Union[_WarningUnsupported, _WarningCompatibility, _WarningDeprecated, _WarningOther],
     Field(discriminator="type"),
 ]
 
 
-class Warning(RootModel[_WarningUnion]):
-    """A provider warning (Rust ``Warning``)."""
-
-    root: _WarningUnion
-
-    @model_validator(mode="before")
-    @classmethod
-    def _ext_to_int(cls, data: Any) -> Any:
-        return _external_tag_before(data)
-
-    @model_serializer(mode="wrap")
-    def _int_to_ext(self, handler):  # type: ignore[no-untyped-def]
-        return _external_tag_serializer(handler(self))
-
-
 # ── FileBytes / FileData ────────────────────────────────────────────────────
-#
-# ``FileBytes`` and ``FileData`` are externally tagged Rust enums.
-# ``FileData``'s variants are struct variants (dict payloads), so it reuses the
-# shared ``_external_tag_before`` / ``_external_tag_serializer`` helpers.
-# ``FileBytes`` are newtype variants whose payload is *not* a dict
-# (``{"Binary": [1, 2, 3]}`` / ``{"Base64": "..."}``), so it has a dedicated
-# before-validator / serializer.
-
-class _FileBytesBinary(BaseModel):
-    type: Literal["Binary"]
-    data: List[int]
 
+#: Raw bytes (a JSON number array) or a base64 string (Rust ``FileBytes``).
+FileBytes = Union[List[int], str]
 
-class _FileBytesBase64(BaseModel):
-    type: Literal["Base64"]
-    data: str
 
-
-_FileBytesUnion = Annotated[
-    Union[
-        _FileBytesBinary,
-        _FileBytesBase64,
-    ],
-    Field(discriminator="type"),
-]
-
-
-class FileBytes(RootModel[_FileBytesUnion]):
-    """Raw bytes or a base64-encoded string (Rust ``FileBytes``).
-
-    Wire form: ``{"Binary": [1, 2, 3]}`` / ``{"Base64": "..."}``.
-    """
-
-    root: _FileBytesUnion
-
-    @model_validator(mode="before")
-    @classmethod
-    def _ext_to_int(cls, data: Any) -> Any:
-        # {"Binary": [1, 2, 3]} → {"type": "Binary", "data": [1, 2, 3]}
-        # {"Base64": "x"}         → {"type": "Base64", "data": "x"}
-        if isinstance(data, dict) and len(data) == 1:
-            (tag, payload), = data.items()
-            return {"type": tag, "data": payload}
-        return data
-
-    @model_serializer(mode="wrap")
-    def _int_to_ext(self, handler):  # type: ignore[no-untyped-def]
-        out = handler(self)
-        if isinstance(out, dict) and "type" in out and "data" in out:
-            return {out["type"]: out["data"]}
-        return out
-
-
-class _FileDataData(BaseModel):
-    type: Literal["Data"]
+class _FileDataData(_Wire):
+    type: Literal["data"]
     data: FileBytes
 
 
-class _FileDataUrl(BaseModel):
-    type: Literal["Url"]
+class _FileDataUrl(_Wire):
+    type: Literal["url"]
     url: str
     original_url: Optional[str] = None
 
 
-class _FileDataReference(BaseModel):
-    type: Literal["Reference"]
+class _FileDataReference(_Wire):
+    type: Literal["reference"]
     reference: Dict[str, str]
 
 
-class _FileDataText(BaseModel):
-    type: Literal["Text"]
+class _FileDataText(_Wire):
+    type: Literal["text"]
     text: str
 
 
-_FileDataUnion = Annotated[
-    Union[
-        _FileDataData,
-        _FileDataUrl,
-        _FileDataReference,
-        _FileDataText,
-    ],
+#: File data (Rust ``FileData``).
+FileData = Annotated[
+    Union[_FileDataData, _FileDataUrl, _FileDataReference, _FileDataText],
     Field(discriminator="type"),
 ]
 
-
-class FileData(RootModel[_FileDataUnion]):
-    """File data as a tagged union (Rust ``FileData``).
-
-    Wire form: ``{"Data": {"data": <FileBytes>}}``, ``{"Url": {"url": "..."}}``,
-    ``{"Reference": {"reference": {...}}}``, ``{"Text": {"text": "..."}}``.
-    """
-
-    root: _FileDataUnion
-
-    @model_validator(mode="before")
-    @classmethod
-    def _ext_to_int(cls, data: Any) -> Any:
-        return _external_tag_before(data)
-
-    @model_serializer(mode="wrap")
-    def _int_to_ext(self, handler):  # type: ignore[no-untyped-def]
-        return _external_tag_serializer(handler(self))
-
-
-_GeneratedFileDataUnion = Annotated[
+#: Data or a URL returned for a generated file (Rust ``GeneratedFileData``).
+GeneratedFileData = Annotated[
     Union[_FileDataData, _FileDataUrl], Field(discriminator="type")
 ]
 
 
-class GeneratedFileData(RootModel[_GeneratedFileDataUnion]):
-    @model_validator(mode="before")
-    @classmethod
-    def _ext_to_int(cls, data: Any) -> Any:
-        return _external_tag_before(data)
-
-    @model_serializer(mode="wrap")
-    def _int_to_ext(self, handler):
-        return _external_tag_serializer(handler(self))
-
-
-# ── GenerateContent ─────────────────────────────────────────────────────────
-
-class _ContentText(BaseModel):
-    type: Literal["Text"]
-    text: str
-    provider_metadata: Optional[Dict[str, Any]] = None
-
+class GeneratedFile(_Meta):
+    data: GeneratedFileData
+    media_type: str
 
-class _ContentToolCall(BaseModel):
-    type: Literal["ToolCall"]
-    tool_call_id: str
-    tool_name: str
-    input: Any
-    provider_executed: Optional[bool] = None
-    dynamic: Optional[bool] = None
-    provider_metadata: Optional[Any] = None
 
+# ── Content / stream parts ──────────────────────────────────────────────────
 
-class _UrlSource(BaseModel):
+class _UrlSource(_Meta):
     id: str
     source_type: Literal["url"]
     url: str
     title: Optional[str] = None
-    provider_metadata: Optional[Dict[str, Any]] = None
 
 
-class _DocumentSource(BaseModel):
+class _DocumentSource(_Meta):
     id: str
     source_type: Literal["document"]
     media_type: str
     title: str
     filename: Optional[str] = None
-    provider_metadata: Optional[Dict[str, Any]] = None
 
 
 Source = Annotated[Union[_UrlSource, _DocumentSource], Field(discriminator="source_type")]
 
 
 class _ContentUrlSource(_UrlSource):
-    type: Literal["Source"]
+    type: Literal["source"]
 
 
 class _ContentDocumentSource(_DocumentSource):
-    type: Literal["Source"]
+    type: Literal["source"]
 
 
 _ContentSource = Annotated[
@@ -443,334 +315,236 @@ class _ContentDocumentSource(_DocumentSource):
 ]
 
 
-class _ContentReasoning(BaseModel):
-    type: Literal["Reasoning"]
+class _Text(_Meta):
+    type: Literal["text"]
     text: str
-    provider_metadata: Optional[Any] = None
 
 
-class _ContentFile(BaseModel):
-    type: Literal["File", "ReasoningFile"]
-    data: GeneratedFileData
-    media_type: str
-    provider_metadata: Optional[Dict[str, Any]] = None
+class _Reasoning(_Meta):
+    type: Literal["reasoning"]
+    text: str
+
+
+class _Custom(_Meta):
+    type: Literal["custom"]
+    kind: str
+
+
+class _File(GeneratedFile):
+    type: Literal["file"]
+
+
+class _ReasoningFile(GeneratedFile):
+    type: Literal["reasoning-file"]
+
+
+class _ReasoningFileOutput(_Meta):
+    type: Literal["reasoning-file"]
+    file: GeneratedFile
+
+
+class _RawToolCallPart(RawToolCall):
+    type: Literal["tool-call"]
+
+
+class _ToolCallPart(ToolCall):
+    type: Literal["tool-call"]
 
 
-class _ContentToolResult(BaseModel):
-    type: Literal["ToolResult"]
+class _ToolResultPart(_Meta):
+    type: Literal["tool-result"]
     tool_call_id: str
     tool_name: str
     result: Any
     is_error: Optional[bool] = None
     preliminary: Optional[bool] = None
     dynamic: Optional[bool] = None
-    provider_metadata: Optional[Any] = None
 
 
-class _ContentCustom(BaseModel):
-    type: Literal["Custom"]
-    kind: str
-    provider_metadata: Optional[Dict[str, Any]] = None
+class _RawApprovalRequest(_Meta):
+    type: Literal["tool-approval-request"]
+    approval_id: str
+    tool_call_id: str
 
 
-class _ContentToolApprovalRequest(BaseModel):
-    type: Literal["ToolApprovalRequest"]
+class _ApprovalRequest(_Wire):
+    type: Literal["tool-approval-request"]
     approval_id: str
-    tool_call_id: str
-    provider_metadata: Optional[Dict[str, Any]] = None
+    tool_call: ToolCall
+    reason: Optional[str] = None
+    is_automatic: Optional[bool] = None
+    signature: Optional[str] = None
 
 
-_GenerateContentUnion = Annotated[
+#: A content item in the raw generation result (Rust ``GenerateContent``).
+GenerateContent = Annotated[
     Union[
-        _ContentText,
-        _ContentToolCall,
+        _Text,
+        _RawToolCallPart,
         _ContentSource,
-        _ContentReasoning,
-        _ContentFile,
-        _ContentToolResult,
-        _ContentCustom,
-        _ContentToolApprovalRequest,
+        _Reasoning,
+        _File,
+        _ReasoningFile,
+        _Custom,
+        _RawApprovalRequest,
+        _ToolResultPart,
     ],
     Field(discriminator="type"),
 ]
 
-
-class GenerateContent(RootModel[_GenerateContentUnion]):
-    """A content item in the generation result (Rust ``GenerateContent``)."""
-
-    root: _GenerateContentUnion
-
-    @model_validator(mode="before")
-    @classmethod
-    def _ext_to_int(cls, data: Any) -> Any:
-        return _external_tag_before(data)
-
-    @model_serializer(mode="wrap")
-    def _int_to_ext(self, handler):  # type: ignore[no-untyped-def]
-        return _external_tag_serializer(handler(self))
+#: A content item of the user-facing result (Rust ``TextContent``).
+TextContent = Annotated[
+    Union[
+        _Text,
+        _ToolCallPart,
+        _ContentSource,
+        _Reasoning,
+        _File,
+        _ReasoningFileOutput,
+        _Custom,
+        _ApprovalRequest,
+        _ToolResultPart,
+    ],
+    Field(discriminator="type"),
+]
 
 
 # ── StreamPart ───────────────────────────────────────────────────────────────
 
-class _SPTextStart(BaseModel):
-    type: Literal["TextStart"]
+class _SPIdMeta(_Meta):
     id: str
-    provider_metadata: Optional[Dict[str, Any]] = None
 
 
-class _SPTextDelta(BaseModel):
-    type: Literal["TextDelta"]
-    id: str
+class _SPDelta(_SPIdMeta):
     delta: str
-    provider_metadata: Optional[Dict[str, Any]] = None
 
 
-class _SPTextEnd(BaseModel):
-    type: Literal["TextEnd"]
-    id: str
-    provider_metadata: Optional[Dict[str, Any]] = None
+class _SPTextStart(_SPIdMeta):
+    type: Literal["text-start"]
 
 
-class _SPStreamStart(BaseModel):
-    type: Literal["StreamStart"]
-    warnings: List[Warning]
+class _SPTextDelta(_SPDelta):
+    type: Literal["text-delta"]
 
 
-class _SPFinish(BaseModel):
-    type: Literal["Finish"]
-    finish_reason: FinishReason
-    usage: Usage
-    provider_metadata: Optional[Any] = None
+class _SPTextEnd(_SPIdMeta):
+    type: Literal["text-end"]
 
 
-class _SPError(BaseModel):
-    type: Literal["Error"]
-    error: Any  # AiMuxErrorValue
+class _SPStreamStart(_Wire):
+    type: Literal["stream-start"]
+    warnings: List[Warning]
 
 
-class _SPToolInputStart(BaseModel):
-    type: Literal["ToolInputStart"]
-    id: str
-    tool_name: str
-    provider_executed: Optional[bool] = None
-    dynamic: Optional[bool] = None
-    title: Optional[str] = None
-    provider_metadata: Optional[Dict[str, Any]] = None
+class _SPFinish(_Meta):
+    type: Literal["finish"]
+    finish_reason: FinishReason
+    usage: Usage
 
 
-class _SPToolInputDelta(BaseModel):
-    type: Literal["ToolInputDelta"]
-    id: str
-    delta: str
-    provider_metadata: Optional[Dict[str, Any]] = None
+class _SPFinishStep(_SPFinish):
+    type: Literal["finish-step"]
+    response: GenerateResponseMetadata
 
 
-class _SPToolInputEnd(BaseModel):
-    type: Literal["ToolInputEnd"]
-    id: str
-    provider_metadata: Optional[Dict[str, Any]] = None
+class _SPError(_Wire):
+    type: Literal["error"]
+    error: AiMuxErrorValue
 
 
-class _SPToolCall(BaseModel):
-    type: Literal["ToolCall"]
-    tool_call_id: str
+class _SPToolInputStart(_SPIdMeta):
+    type: Literal["tool-input-start"]
     tool_name: str
-    input: Any
     provider_executed: Optional[bool] = None
     dynamic: Optional[bool] = None
-    provider_metadata: Optional[Dict[str, Any]] = None
-    invalid: Optional[bool] = None
-    error: Optional[AiMuxErrorValue] = None
-
-
-class _SPToolResult(BaseModel):
-    type: Literal["ToolResult"]
-    tool_call_id: str
-    tool_name: str
-    result: Any
-    is_error: Optional[bool] = None
-    preliminary: Optional[bool] = None
-    dynamic: Optional[bool] = None
-    provider_metadata: Optional[Dict[str, Any]] = None
+    title: Optional[str] = None
 
 
-class _SPReasoningStart(BaseModel):
-    type: Literal["ReasoningStart"]
-    id: str
-    provider_metadata: Optional[Any] = None
+class _SPToolInputDelta(_SPDelta):
+    type: Literal["tool-input-delta"]
 
 
-class _SPReasoningDelta(BaseModel):
-    type: Literal["ReasoningDelta"]
-    id: str
-    delta: str
-    provider_metadata: Optional[Any] = None
+class _SPToolInputEnd(_SPIdMeta):
+    type: Literal["tool-input-end"]
 
 
-class _SPReasoningEnd(BaseModel):
-    type: Literal["ReasoningEnd"]
-    id: str
-    provider_metadata: Optional[Any] = None
+class _SPReasoningStart(_SPIdMeta):
+    type: Literal["reasoning-start"]
 
 
-class _SPFinishStep(_SPFinish):
-    type: Literal["FinishStep"]
-    response: GenerateResponseMetadata
+class _SPReasoningDelta(_SPDelta):
+    type: Literal["reasoning-delta"]
 
 
+class _SPReasoningEnd(_SPIdMeta):
+    type: Literal["reasoning-end"]
 
 
-class _SPRaw(BaseModel):
-    type: Literal["Raw"]
+class _SPRaw(_Wire):
+    type: Literal["raw"]
     raw_value: Any
 
 
-class _SPFile(BaseModel):
-    type: Literal["File"]
-    data: GeneratedFileData
-    media_type: str
-    provider_metadata: Optional[Dict[str, Any]] = None
-
-
-class _SPToolApprovalRequest(BaseModel):
-    type: Literal["ToolApprovalRequest"]
-    approval_id: str
-    tool_call: ToolCall
-    reason: Optional[str] = None
-    is_automatic: Optional[bool] = None
-    signature: Optional[str] = None
-
-
-class _GeneratedFile(BaseModel):
-    data: GeneratedFileData
-    media_type: str
-    provider_metadata: Optional[Dict[str, Any]] = None
-
-
-class _SPReasoningFile(BaseModel):
-    type: Literal["ReasoningFile"]
-    file: _GeneratedFile
-    provider_metadata: Optional[Dict[str, Any]] = None
-
-
-_TextContentUnion = Annotated[
-    Union[
-        _ContentText, _SPToolCall, _ContentSource, _ContentReasoning,
-        _SPFile, _SPReasoningFile, _ContentToolResult, _ContentCustom, _SPToolApprovalRequest,
-    ],
-    Field(discriminator="type"),
-]
-
-
-class TextContent(RootModel[_TextContentUnion]):
-    @model_validator(mode="before")
-    @classmethod
-    def _ext_to_int(cls, data: Any) -> Any:
-        return _external_tag_before(data)
-
-    @model_serializer(mode="wrap")
-    def _int_to_ext(self, handler):
-        return _external_tag_serializer(handler(self))
-
-
-_StreamPartUnion = Annotated[
+#: A single chunk of the stream returned by ``stream_text`` (Rust
+#: ``TextStreamPart``), e.g. ``{"type": "text-delta", "id": ..., "delta": ...}``.
+StreamPart = Annotated[
     Union[
         _SPTextStart,
         _SPTextDelta,
         _SPTextEnd,
         _SPStreamStart,
         _SPFinish,
+        _SPFinishStep,
         _SPError,
         _SPToolInputStart,
         _SPToolInputDelta,
         _SPToolInputEnd,
-        _SPToolCall,
-        _SPToolResult,
-        _SPFile,
-        _SPReasoningFile,
+        _ToolCallPart,
+        _ToolResultPart,
+        _File,
+        _ReasoningFileOutput,
+        _Custom,
+        _ApprovalRequest,
         _SPReasoningStart,
         _SPReasoningDelta,
         _SPReasoningEnd,
-        _SPFinishStep,
         _ContentSource,
         _SPRaw,
-        _ContentCustom,
-        _SPToolApprovalRequest,
     ],
     Field(discriminator="type"),
 ]
 
 
-class StreamPart(RootModel[_StreamPartUnion]):
-    """A single chunk in the stream returned by ``stream_text`` (Rust ``StreamPart``).
-
-    Externally tagged on the wire: ``{"TextDelta": {"id": ..., "delta": ...}}``.
-    ``parse_stream_part`` / ``StreamPart.model_validate`` accept that shape and
-    :meth:`model_dump` re-emits it.
-    """
-
-    root: _StreamPartUnion
-
-    @model_validator(mode="before")
-    @classmethod
-    def _ext_to_int(cls, data: Any) -> Any:
-        return _external_tag_before(data)
-
-    @model_serializer(mode="wrap")
-    def _int_to_ext(self, handler):  # type: ignore[no-untyped-def]
-        return _external_tag_serializer(handler(self))
-
-
 # ─────────────────────────────────────────────────────────────────────────────
-# ResponseFormat — external-tag enum with a unit variant (``"Text"``)
+# ResponseFormat — ``{"type": "text"}`` | ``{"type": "json", schema?, ...}``
 # ─────────────────────────────────────────────────────────────────────────────
 
-class _ResponseFormatJsonBody(BaseModel):
-    model_config = ConfigDict(populate_by_name=True)
+class _ResponseFormatText(_Wire):
+    type: Literal["text"] = "text"
 
+
+class _ResponseFormatJson(_Wire):
+    type: Literal["json"] = "json"
     schema_: Optional[Any] = Field(default=None, alias="schema")
     name: Optional[str] = None
     description: Optional[str] = None
 
 
-class ResponseFormat(RootModel[Union[Literal["Text"], _ResponseFormatJsonBody]]):
-    """How the model should format its response (Rust ``ResponseFormat``).
-
-    Wire format: ``"Text"`` or ``{"Json": {"schema": ..., "name": ..., "description": ...}}``.
-    """
-
-    root: Union[Literal["Text"], _ResponseFormatJsonBody]
-
-    @model_validator(mode="before")
-    @classmethod
-    def _unwrap(cls, data: Any) -> Any:
-        if isinstance(data, dict) and len(data) == 1:
-            (tag, payload), = data.items()
-            if tag == "Text":
-                return "Text"
-            if tag == "Json":
-                return payload if isinstance(payload, dict) else {}
-        return data
-
-    @model_serializer(mode="wrap")
-    def _ser(self, handler):  # type: ignore[no-untyped-def]
-        out = handler(self)
-        if isinstance(out, dict):
-            return {"Json": out}
-        return out  # bare "Text"
+#: How the model should format its response (Rust ``ResponseFormat``).
+ResponseFormat = Annotated[
+    Union[_ResponseFormatText, _ResponseFormatJson], Field(discriminator="type")
+]
 
 
 # ─────────────────────────────────────────────────────────────────────────────
-# Internal-tag enums (Rust ``#[serde(tag = "type", rename_all = "snake_case")]``)
-# These are already internal-tagged on the wire, so a standard discriminated
-# union with a ``type`` discriminator works directly.
+# Tools
 # ─────────────────────────────────────────────────────────────────────────────
 
-class FunctionToolInputExample(BaseModel):
+class FunctionToolInputExample(_Wire):
     input: Dict[str, Any]
 
 
-class FunctionTool(BaseModel):
+class FunctionTool(_Wire):
     """A user-defined function tool (Rust ``FunctionTool``)."""
 
     type: Literal["function"] = "function"
@@ -782,7 +556,7 @@ class FunctionTool(BaseModel):
     input_examples: Optional[List[FunctionToolInputExample]] = None
 
 
-class ProviderTool(BaseModel):
+class ProviderTool(_Wire):
     """A provider-defined tool (Rust ``ProviderTool``)."""
 
     type: Literal["provider"] = "provider"
@@ -797,96 +571,89 @@ class ProviderTool(BaseModel):
 ]
 
 
-# ── ContentPart (internal tag ``type``, snake_case variants) ────────────────
+# ── ContentPart (prompt parts: internal tag ``type``, kebab-case) ───────────
+
+class _PartOptions(_Wire):
+    provider_options: Optional[Any] = None
+
 
-class TextContentPart(BaseModel):
+class TextContentPart(_PartOptions):
     """A text content part (Rust ``ContentPart::Text``)."""
 
     type: Literal["text"] = "text"
     text: str
-    provider_options: Optional[Any] = None
 
 
-class _ImageContentPart(BaseModel):
+class _ImageContentPart(_PartOptions):
     type: Literal["image"] = "image"
     image: List[int]
     media_type: str
-    provider_options: Optional[Any] = None
 
 
-class _FileContentPart(BaseModel):
+class _FileContentPart(_PartOptions):
     type: Literal["file"] = "file"
     data: List[int]
     media_type: str
     filename: Optional[str] = None
-    provider_options: Optional[Any] = None
 
 
-class _FileBase64ContentPart(BaseModel):
-    type: Literal["file_base64"] = "file_base64"
+class _FileBase64ContentPart(_PartOptions):
+    type: Literal["file-base64"] = "file-base64"
     data: str
     media_type: str
     filename: Optional[str] = None
-    provider_options: Optional[Any] = None
 
 
-class _FileUrlContentPart(BaseModel):
-    type: Literal["file_url"] = "file_url"
+class _FileUrlContentPart(_PartOptions):
+    type: Literal["file-url"] = "file-url"
     url: str
     media_type: str
-    provider_options: Optional[Any] = None
 
 
-class _FileReferenceContentPart(BaseModel):
-    type: Literal["file_reference"] = "file_reference"
+class _FileReferenceContentPart(_PartOptions):
+    type: Literal["file-reference"] = "file-reference"
     media_type: str
     reference: Any
     filename: Optional[str] = None
-    provider_options: Optional[Any] = None
 
 
-class _ReasoningContentPart(BaseModel):
+class _ReasoningContentPart(_PartOptions):
     type: Literal["reasoning"] = "reasoning"
     text: str
     signature: Optional[str] = None
-    provider_options: Optional[Any] = None
 
 
-class _ToolCallContentPart(BaseModel):
-    type: Literal["tool_call"] = "tool_call"
+class _ToolCallContentPart(_PartOptions):
+    type: Literal["tool-call"] = "tool-call"
     tool_call_id: str
     tool_name: str
     input: Any
     provider_executed: Optional[bool] = None
-    provider_options: Optional[Any] = None
 
 
-class _ToolResultContentPart(BaseModel):
-    type: Literal["tool_result"] = "tool_result"
+class _ToolResultContentPart(_PartOptions):
+    """``output`` is a ``{"type": "text" | "json" | "execution-denied" |
+    "error-text" | "error-json" | "content", ...}`` object."""
+
+    type: Literal["tool-result"] = "tool-result"
     tool_call_id: str
-    result: Any
-    tool_name: Optional[str] = None
-    is_error: Optional[bool] = None
-    preliminary: Optional[bool] = None
-    dynamic: Optional[bool] = None
-    provider_options: Optional[Any] = None
+    tool_name: str
+    output: Any
 
 
-class _CustomContentPart(BaseModel):
+class _CustomContentPart(_PartOptions):
     type: Literal["custom"] = "custom"
     kind: str
-    provider_options: Optional[Any] = None
 
 
-class _ReasoningFileContentPart(BaseModel):
-    type: Literal["reasoning_file"] = "reasoning_file"
+class _ReasoningFileContentPart(_PartOptions):
+    type: Literal["reasoning-file"] = "reasoning-file"
     data: GeneratedFileData
     media_type: str
-    provider_options: Optional[Any] = None
 
 
-class _ToolApprovalRequestContentPart(BaseModel):
-    type: Literal["tool_approval_request"] = "tool_approval_request"
+class _ToolApprovalRequestContentPart(_Wire):
+    type: Literal["tool-approval-request"] = "tool-approval-request"
     approval_id: str
     tool_call_id: str
     reason: Optional[str] = None
@@ -919,31 +686,15 @@ class _ToolApprovalRequestContentPart(BaseModel):
 
 
 # ─────────────────────────────────────────────────────────────────────────────
-# ToolChoice — mixed tagged/untagged enum
-# Wire: ``"auto" | "none" | "required" | {"type": "tool", "toolName": "..."}``.
-# Note the camelCase ``toolName`` (hand-rolled serde on the Rust side).
+# ToolChoice — ``"auto" | "none" | "required" | {"type": "tool", "toolName": ...}``
 # ─────────────────────────────────────────────────────────────────────────────
 
-class ToolChoiceTool(BaseModel):
+class ToolChoiceTool(_Wire):
     """The ``tool`` tool-choice variant (forces a specific tool)."""
 
     type: Literal["tool"] = "tool"
     tool_name: str
 
-    @model_validator(mode="before")
-    @classmethod
-    def _accept_camel(cls, data: Any) -> Any:
-        if isinstance(data, dict) and "toolName" in data and "tool_name" not in data:
-            data = {**data, "tool_name": data.pop("toolName")}
-        return data
-
-    @model_serializer(mode="wrap")
-    def _emit_camel(self, handler):  # type: ignore[no-untyped-def]
-        out = handler(self)
-        if isinstance(out, dict):
-            out["toolName"] = out.pop("tool_name", None)
-        return out
-
 
 ToolChoice = Union[Literal["auto", "none", "required"], ToolChoiceTool]
 
@@ -952,7 +703,7 @@ def _emit_camel(self, handler):  # type: ignore[no-untyped-def]
 # User-facing prompt / message models
 # ─────────────────────────────────────────────────────────────────────────────
 
-class ModelMessage(BaseModel):
+class ModelMessage(_Wire):
     """A single user-facing chat message (Rust ``ModelMessage``).
 
     ``content`` is untagged on the wire: a bare string or a list of content
@@ -967,7 +718,7 @@ class ModelMessage(BaseModel):
 # Options + results
 # ─────────────────────────────────────────────────────────────────────────────
 
-class TimeoutConfiguration(BaseModel):
+class TimeoutConfiguration(_Wire):
     """Per-call timeout configuration (Rust ``TimeoutConfiguration``).
 
     All values are milliseconds; ``None`` disables the corresponding limit.
@@ -979,7 +730,7 @@ class TimeoutConfiguration(BaseModel):
     chunk_ms: Optional[int] = None
 
 
-class ToolCallRepairContext(BaseModel):
+class ToolCallRepairContext(_Wire):
     """What a ``repair_tool_call`` function is given (RFC-0035).
 
     Mirrors the AI SDK ``repairToolCall`` argument: the offending call with its
@@ -1002,7 +753,7 @@ class ToolCallRepairContext(BaseModel):
 RepairToolCall = Callable[[ToolCallRepairContext], Optional[RawToolCall]]
 
 
-class GenerateTextOptions(BaseModel):
+class GenerateTextOptions(_Wire):
     """User-facing options for ``generate_text`` / ``stream_text``.
 
     Mirrors Rust ``GenerateTextOptions``. All fields are optional; unset fields
@@ -1059,11 +810,11 @@ class GenerateTextOptions(BaseModel):
     """
 
 
-class RequestInfo(BaseModel):
+class RequestInfo(_Wire):
     body: Optional[Any] = None
 
 
-class GenerateResult(BaseModel):
+class GenerateResult(_Wire):
     """Raw provider result (Rust ``GenerateResult``), exposed via ``raw``."""
 
     content: List[GenerateContent]
@@ -1075,7 +826,7 @@ class GenerateResult(BaseModel):
     response: Optional[GenerateResponseMetadata] = None
 
 
-class GenerateTextResult(BaseModel):
+class GenerateTextResult(_Wire):
     """Result of ``generate_text`` (user-facing, Rust ``GenerateTextResult``)."""
 
     content: List[TextContent] = Field(default_factory=list)
@@ -1105,7 +856,7 @@ class GenerateTextResult(BaseModel):
     total_usage: Usage = Field(default_factory=Usage)
 
 
-class GenerateObjectResult(BaseModel):
+class GenerateObjectResult(_Wire):
     """Result of ``generate_object`` (user-facing, M12, Rust
     ``GenerateObjectResult``). The parsed JSON object plus convenience fields
     from the underlying ``generate_text`` call.
@@ -1126,7 +877,7 @@ class GenerateObjectResult(BaseModel):
     raw: GenerateTextResult
 
 
-class StreamTextResultAggregated(BaseModel):
+class StreamTextResultAggregated(_Wire):
     """Aggregated result of ``stream_text().consume()`` (M11, Rust
     ``StreamTextResultAggregated``). Mirrors ``GenerateTextResult``'s
     user-facing fields (without ``raw``, since streaming has no
@@ -1302,7 +1053,9 @@ def _prompt_to_json(prompt: Prompt) -> str:
             f"prompt must be str or list of messages, got {type(prompt).__name__}"
         )
     msgs = [
-        m.model_dump(mode="json", exclude_none=True) if isinstance(m, BaseModel) else m
+        m.model_dump(mode="json", exclude_none=True, by_alias=True)
+        if isinstance(m, BaseModel)
+        else m
         for m in prompt
     ]
     return json.dumps({"prompt": msgs})
@@ -1320,7 +1073,7 @@ def _opts_to_json(options: Optional[GenerateTextOptions]) -> Optional[str]:
     # ``exclude_none`` mirrors Rust's ``#[serde(skip_serializing_if = "Option::is_none")]``
     # on the inner struct fields, producing clean payloads (absent == None for the
     # Rust deserializer). ``by_alias`` emits the ``schema`` field of the JSON
-    # response-format body (and is a no-op for the snake_case fields elsewhere).
+    # response-format body and the camelCase keys of every model.
     return options.model_dump_json(exclude_none=True, by_alias=True)
 
 
@@ -1335,7 +1088,7 @@ def adapter(context: dict) -> Optional[dict]:
         replacement = _repair.run_hook(fn, typed)
         if replacement is None:
             return None
-        return replacement.model_dump(mode="json", exclude_none=True)
+        return replacement.model_dump(mode="json", exclude_none=True, by_alias=True)
 
     return adapter
 
@@ -1376,15 +1129,15 @@ def stream_text(
 ) -> Iterator[dict]:
     """Stream text from a model, yielding parsed ``StreamPart`` dicts.
 
-    Each yielded item is the external-tagged dict form produced by the Rust
-    core, e.g. ``{"TextDelta": {"id": ..., "delta": ...}}``. For typed access,
-    pass the dict through :func:`parse_stream_part`.
+    Each yielded item is the dict form produced by the Rust core, tagged by
+    ``type``, e.g. ``{"type": "text-delta", "id": ..., "delta": ...}``. For
+    typed access, pass the dict through :func:`parse_stream_part`.
 
     Usage::
 
         for part in stream_text(model, "Write a haiku about Rust."):
-            if "TextDelta" in part:
-                print(part["TextDelta"]["delta"], end="")
+            if part["type"] == "text-delta":
+                print(part["delta"], end="")
     """
     prompt_json = _prompt_to_json(prompt)
     opts_json = _opts_to_json(options)
@@ -1397,9 +1150,12 @@ def stream_text(
         yield json.loads(part_json)
 
 
-def parse_stream_part(part: dict) -> StreamPart:
-    """Validate a yielded stream-part dict into a typed :class:`StreamPart`."""
-    return StreamPart.model_validate(part)
+_STREAM_PART = TypeAdapter(StreamPart)
+
+
+def parse_stream_part(part: dict) -> BaseModel:
+    """Validate a yielded stream-part dict into its typed model (see ``StreamPart``)."""
+    return _STREAM_PART.validate_python(part)
 
 
 def generate_text_as_openai(
diff --git a/bindings/python/tests/test_aimux.py b/bindings/python/tests/test_aimux.py
index 60b3397f9..ccb1bb8ce 100644
--- a/bindings/python/tests/test_aimux.py
+++ b/bindings/python/tests/test_aimux.py
@@ -101,7 +101,7 @@ def test_provider_invalid_config_raises():
 
     with pytest.raises(Exception, match="invalid config"):
         provider(
-            "groq",
+            "abacus",
             "sk-test-fake-key",
             "llama-3.3-70b",
             config={"headers": "not-a-map"},
@@ -109,14 +109,13 @@ def test_provider_invalid_config_raises():
 
 
 def test_provider_missing_env_key_raises(monkeypatch):
-    """provider() with api_key=None reads the env var when a request is made, and
-    fails clearly when it is unset (the AI SDK loads the key lazily)."""
+    """provider() with api_key=None reads the env var at call time and fails clearly when unset."""
     from aimux import LoadAPIKeyError, generate_text, provider
 
     monkeypatch.delenv("ABACUS_API_KEY", raising=False)
     model = provider("abacus", None, "m")
     with pytest.raises(LoadAPIKeyError, match="(?i)api key"):
-        generate_text(model, "hi")
+        generate_text(model, "hi", {"maxRetries": 0})
 
 
 def test_provider_params_fill_a_preset_template(monkeypatch):
@@ -127,7 +126,7 @@ def test_provider_params_fill_a_preset_template(monkeypatch):
     # Without the parameter the preset cannot build its base URL (reported at call time).
     unfilled = provider("cloudflare_workers_ai", "sk-test-fake-key", "m")
     with pytest.raises(InvalidArgumentError, match="account_id"):
-        generate_text(unfilled, "hi", {"max_retries": 0})
+        generate_text(unfilled, "hi", {"maxRetries": 0})
     model = provider(
         "cloudflare_workers_ai",
         "sk-test-fake-key",
diff --git a/bindings/python/tests/test_cassette.py b/bindings/python/tests/test_cassette.py
index 8a4ab70f7..d168de9ec 100644
--- a/bindings/python/tests/test_cassette.py
+++ b/bindings/python/tests/test_cassette.py
@@ -22,16 +22,16 @@ def test_generate(self):
             assert isinstance(result["text"], str)
             assert len(result["text"]) > 0, "text should be non-empty"
             assert "usage" in result
-            assert "finish_reason" in result
+            assert "finishReason" in result
 
     def test_stream(self):
         with CassetteServer("openai") as srv:
             model = openai("test-key", "gpt-4o", f"{srv.url}/v1")
             parts = list(stream_text(model, "Hello"))
             assert len(parts) > 0, "should receive stream parts"
-            types = [list(p.keys())[0] for p in parts]
-            assert "StreamStart" in types, "should have StreamStart"
-            assert "Finish" in types, "should have Finish"
+            types = [p["type"] for p in parts]
+            assert "stream-start" in types, "should have stream-start"
+            assert "finish" in types, "should have finish"
 
 
 class TestAnthropicCassette:
@@ -75,7 +75,7 @@ def test_generate_usage_raw_carries_vendor_fields(self):
         assert isinstance(raw.get("prompt_cache_hit_tokens"), int)
         assert isinstance(raw.get("prompt_cache_miss_tokens"), int)
         # Typed totals still work alongside the raw object.
-        assert result["usage"]["input_tokens"]["total"] is not None
+        assert result["usage"]["inputTokens"]["total"] is not None
 
 
 class TestGroqCassette:
diff --git a/bindings/python/tests/test_decision.py b/bindings/python/tests/test_decision.py
index afc16a439..c1cb851a1 100644
--- a/bindings/python/tests/test_decision.py
+++ b/bindings/python/tests/test_decision.py
@@ -62,7 +62,7 @@ def test_decision_uses_unified_runtime_recording(server, tmp_path):
     assert record['complete'] is True
     assert record['input']['operation'] == 'decision'
     assert record['input']['options']['questions'] == QUESTIONS
-    assert record['outcome']['decision_result']['answers'] == result['answers']
+    assert record['outcome']['decisionResult']['answers'] == result['answers']
     assert record['exchanges'][0]['attempt'] == 1
     assert 'test-key' not in json.dumps(record)
 
@@ -82,8 +82,8 @@ def test_optional_usage_counts_preserve_answers(server, monkeypatch, usage):
     result = decide(jev_decision('test-key', 'jev-latest', base + '/v1/systemone'),
                     {'message': 'Billed twice'}, QUESTIONS, max_retries=0)
     assert len(result['answers']) == 3
-    assert result['usage']['input_tokens']['total'] == usage.get('input_tokens')
-    assert result['usage']['output_tokens']['total'] == usage.get('output_tokens')
+    assert result['usage']['inputTokens'].get('total') == usage.get('input_tokens')
+    assert result['usage']['outputTokens'].get('total') == usage.get('output_tokens')
     assert result['usage']['raw'] == usage
 
 
diff --git a/bindings/python/tests/test_e2e.py b/bindings/python/tests/test_e2e.py
index edfac45dc..c2d0c0232 100644
--- a/bindings/python/tests/test_e2e.py
+++ b/bindings/python/tests/test_e2e.py
@@ -154,14 +154,14 @@ def test_stream_text(self):
             parts = list(stream_text(model, "Say hello"))
             assert len(parts) > 0
             text = "".join(
-                p["TextDelta"]["delta"] for p in parts if "TextDelta" in p
+                p["delta"] for p in parts if p["type"] == "text-delta"
             )
             assert text == "Hello world"
 
     def test_generate_text_with_options(self):
         with MockServer(OPENAI_CHAT) as mock:
             model = deepseek("test-key", "gpt-4o", mock.url)
-            generate_text(model, "Hello", {"max_output_tokens": 100, "temperature": 0.5})
+            generate_text(model, "Hello", {"maxOutputTokens": 100, "temperature": 0.5})
             # Options should reach the HTTP request (tested implicitly — if they
             # don't serialize, the Rust side would error)
 
@@ -181,7 +181,7 @@ def test_stream_text(self):
             parts = list(stream_text(model, "Hello"))
             assert len(parts) > 0
             text = "".join(
-                p["TextDelta"]["delta"] for p in parts if "TextDelta" in p
+                p["delta"] for p in parts if p["type"] == "text-delta"
             )
             assert text == "Hello from Claude"
 
@@ -350,7 +350,7 @@ def test_generate_text_parses_tool_calls(self):
                     "type": "function",
                     "name": "get_weather",
                     "description": "Get weather for a location",
-                    "input_schema": {
+                    "inputSchema": {
                         "type": "object",
                         "properties": {"location": {"type": "string"}},
                         "required": ["location"],
@@ -359,22 +359,22 @@ def test_generate_text_parses_tool_calls(self):
             }
             result = generate_text(model, "What's the weather in Tokyo?", opts)
 
-            # Convenience field: tool_calls extracted
-            assert len(result["tool_calls"]) == 1
-            assert result["tool_calls"][0]["tool_name"] == "get_weather"
-            assert result["tool_calls"][0]["tool_call_id"] == "call_abc"
-            assert result["tool_calls"][0]["input"] == {"location": "Tokyo"}
+            # Convenience field: toolCalls extracted
+            assert len(result["toolCalls"]) == 1
+            assert result["toolCalls"][0]["toolName"] == "get_weather"
+            assert result["toolCalls"][0]["toolCallId"] == "call_abc"
+            assert result["toolCalls"][0]["input"] == {"location": "Tokyo"}
 
             # Structured content: raw.content contains the ToolCall variant
             assert "raw" in result
             assert isinstance(result["raw"]["content"], list)
-            tc = next((c for c in result["raw"]["content"] if "ToolCall" in c), None)
-            assert tc is not None, "raw.content must contain a ToolCall variant"
-            assert tc["ToolCall"]["tool_name"] == "get_weather"
-            assert tc["ToolCall"]["tool_call_id"] == "call_abc"
+            tc = next((c for c in result["raw"]["content"] if c["type"] == "tool-call"), None)
+            assert tc is not None, "raw.content must contain a tool-call variant"
+            assert tc["toolName"] == "get_weather"
+            assert tc["toolCallId"] == "call_abc"
             # raw content keeps the provider's argument text; parsing happens
-            # at the Core boundary (top-level tool_calls carry the object).
-            assert tc["ToolCall"]["input"] == '{"location":"Tokyo"}'
+            # at the Core boundary (top-level toolCalls carry the object).
+            assert tc["input"] == '{"location":"Tokyo"}'
 
     def test_multi_role_messages_reach_provider(self):
         with RecordingMockServer(OPENAI_CHAT) as mock:
@@ -401,12 +401,12 @@ def test_tool_choice_reaches_provider(self):
                 "tools": [{
                     "type": "function",
                     "name": "get_weather",
-                    "input_schema": {
+                    "inputSchema": {
                         "type": "object",
                         "properties": {"location": {"type": "string"}},
                     },
                 }],
-                "tool_choice": "required",
+                "toolChoice": "required",
             }
             generate_text(model, "Hello", opts)
 
@@ -414,7 +414,7 @@ def test_tool_choice_reaches_provider(self):
             assert body["tool_choice"] == "required"
 
     def test_stream_text_parses_tool_call_parts(self):
-        """Stream text should surface ToolCall/ToolInputDelta parts (not just TextDelta/Finish)."""
+        """Stream text should surface tool-call/tool-input-delta parts (not just text-delta/finish)."""
         sse_body = (
             'data: {"id":"1","model":"gpt-4o","choices":[{"delta":{"role":"assistant","tool_calls":[{"index":0,"id":"call_xyz","type":"function","function":{"name":"get_weather","arguments":""}}]}}]}\n\n'
             'data: {"id":"1","model":"gpt-4o","choices":[{"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\\"location\\":\\"Tokyo\\"}"}}]}}]}\n\n'
@@ -427,7 +427,7 @@ def test_stream_text_parses_tool_call_parts(self):
                 "tools": [{
                     "type": "function",
                     "name": "get_weather",
-                    "input_schema": {
+                    "inputSchema": {
                         "type": "object",
                         "properties": {"location": {"type": "string"}},
                     },
@@ -436,18 +436,17 @@ def test_stream_text_parses_tool_call_parts(self):
             parts = list(stream_text(model, "What's the weather in Tokyo?", opts))
             assert len(parts) > 0
 
-            # Stream must contain a tool-related StreamPart (not just TextDelta/Finish)
+            # Stream must contain a tool-related StreamPart (not just text-delta/finish)
             has_tool_part = any(
-                "ToolCall" in p or "ToolInputDelta" in p or "ToolInputStart" in p
+                p["type"] in ("tool-call", "tool-input-delta", "tool-input-start")
                 for p in parts
             )
             assert has_tool_part, f"expected a tool stream part; got: {parts}"
 
             # If a complete ToolCall part is present, verify its fields
-            tool_call_parts = [p for p in parts if "ToolCall" in p]
+            tool_call_parts = [p for p in parts if p["type"] == "tool-call"]
             if tool_call_parts:
-                tc = tool_call_parts[0]["ToolCall"]
-                assert tc["tool_name"] == "get_weather"
+                assert tool_call_parts[0]["toolName"] == "get_weather"
 
     def test_tool_call_round_trip(self):
         """Full tool-call round trip: two generate_text calls with a ToolResult
@@ -458,7 +457,7 @@ def test_tool_call_round_trip(self):
         Second call → model returns final text (OPENAI_CHAT).
 
         The messages use the framework's user-facing content-part shape
-        (``tool_call`` / ``tool_result`` parts). The OpenAI converter turns
+        (``tool-call`` / ``tool-result`` parts). The OpenAI converter turns
         these into the provider-facing body (assistant ``tool_calls`` array +
         ``tool`` role message with ``tool_call_id``), which is what we assert
         on for the second request.
@@ -471,7 +470,7 @@ def test_tool_call_round_trip(self):
                     "type": "function",
                     "name": "get_weather",
                     "description": "Get weather for a location",
-                    "input_schema": {
+                    "inputSchema": {
                         "type": "object",
                         "properties": {"location": {"type": "string"}},
                         "required": ["location"],
@@ -481,21 +480,23 @@ def test_tool_call_round_trip(self):
 
             # Step 1: first call — the model requests a tool call.
             result = generate_text(model, "What's the weather in Tokyo?", opts)
-            assert len(result["tool_calls"]) == 1
-            assert result["tool_calls"][0]["tool_name"] == "get_weather"
-            assert result["tool_calls"][0]["tool_call_id"] == "call_abc"
+            assert len(result["toolCalls"]) == 1
+            assert result["toolCalls"][0]["toolName"] == "get_weather"
+            assert result["toolCalls"][0]["toolCallId"] == "call_abc"
 
             # Step 2: back-fill the ToolResult and re-send the conversation.
             # Roles: user → assistant(tool_call) → tool(result).
             messages = [
                 {"role": "user", "content": "What's the weather in Tokyo?"},
                 {"role": "assistant", "content": [
-                    {"type": "tool_call", "tool_call_id": "call_abc",
-                     "tool_name": "get_weather", "input": {"location": "Tokyo"}},
+                    {"type": "tool-call", "toolCallId": "call_abc",
+                     "toolName": "get_weather", "input": {"location": "Tokyo"}},
                 ]},
                 {"role": "tool", "content": [
-                    {"type": "tool_result", "tool_call_id": "call_abc",
-                     "result": {"temperature": 22, "condition": "sunny"}},
+                    {"type": "tool-result", "toolCallId": "call_abc",
+                     "toolName": "get_weather",
+                     "output": {"type": "json",
+                                "value": {"temperature": 22, "condition": "sunny"}}},
                 ]},
             ]
 
diff --git a/bindings/python/tests/test_error.py b/bindings/python/tests/test_error.py
index c76b09851..cb1ebd0ce 100644
--- a/bindings/python/tests/test_error.py
+++ b/bindings/python/tests/test_error.py
@@ -100,7 +100,7 @@ def test_api_call_error_carries_every_field_the_response_produced():
         model = openai("test-key", "gpt-4o", f"http://127.0.0.1:{port}")
         # 429 is retryable; disable retries to observe the bare call failure.
         with pytest.raises(APICallError) as excinfo:
-            generate_text(model, "hi", {"max_retries": 0})
+            generate_text(model, "hi", {"maxRetries": 0})
         err = excinfo.value
 
         assert type(err) is APICallError
@@ -135,7 +135,7 @@ def test_retry_error_carries_reason_and_typed_attempt_history():
 
         model = openai("test-key", "gpt-4o", f"http://127.0.0.1:{port}")
         with pytest.raises(RetryError) as excinfo:
-            generate_text(model, "hi", {"max_retries": 1})
+            generate_text(model, "hi", {"maxRetries": 1})
         err = excinfo.value
 
         assert type(err) is RetryError
@@ -172,7 +172,7 @@ def test_missing_api_key_is_load_api_key_error(monkeypatch):
     monkeypatch.delenv("GROQ_API_KEY", raising=False)
     model = provider("groq", None, "some-model")
     with pytest.raises(LoadAPIKeyError) as excinfo:
-        generate_text(model, "hi", {"max_retries": 0})
+        generate_text(model, "hi", {"maxRetries": 0})
     err = excinfo.value
 
     assert type(err) is LoadAPIKeyError
@@ -200,7 +200,7 @@ def test_missing_setting_is_load_setting_error_at_call_time(monkeypatch):
     monkeypatch.delenv("AZURE_RESOURCE_NAME", raising=False)
     model = azure("test-key", "", "some-deployment")
     with pytest.raises(LoadSettingError) as excinfo:
-        generate_text(model, "hi", {"max_retries": 0})
+        generate_text(model, "hi", {"maxRetries": 0})
     err = excinfo.value
 
     assert type(err) is LoadSettingError
diff --git a/bindings/python/tests/test_multimodal_e2e.py b/bindings/python/tests/test_multimodal_e2e.py
index 2146cb4da..a9640c86c 100644
--- a/bindings/python/tests/test_multimodal_e2e.py
+++ b/bindings/python/tests/test_multimodal_e2e.py
@@ -141,7 +141,7 @@ def __exit__(self, *args):
 # Google Video uses a multi-step async API; we only parse a canned result here.
 VIDEO_RESULT_JSON = json.dumps({
     "videos": [
-        {"Url": {"url": "https://example.com/v.mp4", "media_type": "video/mp4"}},
+        {"type": "url", "url": "https://example.com/v.mp4", "mediaType": "video/mp4"},
     ],
 })
 
@@ -167,23 +167,23 @@ def test_speech(self):
         # Mock returns raw binary with content-type audio/mpeg.
         with MockServer(SPEECH_RESP, content_type='audio/mpeg') as mock:
             speaker = openai_speech("test-key", "tts-1", base_url=mock.url)
-            opts = json.dumps({"text": "Hi", "voice": "alloy", "output_format": "mp3"})
+            opts = json.dumps({"text": "Hi", "voice": "alloy", "outputFormat": "mp3"})
             result = json.loads(speaker.generate(opts))
-            # Binary audio comes back as AudioData::Binary -> {"Binary": [bytes...]}
-            assert "Binary" in result["audio"]
-            assert len(result["audio"]["Binary"]) > 0
+            # Binary audio comes back as AudioData::Binary -> a JSON array of bytes.
+            assert isinstance(result["audio"], list)
+            assert len(result["audio"]) > 0
 
     def test_image(self):
         with MockServer(IMAGE_RESP) as mock:
             imager = openai_image("test-key", "dall-e-3", base_url=mock.url)
-            # ImageCallOptions.provider_options is a required (non-Option) field,
+            # ImageCallOptions.providerOptions is a required (non-Option) field,
             # so it must be present in the serialized options.
             opts = json.dumps(
-                {"prompt": "otter", "n": 1, "provider_options": {}}
+                {"prompt": "otter", "n": 1, "providerOptions": {}}
             )
             result = json.loads(imager.generate(opts))
-            assert len(result["images"]["Base64"]) == 1
-            assert result["images"]["Base64"][0] == "aW1hZ2Ux"
+            assert len(result["images"]) == 1
+            assert result["images"][0] == "aW1hZ2Ux"
 
     def test_transcription(self):
         with MockServer(TRANSCRIPTION_RESP) as mock:
@@ -198,12 +198,12 @@ def test_reranking(self):
             reranker = cohere_reranking(
                 "test-key", "rerank-v3.0", base_url=mock.url
             )
-            # RerankingDocuments is an externally-tagged enum; the text-documents
-            # variant is {"Text": {"values": [...]}} (a bare array is rejected).
-            docs = json.dumps({"Text": {"values": ["doc1", "doc2"]}})
+            # RerankingDocuments is tagged by ``type``; the text-documents
+            # variant is {"type": "text", "values": [...]} (a bare array is rejected).
+            docs = json.dumps({"type": "text", "values": ["doc1", "doc2"]})
             result = json.loads(reranker.rerank("which?", docs))
             assert len(result["ranking"]) == 2
-            assert result["ranking"][0]["relevance_score"] == 0.95
+            assert result["ranking"][0]["relevanceScore"] == 0.95
 
     def test_search(self):
         with MockServer(SEARCH_RESP) as mock:
@@ -217,7 +217,7 @@ def test_files(self):
         with MockServer(FILES_RESP) as mock:
             files = openai_files("test-key", base_url=mock.url)
             result = json.loads(files.upload_file("dGVzdA==", "application/pdf"))
-            assert result["provider_reference"]["openai"] == "file-abc"
+            assert result["providerReference"]["openai"] == "file-abc"
 
     def test_video_construction_and_result_parsing(self):
         # Google Video uses a multi-step async API (POST predict → poll operation
@@ -230,5 +230,5 @@ def test_video_construction_and_result_parsing(self):
 
         result = json.loads(VIDEO_RESULT_JSON)
         assert len(result["videos"]) == 1
-        assert result["videos"][0]["Url"]["url"] == "https://example.com/v.mp4"
-        assert result["videos"][0]["Url"]["media_type"] == "video/mp4"
+        assert result["videos"][0]["url"] == "https://example.com/v.mp4"
+        assert result["videos"][0]["mediaType"] == "video/mp4"
diff --git a/bindings/python/tests/test_session.py b/bindings/python/tests/test_session.py
index e04ee87c9..61fc1e630 100644
--- a/bindings/python/tests/test_session.py
+++ b/bindings/python/tests/test_session.py
@@ -35,14 +35,14 @@ def test_explicit_session_groups_calls(self):
             calls = session_calls("sess-1")
             assert len(calls) == 2
             assert [c["step"] for c in calls] == [0, 1]
-            assert calls[0]["call_id"] != calls[1]["call_id"]
-            assert calls[0]["recorded_at"].endswith("Z")
+            assert calls[0]["callId"] != calls[1]["callId"]
+            assert calls[0]["recordedAt"].endswith("Z")
 
             # A call without session_id (inference off) is not grouped.
             generate_text(model, "third")
             sessions = list_sessions()
             assert len(sessions) == 1
-            assert sessions[0]["session_id"] == "sess-1"
+            assert sessions[0]["sessionId"] == "sess-1"
             assert sessions[0]["source"] == "Explicit"
 
             # Unknown session → empty.
@@ -71,7 +71,7 @@ def test_opt_in_inferer_groups_prefix_continuations(self):
             )
 
             sessions = list_sessions()
-            autos = [s for s in sessions if s["session_id"].startswith("auto-")]
+            autos = [s for s in sessions if s["sessionId"].startswith("auto-")]
             assert len(autos) == 1, "prefix continuation stays in one auto session"
             assert autos[0]["source"] == "Inferred"
             assert len(autos[0]["calls"]) == 2
diff --git a/bindings/python/tests/test_trace.py b/bindings/python/tests/test_trace.py
index 513b97f2b..c55759de9 100644
--- a/bindings/python/tests/test_trace.py
+++ b/bindings/python/tests/test_trace.py
@@ -40,18 +40,18 @@ def test_trace_records_calls_and_query_api(self):
             assert len(stats) == 1
             assert stats[0]["requests"] == 2
             assert stats[0]["provider"] == "deepseek.chat"
-            assert "verdict_counts" in stats[0]
+            assert "verdictCounts" in stats[0]
 
             chain = json.loads(traced.trace_session_chain("sess-1"))
-            assert len(chain["record_ids"]) == 2
-            assert chain["prefix_stability"] > 0.5
+            assert len(chain["recordIds"]) == 2
+            assert chain["prefixStability"] > 0.5
 
             jsonl = traced.trace_export_jsonl()
             lines = [l for l in jsonl.strip().split("\n") if l]
             assert len(lines) == 2, "one TraceRecord per line"
             first = json.loads(lines[0])
-            assert first["fingerprint"]["body_hash"]
-            assert first["session_id"] == "sess-1"
+            assert first["fingerprint"]["bodyHash"]
+            assert first["sessionId"] == "sess-1"
 
             traced.trace_clear()
             assert traced.trace_export_jsonl().strip() == ""
diff --git a/bindings/python/tests/test_transcription_session.py b/bindings/python/tests/test_transcription_session.py
index ac98385ba..e1ae4f807 100644
--- a/bindings/python/tests/test_transcription_session.py
+++ b/bindings/python/tests/test_transcription_session.py
@@ -185,7 +185,7 @@ def test_next_part_timeout_preserves_the_live_session(self):
             try:
                 first = session.next_part(timeout_ms=3000)
                 assert first is not None
-                assert "StreamStart" in json.loads(first)
+                assert json.loads(first)["type"] == "stream-start"
 
                 with pytest.raises(APITimeoutError) as ei:
                     session.next_part(timeout_ms=250)
@@ -198,7 +198,7 @@ def test_next_part_timeout_preserves_the_live_session(self):
                 # The session stayed live across the timeout.
                 late = session.next_part(timeout_ms=5000)
                 assert late is not None
-                assert "TranscriptDelta" in json.loads(late)
+                assert json.loads(late)["type"] == "transcript-delta"
             finally:
                 session.close()
 
@@ -212,7 +212,7 @@ def test_server_error_event_raises(self):
             try:
                 first = session.next_part(timeout_ms=5000)
                 assert first is not None
-                assert "StreamStart" in json.loads(first)
+                assert json.loads(first)["type"] == "stream-start"
 
                 with pytest.raises(APICallError) as ei:
                     session.next_part(timeout_ms=5000)
diff --git a/bindings/python/tests/test_wrapper.py b/bindings/python/tests/test_wrapper.py
index 39177f4de..86083ef43 100644
--- a/bindings/python/tests/test_wrapper.py
+++ b/bindings/python/tests/test_wrapper.py
@@ -19,7 +19,6 @@
     TextContentPart,
     FunctionTool,
     ToolChoiceTool,
-    StreamPart,
     generate_text,
     stream_text,
     parse_stream_part,
@@ -81,12 +80,12 @@ def test_parses_tool_calls(self):
 
         # raw.content carries the ToolCall variant as a typed GenerateContent.
         assert isinstance(result.raw.content, list)
-        tool_call_parts = [c for c in result.raw.content if c.root.type == "ToolCall"]
-        assert tool_call_parts, "raw.content must contain a ToolCall variant"
-        assert tool_call_parts[0].root.tool_name == "get_weather"
-        assert tool_call_parts[0].root.tool_call_id == "call_abc"
+        tool_call_parts = [c for c in result.raw.content if c.type == "tool-call"]
+        assert tool_call_parts, "raw.content must contain a tool-call variant"
+        assert tool_call_parts[0].tool_name == "get_weather"
+        assert tool_call_parts[0].tool_call_id == "call_abc"
         # raw content keeps the provider's argument text (see test_e2e).
-        assert tool_call_parts[0].root.input == '{"location":"Tokyo"}'
+        assert tool_call_parts[0].input == '{"location":"Tokyo"}'
 
     def test_raw_content_is_text_variant(self):
         with MockServer(OPENAI_CHAT) as mock:
@@ -95,8 +94,8 @@ def test_raw_content_is_text_variant(self):
 
         # raw.content[0] is a typed Text content part.
         first = result.raw.content[0]
-        assert first.root.type == "Text"
-        assert first.root.text == "Rust is a systems programming language."
+        assert first.type == "text"
+        assert first.text == "Rust is a systems programming language."
 
 
 # ── stream_text: typed-ish dicts ─────────────────────────────────────────────
@@ -109,7 +108,7 @@ def test_yields_text_delta_dicts(self):
             parts = list(stream_text(model, "Say hello"))
 
         assert len(parts) > 0
-        text = "".join(p["TextDelta"]["delta"] for p in parts if "TextDelta" in p)
+        text = "".join(p["delta"] for p in parts if p["type"] == "text-delta")
         assert text == "Hello world"
 
     def test_stream_part_can_be_typed(self):
@@ -117,15 +116,14 @@ def test_stream_part_can_be_typed(self):
         with MockServer(OPENAI_STREAM, content_type="text/event-stream") as mock:
             model = deepseek("test-key", "gpt-4o", mock.url)
             first_delta = next(
-                p for p in stream_text(model, "Say hello") if "TextDelta" in p
+                p for p in stream_text(model, "Say hello") if p["type"] == "text-delta"
             )
 
         sp = parse_stream_part(first_delta)
-        assert isinstance(sp, StreamPart)
-        assert sp.root.type == "TextDelta"
-        assert sp.root.delta == "Hello"
-        # Round-trips back to the external-tag dict form.
-        assert sp.model_dump(exclude_none=True) == first_delta
+        assert sp.type == "text-delta"
+        assert sp.delta == "Hello"
+        # Round-trips back to the wire dict form.
+        assert sp.model_dump(exclude_none=True, by_alias=True) == first_delta
 
     # ── RFC-0016 M2: include_raw_chunks ─────────────────────────────────────
 
@@ -139,19 +137,19 @@ def test_include_raw_chunks_emits_raw_parts(self):
                             GenerateTextOptions(include_raw_chunks=True))
             )
 
-        raw_parts = [p for p in parts if "Raw" in p]
+        raw_parts = [p for p in parts if p["type"] == "raw"]
         # OPENAI_STREAM has 3 JSON events (2 content chunks + 1 usage chunk);
         # the [DONE] sentinel emits no Raw.
         assert len(raw_parts) == 3
         # Each Raw carries the parsed JSON payload of the SSE event.
         assert (
-            raw_parts[0]["Raw"]["raw_value"]["choices"][0]["delta"]["content"]
+            raw_parts[0]["rawValue"]["choices"][0]["delta"]["content"]
             == "Hello"
         )
         # Raw for the first event precedes its TextDelta.
         first_raw_idx = parts.index(raw_parts[0])
         first_text_idx = next(
-            i for i, p in enumerate(parts) if "TextDelta" in p
+            i for i, p in enumerate(parts) if p["type"] == "text-delta"
         )
         assert first_raw_idx < first_text_idx
 
@@ -161,13 +159,13 @@ def test_raw_chunks_off_by_default(self):
             model = deepseek("test-key", "gpt-4o", mock.url)
             parts = list(stream_text(model, "Say hello"))
 
-        assert all("Raw" not in p for p in parts)
+        assert all(p["type"] != "raw" for p in parts)
 
     def test_include_raw_chunks_option_roundtrip(self):
         """Pure model-layer round-trip of the include_raw_chunks option."""
         opts = GenerateTextOptions(include_raw_chunks=True)
-        dumped = opts.model_dump_json(exclude_none=True)
-        assert '"include_raw_chunks":true' in dumped
+        dumped = _opts_to_json(opts)
+        assert '"includeRawChunks":true' in dumped
         back = GenerateTextOptions.model_validate_json(dumped)
         assert back.include_raw_chunks is True
 
@@ -230,18 +228,18 @@ def test_tool_choice_tool_variant_forces_tool_at_provider(self):
     def test_tool_choice_serializes_camel_case_tool_name(self):
         """At the wrapper→Rust JSON boundary, the ``tool`` tool-choice variant
         must serialize as ``{"type":"tool","toolName":"..."}`` (camelCase), the
-        hand-rolled serde wire format Rust expects. This is a deterministic
+        AI SDK shape Rust expects. This is a deterministic
         unit test of the boundary, independent of provider conversion."""
         opts = GenerateTextOptions(
             tools=[FunctionTool(name="get_weather", input_schema={"type": "object"})],
             tool_choice=ToolChoiceTool(tool_name="get_weather"),
         )
         wire = json.loads(_opts_to_json(opts))
-        assert wire["tool_choice"] == {"type": "tool", "toolName": "get_weather"}
+        assert wire["toolChoice"] == {"type": "tool", "toolName": "get_weather"}
 
         # And it round-trips back into the typed model.
         from aimux.wrapper import ToolChoiceTool as _TCT
-        back = _TCT.model_validate(wire["tool_choice"])
+        back = _TCT.model_validate(wire["toolChoice"])
         assert back.tool_name == "get_weather"
 
 
@@ -331,9 +329,9 @@ def test_round_trip_with_typed_content_parts(self):
                     role="assistant",
                     content=[
                         {
-                            "type": "tool_call",
-                            "tool_call_id": "call_abc",
-                            "tool_name": "get_weather",
+                            "type": "tool-call",
+                            "toolCallId": "call_abc",
+                            "toolName": "get_weather",
                             "input": {"location": "Tokyo"},
                         }
                     ],
@@ -342,9 +340,13 @@ def test_round_trip_with_typed_content_parts(self):
                     role="tool",
                     content=[
                         {
-                            "type": "tool_result",
-                            "tool_call_id": "call_abc",
-                            "result": {"temperature": 22, "condition": "sunny"},
+                            "type": "tool-result",
+                            "toolCallId": "call_abc",
+                            "toolName": "get_weather",
+                            "output": {
+                                "type": "json",
+                                "value": {"temperature": 22, "condition": "sunny"},
+                            },
                         }
                     ],
                 ),
diff --git a/bindings/swift/Sources/Aimux/Aimux.swift b/bindings/swift/Sources/Aimux/Aimux.swift
index 28962b8aa..a34384baf 100644
--- a/bindings/swift/Sources/Aimux/Aimux.swift
+++ b/bindings/swift/Sources/Aimux/Aimux.swift
@@ -176,7 +176,7 @@ public enum AimuxError: Error, LocalizedError, CustomStringConvertible, Equatabl
     /// The model produced tool arguments that fail to parse or validate.
     case invalidToolInput(message: String, status: Int, retryMs: Int64, retryable: Bool, toolName: String, toolInput: String)
     /// A `repairToolCall` hook itself failed; `originalError` is the error it
-    /// was repairing, as externally-tagged wire JSON (the same encoding as
+    /// was repairing, as `AiMuxError` JSON tagged by `name` (the same encoding as
     /// `ToolCall.error`).
     case toolCallRepair(message: String, status: Int, retryMs: Int64, retryable: Bool, originalError: String)
     /// No API key was passed and the fallback environment variable is unset
@@ -367,7 +367,7 @@ public enum AimuxError: Error, LocalizedError, CustomStringConvertible, Equatabl
     }
 
     /// `.toolCallRepair` only: the original lookup/parse/validation error as
-    /// externally-tagged wire JSON (the same encoding as `ToolCall.error`).
+    /// `AiMuxError` JSON tagged by `name` (the same encoding as `ToolCall.error`).
     public var originalError: String? {
         if case .toolCallRepair(_, _, _, _, let v) = self { return v }
         return nil
@@ -752,10 +752,10 @@ public final class Model: @unchecked Sendable {
     ///     registry entry.
     ///   - modelId: Model id.
     ///   - configJson: Optional JSON object of `ProviderOptions`
-    ///     (`{"base_url": "...", "headers": {...}, "organization": "...",
-    ///     "project": "...", "params": {...}}`); `nil` for defaults.
-    ///     `max_retries` (a per-call option) and `body_overrides` (removed) are
-    ///     rejected as `.invalidArgument`.
+    ///     (`{"baseUrl": "...", "headers": {...}, "organization": "...",
+    ///     "project": "...", "params": {...}}`); `nil` for defaults. Any
+    ///     other key (`maxRetries` is a per-call option) is rejected as
+    ///     `.invalidArgument`.
     public static func provider(
         name: String, apiKey: String? = nil, modelId: String, configJson: String? = nil
     ) throws -> Model {
@@ -850,8 +850,8 @@ public final class Model: @unchecked Sendable {
     ///     it does not trap: the array is as likely to come from a `.filter`
     ///     as from a literal, so C's zero-children failure surfaces instead).
     ///   - configJson: optional config: `{"router": "rule"|"weighted",
-    ///     "weights": [...], "fallback": "on_error"|"none", "provider_name",
-    ///     "model_id"}`.
+    ///     "weights": [...], "fallback": "on_error"|"none", "providerName",
+    ///     "modelId"}`.
     /// - Returns: a new RouterModel wrapping the children.
     public static func router(
         _ models: [Model],
@@ -900,21 +900,6 @@ public final class Model: @unchecked Sendable {
         return Model(handle: handle)
     }
 
-    /// Register external OpenAI-compatible providers from a JSON config string
-    /// (RFC-0020).
-    ///
-    /// `configJSON` is `{ "providers": [ { "name", "base_url", ... } ] }`.
-    /// Entries override same-named built-ins or add new ones. Like
-    /// `initRecording`, this mutates process-global registry state.
-    ///
-    /// - Parameter configJSON: Provider registry config JSON.
-    /// - Throws: `AimuxError` (`.invalidArgument`) when the registry rejects
-    ///   the document.
-    public static func registerProviders(_ configJSON: String) throws {
-        try validateJson(configJSON, parameter: "configJSON")
-        if let e = aimux_register_providers(configJSON) { throw expectAimuxError(e, context: "registerProviders") }
-    }
-
     /// Set the global proxy configuration (M6, RFC-0016). Must be called before
     /// the first `generateText` / `streamText` call; a no-op (returns without
     /// error) if the shared HTTP client is already initialised.
@@ -963,7 +948,7 @@ public final class Model: @unchecked Sendable {
     /// Generate a structured JSON object (M12, RFC-0016).
     ///
     /// Same signature as `generateText`; returns a JSON-serialized
-    /// `GenerateObjectResult`. Pass `response_format: { "Json": { ... } }`
+    /// `GenerateObjectResult`. Pass `responseFormat: { "type": "json", ... }`
     /// via `options` for schema control; aimux-core applies JSON repair
     /// before parsing.
     ///
@@ -1028,7 +1013,7 @@ public final class Model: @unchecked Sendable {
     /// - Parameters:
     ///   - prompt: A prompt string (serialized as JSON).
     ///   - options: Optional GenerateTextOptions (serialized as JSON).
-    ///   - onPart: Called for each StreamPart (JSON string).
+    ///   - onPart: Called for each TextStreamPart (JSON string).
     ///   - onDone: Called when the stream completes normally.
     ///   - onError: Called on stream failure with the decoded error unchanged
     ///     (`AimuxError`, or `DecodingError.dataCorrupted` for unparseable
@@ -1099,7 +1084,7 @@ public final class Model: @unchecked Sendable {
         }
     }
 
-    /// Stream text as an AsyncSequence of StreamPart JSON strings.
+    /// Stream text as an AsyncSequence of TextStreamPart JSON strings.
     ///
     /// Usage:
     /// ```swift
diff --git a/bindings/swift/Sources/Aimux/Multimodal.swift b/bindings/swift/Sources/Aimux/Multimodal.swift
index 7278d9e8f..1f5535023 100644
--- a/bindings/swift/Sources/Aimux/Multimodal.swift
+++ b/bindings/swift/Sources/Aimux/Multimodal.swift
@@ -201,8 +201,8 @@ public final class TranscriptionModel: @unchecked Sendable {
     ///
     /// - Parameters:
     ///   - options: Optional session options JSON
-    ///     (`{"input_audio_format": {...}, "provider_options", "headers",
-    ///     "include_raw_chunks"}`).
+    ///     (`{"inputAudioFormat": {...}, "providerOptions", "headers",
+    ///     "includeRawChunks"}`).
     ///   - abortHandle: Optional abort handle (`aimux_abort_signal_new`);
     ///     firing it aborts the session.
     public func startStream(options: String? = nil, abortHandle: UInt64 = 0) throws -> TranscriptionSession {
@@ -490,108 +490,41 @@ public final class Files: @unchecked Sendable {
 // ─────────────────────────────────────────────────────────────────────────────
 // Multimodal result types (Codable)
 //
-// These mirror the Go structs in bindings/go/multimodal_types.go. Field names
-// are snake_case on the wire; Swift camelCase is restored via CodingKeys (the
-// same approach used in Types.swift). `provider_metadata` uses the existing
-// `JSONValue` type; `warnings` is decoded leniently as `[JSONValue]` (matching
-// Go's `json.RawMessage`), so unknown warning shapes don't break decoding.
+// These decode the JSON of the generated `bindings/node/src/types/*Result.ts`
+// types: camelCase fields (so the structs synthesize their `Codable`),
+// optional fields absent rather than null, `warnings` typed as `[Warning]`.
 // ─────────────────────────────────────────────────────────────────────────────
 
-/// A `CodingKey` accepting any string, used to decode/encode the externally
-/// tagged multimodal enums (`AudioData`, `ImageOutputs`, `VideoData`). Mirrors
-/// `AnyCodingKey` in Types.swift, which is file-private and unavailable here.
-fileprivate struct MMCodingKey: CodingKey {
-    let stringValue: String
-    init?(stringValue: String) { self.stringValue = stringValue }
-    var intValue: Int? { nil }
-    init?(intValue: Int) { return nil }
-    init(_ s: String) { self.stringValue = s }
-}
-
-/// Build a `DecodingError.dataCorrupted` for a coding path + message.
-fileprivate func mmDecodingError(_ codingPath: [any CodingKey], _ message: String) -> DecodingError {
-    DecodingError.dataCorrupted(.init(codingPath: codingPath, debugDescription: message, underlyingError: nil))
-}
-
 // MARK: - Embedding
 
-/// Token usage for an embedding call (input tokens only).
+/// Token usage for an embedding call.
 public struct EmbeddingUsage: Codable, Equatable {
-    public var tokens: UInt32?
-
-    public init(tokens: UInt32? = nil) {
-        self.tokens = tokens
-    }
+    public var tokens: UInt32
 }
 
 /// Provider response metadata for embeddings.
 public struct EmbeddingResponse: Codable, Equatable {
     public var headers: [String: String]?
     public var body: JSONValue?
-
-    enum CodingKeys: String, CodingKey {
-        case headers
-        case body
-    }
-
-    public init(headers: [String: String]? = nil, body: JSONValue? = nil) {
-        self.headers = headers; self.body = body
-    }
 }
 
 /// The result of an embedding call.
 public struct EmbeddingResult: Codable, Equatable {
     public var embeddings: [[Float]]
     public var usage: EmbeddingUsage?
+    public var providerMetadata: ProviderMetadata?
     public var response: EmbeddingResponse?
-    public var providerMetadata: JSONValue?
-    public var warnings: [JSONValue]?
-
-    enum CodingKeys: String, CodingKey {
-        case embeddings, usage, response
-        case providerMetadata = "provider_metadata"
-        case warnings
-    }
-
-    public init(embeddings: [[Float]], usage: EmbeddingUsage? = nil, response: EmbeddingResponse? = nil,
-                providerMetadata: JSONValue? = nil, warnings: [JSONValue]? = nil) {
-        self.embeddings = embeddings; self.usage = usage; self.response = response
-        self.providerMetadata = providerMetadata; self.warnings = warnings
-    }
+    public var warnings: [Warning]
 }
 
 // MARK: - Speech (TTS)
 
-/// Generated audio: a base64-encoded string or raw binary bytes.
-/// Wire: `{"Base64":"…"}` | `{"Binary":[…]}`.
-public enum AudioData: Codable, Equatable {
-    case base64(String)
-    case binary([UInt8])
+/// Generated audio: a base64-encoded string or raw bytes. Wire: untagged
+/// (`"…"` | `[…]`), the same shape as ``FileBytes``.
+public typealias AudioData = FileBytes
 
-    public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: MMCodingKey.self)
-        guard let key = c.allKeys.first?.stringValue else {
-            throw mmDecodingError(c.codingPath, "missing audio-data tag")
-        }
-        switch key {
-        case "Base64":
-            self = .base64(try c.decode(String.self, forKey: MMCodingKey(key)))
-        case "Binary":
-            self = .binary(try c.decode([UInt8].self, forKey: MMCodingKey(key)))
-        default:
-            throw mmDecodingError(c.codingPath, "unknown audio-data tag \(key)")
-        }
-    }
-
-    public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: MMCodingKey.self)
-        switch self {
-        case .base64(let s):
-            try c.encode(s, forKey: MMCodingKey("Base64"))
-        case .binary(let bytes):
-            try c.encode(bytes, forKey: MMCodingKey("Binary"))
-        }
-    }
+public struct SpeechRequest: Codable, Equatable {
+    public var body: JSONValue?
 }
 
 /// Response information for a speech call.
@@ -600,71 +533,37 @@ public struct SpeechResponse: Codable, Equatable {
     public var modelId: String?
     public var headers: [String: String]?
     public var body: JSONValue?
-
-    enum CodingKeys: String, CodingKey {
-        case timestamp
-        case modelId = "model_id"
-        case headers
-        case body
-    }
-
-    public init(timestamp: String? = nil, modelId: String? = nil,
-                headers: [String: String]? = nil, body: JSONValue? = nil) {
-        self.timestamp = timestamp; self.modelId = modelId
-        self.headers = headers; self.body = body
-    }
 }
 
 /// The result of a speech generation call.
 public struct SpeechResult: Codable, Equatable {
     public var audio: AudioData
+    public var warnings: [Warning]
+    public var request: SpeechRequest?
     public var response: SpeechResponse
-    public var providerMetadata: JSONValue?
-    public var warnings: [JSONValue]?
-
-    enum CodingKeys: String, CodingKey {
-        case audio, response
-        case providerMetadata = "provider_metadata"
-        case warnings
-    }
-
-    public init(audio: AudioData, response: SpeechResponse,
-                providerMetadata: JSONValue? = nil, warnings: [JSONValue]? = nil) {
-        self.audio = audio; self.response = response
-        self.providerMetadata = providerMetadata; self.warnings = warnings
-    }
+    public var providerMetadata: ProviderMetadata?
 }
 
 // MARK: - Image
 
-/// Generated images: all base64 strings or all binary byte arrays.
-/// Wire: `{"Base64":["…"]}` | `{"Binary":[[…]]}`.
+/// Generated images: all base64 strings or all binary byte arrays. Wire:
+/// untagged (`["…"]` | `[[…]]`).
 public enum ImageOutputs: Codable, Equatable {
     case base64([String])
     case binary([[UInt8]])
 
     public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: MMCodingKey.self)
-        guard let key = c.allKeys.first?.stringValue else {
-            throw mmDecodingError(c.codingPath, "missing image-outputs tag")
-        }
-        switch key {
-        case "Base64":
-            self = .base64(try c.decode([String].self, forKey: MMCodingKey(key)))
-        case "Binary":
-            self = .binary(try c.decode([[UInt8]].self, forKey: MMCodingKey(key)))
-        default:
-            throw mmDecodingError(c.codingPath, "unknown image-outputs tag \(key)")
-        }
+        let c = try decoder.singleValueContainer()
+        if let strings = try? c.decode([String].self) { self = .base64(strings); return }
+        if let batches = try? c.decode([[UInt8]].self) { self = .binary(batches); return }
+        throw aimuxDecodingError(c.codingPath, "expected base64 strings or byte arrays")
     }
 
     public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: MMCodingKey.self)
+        var c = encoder.singleValueContainer()
         switch self {
-        case .base64(let strings):
-            try c.encode(strings, forKey: MMCodingKey("Base64"))
-        case .binary(let batches):
-            try c.encode(batches, forKey: MMCodingKey("Binary"))
+        case .base64(let strings): try c.encode(strings)
+        case .binary(let batches): try c.encode(batches)
         }
     }
 }
@@ -674,36 +573,21 @@ public struct ImageResponse: Codable, Equatable {
     public var timestamp: String?
     public var modelId: String?
     public var headers: [String: String]?
+}
 
-    enum CodingKeys: String, CodingKey {
-        case timestamp
-        case modelId = "model_id"
-        case headers
-    }
-
-    public init(timestamp: String? = nil, modelId: String? = nil, headers: [String: String]? = nil) {
-        self.timestamp = timestamp; self.modelId = modelId; self.headers = headers
-    }
+public struct ImageUsage: Codable, Equatable {
+    public var inputTokens: UInt32?
+    public var outputTokens: UInt32?
+    public var totalTokens: UInt32?
 }
 
 /// The result of an image generation call.
 public struct ImageResult: Codable, Equatable {
     public var images: ImageOutputs
+    public var warnings: [Warning]
+    public var providerMetadata: ProviderMetadata?
     public var response: ImageResponse
-    public var providerMetadata: JSONValue?
-    public var warnings: [JSONValue]?
-
-    enum CodingKeys: String, CodingKey {
-        case images, response
-        case providerMetadata = "provider_metadata"
-        case warnings
-    }
-
-    public init(images: ImageOutputs, response: ImageResponse,
-                providerMetadata: JSONValue? = nil, warnings: [JSONValue]? = nil) {
-        self.images = images; self.response = response
-        self.providerMetadata = providerMetadata; self.warnings = warnings
-    }
+    public var usage: ImageUsage?
 }
 
 // MARK: - Transcription (STT)
@@ -711,28 +595,12 @@ public struct ImageResult: Codable, Equatable {
 /// A transcript segment with timing.
 public struct TranscriptionSegment: Codable, Equatable {
     public var text: String
-    /// Required `f64` in the core, not an `Option` — see `TranscriptionSegment.ts` —
-    /// so these stay non-optional and are always encoded; a provider that omits
-    /// them decodes to `0` rather than failing.
     public var startSecond: Double
     public var endSecond: Double
+}
 
-    enum CodingKeys: String, CodingKey {
-        case text
-        case startSecond = "start_second"
-        case endSecond = "end_second"
-    }
-
-    public init(text: String, startSecond: Double, endSecond: Double) {
-        self.text = text; self.startSecond = startSecond; self.endSecond = endSecond
-    }
-
-    public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: CodingKeys.self)
-        self.text = try c.decodeIfPresent(String.self, forKey: .text) ?? ""
-        self.startSecond = try c.decodeIfPresent(Double.self, forKey: .startSecond) ?? 0
-        self.endSecond = try c.decodeIfPresent(Double.self, forKey: .endSecond) ?? 0
-    }
+public struct TranscriptionRequest: Codable, Equatable {
+    public var body: String?
 }
 
 /// Response information for a transcription call.
@@ -741,125 +609,55 @@ public struct TranscriptionResponse: Codable, Equatable {
     public var modelId: String?
     public var headers: [String: String]?
     public var body: JSONValue?
-
-    enum CodingKeys: String, CodingKey {
-        case timestamp
-        case modelId = "model_id"
-        case headers
-        case body
-    }
-
-    public init(timestamp: String? = nil, modelId: String? = nil,
-                headers: [String: String]? = nil, body: JSONValue? = nil) {
-        self.timestamp = timestamp; self.modelId = modelId
-        self.headers = headers; self.body = body
-    }
 }
 
 /// The result of a transcription call.
 public struct TranscriptionResult: Codable, Equatable {
     public var text: String
-    public var segments: [TranscriptionSegment]?
+    public var segments: [TranscriptionSegment]
     public var language: String?
+    public var durationInSeconds: Double?
+    public var warnings: [Warning]
+    public var request: TranscriptionRequest?
     public var response: TranscriptionResponse
-    public var providerMetadata: JSONValue?
-    public var warnings: [JSONValue]?
-
-    enum CodingKeys: String, CodingKey {
-        case text, segments, language, response
-        case providerMetadata = "provider_metadata"
-        case warnings
-    }
-
-    public init(text: String, response: TranscriptionResponse,
-                segments: [TranscriptionSegment]? = nil, language: String? = nil,
-                providerMetadata: JSONValue? = nil, warnings: [JSONValue]? = nil) {
-        self.text = text; self.response = response
-        self.segments = segments; self.language = language
-        self.providerMetadata = providerMetadata; self.warnings = warnings
-    }
+    public var providerMetadata: ProviderMetadata?
 }
 
 // MARK: - Video
 
-/// The URL variant of `VideoData`.
-public struct VideoUrlData: Codable, Equatable {
-    public var url: String
-    public var mediaType: String?
-
-    enum CodingKeys: String, CodingKey {
-        case url
-        case mediaType = "media_type"
-    }
-
-    public init(url: String, mediaType: String? = nil) {
-        self.url = url; self.mediaType = mediaType
-    }
-}
-
-/// The base64 variant of `VideoData`.
-public struct VideoBase64Data: Codable, Equatable {
-    public var data: String
-    public var mediaType: String?
-
-    enum CodingKeys: String, CodingKey {
-        case data
-        case mediaType = "media_type"
-    }
-
-    public init(data: String, mediaType: String? = nil) {
-        self.data = data; self.mediaType = mediaType
-    }
-}
-
-/// The binary variant of `VideoData`.
-public struct VideoBinaryData: Codable, Equatable {
-    public var data: [UInt8]
-    public var mediaType: String?
-
-    enum CodingKeys: String, CodingKey {
-        case data
-        case mediaType = "media_type"
-    }
-
-    public init(data: [UInt8], mediaType: String? = nil) {
-        self.data = data; self.mediaType = mediaType
-    }
-}
-
-/// Generated video: a download URL, a base64 string, or raw binary bytes.
-/// Wire: `{"Url":{…}}` | `{"Base64":{…}}` | `{"Binary":{…}}`.
+/// Generated video. Wire: `{"type":"url","url","mediaType"}`,
+/// `{"type":"base64","data":"…","mediaType"}` or
+/// `{"type":"binary","data":[…],"mediaType"}`.
 public enum VideoData: Codable, Equatable {
-    case url(VideoUrlData)
-    case base64(VideoBase64Data)
-    case binary(VideoBinaryData)
+    case url(url: String, mediaType: String)
+    case base64(data: String, mediaType: String)
+    case binary(data: [UInt8], mediaType: String)
+
+    private enum Field: String, CodingKey { case type, url, data, mediaType }
 
     public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: MMCodingKey.self)
-        guard let key = c.allKeys.first?.stringValue else {
-            throw mmDecodingError(c.codingPath, "missing video-data tag")
-        }
-        switch key {
-        case "Url":
-            self = .url(try c.decode(VideoUrlData.self, forKey: MMCodingKey(key)))
-        case "Base64":
-            self = .base64(try c.decode(VideoBase64Data.self, forKey: MMCodingKey(key)))
-        case "Binary":
-            self = .binary(try c.decode(VideoBinaryData.self, forKey: MMCodingKey(key)))
-        default:
-            throw mmDecodingError(c.codingPath, "unknown video-data tag \(key)")
+        let c = try decoder.container(keyedBy: Field.self)
+        let mediaType = try c.decode(String.self, forKey: .mediaType)
+        switch try c.decode(String.self, forKey: .type) {
+        case "url": self = .url(url: try c.decode(String.self, forKey: .url), mediaType: mediaType)
+        case "base64": self = .base64(data: try c.decode(String.self, forKey: .data), mediaType: mediaType)
+        case "binary": self = .binary(data: try c.decode([UInt8].self, forKey: .data), mediaType: mediaType)
+        case let t: throw decoder.unknownTag(t, in: "video data")
         }
     }
 
     public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: MMCodingKey.self)
+        var c = encoder.container(keyedBy: Field.self)
         switch self {
-        case .url(let d):
-            try c.encode(d, forKey: MMCodingKey("Url"))
-        case .base64(let d):
-            try c.encode(d, forKey: MMCodingKey("Base64"))
-        case .binary(let d):
-            try c.encode(d, forKey: MMCodingKey("Binary"))
+        case .url(let url, let mediaType):
+            try c.encode("url", forKey: .type); try c.encode(url, forKey: .url)
+            try c.encode(mediaType, forKey: .mediaType)
+        case .base64(let data, let mediaType):
+            try c.encode("base64", forKey: .type); try c.encode(data, forKey: .data)
+            try c.encode(mediaType, forKey: .mediaType)
+        case .binary(let data, let mediaType):
+            try c.encode("binary", forKey: .type); try c.encode(data, forKey: .data)
+            try c.encode(mediaType, forKey: .mediaType)
         }
     }
 }
@@ -869,36 +667,14 @@ public struct VideoResponse: Codable, Equatable {
     public var timestamp: String?
     public var modelId: String?
     public var headers: [String: String]?
-
-    enum CodingKeys: String, CodingKey {
-        case timestamp
-        case modelId = "model_id"
-        case headers
-    }
-
-    public init(timestamp: String? = nil, modelId: String? = nil, headers: [String: String]? = nil) {
-        self.timestamp = timestamp; self.modelId = modelId; self.headers = headers
-    }
 }
 
 /// The result of a video generation call.
 public struct VideoResult: Codable, Equatable {
     public var videos: [VideoData]
+    public var warnings: [Warning]
+    public var providerMetadata: ProviderMetadata?
     public var response: VideoResponse
-    public var providerMetadata: JSONValue?
-    public var warnings: [JSONValue]?
-
-    enum CodingKeys: String, CodingKey {
-        case videos, response
-        case providerMetadata = "provider_metadata"
-        case warnings
-    }
-
-    public init(videos: [VideoData], response: VideoResponse,
-                providerMetadata: JSONValue? = nil, warnings: [JSONValue]? = nil) {
-        self.videos = videos; self.response = response
-        self.providerMetadata = providerMetadata; self.warnings = warnings
-    }
 }
 
 // MARK: - Reranking
@@ -907,15 +683,6 @@ public struct VideoResult: Codable, Equatable {
 public struct RerankingRank: Codable, Equatable {
     public var index: Int
     public var relevanceScore: Double
-
-    enum CodingKeys: String, CodingKey {
-        case index
-        case relevanceScore = "relevance_score"
-    }
-
-    public init(index: Int, relevanceScore: Double) {
-        self.index = index; self.relevanceScore = relevanceScore
-    }
 }
 
 /// Optional response information for a reranking call.
@@ -925,40 +692,14 @@ public struct RerankingResponse: Codable, Equatable {
     public var modelId: String?
     public var headers: [String: String]?
     public var body: JSONValue?
-
-    enum CodingKeys: String, CodingKey {
-        case id
-        case timestamp
-        case modelId = "model_id"
-        case headers
-        case body
-    }
-
-    public init(id: String? = nil, timestamp: String? = nil, modelId: String? = nil,
-                headers: [String: String]? = nil, body: JSONValue? = nil) {
-        self.id = id; self.timestamp = timestamp; self.modelId = modelId
-        self.headers = headers; self.body = body
-    }
 }
 
 /// The result of a reranking call.
 public struct RerankingResult: Codable, Equatable {
     public var ranking: [RerankingRank]
+    public var providerMetadata: ProviderMetadata?
+    public var warnings: [Warning]?
     public var response: RerankingResponse?
-    public var providerMetadata: JSONValue?
-    public var warnings: [JSONValue]?
-
-    enum CodingKeys: String, CodingKey {
-        case ranking, response
-        case providerMetadata = "provider_metadata"
-        case warnings
-    }
-
-    public init(ranking: [RerankingRank], response: RerankingResponse? = nil,
-                providerMetadata: JSONValue? = nil, warnings: [JSONValue]? = nil) {
-        self.ranking = ranking; self.response = response
-        self.providerMetadata = providerMetadata; self.warnings = warnings
-    }
 }
 
 // MARK: - Search
@@ -970,56 +711,22 @@ public struct SearchResultItem: Codable, Equatable {
     public var content: String?
     public var rawContent: String?
     public var score: Double?
-    public var providerMetadata: [String: JSONValue]?
-
-    enum CodingKeys: String, CodingKey {
-        case title, url, content
-        case rawContent = "raw_content"
-        case score
-        case providerMetadata = "provider_metadata"
-    }
-
-    public init(title: String? = nil, url: String? = nil, content: String? = nil,
-                rawContent: String? = nil, score: Double? = nil, providerMetadata: [String: JSONValue]? = nil) {
-        self.title = title; self.url = url; self.content = content
-        self.rawContent = rawContent; self.score = score; self.providerMetadata = providerMetadata
-    }
+    public var providerMetadata: ProviderMetadata?
 }
 
 /// Optional response information for a search call.
 public struct SearchResponse: Codable, Equatable {
     public var headers: [String: String]?
     public var body: JSONValue?
-
-    enum CodingKeys: String, CodingKey {
-        case headers
-        case body
-    }
-
-    public init(headers: [String: String]? = nil, body: JSONValue? = nil) {
-        self.headers = headers; self.body = body
-    }
 }
 
 /// The result of a search call.
 public struct SearchResult: Codable, Equatable {
     public var results: [SearchResultItem]
     public var answer: String?
+    public var providerMetadata: ProviderMetadata?
+    public var warnings: [Warning]
     public var response: SearchResponse?
-    public var providerMetadata: JSONValue?
-    public var warnings: [JSONValue]?
-
-    enum CodingKeys: String, CodingKey {
-        case results, answer, response
-        case providerMetadata = "provider_metadata"
-        case warnings
-    }
-
-    public init(results: [SearchResultItem], answer: String? = nil, response: SearchResponse? = nil,
-                providerMetadata: JSONValue? = nil, warnings: [JSONValue]? = nil) {
-        self.results = results; self.answer = answer; self.response = response
-        self.providerMetadata = providerMetadata; self.warnings = warnings
-    }
 }
 
 // MARK: - Files
@@ -1029,31 +736,17 @@ public struct UploadFileResult: Codable, Equatable {
     public var providerReference: [String: String]
     public var mediaType: String?
     public var filename: String?
-    public var providerMetadata: JSONValue?
-    public var warnings: [JSONValue]?
-
-    enum CodingKeys: String, CodingKey {
-        case providerReference = "provider_reference"
-        case mediaType = "media_type"
-        case filename
-        case providerMetadata = "provider_metadata"
-        case warnings
-    }
-
-    public init(providerReference: [String: String], mediaType: String? = nil, filename: String? = nil,
-                providerMetadata: JSONValue? = nil, warnings: [JSONValue]? = nil) {
-        self.providerReference = providerReference
-        self.mediaType = mediaType; self.filename = filename
-        self.providerMetadata = providerMetadata; self.warnings = warnings
-    }
+    public var providerMetadata: ProviderMetadata?
+    public var warnings: [Warning]
 }
 
+
 // ─────────────────────────────────────────────────────────────────────────────
 // TranscriptionSession (RFC-0028 streaming)
 // ─────────────────────────────────────────────────────────────────────────────
 
 /// Thrown by `TranscriptionSession.nextPart` when the stream ended normally
-/// (a `Finish` part was delivered earlier).
+/// (a `finish` part was delivered earlier).
 ///
 /// A session the caller `close()`d does NOT throw this — it throws the Swift
 /// boundary projection (`DecodingError.dataCorrupted`, "transcription session
diff --git a/bindings/swift/Sources/Aimux/ToolCallRepair.swift b/bindings/swift/Sources/Aimux/ToolCallRepair.swift
index af5bd09a7..5d227a7c9 100644
--- a/bindings/swift/Sources/Aimux/ToolCallRepair.swift
+++ b/bindings/swift/Sources/Aimux/ToolCallRepair.swift
@@ -33,20 +33,11 @@ public struct RawToolCall: Codable, Equatable {
     public var input: String
     public var providerExecuted: Bool?
     public var dynamic: Bool?
-    public var providerMetadata: JSONValue?
-
-    enum CodingKeys: String, CodingKey {
-        case toolCallId = "tool_call_id"
-        case toolName = "tool_name"
-        case input
-        case providerExecuted = "provider_executed"
-        case dynamic
-        case providerMetadata = "provider_metadata"
-    }
+    public var providerMetadata: ProviderMetadata?
 
     public init(toolCallId: String, toolName: String, input: String,
                 providerExecuted: Bool? = nil, dynamic: Bool? = nil,
-                providerMetadata: JSONValue? = nil) {
+                providerMetadata: ProviderMetadata? = nil) {
         self.toolCallId = toolCallId; self.toolName = toolName; self.input = input
         self.providerExecuted = providerExecuted; self.dynamic = dynamic
         self.providerMetadata = providerMetadata
@@ -58,9 +49,9 @@ public struct RawToolCall: Codable, Equatable {
 public struct ToolCallRepairContext: Codable, Equatable {
     /// The failing call, with the provider's raw argument text.
     public var toolCall: RawToolCall
-    /// The failure the call already carries, as externally-tagged wire JSON
-    /// (the same encoding as ``ToolCall/error``), e.g.
-    /// `{"InvalidToolInput":{"tool_name":…,"tool_input":…,"cause":…}}`.
+    /// The failure the call already carries: an `AiMuxError` object tagged by
+    /// `name` (the same encoding as ``ToolCall/error``), e.g.
+    /// `{"name":"AI_InvalidToolInputError","toolName":…,"toolInput":…,"cause":…}`.
     public var error: JSONValue
     /// JSON Schema of the named tool. A name that does not resolve to a
     /// function tool yields the AI SDK's default empty-object schema.
@@ -72,13 +63,6 @@ public struct ToolCallRepairContext: Codable, Equatable {
     /// The instructions the call was generated with, if any.
     public var instructions: String?
 
-    enum CodingKeys: String, CodingKey {
-        case toolCall = "tool_call"
-        case error
-        case inputSchema = "input_schema"
-        case tools, messages, instructions
-    }
-
     public init(toolCall: RawToolCall, error: JSONValue, inputSchema: JSONValue,
                 tools: [Tool], messages: [ModelMessage], instructions: String? = nil) {
         self.toolCall = toolCall; self.error = error; self.inputSchema = inputSchema
@@ -106,7 +90,7 @@ struct RepairToolCallBox: Equatable {
 /// Build the repair argument for one invalid tool call.
 ///
 /// - Parameters:
-///   - toolCall: One `GenerateTextResult.tool_calls` entry with `"invalid": true`.
+///   - toolCall: One `GenerateTextResult.toolCalls` entry with `"invalid": true`.
 ///   - prompt: The prompt JSON the call was generated with.
 ///   - options: The options JSON the call was generated with.
 /// - Returns: The serialized ``ToolCallRepairContext``, or the JSON literal
@@ -138,8 +122,8 @@ public func applyToolCallRepair(
 }
 
 /// Apply a repair reply to a serialized `GenerateTextResult` or
-/// `GenerateObjectResult`. Both `tool_calls` and the matching
-/// `response_messages` tool-call part are rewritten — the transcript must keep
+/// `GenerateObjectResult`. Both `toolCalls` and the matching
+/// `responseMessages` tool-call part are rewritten — the transcript must keep
 /// pointing at the same entry, or the next turn replays the unrepaired
 /// arguments.
 public func applyToolCallRepairToResult(
@@ -161,7 +145,7 @@ public func applyToolCallRepairToResult(
 /// the patched JSON (the input unchanged when no repair function is set).
 ///
 /// Handles `GenerateTextResult` and `StreamTextResultAggregated` (top-level
-/// `tool_calls`) as well as `GenerateObjectResult` (`raw.tool_calls`). Calls
+/// `toolCalls`) as well as `GenerateObjectResult` (`raw.toolCalls`). Calls
 /// are visited in document order and each gets at most one repair attempt, as
 /// in the AI SDK.
 func repairedResultJson(
@@ -172,12 +156,12 @@ func repairedResultJson(
     guard let document = (try? JSONSerialization.jsonObject(with: Data(resultJson.utf8)))
         as? [String: Any] else { return resultJson }
     // A GenerateObjectResult keeps the whole text result under `raw`.
-    let target = document["tool_calls"] != nil ? document : document["raw"] as? [String: Any]
-    guard let calls = target?["tool_calls"] as? [[String: Any]] else { return resultJson }
+    let target = document["toolCalls"] != nil ? document : document["raw"] as? [String: Any]
+    guard let calls = target?["toolCalls"] as? [[String: Any]] else { return resultJson }
 
     var patched = resultJson
     for call in calls where call["invalid"] as? Bool == true {
-        guard let id = call["tool_call_id"] as? String, let callJson = jsonString(call) else { continue }
+        guard let id = call["toolCallId"] as? String, let callJson = jsonString(call) else { continue }
         let contextJson = try toolCallRepairContext(
             toolCall: callJson, prompt: promptJson, options: optsJson
         )
@@ -190,9 +174,10 @@ func repairedResultJson(
     return patched
 }
 
-/// Repair a `{"ToolCall": …}` stream part, returning the replacement part JSON.
+/// Repair a `{"type":"tool-call", …}` stream part, returning the replacement
+/// part JSON.
 ///
-/// Any other part — `ToolInputDelta` above all — is returned untouched and
+/// Any other part — `tool-input-delta` above all — is returned untouched and
 /// immediately: the AI SDK forwards argument deltas verbatim even when a repair
 /// function is configured.
 func repairedStreamPartJson(
@@ -201,9 +186,9 @@ func repairedStreamPartJson(
 ) throws -> String {
     guard let repair = options?.repairToolCall,
           let part = (try? JSONSerialization.jsonObject(with: Data(partJson.utf8))) as? [String: Any],
-          let call = part["ToolCall"] as? [String: Any],
-          call["invalid"] as? Bool == true,
-          let callJson = jsonString(call)
+          part["type"] as? String == "tool-call",
+          part["invalid"] as? Bool == true,
+          let callJson = jsonString(part)
     else { return partJson }
 
     // The two library calls are pure — no runtime, no `ffi_block_on` — so they
@@ -217,8 +202,10 @@ func repairedStreamPartJson(
     let repairedJson = try applyToolCallRepair(
         toolCall: callJson, options: optsJson, reply: reply
     )
-    let repaired = try JSONSerialization.jsonObject(with: Data(repairedJson.utf8))
-    return jsonString(["ToolCall": repaired]) ?? partJson
+    guard var repaired = try JSONSerialization.jsonObject(with: Data(repairedJson.utf8)) as? [String: Any]
+    else { return partJson }
+    repaired["type"] = "tool-call"
+    return jsonString(repaired) ?? partJson
 }
 
 /// Run `body` on another thread and block until it answers.
@@ -263,7 +250,7 @@ private func repairReply(_ context: ToolCallRepairContext, _ repair: RepairToolC
     do {
         guard let replacement = try repair(context) else { return #"{"type":"unchanged"}"# }
         let encoded = try JSONEncoder().encode(replacement)
-        return #"{"type":"repaired","tool_call":"# + String(decoding: encoded, as: UTF8.self) + "}"
+        return #"{"type":"repaired","toolCall":"# + String(decoding: encoded, as: UTF8.self) + "}"
     } catch {
         let message = (error as? LocalizedError)?.errorDescription ?? String(describing: error)
         return jsonString(["type": "failed", "message": message])
diff --git a/bindings/swift/Sources/Aimux/Types.swift b/bindings/swift/Sources/Aimux/Types.swift
index 2b6b8b59b..355b470a6 100644
--- a/bindings/swift/Sources/Aimux/Types.swift
+++ b/bindings/swift/Sources/Aimux/Types.swift
@@ -2,20 +2,19 @@
 //
 // The raw API in `Aimux.swift` exchanges JSON strings across the Swift↔C
 // boundary (`generateText(prompt:options:) -> String`, `streamText` yields
-// JSON-string `StreamPart`s). This file adds a thin, *typed* layer on top:
-// inputs and outputs are `Codable` Swift structs/enums mirroring
-// `bindings/node/src/types/*.ts` (the ts-rs types generated from the Rust serde
-// definitions). The raw API is left untouched; the typed methods live in a
-// `Model` extension and delegate to the raw ones.
+// JSON-string parts). This file adds a thin, *typed* layer on top: inputs and
+// outputs are `Codable` Swift types decoding and encoding exactly the JSON of
+// `bindings/node/src/types/*.ts` (the ts-rs types generated from the Rust
+// serde definitions, which equal the AI SDK JSON). The raw API is left
+// untouched; the typed methods live in a `Model` extension and delegate to it.
 //
-// Wire conventions mirrored here (all derived from the Rust serde attributes):
-//   • struct fields are snake_case on the wire → Swift camelCase via CodingKeys
-//   • `Tool` / `ContentPart`  → internally tagged by `type` (variant = snake_case)
-//   • `ToolChoice`           → mixed: "auto"|"none"|"required" or {"type":"tool","toolName":...}
-//                              (note: `toolName` is camelCase on the wire)
-//   • `ResponseFormat`       → external tag with a unit variant: "Text" or {"Json":{...}}
-//   • `ModelPrompt` / `MessageContent` → untagged (string | array)
-//   • `GenerateContent` / `StreamPart` / `Warning` → external tag {"Variant":{...}}
+// Wire conventions mirrored here:
+//   • field names are camelCase, so structs synthesize their `Codable`
+//   • optional fields are absent, never `null`
+//   • unions are internally tagged: `type` (kebab-case values); `Source` uses
+//     `sourceType`, `Role`-keyed messages use `role`
+//   • `ToolChoice` is mixed: "auto"|"none"|"required" or {"type":"tool","toolName":…}
+//   • `ModelPrompt` / `MessageContent` / `FileBytes` are untagged (string | array)
 
 import CAimuxFFI
 import Foundation
@@ -75,11 +74,16 @@ public enum JSONValue: Codable, Equatable, Sendable {
     public var objectValue: [String: JSONValue]? { if case .object(let o) = self { return o } else { return nil } }
 }
 
-// MARK: - AnyCodingKey (a CodingKey that can hold any string tag/field)
+/// Provider-keyed options: `{ "<provider>": { "<option>": <json> } }`.
+public typealias ProviderOptions = [String: [String: JSONValue]]
+/// Provider-keyed metadata: `{ "<provider>": { "<field>": <json> } }`.
+public typealias ProviderMetadata = [String: [String: JSONValue]]
 
-/// A `CodingKey` accepting any string, used to decode/encode externally- and
-/// internally-tagged enums whose tag/field names are not known at compile time.
-fileprivate struct AnyCodingKey: CodingKey {
+// MARK: - Tagged-union plumbing
+
+/// A `CodingKey` accepting any string, used for the `type` tag of internally
+/// tagged unions (and by `Multimodal.swift`).
+struct AnyCodingKey: CodingKey {
     let stringValue: String
     init?(stringValue: String) { self.stringValue = stringValue }
     var intValue: Int? { nil }
@@ -92,10 +96,29 @@ fileprivate struct AnyCodingKey: CodingKey {
 /// corelibs-foundation lacks the `dataCorruptedError(in: KeyedDecodingContainer)`
 /// and `dataCorrupted(codingPath:debugDescription:)` helpers, so construct the
 /// `Context` directly.
-fileprivate func aimuxDecodingError(_ codingPath: [any CodingKey], _ message: String) -> DecodingError {
+func aimuxDecodingError(_ codingPath: [any CodingKey], _ message: String) -> DecodingError {
     DecodingError.dataCorrupted(.init(codingPath: codingPath, debugDescription: message, underlyingError: nil))
 }
 
+extension Decoder {
+    /// The tag of an internally tagged union (`type` unless told otherwise).
+    func tag(_ key: String = "type") throws -> String {
+        try container(keyedBy: AnyCodingKey.self).decode(String.self, forKey: AnyCodingKey(key))
+    }
+
+    func unknownTag(_ tag: String, in union: String) -> DecodingError {
+        aimuxDecodingError(codingPath, "unknown \(union) tag \(tag)")
+    }
+}
+
+extension Encoder {
+    /// Add the tag to the keyed container a flattened payload already wrote.
+    func putTag(_ tag: String, key: String = "type") throws {
+        var c = container(keyedBy: AnyCodingKey.self)
+        try c.encode(tag, forKey: AnyCodingKey(key))
+    }
+}
+
 // MARK: - String-backed enums
 
 /// Who sent a message. Wire: lowercase ("system"|"user"|"assistant"|"tool").
@@ -131,11 +154,6 @@ public struct FinishReason: Codable, Equatable {
     public var unified: FinishReasonUnified
     /// Raw provider-specific reason (`nil` when the provider had none).
     public var raw: String?
-
-    public init(unified: FinishReasonUnified, raw: String? = nil) {
-        self.unified = unified
-        self.raw = raw
-    }
 }
 
 public struct InputTokenUsage: Codable, Equatable {
@@ -143,22 +161,12 @@ public struct InputTokenUsage: Codable, Equatable {
     public var noCache: UInt32?
     public var cacheRead: UInt32?
     public var cacheWrite: UInt32?
-    enum CodingKeys: String, CodingKey {
-        case total, noCache = "no_cache", cacheRead = "cache_read", cacheWrite = "cache_write"
-    }
-    public init(total: UInt32? = nil, noCache: UInt32? = nil, cacheRead: UInt32? = nil,
-                cacheWrite: UInt32? = nil) {
-        self.total = total; self.noCache = noCache; self.cacheRead = cacheRead; self.cacheWrite = cacheWrite
-    }
 }
 
 public struct OutputTokenUsage: Codable, Equatable {
     public var total: UInt32?
     public var text: UInt32?
     public var reasoning: UInt32?
-    public init(total: UInt32? = nil, text: UInt32? = nil, reasoning: UInt32? = nil) {
-        self.total = total; self.text = text; self.reasoning = reasoning
-    }
 }
 
 /// Token usage statistics.
@@ -167,82 +175,65 @@ public struct Usage: Codable, Equatable {
     public var outputTokens: OutputTokenUsage
     /// Opaque, provider-specific raw usage.
     public var raw: [String: JSONValue]?
-
-    enum CodingKeys: String, CodingKey {
-        case inputTokens = "input_tokens"
-        case outputTokens = "output_tokens"
-        case raw
-    }
-
-    public init(inputTokens: InputTokenUsage, outputTokens: OutputTokenUsage, raw: [String: JSONValue]? = nil) {
-        self.inputTokens = inputTokens
-        self.outputTokens = outputTokens
-        self.raw = raw
-    }
 }
 
-/// Metadata about the API response.
+/// Metadata about the API response (the payload of a `response-metadata` part).
 public struct ResponseMetadata: Codable, Equatable {
     public var id: String?
     public var timestamp: String?
     public var modelId: String?
+}
 
-    enum CodingKeys: String, CodingKey {
-        case id, timestamp
-        case modelId = "model_id"
-    }
+/// Response information for telemetry and debugging.
+public struct ResponseInfo: Codable, Equatable {
+    public var id: String?
+    public var timestamp: String?
+    public var modelId: String?
+    public var headers: [String: String]?
+    public var body: JSONValue?
+}
 
-    public init(id: String? = nil, timestamp: String? = nil, modelId: String? = nil) {
-        self.id = id; self.timestamp = timestamp; self.modelId = modelId
-    }
+public struct RequestInfo: Codable, Equatable {
+    public var body: JSONValue?
 }
 
-/// A provider warning. Wire: externally tagged.
+/// A provider warning. Wire: internally tagged by `type`.
 public enum Warning: Codable, Equatable {
     case unsupported(feature: String, details: String?)
     case compatibility(feature: String, details: String?)
     case deprecated(setting: String, message: String)
     case other(message: String)
 
-    private enum Tag { static let unsupported = "Unsupported"; static let compatibility = "Compatibility"
-        static let deprecated = "Deprecated"; static let other = "Other" }
-    private enum Field: String, CodingKey {
-        case feature, details, setting, message
-    }
+    private enum Field: String, CodingKey { case type, feature, details, setting, message }
 
     public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: AnyCodingKey.self)
-        guard let key = c.allKeys.first?.stringValue else {
-            throw aimuxDecodingError(c.codingPath, "missing warning tag")
-        }
-        let n = try c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey(key))
-        switch key {
-        case Tag.unsupported: self = .unsupported(feature: try n.decode(String.self, forKey: .feature),
-                                                  details: try n.decodeIfPresent(String.self, forKey: .details))
-        case Tag.compatibility: self = .compatibility(feature: try n.decode(String.self, forKey: .feature),
-                                                      details: try n.decodeIfPresent(String.self, forKey: .details))
-        case Tag.deprecated: self = .deprecated(setting: try n.decode(String.self, forKey: .setting),
-                                                message: try n.decode(String.self, forKey: .message))
-        case Tag.other: self = .other(message: try n.decode(String.self, forKey: .message))
-        default: throw aimuxDecodingError(c.codingPath, "unknown warning tag \(key)")
+        let c = try decoder.container(keyedBy: Field.self)
+        switch try c.decode(String.self, forKey: .type) {
+        case "unsupported": self = .unsupported(feature: try c.decode(String.self, forKey: .feature),
+                                                details: try c.decodeIfPresent(String.self, forKey: .details))
+        case "compatibility": self = .compatibility(feature: try c.decode(String.self, forKey: .feature),
+                                                    details: try c.decodeIfPresent(String.self, forKey: .details))
+        case "deprecated": self = .deprecated(setting: try c.decode(String.self, forKey: .setting),
+                                              message: try c.decode(String.self, forKey: .message))
+        case "other": self = .other(message: try c.decode(String.self, forKey: .message))
+        case let t: throw decoder.unknownTag(t, in: "warning")
         }
     }
 
     public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: AnyCodingKey.self)
+        var c = encoder.container(keyedBy: Field.self)
         switch self {
         case .unsupported(let f, let d):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey(Tag.unsupported))
-            try n.encode(f, forKey: .feature); try n.encodeIfPresent(d, forKey: .details)
+            try c.encode("unsupported", forKey: .type); try c.encode(f, forKey: .feature)
+            try c.encodeIfPresent(d, forKey: .details)
         case .compatibility(let f, let d):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey(Tag.compatibility))
-            try n.encode(f, forKey: .feature); try n.encodeIfPresent(d, forKey: .details)
+            try c.encode("compatibility", forKey: .type); try c.encode(f, forKey: .feature)
+            try c.encodeIfPresent(d, forKey: .details)
         case .deprecated(let s, let m):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey(Tag.deprecated))
-            try n.encode(s, forKey: .setting); try n.encode(m, forKey: .message)
+            try c.encode("deprecated", forKey: .type); try c.encode(s, forKey: .setting)
+            try c.encode(m, forKey: .message)
         case .other(let m):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey(Tag.other))
-            try n.encode(m, forKey: .message)
+            try c.encode("other", forKey: .type); try c.encode(m, forKey: .message)
         }
     }
 }
@@ -262,19 +253,11 @@ public struct FunctionTool: Codable, Equatable {
     /// JSON Schema describing the tool's parameters.
     public var inputSchema: JSONValue
     public var strict: Bool?
-    public var providerOptions: JSONValue?
+    public var providerOptions: ProviderOptions?
     public var inputExamples: [FunctionToolInputExample]?
 
-    enum CodingKeys: String, CodingKey {
-        case name, description
-        case inputSchema = "input_schema"
-        case strict
-        case providerOptions = "provider_options"
-        case inputExamples = "input_examples"
-    }
-
     public init(name: String, inputSchema: JSONValue, description: String? = nil,
-                strict: Bool? = nil, providerOptions: JSONValue? = nil,
+                strict: Bool? = nil, providerOptions: ProviderOptions? = nil,
                 inputExamples: [FunctionToolInputExample]? = nil) {
         self.name = name; self.inputSchema = inputSchema; self.description = description
         self.strict = strict; self.providerOptions = providerOptions; self.inputExamples = inputExamples
@@ -294,58 +277,30 @@ public struct ProviderTool: Codable, Equatable {
 
 /// A tool: a function tool or a provider tool.
 ///
-/// Wire: internally tagged by `type`, variant names snake_case
-/// (`{"type":"function", ...}`, `{"type":"provider", ...}`).
+/// Wire: internally tagged by `type` (`{"type":"function", …}`, `{"type":"provider", …}`).
 public enum Tool: Codable, Equatable {
     case function(FunctionTool)
     case provider(ProviderTool)
 
     public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: AnyCodingKey.self)
-        let type = try c.decode(String.self, forKey: AnyCodingKey("type"))
-        switch type {
-        case "function":
-            self = .function(FunctionTool(
-                name: try c.decode(String.self, forKey: AnyCodingKey("name")),
-                inputSchema: try c.decode(JSONValue.self, forKey: AnyCodingKey("input_schema")),
-                description: try c.decodeIfPresent(String.self, forKey: AnyCodingKey("description")),
-                strict: try c.decodeIfPresent(Bool.self, forKey: AnyCodingKey("strict")),
-                providerOptions: try c.decodeIfPresent(JSONValue.self, forKey: AnyCodingKey("provider_options")),
-                inputExamples: try c.decodeIfPresent([FunctionToolInputExample].self, forKey: AnyCodingKey("input_examples"))))
-        case "provider":
-            self = .provider(ProviderTool(
-                id: try c.decode(String.self, forKey: AnyCodingKey("id")),
-                name: try c.decode(String.self, forKey: AnyCodingKey("name")),
-                args: try c.decode([String: JSONValue].self, forKey: AnyCodingKey("args"))))
-        default:
-            throw aimuxDecodingError(c.codingPath, "unknown tool type \(type)")
+        switch try decoder.tag() {
+        case "function": self = .function(try FunctionTool(from: decoder))
+        case "provider": self = .provider(try ProviderTool(from: decoder))
+        case let t: throw decoder.unknownTag(t, in: "tool")
         }
     }
 
     public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: AnyCodingKey.self)
         switch self {
-        case .function(let ft):
-            try c.encode("function", forKey: AnyCodingKey("type"))
-            try c.encode(ft.name, forKey: AnyCodingKey("name"))
-            try c.encode(ft.inputSchema, forKey: AnyCodingKey("input_schema"))
-            try c.encodeIfPresent(ft.description, forKey: AnyCodingKey("description"))
-            try c.encodeIfPresent(ft.strict, forKey: AnyCodingKey("strict"))
-            try c.encodeIfPresent(ft.providerOptions, forKey: AnyCodingKey("provider_options"))
-            try c.encodeIfPresent(ft.inputExamples, forKey: AnyCodingKey("input_examples"))
-        case .provider(let pt):
-            try c.encode("provider", forKey: AnyCodingKey("type"))
-            try c.encode(pt.id, forKey: AnyCodingKey("id"))
-            try c.encode(pt.name, forKey: AnyCodingKey("name"))
-            try c.encode(pt.args, forKey: AnyCodingKey("args"))
+        case .function(let v): try v.encode(to: encoder); try encoder.putTag("function")
+        case .provider(let v): try v.encode(to: encoder); try encoder.putTag("provider")
         }
     }
 }
 
 /// How the model should choose tools.
 ///
-/// Wire: `"auto" | "none" | "required" | {"type":"tool","toolName":"..."}`
-/// (the `toolName` field is camelCase on the wire).
+/// Wire: `"auto" | "none" | "required" | {"type":"tool","toolName":"..."}`.
 public enum ToolChoice: Codable, Equatable {
     case auto
     case none
@@ -385,214 +340,414 @@ public enum ToolChoice: Codable, Equatable {
 
 /// How the model should format its response.
 ///
-/// Wire: external tag with a unit variant — `"Text"` or `{"Json":{...}}`.
+/// Wire: `{"type":"text"}` or `{"type":"json","schema"?,"name"?,"description"?}`.
 public enum ResponseFormat: Codable, Equatable {
     case text
-    case json(schema: JSONValue?, name: String?, description: String?)
+    case json(schema: JSONValue? = nil, name: String? = nil, description: String? = nil)
+
+    private enum Field: String, CodingKey { case type, schema, name, description }
+
+    public init(from decoder: Decoder) throws {
+        let c = try decoder.container(keyedBy: Field.self)
+        switch try c.decode(String.self, forKey: .type) {
+        case "text": self = .text
+        case "json":
+            self = .json(schema: try c.decodeIfPresent(JSONValue.self, forKey: .schema),
+                         name: try c.decodeIfPresent(String.self, forKey: .name),
+                         description: try c.decodeIfPresent(String.self, forKey: .description))
+        case let t: throw decoder.unknownTag(t, in: "response format")
+        }
+    }
+
+    public func encode(to encoder: Encoder) throws {
+        var c = encoder.container(keyedBy: Field.self)
+        switch self {
+        case .text: try c.encode("text", forKey: .type)
+        case .json(let schema, let name, let description):
+            try c.encode("json", forKey: .type)
+            try c.encodeIfPresent(schema, forKey: .schema)
+            try c.encodeIfPresent(name, forKey: .name)
+            try c.encodeIfPresent(description, forKey: .description)
+        }
+    }
+}
 
-    private enum Field: String, CodingKey { case schema, name, description }
+// MARK: - File data
+
+/// Either raw bytes or a base64-encoded string. Wire: untagged
+/// (`[…]` | `"…"`).
+public enum FileBytes: Codable, Equatable {
+    case binary([UInt8])
+    case base64(String)
 
     public init(from decoder: Decoder) throws {
         let c = try decoder.singleValueContainer()
-        if let s = try? c.decode(String.self) {
-            if s == "Text" { self = .text; return }
-            throw aimuxDecodingError(c.codingPath, "unknown response format \(s)")
+        if let bytes = try? c.decode([UInt8].self) { self = .binary(bytes); return }
+        if let s = try? c.decode(String.self) { self = .base64(s); return }
+        throw aimuxDecodingError(c.codingPath, "expected a byte array or a base64 string")
+    }
+
+    public func encode(to encoder: Encoder) throws {
+        var c = encoder.singleValueContainer()
+        switch self {
+        case .binary(let bytes): try c.encode(bytes)
+        case .base64(let s): try c.encode(s)
         }
-        let o = try decoder.container(keyedBy: AnyCodingKey.self)
-        guard let key = o.allKeys.first?.stringValue else {
-            throw aimuxDecodingError(o.codingPath, "missing response format tag")
+    }
+}
+
+/// File data as a tagged union. Wire: `{"type":"data","data":<FileBytes>}`,
+/// `{"type":"url","url":…,"originalUrl"?}`, `{"type":"reference","reference":{…}}`,
+/// `{"type":"text","text":…}`.
+public enum FileData: Codable, Equatable {
+    case data(data: FileBytes)
+    case url(url: String, originalUrl: String? = nil)
+    case reference(reference: [String: String])
+    case text(text: String)
+
+    private enum Field: String, CodingKey { case type, data, url, originalUrl, reference, text }
+
+    public init(from decoder: Decoder) throws {
+        let c = try decoder.container(keyedBy: Field.self)
+        switch try c.decode(String.self, forKey: .type) {
+        case "data": self = .data(data: try c.decode(FileBytes.self, forKey: .data))
+        case "url": self = .url(url: try c.decode(String.self, forKey: .url),
+                                originalUrl: try c.decodeIfPresent(String.self, forKey: .originalUrl))
+        case "reference": self = .reference(reference: try c.decode([String: String].self, forKey: .reference))
+        case "text": self = .text(text: try c.decode(String.self, forKey: .text))
+        case let t: throw decoder.unknownTag(t, in: "file data")
         }
-        guard key == "Json" else {
-            throw aimuxDecodingError(o.codingPath, "unknown response format tag \(key)")
+    }
+
+    public func encode(to encoder: Encoder) throws {
+        var c = encoder.container(keyedBy: Field.self)
+        switch self {
+        case .data(let d): try c.encode("data", forKey: .type); try c.encode(d, forKey: .data)
+        case .url(let u, let o):
+            try c.encode("url", forKey: .type); try c.encode(u, forKey: .url)
+            try c.encodeIfPresent(o, forKey: .originalUrl)
+        case .reference(let r): try c.encode("reference", forKey: .type); try c.encode(r, forKey: .reference)
+        case .text(let t): try c.encode("text", forKey: .type); try c.encode(t, forKey: .text)
+        }
+    }
+}
+
+/// Data or a URL returned for a generated file. Wire: `{"type":"data","data":…}`
+/// or `{"type":"url","url":…,"originalUrl"?}`.
+public enum GeneratedFileData: Codable, Equatable {
+    case data(data: FileBytes)
+    case url(url: String, originalUrl: String? = nil)
+
+    private enum Field: String, CodingKey { case type, data, url, originalUrl }
+
+    public init(from decoder: Decoder) throws {
+        let c = try decoder.container(keyedBy: Field.self)
+        switch try c.decode(String.self, forKey: .type) {
+        case "data": self = .data(data: try c.decode(FileBytes.self, forKey: .data))
+        case "url": self = .url(url: try c.decode(String.self, forKey: .url),
+                                originalUrl: try c.decodeIfPresent(String.self, forKey: .originalUrl))
+        case let t: throw decoder.unknownTag(t, in: "generated file data")
         }
-        let n = try o.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey(key))
-        self = .json(schema: try n.decodeIfPresent(JSONValue.self, forKey: .schema),
-                     name: try n.decodeIfPresent(String.self, forKey: .name),
-                     description: try n.decodeIfPresent(String.self, forKey: .description))
     }
 
     public func encode(to encoder: Encoder) throws {
+        var c = encoder.container(keyedBy: Field.self)
         switch self {
-        case .text:
-            var c = encoder.singleValueContainer(); try c.encode("Text")
-        case .json(let schema, let name, let description):
-            var c = encoder.container(keyedBy: AnyCodingKey.self)
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Json"))
-            try n.encodeIfPresent(schema, forKey: .schema)
-            try n.encodeIfPresent(name, forKey: .name)
-            try n.encodeIfPresent(description, forKey: .description)
+        case .data(let d): try c.encode("data", forKey: .type); try c.encode(d, forKey: .data)
+        case .url(let u, let o):
+            try c.encode("url", forKey: .type); try c.encode(u, forKey: .url)
+            try c.encodeIfPresent(o, forKey: .originalUrl)
+        }
+    }
+}
+
+public struct GeneratedFile: Codable, Equatable {
+    public var data: GeneratedFileData
+    public var mediaType: String
+    public var providerMetadata: ProviderMetadata?
+
+    public init(data: GeneratedFileData, mediaType: String, providerMetadata: ProviderMetadata? = nil) {
+        self.data = data; self.mediaType = mediaType; self.providerMetadata = providerMetadata
+    }
+}
+
+/// A source / citation. Wire: `{"sourceType":"url","id","url","title"?,…}` or
+/// `{"sourceType":"document","id","mediaType","title","filename"?,…}`.
+public enum Source: Codable, Equatable {
+    case url(id: String, url: String, title: String?, providerMetadata: ProviderMetadata?)
+    case document(id: String, mediaType: String, title: String, filename: String?,
+                  providerMetadata: ProviderMetadata?)
+
+    private enum Field: String, CodingKey {
+        case id, sourceType, url, title, filename, mediaType, providerMetadata
+    }
+
+    public init(from decoder: Decoder) throws {
+        let c = try decoder.container(keyedBy: Field.self)
+        let id = try c.decode(String.self, forKey: .id)
+        let pm = try c.decodeIfPresent(ProviderMetadata.self, forKey: .providerMetadata)
+        switch try c.decode(String.self, forKey: .sourceType) {
+        case "url":
+            self = .url(id: id, url: try c.decode(String.self, forKey: .url),
+                        title: try c.decodeIfPresent(String.self, forKey: .title), providerMetadata: pm)
+        case "document":
+            self = .document(id: id, mediaType: try c.decode(String.self, forKey: .mediaType),
+                             title: try c.decode(String.self, forKey: .title),
+                             filename: try c.decodeIfPresent(String.self, forKey: .filename),
+                             providerMetadata: pm)
+        case let t: throw decoder.unknownTag(t, in: "source")
+        }
+    }
+
+    public func encode(to encoder: Encoder) throws {
+        var c = encoder.container(keyedBy: Field.self)
+        switch self {
+        case .url(let id, let url, let title, let pm):
+            try c.encode("url", forKey: .sourceType); try c.encode(id, forKey: .id)
+            try c.encode(url, forKey: .url); try c.encodeIfPresent(title, forKey: .title)
+            try c.encodeIfPresent(pm, forKey: .providerMetadata)
+        case .document(let id, let mediaType, let title, let filename, let pm):
+            try c.encode("document", forKey: .sourceType); try c.encode(id, forKey: .id)
+            try c.encode(mediaType, forKey: .mediaType); try c.encode(title, forKey: .title)
+            try c.encodeIfPresent(filename, forKey: .filename)
+            try c.encodeIfPresent(pm, forKey: .providerMetadata)
         }
     }
 }
 
 // MARK: - Messages
 
+/// Text content part of a tool-result `content` output.
+public struct TextPart: Codable, Equatable {
+    public var text: String
+    public var providerOptions: ProviderOptions?
+
+    public init(text: String, providerOptions: ProviderOptions? = nil) {
+        self.text = text; self.providerOptions = providerOptions
+    }
+}
+
+/// File content part of a tool-result `content` output.
+public struct FilePart: Codable, Equatable {
+    public var data: FileData
+    public var mediaType: String
+    public var filename: String?
+    public var providerOptions: ProviderOptions?
+
+    public init(data: FileData, mediaType: String, filename: String? = nil,
+                providerOptions: ProviderOptions? = nil) {
+        self.data = data; self.mediaType = mediaType; self.filename = filename
+        self.providerOptions = providerOptions
+    }
+}
+
+/// One item of a `content` tool-result output. Wire: tagged by `type`.
+public enum ToolResultContent: Codable, Equatable {
+    case text(TextPart)
+    case file(FilePart)
+    case custom(providerOptions: ProviderOptions? = nil)
+
+    private enum Field: String, CodingKey { case type, providerOptions }
+
+    public init(from decoder: Decoder) throws {
+        switch try decoder.tag() {
+        case "text": self = .text(try TextPart(from: decoder))
+        case "file": self = .file(try FilePart(from: decoder))
+        case "custom":
+            self = .custom(providerOptions: try decoder.container(keyedBy: Field.self)
+                .decodeIfPresent(ProviderOptions.self, forKey: .providerOptions))
+        case let t: throw decoder.unknownTag(t, in: "tool result content")
+        }
+    }
+
+    public func encode(to encoder: Encoder) throws {
+        switch self {
+        case .text(let v): try v.encode(to: encoder); try encoder.putTag("text")
+        case .file(let v): try v.encode(to: encoder); try encoder.putTag("file")
+        case .custom(let po):
+            var c = encoder.container(keyedBy: Field.self)
+            try c.encode("custom", forKey: .type); try c.encodeIfPresent(po, forKey: .providerOptions)
+        }
+    }
+}
+
+/// The result of a tool call, as sent back to the model.
+///
+/// Wire: `{"type":"text","value"}`, `{"type":"json","value"}`,
+/// `{"type":"execution-denied","reason"?}`, `{"type":"error-text","value"}`,
+/// `{"type":"error-json","value"}`, `{"type":"content","value":[…]}`; all but
+/// `content` may carry `providerOptions`.
+public enum ToolResultOutput: Codable, Equatable {
+    case text(String, providerOptions: ProviderOptions? = nil)
+    case json(JSONValue, providerOptions: ProviderOptions? = nil)
+    case executionDenied(reason: String? = nil, providerOptions: ProviderOptions? = nil)
+    case errorText(String, providerOptions: ProviderOptions? = nil)
+    case errorJson(JSONValue, providerOptions: ProviderOptions? = nil)
+    case content([ToolResultContent])
+
+    private enum Field: String, CodingKey { case type, value, reason, providerOptions }
+
+    public init(from decoder: Decoder) throws {
+        let c = try decoder.container(keyedBy: Field.self)
+        let po = try c.decodeIfPresent(ProviderOptions.self, forKey: .providerOptions)
+        switch try c.decode(String.self, forKey: .type) {
+        case "text": self = .text(try c.decode(String.self, forKey: .value), providerOptions: po)
+        case "json": self = .json(try c.decode(JSONValue.self, forKey: .value), providerOptions: po)
+        case "execution-denied":
+            self = .executionDenied(reason: try c.decodeIfPresent(String.self, forKey: .reason), providerOptions: po)
+        case "error-text": self = .errorText(try c.decode(String.self, forKey: .value), providerOptions: po)
+        case "error-json": self = .errorJson(try c.decode(JSONValue.self, forKey: .value), providerOptions: po)
+        case "content": self = .content(try c.decode([ToolResultContent].self, forKey: .value))
+        case let t: throw decoder.unknownTag(t, in: "tool result output")
+        }
+    }
+
+    public func encode(to encoder: Encoder) throws {
+        var c = encoder.container(keyedBy: Field.self)
+        switch self {
+        case .text(let v, let po):
+            try c.encode("text", forKey: .type); try c.encode(v, forKey: .value)
+            try c.encodeIfPresent(po, forKey: .providerOptions)
+        case .json(let v, let po):
+            try c.encode("json", forKey: .type); try c.encode(v, forKey: .value)
+            try c.encodeIfPresent(po, forKey: .providerOptions)
+        case .executionDenied(let r, let po):
+            try c.encode("execution-denied", forKey: .type); try c.encodeIfPresent(r, forKey: .reason)
+            try c.encodeIfPresent(po, forKey: .providerOptions)
+        case .errorText(let v, let po):
+            try c.encode("error-text", forKey: .type); try c.encode(v, forKey: .value)
+            try c.encodeIfPresent(po, forKey: .providerOptions)
+        case .errorJson(let v, let po):
+            try c.encode("error-json", forKey: .type); try c.encode(v, forKey: .value)
+            try c.encodeIfPresent(po, forKey: .providerOptions)
+        case .content(let v):
+            try c.encode("content", forKey: .type); try c.encode(v, forKey: .value)
+        }
+    }
+}
+
 /// A part of a multi-part message.
 ///
-/// Wire: internally tagged by `type` (variant names snake_case):
-/// `{"type":"text","text":"..."}`, `{"type":"tool_call","tool_call_id":...}`, …
+/// Wire: internally tagged by `type` (kebab-case):
+/// `{"type":"text","text":"..."}`, `{"type":"tool-call","toolCallId":…}`,
+/// `{"type":"tool-result","toolCallId","toolName","output":<ToolResultOutput>}`, …
 public enum ContentPart: Codable, Equatable {
-    case text(text: String, providerOptions: JSONValue?)
-    case image(image: [UInt8], mediaType: String, providerOptions: JSONValue?)
-    case file(data: [UInt8], mediaType: String, filename: String?, providerOptions: JSONValue?)
-    case fileBase64(data: String, mediaType: String, filename: String?, providerOptions: JSONValue?)
-    case fileUrl(url: String, mediaType: String, providerOptions: JSONValue?)
-    case fileReference(mediaType: String, reference: JSONValue, filename: String?, providerOptions: JSONValue?)
-    case reasoning(text: String, signature: String?, providerOptions: JSONValue?)
+    case text(text: String, providerOptions: ProviderOptions? = nil)
+    case image(image: [UInt8], mediaType: String, providerOptions: ProviderOptions? = nil)
+    case file(data: [UInt8], mediaType: String, filename: String? = nil, providerOptions: ProviderOptions? = nil)
+    case fileBase64(data: String, mediaType: String, filename: String? = nil, providerOptions: ProviderOptions? = nil)
+    case fileUrl(url: String, mediaType: String, providerOptions: ProviderOptions? = nil)
+    case fileReference(mediaType: String, reference: JSONValue, filename: String? = nil,
+                       providerOptions: ProviderOptions? = nil)
+    case reasoning(text: String, signature: String? = nil, providerOptions: ProviderOptions? = nil)
     case toolCall(toolCallId: String, toolName: String, input: JSONValue,
-                  providerExecuted: Bool? = nil,
-                  providerOptions: JSONValue?)
-    case toolResult(toolCallId: String, result: JSONValue, toolName: String?,
-                     isError: Bool?, preliminary: Bool?, dynamic: Bool?, providerOptions: JSONValue?)
-
-    case custom(kind: String, providerOptions: JSONValue?)
-    case reasoningFile(data: GeneratedFileData, mediaType: String, providerOptions: JSONValue?)
-    case toolApprovalRequest(approvalId: String, toolCallId: String, reason: String?, isAutomatic: Bool?, signature: String?, inputSchemaInput: JSONValue?)
+                  providerExecuted: Bool? = nil, providerOptions: ProviderOptions? = nil)
+    case toolResult(toolCallId: String, toolName: String, output: ToolResultOutput,
+                    providerOptions: ProviderOptions? = nil)
+    case custom(kind: String, providerOptions: ProviderOptions? = nil)
+    case reasoningFile(data: GeneratedFileData, mediaType: String, providerOptions: ProviderOptions? = nil)
+    case toolApprovalRequest(approvalId: String, toolCallId: String, reason: String? = nil,
+                             isAutomatic: Bool? = nil, signature: String? = nil,
+                             inputSchemaInput: JSONValue? = nil)
 
     private enum Field: String, CodingKey {
-        case kind, approvalId = "approval_id", reason, isAutomatic = "is_automatic", inputSchemaInput = "input_schema_input"
-        case text, image, data, mediaType = "media_type", filename, url, reference
-        case signature, toolCallId = "tool_call_id", toolName = "tool_name", input, result
-        case providerExecuted = "provider_executed"
-        case isError = "is_error", preliminary, dynamic
-        case providerOptions = "provider_options"
+        case type, text, image, data, mediaType, filename, url, reference, signature
+        case toolCallId, toolName, input, output, providerExecuted, providerOptions
+        case kind, approvalId, reason, isAutomatic, inputSchemaInput
     }
 
     public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: AnyCodingKey.self)
-        let type = try c.decode(String.self, forKey: AnyCodingKey("type"))
-        func po() throws -> JSONValue? { try c.decodeIfPresent(JSONValue.self, forKey: AnyCodingKey("provider_options")) }
-        switch type {
-        case "custom":
-            self = .custom(kind: try c.decode(String.self, forKey: AnyCodingKey("kind")),
-                           providerOptions: try c.decodeIfPresent(JSONValue.self, forKey: AnyCodingKey("provider_options")))
-        case "reasoning_file":
-            self = .reasoningFile(data: try c.decode(GeneratedFileData.self, forKey: AnyCodingKey("data")),
-                                  mediaType: try c.decode(String.self, forKey: AnyCodingKey("media_type")),
-                                  providerOptions: try c.decodeIfPresent(JSONValue.self, forKey: AnyCodingKey("provider_options")))
-        case "tool_approval_request":
-            self = .toolApprovalRequest(approvalId: try c.decode(String.self, forKey: AnyCodingKey("approval_id")),
-                                         toolCallId: try c.decode(String.self, forKey: AnyCodingKey("tool_call_id")),
-                                         reason: try c.decodeIfPresent(String.self, forKey: AnyCodingKey("reason")),
-                                         isAutomatic: try c.decodeIfPresent(Bool.self, forKey: AnyCodingKey("is_automatic")),
-                                         signature: try c.decodeIfPresent(String.self, forKey: AnyCodingKey("signature")),
-                                         inputSchemaInput: try c.decodeIfPresent(JSONValue.self, forKey: AnyCodingKey("input_schema_input")))
+        let c = try decoder.container(keyedBy: Field.self)
+        let po = try c.decodeIfPresent(ProviderOptions.self, forKey: .providerOptions)
+        switch try c.decode(String.self, forKey: .type) {
         case "text":
-            self = .text(text: try c.decode(String.self, forKey: AnyCodingKey("text")),
-                         providerOptions: try po())
+            self = .text(text: try c.decode(String.self, forKey: .text), providerOptions: po)
         case "image":
-            self = .image(image: try c.decode([UInt8].self, forKey: AnyCodingKey("image")),
-                          mediaType: try c.decode(String.self, forKey: AnyCodingKey("media_type")),
-                          providerOptions: try po())
+            self = .image(image: try c.decode([UInt8].self, forKey: .image),
+                          mediaType: try c.decode(String.self, forKey: .mediaType), providerOptions: po)
         case "file":
-            self = .file(data: try c.decode([UInt8].self, forKey: AnyCodingKey("data")),
-                         mediaType: try c.decode(String.self, forKey: AnyCodingKey("media_type")),
-                         filename: try c.decodeIfPresent(String.self, forKey: AnyCodingKey("filename")),
-                         providerOptions: try po())
-        case "file_base64":
-            self = .fileBase64(data: try c.decode(String.self, forKey: AnyCodingKey("data")),
-                               mediaType: try c.decode(String.self, forKey: AnyCodingKey("media_type")),
-                               filename: try c.decodeIfPresent(String.self, forKey: AnyCodingKey("filename")),
-                               providerOptions: try po())
-        case "file_url":
-            self = .fileUrl(url: try c.decode(String.self, forKey: AnyCodingKey("url")),
-                            mediaType: try c.decode(String.self, forKey: AnyCodingKey("media_type")),
-                            providerOptions: try po())
-        case "file_reference":
-            self = .fileReference(mediaType: try c.decode(String.self, forKey: AnyCodingKey("media_type")),
-                                  reference: try c.decode(JSONValue.self, forKey: AnyCodingKey("reference")),
-                                  filename: try c.decodeIfPresent(String.self, forKey: AnyCodingKey("filename")),
-                                  providerOptions: try po())
+            self = .file(data: try c.decode([UInt8].self, forKey: .data),
+                         mediaType: try c.decode(String.self, forKey: .mediaType),
+                         filename: try c.decodeIfPresent(String.self, forKey: .filename), providerOptions: po)
+        case "file-base64":
+            self = .fileBase64(data: try c.decode(String.self, forKey: .data),
+                               mediaType: try c.decode(String.self, forKey: .mediaType),
+                               filename: try c.decodeIfPresent(String.self, forKey: .filename), providerOptions: po)
+        case "file-url":
+            self = .fileUrl(url: try c.decode(String.self, forKey: .url),
+                            mediaType: try c.decode(String.self, forKey: .mediaType), providerOptions: po)
+        case "file-reference":
+            self = .fileReference(mediaType: try c.decode(String.self, forKey: .mediaType),
+                                  reference: try c.decode(JSONValue.self, forKey: .reference),
+                                  filename: try c.decodeIfPresent(String.self, forKey: .filename), providerOptions: po)
         case "reasoning":
-            self = .reasoning(text: try c.decode(String.self, forKey: AnyCodingKey("text")),
-                             signature: try c.decodeIfPresent(String.self, forKey: AnyCodingKey("signature")),
-                             providerOptions: try po())
-        case "tool_call":
-            self = .toolCall(toolCallId: try c.decode(String.self, forKey: AnyCodingKey("tool_call_id")),
-                             toolName: try c.decode(String.self, forKey: AnyCodingKey("tool_name")),
-                             input: try c.decode(JSONValue.self, forKey: AnyCodingKey("input")),
-                             providerExecuted: try c.decodeIfPresent(Bool.self, forKey: AnyCodingKey("provider_executed")),
-                             providerOptions: try po())
-        case "tool_result":
-            self = .toolResult(toolCallId: try c.decode(String.self, forKey: AnyCodingKey("tool_call_id")),
-                               result: try c.decode(JSONValue.self, forKey: AnyCodingKey("result")),
-                               toolName: try c.decodeIfPresent(String.self, forKey: AnyCodingKey("tool_name")),
-                               isError: try c.decodeIfPresent(Bool.self, forKey: AnyCodingKey("is_error")),
-                               preliminary: try c.decodeIfPresent(Bool.self, forKey: AnyCodingKey("preliminary")),
-                               dynamic: try c.decodeIfPresent(Bool.self, forKey: AnyCodingKey("dynamic")),
-                               providerOptions: try po())
-        default:
-            throw aimuxDecodingError(c.codingPath, "unknown content part type \(type)")
+            self = .reasoning(text: try c.decode(String.self, forKey: .text),
+                              signature: try c.decodeIfPresent(String.self, forKey: .signature), providerOptions: po)
+        case "tool-call":
+            self = .toolCall(toolCallId: try c.decode(String.self, forKey: .toolCallId),
+                             toolName: try c.decode(String.self, forKey: .toolName),
+                             input: try c.decode(JSONValue.self, forKey: .input),
+                             providerExecuted: try c.decodeIfPresent(Bool.self, forKey: .providerExecuted),
+                             providerOptions: po)
+        case "tool-result":
+            self = .toolResult(toolCallId: try c.decode(String.self, forKey: .toolCallId),
+                               toolName: try c.decode(String.self, forKey: .toolName),
+                               output: try c.decode(ToolResultOutput.self, forKey: .output), providerOptions: po)
+        case "custom":
+            self = .custom(kind: try c.decode(String.self, forKey: .kind), providerOptions: po)
+        case "reasoning-file":
+            self = .reasoningFile(data: try c.decode(GeneratedFileData.self, forKey: .data),
+                                  mediaType: try c.decode(String.self, forKey: .mediaType), providerOptions: po)
+        case "tool-approval-request":
+            self = .toolApprovalRequest(approvalId: try c.decode(String.self, forKey: .approvalId),
+                                        toolCallId: try c.decode(String.self, forKey: .toolCallId),
+                                        reason: try c.decodeIfPresent(String.self, forKey: .reason),
+                                        isAutomatic: try c.decodeIfPresent(Bool.self, forKey: .isAutomatic),
+                                        signature: try c.decodeIfPresent(String.self, forKey: .signature),
+                                        inputSchemaInput: try c.decodeIfPresent(JSONValue.self, forKey: .inputSchemaInput))
+        case let t: throw decoder.unknownTag(t, in: "content part")
         }
     }
 
     public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: AnyCodingKey.self)
+        var c = encoder.container(keyedBy: Field.self)
+        func tag(_ t: String, _ po: ProviderOptions?) throws {
+            try c.encode(t, forKey: .type); try c.encodeIfPresent(po, forKey: .providerOptions)
+        }
         switch self {
-        case .custom(let kind, let po):
-            try c.encode("custom", forKey: AnyCodingKey("type")); try c.encode(kind, forKey: AnyCodingKey("kind"))
-            try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
-        case .reasoningFile(let data, let mediaType, let po):
-            try c.encode("reasoning_file", forKey: AnyCodingKey("type")); try c.encode(data, forKey: AnyCodingKey("data"))
-            try c.encode(mediaType, forKey: AnyCodingKey("media_type")); try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
-        case .toolApprovalRequest(let approvalId, let toolCallId, let reason, let isAutomatic, let signature, let inputSchemaInput):
-            try c.encode("tool_approval_request", forKey: AnyCodingKey("type")); try c.encode(approvalId, forKey: AnyCodingKey("approval_id"))
-            try c.encode(toolCallId, forKey: AnyCodingKey("tool_call_id")); try c.encodeIfPresent(reason, forKey: AnyCodingKey("reason"))
-            try c.encodeIfPresent(isAutomatic, forKey: AnyCodingKey("is_automatic")); try c.encodeIfPresent(signature, forKey: AnyCodingKey("signature"))
-            try c.encodeIfPresent(inputSchemaInput, forKey: AnyCodingKey("input_schema_input"))
         case .text(let text, let po):
-            try c.encode("text", forKey: AnyCodingKey("type"))
-            try c.encode(text, forKey: AnyCodingKey("text"))
-            try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
+            try tag("text", po); try c.encode(text, forKey: .text)
         case .image(let image, let mediaType, let po):
-            try c.encode("image", forKey: AnyCodingKey("type"))
-            try c.encode(image, forKey: AnyCodingKey("image"))
-            try c.encode(mediaType, forKey: AnyCodingKey("media_type"))
-            try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
+            try tag("image", po); try c.encode(image, forKey: .image); try c.encode(mediaType, forKey: .mediaType)
         case .file(let data, let mediaType, let filename, let po):
-            try c.encode("file", forKey: AnyCodingKey("type"))
-            try c.encode(data, forKey: AnyCodingKey("data"))
-            try c.encode(mediaType, forKey: AnyCodingKey("media_type"))
-            try c.encodeIfPresent(filename, forKey: AnyCodingKey("filename"))
-            try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
+            try tag("file", po); try c.encode(data, forKey: .data); try c.encode(mediaType, forKey: .mediaType)
+            try c.encodeIfPresent(filename, forKey: .filename)
         case .fileBase64(let data, let mediaType, let filename, let po):
-            try c.encode("file_base64", forKey: AnyCodingKey("type"))
-            try c.encode(data, forKey: AnyCodingKey("data"))
-            try c.encode(mediaType, forKey: AnyCodingKey("media_type"))
-            try c.encodeIfPresent(filename, forKey: AnyCodingKey("filename"))
-            try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
+            try tag("file-base64", po); try c.encode(data, forKey: .data); try c.encode(mediaType, forKey: .mediaType)
+            try c.encodeIfPresent(filename, forKey: .filename)
         case .fileUrl(let url, let mediaType, let po):
-            try c.encode("file_url", forKey: AnyCodingKey("type"))
-            try c.encode(url, forKey: AnyCodingKey("url"))
-            try c.encode(mediaType, forKey: AnyCodingKey("media_type"))
-            try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
+            try tag("file-url", po); try c.encode(url, forKey: .url); try c.encode(mediaType, forKey: .mediaType)
         case .fileReference(let mediaType, let reference, let filename, let po):
-            try c.encode("file_reference", forKey: AnyCodingKey("type"))
-            try c.encode(mediaType, forKey: AnyCodingKey("media_type"))
-            try c.encode(reference, forKey: AnyCodingKey("reference"))
-            try c.encodeIfPresent(filename, forKey: AnyCodingKey("filename"))
-            try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
+            try tag("file-reference", po); try c.encode(mediaType, forKey: .mediaType)
+            try c.encode(reference, forKey: .reference); try c.encodeIfPresent(filename, forKey: .filename)
         case .reasoning(let text, let signature, let po):
-            try c.encode("reasoning", forKey: AnyCodingKey("type"))
-            try c.encode(text, forKey: AnyCodingKey("text"))
-            try c.encodeIfPresent(signature, forKey: AnyCodingKey("signature"))
-            try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
-        case .toolCall(let toolCallId, let toolName, let input, let providerExecuted, let po):
-            try c.encode("tool_call", forKey: AnyCodingKey("type"))
-            try c.encode(toolCallId, forKey: AnyCodingKey("tool_call_id"))
-            try c.encode(toolName, forKey: AnyCodingKey("tool_name"))
-            try c.encode(input, forKey: AnyCodingKey("input"))
-            try c.encodeIfPresent(providerExecuted, forKey: AnyCodingKey("provider_executed"))
-            try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
-        case .toolResult(let toolCallId, let result, let toolName, let isError, let preliminary, let dynamic, let po):
-            try c.encode("tool_result", forKey: AnyCodingKey("type"))
-            try c.encode(toolCallId, forKey: AnyCodingKey("tool_call_id"))
-            try c.encode(result, forKey: AnyCodingKey("result"))
-            try c.encodeIfPresent(toolName, forKey: AnyCodingKey("tool_name"))
-            try c.encodeIfPresent(isError, forKey: AnyCodingKey("is_error"))
-            try c.encodeIfPresent(preliminary, forKey: AnyCodingKey("preliminary"))
-            try c.encodeIfPresent(dynamic, forKey: AnyCodingKey("dynamic"))
-            try c.encodeIfPresent(po, forKey: AnyCodingKey("provider_options"))
+            try tag("reasoning", po); try c.encode(text, forKey: .text)
+            try c.encodeIfPresent(signature, forKey: .signature)
+        case .toolCall(let id, let name, let input, let pe, let po):
+            try tag("tool-call", po); try c.encode(id, forKey: .toolCallId); try c.encode(name, forKey: .toolName)
+            try c.encode(input, forKey: .input); try c.encodeIfPresent(pe, forKey: .providerExecuted)
+        case .toolResult(let id, let name, let output, let po):
+            try tag("tool-result", po); try c.encode(id, forKey: .toolCallId); try c.encode(name, forKey: .toolName)
+            try c.encode(output, forKey: .output)
+        case .custom(let kind, let po):
+            try tag("custom", po); try c.encode(kind, forKey: .kind)
+        case .reasoningFile(let data, let mediaType, let po):
+            try tag("reasoning-file", po); try c.encode(data, forKey: .data); try c.encode(mediaType, forKey: .mediaType)
+        case .toolApprovalRequest(let approvalId, let toolCallId, let reason, let isAutomatic, let signature, let schemaInput):
+            try tag("tool-approval-request", nil); try c.encode(approvalId, forKey: .approvalId)
+            try c.encode(toolCallId, forKey: .toolCallId); try c.encodeIfPresent(reason, forKey: .reason)
+            try c.encodeIfPresent(isAutomatic, forKey: .isAutomatic); try c.encodeIfPresent(signature, forKey: .signature)
+            try c.encodeIfPresent(schemaInput, forKey: .inputSchemaInput)
         }
     }
 }
@@ -661,33 +816,23 @@ public enum ModelPrompt: Codable, Equatable {
 public struct ToolCall: Codable, Equatable {
     public var toolCallId: String
     public var toolName: String
-    /// Arguments as a JSON value (usually an object).
+    /// Parsed arguments (usually an object); the original string when `invalid`
+    /// is true and the provider input was not valid JSON.
     public var input: JSONValue
     public var providerExecuted: Bool?
     public var dynamic: Bool?
     /// Additional provider-specific metadata associated with this call.
-    public var providerMetadata: JSONValue?
-    /// Set by Core when the tool call stays invalid after optional repair.
+    public var providerMetadata: ProviderMetadata?
+    /// Set when the tool call stays invalid after the optional repair.
     public var invalid: Bool?
-    /// The typed lookup, parse, schema, or repair failure for an invalid call.
+    /// The typed lookup, parse, schema, or repair failure for an invalid call:
+    /// an `AiMuxError` object tagged by `name` (e.g. `"AI_InvalidToolInputError"`).
     public var error: JSONValue?
 
-    enum CodingKeys: String, CodingKey {
-        case toolCallId = "tool_call_id"
-        case toolName = "tool_name"
-        case input
-        case providerExecuted = "provider_executed"
-        case dynamic
-        case providerMetadata = "provider_metadata"
-        case invalid
-        case error
-    }
-
     public init(toolCallId: String, toolName: String, input: JSONValue,
                 providerExecuted: Bool? = nil, dynamic: Bool? = nil,
-                providerMetadata: JSONValue? = nil,
-                invalid: Bool? = nil,
-                error: JSONValue? = nil) {
+                providerMetadata: ProviderMetadata? = nil,
+                invalid: Bool? = nil, error: JSONValue? = nil) {
         self.toolCallId = toolCallId; self.toolName = toolName; self.input = input
         self.providerExecuted = providerExecuted; self.dynamic = dynamic
         self.providerMetadata = providerMetadata
@@ -695,559 +840,213 @@ public struct ToolCall: Codable, Equatable {
     }
 }
 
-// MARK: - File data (payload of `GenerateContent.file` / `StreamPart.file`)
-
-/// Either raw bytes or a base64-encoded string (mirrors `aimux_core::FileBytes`).
-///
-/// Wire: externally tagged newtype — `{"Binary":[…]}`, `{"Base64":"…"}`.
-public enum FileBytes: Codable, Equatable {
-    case binary([UInt8])
-    case base64(String)
-
-    public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: AnyCodingKey.self)
-        guard let key = c.allKeys.first?.stringValue else {
-            throw aimuxDecodingError(c.codingPath, "missing file-bytes tag")
-        }
-        switch key {
-        case "Binary":
-            self = .binary(try c.decode([UInt8].self, forKey: AnyCodingKey(key)))
-        case "Base64":
-            self = .base64(try c.decode(String.self, forKey: AnyCodingKey(key)))
-        default:
-            throw aimuxDecodingError(c.codingPath, "unknown file-bytes tag \(key)")
-        }
-    }
+/// A tool result produced by a provider-executed tool, next to the call it answers.
+public struct ToolResult: Codable, Equatable {
+    public var toolCallId: String
+    public var toolName: String
+    public var result: JSONValue
+    public var isError: Bool?
+    public var preliminary: Bool?
+    public var dynamic: Bool?
+    public var providerMetadata: ProviderMetadata?
+}
 
-    public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: AnyCodingKey.self)
-        switch self {
-        case .binary(let bytes):
-            try c.encode(bytes, forKey: AnyCodingKey("Binary"))
-        case .base64(let s):
-            try c.encode(s, forKey: AnyCodingKey("Base64"))
-        }
-    }
+/// A reasoning / thinking segment (provider metadata carries e.g. the
+/// Anthropic signature at `anthropic.signature`).
+public struct ReasoningOutput: Codable, Equatable {
+    public var text: String
+    public var providerMetadata: ProviderMetadata?
 }
 
-/// File data as a tagged discriminated union (mirrors `aimux_core::FileData`).
-///
-/// Wire: externally tagged — `{"Data":{"data":<FileBytes>}}`,
-/// `{"Url":{"url":"…"}}`, `{"Reference":{"reference":{…}}}`,
-/// `{"Text":{"text":"…"}}`.
-public enum FileData: Codable, Equatable {
-    case data(data: FileBytes)
-    case url(url: String, originalUrl: String? = nil)
-    case reference(reference: [String: String])
-    case text(text: String)
+/// A file generated as part of reasoning (call-layer shape).
+public struct ReasoningFileOutput: Codable, Equatable {
+    public var file: GeneratedFile
+    public var providerMetadata: ProviderMetadata?
+}
 
-    private enum Field: String, CodingKey {
-        case data, url, reference, text, originalUrl = "original_url"
-    }
+/// Reasoning text or a generated file produced during reasoning.
+public enum ReasoningPart: Codable, Equatable {
+    case reasoning(ReasoningOutput)
+    case reasoningFile(ReasoningFileOutput)
 
     public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: AnyCodingKey.self)
-        guard let key = c.allKeys.first?.stringValue else {
-            throw aimuxDecodingError(c.codingPath, "missing file-data tag")
-        }
-        let n = try c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey(key))
-        switch key {
-        case "Data":
-            self = .data(data: try n.decode(FileBytes.self, forKey: .data))
-        case "Url":
-            self = .url(url: try n.decode(String.self, forKey: .url),
-                        originalUrl: try n.decodeIfPresent(String.self, forKey: .originalUrl))
-        case "Reference":
-            self = .reference(reference: try n.decode([String: String].self, forKey: .reference))
-        case "Text":
-            self = .text(text: try n.decode(String.self, forKey: .text))
-        default:
-            throw aimuxDecodingError(c.codingPath, "unknown file-data tag \(key)")
+        switch try decoder.tag() {
+        case "reasoning": self = .reasoning(try ReasoningOutput(from: decoder))
+        case "reasoning-file": self = .reasoningFile(try ReasoningFileOutput(from: decoder))
+        case let t: throw decoder.unknownTag(t, in: "reasoning part")
         }
     }
 
     public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: AnyCodingKey.self)
-        switch self {
-        case .data(let data):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Data"))
-            try n.encode(data, forKey: .data)
-        case .url(let url, let originalUrl):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Url"))
-            try n.encode(url, forKey: .url)
-            try n.encodeIfPresent(originalUrl, forKey: .originalUrl)
-        case .reference(let reference):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Reference"))
-            try n.encode(reference, forKey: .reference)
-        case .text(let text):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Text"))
-            try n.encode(text, forKey: .text)
-        }
-    }
-}
-
-/// Data or a URL returned for a generated file.
-public enum GeneratedFileData: Codable, Equatable {
-    case data(data: FileBytes)
-    case url(url: String, originalUrl: String? = nil)
-    private enum Field: String, CodingKey {
-        case data, url, originalUrl = "original_url"
-    }
-    public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: AnyCodingKey.self)
-        guard let key = c.allKeys.first?.stringValue else {
-            throw aimuxDecodingError(c.codingPath, "missing generated-file-data tag")
-        }
-        let n = try c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey(key))
-        switch key {
-        case "Data": self = .data(data: try n.decode(FileBytes.self, forKey: .data))
-        case "Url": self = .url(url: try n.decode(String.self, forKey: .url),
-                                originalUrl: try n.decodeIfPresent(String.self, forKey: .originalUrl))
-        default: throw aimuxDecodingError(c.codingPath, "invalid generated-file-data tag")
-        }
-    }
-    public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: AnyCodingKey.self)
         switch self {
-        case .data(let data):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Data"))
-            try n.encode(data, forKey: .data)
-        case .url(let url, let originalUrl):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Url"))
-            try n.encode(url, forKey: .url); try n.encodeIfPresent(originalUrl, forKey: .originalUrl)
+        case .reasoning(let v): try v.encode(to: encoder); try encoder.putTag("reasoning")
+        case .reasoningFile(let v): try v.encode(to: encoder); try encoder.putTag("reasoning-file")
         }
     }
 }
 
-public struct GeneratedFile: Codable, Equatable {
-    public var data: GeneratedFileData
-    public var mediaType: String
-    public var providerMetadata: JSONValue?
-    enum CodingKeys: String, CodingKey {
-        case data, mediaType = "media_type", providerMetadata = "provider_metadata"
-    }
-    public init(data: GeneratedFileData, mediaType: String, providerMetadata: JSONValue? = nil) {
-        self.data = data; self.mediaType = mediaType; self.providerMetadata = providerMetadata
-    }
+/// Provider approval request for a provider-executed call.
+public struct RawToolApprovalRequest: Codable, Equatable {
+    public var approvalId: String
+    public var toolCallId: String
+    public var providerMetadata: ProviderMetadata?
 }
 
-public enum Source: Codable, Equatable {
-    case url(id: String, url: String, title: String?, providerMetadata: JSONValue?)
-    case document(id: String, mediaType: String, title: String, filename: String?,
-                  providerMetadata: JSONValue?)
-
-    private enum Field: String, CodingKey {
-        case id, sourceType = "source_type", url, title, filename
-        case mediaType = "media_type", providerMetadata = "provider_metadata"
-    }
-
-    public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: Field.self)
-        let id = try c.decode(String.self, forKey: .id)
-        let pm = try c.decodeIfPresent(JSONValue.self, forKey: .providerMetadata)
-        switch try c.decode(String.self, forKey: .sourceType) {
-        case "url":
-            self = .url(id: id, url: try c.decode(String.self, forKey: .url),
-                        title: try c.decodeIfPresent(String.self, forKey: .title), providerMetadata: pm)
-        case "document":
-            self = .document(id: id, mediaType: try c.decode(String.self, forKey: .mediaType),
-                             title: try c.decode(String.self, forKey: .title),
-                             filename: try c.decodeIfPresent(String.self, forKey: .filename),
-                             providerMetadata: pm)
-        default: throw aimuxDecodingError(c.codingPath, "invalid source type")
-        }
-    }
-
-    public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: Field.self)
-        switch self {
-        case .url(let id, let url, let title, let pm):
-            try c.encode("url", forKey: .sourceType); try c.encode(id, forKey: .id)
-            try c.encode(url, forKey: .url); try c.encodeIfPresent(title, forKey: .title)
-            try c.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .document(let id, let mediaType, let title, let filename, let pm):
-            try c.encode("document", forKey: .sourceType); try c.encode(id, forKey: .id)
-            try c.encode(mediaType, forKey: .mediaType); try c.encode(title, forKey: .title)
-            try c.encodeIfPresent(filename, forKey: .filename)
-            try c.encodeIfPresent(pm, forKey: .providerMetadata)
-        }
-    }
+/// Approval request exposed by text generation.
+public struct ToolApprovalRequestOutput: Codable, Equatable {
+    public var approvalId: String
+    public var toolCall: ToolCall
+    public var reason: String?
+    public var isAutomatic: Bool?
+    public var signature: String?
 }
 
-/// A content item in the generation result.
-///
-/// Wire: externally tagged — `{"Text":{...}}`, `{"ToolCall":{...}}`, …
-public enum GenerateContent: Codable, Equatable {
-    case text(text: String, providerMetadata: JSONValue?)
-    case toolCall(toolCallId: String, toolName: String, input: JSONValue,
-                  providerExecuted: Bool?, dynamic: Bool?,
-                  providerMetadata: JSONValue?)
+/// A content item in a generation result. Wire: internally tagged by `type`;
+/// the payload types of `toolCall`, `toolApprovalRequest` and
+/// `reasoningFile` differ between the provider-facing and the call-layer
+/// shape, hence the generics (see ``RawGenerateContent`` / ``TextContent``).
+public enum GenerateContent<
+    C: Codable & Equatable, A: Codable & Equatable, R: Codable & Equatable
+>: Codable, Equatable {
+    case text(text: String, providerMetadata: ProviderMetadata?)
+    case toolCall(C)
     case source(Source)
-    case reasoning(text: String, providerMetadata: JSONValue?)
-    case file(data: GeneratedFileData, mediaType: String, providerMetadata: JSONValue?)
-    case toolResult(toolCallId: String, toolName: String, result: JSONValue,
-                    isError: Bool?, preliminary: Bool?, dynamic: Bool?, providerMetadata: JSONValue?)
-
-    case custom(kind: String, providerMetadata: JSONValue?)
-    case reasoningFile(data: GeneratedFileData, mediaType: String, providerMetadata: JSONValue?)
-    case toolApprovalRequest(approvalId: String, toolCallId: String, providerMetadata: JSONValue?)
+    case reasoning(ReasoningOutput)
+    case file(GeneratedFile)
+    case reasoningFile(R)
+    case custom(kind: String, providerMetadata: ProviderMetadata?)
+    case toolApprovalRequest(A)
+    case toolResult(ToolResult)
 
-    private enum Field: String, CodingKey {
-        case text
-        case kind, approvalId = "approval_id", toolCall = "tool_call", reason, isAutomatic = "is_automatic", signature
-        case toolCallId = "tool_call_id", toolName = "tool_name", input, result
-        case providerExecuted = "provider_executed", dynamic
-        case providerMetadata = "provider_metadata"
-        case id, sourceType = "source_type", url, title
-        case isError = "is_error", preliminary
-        case data, mediaType = "media_type"
-    }
+    private enum Field: String, CodingKey { case type, text, kind, providerMetadata }
 
     public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: AnyCodingKey.self)
-        guard let key = c.allKeys.first?.stringValue else {
-            throw aimuxDecodingError(c.codingPath, "missing generate-content tag")
-        }
-        let n = try c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey(key))
-        switch key {
-        case "Text":
-            self = .text(text: try n.decode(String.self, forKey: .text),
-                         providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ToolCall":
-            self = .toolCall(toolCallId: try n.decode(String.self, forKey: .toolCallId),
-                             toolName: try n.decode(String.self, forKey: .toolName),
-                             input: try n.decode(JSONValue.self, forKey: .input),
-                             providerExecuted: try n.decodeIfPresent(Bool.self, forKey: .providerExecuted),
-                             dynamic: try n.decodeIfPresent(Bool.self, forKey: .dynamic),
-                             providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "Source":
-            self = .source(try Source(from: c.superDecoder(forKey: AnyCodingKey("Source"))))
-        case "Reasoning":
-            self = .reasoning(text: try n.decode(String.self, forKey: .text),
-                             providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "Custom":
-            self = .custom(kind: try n.decode(String.self, forKey: .kind),
-                           providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ReasoningFile":
-            self = .reasoningFile(data: try n.decode(GeneratedFileData.self, forKey: .data),
-                                  mediaType: try n.decode(String.self, forKey: .mediaType),
-                                  providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ToolApprovalRequest":
-            self = .toolApprovalRequest(approvalId: try n.decode(String.self, forKey: .approvalId),
-                                         toolCallId: try n.decode(String.self, forKey: .toolCallId),
-                                         providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "File":
-            self = .file(data: try n.decode(GeneratedFileData.self, forKey: .data),
-                         mediaType: try n.decode(String.self, forKey: .mediaType),
-                         providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ToolResult":
-            self = .toolResult(toolCallId: try n.decode(String.self, forKey: .toolCallId),
-                               toolName: try n.decode(String.self, forKey: .toolName),
-                               result: try n.decode(JSONValue.self, forKey: .result),
-                               isError: try n.decodeIfPresent(Bool.self, forKey: .isError),
-                               preliminary: try n.decodeIfPresent(Bool.self, forKey: .preliminary),
-                               dynamic: try n.decodeIfPresent(Bool.self, forKey: .dynamic),
-                               providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        default:
-            throw aimuxDecodingError(c.codingPath, "unknown generate-content tag \(key)")
+        let t = try decoder.tag()
+        switch t {
+        case "text", "custom":
+            let c = try decoder.container(keyedBy: Field.self)
+            let pm = try c.decodeIfPresent(ProviderMetadata.self, forKey: .providerMetadata)
+            self = t == "text"
+                ? .text(text: try c.decode(String.self, forKey: .text), providerMetadata: pm)
+                : .custom(kind: try c.decode(String.self, forKey: .kind), providerMetadata: pm)
+        case "tool-call": self = .toolCall(try C(from: decoder))
+        case "source": self = .source(try Source(from: decoder))
+        case "reasoning": self = .reasoning(try ReasoningOutput(from: decoder))
+        case "file": self = .file(try GeneratedFile(from: decoder))
+        case "reasoning-file": self = .reasoningFile(try R(from: decoder))
+        case "tool-approval-request": self = .toolApprovalRequest(try A(from: decoder))
+        case "tool-result": self = .toolResult(try ToolResult(from: decoder))
+        default: throw decoder.unknownTag(t, in: "content")
         }
     }
 
     public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: AnyCodingKey.self)
         switch self {
         case .text(let text, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Text"))
-            try n.encode(text, forKey: .text)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .toolCall(let toolCallId, let toolName, let input, let pe, let dyn, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ToolCall"))
-            try n.encode(toolCallId, forKey: .toolCallId)
-            try n.encode(toolName, forKey: .toolName)
-            try n.encode(input, forKey: .input)
-            try n.encodeIfPresent(pe, forKey: .providerExecuted)
-            try n.encodeIfPresent(dyn, forKey: .dynamic)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .source(let source):
-            try source.encode(to: c.superEncoder(forKey: AnyCodingKey("Source")))
-        case .reasoning(let text, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Reasoning"))
-            try n.encode(text, forKey: .text)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
+            var c = encoder.container(keyedBy: Field.self)
+            try c.encode("text", forKey: .type); try c.encode(text, forKey: .text)
+            try c.encodeIfPresent(pm, forKey: .providerMetadata)
         case .custom(let kind, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Custom"))
-            try n.encode(kind, forKey: .kind); try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .reasoningFile(let data, let mediaType, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ReasoningFile"))
-            try n.encode(data, forKey: .data); try n.encode(mediaType, forKey: .mediaType)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .toolApprovalRequest(let approvalId, let toolCallId, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ToolApprovalRequest"))
-            try n.encode(approvalId, forKey: .approvalId); try n.encode(toolCallId, forKey: .toolCallId)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .file(let data, let mediaType, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("File"))
-            try n.encode(data, forKey: .data)
-            try n.encode(mediaType, forKey: .mediaType)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .toolResult(let toolCallId, let toolName, let result, let ie, let prel, let dyn, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ToolResult"))
-            try n.encode(toolCallId, forKey: .toolCallId)
-            try n.encode(toolName, forKey: .toolName)
-            try n.encode(result, forKey: .result)
-            try n.encodeIfPresent(ie, forKey: .isError)
-            try n.encodeIfPresent(prel, forKey: .preliminary)
-            try n.encodeIfPresent(dyn, forKey: .dynamic)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
+            var c = encoder.container(keyedBy: Field.self)
+            try c.encode("custom", forKey: .type); try c.encode(kind, forKey: .kind)
+            try c.encodeIfPresent(pm, forKey: .providerMetadata)
+        case .toolCall(let v): try v.encode(to: encoder); try encoder.putTag("tool-call")
+        case .source(let v): try v.encode(to: encoder); try encoder.putTag("source")
+        case .reasoning(let v): try v.encode(to: encoder); try encoder.putTag("reasoning")
+        case .file(let v): try v.encode(to: encoder); try encoder.putTag("file")
+        case .reasoningFile(let v): try v.encode(to: encoder); try encoder.putTag("reasoning-file")
+        case .toolApprovalRequest(let v): try v.encode(to: encoder); try encoder.putTag("tool-approval-request")
+        case .toolResult(let v): try v.encode(to: encoder); try encoder.putTag("tool-result")
         }
     }
 }
 
-public struct ResponseInfo: Codable, Equatable {
-    public var id: String?
-    public var timestamp: String?
-    public var modelId: String?
-    public var headers: [String: String]?
-    public var body: JSONValue?
-    enum CodingKeys: String, CodingKey {
-        case id, timestamp, headers, body, modelId = "model_id"
-    }
-    public init(id: String? = nil, timestamp: String? = nil, modelId: String? = nil, headers: [String: String]? = nil, body: JSONValue? = nil) {
-        self.id = id; self.timestamp = timestamp; self.modelId = modelId; self.headers = headers; self.body = body
-    }
-}
-
-public struct RequestInfo: Codable, Equatable {
-    public var body: JSONValue?
-
-    public init(body: JSONValue? = nil) { self.body = body }
-}
+/// Provider-facing content (`GenerateResult.content`): raw tool calls.
+public typealias RawGenerateContent = GenerateContent<RawToolCall, RawToolApprovalRequest, GeneratedFile>
+/// Call-layer content (`GenerateTextResult.content`): parsed tool calls.
+public typealias TextContent = GenerateContent<ToolCall, ToolApprovalRequestOutput, ReasoningFileOutput>
 
 /// Raw provider result (the `raw` field of `GenerateTextResult`).
 public struct GenerateResult: Codable, Equatable {
-    public var content: [GenerateContent]
-    public var finishReason: FinishReason?
-    public var usage: Usage?
-    public var warnings: [Warning]?
-    public var providerMetadata: JSONValue?
-    public var response: ResponseInfo?
+    public var content: [RawGenerateContent]
+    public var finishReason: FinishReason
+    public var usage: Usage
+    public var warnings: [Warning]
+    public var providerMetadata: ProviderMetadata?
     public var request: RequestInfo?
-
-    enum CodingKeys: String, CodingKey {
-        case content
-        case finishReason = "finish_reason"
-        case usage, warnings
-        case providerMetadata = "provider_metadata"
-        case response, request
-    }
-
-    public init(content: [GenerateContent], finishReason: FinishReason? = nil, usage: Usage? = nil,
-                warnings: [Warning]? = nil, providerMetadata: JSONValue? = nil,
-                response: ResponseInfo? = nil, request: RequestInfo? = nil) {
-        self.content = content; self.finishReason = finishReason; self.usage = usage
-        self.warnings = warnings; self.providerMetadata = providerMetadata
-        self.response = response; self.request = request
-    }
+    public var response: ResponseInfo?
 }
 
-/// Result of `generate_text` (user-facing).
+/// Result of `generateText` (user-facing).
 public struct GenerateTextResult: Codable, Equatable {
+    /// Ordered generated content with parsed tool calls.
+    public var content: [TextContent]?
     /// The generated text (concatenated from all text content parts).
-    public var content: [JSONValue]?
     public var text: String
     /// Tool calls requested by the model.
     public var toolCalls: [ToolCall]
-    /// Why generation stopped.
     public var finishReason: FinishReason
-    /// Token usage.
     public var usage: Usage
-    /// Warnings produced while generating the response.
     public var warnings: [Warning]
     /// Raw provider result (for advanced use).
     public var raw: GenerateResult
-    /// Reasoning / thinking segments (M7). Weak type — use `raw.content` for full typing.
-    public var reasoning: [JSONValue]
-    /// Concatenated reasoning text (M7).
+    public var reasoning: [ReasoningPart]
+    /// Concatenated reasoning text.
     public var reasoningText: String
-    /// Sources / citations (M7).
     public var sources: [Source]
-    /// Files generated by the model (M7). Weak type.
-    public var files: [JSONValue]
-    /// Assistant messages ready to append for the next turn (M7).
+    public var files: [GeneratedFile]
+    /// Assistant messages ready to append for the next turn.
     public var responseMessages: [ModelMessage]
-    /// Raw provider-specific finish reason string (M12, e.g. "stop", "end_turn").
+    /// Raw provider-specific finish reason string (e.g. "stop", "end_turn").
     public var rawFinishReason: String?
-    /// Provider-specific metadata (e.g. Anthropic cache info). Mirrored from
-    /// `raw.provider_metadata` for top-level convenience. Weak type.
-    public var providerMetadata: JSONValue?
-    /// Response metadata (id, timestamp, model_id). Mirrored from `raw.response`
-    /// for top-level convenience.
-    public var response: ResponseMetadata
-    /// Total token usage across all steps. In single-step mode (aimux's
-    /// default), equals `usage`. Provided for AI SDK parity.
+    public var providerMetadata: ProviderMetadata?
+    public var request: RequestInfo
+    public var response: ResponseInfo
+    /// Total token usage across all steps (single step: equals `usage`).
     public var totalUsage: Usage
-
-    enum CodingKeys: String, CodingKey {
-        case content, text
-        case toolCalls = "tool_calls"
-        case finishReason = "finish_reason"
-        case usage, warnings, raw
-        case reasoning
-        case reasoningText = "reasoning_text"
-        case sources, files
-        case responseMessages = "response_messages"
-        case rawFinishReason = "raw_finish_reason"
-        case providerMetadata = "provider_metadata"
-        case response
-        case totalUsage = "total_usage"
-    }
-
-    public init(text: String, toolCalls: [ToolCall], finishReason: FinishReason,
-                usage: Usage, warnings: [Warning] = [], raw: GenerateResult,
-                reasoning: [JSONValue] = [], reasoningText: String = "",
-                sources: [Source] = [], files: [JSONValue] = [],
-                responseMessages: [ModelMessage] = [],
-                rawFinishReason: String? = nil,
-                providerMetadata: JSONValue? = nil,
-                response: ResponseMetadata = ResponseMetadata(),
-                totalUsage: Usage = Usage(inputTokens: InputTokenUsage(), outputTokens: OutputTokenUsage()), content: [JSONValue]? = nil) {
-        self.content = content
-        self.text = text; self.toolCalls = toolCalls; self.finishReason = finishReason
-        self.usage = usage; self.warnings = warnings; self.raw = raw
-        self.reasoning = reasoning; self.reasoningText = reasoningText
-        self.sources = sources; self.files = files
-        self.responseMessages = responseMessages
-        self.rawFinishReason = rawFinishReason
-        self.providerMetadata = providerMetadata
-        self.response = response
-        self.totalUsage = totalUsage
-    }
 }
 
-/// Result of `generate_object` (user-facing, M12). The parsed JSON object plus
-/// convenience fields from the underlying `generate_text` call.
+/// Result of `generateObject`: the parsed JSON object plus convenience fields
+/// from the underlying `generateText` call.
 public struct GenerateObjectResult: Codable, Equatable {
-    /// The parsed JSON object returned by the model (arbitrary JSON, weak type).
+    /// The parsed JSON object returned by the model.
     public var object: JSONValue
-    /// Why generation stopped.
     public var finishReason: FinishReason
-    /// Raw provider-specific finish reason string.
     public var rawFinishReason: String?
-    /// Token usage.
     public var usage: Usage
-    /// Warnings from the provider.
     public var warnings: [Warning]
     /// Concatenated reasoning text (if the model produced reasoning/thinking).
     public var reasoning: String?
-    /// Provider-specific metadata (e.g. Anthropic cache info). Weak type.
-    public var providerMetadata: JSONValue?
-    /// Response metadata (id, timestamp, model_id).
+    public var providerMetadata: ProviderMetadata?
     public var response: ResponseMetadata
-    /// The full `generate_text` result (for advanced use).
+    /// The full `generateText` result (for advanced use).
     public var raw: GenerateTextResult
+}
 
-    enum CodingKeys: String, CodingKey {
-        case object
-        case finishReason = "finish_reason"
-        case rawFinishReason = "raw_finish_reason"
-        case usage, warnings
-        case reasoning
-        case providerMetadata = "provider_metadata"
-        case response, raw
-    }
-
-    public init(object: JSONValue, finishReason: FinishReason,
-                rawFinishReason: String? = nil,
-                usage: Usage = Usage(inputTokens: InputTokenUsage(), outputTokens: OutputTokenUsage()),
-                warnings: [Warning] = [],
-                reasoning: String? = nil,
-                providerMetadata: JSONValue? = nil,
-                response: ResponseMetadata = ResponseMetadata(),
-                raw: GenerateTextResult) {
-        self.object = object; self.finishReason = finishReason
-        self.rawFinishReason = rawFinishReason; self.usage = usage
-        self.warnings = warnings; self.reasoning = reasoning
-        self.providerMetadata = providerMetadata; self.response = response
-        self.raw = raw
-    }
-}
-
-/// Aggregated result of `stream_text().consume()` (M11). Mirrors
-/// `GenerateTextResult`'s user-facing fields (without `raw`, since streaming
-/// has no `GenerateResult` equivalent).
+/// Aggregated result of `consumeStreamText`; `GenerateTextResult`'s fields
+/// without `raw`.
 public struct StreamTextResultAggregated: Codable, Equatable {
-    /// The generated text (concatenated `TextDelta`).
-    public var content: [JSONValue]?
+    public var content: [TextContent]?
     public var text: String
-    /// Reasoning / thinking segments. Weak type.
-    public var reasoning: [JSONValue]
-    /// Concatenated reasoning text.
+    public var reasoning: [ReasoningPart]
     public var reasoningText: String
-    /// Tool calls requested by the model.
     public var toolCalls: [ToolCall]
-    /// Sources / citations.
     public var sources: [Source]
-    /// Files generated by the model. Weak type.
-    public var files: [JSONValue]
-    /// Why generation stopped.
+    public var files: [GeneratedFile]
     public var finishReason: FinishReason
-    /// Raw provider-specific finish reason string.
     public var rawFinishReason: String?
-    /// Token usage.
     public var usage: Usage
-    /// Total token usage across all steps. In single-step mode (aimux's
-    /// default), equals `usage`. Provided for AI SDK parity.
     public var totalUsage: Usage
-    /// Warnings from the provider.
     public var warnings: [Warning]
-    /// Provider-specific metadata from the Finish chunk. Weak type.
-    public var providerMetadata: JSONValue?
-    /// Response metadata (id, timestamp, model_id) if emitted by the stream.
-    public var response: ResponseMetadata?
-    /// Assistant messages ready to append for the next turn.
+    public var providerMetadata: ProviderMetadata?
+    public var request: RequestInfo
+    public var response: ResponseInfo
     public var responseMessages: [ModelMessage]
-
-    enum CodingKeys: String, CodingKey {
-        case content, text
-        case reasoning
-        case reasoningText = "reasoning_text"
-        case toolCalls = "tool_calls"
-        case sources, files
-        case finishReason = "finish_reason"
-        case rawFinishReason = "raw_finish_reason"
-        case usage
-        case totalUsage = "total_usage"
-        case warnings
-        case providerMetadata = "provider_metadata"
-        case response
-        case responseMessages = "response_messages"
-    }
-
-    public init(text: String = "", reasoning: [JSONValue] = [],
-                reasoningText: String = "", toolCalls: [ToolCall] = [],
-                sources: [Source] = [], files: [JSONValue] = [],
-                finishReason: FinishReason, rawFinishReason: String? = nil,
-                usage: Usage = Usage(inputTokens: InputTokenUsage(), outputTokens: OutputTokenUsage()),
-                totalUsage: Usage = Usage(inputTokens: InputTokenUsage(), outputTokens: OutputTokenUsage()),
-                warnings: [Warning] = [],
-                providerMetadata: JSONValue? = nil,
-                response: ResponseMetadata? = nil,
-                responseMessages: [ModelMessage] = [], content: [JSONValue]? = nil) {
-        self.content = content
-        self.text = text; self.reasoning = reasoning; self.reasoningText = reasoningText
-        self.toolCalls = toolCalls; self.sources = sources; self.files = files
-        self.finishReason = finishReason; self.rawFinishReason = rawFinishReason
-        self.usage = usage; self.totalUsage = totalUsage; self.warnings = warnings
-        self.providerMetadata = providerMetadata; self.response = response
-        self.responseMessages = responseMessages
-    }
 }
 
 // MARK: - TimeoutConfiguration
 
-/// Per-call timeout configuration.
-///
-/// Mirrors `TimeoutConfiguration.ts`. All values are milliseconds; `nil`
+/// Per-call timeout configuration. All values are milliseconds; `nil`
 /// disables the corresponding limit. A `total` timeout also covers retry
 /// backoff and the whole streamed response.
 public struct TimeoutConfiguration: Codable, Equatable {
@@ -1256,13 +1055,6 @@ public struct TimeoutConfiguration: Codable, Equatable {
     public var firstChunkMs: UInt64?
     public var chunkMs: UInt64?
 
-    enum CodingKeys: String, CodingKey {
-        case totalMs = "total_ms"
-        case stepMs = "step_ms"
-        case firstChunkMs = "first_chunk_ms"
-        case chunkMs = "chunk_ms"
-    }
-
     public init(totalMs: UInt64? = nil, stepMs: UInt64? = nil,
                 firstChunkMs: UInt64? = nil, chunkMs: UInt64? = nil) {
         self.totalMs = totalMs; self.stepMs = stepMs
@@ -1272,11 +1064,10 @@ public struct TimeoutConfiguration: Codable, Equatable {
 
 // MARK: - GenerateTextOptions
 
-/// User-facing options for `generate_text` / `stream_text`.
+/// User-facing options for `generateText` / `streamText`.
 ///
-/// All fields are optional. Encoding omits `nil` fields (the Rust side decodes
-/// missing `Option` fields as `None`), matching the partial-options usage of
-/// the raw API.
+/// All fields are optional. Encoding omits `nil` fields, matching the
+/// partial-options usage of the raw API.
 public struct GenerateTextOptions: Codable, Equatable {
     public var maxOutputTokens: UInt32?
     public var temperature: Double?
@@ -1290,7 +1081,7 @@ public struct GenerateTextOptions: Codable, Equatable {
     public var tools: [Tool]?
     public var toolChoice: ToolChoice?
     public var headers: [String: String]?
-    public var providerOptions: JSONValue?
+    public var providerOptions: ProviderOptions?
     public var reasoning: ReasoningEffort?
     public var instructions: String?
     public var maxRetries: UInt32?
@@ -1332,23 +1123,10 @@ public struct GenerateTextOptions: Codable, Equatable {
     var repairToolCallBox: RepairToolCallBox? = nil
 
     enum CodingKeys: String, CodingKey {
-        case maxOutputTokens = "max_output_tokens"
-        case temperature
-        case stopSequences = "stop_sequences"
-        case topP = "top_p"
-        case topK = "top_k"
-        case presencePenalty = "presence_penalty"
-        case frequencyPenalty = "frequency_penalty"
-        case responseFormat = "response_format"
-        case seed, tools
-        case toolChoice = "tool_choice"
-        case headers
-        case providerOptions = "provider_options"
-        case reasoning, instructions
-        case maxRetries = "max_retries"
-        case timeout
-        case includeRawChunks = "include_raw_chunks"
-        case sessionId = "session_id"
+        case maxOutputTokens, temperature, stopSequences, topP, topK, presencePenalty
+        case frequencyPenalty, responseFormat, seed, tools, toolChoice, headers
+        case providerOptions, reasoning, instructions, maxRetries, timeout
+        case includeRawChunks, sessionId
     }
 
     public init(maxOutputTokens: UInt32? = nil, temperature: Double? = nil,
@@ -1356,7 +1134,7 @@ public struct GenerateTextOptions: Codable, Equatable {
                 presencePenalty: Double? = nil, frequencyPenalty: Double? = nil,
                 responseFormat: ResponseFormat? = nil, seed: UInt64? = nil,
                 tools: [Tool]? = nil, toolChoice: ToolChoice? = nil,
-                headers: [String: String]? = nil, providerOptions: JSONValue? = nil,
+                headers: [String: String]? = nil, providerOptions: ProviderOptions? = nil,
                 reasoning: ReasoningEffort? = nil, instructions: String? = nil,
                 maxRetries: UInt32? = nil,
                 timeout: TimeoutConfiguration? = nil,
@@ -1376,242 +1154,119 @@ public struct GenerateTextOptions: Codable, Equatable {
     }
 }
 
-// MARK: - StreamPart
+// MARK: - TextStreamPart
 
-/// A single chunk in the stream returned by `stream_text`.
+/// A single chunk in the stream returned by `streamText` (the call-layer
+/// `TextStreamPart`: parsed tool calls).
 ///
-/// Wire: externally tagged — `{"TextDelta":{"id":"…","delta":"…"}}`, …
-public enum StreamPart: Codable, Equatable {
-    // P0: text
-    case textStart(id: String, providerMetadata: JSONValue?)
-    case textDelta(id: String, delta: String, providerMetadata: JSONValue?)
-    case textEnd(id: String, providerMetadata: JSONValue?)
-    // P0: stream lifecycle
+/// Wire: internally tagged by `type` — `{"type":"text-delta","id":"…","delta":"…"}`, …
+public enum TextStreamPart: Codable, Equatable {
+    case textStart(id: String, providerMetadata: ProviderMetadata?)
+    case textDelta(id: String, delta: String, providerMetadata: ProviderMetadata?)
+    case textEnd(id: String, providerMetadata: ProviderMetadata?)
     case streamStart(warnings: [Warning])
-    case finish(finishReason: FinishReason, usage: Usage, providerMetadata: JSONValue?)
+    case finish(finishReason: FinishReason, usage: Usage, providerMetadata: ProviderMetadata?)
+    case finishStep(finishReason: FinishReason, usage: Usage, providerMetadata: ProviderMetadata?,
+                    response: ResponseInfo)
+    /// `error` is an `AiMuxError` object tagged by `name`.
     case error(error: JSONValue)
-    // P1: tool calls
     case toolInputStart(id: String, toolName: String, providerExecuted: Bool?, dynamic: Bool?,
-                        title: String?, providerMetadata: JSONValue?)
-    case toolInputDelta(id: String, delta: String, providerMetadata: JSONValue?)
-    case toolInputEnd(id: String, providerMetadata: JSONValue?)
-    case toolCall(toolCallId: String, toolName: String, input: JSONValue,
-                  providerExecuted: Bool?, dynamic: Bool?,
-                  providerMetadata: JSONValue?,
-                  invalid: Bool?, error: JSONValue?)
-    case toolResult(toolCallId: String, toolName: String, result: JSONValue,
-                    isError: Bool?, preliminary: Bool?, dynamic: Bool?, providerMetadata: JSONValue?)
-    // P2: file
-    case file(data: GeneratedFileData, mediaType: String, providerMetadata: JSONValue?)
-    // P2: reasoning
-    case reasoningStart(id: String, providerMetadata: JSONValue?)
-    case reasoningDelta(id: String, delta: String, providerMetadata: JSONValue?)
-    case reasoningEnd(id: String, providerMetadata: JSONValue?)
-    // P2: metadata
-    case responseMetadata(id: String?, timestamp: String?, modelId: String?)
+                        title: String?, providerMetadata: ProviderMetadata?)
+    case toolInputDelta(id: String, delta: String, providerMetadata: ProviderMetadata?)
+    case toolInputEnd(id: String, providerMetadata: ProviderMetadata?)
+    case toolCall(ToolCall)
+    case toolResult(ToolResult)
+    case file(GeneratedFile)
+    case reasoningFile(ReasoningFileOutput)
+    case custom(kind: String, providerMetadata: ProviderMetadata?)
+    case toolApprovalRequest(ToolApprovalRequestOutput)
+    case reasoningStart(id: String, providerMetadata: ProviderMetadata?)
+    case reasoningDelta(id: String, delta: String, providerMetadata: ProviderMetadata?)
+    case reasoningEnd(id: String, providerMetadata: ProviderMetadata?)
     case source(Source)
     case raw(rawValue: JSONValue)
 
-    case custom(kind: String, providerMetadata: JSONValue?)
-    case reasoningFile(file: GeneratedFile, providerMetadata: JSONValue?)
-    case toolApprovalRequest(approvalId: String, toolCall: ToolCall, reason: String?, isAutomatic: Bool?, signature: String?)
-
     private enum Field: String, CodingKey {
-        case file, kind, approvalId = "approval_id", toolCall = "tool_call", reason, isAutomatic = "is_automatic", signature
-        case id, delta, warnings, usage, text
-        case finishReason = "finish_reason", providerMetadata = "provider_metadata"
-        case error
-        case toolName = "tool_name", toolCallId = "tool_call_id", input, result
-        case providerExecuted = "provider_executed", dynamic, invalid
-        case isError = "is_error", preliminary
-        case timestamp, modelId = "model_id"
-        case sourceType = "source_type", url, title
-        case data, mediaType = "media_type"
-        case rawValue = "raw_value"
+        case type, id, delta, warnings, finishReason, usage, providerMetadata, response, error
+        case toolName, providerExecuted, dynamic, title, kind, rawValue
     }
 
     public init(from decoder: Decoder) throws {
-        let c = try decoder.container(keyedBy: AnyCodingKey.self)
-        guard let key = c.allKeys.first?.stringValue else {
-            throw aimuxDecodingError(c.codingPath, "missing stream-part tag")
-        }
-        // Unit-ish variants with no nested payload (none in StreamPart) would be
-        // bare strings; every variant here carries a payload object.
-        let n = try c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey(key))
-        switch key {
-        case "TextStart":
-            self = .textStart(id: try n.decode(String.self, forKey: .id),
-                              providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "TextDelta":
-            self = .textDelta(id: try n.decode(String.self, forKey: .id),
-                              delta: try n.decode(String.self, forKey: .delta),
-                              providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "TextEnd":
-            self = .textEnd(id: try n.decode(String.self, forKey: .id),
-                            providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "StreamStart":
-            self = .streamStart(warnings: try n.decode([Warning].self, forKey: .warnings))
-        case "Finish":
-            self = .finish(finishReason: try n.decode(FinishReason.self, forKey: .finishReason),
-                           usage: try n.decode(Usage.self, forKey: .usage),
-                           providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "Error":
-            self = .error(error: try n.decode(JSONValue.self, forKey: .error))
-        case "ToolInputStart":
-            self = .toolInputStart(id: try n.decode(String.self, forKey: .id),
-                                   toolName: try n.decode(String.self, forKey: .toolName),
-                                   providerExecuted: try n.decodeIfPresent(Bool.self, forKey: .providerExecuted),
-                                   dynamic: try n.decodeIfPresent(Bool.self, forKey: .dynamic),
-                                   title: try n.decodeIfPresent(String.self, forKey: .title),
-                                   providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ToolInputDelta":
-            self = .toolInputDelta(id: try n.decode(String.self, forKey: .id),
-                                   delta: try n.decode(String.self, forKey: .delta),
-                                   providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ToolInputEnd":
-            self = .toolInputEnd(id: try n.decode(String.self, forKey: .id),
-                                 providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ToolCall":
-            self = .toolCall(toolCallId: try n.decode(String.self, forKey: .toolCallId),
-                             toolName: try n.decode(String.self, forKey: .toolName),
-                             input: try n.decode(JSONValue.self, forKey: .input),
-                             providerExecuted: try n.decodeIfPresent(Bool.self, forKey: .providerExecuted),
-                             dynamic: try n.decodeIfPresent(Bool.self, forKey: .dynamic),
-                             providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata),
-                             invalid: try n.decodeIfPresent(Bool.self, forKey: .invalid),
-                             error: try n.decodeIfPresent(JSONValue.self, forKey: .error))
-        case "ToolResult":
-            self = .toolResult(toolCallId: try n.decode(String.self, forKey: .toolCallId),
-                               toolName: try n.decode(String.self, forKey: .toolName),
-                               result: try n.decode(JSONValue.self, forKey: .result),
-                               isError: try n.decodeIfPresent(Bool.self, forKey: .isError),
-                               preliminary: try n.decodeIfPresent(Bool.self, forKey: .preliminary),
-                               dynamic: try n.decodeIfPresent(Bool.self, forKey: .dynamic),
-                               providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "Custom":
-            self = .custom(kind: try n.decode(String.self, forKey: .kind),
-                           providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ReasoningFile":
-            self = .reasoningFile(file: try n.decode(GeneratedFile.self, forKey: .file),
-                                  providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ToolApprovalRequest":
-            self = .toolApprovalRequest(approvalId: try n.decode(String.self, forKey: .approvalId),
-                                         toolCall: try n.decode(ToolCall.self, forKey: .toolCall),
-                                         reason: try n.decodeIfPresent(String.self, forKey: .reason),
-                                         isAutomatic: try n.decodeIfPresent(Bool.self, forKey: .isAutomatic),
-                                         signature: try n.decodeIfPresent(String.self, forKey: .signature))
-        case "File":
-            self = .file(data: try n.decode(GeneratedFileData.self, forKey: .data),
-                         mediaType: try n.decode(String.self, forKey: .mediaType),
-                         providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ReasoningStart":
-            self = .reasoningStart(id: try n.decode(String.self, forKey: .id),
-                                   providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ReasoningDelta":
-            self = .reasoningDelta(id: try n.decode(String.self, forKey: .id),
-                                   delta: try n.decode(String.self, forKey: .delta),
-                                   providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ReasoningEnd":
-            self = .reasoningEnd(id: try n.decode(String.self, forKey: .id),
-                                 providerMetadata: try n.decodeIfPresent(JSONValue.self, forKey: .providerMetadata))
-        case "ResponseMetadata":
-            self = .responseMetadata(id: try n.decodeIfPresent(String.self, forKey: .id),
-                                     timestamp: try n.decodeIfPresent(String.self, forKey: .timestamp),
-                                     modelId: try n.decodeIfPresent(String.self, forKey: .modelId))
-        case "Source":
-            self = .source(try Source(from: c.superDecoder(forKey: AnyCodingKey("Source"))))
-        case "Raw":
-            self = .raw(rawValue: try n.decode(JSONValue.self, forKey: .rawValue))
-        default:
-            throw aimuxDecodingError(c.codingPath, "unknown stream-part tag \(key)")
+        let t = try decoder.tag()
+        let c = try decoder.container(keyedBy: Field.self)
+        func pm() throws -> ProviderMetadata? { try c.decodeIfPresent(ProviderMetadata.self, forKey: .providerMetadata) }
+        func id() throws -> String { try c.decode(String.self, forKey: .id) }
+        func delta() throws -> String { try c.decode(String.self, forKey: .delta) }
+        switch t {
+        case "text-start": self = .textStart(id: try id(), providerMetadata: try pm())
+        case "text-delta": self = .textDelta(id: try id(), delta: try delta(), providerMetadata: try pm())
+        case "text-end": self = .textEnd(id: try id(), providerMetadata: try pm())
+        case "stream-start": self = .streamStart(warnings: try c.decode([Warning].self, forKey: .warnings))
+        case "finish":
+            self = .finish(finishReason: try c.decode(FinishReason.self, forKey: .finishReason),
+                           usage: try c.decode(Usage.self, forKey: .usage), providerMetadata: try pm())
+        case "finish-step":
+            self = .finishStep(finishReason: try c.decode(FinishReason.self, forKey: .finishReason),
+                               usage: try c.decode(Usage.self, forKey: .usage), providerMetadata: try pm(),
+                               response: try c.decode(ResponseInfo.self, forKey: .response))
+        case "error": self = .error(error: try c.decode(JSONValue.self, forKey: .error))
+        case "tool-input-start":
+            self = .toolInputStart(id: try id(), toolName: try c.decode(String.self, forKey: .toolName),
+                                   providerExecuted: try c.decodeIfPresent(Bool.self, forKey: .providerExecuted),
+                                   dynamic: try c.decodeIfPresent(Bool.self, forKey: .dynamic),
+                                   title: try c.decodeIfPresent(String.self, forKey: .title), providerMetadata: try pm())
+        case "tool-input-delta": self = .toolInputDelta(id: try id(), delta: try delta(), providerMetadata: try pm())
+        case "tool-input-end": self = .toolInputEnd(id: try id(), providerMetadata: try pm())
+        case "tool-call": self = .toolCall(try ToolCall(from: decoder))
+        case "tool-result": self = .toolResult(try ToolResult(from: decoder))
+        case "file": self = .file(try GeneratedFile(from: decoder))
+        case "reasoning-file": self = .reasoningFile(try ReasoningFileOutput(from: decoder))
+        case "custom": self = .custom(kind: try c.decode(String.self, forKey: .kind), providerMetadata: try pm())
+        case "tool-approval-request": self = .toolApprovalRequest(try ToolApprovalRequestOutput(from: decoder))
+        case "reasoning-start": self = .reasoningStart(id: try id(), providerMetadata: try pm())
+        case "reasoning-delta": self = .reasoningDelta(id: try id(), delta: try delta(), providerMetadata: try pm())
+        case "reasoning-end": self = .reasoningEnd(id: try id(), providerMetadata: try pm())
+        case "source": self = .source(try Source(from: decoder))
+        case "raw": self = .raw(rawValue: try c.decode(JSONValue.self, forKey: .rawValue))
+        default: throw decoder.unknownTag(t, in: "stream part")
         }
     }
 
     public func encode(to encoder: Encoder) throws {
-        var c = encoder.container(keyedBy: AnyCodingKey.self)
+        var c = encoder.container(keyedBy: Field.self)
+        func head(_ t: String, id: String? = nil, delta: String? = nil, _ pm: ProviderMetadata? = nil) throws {
+            try c.encode(t, forKey: .type)
+            try c.encodeIfPresent(id, forKey: .id); try c.encodeIfPresent(delta, forKey: .delta)
+            try c.encodeIfPresent(pm, forKey: .providerMetadata)
+        }
         switch self {
-        case .textStart(let id, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("TextStart"))
-            try n.encode(id, forKey: .id); try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .textDelta(let id, let delta, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("TextDelta"))
-            try n.encode(id, forKey: .id); try n.encode(delta, forKey: .delta)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .textEnd(let id, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("TextEnd"))
-            try n.encode(id, forKey: .id); try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .streamStart(let warnings):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("StreamStart"))
-            try n.encode(warnings, forKey: .warnings)
-        case .finish(let fr, let usage, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Finish"))
-            try n.encode(fr, forKey: .finishReason); try n.encode(usage, forKey: .usage)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .error(let error):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Error"))
-            try n.encode(error, forKey: .error)
-        case .toolInputStart(let id, let toolName, let pe, let dyn, let title, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ToolInputStart"))
-            try n.encode(id, forKey: .id); try n.encode(toolName, forKey: .toolName)
-            try n.encodeIfPresent(pe, forKey: .providerExecuted); try n.encodeIfPresent(dyn, forKey: .dynamic)
-            try n.encodeIfPresent(title, forKey: .title); try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .toolInputDelta(let id, let delta, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ToolInputDelta"))
-            try n.encode(id, forKey: .id); try n.encode(delta, forKey: .delta)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .toolInputEnd(let id, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ToolInputEnd"))
-            try n.encode(id, forKey: .id); try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .toolCall(let toolCallId, let toolName, let input, let pe, let dyn, let pm, let inv, let err):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ToolCall"))
-            try n.encode(toolCallId, forKey: .toolCallId); try n.encode(toolName, forKey: .toolName)
-            try n.encode(input, forKey: .input)
-            try n.encodeIfPresent(pe, forKey: .providerExecuted); try n.encodeIfPresent(dyn, forKey: .dynamic)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-            try n.encodeIfPresent(inv, forKey: .invalid); try n.encodeIfPresent(err, forKey: .error)
-        case .toolResult(let toolCallId, let toolName, let result, let ie, let prel, let dyn, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ToolResult"))
-            try n.encode(toolCallId, forKey: .toolCallId); try n.encode(toolName, forKey: .toolName)
-            try n.encode(result, forKey: .result)
-            try n.encodeIfPresent(ie, forKey: .isError); try n.encodeIfPresent(prel, forKey: .preliminary)
-            try n.encodeIfPresent(dyn, forKey: .dynamic)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .custom(let kind, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Custom"))
-            try n.encode(kind, forKey: .kind); try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .reasoningFile(let file, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ReasoningFile"))
-            try n.encode(file, forKey: .file)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .toolApprovalRequest(let approvalId, let toolCall, let reason, let isAutomatic, let signature):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ToolApprovalRequest"))
-            try n.encode(approvalId, forKey: .approvalId); try n.encode(toolCall, forKey: .toolCall)
-            try n.encodeIfPresent(reason, forKey: .reason); try n.encodeIfPresent(isAutomatic, forKey: .isAutomatic)
-            try n.encodeIfPresent(signature, forKey: .signature)
-        case .file(let data, let mediaType, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("File"))
-            try n.encode(data, forKey: .data); try n.encode(mediaType, forKey: .mediaType)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .reasoningStart(let id, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ReasoningStart"))
-            try n.encode(id, forKey: .id); try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .reasoningDelta(let id, let delta, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ReasoningDelta"))
-            try n.encode(id, forKey: .id); try n.encode(delta, forKey: .delta)
-            try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .reasoningEnd(let id, let pm):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ReasoningEnd"))
-            try n.encode(id, forKey: .id); try n.encodeIfPresent(pm, forKey: .providerMetadata)
-        case .responseMetadata(let id, let timestamp, let modelId):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("ResponseMetadata"))
-            try n.encodeIfPresent(id, forKey: .id)
-            try n.encodeIfPresent(timestamp, forKey: .timestamp)
-            try n.encodeIfPresent(modelId, forKey: .modelId)
-        case .source(let source):
-            try source.encode(to: c.superEncoder(forKey: AnyCodingKey("Source")))
-        case .raw(let rawValue):
-            var n = c.nestedContainer(keyedBy: Field.self, forKey: AnyCodingKey("Raw"))
-            try n.encode(rawValue, forKey: .rawValue)
+        case .textStart(let i, let pm): try head("text-start", id: i, pm)
+        case .textDelta(let i, let d, let pm): try head("text-delta", id: i, delta: d, pm)
+        case .textEnd(let i, let pm): try head("text-end", id: i, pm)
+        case .streamStart(let w): try head("stream-start"); try c.encode(w, forKey: .warnings)
+        case .finish(let fr, let u, let pm):
+            try head("finish", pm); try c.encode(fr, forKey: .finishReason); try c.encode(u, forKey: .usage)
+        case .finishStep(let fr, let u, let pm, let r):
+            try head("finish-step", pm); try c.encode(fr, forKey: .finishReason); try c.encode(u, forKey: .usage)
+            try c.encode(r, forKey: .response)
+        case .error(let e): try head("error"); try c.encode(e, forKey: .error)
+        case .toolInputStart(let i, let n, let pe, let dyn, let title, let pm):
+            try head("tool-input-start", id: i, pm); try c.encode(n, forKey: .toolName)
+            try c.encodeIfPresent(pe, forKey: .providerExecuted); try c.encodeIfPresent(dyn, forKey: .dynamic)
+            try c.encodeIfPresent(title, forKey: .title)
+        case .toolInputDelta(let i, let d, let pm): try head("tool-input-delta", id: i, delta: d, pm)
+        case .toolInputEnd(let i, let pm): try head("tool-input-end", id: i, pm)
+        case .toolCall(let v): try v.encode(to: encoder); try encoder.putTag("tool-call")
+        case .toolResult(let v): try v.encode(to: encoder); try encoder.putTag("tool-result")
+        case .file(let v): try v.encode(to: encoder); try encoder.putTag("file")
+        case .reasoningFile(let v): try v.encode(to: encoder); try encoder.putTag("reasoning-file")
+        case .custom(let k, let pm): try head("custom", pm); try c.encode(k, forKey: .kind)
+        case .toolApprovalRequest(let v): try v.encode(to: encoder); try encoder.putTag("tool-approval-request")
+        case .reasoningStart(let i, let pm): try head("reasoning-start", id: i, pm)
+        case .reasoningDelta(let i, let d, let pm): try head("reasoning-delta", id: i, delta: d, pm)
+        case .reasoningEnd(let i, let pm): try head("reasoning-end", id: i, pm)
+        case .source(let v): try v.encode(to: encoder); try encoder.putTag("source")
+        case .raw(let v): try head("raw"); try c.encode(v, forKey: .rawValue)
         }
     }
 }
@@ -1884,7 +1539,7 @@ public extension Model {
     /// Generate a structured JSON object with typed inputs/outputs (M12, RFC-0016).
     ///
     /// Same signature as the typed `generateText`; returns a decoded
-    /// `GenerateObjectResult`. Pass `response_format: { "Json": { ... } }`
+    /// `GenerateObjectResult`. Pass `responseFormat: .json(schema: …)`
     /// via `options` for schema control; the engine applies JSON repair
     /// before parsing.
     ///
@@ -1927,15 +1582,15 @@ public extension Model {
         return try JSONDecoder().decode(StreamTextResultAggregated.self, from: Data(resultJson.utf8))
     }
 
-    /// Stream text with typed `StreamPart`s.
+    /// Stream text with typed `TextStreamPart`s.
     ///
-    /// Each raw JSON-string part is decoded into a `StreamPart` before being
+    /// Each raw JSON-string part is decoded into a `TextStreamPart` before being
     /// passed to `onPart`. A part that fails to decode is reported via
     /// `onError` (the stream otherwise continues until done/error).
     func streamText(
         prompt: ModelPrompt,
         options: GenerateTextOptions? = nil,
-        onPart: @escaping (StreamPart) -> Void,
+        onPart: @escaping (TextStreamPart) -> Void,
         onDone: @escaping () -> Void,
         onError: @escaping (any Error) -> Void
     ) {
@@ -1957,7 +1612,7 @@ public extension Model {
                            let partJson = try repairedStreamPartJson(
                                json, promptJson: promptJson, optsJson: optsJson, options: options
                            )
-                           try onPart(JSONDecoder().decode(StreamPart.self, from: Data(partJson.utf8)))
+                           try onPart(JSONDecoder().decode(TextStreamPart.self, from: Data(partJson.utf8)))
                        } catch {
                            onError(error) // DecodingError from JSONDecoder, AimuxError from repair
                        }
@@ -1966,7 +1621,7 @@ public extension Model {
                    onError: onError)
     }
 
-    /// Stream text as an `AsyncSequence` of typed `StreamPart`s.
+    /// Stream text as an `AsyncSequence` of typed `TextStreamPart`s.
     ///
     /// The stream finishes on normal completion and throws `AimuxError`
     /// (AiMuxError failure, C codes preserved via `fromC`) or the native
@@ -1974,7 +1629,7 @@ public extension Model {
     func streamTextAsync(
         prompt: ModelPrompt,
         options: GenerateTextOptions? = nil
-    ) -> AsyncThrowingStream<StreamPart, Error> {
+    ) -> AsyncThrowingStream<TextStreamPart, Error> {
         AsyncThrowingStream { continuation in
             self.streamText(
                 prompt: prompt, options: options,
@@ -2070,8 +1725,8 @@ public extension Model {
 
 /// JSON (de)serialization helpers shared by the typed wrapper.
 fileprivate enum AimuxCodable {
-    /// Encode an `Encodable` value to a JSON string using the default key
-    /// strategy (snake_case mapping is handled per-type via `CodingKeys`).
+    /// Encode an `Encodable` value to a JSON string (the wire is camelCase, so
+    /// the default key strategy applies).
     static func jsonString<T: Encodable>(for value: T) throws -> String {
         let data = try JSONEncoder().encode(value)
         return String(data: data, encoding: .utf8) ?? ""
diff --git a/bindings/swift/Tests/AimuxTests/AimuxTests.swift b/bindings/swift/Tests/AimuxTests/AimuxTests.swift
index 77d22ed00..82b0df8b7 100644
--- a/bindings/swift/Tests/AimuxTests/AimuxTests.swift
+++ b/bindings/swift/Tests/AimuxTests/AimuxTests.swift
@@ -379,9 +379,7 @@ final class AimuxTests: XCTestCase {
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let resultJson = try model.generateText(prompt: jsonEncodeString("What is Rust?"))
         let r = parseJSON(resultJson)
         XCTAssertEqual(r["text"] as? String, "Rust is a systems programming language.")
@@ -390,21 +388,19 @@ final class AimuxTests: XCTestCase {
     }
 
     /// Tool-call parsing: mock returns tool_calls; verify convenience field
-    /// `tool_calls` and the structured `raw.content` ToolCall variant.
+    /// `toolCalls` and the structured `raw.content` `tool-call` item.
     func testE2EOpenAIToolCallParsing() throws {
         let server = MockHTTPServer(response: .json(openaiToolCallResponse))
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let opts = jsonEncode([
             "tools": [[
                 "type": "function",
                 "name": "get_weather",
                 "description": "Get weather for a location",
-                "input_schema": [
+                "inputSchema": [
                     "type": "object",
                     "properties": ["location": ["type": "string"]],
                     "required": ["location"],
@@ -415,22 +411,21 @@ final class AimuxTests: XCTestCase {
             prompt: jsonEncodeString("What's the weather in Tokyo?"), options: opts
         ))
 
-        // Convenience field: tool_calls extracted
-        let toolCalls = (r["tool_calls"] as? [[String: Any]]) ?? []
+        // Convenience field: toolCalls extracted
+        let toolCalls = (r["toolCalls"] as? [[String: Any]]) ?? []
         XCTAssertEqual(toolCalls.count, 1)
-        XCTAssertEqual(toolCalls[0]["tool_name"] as? String, "get_weather")
-        XCTAssertEqual(toolCalls[0]["tool_call_id"] as? String, "call_abc")
+        XCTAssertEqual(toolCalls[0]["toolName"] as? String, "get_weather")
+        XCTAssertEqual(toolCalls[0]["toolCallId"] as? String, "call_abc")
         let input = toolCalls[0]["input"] as? [String: Any]
         XCTAssertEqual(input?["location"] as? String, "Tokyo")
 
-        // Structured content: raw.content contains a ToolCall variant
+        // Structured content: raw.content contains a tool-call item
         let raw = r["raw"] as? [String: Any]
         let content = (raw?["content"] as? [[String: Any]]) ?? []
-        let tcVariant = content.first(where: { $0["ToolCall"] != nil })
-        XCTAssertNotNil(tcVariant, "raw.content should contain a ToolCall variant")
-        let tc = tcVariant?["ToolCall"] as? [String: Any]
-        XCTAssertEqual(tc?["tool_name"] as? String, "get_weather")
-        XCTAssertEqual(tc?["tool_call_id"] as? String, "call_abc")
+        let tc = content.first(where: { $0["type"] as? String == "tool-call" })
+        XCTAssertNotNil(tc, "raw.content should contain a tool-call item")
+        XCTAssertEqual(tc?["toolName"] as? String, "get_weather")
+        XCTAssertEqual(tc?["toolCallId"] as? String, "call_abc")
     }
 
     /// Multi-role messages: a system+user array reaches the provider verbatim.
@@ -439,9 +434,7 @@ final class AimuxTests: XCTestCase {
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let prompt = jsonEncode([
             ["role": "system", "content": "You are a helpful assistant."],
             ["role": "user", "content": "What is Rust?"],
@@ -459,25 +452,23 @@ final class AimuxTests: XCTestCase {
         XCTAssertEqual(msgs[1]["content"] as? String, "What is Rust?")
     }
 
-    /// ToolChoice: passing `tool_choice: "required"` reaches the provider body.
+    /// ToolChoice: passing `toolChoice: "required"` reaches the provider body.
     func testE2EOpenAIToolChoiceRequired() throws {
         let server = MockHTTPServer(response: .json(openaiToolCallResponse))
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let opts = jsonEncode([
             "tools": [[
                 "type": "function",
                 "name": "get_weather",
-                "input_schema": [
+                "inputSchema": [
                     "type": "object",
                     "properties": ["location": ["type": "string"]],
                 ],
             ]],
-            "tool_choice": "required",
+            "toolChoice": "required",
         ])
         // Must not throw.
         _ = try model.generateText(prompt: jsonEncodeString("Hello"), options: opts)
@@ -493,9 +484,7 @@ final class AimuxTests: XCTestCase {
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         var parts: [String] = []
         var streamErr: (any Error)?
         model.streamText(
@@ -509,26 +498,24 @@ final class AimuxTests: XCTestCase {
 
         let parsed = parts.map { parseJSON($0) }
         let text = parsed
-            .compactMap { ($0["TextDelta"] as? [String: Any])?["delta"] as? String }
+            .compactMap { $0["type"] as? String == "text-delta" ? $0["delta"] as? String : nil }
             .joined()
         XCTAssertEqual(text, "Hello world")
     }
 
-    /// Streaming tool call: the stream emits a tool-related StreamPart
-    /// (ToolCall / ToolInputStart / ToolInputDelta).
+    /// Streaming tool call: the stream emits a tool-related part
+    /// (`tool-call` / `tool-input-start` / `tool-input-delta`).
     func testE2EOpenAIStreamToolCall() throws {
         let server = MockHTTPServer(response: .sse(sse(openaiStreamToolEvents)))
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let opts = jsonEncode([
             "tools": [[
                 "type": "function",
                 "name": "get_weather",
-                "input_schema": [
+                "inputSchema": [
                     "type": "object",
                     "properties": ["location": ["type": "string"]],
                 ],
@@ -548,14 +535,12 @@ final class AimuxTests: XCTestCase {
 
         let parsed = parts.map { parseJSON($0) }
         let hasToolPart = parsed.contains {
-            $0["ToolCall"] != nil
-                || $0["ToolInputDelta"] != nil
-                || $0["ToolInputStart"] != nil
+            ["tool-call", "tool-input-delta", "tool-input-start"].contains($0["type"] as? String ?? "")
         }
-        XCTAssertTrue(hasToolPart, "stream should contain a tool-related StreamPart")
+        XCTAssertTrue(hasToolPart, "stream should contain a tool-related part")
 
-        if let toolCall = parsed.first(where: { $0["ToolCall"] != nil })?["ToolCall"] as? [String: Any] {
-            XCTAssertEqual(toolCall["tool_name"] as? String, "get_weather")
+        if let toolCall = parsed.first(where: { $0["type"] as? String == "tool-call" }) {
+            XCTAssertEqual(toolCall["toolName"] as? String, "get_weather")
         }
     }
 
@@ -575,7 +560,7 @@ final class AimuxTests: XCTestCase {
 
         let model = try Model.anthropic(
             apiKey: "test-key", modelId: "claude-3-5-sonnet-20241022",
-            baseUrl: server.baseURL
+            baseUrl: server.baseURL + "/v1"
         )
         let r = parseJSON(try model.generateText(prompt: jsonEncodeString("Hello")))
         XCTAssertEqual(r["text"] as? String, "Hello from Claude!")
@@ -589,9 +574,10 @@ final class AimuxTests: XCTestCase {
     /// end-to-end (Swift → FFI → reqwest → mock) and that the second request
     /// body carries the tool-result message.
     ///
-    /// The replayed messages use the aimux content-part shape (type-tagged
-    /// `ToolCall` / `ToolResult`), which is what `generate_text` deserializes
-    /// and the OpenAI converter maps to the wire `tool_calls` / `tool` format.
+    /// The replayed messages use the content-part shape (`tool-call` /
+    /// `tool-result` with a typed `output`), which is what `generateText`
+    /// deserializes and the OpenAI converter maps to the wire `tool_calls` /
+    /// `tool` format.
     func testE2EToolCallRoundTrip() throws {
         let server = MockHTTPServer(responses: [
             .json(openaiToolCallResponse),
@@ -600,15 +586,13 @@ final class AimuxTests: XCTestCase {
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let opts = jsonEncode([
             "tools": [[
                 "type": "function",
                 "name": "get_weather",
                 "description": "Get weather for a location",
-                "input_schema": [
+                "inputSchema": [
                     "type": "object",
                     "properties": ["location": ["type": "string"]],
                     "required": ["location"],
@@ -620,24 +604,25 @@ final class AimuxTests: XCTestCase {
         let r1 = parseJSON(try model.generateText(
             prompt: jsonEncodeString("What's the weather in Tokyo?"), options: opts
         ))
-        let toolCalls = (r1["tool_calls"] as? [[String: Any]]) ?? []
+        let toolCalls = (r1["toolCalls"] as? [[String: Any]]) ?? []
         XCTAssertEqual(toolCalls.count, 1)
-        XCTAssertEqual(toolCalls[0]["tool_name"] as? String, "get_weather")
-        XCTAssertEqual(toolCalls[0]["tool_call_id"] as? String, "call_abc")
+        XCTAssertEqual(toolCalls[0]["toolName"] as? String, "get_weather")
+        XCTAssertEqual(toolCalls[0]["toolCallId"] as? String, "call_abc")
 
         // 2) Second call: replay the conversation with the ToolResult filled in.
         let secondPrompt = jsonEncode([
             ["role": "user", "content": "What's the weather in Tokyo?"],
             ["role": "assistant", "content": [[
-                "type": "tool_call",
-                "tool_call_id": "call_abc",
-                "tool_name": "get_weather",
+                "type": "tool-call",
+                "toolCallId": "call_abc",
+                "toolName": "get_weather",
                 "input": ["location": "Tokyo"],
             ]]],
             ["role": "tool", "content": [[
-                "type": "tool_result",
-                "tool_call_id": "call_abc",
-                "output": ["temperature": 22, "condition": "sunny"],
+                "type": "tool-result",
+                "toolCallId": "call_abc",
+                "toolName": "get_weather",
+                "output": ["type": "json", "value": ["temperature": 22, "condition": "sunny"]],
             ]]],
         ])
         let r2 = parseJSON(try model.generateText(prompt: secondPrompt, options: opts))
diff --git a/bindings/swift/Tests/AimuxTests/MockHTTPServer.swift b/bindings/swift/Tests/AimuxTests/MockHTTPServer.swift
index f77fdb474..abbce9213 100644
--- a/bindings/swift/Tests/AimuxTests/MockHTTPServer.swift
+++ b/bindings/swift/Tests/AimuxTests/MockHTTPServer.swift
@@ -11,12 +11,23 @@
 // close` so each FFI call maps to exactly one connection.
 
 import Foundation
+@testable import Aimux
 #if canImport(Glibc)
 import Glibc
 #elseif canImport(Darwin)
 import Darwin
 #endif
 
+/// A model that speaks the OpenAI chat-completions wire format at `baseURL`:
+/// the registry's OpenAI-compatible preset (`Model.openai` talks to the
+/// Responses API).
+func chatModel(_ baseURL: String) throws -> Model {
+    try Model.provider(
+        name: "deepseek", apiKey: "test-key", modelId: "gpt-4o",
+        configJson: #"{"baseUrl":"\#(baseURL)"}"#
+    )
+}
+
 /// A preset HTTP response returned by `MockHTTPServer`.
 struct MockResponse {
     let status: Int
diff --git a/bindings/swift/Tests/AimuxTests/MultimodalTests.swift b/bindings/swift/Tests/AimuxTests/MultimodalTests.swift
index ea416bbc6..8fb4b2b64 100644
--- a/bindings/swift/Tests/AimuxTests/MultimodalTests.swift
+++ b/bindings/swift/Tests/AimuxTests/MultimodalTests.swift
@@ -8,14 +8,14 @@
 //
 // Important: the methods return the aimux *canonical* wire shape produced by
 // the FFI, NOT the raw provider response. The FFI converts provider shapes,
-// e.g. OpenAI `data[].b64_json` → `images.Base64[]`, an OpenAI file object →
-// `provider_reference.openai`, raw audio bytes → `audio.Binary[]`. The canned
-// responses below are the provider shapes (what the mock returns); the
-// assertions check the converted aimux shapes (what the FFI returns).
+// e.g. OpenAI `data[].b64_json` → `images` (base64 strings), an OpenAI file
+// object → `providerReference.openai`, raw audio bytes → `audio` (a byte
+// array). The canned responses below are the provider shapes (what the mock
+// returns); the assertions decode the result into the typed result structs
+// (what the FFI returns).
 //
-// Call-options JSON always includes `"provider_options":{}` because the Rust
-// SharedProviderOptions (HashMap) requires a map, not null — matching the Go
-// bindings' non-omitempty `jsonObj` (see bindings/go/multimodal_types.go).
+// Image, video and file call options carry `"providerOptions":{}` where the
+// generated options type requires it (`ImageCallOptions`, `VideoCallOptions`).
 
 import XCTest
 @testable import Aimux
@@ -36,19 +36,17 @@ final class MultimodalTests: XCTestCase {
         let model = try EmbeddingModel.openai(
             apiKey: "test", modelId: "text-embedding-3-small", baseUrl: server.baseURL
         )
-        let r = parseJSON(try model.embed(values: jsonEncode(["hello"])))
+        let r = try decode(EmbeddingResult.self, try model.embed(values: jsonEncode(["hello"])))
 
         // `embeddings` is [[float]]; verify one embedding vector of dimension 3.
-        let embeddings = (r["embeddings"] as? [Any]) ?? []
-        XCTAssertEqual(embeddings.count, 1)
-        let firstDims = (embeddings.first as? [Any]) ?? []
-        XCTAssertEqual(firstDims.count, 3, "first embedding should have 3 dimensions")
+        XCTAssertEqual(r.embeddings.count, 1)
+        XCTAssertEqual(r.embeddings.first?.count, 3, "first embedding should have 3 dimensions")
     }
 
     // MARK: - Image
 
     /// OpenAI image: mock returns `data[].b64_json`; the FFI exposes it as
-    /// `images.Base64[]`.
+    /// `images` (base64 strings).
     func testE2E_ImageViaMock() throws {
         let server = MockHTTPServer(response: jsonResponse(
             #"{"data":[{"b64_json":"aW1hZ2Ux"}]}"#
@@ -62,14 +60,11 @@ final class MultimodalTests: XCTestCase {
         let opts = jsonEncode([
             "prompt": "otter",
             "n": 1,
-            "provider_options": [String: Any](),
+            "providerOptions": [String: Any](),
         ])
-        let r = parseJSON(try model.generate(options: opts))
+        let r = try decode(ImageResult.self, try model.generate(options: opts))
 
-        let images = (r["images"] as? [String: Any]) ?? [:]
-        let base64 = (images["Base64"] as? [String]) ?? []
-        XCTAssertEqual(base64.count, 1)
-        XCTAssertEqual(base64[0], "aW1hZ2Ux")
+        XCTAssertEqual(r.images, .base64(["aW1hZ2Ux"]))
     }
 
     // MARK: - Transcription
@@ -86,15 +81,16 @@ final class MultimodalTests: XCTestCase {
         let model = try TranscriptionModel.openai(
             apiKey: "test", modelId: "whisper-1", baseUrl: server.baseURL
         )
-        let r = parseJSON(try model.generate(audioBase64: "dGVzdA==", mediaType: "audio/mp3"))
+        let r = try decode(TranscriptionResult.self,
+                           try model.generate(audioBase64: "dGVzdA==", mediaType: "audio/mp3"))
 
-        XCTAssertEqual(r["text"] as? String, "Hello world")
+        XCTAssertEqual(r.text, "Hello world")
     }
 
     // MARK: - Reranking
 
     /// Cohere reranking: mock returns `results[]`; the FFI exposes it as
-    /// `ranking[]` with `index` and `relevance_score`.
+    /// `ranking[]` with `index` and `relevanceScore`.
     func testE2E_RerankingViaMock() throws {
         let server = MockHTTPServer(response: jsonResponse(
             #"{"results":[{"index":1,"relevance_score":0.95},{"index":0,"relevance_score":0.3}]}"#
@@ -107,16 +103,14 @@ final class MultimodalTests: XCTestCase {
         )
         let opts = jsonEncode([
             "query": "which?",
-            "documents": ["Text": ["values": ["doc1", "doc2"]]],
-            "top_n": 2,
-            "provider_options": [String: Any](),
+            "documents": ["type": "text", "values": ["doc1", "doc2"]],
+            "topN": 2,
         ])
-        let r = parseJSON(try model.rerank(options: opts))
+        let r = try decode(RerankingResult.self, try model.rerank(options: opts))
 
-        let ranking = (r["ranking"] as? [[String: Any]]) ?? []
-        XCTAssertEqual(ranking.count, 2)
-        XCTAssertEqual(ranking[0]["index"] as? Int, 1)
-        XCTAssertEqual(ranking[0]["relevance_score"] as? Double, 0.95)
+        XCTAssertEqual(r.ranking.count, 2)
+        XCTAssertEqual(r.ranking[0].index, 1)
+        XCTAssertEqual(r.ranking[0].relevanceScore, 0.95)
     }
 
     // MARK: - Search
@@ -133,20 +127,18 @@ final class MultimodalTests: XCTestCase {
         let model = try SearchModel.tavily(apiKey: "test", baseUrl: server.baseURL)
         let opts = jsonEncode([
             "query": "What is Rust?",
-            "provider_options": [String: Any](),
         ])
-        let r = parseJSON(try model.search(options: opts))
+        let r = try decode(SearchResult.self, try model.search(options: opts))
 
-        let results = (r["results"] as? [[String: Any]]) ?? []
-        XCTAssertEqual(results.count, 1)
-        XCTAssertEqual(results[0]["title"] as? String, "Rust")
-        XCTAssertEqual(r["answer"] as? String, "Rust is a systems language.")
+        XCTAssertEqual(r.results.count, 1)
+        XCTAssertEqual(r.results[0].title, "Rust")
+        XCTAssertEqual(r.answer, "Rust is a systems language.")
     }
 
     // MARK: - Files
 
     /// OpenAI files: mock returns an OpenAI file object; the FFI exposes the
-    /// file id as `provider_reference.openai`.
+    /// file id as `providerReference.openai`.
     func testE2E_FilesViaMock() throws {
         let server = MockHTTPServer(response: jsonResponse(
             #"{"id":"file-abc","object":"file","bytes":1024,"created_at":1234,"filename":"test.pdf","purpose":"assistants"}"#
@@ -155,10 +147,10 @@ final class MultimodalTests: XCTestCase {
         defer { server.stop() }
 
         let files = try Files.openai(apiKey: "test", baseUrl: server.baseURL)
-        let r = parseJSON(try files.upload(dataBase64: "dGVzdA==", mediaType: "application/pdf"))
+        let r = try decode(UploadFileResult.self,
+                           try files.upload(dataBase64: "dGVzdA==", mediaType: "application/pdf"))
 
-        let providerRef = (r["provider_reference"] as? [String: Any]) ?? [:]
-        XCTAssertEqual(providerRef["openai"] as? String, "file-abc")
+        XCTAssertEqual(r.providerReference["openai"], "file-abc")
     }
 
     // MARK: - Speech (TTS)
@@ -167,7 +159,7 @@ final class MultimodalTests: XCTestCase {
     /// `Content-Type: audio/mpeg`. Unlike a URLProtocol mock, this POSIX-socket
     /// MockHTTPServer supports arbitrary content-types via direct `MockResponse`
     /// construction, so we exercise the full round trip (Swift → FFI → reqwest →
-    /// mock). The FFI returns the bytes as `audio.Binary`.
+    /// mock). The FFI returns the bytes as `audio` (a byte array).
     func testE2E_SpeechViaMock() throws {
         // A non-empty binary payload (the literal bytes of a base64-looking
         // string, matching bindings/go/multimodal_withbase_test.go).
@@ -184,14 +176,14 @@ final class MultimodalTests: XCTestCase {
         let opts = jsonEncode([
             "text": "Hi",
             "voice": "alloy",
-            "output_format": "mp3",
-            "provider_options": [String: Any](),
+            "outputFormat": "mp3",
         ])
-        let r = parseJSON(try model.generate(options: opts))
+        let r = try decode(SpeechResult.self, try model.generate(options: opts))
 
-        let audio = (r["audio"] as? [String: Any]) ?? [:]
-        let binary = (audio["Binary"] as? [Any]) ?? []
-        XCTAssertFalse(binary.isEmpty, "speech audio should contain binary bytes")
+        guard case .binary(let bytes) = r.audio else {
+            return XCTFail("expected binary audio, got \(r.audio)")
+        }
+        XCTAssertFalse(bytes.isEmpty, "speech audio should contain binary bytes")
     }
 
     // MARK: - Video (construction + result parsing only)
@@ -199,22 +191,19 @@ final class MultimodalTests: XCTestCase {
     /// Google video uses a multi-step async API (POST predict → poll operation →
     /// fetch result) that a single-response mock server can't cover. We verify
     /// construction with a base URL (pointing at a non-listening port, never
-    /// contacted) and parse a canned VideoResult JSON, checking the `Url`
-    /// variant's `url`/`media_type` — mirroring the Go test.
+    /// contacted) and decode a canned VideoResult JSON, checking the `url`
+    /// variant's `url`/`mediaType` — mirroring the Go test.
     func testE2E_VideoConstructionAndParse() throws {
         let model = try VideoModel.google(
             apiKey: "test", modelId: "veo-3.0", baseUrl: "http://localhost:9999"
         )
         XCTAssertNotNil(model)
 
-        let r = parseJSON(
-            #"{"videos":[{"Url":{"url":"https://example.com/v.mp4","media_type":"video/mp4"}}]}"#
-        )
-        let videos = (r["videos"] as? [[String: Any]]) ?? []
-        XCTAssertEqual(videos.count, 1)
-        let urlData = (videos[0]["Url"] as? [String: Any]) ?? [:]
-        XCTAssertEqual(urlData["url"] as? String, "https://example.com/v.mp4")
-        XCTAssertEqual(urlData["media_type"] as? String, "video/mp4")
+        let r = try decode(VideoResult.self, """
+            {"videos":[{"type":"url","url":"https://example.com/v.mp4","mediaType":"video/mp4"}],
+             "warnings":[],"response":{}}
+            """)
+        XCTAssertEqual(r.videos, [.url(url: "https://example.com/v.mp4", mediaType: "video/mp4")])
     }
 
     // MARK: - TranscriptionSession use-after-close
@@ -296,11 +285,7 @@ private func jsonEncode(_ object: Any) -> String {
     return String(data: data, encoding: .utf8)!
 }
 
-/// Parse a JSON object string into `[String: Any]` (empty on failure).
-private func parseJSON(_ s: String) -> [String: Any] {
-    guard let data = s.data(using: .utf8),
-          let obj = try? JSONSerialization.jsonObject(with: data) as? [String: Any] else {
-        return [:]
-    }
-    return obj
+/// Decode a JSON result string into its typed result struct.
+private func decode<T: Decodable>(_ type: T.Type, _ json: String) throws -> T {
+    try JSONDecoder().decode(type, from: Data(json.utf8))
 }
diff --git a/bindings/swift/Tests/AimuxTests/WrapperTests.swift b/bindings/swift/Tests/AimuxTests/WrapperTests.swift
index 428579bbf..389f3cdc4 100644
--- a/bindings/swift/Tests/AimuxTests/WrapperTests.swift
+++ b/bindings/swift/Tests/AimuxTests/WrapperTests.swift
@@ -1,7 +1,7 @@
 // WrapperTests.swift — tests for the typed Codable wrapper layer (Types.swift).
 //
 // Exercises the `Model` extension methods that take `ModelPrompt` /
-// `GenerateTextOptions` and return `GenerateTextResult` / `StreamPart`,
+// `GenerateTextOptions` and return `GenerateTextResult` / `TextStreamPart`,
 // verifying the JSON↔Codable boundary is handled correctly end-to-end
 // (Swift typed API → FFI → reqwest → MockHTTPServer). The raw string-based
 // API is left untouched; these tests use only the typed layer.
@@ -21,9 +21,7 @@ final class WrapperTests: XCTestCase {
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let result = try model.generateText(prompt: .text("What is Rust?"))
 
         XCTAssertEqual(result.text, "Rust is a systems programming language.")
@@ -47,12 +45,12 @@ final class WrapperTests: XCTestCase {
             toolCallId: "call_1",
             toolName: "get_weather",
             input: jv(#"{"location":"Tokyo"}"#),
-            providerMetadata: jv(#"{"openai":{"item_id":"item_1"}}"#)
+            providerMetadata: ["openai": ["itemId": .string("item_1")]]
         )
 
         let encoded = try JSONEncoder().encode(original)
         let wire = try JSONSerialization.jsonObject(with: encoded) as? [String: Any]
-        let metadata = wire?["provider_metadata"] as? [String: Any]
+        let metadata = wire?["providerMetadata"] as? [String: Any]
         XCTAssertNotNil(metadata?["openai"])
 
         let decoded = try JSONDecoder().decode(ToolCall.self, from: encoded)
@@ -67,9 +65,7 @@ final class WrapperTests: XCTestCase {
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let options = GenerateTextOptions(tools: [
             .function(FunctionTool(
                 name: "get_weather",
@@ -87,11 +83,9 @@ final class WrapperTests: XCTestCase {
         XCTAssertEqual(result.toolCalls[0].toolCallId, "call_abc")
         XCTAssertEqual(result.toolCalls[0].input["location"]?.stringValue, "Tokyo")
 
-        // Structured raw.content contains a ToolCall variant mirroring the call.
-        let toolContents = result.raw.content.compactMap { part -> ToolCall? in
-            if case .toolCall(let id, let name, let input, _, _, _) = part {
-                return ToolCall(toolCallId: id, toolName: name, input: input)
-            }
+        // Structured raw.content contains a tool-call item mirroring the call.
+        let toolContents = result.raw.content.compactMap { part -> RawToolCall? in
+            if case .toolCall(let call) = part { return call }
             return nil
         }
         XCTAssertEqual(toolContents.count, 1)
@@ -99,7 +93,7 @@ final class WrapperTests: XCTestCase {
         XCTAssertEqual(toolContents[0].toolCallId, "call_abc")
         // Raw content keeps the provider's argument text; the parsed object
         // lives on the top-level toolCalls (asserted above).
-        XCTAssertEqual(toolContents[0].input.stringValue, "{\"location\":\"Tokyo\"}")
+        XCTAssertEqual(toolContents[0].input, "{\"location\":\"Tokyo\"}")
     }
 
     // MARK: - generateText: multi-role messages
@@ -111,9 +105,7 @@ final class WrapperTests: XCTestCase {
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let result = try model.generateText(prompt: .messages([
             .system("You are a helpful assistant."),
             .user("What is Rust?"),
@@ -139,9 +131,7 @@ final class WrapperTests: XCTestCase {
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let options = GenerateTextOptions(
             tools: [.function(FunctionTool(
                 name: "get_weather",
@@ -181,19 +171,17 @@ final class WrapperTests: XCTestCase {
         XCTAssertEqual(try JSONDecoder().decode(ToolChoice.self, from: data), tool)
     }
 
-    // MARK: - streamText: typed StreamPart
+    // MARK: - streamText: typed TextStreamPart
 
-    /// Streaming text deltas are delivered as typed `StreamPart.textDelta`
+    /// Streaming text deltas are delivered as typed `TextStreamPart.textDelta`
     /// values, reassembling into the full text.
     func testTypedStreamTextYieldsStreamParts() throws {
         let server = MockHTTPServer(response: .sse(sse(openaiStreamTextEvents)))
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
-        var parts: [StreamPart] = []
+        let model = try chatModel(server.baseURL)
+        var parts: [TextStreamPart] = []
         var streamErr: (any Error)?
         model.streamText(
             prompt: .text("Say hello"),
@@ -210,27 +198,25 @@ final class WrapperTests: XCTestCase {
         }
         XCTAssertEqual(text, "Hello world")
 
-        // Every emitted part decoded into a concrete StreamPart variant.
+        // Every emitted part decoded into a concrete TextStreamPart variant.
         XCTAssertTrue(parts.contains { if case .textDelta = $0 { true } else { false } })
     }
 
     /// Streaming tool-call fragments are delivered as typed tool-related
-    /// `StreamPart`s, including a final `.toolCall` carrying the tool name.
+    /// `TextStreamPart`s, including a final `.toolCall` carrying the tool name.
     func testTypedStreamTextToolCall() throws {
         let server = MockHTTPServer(response: .sse(sse(openaiStreamToolEvents)))
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let options = GenerateTextOptions(tools: [
             .function(FunctionTool(
                 name: "get_weather",
                 inputSchema: jv(#"{"type":"object","properties":{"location":{"type":"string"}}}"#)
             ))
         ])
-        var parts: [StreamPart] = []
+        var parts: [TextStreamPart] = []
         var streamErr: (any Error)?
         model.streamText(
             prompt: .text("What's the weather?"),
@@ -250,11 +236,11 @@ final class WrapperTests: XCTestCase {
             default: return false
             }
         }
-        XCTAssertTrue(hasToolPart, "stream should contain a tool-related StreamPart")
+        XCTAssertTrue(hasToolPart, "stream should contain a tool-related TextStreamPart")
 
         // The complete ToolCall part carries the tool name and structured input.
         let toolCall = parts.compactMap { part -> (String, JSONValue)? in
-            if case .toolCall(_, let name, let input, _, _, _, _, _) = part { return (name, input) }
+            if case .toolCall(let call) = part { return (call.toolName, call.input) }
             return nil
         }.first
         XCTAssertEqual(toolCall?.0, "get_weather")
@@ -267,9 +253,7 @@ final class WrapperTests: XCTestCase {
         try server.start()
         defer { server.stop() }
 
-        let model = try Model.openai(
-            apiKey: "test-key", modelId: "gpt-4o", baseUrl: server.baseURL
-        )
+        let model = try chatModel(server.baseURL)
         let parts = try await collect(model.streamTextAsync(prompt: .text("Say hello")))
         XCTAssertFalse(parts.isEmpty)
         let text = parts.reduce(into: "") { acc, part in
@@ -286,33 +270,42 @@ final class WrapperTests: XCTestCase {
         let json = """
         {
           "text": "hello",
-          "tool_calls": [
-            {"tool_call_id":"call_1","tool_name":"get_weather","input":{"location":"Tokyo"}}
+          "toolCalls": [
+            {"toolCallId":"call_1","toolName":"get_weather","input":{"location":"Tokyo"}}
           ],
-          "finish_reason": {"unified":"stop","raw":"stop"},
-          "usage": {"input_tokens":{"total":3},"output_tokens":{"total":2}},
-          "warnings": [],
-          "reasoning": [],
-          "reasoning_text": "",
-          "sources": [],
+          "finishReason": {"unified":"stop","raw":"stop"},
+          "usage": {"inputTokens":{"total":3},"outputTokens":{"total":2}},
+          "warnings": [{"type":"other","message":"careful"}],
+          "reasoning": [{"type":"reasoning","text":"hmm"}],
+          "reasoningText": "hmm",
+          "sources": [{"sourceType":"url","id":"s1","url":"https://example.com"}],
           "files": [],
-          "response_messages": [],
-          "raw_finish_reason": "stop",
-          "provider_metadata": null,
-          "response": {"id":"resp_1","timestamp":null,"model_id":"gpt-4o"},
-          "total_usage": {"input_tokens":{"total":3},"output_tokens":{"total":2}},
+          "responseMessages": [
+            {"role":"assistant","content":[
+              {"type":"tool-call","toolCallId":"call_1","toolName":"get_weather","input":{"location":"Tokyo"}}
+            ]},
+            {"role":"tool","content":[
+              {"type":"tool-result","toolCallId":"call_1","toolName":"get_weather",
+               "output":{"type":"json","value":{"temp":22}}}
+            ]}
+          ],
+          "rawFinishReason": "stop",
+          "request": {},
+          "response": {"id":"resp_1","modelId":"gpt-4o"},
+          "totalUsage": {"inputTokens":{"total":3},"outputTokens":{"total":2}},
+          "content": [
+            {"type":"text","text":"hello"},
+            {"type":"tool-call","toolCallId":"call_1","toolName":"get_weather","input":{"location":"Tokyo"}}
+          ],
           "raw": {
             "content": [
-              {"Text":{"text":"hello"}},
-              {"ToolCall":{"tool_call_id":"call_1","tool_name":"get_weather","input":{"location":"Tokyo"}}}
+              {"type":"text","text":"hello"},
+              {"type":"tool-call","toolCallId":"call_1","toolName":"get_weather","input":"{\\"location\\":\\"Tokyo\\"}"}
             ],
-            "finish_reason": {"unified":"tool-calls","raw":"tool_calls"},
-            "usage": {"input_tokens":{"total":3},"output_tokens":{"total":2}},
+            "finishReason": {"unified":"tool-calls","raw":"tool_calls"},
+            "usage": {"inputTokens":{"total":3},"outputTokens":{"total":2}},
             "warnings": [],
-            "provider_metadata": null,
-            "response": {"id":"resp_1","timestamp":null,"model_id":"gpt-4o"},
-            "request_body": null,
-            "response_headers": null
+            "response": {"id":"resp_1","modelId":"gpt-4o"}
           }
         }
         """
@@ -321,9 +314,16 @@ final class WrapperTests: XCTestCase {
         XCTAssertEqual(original.text, "hello")
         XCTAssertEqual(original.toolCalls[0].toolName, "get_weather")
         XCTAssertEqual(original.finishReason.unified, .stop)
-        XCTAssertEqual(original.raw.finishReason?.unified, .toolCalls)
+        XCTAssertEqual(original.raw.finishReason.unified, .toolCalls)
         XCTAssertEqual(original.raw.response?.id, "resp_1")
         XCTAssertEqual(original.raw.response?.modelId, "gpt-4o")
+        XCTAssertEqual(original.warnings, [.other(message: "careful")])
+        XCTAssertEqual(original.reasoning, [.reasoning(ReasoningOutput(text: "hmm"))])
+        guard case .parts(let toolContent) = original.responseMessages[1].content,
+              case .toolResult(_, _, let output, _) = toolContent[0] else {
+            return XCTFail("expected a tool-result part, got \(original.responseMessages[1].content)")
+        }
+        XCTAssertEqual(output, .json(.object(["temp": .number(22)])))
 
         // Encode → decode yields an equal value.
         let reencoded = try JSONEncoder().encode(original)
@@ -334,10 +334,10 @@ final class WrapperTests: XCTestCase {
 
 // MARK: - Helpers
 
-/// Drain an `AsyncThrowingStream<StreamPart, Error>` into an array (the mock stream completes
+/// Drain an `AsyncThrowingStream<TextStreamPart, Error>` into an array (the mock stream completes
 /// synchronously inside `streamText`, so this returns promptly).
-private func collect(_ stream: AsyncThrowingStream<StreamPart, Error>) async throws -> [StreamPart] {
-    var parts: [StreamPart] = []
+private func collect(_ stream: AsyncThrowingStream<TextStreamPart, Error>) async throws -> [TextStreamPart] {
+    var parts: [TextStreamPart] = []
     for try await part in stream { parts.append(part) }
     return parts
 }
diff --git a/contract-tests/fixtures/decision-openai-full.json b/contract-tests/fixtures/decision-openai-full.json
index a8c110b57..beebd788c 100644
--- a/contract-tests/fixtures/decision-openai-full.json
+++ b/contract-tests/fixtures/decision-openai-full.json
@@ -6,9 +6,7 @@
     "images": [
       {
         "media_type": "image/png",
-        "data": {
-          "Base64": "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mP8/x8AAwMCAO+aZ1sAAAAASUVORK5CYII="
-        },
+        "data": "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mP8/x8AAwMCAO+aZ1sAAAAASUVORK5CYII=",
         "detail": "low"
       }
     ],
diff --git a/docs/api/decision.md b/docs/api/decision.md
index 170420d3f..17a875b83 100644
--- a/docs/api/decision.md
+++ b/docs/api/decision.md
@@ -37,7 +37,7 @@ Choice 支持 2–255 项，`label` 为规范化结果的 map key，可选 `valu
 ```ts
 const result = await decide(model, {
   state: 'Rate the picture',
-  images: [{ data: { Base64: imageBase64 }, media_type: 'image/png', detail: 'low' }],
+  images: [{ data: imageBase64, media_type: 'image/png', detail: 'low' }],
   questions: [
     { id: 'choice', type: 'choice', instructions: 'Pick a value', options: [
       { label: 'boolean_true', value: true },
@@ -52,8 +52,8 @@ const result = await decide(model, {
 // Choice: selected='boolean_true', value=true；两个原生值使用不同概率 map key。
 ```
 
-最多 128 张内嵌图，detail 支持 low/high/auto/original。也可使用
-`data: { Binary: [137, 80, ...] }`，与已有 FileBytes contract 一致。
+最多 128 张内嵌图，detail 支持 low/high/auto/original。`data` 也可以是字节数组
+`[137, 80, ...]`，与公共 JSON 的 FileBytes 形状一致（base64 字符串或字节数组）。
 对象/数组 instructions 和 Boolean criteria 返回 Unsupported。
 `openai` options 仅接受可选 safety_identifier（最多 128 字符）。
 原生概率/confidence 保留原值，不套用 Jev 两位小数规则；usage 缺失或合计不一致
@@ -322,10 +322,10 @@ finally:
 ```
 
 标准 `recordings.jsonl` 每行记录一次逻辑调用：`input.operation = "decision"`，
-`input.options` 保留 state/questions/images 以及协议、编码 options，`input.decision_capabilities` 保留能力声明；
-provider 只记录 provider_id/provider/model_id，endpoint 与凭证来自调用方配置；
+`input.options` 保留 state/questions/images 以及协议、编码 options，`input.decisionCapabilities` 保留能力声明；
+provider 只记录 providerId/provider/modelId，endpoint 与凭证来自调用方配置；
 exchanges 记录各重试 attempt 的 HTTP 请求、响应和时延；
-`outcome.decision_result` 保存规范化结果。凭据沿用统一脱敏规则。
+`outcome.decisionResult` 保存规范化结果。凭据沿用统一脱敏规则。
 录制使用 schema 3；语言录制缺少 operation 时仍按 language model 解析。
 
 CLI 离线回放不需要 key：
diff --git a/docs/api/flutter.md b/docs/api/flutter.md
index dfcb56b90..4a4c64bf2 100644
--- a/docs/api/flutter.md
+++ b/docs/api/flutter.md
@@ -58,8 +58,11 @@ final model = Model.provider('groq', 'llama-3.3-70b');
 final result = model.generateText('Hello');
 model.close();
 
-// Optional config JSON ({"base_url": "..."}):
-final model2 = Model.provider('groq', 'llama-3.3-70b', apiKey: 'sk-...');
+// Optional settings JSON; keys are camelCase ({"baseUrl": "...", "headers": {...},
+// "organization", "project", "params"}) and an unknown key is an
+// InvalidArgumentError:
+final model2 = Model.provider('groq', 'llama-3.3-70b',
+    apiKey: 'sk-...', configJson: '{"baseUrl": "http://localhost:3000"}');
 model2.close();
 ```
 
@@ -115,15 +118,13 @@ Every instance has `message`, `code` (`AimuxErrorCode` constants matching C
 subclass only: `APICallError.providerCode` / `.providerMessage` / `.responseBody`
 (`String?`), `NoSuchModelError.modelId` / `.modelType`, and
 `NoSuchProviderError.providerId`. A code outside the enum is a header/library
-mismatch and fails with `StateError`, not an error type.
-subclass only: `APICallError.providerCode` / `.providerMessage` / `.requestId` / `.responseBody`
-(`String?`), `NoSuchModelError.modelId` / `.modelType`,
-`NoSuchProviderError.providerId`, `NoSuchToolError.toolName` /
+mismatch and fails with `StateError`, not an error type. Further payload:
+`NoSuchToolError.toolName` /
 `.availableTools` (`List<String>?`, null when no tool set was supplied),
 `InvalidToolInputError.toolName` / `.toolInput` (the raw argument text), and
 `ToolCallRepairError.originalError` (the repaired-over error decoded from its
-wire JSON — the same shape as `ToolCall.error`). A code outside the enum is a
-header/library mismatch and fails with `StateError`, not an error type.
+wire JSON: an `AiMuxError` object keyed by `name`, the same shape as
+`ToolCall.error`).
 
 `APICallError` additionally exposes the sanitized URL/request values,
 response headers/body, parsed provider data/code, and retryability.
@@ -151,7 +152,7 @@ errors, never an aimux type. Raw JSON string parameters (`configJson`,
 `optsJson`, `valuesJson`, …) are validated with `jsonDecode` in Dart *before*
 the C call. Empty/blank is rejected for required raw-JSON params (`valuesJson`,
 the `optsJson` of speech/image/video/rerank/search, and the `configJson` of
-`registerProviders` / `initProxy`) and treated as "defaults" for optional ones
+`initProxy`) and treated as "defaults" for optional ones
 (nullable in the signature — provider/router/moa `configJson`,
 embed/upload/transcription `optsJson`), matching the C ABI; use-after-close is
 guarded locally. A returned C code in 200..206 is a binding/library invariant:
@@ -204,7 +205,7 @@ models, `TranscriptionSession`, `Files`) throws `StateError`.
 Stream terminal failures surface via `Stream.addError` with whatever the
 decoder produced (`AimuxException`, or the native `StateError` for a C ABI
 failure) and the stream is then closed — there is no `on_error` callback;
-provider mid-stream `StreamPart::Error` is data on `on_part`. A
+a provider mid-stream failure is an `error` part, data on `on_part`. A
 `TranscriptionSession.nextPart` timeout is a poll state, not an error:
 `AimuxTranscriptionTimeoutException` (session still live); a normal end is
 `AimuxTranscriptionEndedException`.
@@ -227,7 +228,7 @@ model.close();
 final model = Model.openai('sk-...', 'gpt-4o');
 final stream = model.streamText('Write a haiku');
 await for (final part in stream) {
-  if (part.containsKey('TextDelta')) print(part['TextDelta']['delta']);
+  if (part['type'] == 'text-delta') print(part['delta']);
 }
 model.close();
 ```
@@ -288,7 +289,7 @@ instance). Throwing means the repair failed: the call stays invalid and carries
 a `ToolCallRepairError` whose cause is the thrown object's `toString()`. A
 replacement that still fails schema validation does the same.
 
-`streamText` repairs the `ToolCall` part in flight — tool-input deltas are
+`streamText` repairs the `tool-call` part in flight — tool-input deltas are
 forwarded verbatim and in order — and is the only entry point that awaits an
 `async` hook; the synchronous ones reject a `Future` with a `StateError`.
 `generateTextAsOpenAI` repairs too: `ChatCompletion` carries no invalid
@@ -297,23 +298,66 @@ marker, so it repairs the native result and converts it
 repair — the chunk stream forwards the provider's argument deltas verbatim.
 
 `RawToolCall` carries every provider field (`providerExecuted`,
-`thoughtSignature`, `providerMetadata`), and the core adopts a replacement as
+`dynamic`, `providerMetadata`), and the core adopts a replacement as
 returned — build it with `context.toolCall.copyWith(input: …)` so those
 fields survive the repair.
 
+## Wire format
+
+The JSON that crosses the C ABI is the AI SDK's JSON, and `types.dart` reads
+and writes exactly the shapes of the generated TypeScript types
+(`bindings/node/src/types/*.ts`):
+
+- field names are camelCase (`toolCalls`, `finishReason`, `providerOptions`,
+  `maxOutputTokens`, ...); `config_json` / router / moa settings follow suit
+  (`baseUrl`, `providerName`, ...);
+- unions carry a `type` tag with kebab-case values (`text-delta`, `tool-call`,
+  `file-base64`, ...); sources use `sourceType`; errors use `name`
+  (`{"name": "AI_APICallError", ...}`);
+- optional fields are absent, never `null`;
+- `FileBytes`, audio and image outputs are untagged (a byte array or a base64
+  string);
+- image `size` is a `"WxH"` string and `aspectRatio` a `"W:H"` string.
+
+A tool result in a message is a `ContentPartToolResult(toolCallId, toolName,
+output)`; `output` is a `ToolResultOutput` (`text`, `json`, `executionDenied`,
+`errorText`, `errorJson` or `content`):
+
+```dart
+final followUp = [
+  ModelMessage(role: 'tool', content: [
+    ContentPartToolResult(
+      toolCallId: call.toolCallId,
+      toolName: call.toolName,
+      output: ToolResultOutput.json({'temperature': 22}),
+    ).toJson(),
+  ]),
+];
+```
+
+There is no provider-registration API: `aimux_register_providers` is gone, and
+so is `registerProviders`. Use `Model.provider` / `createProvider` with a
+`baseUrl` to reach a custom endpoint.
+
 ## Types
 
 `bindings/flutter/lib/types.dart` declares the typed model surface (with
 `toJson` / `fromJson` on each): `Role`, `FinishReasonUnified`,
-`ReasoningEffort`, `TokenUsage`, `Usage`, `FinishReason`, `ToolCall`,
-`FunctionTool`, `Tool`, `ToolChoice`, `ResponseMetadata`, `GenerateContent`
-(sealed), `GenerateResult`, `GenerateTextResult`, `GenerateTextOptions`,
-`ModelMessage`, `StreamPart` (sealed), `FileBytes`, `FileData`, `ContentPart`,
-`RawToolCall` / `ToolCallRepairContext` / `RepairToolCall` (tool-call repair).
+`ReasoningEffort`, `InputTokenUsage` / `OutputTokenUsage` / `Usage`,
+`FinishReason`, `ToolCall`, `FunctionTool`, `Tool`, `ToolChoice`,
+`RequestInfo` / `ResponseInfo` / `ResponseMetadata`, `GeneratedFile`,
+`Source`, `ToolResult`, `GenerateContent` (sealed), `GenerateResult`,
+`GenerateTextResult`, `GenerateObjectResult`, `StreamTextResultAggregated`,
+`GenerateTextOptions`, `ModelMessage`, `StreamPart` (sealed; the
+`TextStreamPart` shapes the stream emits), `FileBytes`, `FileData`,
+`ContentPart` (sealed), `ToolResultOutput`, and `RawToolCall` /
+`ToolCallRepairContext` / `RepairToolCall` (tool-call repair). A `type` the
+binding does not model decodes to the `...Unknown` fallback of each union and
+re-encodes verbatim.
 
 `ToolCall` carries `providerMetadata` plus `invalid` (set by Core when tool
 lookup, input parse, or schema validation fails, even after repair) and `error`
-(the serialized `AiMuxError` for that failure).
+(the `AiMuxError` object for that failure).
 
 ## Coverage
 
diff --git a/docs/api/go.md b/docs/api/go.md
index 620b04fa4..15fc8c052 100644
--- a/docs/api/go.md
+++ b/docs/api/go.md
@@ -94,7 +94,7 @@ no Go type of their own; the binding maps them to native Go errors:
 |---------|----|
 | bad raw JSON in `promptJson` / `optsJson` / `configJSON` | plain `error` naming the parameter, e.g. `aimux: prompt_json: invalid JSON` — checked in Go before the C call with `json.Valid` **plus** a surrogate-pairing scan, because `json.Valid` accepts unpaired `\uD800`–`\uDFFF` escapes that serde_json rejects (required parameters reject `""`; optional `""` = default; JSONL by line) |
 | a string parameter that is not valid UTF-8, or contains a NUL | plain `error`: `aimux: <param>: must be valid UTF-8` / `aimux: <param>: must not contain NUL` — a Go string is an arbitrary byte sequence and `C.CString` passes it through verbatim (a NUL would truncate the argument silently), so every user-supplied string is checked before the C call. Two exceptions, both deliberate: the five `mustNew` constructors panic instead (see below), and `InitLogging` has no error channel and falls back to `"warn"` |
-| a typed option struct (`ProviderConfig`, `EmbeddingCallOptions`, `SpeechCallOptions`, … , `TranscriptionSessionOpts`) whose raw JSON field carries bad bytes | plain `error` naming the marshalled parameter, e.g. `aimux: opts: invalid JSON: unpaired high surrogate \uD800`, `aimux: config_json: must be valid UTF-8` — `json.Marshal` is **not** self-validating: it coerces invalid UTF-8 in a Go *string* field to U+FFFD and escapes NUL, but a `json.RawMessage` field (`ProviderOptions`, `Audio`, `Documents`, `Files`, `Mask`, `Data`, `InputSchema`) is emitted through `compact()`, which checks JSON *syntax only*. Raw non-UTF-8 bytes and lone `\uD800`–`\uDFFF` escapes therefore survive marshalling, so every marshalled C argument gets the same `checkJSON` as a raw-string parameter |
+| a typed option struct (`ProviderConfig`, `EmbeddingCallOptions`, `SpeechCallOptions`, … , `TranscriptionSessionOpts`) whose raw JSON field carries bad bytes | plain `error` naming the marshalled parameter, e.g. `aimux: opts: invalid JSON: unpaired high surrogate \uD800`, `aimux: config_json: must be valid UTF-8` — `json.Marshal` is **not** self-validating: it coerces invalid UTF-8 in a Go *string* field to U+FFFD and escapes NUL, but a `json.RawMessage` field (`ProviderOptions`, `Documents`, `Files`, `Mask`, `Data`, `InputSchema`) is emitted through `compact()`, which checks JSON *syntax only*. Raw non-UTF-8 bytes and lone `\uD800`–`\uDFFF` escapes therefore survive marshalling, so every marshalled C argument gets the same `checkJSON` as a raw-string parameter |
 | use-after-close (`Model`, `ProviderHandle`, multimodal models, `TranscriptionSession`) | `errors.Is(err, aimux.ErrClosed)` — guarded in Go before the C call |
 | a `nil` `*Model` handed to a composite constructor — a `NewRouter` child, a `NewMoa` reference, or the `NewMoa` aggregator | plain `error` naming the position: `aimux: router: models[2] is nil`, `aimux: moa: references[1] is nil`, `aimux: moa: aggregator is nil` — checked in Go before the C call, so a nil element is a returned error rather than the nil-pointer dereference that would otherwise take the process down |
 | trace query (`TraceAggregate`, `TraceSessionChain`, `TraceExportJsonl`, `TraceClear`) on a model that never went through `Trace` / `TraceAudited` | `errors.Is(err, aimux.ErrNotTraced)` — guarded in Go before the C call, ahead of argument validation; the trace store is keyed on the wrapper handle, so C can only report it as a missing handle |
@@ -143,7 +143,7 @@ fmt.Println(result)
 // streaming (typed: model.Stream(prompt, opts) yields *StreamPart values)
 stream := model.StreamText(`"Write a haiku"`, "")
 for part := range stream.Parts() {
-    fmt.Println(part) // StreamPart JSON
+    fmt.Println(part) // StreamPart JSON, e.g. {"type":"text-delta","id":"0","delta":"..."}
 }
 if err := stream.Err(); err != nil { // only after Parts has closed
     log.Fatal(err)
@@ -172,10 +172,10 @@ defer model.Close()
 model2, err := aimux.ProviderWithBase("groq", "sk-...", "llama-3.3-70b", "https://relay.example/v1")
 defer model2.Close()
 
-// Full ProviderOptions (base_url / headers / organization / project / params).
-// Retry is a per-call option (CallOptions.MaxRetries) and request-body overrides
-// no longer exist: a config_json carrying max_retries or body_overrides is
-// rejected as CodeInvalidArgument, so ProviderConfig has neither.
+// Full settings (baseUrl / headers / organization / project / params).
+// Retry is a per-call option (GenerateTextOptions.MaxRetries) and request-body
+// overrides no longer exist: any other key in config_json is rejected as
+// CodeInvalidArgument, so ProviderConfig has neither.
 model3, err := aimux.ProviderWithConfig("groq", "sk-...", "llama-3.3-70b", &aimux.ProviderConfig{
 	Headers: map[string]string{"X-Custom": "1"},
 })
@@ -226,6 +226,20 @@ if err := stream.Err(); err != nil {
 }
 ```
 
+Each part is an object tagged by `type` (`text-delta`, `tool-call`, `finish`,
+...). The typed `model.Stream(prompt, opts)` yields `*StreamPart{Type, Raw}`
+instead; decode `Raw` (the whole part) into the struct you need:
+
+```go
+for part := range typed.Parts() {
+    if part.Type == "text-delta" {
+        var td aimux.TextDeltaPayload
+        json.Unmarshal(part.Raw, &td)
+        fmt.Print(td.Delta)
+    }
+}
+```
+
 > Stream part variants are documented in the [API overview](../API.md#streaming-generation).
 > Drain `Parts()` before calling `Err()`. If you stop reading early, call
 > `Cancel()` so the native stream does not keep running.
@@ -270,7 +284,7 @@ itself call aimux — e.g. ask a model to rewrite the arguments. Each call is
 repaired at most once.
 
 It applies to the typed entry points `Generate`, `GenerateObj`,
-`ConsumeStream` and `Stream` (which delivers the repaired `ToolCall` part;
+`ConsumeStream` and `Stream` (which delivers the repaired `tool-call` part;
 input deltas are always forwarded verbatim), and to `GenerateAsOpenAI`: a
 `ChatCompletion` carries no invalid marker, so it repairs the native result
 and converts it (`Model.GenerateTextResultAsOpenAI`). `StreamAsOpenAI` does
@@ -329,7 +343,7 @@ result, err := aimux.ParseSpeechResult(resultJSON)
 if err != nil {
     log.Fatal(err)
 }
-// audio bytes: *result.Audio.Base64 (base64 string) or result.Audio.Binary
+// audio bytes: []byte(result.Audio) (base64 string or byte array on the wire)
 ```
 
 ## Speech to Text (STT)
@@ -361,7 +375,6 @@ fmt.Println(*result.Language)         // detected language
 
 ```go
 prompt := "A cute baby sea otter"
-n := 1
 
 imager, err := aimux.NewOpenAIImage("sk-...", "dall-e-3")
 if err != nil {
@@ -371,7 +384,7 @@ defer imager.Close()
 
 resultJSON, err := imager.Generate(&aimux.ImageCallOptions{
     Prompt: &prompt,
-    N:      &n,
+    N:      1, // required
 })
 if err != nil {
     log.Fatal(err)
@@ -380,7 +393,7 @@ result, err := aimux.ParseImageResult(resultJSON)
 if err != nil {
     log.Fatal(err)
 }
-// result.Images.Base64[0] (base64) or result.Images.Binary[0] (raw bytes)
+// []byte(result.Images[0]) — the image bytes
 ```
 
 ## Video Generation
@@ -414,7 +427,7 @@ result, err := aimux.ParseVideoResult(resultJSON)
 if err != nil {
     log.Fatal(err)
 }
-// result.Videos[0].Url.URL — video URL
+// result.Videos[0].URL — video URL (Type "url"; "base64"/"binary" carry Data)
 ```
 
 ## Reranking
@@ -432,7 +445,7 @@ defer reranker.Close()
 
 resultJSON, err := reranker.Rerank(&aimux.RerankingCallOptions{
     Query:     "What is Rust?",
-    Documents: json.RawMessage(`[{"text":"Rust is a systems programming language."},{"text":"Rust is a chemical element."}]`),
+    Documents: json.RawMessage(`{"type":"text","values":["Rust is a systems programming language.","Rust is a chemical element."]}`),
     TopN:      &topN,
 })
 if err != nil {
@@ -578,6 +591,37 @@ Typed structs live in `bindings/go/types.go` (text) and
 The multimodal methods return JSON strings through the C ABI; the
 `ParseXxxResult` functions decode them into the typed structs. All call-option
 pointer fields (`*string`, `*bool`, `*int`) are optional — pass `nil` to omit.
+Image `Size` and video `Resolution` are `"WxH"` strings (`"1024x1024"`);
+`AspectRatio` is `"W:H"` (`"16:9"`). Binary data (`Audio`, `Images`, `VideoData.Data`) is `aimux.Bytes`: a
+base64 string or a byte array on the wire, always `[]byte` in Go.
+
+### Wire format
+
+The JSON the binding reads and writes is the AI SDK JSON, the same as the
+generated TypeScript types in `bindings/node/src/types/`: camelCase field names
+(`toolCallId`, `finishReason`, `inputTokens`, `providerOptions`, ...), unions
+tagged by a `type` key (`"tool-call"`, `"text-delta"`, ...; errors by `name`),
+and optional fields absent rather than `null`. Go field tags follow it, so a
+typed struct marshals to exactly that JSON.
+
+A tool result goes back to the model as a `tool-result` content part in a
+`tool` message; its `output` is a union:
+
+```go
+msg := aimux.ModelMessage{
+    Role: aimux.RoleTool,
+    Content: []map[string]any{{
+        "type":       "tool-result",
+        "toolCallId": call.ToolCallID,
+        "toolName":   call.ToolName,
+        // {"type":"text"|"error-text","value":string},
+        // {"type":"json"|"error-json","value":any},
+        // {"type":"execution-denied","reason":string},
+        // {"type":"content","value":[...]}
+        "output": map[string]any{"type": "json", "value": map[string]any{"temp": "20C"}},
+    }},
+}
+```
 
 `ToolCall` carries `ProviderMetadata` plus `Invalid` (set by Core when tool
 lookup, input parse, or schema validation fails, even after repair) and `Error`
diff --git a/docs/api/java.md b/docs/api/java.md
index 6dc3a810f..f7ef812d2 100644
--- a/docs/api/java.md
+++ b/docs/api/java.md
@@ -46,7 +46,7 @@ try (Model model = Model.providerFromEnv("groq", "llama-3.3-70b")) {
     String result = model.generateText("\"Hello\"");
 }
 
-// Optional config JSON ({"base_url": "..."}):
+// Optional config JSON ({"baseUrl": "..."}); an unknown key is InvalidArgumentError:
 try (Model model = Model.provider("groq", "sk-...", "llama-3.3-70b", null)) {
     String result = model.generateText("\"Hello\"");
 }
@@ -54,6 +54,9 @@ try (Model model = Model.provider("groq", "sk-...", "llama-3.3-70b", null)) {
 
 `deepseek(apiKey, modelId)` remains as a by-name shortcut to the vendor package.
 Unknown names throw `NoSuchProviderError` naming the requested provider.
+There is no provider-registration API: the provider set is the built-in
+registry, and a custom endpoint is a `baseUrl` in the `configJson` above (or a
+`...WithBase` factory).
 
 ## Errors
 
@@ -76,6 +79,7 @@ RuntimeException
       ├── ToolCallRepairError        // code 17: tool-call repair itself failed
       ├── LoadAPIKeyError            // code 18: no API key, fallback env var unset (getEnvVar())
       ├── LoadSettingError           // code 19: a required setting is missing (getEnvVar())
+      ├── NoOutputGeneratedError     // code 20: the call produced no output
       └── OtherError
 ```
 
@@ -84,7 +88,7 @@ Every instance has:
 | Field | Meaning |
 |-------|---------|
 | `getMessage()` | human-readable text from C |
-| `getCode()` | `aimux_error_code_t` value 1–19 (4 retired, 14 = `Retry`; matches `aimux-error.h`) |
+| `getCode()` | `aimux_error_code_t` value 1–20 (4 retired, 14 = `Retry`; matches `aimux-error.h`) |
 | `getStatusCode()` | HTTP status, or `-1` |
 | `getRetryMs()` | rate-limit hint, or `-1` (`0` = retry now) |
 | `isRetryable()` | the `AiMuxError` retry verdict (not derivable from status) |
@@ -343,7 +347,7 @@ try (SpeechModel model = SpeechModel.openai("sk-...", "tts-1")) {
     String opts = new JSONObject()
         .put("text", "Hello")
         .put("voice", "alloy")
-        .put("output_format", "mp3")
+        .put("outputFormat", "mp3")
         .toString();
     String result = model.generate(opts);
 }
@@ -388,7 +392,7 @@ Factories: `VideoModel.google` / `googleWithBase`.
 
 ```java
 try (RerankingModel model = RerankingModel.cohere("sk-...", "rerank-v3.0")) {
-    String result = model.rerank("{\"query\":\"...\",\"documents\":{...},\"top_n\":2}");
+    String result = model.rerank("{\"query\":\"...\",\"documents\":{\"type\":\"text\",\"values\":[...]},\"topN\":2}");
 }
 ```
 
@@ -398,7 +402,7 @@ Factories: `RerankingModel.cohere` / `cohereWithBase`.
 
 ```java
 try (SearchModel model = SearchModel.tavily("sk-...")) {
-    String result = model.search("{\"query\":\"What is Rust?\",\"max_results\":5}");
+    String result = model.search("{\"query\":\"What is Rust?\",\"maxResults\":5}");
 }
 // result: {"results":[{"title":"Rust",...}],"answer":"Rust is a systems language."}
 ```
@@ -411,7 +415,7 @@ Factories: `SearchModel.tavily` / `tavilyWithBase`.
 try (Files files = Files.openai("sk-...")) {
     String result = files.uploadFile(base64Data, "application/pdf");
 }
-// result: {"provider_reference":{"openai":"file-abc"}, ...}
+// result: {"providerReference":{"openai":"file-abc"}, ...}
 ```
 
 Factories: `Files.openai` / `openaiWithBase`.
@@ -426,8 +430,24 @@ Factories: `Files.openai` / `openaiWithBase`.
 custom serializer for the scalar-or-object wire form), `ContentPart` (sealed),
 `MessageContent` (sealed), `ModelMessage`, `GenerateTextOptions`,
 `FileBytes` / `FileData`, `GenerateContent`, `GenerateResult`,
-`GenerateTextResult`, `StreamPart` (sealed). All sealed hierarchies serialize
-in the wrapper-object wire form (e.g. `{"TextDelta":{...}}`).
+`GenerateTextResult`, `StreamPart` (sealed), `ToolResultOutput`. Field names
+are the AI SDK's camelCase (`toolCallId`, `providerOptions`, `finishReason`, ...)
+and optional fields are absent, never `null`. Sealed hierarchies are tagged on
+`type` with the upstream kebab-case names (e.g. `{"type":"text-delta",...}`,
+`{"type":"tool-call",...}`); `FileBytes` is untagged (an int array or a base64
+string).
+
+A `ContentPart.ToolResult` (`type: "tool-result"`) carries `toolCallId`,
+`toolName`, `output` and `providerOptions`; `output` is a `ToolResultOutput`
+built with `text` / `json` / `executionDenied` / `errorText` / `errorJson` /
+`content`. The stream-side `StreamPart.ToolResult` and
+`GenerateContent.ToolResult` are the executed-result shape and keep `result`,
+`isError`, `preliminary` and `dynamic`.
+
+`GenerateTextOptions.responseFormat` is a JSON node: `{"type":"text"}` or
+`{"type":"json","schema":{...},"name":"...","description":"..."}`.
+`ImageCallOptions` / `VideoCallOptions` take `size` / `resolution` as `"WxH"`
+strings (e.g. `"1024x1024"`) and `aspectRatio` as `"W:H"` (e.g. `"16:9"`).
 
 `RawToolCall` and `ToolCallRepairContext` serve
 [tool-call repair](#tool-call-repair); `ToolCallRepair` (the hook itself) is a
@@ -444,8 +464,9 @@ validation fails, even after repair) and `getError()` (the serialized
 `TranscriptionResult`, `RerankingCallOptions` / `RerankingResult`,
 `VideoCallOptions` / `VideoPollOptions` / `VideoResult`, `SearchCallOptions` / `SearchResult`,
 `UploadFileCallOptions` / `UploadFileResult`. Sealed unions (`AudioData`,
-`ImageOutputs`, `VideoData`) use custom Jackson serializers with the
-externally-tagged wire form. Response types (`EmbeddingResponse`, etc.) match
+`ImageOutputs`, `VideoFileData`) are untagged (a base64 string or an int array)
+and `VideoData` / `VideoFile` are tagged on `type`, all with custom Jackson
+serializers. Response types (`EmbeddingResponse`, etc.) match
 the `aimux-core` `.ts` wire definitions exactly (`{headers, body, ...}`).
 
 ## Error Handling
@@ -474,5 +495,7 @@ cargo build -p aimux-ffi --release
 cd bindings/java
 export JAVA_HOME=...   # JDK 9+ (bytecode targets Java 8 via --release 8)
 export LD_LIBRARY_PATH="$(pwd)/../../target/release:${LD_LIBRARY_PATH}"
+# or point the tests at another build (e.g. a debug build in a shared target dir):
+# export AIMUX_FFI_LIB_DIR=/path/to/target/debug
 gradle test
 ```
diff --git a/docs/api/kotlin.md b/docs/api/kotlin.md
index a4e84a4ec..2c7c7ed8d 100644
--- a/docs/api/kotlin.md
+++ b/docs/api/kotlin.md
@@ -38,12 +38,20 @@ Model.provider(name = "groq", modelId = "llama-3.3-70b").use { model ->
     val result = model.generateText("\"Hello\"")
 }
 
-// Optional config JSON ({"base_url": "..."}):
-Model.provider(name = "groq", apiKey = "sk-...", modelId = "llama-3.3-70b").use { model ->
+// Optional config JSON: baseUrl, headers, organization, project, params.
+Model.provider(
+    name = "groq", apiKey = "sk-...", modelId = "llama-3.3-70b",
+    configJson = """{"baseUrl": "https://proxy.example.com/v1"}""",
+).use { model ->
     val result = model.generateText("\"Hello\"")
 }
 ```
 
+`configJson` (also for `Model.createProvider`) rejects unknown keys as an invalid
+argument; per-call settings such as `maxRetries` belong in `GenerateTextOptions`.
+There is no `registerProviders` / overlay API: providers come from the built-in
+factories and the preset rows only.
+
 Unknown names throw `NoSuchProviderError` naming the requested provider.
 
 ## Errors
@@ -80,7 +88,7 @@ RuntimeException
       ├── TimeoutError / RequestAbortedError
       ├── NoSuchToolError            // toolName + availableTools (null = no tool set supplied)
       ├── InvalidToolInputError      // toolName + toolInput (the raw argument text)
-      ├── ToolCallRepairError        // originalError (wire JSON, same encoding as ToolCall.error)
+      ├── ToolCallRepairError        // originalError (name-tagged AiMuxError JSON, same encoding as ToolCall.error)
       ├── LoadAPIKeyError            // code 18: no API key, fallback env var unset (envVar)
       ├── LoadSettingError           // code 19: a required setting is missing (envVar)
       └── OtherError
@@ -177,7 +185,7 @@ The raw `Model` speaks JSON strings. `TypedModel` wraps it with typed objects:
 val model = TypedModel.openai("sk-...", "gpt-4o")
 val result = model.generateText("What is Rust?")
 println(result.text)          // typed GenerateTextResult
-println(result.usage?.inputTokens?.total)
+println(result.usage.inputTokens.total)
 ```
 
 | API | Signature |
@@ -286,20 +294,44 @@ session (idempotent).
 ## Types
 
 `bindings/kotlin/src/main/kotlin/ai/arcships/aimux/Types.kt` declares the
-typed model surface: `Role`, `FinishReasonUnified`, `ReasoningEffort`,
-`TokenUsage`, `Usage`, `FinishReason`, `ResponseMetadata`, `ToolCall`,
-`FunctionTool`, `ProviderTool`, `Tool` (sealed), `ToolChoice` (sealed), `ContentPart`
-(sealed), `MessageContent` (sealed), `ModelMessage`, `GenerateTextOptions`,
-`FileBytes` / `FileData` (sealed), `GenerateContent` (sealed),
-`GenerateResult`, `GenerateTextResult`, `StreamPart` (sealed).
-
-`MultimodalTypes.kt` includes `VideoCallOptions.poll: VideoPollOptions?`;
-`intervalMs` and `timeoutMs` serialize as `interval_ms` / `timeout_ms` for the
-Core-owned video status loop.
-`ToolCall` (top-level and `StreamPart.ToolCall`) carries `providerMetadata`
-plus `invalid` (set by Core when tool lookup, input parse, or schema validation
-fails, even after repair) and `error` (the serialized `AiMuxError` for that
-failure).
+typed model surface. The wire JSON is the AI SDK's and equals the generated
+TypeScript types in `bindings/node/src/types`: camelCase field names (Kotlin
+properties carry the same names), unions tagged on `type` (`Source` on
+`sourceType`), optional fields absent rather than null. Decoding is lenient
+(unknown keys ignored).
+
+| Kotlin | Wire |
+|--------|------|
+| `Usage(inputTokens: InputTokenUsage(total, noCache, cacheRead, cacheWrite), outputTokens: OutputTokenUsage(total, text, reasoning), raw)` | `inputTokens` / `noCache` / `cacheRead` / `cacheWrite` |
+| `FinishReason(unified, raw)`, `ResponseMetadata` / `ResponseInfo(id, timestamp, modelId, headers, body)` | `modelId` |
+| `ToolCall(toolCallId, toolName, input, providerExecuted, dynamic, providerMetadata, invalid, error)` | `error` is the `name`-tagged `AiMuxError` JSON |
+| `Tool` (`Function` / `Provider`), `ToolChoice` | `{"type":"function"}`, `inputSchema`, `toolChoice` |
+| `ContentPart` (sealed) | `type`: `text`, `custom`, `reasoning`, `reasoning-file`, `image`, `file`, `file-base64`, `file-url`, `file-reference`, `tool-call`, `tool-result`, `tool-approval-request` |
+| `ContentPart.ToolResult(toolCallId, toolName, output, providerOptions)` | `output` is a `ToolResultOutput`: `text`, `json` (`ToolResultOutput.JsonValue`), `execution-denied`, `error-text`, `error-json`, `content` |
+| `FileBytes` | untagged: JSON array = binary, JSON string = base64 (also audio / image / video file data) |
+| `FileData` / `GeneratedFileData` | `type`: `data`, `url`, `reference`, `text` |
+| `GenerateContent` (provider-level `raw.content`) | `type`: `text`, `tool-call` (`input` is the raw argument text), `source`, `reasoning`, `file`, `reasoning-file`, `custom`, `tool-approval-request`, `tool-result` |
+| `StreamPart` (sealed) | `type`: `text-start/delta/end`, `stream-start`, `finish`, `finish-step`, `error`, `tool-input-start/delta/end`, `tool-call`, `tool-result`, `file`, `reasoning-file`, `custom`, `tool-approval-request`, `reasoning-start/delta/end`, `source`, `raw` |
+| `Warning` (sealed) | `type`: `unsupported`, `compatibility`, `deprecated`, `other` |
+| `GenerateTextResult`, `GenerateObjectResult`, `StreamTextResultAggregated` | `toolCalls`, `finishReason`, `totalUsage`, `responseMessages`, `reasoningText`, `request`, `response` |
+| `GenerateTextOptions` | `maxOutputTokens`, `stopSequences`, `topP`, `providerOptions`, `responseFormat` (`ResponseFormat.Text` / `.Json(schema, name, description)`), `maxRetries`, `includeRawChunks`, `sessionId`, ... |
+
+Provider options / metadata are `Map<String, Map<String, JsonElement>>`
+(`ProviderOptions` / `ProviderMetadata`).
+
+`MultimodalTypes.kt` holds the other modalities with the same conventions:
+`AudioData` and `ImageOutputs` are untagged (a string is base64, an array is
+binary); `VideoData` is tagged on `type` (`url`, `base64`, `binary`);
+`Size` and `AspectRatio` serialize as the strings `"WxH"` and `"W:H"`;
+`ImageCallOptions` / `VideoCallOptions` always encode `n` and `providerOptions`
+(required on the wire); `VideoCallOptions.poll: VideoPollOptions?` carries
+`intervalMs` / `timeoutMs` for the Core-owned video status loop.
+
+## Native library in tests
+
+`build.gradle.kts` points `jna.library.path` at `../../target/release`;
+set `AIMUX_FFI_LIB_DIR` to use a library built elsewhere (for example
+`$CARGO_TARGET_DIR/debug`).
 
 ## Coverage
 
diff --git a/docs/api/node.md b/docs/api/node.md
index a9150ebc0..15a56bf08 100644
--- a/docs/api/node.md
+++ b/docs/api/node.md
@@ -44,7 +44,7 @@ The 3rd argument of every constructor is a base URL string or a
 parameters, e.g. `{ account_id: '…' }` for `cloudflare_workers_ai`).
 `maxRetries` and `bodyOverrides` are kept in the type only so that passing them
 throws `InvalidArgumentError`; retry is a per-call option
-(`max_retries` in the call options).
+(`maxRetries` in the call options).
 
 > **Scope:** `provider(name)` reaches vendor packages and the 282 preset rows;
 > typed factories remain available. Custom endpoints use the `baseUrl` override.
@@ -178,16 +178,16 @@ const { openai, generateText } = require('@arcships/aimux')
 
 const model = await openai('sk-...', 'gpt-4o', 'https://api.openai.com/v1')
 const result = await generateText(model, 'Explain Rust ownership.', {
-  max_output_tokens: 100,
+  maxOutputTokens: 100,
   temperature: 0.7,
-  max_retries: 0,                          // disable retries for this call
-  timeout: { total_ms: 30_000, first_chunk_ms: 5_000, chunk_ms: 2_000 },
+  maxRetries: 0,                           // disable retries for this call
+  timeout: { totalMs: 30_000, firstChunkMs: 5_000, chunkMs: 2_000 },
 })
 
 console.log(result.text)           // generated text
 console.log(result.usage)          // token usage
-console.log(result.finish_reason)  // finish reason
-console.log(result.tool_calls)     // tool calls (if any)
+console.log(result.finishReason)   // finish reason
+console.log(result.toolCalls)      // tool calls (if any)
 ```
 
 Cancellation via `AbortSignal` (4th argument — works for both
@@ -211,8 +211,8 @@ optional `AbortBridge` (wrap a JS `AbortSignal`) as their last argument:
 // access structured content
 const result = await generateText(model, "...", { tools })
 const rawContent = result.raw.content
-const toolCallPart = rawContent.find(c => c.ToolCall)
-const reasoningPart = rawContent.find(c => c.Reasoning)
+const toolCallPart = rawContent.find(c => c.type === 'tool-call')
+const reasoningPart = rawContent.find(c => c.type === 'reasoning')
 ```
 
 ## Streaming Generation
@@ -224,10 +224,10 @@ const { openai, streamText } = require('@arcships/aimux')
 
 const model = await openai('sk-...', 'gpt-4o')
 for await (const part of streamText(model, 'Write a haiku about Rust.')) {
-  if (part.TextDelta) {
-    process.stdout.write(part.TextDelta.delta)
+  if (part.type === 'text-delta') {
+    process.stdout.write(part.delta)
   }
-  if (part.Finish) {
+  if (part.type === 'finish') {
     console.log('\n[done]')
   }
 }
@@ -247,7 +247,7 @@ const tools = [{
   type: 'function',
   name: 'get_weather',
   description: 'Get current weather',
-  input_schema: {
+  inputSchema: {
     type: 'object',
     properties: {
       location: { type: 'string', description: 'City name' }
@@ -257,9 +257,9 @@ const tools = [{
 }]
 
 const result = await generateText(model, "What's the weather in Tokyo?", { tools })
-if (result.tool_calls.length > 0) {
-  const call = result.tool_calls[0]
-  console.log(call.tool_name)     // get_weather
+if (result.toolCalls.length > 0) {
+  const call = result.toolCalls[0]
+  console.log(call.toolName)     // get_weather
   console.log(call.input)         // { location: "Tokyo" }
 }
 ```
@@ -277,13 +277,13 @@ JavaScript, after the model call and before the result is decoded, so it can
 ```typescript
 const result = await generateText(model, "What's the weather in Tokyo?", {
   tools,
-  repairToolCall: async ({ tool_call, error, input_schema, messages }) => {
-    // tool_call.input is the model's raw argument TEXT, not a parsed object.
+  repairToolCall: async ({ toolCall, error, inputSchema, messages }) => {
+    // toolCall.input is the model's raw argument TEXT, not a parsed object.
     const fixed = await generateText(model, [
       ...messages,
-      { role: 'user', content: `Rewrite these arguments to match ${JSON.stringify(input_schema)}: ${tool_call.input}` },
+      { role: 'user', content: `Rewrite these arguments to match ${JSON.stringify(inputSchema)}: ${toolCall.input}` },
     ])
-    return { ...tool_call, input: fixed.text }   // null → leave the call invalid
+    return { ...toolCall, input: fixed.text }   // null → leave the call invalid
   },
 })
 ```
@@ -295,7 +295,7 @@ Each invalid call is repaired at most once, and a call made without a tool set
 is never repaired.
 
 Both `generateText` and `streamText` support it — in the stream, the settled
-`ToolCall` part is replaced, while `ToolInputDelta` parts are the provider's raw
+`tool-call` part is replaced, while `tool-input-delta` parts are the provider's raw
 text and pass through untouched. `generateTextAsOpenai` supports it too: the
 OpenAI shape carries no `invalid` marker, so it repairs the native result and
 converts it (`Model.generateTextResultAsOpenai`). `streamTextAsOpenai` does not
@@ -307,7 +307,7 @@ reflect repair — its argument deltas are the provider's text, as in the AI SDK
 ```typescript
 const opts = {
   tools,
-  tool_choice: 'auto'        // 'auto' | 'none' | 'required' | { type: 'tool', toolName: 'get_weather' }
+  toolChoice: 'auto'         // 'auto' | 'none' | 'required' | { type: 'tool', toolName: 'get_weather' }
 }
 ```
 
@@ -320,12 +320,12 @@ const opts = {
 const result = await generateText(model, [
   { role: 'user', content: "What's the weather in Tokyo?" },
   { role: 'assistant', content: [{
-    type: 'tool_call', tool_call_id: 'call_abc',
-    tool_name: 'get_weather', input: { location: 'Tokyo' },
+    type: 'tool-call', toolCallId: 'call_abc',
+    toolName: 'get_weather', input: { location: 'Tokyo' },
   }] },
   { role: 'tool', content: [{
-    type: 'tool_result', tool_call_id: 'call_abc', tool_name: 'get_weather',
-    result: { temperature: 22, condition: 'sunny' },
+    type: 'tool-result', toolCallId: 'call_abc', toolName: 'get_weather',
+    output: { type: 'json', value: { temperature: 22, condition: 'sunny' } },
   }] },
 ], { tools })
 ```
@@ -358,14 +358,13 @@ const speaker = await openaiSpeech('sk-...', 'tts-1')
 const resultJson = await speaker.generate(JSON.stringify({
   text: 'Hello world!',
   voice: 'alloy',
-  output_format: 'mp3',
+  outputFormat: 'mp3',
 }))
 const result = JSON.parse(resultJson)
 
-// audio is in result.audio (base64 or binary)
-if (result.audio.Base64) {
-  fs.writeFileSync('out.mp3', Buffer.from(result.audio.Base64, 'base64'))
-}
+// result.audio is a base64 string or an array of byte values
+const audio = result.audio
+fs.writeFileSync('out.mp3', typeof audio === 'string' ? Buffer.from(audio, 'base64') : Buffer.from(audio))
 ```
 
 ## Speech to Text (STT)
@@ -396,13 +395,13 @@ const imager = await openaiImage('sk-...', 'dall-e-3')
 const resultJson = await imager.generate(JSON.stringify({
   prompt: 'A cute baby sea otter',
   n: 1,
-  provider_options: {},
+  providerOptions: {},
 }))
 const result = JSON.parse(resultJson)
 
-if (result.images.Base64) {
-  fs.writeFileSync('out.png', Buffer.from(result.images.Base64[0], 'base64'))
-}
+// result.images is an array of base64 strings, or an array of byte arrays
+const image = result.images[0]
+fs.writeFileSync('out.png', typeof image === 'string' ? Buffer.from(image, 'base64') : Buffer.from(image))
 ```
 
 Multimodal calls accept an optional `AbortBridge` as their last argument —
@@ -411,7 +410,7 @@ wrap the JS `AbortSignal` in one:
 ```typescript
 const controller = new AbortController()
 const resultJson = await imager.generate(
-  JSON.stringify({ prompt: 'A cute baby sea otter', n: 1, provider_options: {} }),
+  JSON.stringify({ prompt: 'A cute baby sea otter', n: 1, providerOptions: {} }),
   new AbortBridge(controller.signal),
 )
 controller.abort() // cancels the image call
@@ -428,14 +427,14 @@ const videor = await googleVideo('sk-...', 'veo-3.0')
 const resultJson = await videor.generate(JSON.stringify({
   prompt: 'A cat playing piano',
   n: 1,
-  poll: { interval_ms: 1_000, timeout_ms: 120_000 },
-  provider_options: {},
+  poll: { intervalMs: 1_000, timeoutMs: 120_000 },
+  providerOptions: {},
 }))
 const result = JSON.parse(resultJson)
 
-// result.videos is usually [{ Url: { url, media_type } }]
-if (result.videos[0].Url) {
-  console.log('Video URL:', result.videos[0].Url.url)
+// result.videos is usually [{ type: 'url', url, mediaType }]
+if (result.videos[0].type === 'url') {
+  console.log('Video URL:', result.videos[0].url)
 }
 ```
 
@@ -452,17 +451,17 @@ const { cohereReranking } = require('@arcships/aimux/raw')
 const reranker = await cohereReranking('sk-...', 'rerank-v3.0')
 const resultJson = await reranker.rerank(
   'What is Rust?',
-  // docs_json is the externally-tagged `RerankingDocuments` enum —
-  // `{ Object: { values } }` for JSON documents, `{ Text: { values } }` for plain strings
-  JSON.stringify({ Object: { values: [
+  // docs_json is the `type`-tagged `RerankingDocuments` enum —
+  // `{ type: 'object', values }` for JSON documents, `{ type: 'text', values }` for plain strings
+  JSON.stringify({ type: 'object', values: [
     { text: 'Rust is a systems programming language.' },
     { text: 'Rust is a chemical element.' },
-  ] } }),
+  ] }),
 )
 const result = JSON.parse(resultJson)
 
-// result.ranking sorted by relevance (each rank: { index, relevance_score })
-result.ranking.forEach(r => console.log(r.index, r.relevance_score))
+// result.ranking sorted by relevance (each rank: { index, relevanceScore })
+result.ranking.forEach(r => console.log(r.index, r.relevanceScore))
 ```
 
 ## Search
@@ -491,7 +490,7 @@ const fileBase64 = fs.readFileSync('doc.pdf').toString('base64')
 const resultJson = await files.uploadFile(fileBase64, 'application/pdf')
 const result = JSON.parse(resultJson)
 
-console.log(result.provider_reference)  // { openai: 'file-xxx' }
+console.log(result.providerReference)  // { openai: 'file-xxx' }
 ```
 
 ## API Surface
@@ -541,37 +540,38 @@ import type {
 ```typescript
 // bindings/node/src/types/GenerateTextResult.ts (ts-rs generated)
 export type GenerateTextResult = {
-  text: string                            // generated text (all Text variants concatenated)
-  tool_calls: Array<ToolCall>             // tool call list (extracted from content)
-  finish_reason: FinishReason             // finish reason
+  text: string                            // generated text (all text parts concatenated)
+  toolCalls: Array<ToolCall>              // tool call list (extracted from content)
+  finishReason: FinishReason              // finish reason
   usage: Usage                            // token usage
   warnings: Array<Warning>                // warnings
   raw: GenerateResult                     // raw provider result (includes full content)
-  reasoning: Array<ReasoningOutput>       // reasoning / thinking segments
-  reasoning_text: string                  // the reasoning segments concatenated
+  reasoning: Array<ReasoningPart>         // reasoning / thinking segments
+  reasoningText: string                   // the reasoning segments concatenated
   sources: Array<Source>                  // sources / citations (search-preview models)
   files: Array<GeneratedFile>             // files generated by the model
-  response_messages: Array<ModelMessage>  // assistant messages ready for the next turn
-  raw_finish_reason: string | null        // provider's own finish-reason string
-  provider_metadata: JsonValue | null     // mirrored from raw.provider_metadata
+  responseMessages: Array<ModelMessage>   // assistant messages ready for the next turn
+  rawFinishReason?: string                // provider's own finish-reason string
+  providerMetadata?: Record<string, Record<string, JsonValue>>  // mirrored from raw.providerMetadata
   request: RequestInfo                    // sent request body
-  response: ResponseInfo                  // id, timestamp, model_id, headers and body
-  total_usage: Usage                      // usage across all steps (equals usage in single-step mode)
+  response: ResponseInfo                  // id, timestamp, modelId, headers and body
+  totalUsage: Usage                       // usage across all steps (equals usage in single-step mode)
 }
 ```
 
-`streamText` emits `TextStreamPart`, an external-tagged union (each is a one-key
-object — type narrowing via `part.TextDelta` etc. works out of the box):
+`streamText` emits `TextStreamPart`, a union tagged by `type` (type narrowing via
+`part.type === 'text-delta'` etc. works out of the box):
 
 ```typescript
 // bindings/node/src/types/TextStreamPart.ts (variants, abridged)
 export type TextStreamPart =
-  | { StreamStart: ... } | { TextStart: ... } | { TextDelta: ... } | { TextEnd: ... }
-  | { ToolInputStart: ... } | { ToolInputDelta: ... } | { ToolInputEnd: ... }
-  | { ToolCall: ... } | { ToolResult: ... }
-  | { ReasoningStart: ... } | { ReasoningDelta: ... } | { ReasoningEnd: ... }
-  | { Source: ... } | { Finish: ... }
-  | { Error: ... } | { Raw: ... } | { File: ... }
+  | { type: 'stream-start', ... } | { type: 'text-start', ... } | { type: 'text-delta', ... } | { type: 'text-end', ... }
+  | { type: 'tool-input-start', ... } | { type: 'tool-input-delta', ... } | { type: 'tool-input-end', ... }
+  | { type: 'tool-call', ... } | { type: 'tool-result', ... }
+  | { type: 'reasoning-start', ... } | { type: 'reasoning-delta', ... } | { type: 'reasoning-end', ... }
+  | { type: 'source', ... } | { type: 'finish', ... } | { type: 'finish-step', ... }
+  | { type: 'error', ... } | { type: 'raw', ... } | { type: 'file', ... }
+  | { type: 'reasoning-file', ... } | { type: 'custom', ... } | { type: 'tool-approval-request', ... }
 ```
 
 Provider-layer `StreamPart` also has `ResponseMetadata`; the call layer
diff --git a/docs/api/python.md b/docs/api/python.md
index f8f9c32e3..beb61ffc1 100644
--- a/docs/api/python.md
+++ b/docs/api/python.md
@@ -42,7 +42,7 @@ result = generate_text(model, "Hello")
 `ProviderOptions` shape (`base_url` / `headers` / `organization` / `project` /
 `params`); the `base_url` parameter wins over `config["base_url"]`.
 `max_retries` (a per-call option) and `body_overrides` (removed) in `config`
-raise `InvalidArgumentError`. A missing key or setting raises `LoadAPIKeyError`
+raise `InvalidArgumentError`. A missing key or setting raises (on the first call) `LoadAPIKeyError`
 (`env_var`, `description`) / `LoadSettingError` (`env_var`, `setting_name`). `openai` / `anthropic` / `deepseek` factories
 remain; DeepSeek now uses its own vendor chat model. `openai` and by-name
 `provider("openai", …)` select the Responses API.
@@ -60,13 +60,13 @@ from aimux import openai, generate_text
 
 model = openai("sk-...", "gpt-4o")
 result = generate_text(model, "Explain Rust ownership.", {
-    "max_output_tokens": 100,
+    "maxOutputTokens": 100,
     "temperature": 0.7,
 })
 
 print(result["text"])
 print(result["usage"])
-print(result["finish_reason"])
+print(result["finishReason"])
 ```
 
 > Parameters, return value, and the `raw.content` variants are documented in
@@ -81,9 +81,9 @@ from aimux import openai, stream_text
 
 model = openai("sk-...", "gpt-4o")
 for part in stream_text(model, "Write a haiku about Rust."):
-    if "TextDelta" in part:
-        print(part["TextDelta"]["delta"], end="")
-    if "Finish" in part:
+    if part["type"] == "text-delta":
+        print(part["delta"], end="")
+    if part["type"] == "finish":
         print("\n[done]")
 ```
 
@@ -101,7 +101,7 @@ tools = [{
     "type": "function",
     "name": "get_weather",
     "description": "Get current weather",
-    "input_schema": {
+    "inputSchema": {
         "type": "object",
         "properties": {"location": {"type": "string", "description": "City name"}},
         "required": ["location"]
@@ -109,30 +109,30 @@ tools = [{
 }]
 
 result = generate_text(model, "What's the weather in Tokyo?", {"tools": tools})
-if len(result["tool_calls"]) > 0:
-    call = result["tool_calls"][0]
-    print(call["tool_name"])      # get_weather
+if len(result["toolCalls"]) > 0:
+    call = result["toolCalls"][0]
+    print(call["toolName"])      # get_weather
     print(call["input"])          # {"location": "Tokyo"}
 ```
 
 ### Repairing Invalid Tool Calls
 
 A tool call the model got wrong never fails generation: it comes back with
-`invalid: true` and a typed `error`. Pass `repair_tool_call` to get one shot at
+`invalid: true` and a typed `error`. Pass `repairToolCall` to get one shot at
 fixing it (RFC-0035, the equivalent of the AI SDK `repairToolCall`):
 
 ```python
 def repair(ctx):
-    # ctx = {tool_call, error, input_schema, tools, messages, instructions};
-    # ctx["tool_call"]["input"] is the provider's raw argument text.
-    args = json.loads(ctx["tool_call"]["input"])
+    # ctx = {toolCall, error, inputSchema, tools, messages, instructions};
+    # ctx["toolCall"]["input"] is the provider's raw argument text.
+    args = json.loads(ctx["toolCall"]["input"])
     if "location" not in args:
         return None                      # None = leave the call invalid
-    return dict(ctx["tool_call"], input=json.dumps({"city": args["location"]}))
+    return dict(ctx["toolCall"], input=json.dumps({"city": args["location"]}))
 
 result = generate_text(model, "What's the weather in Tokyo?",
-                       {"tools": tools, "repair_tool_call": repair})
-print(result["tool_calls"][0]["input"])   # {"city": "Tokyo"}
+                       {"tools": tools, "repairToolCall": repair})
+print(result["toolCalls"][0]["input"])   # {"city": "Tokyo"}
 ```
 
 The function is called once per invalid call, after generation, and its result
@@ -140,13 +140,13 @@ is re-validated against the tool's schema: a call that is still wrong — and on
 whose repair raises — comes back invalid with a `ToolCallRepair` error carrying
 the original one. Returning `None` leaves the call untouched. It runs on the
 Python side, outside any native call, so it may itself call `generate_text` to
-ask a model for better arguments. Both `tool_calls` and `response_messages` are
+ask a model for better arguments. Both `toolCalls` and `responseMessages` are
 patched, so the next turn replays the repaired arguments.
 
-The same option exists on the typed wrapper's `GenerateTextOptions`, where it
+The same option exists on the typed wrapper's `GenerateTextOptions` (as `repair_tool_call`), where it
 takes a `ToolCallRepairContext` and returns a `RawToolCall`. It applies to
 `generate_text`, `generate_object`, `consume_stream_text` and `stream_text`
-(which yields the repaired `ToolCall` part; tool-input deltas are the
+(which yields the repaired `tool-call` part; tool-input deltas are the
 provider's own text and pass through untouched), and to
 `generate_text_as_openai`: a `chat.completion` carries no `invalid` marker, so
 it repairs the native result and converts it
@@ -159,12 +159,12 @@ returns something other than a `RawToolCall` or `None` — propagates as itself.
 
 ### Tool Selection Strategy
 
-Pass `tool_choice` through the options dict:
+Pass `toolChoice` through the options dict:
 
 ```python
 opts = {
     "tools": tools,
-    "tool_choice": "auto"   # "auto" | "none" | "required" | {"type": "tool", "toolName": "get_weather"}
+    "toolChoice": "auto"   # "auto" | "none" | "required" | {"type": "tool", "toolName": "get_weather"}
 }
 ```
 
@@ -207,11 +207,13 @@ speaker = openai_speech("sk-...", "tts-1")
 result = json.loads(speaker.generate(json.dumps({
     "text": "Hello world!",
     "voice": "alloy",
-    "output_format": "mp3",
+    "outputFormat": "mp3",
 })))
 
-if "Base64" in result["audio"]:
-    audio_bytes = base64.b64decode(result["audio"]["Base64"])
+# result["audio"] is a list of byte values, or a base64 string
+audio = result["audio"]
+audio_bytes = bytes(audio) if isinstance(audio, list) else base64.b64decode(audio)
+if audio_bytes:
     with open("out.mp3", "wb") as f:
         f.write(audio_bytes)
 ```
@@ -242,12 +244,13 @@ imager = openai_image("sk-...", "dall-e-3")
 result = json.loads(imager.generate(json.dumps({
     "prompt": "A cute baby sea otter",
     "n": 1,
-    "provider_options": {},
+    "providerOptions": {},
 })))
 
-if "Base64" in result["images"]:
-    with open("out.png", "wb") as f:
-        f.write(base64.b64decode(result["images"]["Base64"][0]))
+# result["images"] is a list of base64 strings, or a list of byte arrays
+image = result["images"][0]
+with open("out.png", "wb") as f:
+    f.write(base64.b64decode(image) if isinstance(image, str) else bytes(image))
 ```
 
 ## Video Generation
@@ -262,12 +265,12 @@ videor = google_video("sk-...", "veo-3.0")
 result = json.loads(videor.generate(json.dumps({
     "prompt": "A cat playing piano",
     "n": 1,
-    "provider_options": {},
+    "providerOptions": {},
 })))
 
-# result["videos"] is usually [{"Url": {"url": "...", "media_type": "..."}}]
-if "Url" in result["videos"][0]:
-    print(result["videos"][0]["Url"]["url"])
+# result["videos"] is usually [{"type": "url", "url": "...", "mediaType": "..."}]
+if result["videos"][0]["type"] == "url":
+    print(result["videos"][0]["url"])
 ```
 
 ## Reranking
@@ -279,23 +282,23 @@ from aimux import cohere_reranking
 import json
 
 reranker = cohere_reranking("sk-...", "rerank-v3.0")
-# docs_json is the externally-tagged `RerankingDocuments` enum —
-# {"Object": {"values": [...]}} for JSON documents, {"Text": {"values": [...]}} for strings
-docs = {"Object": {"values": [
+# docs_json is the `type`-tagged `RerankingDocuments` enum —
+# {"type": "object", "values": [...]} for JSON documents, {"type": "text", "values": [...]} for strings
+docs = {"type": "object", "values": [
     {"text": "Rust is a systems programming language."},
     {"text": "Rust is a chemical element."},
-]}}
+]}
 result = json.loads(reranker.rerank(
     "What is Rust?",
     json.dumps(docs),
-    # opts_json is a whole RerankingCallOptions: only top_n and provider_options
+    # opts_json is a whole RerankingCallOptions: only topN and providerOptions
     # are read from it, but query and documents must be present to deserialize
-    json.dumps({"query": "What is Rust?", "documents": docs, "top_n": 3}),
+    json.dumps({"query": "What is Rust?", "documents": docs, "topN": 3}),
 ))
 
-# result["ranking"] sorted by relevance_score
+# result["ranking"] sorted by relevanceScore
 for rank in result["ranking"]:
-    print(rank["index"], rank["relevance_score"])
+    print(rank["index"], rank["relevanceScore"])
 ```
 
 ## Search
@@ -323,7 +326,7 @@ files = openai_files("sk-...")
 file_b64 = base64.b64encode(open("doc.pdf", "rb").read()).decode()
 result = json.loads(files.upload_file(file_b64, "application/pdf"))
 
-print(result["provider_reference"])  # {"openai": "file-xxx"}
+print(result["providerReference"])  # {"openai": "file-xxx"}
 ```
 
 ## API Surface
@@ -354,7 +357,7 @@ Three module-level native functions back tool-call repair (all pure and
 synchronous, JSON in / JSON out): `tool_call_repair_context(tool_call_json,
 prompt_json, opts_json=None)`, `apply_tool_call_repair(tool_call_json,
 opts_json, reply_json)` and `apply_tool_call_repair_to_result(result_json,
-opts_json, tool_call_id, reply_json)`. The `repair_tool_call` option drives
+opts_json, tool_call_id, reply_json)`. The `repairToolCall` option drives
 them for you; call them directly only to build your own loop.
 
 All factories accept an optional `base_url` and return instances synchronously
@@ -375,7 +378,7 @@ from aimux.wrapper import (
     Tool, ToolChoice, ResponseFormat, StreamPart, GenerateContent,
     FileData, FileBytes, Warning, AiMuxErrorValue,
     # pydantic models
-    TokenUsage, Usage, FinishReason, ResponseMetadata, ToolCall,
+    InputTokenUsage, OutputTokenUsage, Usage, FinishReason, ResponseMetadata, ToolCall,
     ModelMessage, FunctionTool, ProviderTool, TextContentPart,
     GenerateTextOptions, GenerateTextResult, GenerateResult,
     RawToolCall, ToolCallRepairContext, RepairToolCall,
@@ -464,6 +467,12 @@ except AimuxError:
 (The pydantic wire type named `AiMuxErrorValue` in `aimux.wrapper` is only for
 stream/payload shapes, not the raised exception.)
 
+The wire format is the AI SDK's: JSON keys are camelCase and unions are tagged
+by `type`. The raw dict API exchanges that JSON as-is (`result["finishReason"]`,
+`{"type": "text-delta", ...}`). The typed wrapper's models keep `snake_case`
+Python attributes and map them to the camelCase keys through a pydantic alias
+generator (both spellings validate; serialization uses camelCase).
+
 Key shapes (mirroring the shared JSON schema):
 
 ```python
@@ -478,12 +487,11 @@ class GenerateTextResult(BaseModel):
     response: ResponseInfo
 ```
 
-`stream_text` emits the Rust call-layer `TextStreamPart` JSON. The Python
-wrapper calls its validating `RootModel` `StreamPart`, over the external-tagged
-union dict, e.g.
-`{"TextDelta": {"id": ..., "delta": ...}}`. Iterate dicts with
-`if "TextDelta" in part:` (as in [Streaming Generation](#streaming-generation))
-or validate them with `parse_stream_part(part)` for attribute access.
+`stream_text` emits the Rust call-layer `TextStreamPart` JSON: a dict tagged by
+`type`, e.g. `{"type": "text-delta", "id": ..., "delta": ...}`. Iterate dicts with
+`if part["type"] == "text-delta":` (as in [Streaming Generation](#streaming-generation))
+or validate them with `parse_stream_part(part)` for attribute access (the
+wrapper's `StreamPart` is a pydantic discriminated union on `type`).
 
 Provider-layer `StreamPart::ResponseMetadata` events are consumed internally
 and are not emitted by `stream_text`.
diff --git a/docs/api/swift.md b/docs/api/swift.md
index 7faa6da12..804f82c9f 100644
--- a/docs/api/swift.md
+++ b/docs/api/swift.md
@@ -41,8 +41,9 @@ The vendor packages and 282 registry-backed OpenAI-compatible providers are reac
 let model = try Model.provider(name: "groq", modelId: "llama-3.3-70b")
 let result = try model.generateText(prompt: "\"Hello\"")
 
-// Optional config JSON ({"base_url": "..."}):
-let model2 = try Model.provider(name: "groq", apiKey: "sk-...", modelId: "llama-3.3-70b")
+// Optional config JSON ({"baseUrl": "..."}; unknown keys are rejected):
+let model2 = try Model.provider(name: "groq", apiKey: "sk-...", modelId: "llama-3.3-70b",
+                                configJson: #"{"baseUrl":"https://proxy.example/v1"}"#)
 ```
 
 Unknown names throw `.noSuchProvider` (payload: the provider id).
@@ -65,14 +66,44 @@ print(result2)
 > Parameters, return value, and the `raw.content` variants are documented in
 > the [API overview](../API.md#text-generation).
 
+The JSON is the AI SDK JSON: camelCase fields, optional fields absent (never
+`null`), unions tagged by `type` (`{"type":"tool-call","toolCallId":…}`). The
+typed overloads take and return Codable values with exactly that shape:
+
+```swift
+let options = GenerateTextOptions(
+    tools: [.function(FunctionTool(name: "get_weather", inputSchema: schema))],
+    toolChoice: .required
+)
+let result = try model.generateText(prompt: .text("Weather in Tokyo?"), options: options)
+print(result.text, result.toolCalls.first?.toolName ?? "-", result.usage.inputTokens.total ?? 0)
+
+// Send a tool result back: the part carries `toolName` and a typed `output`.
+let followUp: [ModelMessage] = [
+    .user("Weather in Tokyo?"),
+    ModelMessage(role: .assistant, content: .parts([
+        .toolCall(toolCallId: "call_1", toolName: "get_weather", input: .object(["location": .string("Tokyo")]))
+    ])),
+    ModelMessage(role: .tool, content: .parts([
+        .toolResult(toolCallId: "call_1", toolName: "get_weather",
+                    output: .json(.object(["temperature": .number(22)])))
+    ])),
+]
+```
+
+`ToolResultOutput` is `.text`, `.json`, `.executionDenied(reason:)`,
+`.errorText`, `.errorJson` or `.content([ToolResultContent])`
+(`{"type":"text"|"json"|"execution-denied"|"error-text"|"error-json"|"content"}`).
+
 ## Streaming Generation
 
 `streamText(prompt:options:onPart:onDone:onError:)` delivers each part as a
-`StreamPart` JSON string.
+`TextStreamPart` JSON string (`{"type":"text-delta","id":…,"delta":…}`); the
+typed overload decodes it into `TextStreamPart`.
 
 ```swift
 model.streamText(prompt: "\"Write a haiku\"") { part in
-    print(part) // StreamPart JSON string
+    print(part) // TextStreamPart JSON string
 } onDone: {
     print("[done]")
 } onError: { error in
@@ -178,7 +209,7 @@ and yields the invariant `DecodingError.dataCorrupted("aimux ffi: <context>:
 | `.aborted` | `AIMUX_E_ABORTED` (13) | Request aborted |
 | `.noSuchTool` | `AIMUX_E_NO_SUCH_TOOL` (15) | The model called a tool outside the supplied tool set; carries `toolName` and `availableTools` (`[String]?`, `nil` when no tool set was supplied) |
 | `.invalidToolInput` | `AIMUX_E_INVALID_TOOL_INPUT` (16) | Tool arguments failed to parse/validate; carries `toolName` and `toolInput` (the raw argument text) |
-| `.toolCallRepair` | `AIMUX_E_TOOL_CALL_REPAIR` (17) | A `repairToolCall` hook itself failed; carries `originalError` (the repaired-over error as wire JSON, the `ToolCall.error` encoding) |
+| `.toolCallRepair` | `AIMUX_E_TOOL_CALL_REPAIR` (17) | A `repairToolCall` hook itself failed; carries `originalError` (the repaired-over error as `AiMuxError` JSON tagged by `name`, the `ToolCall.error` encoding) |
 | `.loadApiKey` | `AIMUX_E_LOAD_API_KEY` (18) | No API key passed and the fallback environment variable is unset; carries `envVar` |
 | `.loadSetting` | `AIMUX_E_LOAD_SETTING` (19) | A required provider setting is missing; carries `envVar` |
 | `.other` | `AIMUX_E_OTHER` (1) | Unclassified core error |
@@ -256,23 +287,33 @@ throws the same.
 
 ## Types
 
-`bindings/swift/Sources/Aimux/Types.swift` declares lightweight Codable types
-mirroring the shared JSON shape — usable with the JSON-string APIs:
+`bindings/swift/Sources/Aimux/Types.swift` declares Codable types decoding and
+encoding exactly the JSON of the generated `bindings/node/src/types/*.ts`
+(camelCase fields, unions tagged by `type`) — usable with the JSON-string APIs:
 
 `JSONValue` (recursive JSON enum with `stringValue` / `boolValue` /
-`doubleValue` / `intValue` / `arrayValue` / `objectValue` accessors), `Role`,
-`FinishReasonUnified`, `ReasoningEffort`, `FinishReason`, `TokenUsage`,
-`Usage`, `ResponseMetadata`, `Warning`, `FunctionTool`, `ProviderTool`,
-`Tool`, `ToolChoice`, `ResponseFormat`, `ContentPart`, `MessageContent`,
-`ModelMessage`, `ModelPrompt`, `ToolCall`, `FileBytes`, `FileData`,
-`GenerateContent`, `GenerateResult`, `GenerateTextResult`,
-`GenerateTextOptions`, `StreamPart`, `RawToolCall`, `ToolCallRepairContext`
-(all `Codable, Equatable`).
-
-`ToolCall` (top-level and `StreamPart.toolCall`) carries `providerMetadata`
+`doubleValue` / `intValue` / `arrayValue` / `objectValue` accessors),
+`ProviderOptions` / `ProviderMetadata` (`[String: [String: JSONValue]]`),
+`Role`, `FinishReasonUnified`, `ReasoningEffort`, `FinishReason`,
+`InputTokenUsage`, `OutputTokenUsage`, `Usage`, `ResponseMetadata`,
+`ResponseInfo`, `RequestInfo`, `Warning`, `FunctionTool`, `ProviderTool`,
+`Tool`, `ToolChoice`, `ResponseFormat`, `ContentPart`, `ToolResultOutput`,
+`ToolResultContent`, `MessageContent`, `ModelMessage`, `ModelPrompt`,
+`ToolCall`, `ToolResult`, `FileBytes`, `FileData`, `GeneratedFile`, `Source`,
+`GenerateContent` (`RawGenerateContent` / `TextContent`), `GenerateResult`,
+`GenerateTextResult`, `GenerateObjectResult`, `StreamTextResultAggregated`,
+`GenerateTextOptions`, `TimeoutConfiguration`, `TextStreamPart`, `RawToolCall`,
+`ToolCallRepairContext`, the OpenAI-shaped `ChatCompletion` /
+`ChatCompletionChunk` (vendor shape, snake_case), and the multimodal results
+(`EmbeddingResult`, `SpeechResult`, `ImageResult`, `TranscriptionResult`,
+`VideoResult`, `RerankingResult`, `SearchResult`, `UploadFileResult`), all
+`Codable, Equatable`.
+
+`ToolCall` (top-level and `TextStreamPart.toolCall`) carries `providerMetadata`
 plus `invalid` (set by Core when tool lookup, input parse, or schema validation
-fails, even after repair) and `error` (the serialized `AiMuxError` for that
-failure).
+fails, even after repair) and `error` (the `AiMuxError` object for that
+failure, tagged by `name`, e.g. `"AI_InvalidToolInputError"`; held as a
+`JSONValue`, as is `TextStreamPart.error`).
 
 Example:
 
@@ -284,7 +325,7 @@ print(decoded.inputTokens.total ?? 0)
 
 ## Coverage
 
-Text generation and streaming are supported. Multimodal features (embedding,
-TTS, STT, image, video, rerank, search, files) are reachable only through the
-raw [C ABI](c.md) until the wrappers are extended — see the
-[coverage matrix](../API.md#feature-coverage).
+Text generation and streaming have typed wrappers. The multimodal features
+(embedding, TTS, STT, image, video, rerank, search, files) take their options
+as JSON strings and return JSON strings that decode into the typed result
+structs above; see the [coverage matrix](../API.md#feature-coverage).
diff --git a/rfc/0037-decision-api.md b/rfc/0037-decision-api.md
index 3c6d0e118..3a5c7e62b 100644
--- a/rfc/0037-decision-api.md
+++ b/rfc/0037-decision-api.md
@@ -117,8 +117,8 @@ server 验证；另于 2026-10-06 使用 Python binding 对官方 endpoint 完
 CI 无需线上 key。
 
 `decide` 复用统一 JSONL/ring recorder，输入标记为 decision（旧记录默认
-language model），state/questions 存入 input.options，能力声明存入 input.decision_capabilities，provider 仅保留身份，
-规范化结果存入 outcome.decision_result。每次 retry 分配独立 attempt，
+language model），state/questions 存入 input.options，能力声明存入 input.decisionCapabilities，provider 仅保留身份，
+规范化结果存入 outcome.decisionResult。每次 retry 分配独立 attempt，
 录制器快照跨 HTTP、取消和收尾保持一致。覆盖成功、错误、超时、取消、
 脱敏、无重复占位记录、严格匹配和 recorder 替换。
 `aimux-replay --mock` 支持标准 JSONL 的离线 decision 回放；不带 --mock
diff --git a/rfc/0038-decision-provider-expansion.md b/rfc/0038-decision-provider-expansion.md
index cc92cbd86..3ee60238f 100644
--- a/rfc/0038-decision-provider-expansion.md
+++ b/rfc/0038-decision-provider-expansion.md
@@ -53,7 +53,7 @@ projector、endpoint 是否可用由服务返回实际错误。vLLM 示例必须
 
 - `state` 是文本或 JSON 证据；`images` 是独立的内嵌图片列表，不从 state 猜媒体。
 - `DecisionImage` 为 `{data: FileBytes, media_type, detail?}`，复用已有
-  `FileBytes` 的 `{"Base64":"..."}` / `{"Binary":[...]}` 表达。
+  `FileBytes` 的表达：base64 字符串或字节数组。
 - Choice 的 `label` 始终是答案及概率 map 的稳定字符串键。新增可选 `value`，
   类型为 string / boolean；缺省使用 label。原生 boolean `true` 与字符串
   `"true"` 可同时出现，只需不同 label。响应保留 selected label 和显式 value。
diff --git a/tools/aimux-cli/src/probe/offline.rs b/tools/aimux-cli/src/probe/offline.rs
index 3ad13eea2..c848f24a8 100644
--- a/tools/aimux-cli/src/probe/offline.rs
+++ b/tools/aimux-cli/src/probe/offline.rs
@@ -88,31 +88,29 @@ mod tests {
             r#"{{
                 "provider": "openai",
                 "model": "gpt-4o",
-                "session_id": "{session}",
-                "call_id": "t-{session}-{claimed}",
-                "sent_at_unix_ms": 1785900000000,
-                "monotonic_sent_ms": 0,
+                "sessionId": "{session}",
+                "callId": "t-{session}-{claimed}",
+                "sentAtUnixMs": 1785900000000,
                 "fingerprint": {{
-                    "body_hash": "",
-                    "len_bytes": 8192,
-                    "block_size": 4096,
-                    "block_hashes": [],
-                    "token_estimate": 2048
+                    "bodyHash": "",
+                    "lenBytes": 8192,
+                    "blockSize": 4096,
+                    "blockHashes": [],
+                    "tokenEstimate": 2048
                 }},
                 "usage": {{
-                    "input_total": 2048,
-                    "cache_read": {claimed}
+                    "inputTotal": 2048,
+                    "cacheRead": {claimed}
                 }},
                 "verdict": {{
                     "kind": "Trusted",
                     "confidence": "High",
                     "violated": [],
-                    "expected_max": 2048,
+                    "expectedMax": 2048,
                     "claimed": {claimed},
-                    "lcp_bytes": 8192,
+                    "lcpBytes": 8192,
                     "notes": []
-                }},
-                "scope_key": 1
+                }}
             }}"#
         ))
         .unwrap()
@@ -140,7 +138,7 @@ mod tests {
         // 捕获 stdout 断言 JSON 输出。
         let out = run_captured(&args);
         assert!(out.contains("\"requests\":2"), "aggregate count: {out}");
-        assert!(out.contains("\"claimed_cache_read_total\":1536"), "{out}");
+        assert!(out.contains("\"claimedCacheReadTotal\":1536"), "{out}");
         assert!(out.contains("\"provider\":\"openai\""), "{out}");
 
         // session 过滤。
diff --git a/tools/aimux-cli/src/probe/session.rs b/tools/aimux-cli/src/probe/session.rs
index e74b6199d..a27170be1 100644
--- a/tools/aimux-cli/src/probe/session.rs
+++ b/tools/aimux-cli/src/probe/session.rs
@@ -40,17 +40,17 @@ mod tests {
         serde_json::json!({
             "provider": "openai",
             "model": "gpt-4o",
-            "session_id": session,
-            "call_id": format!("t-{idx}"),
-            "sent_at_unix_ms": 1785900000000 + idx as i64 * 1000,
+            "sessionId": session,
+            "callId": format!("t-{idx}"),
+            "sentAtUnixMs": 1785900000000 + idx as i64 * 1000,
             "fingerprint": {
-                "body_hash": format!("{idx:032x}"),
-                "len_bytes": body_len,
-                "block_size": 4096,
-                "block_hashes": vec![format!("{:032x}", idx)],
-                "token_estimate": body_len / 4,
+                "bodyHash": format!("{idx:032x}"),
+                "lenBytes": body_len,
+                "blockSize": 4096,
+                "blockHashes": vec![format!("{:032x}", idx)],
+                "tokenEstimate": body_len / 4,
             },
-            "usage": {"input_total": body_len / 4},
+            "usage": {"inputTotal": body_len / 4},
         })
     }
 
@@ -80,7 +80,7 @@ mod tests {
         let chain = store.session_chain("sess-1").unwrap();
         let json = serde_json::to_string(&chain).unwrap();
         assert_eq!(chain.record_ids.len(), 2);
-        assert!(json.contains("\"session_id\":\"sess-1\""), "{json}");
+        assert!(json.contains("\"sessionId\":\"sess-1\""), "{json}");
 
         // 未知 session → 无输出错误路径。
         let args2 = SessionArgs {
diff --git a/tools/aimux-web/examples/gen_demo_fixtures.rs b/tools/aimux-web/examples/gen_demo_fixtures.rs
index 242628865..4d180e832 100644
--- a/tools/aimux-web/examples/gen_demo_fixtures.rs
+++ b/tools/aimux-web/examples/gen_demo_fixtures.rs
@@ -99,7 +99,7 @@ fn main() {
         .nth(1)
         .unwrap_or_else(|| "demo.jsonl".to_string());
 
-    // OpenAI chat.completions SSE bodies so mock streaming emits TextDelta.
+    // OpenAI chat.completions SSE bodies so mock streaming emits text-delta parts.
     let sse_pg = r#"data: {"id":"chatcmpl-demo-pg","model":"gpt-4o-mini"}
 
 data: {"choices":[{"index":0,"delta":{"role":"assistant","content":"1 + 1 = 2"}}]}
@@ -132,8 +132,8 @@ data: [DONE]"#;
             sse_pg,
             320,
             serde_json::json!({
-                "input_tokens": { "total": 12, "cache_read": 0 },
-                "output_tokens": { "total": 5 }
+                "inputTokens": { "total": 12, "cacheRead": 0 },
+                "outputTokens": { "total": 5 }
             }),
         ),
         // Agent mock target: [system (agent), user] — matches the Agent page's default def.
@@ -156,8 +156,8 @@ data: [DONE]"#;
             sse_agent,
             412,
             serde_json::json!({
-                "input_tokens": { "total": 64, "cache_read": 0 },
-                "output_tokens": { "total": 9 }
+                "inputTokens": { "total": 64, "cacheRead": 0 },
+                "outputTokens": { "total": 9 }
             }),
         ),
         // Enrichment: a multi-step session (steps 1 & 2) on deepseek.
@@ -180,8 +180,8 @@ data: [DONE]"#;
             "{\"choices\":[{\"message\":{\"content\":\"Prefix caching reuses the KV cache...\"}}]}",
             810,
             serde_json::json!({
-                "input_tokens": { "total": 2300, "cache_read": 0 },
-                "output_tokens": { "total": 210 }
+                "inputTokens": { "total": 2300, "cacheRead": 0 },
+                "outputTokens": { "total": 210 }
             }),
         ),
         recording(
@@ -211,8 +211,8 @@ data: [DONE]"#;
             "{\"choices\":[{\"message\":{\"content\":\"cache_read / input_total\"}}]}",
             796,
             serde_json::json!({
-                "input_tokens": { "total": 2650, "cache_read": 2300 },
-                "output_tokens": { "total": 18 }
+                "inputTokens": { "total": 2650, "cacheRead": 2300 },
+                "outputTokens": { "total": 18 }
             }),
         ),
         // Enrichment: an errored call (shows the error badge).
@@ -228,7 +228,7 @@ data: [DONE]"#;
             None,
             "{}",
             1200,
-            serde_json::json!({ "input_tokens": null, "output_tokens": null }),
+            serde_json::json!({ "inputTokens": null, "outputTokens": null }),
         ),
         // Enrichment: an older successful call on a different model.
         recording(
@@ -244,8 +244,8 @@ data: [DONE]"#;
             "{\"choices\":[{\"message\":{\"content\":\"Ownership, borrowing — memory safe by design...\"}}]}",
             540,
             serde_json::json!({
-                "input_tokens": { "total": 42, "cache_read": 0 },
-                "output_tokens": { "total": 24 }
+                "inputTokens": { "total": 42, "cacheRead": 0 },
+                "outputTokens": { "total": 24 }
             }),
         ),
     ];
diff --git a/tools/aimux-web/examples/gen_fixture.rs b/tools/aimux-web/examples/gen_fixture.rs
index ac1e394b0..8c8a8525d 100644
--- a/tools/aimux-web/examples/gen_fixture.rs
+++ b/tools/aimux-web/examples/gen_fixture.rs
@@ -72,8 +72,8 @@ fn main() {
             error_value: None,
             decision_result: None,
             usage: Some(serde_json::json!({
-                "input_tokens": { "total": 12, "cache_read": 0 },
-                "output_tokens": { "total": 5 }
+                "inputTokens": { "total": 12, "cacheRead": 0 },
+                "outputTokens": { "total": 5 }
             })),
         },
         complete: true,
diff --git a/tools/aimux-web/src/api/calls.rs b/tools/aimux-web/src/api/calls.rs
index 7cd390a11..5106fc34c 100644
--- a/tools/aimux-web/src/api/calls.rs
+++ b/tools/aimux-web/src/api/calls.rs
@@ -21,7 +21,7 @@ use crate::wire::{self, WireCallRequest, WireCallResponse};
 
 /// `POST /api/calls` — validates the wire request, builds (or reuses) the
 /// model, and returns either a non-streaming JSON result or an SSE stream of
-/// `StreamPart` events + a final `meta` event (trace anchor).
+/// `TextStreamPart` events + a final `meta` event (trace anchor).
 pub async fn call(State(state): State<AppState>, Json(req): Json<WireCallRequest>) -> Response {
     match run(state, req).await {
         Ok(resp) => resp,
@@ -70,7 +70,7 @@ async fn run(state: AppState, req: WireCallRequest) -> Result<Response, AiMuxErr
     }
 }
 
-/// Non-streaming: JSON `{text, finish_reason, usage, meta}`.
+/// Non-streaming: JSON `{text, finishReason, usage, meta}`.
 async fn non_stream_response(
     state: AppState,
     model: Arc<dyn LanguageModel>,
diff --git a/tools/aimux-web/src/api/replay.rs b/tools/aimux-web/src/api/replay.rs
index 19233ba7e..e96291493 100644
--- a/tools/aimux-web/src/api/replay.rs
+++ b/tools/aimux-web/src/api/replay.rs
@@ -20,6 +20,7 @@ use crate::state::AppState;
 use crate::wire::{self, WireMessage};
 
 #[derive(Deserialize)]
+#[serde(rename_all = "camelCase")]
 pub struct ReplayRequest {
     pub call_id: String,
     /// Fallback for recordings whose key source is `explicit`/`unknown`:
@@ -32,6 +33,7 @@ pub struct ReplayRequest {
 }
 
 #[derive(Deserialize)]
+#[serde(rename_all = "camelCase")]
 pub struct ReplayOverridesWire {
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub messages: Option<Vec<WireMessage>>,
@@ -106,11 +108,11 @@ async fn run_inner(state: AppState, req: ReplayRequest) -> Result<Response, AiMu
 
     let result = replay_with_model(&recording, &*model, Some(&overrides)).await?;
     Ok(Json(json!({
-        "call_id": recording.call_id,
+        "callId": recording.call_id,
         "text": result.text,
-        "finish_reason": serde_json::to_value(result.finish_reason.unified).unwrap_or_default(),
+        "finishReason": serde_json::to_value(result.finish_reason.unified).unwrap_or_default(),
         "usage": serde_json::to_value(&result.usage).unwrap_or_default(),
-        "tool_calls": serde_json::to_value(&result.tool_calls).unwrap_or_default(),
+        "toolCalls": serde_json::to_value(&result.tool_calls).unwrap_or_default(),
         "meta": state.last_meta(recording.session_id.as_deref()),
     }))
     .into_response())
diff --git a/tools/aimux-web/src/api/sessions.rs b/tools/aimux-web/src/api/sessions.rs
index 22859c00c..5c555dd5e 100644
--- a/tools/aimux-web/src/api/sessions.rs
+++ b/tools/aimux-web/src/api/sessions.rs
@@ -21,7 +21,7 @@ pub async fn detail(State(state): State<AppState>, Path(id): Path<String>) -> Re
         .filter(|r| r.session_id.as_deref() == Some(id.as_str()))
         .collect();
     Json(json!({
-        "session_id": id,
+        "sessionId": id,
         "calls": calls,
         "recordings": recordings,
     }))
diff --git a/tools/aimux-web/src/wire.rs b/tools/aimux-web/src/wire.rs
index e9457b0d2..3d40dc275 100644
--- a/tools/aimux-web/src/wire.rs
+++ b/tools/aimux-web/src/wire.rs
@@ -11,7 +11,9 @@ use ts_rs::TS;
 use aimux_core::content::ContentPart;
 use aimux_core::error::AiMuxError;
 use aimux_core::generate::GenerateTextOptions;
+use aimux_core::language_model_message::ToolResultOutput;
 use aimux_core::message::{MessageContent, ModelMessage, ModelPrompt, Role};
+use aimux_core::options::ResponseFormat;
 use aimux_core::tool::{FunctionTool, Tool};
 
 // ─────────────────────────────────────────────────────────────────────────────
@@ -30,6 +32,7 @@ use aimux_core::tool::{FunctionTool, Tool};
 /// One model call from the console (RFC-0029 §5.3).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
 #[cfg_attr(feature = "ts-export", ts(export))]
+#[serde(rename_all = "camelCase")]
 pub struct WireCallRequest {
     /// Registry name or native protocol name ("openai", "deepseek", …).
     pub provider: String,
@@ -38,12 +41,14 @@ pub struct WireCallRequest {
     /// API key: `None` (Settings-saved key or the provider's registered env
     /// var), `Some("env:VAR")`, or a plaintext literal — the latter only
     /// while the console is loopback-bound (RFC-0029 §5.5).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub api_key: Option<String>,
     /// Base URL override (proxies / local endpoints).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub base_url: Option<String>,
-    /// Stream the response as SSE (`StreamPart` events). Default true.
+    /// Stream the response as SSE (`TextStreamPart` events). Default true.
     #[serde(default = "default_true")]
     pub stream: bool,
     /// Route through the loaded mock model (offline mode, RFC-0023 P3).
@@ -52,9 +57,11 @@ pub struct WireCallRequest {
     #[serde(default)]
     pub options: WireOptions,
     /// Session grouping id (RFC-0024). The console reuses one id per agent run.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub session_id: Option<String>,
     /// Informational step marker (the backend assigns the authoritative step).
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub step: Option<u32>,
     pub messages: Vec<WireMessage>,
@@ -66,24 +73,34 @@ fn default_true() -> bool {
 
 #[derive(Debug, Clone, Default, Serialize, Deserialize, TS)]
 #[cfg_attr(feature = "ts-export", ts(export))]
+#[serde(rename_all = "camelCase")]
 pub struct WireOptions {
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub temperature: Option<f64>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_output_tokens: Option<u32>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub top_p: Option<f64>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub stop_sequences: Option<Vec<String>>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub tools: Option<Vec<WireTool>>,
-    /// `"text"` or `{"json": …}` — passed through as `ResponseFormat`.
+    /// A `ResponseFormat` object: `{"type":"text"}` or `{"type":"json",…}`.
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub response_format: Option<Value>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub headers: Option<HashMap<String, String>>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_retries: Option<u32>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub include_raw_chunks: Option<bool>,
 }
@@ -91,8 +108,10 @@ pub struct WireOptions {
 /// A function tool definition (JSON Schema parameters).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
 #[cfg_attr(feature = "ts-export", ts(export))]
+#[serde(rename_all = "camelCase")]
 pub struct WireTool {
     pub name: String,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub description: Option<String>,
     /// JSON Schema describing the parameters.
@@ -102,6 +121,7 @@ pub struct WireTool {
 /// A chat message in the wire format.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
 #[cfg_attr(feature = "ts-export", ts(export))]
+#[serde(rename_all = "camelCase")]
 pub struct WireMessage {
     /// "system" | "user" | "assistant" | "tool".
     pub role: String,
@@ -111,7 +131,11 @@ pub struct WireMessage {
 /// A content part in the wire format.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
 #[cfg_attr(feature = "ts-export", ts(export))]
-#[serde(tag = "type", rename_all = "snake_case")]
+#[serde(
+    tag = "type",
+    rename_all = "kebab-case",
+    rename_all_fields = "camelCase"
+)]
 pub enum WireContentPart {
     Text {
         text: String,
@@ -120,26 +144,31 @@ pub enum WireContentPart {
         tool_call_id: String,
         tool_name: String,
         input: Value,
+        #[ts(optional)]
         #[serde(default, skip_serializing_if = "Option::is_none")]
         provider_executed: Option<bool>,
     },
     ToolResult {
         tool_call_id: String,
-        result: Value,
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        is_error: Option<bool>,
+        tool_name: String,
+        /// The core type, imported by the frontend through the `@aimux/types` alias.
+        #[ts(type = "import(\"@aimux/types/ToolResultOutput\").ToolResultOutput")]
+        output: ToolResultOutput,
     },
 }
 
 /// Response for non-streaming calls.
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
 #[cfg_attr(feature = "ts-export", ts(export))]
+#[serde(rename_all = "camelCase")]
 pub struct WireCallResponse {
     pub text: String,
     pub finish_reason: Value,
     pub usage: Value,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub meta: Option<WireMeta>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub error: Option<String>,
 }
@@ -147,10 +176,13 @@ pub struct WireCallResponse {
 /// SSE `meta` event payload — the frontend trace anchor (RFC-0029 §5.2).
 #[derive(Debug, Clone, Serialize, Deserialize, TS)]
 #[cfg_attr(feature = "ts-export", ts(export))]
+#[serde(rename_all = "camelCase")]
 pub struct WireMeta {
     pub call_id: String,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub session_id: Option<String>,
+    #[ts(optional)]
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub step: Option<u32>,
     pub outcome: String,
@@ -181,15 +213,12 @@ fn to_content_part(p: &WireContentPart) -> ContentPart {
         },
         WireContentPart::ToolResult {
             tool_call_id,
-            result,
-            is_error,
+            tool_name,
+            output,
         } => ContentPart::ToolResult {
             tool_call_id: tool_call_id.clone(),
-            result: result.clone(),
-            tool_name: None,
-            is_error: *is_error,
-            preliminary: None,
-            dynamic: None,
+            tool_name: tool_name.clone(),
+            output: output.clone(),
             provider_options: None,
         },
     }
@@ -242,27 +271,18 @@ pub fn to_generate_options(
     o: &WireOptions,
     session_id: Option<&str>,
 ) -> Result<GenerateTextOptions, AiMuxError> {
+    // `{"type":"text"}` is the default format, so it maps to "unset".
     let response_format = match &o.response_format {
         None => None,
-        Some(Value::String(s)) if s == "text" => None,
-        Some(Value::Object(obj)) => {
-            let schema = obj.get("schema").cloned();
-            let name = obj.get("name").and_then(Value::as_str).map(str::to_string);
-            let description = obj
-                .get("description")
-                .and_then(Value::as_str)
-                .map(str::to_string);
-            Some(aimux_core::options::ResponseFormat::Json {
-                schema,
-                name,
-                description,
-            })
-        }
-        Some(other) => {
-            return Err(AiMuxError::InvalidArgument(format!(
-                "response_format must be \"text\" or an object, got {other}"
-            )));
-        }
+        Some(v) => match serde_json::from_value::<ResponseFormat>(v.clone()) {
+            Ok(ResponseFormat::Text) => None,
+            Ok(json) => Some(json),
+            Err(e) => {
+                return Err(AiMuxError::InvalidArgument(format!(
+                    "invalid responseFormat: {e}"
+                )));
+            }
+        },
     };
 
     Ok(GenerateTextOptions {
@@ -283,7 +303,7 @@ pub fn to_generate_options(
     })
 }
 
-/// Validate and resolve the wire `api_key` field for one call (RFC-0029 §5.5).
+/// Validate and resolve the wire `apiKey` field for one call (RFC-0029 §5.5).
 ///
 /// Priority:
 /// 1. **explicit request spec** — `Some("env:VAR")` reads the environment
@@ -443,8 +463,11 @@ mod tests {
                 role: "tool".into(),
                 content: vec![WireContentPart::ToolResult {
                     tool_call_id: "c1".into(),
-                    result: serde_json::json!({"value": 2}),
-                    is_error: None,
+                    tool_name: "calc".into(),
+                    output: ToolResultOutput::Json {
+                        value: serde_json::json!({"value": 2}),
+                        provider_options: None,
+                    },
                 }],
             },
         ];
diff --git a/tools/aimux-web/web/src/agent/engine.ts b/tools/aimux-web/web/src/agent/engine.ts
index 41f7170c0..f5ccebdc1 100644
--- a/tools/aimux-web/web/src/agent/engine.ts
+++ b/tools/aimux-web/web/src/agent/engine.ts
@@ -8,7 +8,7 @@
 
 import { api, callStream, parseStreamPart } from '../api/client'
 import type { WireCallRequest } from '@/types/generated/WireCallRequest'
-import type { StreamPart } from '@aimux/types/StreamPart'
+import type { TextStreamPart } from '@aimux/types/TextStreamPart'
 
 /** A chat message as sent to the backend (loose typing: the backend wire
  * schema's `JsonValue` is a recursive union that trips TS2589 inside
@@ -32,7 +32,7 @@ export interface AgentStepView {
   usage?: unknown
   latencyMs?: number
   error?: string
-  meta?: { call_id?: string; session_id?: string | null; step?: number | null; outcome?: string }
+  meta?: { callId?: string; sessionId?: string; step?: number; outcome?: string }
 }
 
 export interface AgentDef {
@@ -99,16 +99,16 @@ export async function runAgent(
       const body: WireCallRequest = {
         provider: def.provider,
         model: def.model,
-        api_key: def.api_key || null,
-        base_url: def.base_url || null,
+        apiKey: def.api_key || undefined,
+        baseUrl: def.base_url || undefined,
         stream: true,
         mock: def.mock,
         options: {
           temperature: def.temperature,
-          max_output_tokens: 2048,
+          maxOutputTokens: 2048,
           tools: def.tools.map((t) => toolSchemas.get(t)).filter(Boolean) as never[],
         },
-        session_id: def.session_id,
+        sessionId: def.session_id,
         step,
         messages: run.messages,
       }
@@ -116,7 +116,7 @@ export async function runAgent(
       try {
         for await (const ev of callStream(body, controller.signal)) {
           if (ev.event === 'stream_part') {
-            const part: StreamPart = parseStreamPart(ev.data)
+            const part: TextStreamPart = parseStreamPart(ev.data)
             applyPart(view, part)
           } else if (ev.event === 'meta') {
             view.meta = JSON.parse(ev.data)
@@ -147,9 +147,9 @@ export async function runAgent(
       if (view.text.trim()) assistantParts.push({ type: 'text', text: view.text })
       for (const tc of view.toolCalls) {
         assistantParts.push({
-          type: 'tool_call',
-          tool_call_id: tc.id,
-          tool_name: tc.name,
+          type: 'tool-call',
+          toolCallId: tc.id,
+          toolName: tc.name,
           input: tc.input,
         })
       }
@@ -163,19 +163,19 @@ export async function runAgent(
           tc.result = out.result
           tc.is_error = false
           toolResults.push({
-            type: 'tool_result',
-            tool_call_id: tc.id,
-            result: out.result,
-            is_error: false,
+            type: 'tool-result',
+            toolCallId: tc.id,
+            toolName: tc.name,
+            output: { type: typeof out.result === 'string' ? 'text' : 'json', value: out.result },
           })
         } catch (e) {
           tc.result = String(e)
           tc.is_error = true
           toolResults.push({
-            type: 'tool_result',
-            tool_call_id: tc.id,
-            result: String(e),
-            is_error: true,
+            type: 'tool-result',
+            toolCallId: tc.id,
+            toolName: tc.name,
+            output: { type: 'error-text', value: String(e) },
           })
         } finally {
           tc.executing = false
@@ -191,22 +191,18 @@ export async function runAgent(
   }
 }
 
-function applyPart(view: AgentStepView, part: StreamPart): void {
-  if ('TextDelta' in part) {
-    view.text += part.TextDelta.delta
-  } else if ('ToolCall' in part) {
-    view.toolCalls.push({
-      id: part.ToolCall.tool_call_id,
-      name: part.ToolCall.tool_name,
-      input: part.ToolCall.input,
-    })
-  } else if ('ToolInputStart' in part) {
+function applyPart(view: AgentStepView, part: TextStreamPart): void {
+  if (part.type === 'text-delta') {
+    view.text += part.delta
+  } else if (part.type === 'tool-call') {
+    view.toolCalls.push({ id: part.toolCallId, name: part.toolName, input: part.input })
+  } else if (part.type === 'tool-input-start') {
     // Streamed tool input — collect into a pending call (rare path).
-    view.toolCalls.push({ id: part.ToolInputStart.id, name: part.ToolInputStart.tool_name, input: '' })
-  } else if ('ToolInputDelta' in part) {
-    const pending = view.toolCalls.find((t) => t.id === part.ToolInputDelta.id)
+    view.toolCalls.push({ id: part.id, name: part.toolName, input: '' })
+  } else if (part.type === 'tool-input-delta') {
+    const pending = view.toolCalls.find((t) => t.id === part.id)
     if (pending && typeof pending.input === 'string') {
-      pending.input += part.ToolInputDelta.delta
+      pending.input += part.delta
       // Best-effort parse of accumulated JSON arguments.
       try {
         pending.input = JSON.parse(pending.input)
@@ -214,7 +210,7 @@ function applyPart(view: AgentStepView, part: StreamPart): void {
         /* keep accumulating */
       }
     }
-  } else if ('Finish' in part) {
-    view.usage = part.Finish.usage
+  } else if (part.type === 'finish') {
+    view.usage = part.usage
   }
 }
diff --git a/tools/aimux-web/web/src/api/client.ts b/tools/aimux-web/web/src/api/client.ts
index 97bc7a2ca..202f3979a 100644
--- a/tools/aimux-web/web/src/api/client.ts
+++ b/tools/aimux-web/web/src/api/client.ts
@@ -5,7 +5,7 @@ import type { WireCallResponse } from '@/types/generated/WireCallResponse'
 import type { Recording } from '@aimux/types/Recording'
 import type { TraceRecord } from '@aimux/types/TraceRecord'
 import type { SessionView } from '@aimux/types/SessionView'
-import type { StreamPart } from '@aimux/types/StreamPart'
+import type { TextStreamPart } from '@aimux/types/TextStreamPart'
 
 const API = '/api'
 
@@ -179,15 +179,15 @@ export const api = {
     return fetch(`${API}/sessions`).then((r) => j<SessionView[]>(r))
   },
 
-  sessionDetail(id: string): Promise<{ session_id: string; calls: unknown[]; recordings: Recording[] }> {
+  sessionDetail(id: string): Promise<{ sessionId: string; calls: unknown[]; recordings: Recording[] }> {
     return fetch(`${API}/sessions/${encodeURIComponent(id)}`).then((r) => j(r))
   },
 
   replay(body: {
-    call_id: string
-    api_key?: string | null
-    overrides?: { messages?: unknown[]; temperature?: number | null; max_output_tokens?: number | null } | null
-  }): Promise<WireCallResponse & { call_id: string; tool_calls?: unknown; meta?: unknown }> {
+    callId: string
+    apiKey?: string | null
+    overrides?: { messages?: unknown[]; temperature?: number | null; maxOutputTokens?: number | null } | null
+  }): Promise<WireCallResponse & { callId: string; toolCalls?: unknown }> {
     return fetch(`${API}/replay`, {
       method: 'POST',
       headers: { 'Content-Type': 'application/json' },
@@ -220,7 +220,7 @@ export const api = {
   },
 }
 
-/** Parse a `stream_part` JSON payload into a StreamPart (best-effort). */
-export function parseStreamPart(json: string): StreamPart {
-  return JSON.parse(json) as StreamPart
+/** Parse a `stream_part` JSON payload into a TextStreamPart (best-effort). */
+export function parseStreamPart(json: string): TextStreamPart {
+  return JSON.parse(json) as TextStreamPart
 }
diff --git a/tools/aimux-web/web/src/components/StepCard.vue b/tools/aimux-web/web/src/components/StepCard.vue
index ddfe7bf76..ab618403b 100644
--- a/tools/aimux-web/web/src/components/StepCard.vue
+++ b/tools/aimux-web/web/src/components/StepCard.vue
@@ -16,7 +16,7 @@ const emit = defineEmits<{ trace: [] }>()
       </Badge>
       <span v-if="step.latencyMs != null">{{ step.latencyMs }}ms</span>
       <button
-        v-if="step.meta?.call_id"
+        v-if="step.meta?.callId"
         class="underline decoration-dotted hover:text-foreground cursor-pointer"
         @click="emit('trace')"
       >
diff --git a/tools/aimux-web/web/src/types/generated/WireCallRequest.ts b/tools/aimux-web/web/src/types/generated/WireCallRequest.ts
index 9298e85c1..22822a900 100644
--- a/tools/aimux-web/web/src/types/generated/WireCallRequest.ts
+++ b/tools/aimux-web/web/src/types/generated/WireCallRequest.ts
@@ -19,13 +19,13 @@ model: string,
  * var), `Some("env:VAR")`, or a plaintext literal — the latter only
  * while the console is loopback-bound (RFC-0029 §5.5).
  */
-api_key?: string | null,
+apiKey?: string,
 /**
  * Base URL override (proxies / local endpoints).
  */
-base_url?: string | null,
+baseUrl?: string,
 /**
- * Stream the response as SSE (`StreamPart` events). Default true.
+ * Stream the response as SSE (`TextStreamPart` events). Default true.
  */
 stream: boolean,
 /**
@@ -35,8 +35,8 @@ mock: boolean, options: WireOptions,
 /**
  * Session grouping id (RFC-0024). The console reuses one id per agent run.
  */
-session_id?: string | null,
+sessionId?: string,
 /**
  * Informational step marker (the backend assigns the authoritative step).
  */
-step?: number | null, messages: Array<WireMessage>, };
+step?: number, messages: Array<WireMessage>, };
diff --git a/tools/aimux-web/web/src/types/generated/WireCallResponse.ts b/tools/aimux-web/web/src/types/generated/WireCallResponse.ts
index bd19eb41d..f1128110a 100644
--- a/tools/aimux-web/web/src/types/generated/WireCallResponse.ts
+++ b/tools/aimux-web/web/src/types/generated/WireCallResponse.ts
@@ -5,4 +5,4 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * Response for non-streaming calls.
  */
-export type WireCallResponse = { text: string, finish_reason: JsonValue, usage: JsonValue, meta?: WireMeta | null, error?: string | null, };
+export type WireCallResponse = { text: string, finishReason: JsonValue, usage: JsonValue, meta?: WireMeta, error?: string, };
diff --git a/tools/aimux-web/web/src/types/generated/WireContentPart.ts b/tools/aimux-web/web/src/types/generated/WireContentPart.ts
index c3a3fd32d..241ba4fe8 100644
--- a/tools/aimux-web/web/src/types/generated/WireContentPart.ts
+++ b/tools/aimux-web/web/src/types/generated/WireContentPart.ts
@@ -4,4 +4,8 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * A content part in the wire format.
  */
-export type WireContentPart = { "type": "text", text: string, } | { "type": "tool_call", tool_call_id: string, tool_name: string, input: JsonValue, provider_executed?: boolean | null, } | { "type": "tool_result", tool_call_id: string, result: JsonValue, is_error?: boolean | null, };
+export type WireContentPart = { "type": "text", text: string, } | { "type": "tool-call", toolCallId: string, toolName: string, input: JsonValue, providerExecuted?: boolean, } | { "type": "tool-result", toolCallId: string, toolName: string,
+/**
+ * The core type, imported by the frontend through the `@aimux/types` alias.
+ */
+output: import("@aimux/types/ToolResultOutput").ToolResultOutput, };
diff --git a/tools/aimux-web/web/src/types/generated/WireMeta.ts b/tools/aimux-web/web/src/types/generated/WireMeta.ts
index f11a6d83c..a01d15a44 100644
--- a/tools/aimux-web/web/src/types/generated/WireMeta.ts
+++ b/tools/aimux-web/web/src/types/generated/WireMeta.ts
@@ -3,4 +3,4 @@
 /**
  * SSE `meta` event payload — the frontend trace anchor (RFC-0029 §5.2).
  */
-export type WireMeta = { call_id: string, session_id?: string | null, step?: number | null, outcome: string, };
+export type WireMeta = { callId: string, sessionId?: string, step?: number, outcome: string, };
diff --git a/tools/aimux-web/web/src/types/generated/WireOptions.ts b/tools/aimux-web/web/src/types/generated/WireOptions.ts
index 7d0ae849c..28d0c10b8 100644
--- a/tools/aimux-web/web/src/types/generated/WireOptions.ts
+++ b/tools/aimux-web/web/src/types/generated/WireOptions.ts
@@ -2,8 +2,8 @@
 import type { WireTool } from "./WireTool";
 import type { JsonValue } from "./serde_json/JsonValue";
 
-export type WireOptions = { temperature?: number | null, max_output_tokens?: number | null, top_p?: number | null, stop_sequences?: Array<string> | null, tools?: Array<WireTool> | null,
+export type WireOptions = { temperature?: number, maxOutputTokens?: number, topP?: number, stopSequences?: Array<string>, tools?: Array<WireTool>,
 /**
- * `"text"` or `{"json": …}` — passed through as `ResponseFormat`.
+ * A `ResponseFormat` object: `{"type":"text"}` or `{"type":"json",…}`.
  */
-response_format?: JsonValue | null, headers?: { [key in string]: string } | null, max_retries?: number | null, include_raw_chunks?: boolean | null, };
+responseFormat?: JsonValue, headers?: { [key in string]: string }, maxRetries?: number, includeRawChunks?: boolean, };
diff --git a/tools/aimux-web/web/src/types/generated/WireTool.ts b/tools/aimux-web/web/src/types/generated/WireTool.ts
index 966a8679e..3862aad2c 100644
--- a/tools/aimux-web/web/src/types/generated/WireTool.ts
+++ b/tools/aimux-web/web/src/types/generated/WireTool.ts
@@ -4,7 +4,7 @@ import type { JsonValue } from "./serde_json/JsonValue";
 /**
  * A function tool definition (JSON Schema parameters).
  */
-export type WireTool = { name: string, description?: string | null,
+export type WireTool = { name: string, description?: string,
 /**
  * JSON Schema describing the parameters.
  */
diff --git a/tools/aimux-web/web/src/views/Agent.vue b/tools/aimux-web/web/src/views/Agent.vue
index 62cda07d4..8111b3f0f 100644
--- a/tools/aimux-web/web/src/views/Agent.vue
+++ b/tools/aimux-web/web/src/views/Agent.vue
@@ -93,7 +93,7 @@ function exportDef() {
 }
 
 function openTrace(step: AgentStepView) {
-  if (step.meta?.call_id) router.push({ path: '/traces', query: { call: step.meta.call_id } })
+  if (step.meta?.callId) router.push({ path: '/traces', query: { call: step.meta.callId } })
 }
 
 type BadgeVariant = 'default' | 'secondary' | 'success' | 'warning' | 'destructive'
diff --git a/tools/aimux-web/web/src/views/Playground.vue b/tools/aimux-web/web/src/views/Playground.vue
index 08a95cdb1..c140d9399 100644
--- a/tools/aimux-web/web/src/views/Playground.vue
+++ b/tools/aimux-web/web/src/views/Playground.vue
@@ -73,7 +73,7 @@ interface ChatItem {
   text: string
   toolCalls: Array<{ id: string; name: string; input: unknown }>
   final: boolean
-  meta?: { call_id?: string; session_id?: string | null; step?: number | null; outcome?: string }
+  meta?: { callId?: string; sessionId?: string; step?: number; outcome?: string }
   error?: string
 }
 const chat = ref<ChatItem[]>([])
@@ -96,7 +96,7 @@ function wirePart(role: 'user' | 'assistant', item: ChatItem): WireMessage {
   const content: any[] = []
   if (item.text.trim()) content.push({ type: 'text', text: item.text })
   for (const tc of item.toolCalls) {
-    content.push({ type: 'tool_call', tool_call_id: tc.id, tool_name: tc.name, input: tc.input })
+    content.push({ type: 'tool-call', toolCallId: tc.id, toolName: tc.name, input: tc.input })
   }
   return { role, content } as WireMessage
 }
@@ -123,18 +123,18 @@ async function send() {
   const body = {
     provider: provider.value,
     model: model.value,
-    api_key: apiKey.value || null,
-    base_url: baseUrl.value || null,
+    apiKey: apiKey.value || undefined,
+    baseUrl: baseUrl.value || undefined,
     stream: stream.value,
     mock: mock.value,
     options: {
       temperature: temperature.value,
-      max_output_tokens: maxTokens.value,
+      maxOutputTokens: maxTokens.value,
       tools,
-      response_format: responseFormat.value === 'json' ? { json: {} } : null,
-      headers: safeJson(headersJson.value),
+      responseFormat: { type: responseFormat.value },
+      headers: safeJson(headersJson.value) ?? undefined,
     },
-    session_id: sessionId.value,
+    sessionId: sessionId.value,
     messages: wireMessages.value,
   } as unknown as WireCallRequest
 
@@ -143,9 +143,9 @@ async function send() {
       for await (const ev of callStream(body, controller.signal)) {
         if (ev.event === 'stream_part') {
           const part = parseStreamPart(ev.data)
-          if ('TextDelta' in part) assistant.text += part.TextDelta.delta
-          else if ('ToolCall' in part)
-            assistant.toolCalls.push({ id: part.ToolCall.tool_call_id, name: part.ToolCall.tool_name, input: part.ToolCall.input })
+          if (part.type === 'text-delta') assistant.text += part.delta
+          else if (part.type === 'tool-call')
+            assistant.toolCalls.push({ id: part.toolCallId, name: part.toolName, input: part.input })
         } else if (ev.event === 'meta') {
           assistant.meta = JSON.parse(ev.data)
         } else if (ev.event === 'error') {
@@ -189,8 +189,8 @@ function safeJson(s: string): unknown | null {
 }
 
 function openTrace(item: ChatItem) {
-  if (item.meta?.call_id) {
-    router.push({ path: '/traces', query: { call: item.meta.call_id } })
+  if (item.meta?.callId) {
+    router.push({ path: '/traces', query: { call: item.meta.callId } })
   }
 }
 </script>
@@ -229,7 +229,7 @@ function openTrace(item: ChatItem) {
 
               <div v-if="item.error" class="mt-1 text-xs text-destructive">{{ item.error }}</div>
               <button
-                v-if="item.meta?.call_id"
+                v-if="item.meta?.callId"
                 class="mt-1 text-xs text-muted-foreground underline decoration-dotted hover:text-foreground cursor-pointer"
                 @click="openTrace(item)"
               >
diff --git a/tools/aimux-web/web/src/views/Replay.vue b/tools/aimux-web/web/src/views/Replay.vue
index 43d1f21f0..73f3ef212 100644
--- a/tools/aimux-web/web/src/views/Replay.vue
+++ b/tools/aimux-web/web/src/views/Replay.vue
@@ -31,12 +31,12 @@ const result = shallowRef<ResultView | null>(null)
 
 async function load() {
   traces.value = await api.traces({ limit: 200 })
-  if (!callId.value && traces.value.length) callId.value = traces.value[0].call_id
+  if (!callId.value && traces.value.length) callId.value = traces.value[0].callId
 }
 onMounted(load)
 
 function recLatency(r: Recording): number {
-  return r.exchanges?.reduce((s, e) => s + (e.timing?.latency_ms ?? 0), 0) ?? 0
+  return r.exchanges?.reduce((s, e) => s + (e.timing?.latencyMs ?? 0), 0) ?? 0
 }
 
 function oldText(r: Recording): string {
@@ -57,14 +57,14 @@ async function replay() {
   error.value = null
   running.value = true
   try {
-    const old = traces.value.find((t) => t.call_id === callId.value)
+    const old = traces.value.find((t) => t.callId === callId.value)
     if (!old) throw new Error('recording not found')
     const res = await api.replay({
-      call_id: callId.value,
-      api_key: apiKey.value || null,
+      callId: callId.value,
+      apiKey: apiKey.value || null,
       overrides: {
         temperature: temperature.value,
-        max_output_tokens: maxTokens.value,
+        maxOutputTokens: maxTokens.value,
       },
     })
     result.value = {
@@ -89,7 +89,7 @@ async function replay() {
       <span class="pb-2 text-sm font-semibold">Replay</span>
       <div class="w-72">
         <Label>选择录制（请求回放，重发真实 API）</Label>
-        <Combobox v-model="callId" :options="traces.map((t) => t.call_id)" class="mt-1" placeholder="call_id…" />
+        <Combobox v-model="callId" :options="traces.map((t) => t.callId)" class="mt-1" placeholder="call_id…" />
       </div>
       <div class="w-44">
         <Label>API key（env，explicit 源需要）</Label>
@@ -113,7 +113,7 @@ async function replay() {
 
       <template v-if="result">
         <div class="mb-3 flex items-center gap-3 text-xs text-muted-foreground">
-          <span class="font-mono">{{ result.old.call_id.slice(0, 24) }}…</span>
+          <span class="font-mono">{{ result.old.callId.slice(0, 24) }}…</span>
           <span>旧延迟 {{ recLatency(result.old) }}ms</span>
           <span>·</span>
           <span>新延迟 <span v-if="result.newLatencyMs != null">{{ result.newLatencyMs }}ms</span><span v-else>—</span></span>
diff --git a/tools/aimux-web/web/src/views/Sessions.vue b/tools/aimux-web/web/src/views/Sessions.vue
index c3abb1f56..8b1bb091b 100644
--- a/tools/aimux-web/web/src/views/Sessions.vue
+++ b/tools/aimux-web/web/src/views/Sessions.vue
@@ -31,7 +31,7 @@ async function open(id: string) {
 onMounted(load)
 
 function recLatency(r: Recording): number {
-  return r.exchanges?.reduce((s, e) => s + (e.timing?.latency_ms ?? 0), 0) ?? 0
+  return r.exchanges?.reduce((s, e) => s + (e.timing?.latencyMs ?? 0), 0) ?? 0
 }
 
 function verdictBadge(r: Recording): string {
@@ -55,18 +55,18 @@ function verdictBadge(r: Recording): string {
         <template v-else>
           <button
             v-for="s in sessions"
-            :key="s.session_id"
+            :key="s.sessionId"
             class="block w-full border-b px-4 py-3 text-left hover:bg-accent/40 cursor-pointer"
-            :class="{ 'bg-accent/60': selected === s.session_id }"
-            @click="open(s.session_id)"
+            :class="{ 'bg-accent/60': selected === s.sessionId }"
+            @click="open(s.sessionId)"
           >
             <div class="flex items-center gap-2">
-              <span class="truncate font-mono text-xs">{{ s.session_id }}</span>
+              <span class="truncate font-mono text-xs">{{ s.sessionId }}</span>
               <Badge variant="secondary" class="text-[10px]">{{ s.source }}</Badge>
               <span class="ml-auto text-xs text-muted-foreground">{{ s.calls.length }} calls</span>
             </div>
             <div class="mt-1 text-[11px] text-muted-foreground">
-              最后：{{ s.calls[s.calls.length - 1]?.recorded_at?.replace('T', ' ').slice(0, 19) ?? '—' }}
+              最后：{{ s.calls[s.calls.length - 1]?.recordedAt?.replace('T', ' ').slice(0, 19) ?? '—' }}
             </div>
           </button>
           <div v-if="!sessions.length" class="p-6 text-center text-xs text-muted-foreground">
@@ -88,12 +88,12 @@ function verdictBadge(r: Recording): string {
         <div class="space-y-2">
           <div
             v-for="r in [...recordings].sort((a, b) => (a.step ?? 0) - (b.step ?? 0))"
-            :key="r.call_id"
+            :key="r.callId"
             class="flex cursor-pointer items-center gap-3 rounded-md border px-3 py-2 hover:bg-accent/40"
-            @click="router.push({ path: '/traces', query: { call: r.call_id } })"
+            @click="router.push({ path: '/traces', query: { call: r.callId } })"
           >
             <span class="w-12 font-mono text-xs text-muted-foreground">step{{ r.step }}</span>
-            <span class="w-36 truncate font-mono text-xs">{{ r.provider?.provider }}/{{ r.provider?.model_id }}</span>
+            <span class="w-36 truncate font-mono text-xs">{{ r.provider?.provider }}/{{ r.provider?.modelId }}</span>
             <span class="text-xs text-muted-foreground">{{ recLatency(r) }}ms</span>
             <span class="ml-auto"><Badge :variant="verdictBadge(r) === 'error' ? 'destructive' : verdictBadge(r) === 'success' ? 'success' : 'secondary'">{{ r.outcome?.status }}</Badge></span>
           </div>
diff --git a/tools/aimux-web/web/src/views/Traces.vue b/tools/aimux-web/web/src/views/Traces.vue
index 1fe2835c3..b0e556299 100644
--- a/tools/aimux-web/web/src/views/Traces.vue
+++ b/tools/aimux-web/web/src/views/Traces.vue
@@ -42,7 +42,7 @@ async function load() {
       limit: 300,
     })
     if (searchCallId.value) {
-      traces.value = traces.value.filter((t) => t.call_id.includes(searchCallId.value))
+      traces.value = traces.value.filter((t) => t.callId.includes(searchCallId.value))
     }
   } finally {
     loading.value = false
@@ -73,35 +73,35 @@ function statusVariant(s: string): 'default' | 'success' | 'destructive' | 'warn
 }
 
 function recLatency(r: Recording): number {
-  return r.exchanges?.reduce((s, e) => s + (e.timing?.latency_ms ?? 0), 0) ?? 0
+  return r.exchanges?.reduce((s, e) => s + (e.timing?.latencyMs ?? 0), 0) ?? 0
 }
 
 function recTokens(r: Recording): number | null {
-  const u = r.outcome?.usage as { input_tokens?: { total?: number }; output_tokens?: { total?: number } } | null | undefined
+  const u = r.outcome?.usage as { inputTokens?: { total?: number }; outputTokens?: { total?: number } } | null | undefined
   if (!u) return null
-  const input = u.input_tokens?.total ?? 0
-  const output = u.output_tokens?.total ?? 0
+  const input = u.inputTokens?.total ?? 0
+  const output = u.outputTokens?.total ?? 0
   return input + output
 }
 
 function timeLabel(r: Recording): string {
-  return r.recorded_at.replace('T', ' ').replace('Z', '').slice(0, 19)
+  return r.recordedAt.replace('T', ' ').replace('Z', '').slice(0, 19)
 }
 
 function providerLabel(r: Recording): string {
-  return `${r.provider?.provider ?? '?'}/${r.provider?.model_id ?? '?'}`
+  return `${r.provider?.provider ?? '?'}/${r.provider?.modelId ?? '?'}`
 }
 
 const sessionItems = computed<WaterfallItem[]>(() => {
-  const sid = selected.value?.session_id
+  const sid = selected.value?.sessionId
   if (!sid) return []
   return traces.value
-    .filter((t) => t.session_id === sid)
+    .filter((t) => t.sessionId === sid)
     .map((t) => ({
       step: t.step ?? 0,
-      label: `${t.provider?.model_id ?? 'model'}`,
+      label: `${t.provider?.modelId ?? 'model'}`,
       ms: recLatency(t),
-      ttfb: t.exchanges?.[0]?.timing?.ttfb_ms ?? null,
+      ttfb: t.exchanges?.[0]?.timing?.ttfbMs ?? null,
       status: String(t.outcome?.status ?? '').toLowerCase(),
     }))
 })
@@ -131,9 +131,9 @@ async function doImport() {
 
 function traceMetaOf(r: Recording): WireMeta {
   return {
-    call_id: r.call_id,
-    session_id: r.session_id ?? null,
-    step: r.step ?? null,
+    callId: r.callId,
+    sessionId: r.sessionId,
+    step: r.step,
     outcome: String(r.outcome?.status ?? '').toLowerCase(),
   }
 }
@@ -173,15 +173,15 @@ function traceMetaOf(r: Recording): WireMeta {
             <template v-else>
               <tr
                 v-for="r in traces"
-                :key="r.call_id"
+                :key="r.callId"
                 class="cursor-pointer border-t transition-colors hover:bg-accent/40"
-                :class="{ 'bg-accent/60': selected?.call_id === r.call_id }"
+                :class="{ 'bg-accent/60': selected?.callId === r.callId }"
                 @click="selected = r"
               >
                 <td class="px-3 py-2 font-mono text-xs text-muted-foreground">{{ timeLabel(r) }}</td>
                 <td class="px-3 py-2 font-mono text-xs">{{ providerLabel(r) }}</td>
                 <td class="px-3 py-2 font-mono text-xs text-muted-foreground">
-                  {{ r.session_id ? `${r.session_id.slice(0, 16)}…` : '—' }}<span v-if="r.step != null"> / {{ r.step }}</span>
+                  {{ r.sessionId ? `${r.sessionId.slice(0, 16)}…` : '—' }}<span v-if="r.step != null"> / {{ r.step }}</span>
                 </td>
                 <td class="px-3 py-2 text-right font-mono text-xs">{{ recLatency(r) }}ms</td>
                 <td class="px-3 py-2 text-right font-mono text-xs">{{ recTokens(r) ?? '—' }}</td>
@@ -201,14 +201,14 @@ function traceMetaOf(r: Recording): WireMeta {
     <!-- detail -->
     <aside v-if="selected" class="flex w-[30rem] shrink-0 flex-col border-l">
       <div class="flex items-center gap-2 border-b px-3 py-2">
-        <span class="truncate font-mono text-xs">{{ selected.call_id }}</span>
+        <span class="truncate font-mono text-xs">{{ selected.callId }}</span>
         <div class="flex-1" />
         <Button variant="ghost" size="sm" @click="selected = null">✕</Button>
       </div>
 
       <div class="border-b p-3">
         <div class="mb-2 text-xs font-semibold text-muted-foreground">会话瀑布</div>
-        <Waterfall :items="sessionItems.length ? sessionItems : [{ step: selected.step ?? 0, label: providerLabel(selected), ms: recLatency(selected), ttfb: selected.exchanges?.[0]?.timing?.ttfb_ms ?? null, status: String(selected.outcome?.status ?? '').toLowerCase() }]" />
+        <Waterfall :items="sessionItems.length ? sessionItems : [{ step: selected.step ?? 0, label: providerLabel(selected), ms: recLatency(selected), ttfb: selected.exchanges?.[0]?.timing?.ttfbMs ?? null, status: String(selected.outcome?.status ?? '').toLowerCase() }]" />
       </div>
 
       <div class="flex-1 overflow-auto p-3">
@@ -232,14 +232,14 @@ function traceMetaOf(r: Recording): WireMeta {
               <Badge variant="secondary">attempt {{ ex.attempt }}</Badge>
               <span class="text-muted-foreground">{{ ex.request?.method }} {{ ex.request?.url }}</span>
               <span v-if="ex.response" class="text-muted-foreground">→ {{ ex.response.status }}</span>
-              <span class="ml-auto text-muted-foreground">{{ ex.timing?.latency_ms ?? '—' }}ms</span>
+              <span class="ml-auto text-muted-foreground">{{ ex.timing?.latencyMs ?? '—' }}ms</span>
             </div>
             <Card class="p-2 mb-1">
               <div class="mb-1 text-[11px] text-muted-foreground">请求 body</div>
               <JsonViewer :data="ex.request?.body" />
             </Card>
             <Card class="p-2">
-              <div class="mb-1 text-[11px] text-muted-foreground">响应 body{{ ex.response?.stream_chunks != null ? `（${ex.response.stream_chunks} chunks）` : '' }}</div>
+              <div class="mb-1 text-[11px] text-muted-foreground">响应 body{{ ex.response?.streamChunks != null ? `（${ex.response.streamChunks} chunks）` : '' }}</div>
               <JsonViewer :data="ex.response?.body" />
             </Card>
           </div>