[cunninghamcard-bit]

Requested by li

Summary

Before: core content, provider options, provider metadata and the provider prompt each have aimux-specific shapes, and vendors read options as untyped JSON. User messages are a role plus one content-part enum shared by every role, with no message-level provider options; reasoning parts carry a top-level signature; text results reuse a content shape generic over the provider layer. Provider namespace maps are unordered, and a custom content kind is any string. The prompt goes to the model without upstream's checks, so a tool call with no result or a response that breaks the tool choice is not reported. The router unit tests and the error variant-set test are gone, and a scratch MIGRATION-NS.md sits in the repository root.

After: user messages are the AI SDK's ModelMessage: a union by role, each role with its own content parts and message-level providerOptions, including tool approval responses and the deprecated tool output items (accepted with a deprecation warning). generate_text and stream_text standardize the prompt as upstream does: empty messages are an invalid prompt, consecutive tool messages are combined, a tool call without a result fails with MissingToolResults, and a response that breaks the tool choice fails with ToolChoiceViolation. Text results carry the AI SDK's core ContentPart (11 kinds, including tool errors and approval requests and responses), separate from the V4 provider content. There is one source, file and reasoning type per layer, and reasoning has no top-level signature. Provider options and metadata have one typed, insertion-ordered shape that each vendor package builds and reads; a non-object option is an error and unknown keys no longer leak. The provider prompt is modelled by role with one file part, using upstream's hyphenated part tags, and references, files and tool output convert as upstream does. A custom content kind must be ${string}.${string}, and CallOptions no longer carries a call id. FFI, bindings and their generated types follow; the router and variant-set tests are back.

Type of change

• Bug fix (non-breaking)
• New feature (non-breaking)
• Breaking change
• New provider
• New / updated binding
• Docs / RFC

Checklist

• cargo fmt --all -- --check passes
• cargo clippy --workspace --all-targets -- -D warnings passes
• cargo test --workspace passes (cassette playback — no network/keys)
• Added or updated cassette tests / contract fixtures for behavior changes
• Updated relevant docs (README.md, docs/, rfc/, bindings/README.md)
• Commit messages follow Conventional Commits

Notes for reviewers

• Position: base master, with PR 1 ([1/8] chore: CI on every PR and test fixture cleanup #249) and PR 2 ([2/8] refactor!: tool calls follow the AI SDK #250) merged in; next is PR 4 ([4/8] refactor!: results and streams follow the AI SDK #252).
• Supersedes: [v4/p3-messages] refactor(core)!: messages, content, provider options and prompts follow the AI SDK V4 types #237, which GitHub closed when its branch was renamed to match the title.
• Replaces: part of Draft refactor(core)!: align public core types with the AI SDK V4 shapes #219.
• Aligns to: @ai-sdk/provider-utils 5.0.53 types/ (model messages and content parts; unchanged in 5.0.55); ai 7.0.127 prompt/convert-to-language-model-prompt, prompt/standardize-prompt, generate-text/content-part, to-response-messages; @ai-sdk/provider 4.0.21 shared provider options, metadata and references.
• Closes: V4 alignment: core types, stream parts, results and call pipeline vs ai 7.x #229 B8, C31, C32, C36, C37, C38, C42, NP2-2 (custom content kind); V4 alignment: provider factory, HTTP and auth #230 B20; V4 alignment: public JSON, tools and lost tests #232 B7, B23 (router tests), N31-7.1-1.
• In part:
  • V4 alignment: core types, stream parts, results and call pipeline vs ai 7.x #229 NP2-1: the Rust types keep insertion order here; the bindings keep it in PR 8.
  • V4 alignment: core types, stream parts, results and call pipeline vs ai 7.x #229 C41: the core tool-error and approval parts are here; the stream parts are in PR 4.
  • V4 alignment: provider factory, HTTP and auth #230 C7: call_id and the step helper leave CallOptions here; the retry count, timeout and session id follow in PR 4.
• Fork fixes included: none.
• Not in this PR:
  • V4 alignment: core types, stream parts, results and call pipeline vs ai 7.x #229 C38, two upstream checks: system messages inside messages are still accepted, and instructions stays a string. Both would change the JSON that the bindings, CLI and web console send, so they wait for the binding migration.
  • Dedicated binding error classes for the two new errors also wait for the binding migration. Until then the C ABI reports them as AIMUX_E_INVALID_PROMPT and AIMUX_E_INVALID_RESPONSE_DATA, and Node and Python raise InvalidPromptError and InvalidResponseDataError.
  • tool-output-denied is only a stream part upstream, so it comes with the stream parts in PR 4. So do provider tool results with no matching call (NP2-3) and the stream tool error for an invalid call.
  • title and toolMetadata: aimux has no user-level tool set (as in PR 2).
  • Only the deprecated tool output items are converted here; URL downloads for user files come in PR 5.
• Behaviour changes (breaking):
  • content::ContentPart, MessageContent and Role are replaced by UserContentPart, AssistantContentPart, ToolContentPart and their content types. The user tool result carries a typed output; the earlier result / is_error JSON is still read.
  • GenerateTextResult.content is Vec<ContentPart> (the core part). TextContent is gone in Rust; Python keeps the name for the new union.
  • ReasoningOutput has no signature, and extract_reasoning_signature is removed. Signatures live in each provider's own namespace.
  • SharedProviderOptions and SharedProviderMetadata are one NamespaceMap (IndexMap<String, JsonObject>); SharedProviderReference is ordered as well.
  • convert_to_language_model_prompt takes the provider id as a third argument.
  • CallOptions.call_id is removed; the id comes from the recording context. CallOptions::for_step is now crate-private.
  • New errors MissingToolResults and ToolChoiceViolation, mapped in the FFI, Node and Python.
• Checks: fmt, clippy and cargo test --workspace pass; scripts/gen_ts_types.py --check is clean; the Node binding type-checks with tsc. The other bindings treat result content as opaque JSON; their suites run in CI.
• Merged RFC / docs: NamespaceMap order and aliases (NP2-1); C32 line refs; C37 deprecation notes; C38 ToolChoiceViolation, TOOLDEF-3 line refs and empty tool input {}; part 2 Q1 scope note, the §3.5 speech and transcription tracing note, and the tool-call acceptance row; docs/api/node.md, docs/api/rust.md; CHANGELOG Breaking bullets.
• Restored tests: the trace layer's tests were deleted in an earlier cut of this PR because they constructed the old result types; bfd9fdd restores all 10 on the new types (the time-to-first-output test now checks the value instead of matching null) and adds 4: the call id comes from the recording context on the success, stream and error paths, body/id/headers come from the result, a file part counts as the first output, and the raw usage object is kept.
• Size: release artifacts before and after (CI artifacts, zip size):

CI artifact (zip)	master a8e3a99	[2/8] 496ef24	this PR 35711cb	change vs [2/8]	change vs master
ffi-x86_64-unknown-linux-gnu	3.77 MB	3.78 MB	4.01 MB	+232.0 KiB (+6.29%)	+243.0 KiB (+6.61%)
ffi-aarch64-apple-darwin	3.12 MB	3.13 MB	3.31 MB	+183.2 KiB (+6.00%)	+190.5 KiB (+6.26%)
ffi-x86_64-pc-windows-msvc	3.03 MB	3.03 MB	3.21 MB	+176.5 KiB (+5.96%)	+181.0 KiB (+6.13%)
node-x86_64-unknown-linux-gnu	55.30 MB	55.60 MB	57.58 MB	+1938.5 KiB (+3.57%)	+2232.3 KiB (+4.13%)
node-aarch64-unknown-linux-gnu	53.25 MB	53.55 MB	55.42 MB	+1834.4 KiB (+3.51%)	+2121.4 KiB (+4.08%)
node-aarch64-apple-darwin	15.06 MB	15.11 MB	15.91 MB	+778.4 KiB (+5.28%)	+823.7 KiB (+5.60%)
node-x86_64-apple-darwin	13.60 MB	13.63 MB	14.36 MB	+713.4 KiB (+5.36%)	+743.5 KiB (+5.60%)
node-x86_64-pc-windows-msvc	10.57 MB	10.59 MB	11.21 MB	+604.8 KiB (+5.85%)	+623.7 KiB (+6.04%)
node-aarch64-pc-windows-msvc	9.87 MB	9.90 MB	10.48 MB	+569.4 KiB (+5.89%)	+592.7 KiB (+6.15%)

[cunninghamcard-bit]

eric 的意见	我们的处理	位置
OpenAI 输出适配的四组端到端测试被删，没有替代（阻塞）	五个测试原样恢复，只把 mock 换成当前的结果类型；另加原文保真矩阵（非法 JSON、JSON 字符串、null、带空白、未知工具，各有无 input delta，非流式和流式）（7a726e33）	aimux-core/tests/openai_output_tool_input_test.rs
trace 层回归套件被整组删除，而 call_id 来源恰好改了（阻塞）	按新接口恢复全部 10 个测试，另加 4 个：call_id 取自录制上下文（成功、流式、失败三条路径）、body/id/headers 取自结果、文件部件算首个输出、raw usage 原样保留（bfd9fddd）。恢复时发现并修了一个问题：正常结束的 stream_text 被 trace 记成"流提前丢弃"并跳过 LCP 审计，现在在 Finish 到达时记录（4b212066）	aimux-core/tests/cache_probe_test.rs
metadata 保真缺测试	加了测试：未知 namespace、空 namespace、多层嵌套 null，经文本和流式推理两条路径写回 providerOptions 后和原值完全一致（88ce267f）	aimux-core/src/response_messages.rs 的测试模块
Gemini inlineData 的测试和 cassette	三个测试恢复，两个 cassette 不再是孤儿（62e29675）	aimux-providers/tests/data_loss_regression_test.rs
FFI 的组合、帧错误、会话输入测试	恢复（750f9a0b）	aimux-ffi/src/lib.rs 的测试模块
ToolResultPartJson 还在写旧的 result / is_error，请写明何时移除	这是给绑定旧 JSON 的私有兼容，[8/8] 已经删掉读写两侧，ToolResultPart 只有 output（272a70b2）	[8/8] aimux-core/src/content.rs
reasoning 的 provider_metadata 像是被吞了	没有被吞；那个 #[allow(unused)] 是 master 遗留的属性，[8/8] 已删	—
疑问：上游用户层 tool-result 还接受 result / isError 吗	不接受。provider-utils 5.0.55 的 content-part.ts:218-241 里用户层 ToolResultPart 只有 output；result/isError 只在 provider 层的 LanguageModelV4ToolResult	—
疑问：call_id 改来源后，录制回放链路有没有自动化覆盖	有。新增 trace_recording_correlation_test.rs，在全局录制开启时经 TraceLayer 跑 generate_text 和 stream_text，断言 trace 用的就是录制的 id、能按 id 取到录制、回放时发出同样的请求（6faea338）；四个 replay_with_model 单测也恢复了（9fd89058）	aimux-core/tests/trace_recording_correlation_test.rs

---

Generated by Claude Code

[eric8810]

复核通过，撤回修改请求。逐项核实：openai_output_tool_input_test.rs 恢复（5 个原测试 + 原文保真矩阵）✓；trace 层 10 个测试按新接口恢复并新增 4 个（call_id 取自录制上下文的三条路径、结果取值、文件部件、raw usage）✓；恢复过程中发现并修复"正常结束的 stream_text 被记为流提前丢弃、跳过 LCP 审计"的真实缺陷（4b212066）——这正说明该套件值得恢复 ✓；metadata 未知 namespace/嵌套 null 保真测试、Gemini inlineData 测试与 cassette、FFI 测试、trace_recording_correlation_test.rs（录制回放链路自动化）全部落地 ✓；ToolResultPartJson 旧写侧在 [8/8] 删除、上游 tool-result 语义疑问已答（用户层仅 output）✓。批准。