refactor(core)!: model the provider prompt by role, with one file part

[cunninghamcard-bit]

What

Replace the provider-facing prompt struct with role-specific messages and parts based on LanguageModelV4Prompt and LanguageModelV4Message in @ai-sdk/provider. This removes role/part combinations that providers cannot accept and gives file input and tool output explicit unions. This branch follows rfc-0036/provider-options-types and precedes rfc-0036/provider-result-types; vendor converters, recordings, replay consumers and generated Node types are migrated with it.

Before and after

Concept and upstream source	Before	After
LanguageModelV4Message in language-model-v4-prompt.ts	One LanguageModelPromptMessage struct using the call-layer ContentPart union for every role	LanguageModelMessage has system, user, assistant and tool variants; system content is a string and the other roles have distinct part unions
LanguageModelV4FilePart	Image, binary file, base64 file, URL file and reference file variants reach provider converters separately	One FilePart uses tagged FileData; the call-layer conversion folds the existing file variants into it
LanguageModelV4ToolResultPart and LanguageModelV4ToolResultOutput	Free JSON result and error flags	Required tool_name and structured output: text, json, execution-denied, error-text, error-json or content; content supports text, file and custom parts
LanguageModelV4ReasoningPart and LanguageModelV4ToolCallPart	Dedicated signature fields and a with_signature compatibility path	Provider prompt signatures use provider options; the compatibility path is removed and conversion forwards those options without overriding them
Assistant and tool unions in language-model-v4-prompt.ts	No reasoning-file, custom or tool-approval-response provider parts	Assistant supports reasoning-file and custom; tool messages support tool-approval-response; part tags use upstream's hyphenated names
convert-to-openai-chat-messages.ts and shared chat consumers	Non-object OpenAI tool inputs serialized directly; Groq tool output could consume OpenAI cache options; references always used the OpenAI namespace	OpenAI non-object tool input becomes {}; cache breakpoints are OpenAI-only; DeepSeek resolves its own references and Groq/Compatible reject references
convert-to-openai-responses-input.ts	User references could become empty text; denied approval output could produce another function-call output	References become file IDs; execution-denied output carrying an approval ID is skipped
convert-to-google-messages.ts	Tool content files could become text; original cloud-storage URLs were ignored	Supported Vertex tool content URLs become fileData; gs: URLs preserve original_url; Vertex user references are rejected
Bedrock, Cohere, Mistral and xAI prompt conversion modules	Unsupported inputs could be dropped, coerced or sent under the wrong file shape	Bedrock emits S3 image locations and rejects unsupported tool-content media types; Cohere and Mistral reject unsupported references and document inputs and resolve partial image media types; xAI rejects inline non-image files

convert_to_language_model_prompt rejects parts that are invalid for their role, non-text system content and tool results without a name or a preceding named call. The call-layer message representation remains separate from these provider types. Recordings now serialize system content as a string and all provider file parts through the single file shape.

The ${string}.${string} constraint on custom kinds and the URL type are TypeScript-level constraints, not runtime validation in the upstream type module. Rust keeps String; JSON tool values remain Value.

Tests

Remove hand-written mock-server tests built around the old prompt struct, the hand-written stream-error and tool-result/reasoning suites, and session-inference tests that pin Rust-specific behavior. Keep and migrate upstream-ported tests and cassette replay tests to the role-specific prompt and structured tool output. Restore the upstream Groq tool-argument and reasoning cases, Google model-path cases and Mistral strict-tool case that had been removed with broader test files. No cassette or fixtures/ file changes are part of this branch.

The Groq file-reference test asserts upstream's error text, file parts with provider references.

Not in this pull request

• Serde naming and optionality remain pending a maintainer decision: snake_case fields, PascalCase external tags in the existing shared wire types, Option fields serialized as null where omission is not configured, and generated Node fields that are required-nullable. In particular, message-level provider_options remains required-nullable in generated Node types while upstream providerOptions is optional. This branch does not decide the binding-wide wire contract.

• The call-layer ModelMessage tool-result shape still uses free JSON and flags. Its replacement and the wider call-layer message/stream separation need the maintainer's decision; provider prompt types do not settle that contract.

• Bedrock user-file media detection and validation are fixed on rfc-0036/vendor-alignment, which owns the broader vendor behavior changes. This branch's tool-content validation does not establish full user-file conversion parity.

• OpenAI Responses approval-ID mapping in generated output and custom-tool async replay are fixed on rfc-0036/vendor-alignment; those belong to the later vendor result and tool behavior work.

• A complete upstream test-suite port is outside this prompt migration. Keeping and restoring selected upstream cases does not establish complete coverage of each vendor converter.

• aimux-providers/tests/anthropic_remaining_test.rs, a translation of upstream Anthropic cases, is deleted here; it comes back adapted and un-ignored on vendor-alignment, which also implements the behaviour its cases cover.

Stack refresh (2026-10-07)

Head df272b11 follows rfc-0036/provider-options-types at c22dcfad. The stack includes master at 0b365b39 and the updated #203 (3e2d9fc7), #204 (b897cb21) and #213 (7c0933b1). Updated parents were merged into their children to preserve the existing public merge history; the PR base and Draft state are unchanged.

Verification

• Relative to the updated immediate base: 93 files changed, 4177 insertions(+), 12601 deletions(-).
• Rust formatting and workspace Clippy passed during this layer’s refresh. Final integration at local wire-format 415e37e1 passes cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings, and the provider boundary script.
• Full cargo test --workspace at final wire-format: 2,819 passed, 0 failed, 6 ignored, including doctests. The intermediate factory-source integration also passed 2,434 tests before the final obsolete Decision hook removal. These are integration results, not a claim that the full suite was rerun at every PR head.
• Generated TypeScript synchronization, Node build:typed, and web production build passed at the wire-format layer.

Not verified

No GitHub check runs are attached to this Draft stack head at the time of this update. Go, Java, Kotlin, Swift and Flutter bindings were not compiled locally in this refresh; Node npm test and Python pytest were not run. Live vendor calls and unported upstream cases remain outside this validation.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WGwanvo9WLU9HWD7sRR8TS

[cunninghamcard-bit]

Folded into #219 while consolidating the RFC-0036 stack from twelve pull requests into four. This branch is unchanged; its commits are part of the range #219 now shows (base rfc-0036/integration).

🤖 Generated with Claude Code