fix(providers)!: align ten vendor packages with their upstream counterparts

[cunninghamcard-bit]

What

Align vendor request conversion, response conversion, authentication and streaming with the pinned AI SDK sources. This pull request (#220), on rfc-0036/vendor-alignment, follows rfc-0036/provider-factory (#214) in the stack and uses the protocol types and factory API introduced below it. It ports previously missing vendor features and removes conflicting behavior; the remaining model families and public-surface decisions are listed below. The final branch diff against rfc-0036/provider-factory is 155 files changed, 19,410 insertions and 6,336 deletions.

Before and after

Area and upstream module	Before	After
OpenAI OpenAIEmbeddingModel	Embeddings accepted base64 vector strings.	Requests encoding_format: "float" and reads numeric vectors.
OpenAI OpenAIResponsesLanguageModel and convertToOpenAIResponsesInput	Provider tool requests, events and history were incomplete.	Adds typed provider-tool factories, request and event conversion, approval requests and compaction custom content. Approval IDs map back to the original tool-call IDs in generate and stream; custom-tool history retains async. The local shell environment accepts an omitted type tag, as in upstream shell.ts.
Azure provider and transcription modules	Transcription used only the OpenAI route.	Selects OpenAI or Azure Speech transcription, including the Speech endpoint, authentication headers and base-URL override.
Anthropic AnthropicLanguageModel and message conversion	Document citations and container uploads were dropped; structured-output fallback and computer toolset mapping were incomplete.	Returns document sources and container-upload custom content, supports the JSON-tool fallback, and maps computer toolset members to the caller's tool name with action and toolsetName. History retains the toolset mapping. Mid-stream errors are forwarded without a synthetic finish.
AmazonBedrockChatLanguageModel and message conversion	Structured JSON handling and provider-defined Anthropic tools were missing; inline media could be rejected before detection.	Implements native response format, JSON-tool and prompt-instruction paths, with generate/stream extraction and JSON-tool finish-reason and metadata handling. Adds provider-defined tools and model-family selection. Resolves inline media before format validation, percent-encodes model IDs, and emits metadata under amazonBedrock and bedrock. Response bodies are omitted as upstream does.
Google GoogleLanguageModel and message conversion	Reasoning-file replay was rejected; null function-call arguments could become "null", and null partialArgs could discard a call.	Preserves generated files, reasoning files and their replay metadata. Null args becomes {} without a parameter delta; null partialArgs permits the complete-call path. Provider response model IDs are left unset.
GoogleVertexProvider and GoogleVertexAnthropicProvider	Authentication required an explicit or environment token, and Anthropic models also had an entry on the main Vertex provider.	Adds application default credential resolution and googleAuthOptions, plus a separate Vertex Anthropic factory. Removes the main provider's anthropic_model entry. Vertex Anthropic uses googleVertex metadata and excludes Express mode.
Vertex language, image and transcription modules	Partial function arguments and thought files were incomplete; Speech-to-Text requests lost authentication headers.	Accumulates partial function arguments, preserves thought files, rejects provider references in user files, and carries credentials to Speech-to-Text. Adds Gemini transcription alongside Speech-to-Text. Image generation follows the upstream Gemini path; URL image input remains rejected by media-type resolution, matching the pinned upstream path.
Mistral MistralChatLanguageModel and message conversion	Prompt and tool-result conversion used separate local helpers.	Consolidates conversion, preserving the tool result's name and using the shared tool-output conversion.
Cohere CohereChatLanguageModel	A stream ending without usage could panic.	Returns a parse error for missing message-end usage.
xAI XaiResponsesLanguageModel	Image-generation tool calls and results were dropped.	Returns their calls and results in generate and stream.
Groq GroqChatLanguageModel	Rejects a first SSE error before returning a stream.	The final implementation still rejects it. The retained upstream-ported test now expects stream-start, error and finish, leaving an implementation/test mismatch to fix.

Where upstream expresses constraints only in TypeScript types, such as dotted IDs and kinds, schema types, non-null types, URL/Date values and numeric domains, it does not enforce those constraints at runtime either. Rust retains Value, String and u32 where those representations apply.

Tests

• Keeps existing upstream-ported tests and adjusts them to the upstream behavior and current factory API. Restores the upstream expectation for the Groq first-error SSE sequence and the Bedrock end-to-end toolChoice: "auto" input and assertion. The OpenAI streaming selection asserts the upstream event sequence, and the structured-output selection uses the upstream model's capability path.
• Adds selected upstream cases in the vendor upstream_e2e_test.rs files through the public factories with an injected transport. These cover selected generate, stream, tool, error and newly ported feature paths; they are not complete upstream suites.
• Deletes hand-written tests or assertions that preserved removed Rust behavior, including Anthropic assertions requiring a synthetic finish after a mid-stream error. Remaining upstream ports and cassette replays are retained or updated for the new behavior.
• Retains cassette regression coverage, including binary Bedrock body_base64 stream replay, and restores Google generated-file regression coverage from recordings. Deletes the OpenAI embedding replays whose recordings contain base64 vectors and the Bedrock conformance streaming replay whose text-encoded binary data cannot exercise the upstream decoder. Recording files are unchanged; rerecording is a maintainer decision.
• Neither aimux-providers/tests/cassettes/ nor fixtures/ changes in this branch diff.

Later commits on this branch, after the review fixes: Groq and the openai-compatible chat model no longer read the first SSE event before returning the stream; an error event flows as stream-start, error, finish as upstream does. Bedrock gains the S3 supported-url entry, sends an empty tool-use input as {}, and its migrated image and embedding tests mount the percent-encoded model path. The two upstream-translated files bedrock_remaining_test.rs (38 cases) and anthropic_remaining_test.rs (21 cases), deleted lower in the stack, are back, adapted and no longer ignored; Anthropic gains the mid-conversation tool change handling, argument validation, leading system message handling, beta headers and the empty base-url error text those cases cover. Six Bedrock cases are left out: three whose upstream expectation changed, two environment cases without an upstream counterpart, one duplicated by the e2e selection.

Not in this pull request

• Google speech, transcription and evaluation; xAI image, files, speech, transcription and video; Groq transcription; Mistral speech and transcription; OpenAI skills; and registry evaluationModel and customProvider are not ported. These require additional model families or interfaces, pending the maintainer's decision. xAI Responses image-generation tools do not provide a standalone image-model family.

• Serde naming remains a maintainer decision: snake_case fields, PascalCase external tags, Option serialized as null, and generated Node fields that are required and nullable. This pull request does not settle the binding wire format.

• The by-name entries in aimux_providers::provider, used by FFI, Node and Python, remain pending the maintainer's decision. Package factory changes do not settle those entries.

• The call-layer ModelMessage tool-result shape and remaining tool-result/error behavior remain pending the maintainer's decision. The distinct call-layer stream enum introduced below this branch does not settle those remaining behavior choices.

• The Google authentication port covers credential resolution and token exchange used by the Vertex factories. It does not reproduce every upstream auth-library transporter option, retry policy, environment probe or project-discovery path; those library behaviors require further work.

• Groq first-error SSE handling remains unresolved in the final branch: do_stream still rejects before constructing the stream, although its upstream-ported test now expects stream-start, error and finish. The implementation must be fixed to satisfy that restored expectation.

• The selected vendor tests do not port complete upstream suites. Groq options include runtime schema tests as well as a TypeScript-only inference case; options, usage, message conversion, tool preparation and generate/stream cases remain unported. Completing those ports requires retaining the upstream inputs and assertions.

• The OpenAI chat and Responses models still read the first SSE event before returning the stream (aimux-providers/src/openai/model.rs, openai/responses/mod.rs); the same change as for Groq is owed there. No ported test covers it yet.

Not verified

This documentation review did not run compilation, tests or runtime gates. Node npm test and Python pytest were not run. Other language bindings edited earlier in the stack were not compiled locally. CI YAML was only syntax-checked, which does not establish CI execution. Live credential exchange, vendor requests and the unported upstream cases are not verified here.

Verification

Gate on the final commit 5dc85be6: cargo fmt --check, cargo clippy --workspace --all-targets -D warnings, cargo doc, tests of core, provider-utils, providers, ffi, web, replay and cli: 2,514 passed, 0 failed; Node and Python bindings cargo check; provider boundary script; gen_ts_types.py --check, gen_providers_doc.py --check. Cassettes and fixtures/aisdk unchanged relative to the base. 158 files, +21206 / −6464.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WGwanvo9WLU9HWD7sRR8TS