fix(messages): stop forwarding Anthropic-only fields to OpenAI-compatible upstreams

[jarvis9443]

Problem

When an Anthropic SDK / Claude Code client calls /v1/messages targeting a model whose upstream is OpenAI/Azure OpenAI, Anthropic-only top-level request fields leak into the upstream request body and the upstream rejects them:

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"Unknown parameter: 'context_management'."}}

The chain: parse_inbound_request() copies every non-whitelisted top-level key into ChatFormat.extra; cross_provider_dispatch() only translates tools/tool_choice; OpenAiRequest.extra is #[serde(flatten)], so context_management (and thinking, top_k, mcp_servers, …) land verbatim on the OpenAI wire.

Fix

Replace the tools/tool_choice-only fixup in cross_provider_dispatch with a whitelist translation, translate_extras_to_openai_shape() in the anthropic provider crate:

• tools / tool_choice → existing OpenAI-shape translation (bug: /v1/messages silently drops tools and tool_choice when upstream is OpenAI-compatible #236 behavior unchanged)
• stop_sequences → stop
• metadata.user_id → user (LiteLLM parity)
• thinking → reasoning_effort (LiteLLM-aligned buckets: budget_tokens ≥4096 high / ≥2048 medium / ≥1024 low / below minimal; adaptive → medium; disabled → omitted)
• everything else (context_management, top_k, mcp_servers, container, service_tier, betas, …) is dropped (debug-logged), so future Anthropic SDK fields cannot re-trigger the 400

This mirrors the whitelist-and-drop policy the /v1/responses bridge (#825) already applies in the opposite direction. Both streaming and non-streaming go through the same ChatFormat build, so one fix covers both.

Behavior changes / compatibility

• Anthropic-native /v1/messages passthrough (Anthropic upstream) is untouched — bodies still forward verbatim, context_management keeps working there.
• Bedrock/Vertex bridges only ever read tools/tool_choice from extras, so dropping the rest changes nothing for them.
• LiteLLM baseline (HEAD 88e03e5): its Anthropic→OpenAI adapter is also whitelist-based and maps metadata.user_id→user and thinking→reasoning_effort; it translates context_management into OpenAI Responses-API compaction (or polyfills it in-gateway on the chat path) rather than dropping — that fuller treatment is deferred to a follow-up issue. stop_sequences→stop goes beyond LiteLLM (which silently drops it) since chat/completions supports stop natively.

Tests

• Unit tests for translate_extras_to_openai_shape (drop set, mappings, thinking buckets).
• New E2E messages-cross-provider-extras-e2e.test.ts (non-streaming + streaming): asserts Anthropic-only fields never reach the mock OpenAI upstream while translated fields survive. Fails before this fix (context_management leaked), passes after.

Fixes api7/AISIX-Cloud#953

Summary by CodeRabbit

• New Features

  • Expanded cross-provider request translation so Anthropic message extras are converted into an OpenAI-compatible shape more consistently.
  • Added support for mapping additional fields such as stop sequences, user metadata, and thinking-related settings.

• Bug Fixes

  • Removed Anthropic-only fields from upstream requests to prevent unsupported parameters from being forwarded.
  • Improved handling for both streaming and non-streaming message requests across providers.

• Tests

  • Added end-to-end coverage for cross-provider extras translation and field preservation/removal behavior.

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR adds a translate_extras_to_openai_shape function in the Anthropic provider crate that whitelists and translates Anthropic-specific request extras (tools, tool_choice, stop_sequences, metadata.user_id, thinking) into OpenAI-compatible fields, dropping unrecognized Anthropic-only fields. It replaces prior manual tools/tool_choice translation in cross_provider_dispatch with a single call to this helper, and adds unit and e2e tests validating the behavior.

Changes

Anthropic-to-OpenAI extras translation

Layer / File(s)	Summary
Translation helper and re-export crates/aisix-provider-anthropic/src/wire.rs, crates/aisix-provider-anthropic/src/lib.rs	Adds translate_extras_to_openai_shape (whitelisting tools/tool_choice/stop_sequences/metadata.user_id/thinking, dropping other fields with a debug log) and reasoning_effort_from_thinking (bucketing thinking config into reasoning_effort levels); re-exports the new function from the crate root.
Unit tests for translation crates/aisix-provider-anthropic/src/wire.rs	Adds tests covering dropping of Anthropic-only fields, correct translation of whitelisted fields, metadata without user_id, and thinking bucketing across enabled/adaptive/disabled variants.
Cross-provider dispatch wiring crates/aisix-proxy/src/messages.rs	Replaces manual tools/tool_choice extraction and translation logic in cross_provider_dispatch with a single call to translate_extras_to_openai_shape(&mut chat.extra).
E2E coverage tests/e2e/src/cases/messages-cross-provider-extras-e2e.test.ts	Adds a new e2e test suite with shared Anthropic-only field definitions and helper assertions, plus non-streaming and streaming tests verifying Anthropic-only fields are dropped and translatable fields correctly propagate to an OpenAI-compatible upstream mock.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Sequence Diagram(s)

sequenceDiagram
  participant Client
  participant Dispatch as cross_provider_dispatch
  participant Translator as translate_extras_to_openai_shape
  participant Upstream as OpenAI-compatible upstream

  Client->>Dispatch: POST /v1/messages (Anthropic body with extras)
  Dispatch->>Translator: translate_extras_to_openai_shape(chat.extra)
  Translator->>Translator: map tools/tool_choice, stop_sequences, metadata.user_id, thinking
  Translator->>Translator: drop unrecognized Anthropic-only fields
  Translator-->>Dispatch: translated extras
  Dispatch->>Upstream: POST /v1/chat/completions (translated body)
  Upstream-->>Dispatch: response
  Dispatch-->>Client: response

Loading

Possibly related issues

• api7/AISIX-Cloud#953: Directly addresses the reported bug where Anthropic-only fields like context_management leaked to OpenAI-compatible upstreams, causing 400 errors.

Possibly related PRs

• api7/aisix#254: Prior PR implemented the explicit tools/tool_choice translation logic that this PR consolidates into translate_extras_to_openai_shape.

🚥 Pre-merge checks | ✅ 5 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
E2e Test Quality Review	⚠️ Warning	The new E2E suite uses a permissive OpenAI mock, so it won’t catch the thinking→reasoning_effort regression on non-reasoning models.	Add an E2E case that exercises a non-reasoning OpenAI fixture and asserts reasoning_effort is omitted (or the request still succeeds) for both streaming and non-streaming.

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title is concise and accurately describes the main change: preventing Anthropic-only fields from reaching OpenAI-compatible upstreams.
Linked Issues check	✅ Passed	The changes whitelist/translate extras, dropping Anthropic-only fields on the OpenAI path and preserving Anthropic-native behavior, which matches #953.
Out of Scope Changes check	✅ Passed	The helper, proxy wiring, and E2E tests are all directly related to the same cross-provider request-shaping fix.
Security Check	✅ Passed	No new secret/PII leakage or auth/ownership issues found; the only new log emits dropped field names, not values.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/953-anthropic-extras-cross-provider

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🧹 Nitpick comments (1)

tests/e2e/src/cases/messages-cross-provider-extras-e2e.test.ts (1)

27-40: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Move thinking out of the drop-only set. It maps to reasoning_effort, so this case should assert the translated field instead of treating it as Anthropic-only.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/e2e/src/cases/messages-cross-provider-extras-e2e.test.ts` around lines
27 - 40, Move thinking out of ANTHROPIC_ONLY_FIELDS in
messages-cross-provider-extras-e2e.test.ts, because it translates to
reasoning_effort rather than being Anthropic-only. Update the test case that
uses ANTHROPIC_ONLY_FIELDS to assert the translated reasoning_effort field
instead, and keep only true drop-only extras in the Anthropic-only set. Use the
ANTHROPIC_ONLY_FIELDS constant and the related cross-provider extras assertions
to locate the test logic.

Source: Coding guidelines

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/aisix-provider-anthropic/src/wire.rs`:
- Around line 487-491: The current `thinking` handling in `wire.rs` always maps
to `reasoning_effort`, which can break non-reasoning upstreams like `gpt-4o`.
Update the `thinking` branch in the request translation logic to only emit
`reasoning_effort` when the target model/provider supports it, using the
existing `reasoning_effort_from_thinking` helper plus a capability/model gate.
Keep the `extra.insert("reasoning_effort", ...)` path behind that check so
Anthropic `thinking` is ignored or dropped for unsupported OpenAI models rather
than forwarded as an invalid field.

---

Nitpick comments:
In `@tests/e2e/src/cases/messages-cross-provider-extras-e2e.test.ts`:
- Around line 27-40: Move thinking out of ANTHROPIC_ONLY_FIELDS in
messages-cross-provider-extras-e2e.test.ts, because it translates to
reasoning_effort rather than being Anthropic-only. Update the test case that
uses ANTHROPIC_ONLY_FIELDS to assert the translated reasoning_effort field
instead, and keep only true drop-only extras in the Anthropic-only set. Use the
ANTHROPIC_ONLY_FIELDS constant and the related cross-provider extras assertions
to locate the test logic.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 488f4ddf-0133-42d9-96e8-92bbea4eaf46

📥 Commits

Reviewing files that changed from the base of the PR and between bba94b5 and eb60efa.

📒 Files selected for processing (4)

• crates/aisix-provider-anthropic/src/lib.rs
• crates/aisix-provider-anthropic/src/wire.rs
• crates/aisix-proxy/src/messages.rs
• tests/e2e/src/cases/messages-cross-provider-extras-e2e.test.ts