fix(proxy): apply RequestOverrides on /v1/messages Anthropic passthrough (#335 #337)

[moonming]

Summary

The Anthropic-upstream branch of /v1/messages dispatch bypasses the Hub and the OpenAI bridge — it builds its HTTP request directly via reqwest. That meant the PK's request.* override block (param renames, constraints, default body fields, default headers) silently no-opped on this path: cp-api validated and persisted the override, but the DP never applied it on outbound.

Surfaced by AISIX-Cloud's source-blind E2E matrix (PR #352 D3.3): the param_constraints clamp scenario failed pre-audit; the default_headers injection scenario failed on PR #352's first CI run. Both led to filing #335 and #337 with full repros.

Fix

crates/aisix-proxy/src/messages.rs::dispatch (Anthropic-passthrough branch only):

• Apply aisix_provider_openai::overrides::{apply_param_renames, apply_param_constraints, apply_default_body_fields} to the outbound JSON body before .json(body).
• Build an explicit axum::http::HeaderMap so apply_default_headers can inject operator-supplied headers via the shared apply pipeline.

Bridge-owned headers (x-api-key, anthropic-version, content-type, x-aisix-request-id) are inserted FIRST. apply_default_headers skips keys already present + the RESERVED_DEFAULT_HEADERS blacklist (which DOES include x-api-key), so operator headers can never clobber auth.

aisix-provider-openai moves from [dev-dependencies] to [dependencies] in aisix-proxy/Cargo.toml. Same architectural shape as aisix-provider-anthropic.

What's NOT in this PR (filed as follow-ups)

Audit (CLAUDE.md §7) verified that the cross-provider dispatch branch inherits the apply pipeline ONLY for OpenAiBridge and AzureOpenAiBridge. Three more provider bridges have the same wire-in gap:

• ai-gateway#333 aisix-provider-openai's streaming reasoning_content — actually NOT a gap; closed after verification (false-positive audit on D3.1)
• ai-gateway#339 wire RequestOverrides on aisix-provider-vertex
• ai-gateway#340 wire RequestOverrides on aisix-provider-bedrock

This PR is scoped to the Anthropic-passthrough fix. The cross-provider branch through OpenAI-family bridges (OpenAI / DeepSeek / Google) already works via the Bridge's own prepare_outbound_body / build_request_headers.

Tests

Five new wiremock-driven unit tests on the Anthropic-passthrough path, each strict-matching on body or header. All five fail before the fix, pass after. 19 of 19 messages.rs tests green; clippy clean.

Test	Pins
anthropic_passthrough_applies_param_renames	max_tokens → max_tokens_to_sample on outbound
anthropic_passthrough_clamps_temperature_via_param_constraints	caller 0.9 → upstream 0.5
anthropic_passthrough_fills_default_body_fields_when_caller_omits	missing top_p filled with 0.9
anthropic_passthrough_injects_default_headers	x-tenant-id: acme-prod-42 reaches upstream
anthropic_passthrough_default_headers_cannot_overwrite_x_api_key	defense-in-depth — PK secret survives even if default_headers names x-api-key

Audit findings addressed

Severity	Finding	Resolution
MEDIUM-1	Cross-provider only inherits override pipeline for OpenAI-family bridges	Filed as ai-gateway#339 (Vertex) + ai-gateway#340 (Bedrock); explicitly scoped this PR to Anthropic-passthrough
MEDIUM-2	BridgeError::Config error wording could surface in operator logs	Acceptable as-is (mirrors bridge.rs:347); follow-up to revisit log level for unreachable-by-construction errors deferred
LOW-1	RESERVED_DEFAULT_HEADERS guard not isolated from contains_key guard in test	Acceptable; unit test at overrides.rs:548 covers RESERVED list in isolation
LOW-2	param_renames ordering with model field rewrite	Same as OpenAI bridge ordering; comment update documenting precedence is appropriate but optional

References (per CLAUDE.md §7)

• LiteLLM applies the same primitives on both endpoints via its litellm_pre_call_hooks chain.
• Portkey applies header/body transforms in its transformers layer ahead of the upstream call regardless of provider family.
• Anthropic API does NOT spec a request.* override block server-side — this is the gateway's customer-facing primitive (refactor(server): inline DeepSeek/Google bridge factories + delete wrapper crates (Phase A) #302 §5 RuntimeConfig).

Test plan

• CI green (cargo test + clippy + fmt)
• Once merged, AISIX-Cloud held-back matrix scenarios flip on (D3.3 param_constraints clamp + default_headers injection)

Closes #335
Closes #337
Refs #339 #340

[coderabbitai]

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Free

Run ID: b8cf8b83-ce03-4932-a9b0-b5300e20be28

📥 Commits

Reviewing files that changed from the base of the PR and between e6125b6 and 2e1b01d.

📒 Files selected for processing (2)

• crates/aisix-proxy/Cargo.toml
• crates/aisix-proxy/src/messages.rs

---

📝 Walkthrough

Walkthrough

The PR adds OpenAI provider support to the proxy crate and implements a request override pipeline for Anthropic passthrough. Body parameters are transformed via renames and constraints, default fields are injected, and headers are explicitly constructed with operator-supplied overrides protected from reserved keys.

Changes

Anthropic Passthrough Request Override Pipeline

Layer / File(s)	Summary
OpenAI Provider Dependency Setup crates/aisix-proxy/Cargo.toml	aisix-provider-openai added to proxy dependencies alongside aisix-provider-anthropic with expanded inline documentation explaining why /v1/messages handler requires these provider primitives.
Request Body Override Application crates/aisix-proxy/src/messages.rs	Anthropic passthrough applies provider-key request.* override config to outbound body: parameter renames, parameter constraints, and default body field injection are guarded by override config presence.
Header Construction with Operator Injection crates/aisix-proxy/src/messages.rs	Upstream headers explicitly constructed as HeaderMap with gateway-required headers inserted first, then operator-supplied request.default_headers applied through shared override helper with x-api-key protected by reserved-header blacklist.
Request Override Test Suite crates/aisix-proxy/src/messages.rs	New wiremock-backed tests verify override features: param_renames, param_constraints clamping, default_body_fields population, default_headers injection, and reserved-header protection.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

---

Note

🎁 Summarized by CodeRabbit Free

Your organization has reached its limit of developer seats under the Pro Plan. For new users, CodeRabbit will generate a high-level summary and a walkthrough for each pull request. For a comprehensive line-by-line review, please add seats to your subscription by visiting https://app.coderabbit.ai/login.If you believe this is a mistake and have available seats, please assign one to the pull request author through the subscription management page using the link above.

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

[Copilot]

Pull request overview

This PR wires ProviderKey request.* overrides into the direct Anthropic /v1/messages passthrough path, which previously bypassed the shared Hub/OpenAI bridge override pipeline.

Changes:

• Applies param renames, param constraints, default body fields, and default headers before dispatching Anthropic passthrough requests.
• Builds outbound headers explicitly so default headers can be merged without clobbering bridge-owned auth/version headers.
• Moves aisix-provider-openai to runtime dependencies and adds wiremock coverage for the new passthrough override behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
crates/aisix-proxy/src/messages.rs	Applies request override primitives to Anthropic passthrough body/headers and adds related tests.
crates/aisix-proxy/Cargo.toml	Promotes aisix-provider-openai to a runtime dependency for shared override helpers.

Comments suppressed due to low confidence (1)

crates/aisix-proxy/src/messages.rs:1015

• This test relies on body_partial_json, so it is not strict on the whole body; it would still pass if the old max_tokens field were left alongside max_tokens_to_sample. Please reword the comment or add an exact/captured-body assertion if that absence is intended to be pinned.

        // primitive now fires on outbound. mock-llm matcher is
        // strict on body — the rename MUST be applied or wiremock
        // returns 404.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.