fix(openai): merge additional_params tools into chat completions tool list; record error-envelope coverage

[gold-silver-copper]

Summary

Live-traffic bug hunt against the OpenAI, Anthropic, and Gemini providers: new cassette-backed tests were recorded against the real APIs until a genuine defect surfaced; the defect is fixed here with a cassette-backed regression test, and the passing coverage recorded along the way ships too.

Bug found and fixed

OpenAI Chat Completions: additional_params["tools"] silently replaced builder-registered tools (issue #1890's chat-path half).

• Symptom (reproduced live, recorded): building a request with .tool(...) plus additional_params(json!({"tools": [...]})) sent only the additional_params tools to the wire. With tool_choice: required and a prompt demanding the builder tool, the model was forced to call the wrong tool — the recorded pre-fix traffic shows the request body containing only the raw tools array, and the model answering with lookup_beta when told to call lookup_alpha.
• Root cause: providers/openai/completion/mod.rs flattens additional_params into the serialized CompletionRequest. The body is built via serde_json::to_value, where the flattened tools key wins over the typed tools field — a silent last-write-wins clobber. Every sibling path already merges instead: the Responses API (responses_api/mod.rs), Anthropic (build_tool_definitions), and Gemini (extract_tools_from_additional_params). The chat path was the one hole.
• Fix: extract tools from additional_params in TryFrom<OpenAIRequestParams> for CompletionRequest and append it to the typed tool list, mirroring the Responses-path semantics including the loud RequestError on a malformed payload.
• Proof: tests/providers/openai/cassette/additional_params_tools.rs + tests/cassettes/openai/additional_params_tools/builder_tools_survive_additional_params_tools.yaml — a fresh post-fix recording whose request body carries both tools and whose response calls lookup_alpha. The cassette harness matches the outbound body, so any regression back to the clobber fails replay as a mock miss.

New coverage recorded (independent of the fix)

Error-envelope preservation had zero non-2xx interactions in any committed cassette. Six new recordings (unary + streaming per provider) now pin that a real provider error keeps its status and raw body recoverable through provider_response_status() / provider_response_body():

• openai/error_envelope/* — Responses API, 400 model_not_found (note: OpenAI returns 400, not 404, for an unknown model)
• anthropic/error_envelope/* — Messages API, 404 not_found_error
• gemini/error_envelope/* — generateContent + streamGenerateContent, 404 NOT_FOUND (the streaming error arrives on an SSE-content-type response and still surfaces correctly)

These also document that the streaming paths surface the pre-stream HTTP error lazily (from the first poll of the stream), which the tests capture.

Tracker sweep

All 72 open issues and 52 open PRs were read in full (including comments). Classification:

Addressed here

• feat: support provider-hosted tools in AgentBuilder end-to-end #1890 (chat-path half): fixed as above. The Responses-path half (hosted tools via additional_params) was found to already merge correctly on main (responses_api/mod.rs removes and merges the key); the issue's remaining ask — first-class provider_tools() consumption — is a feature, not fixed here.

In scope, reproduced or verified, left as follow-ups (open fix PRs exist or fix is a larger surface)

• Response headers (incl. Retry-After) are discarded when non-success HTTP errors are constructed #2210 / PR feat(core)!: preserve response headers on non-success HTTP errors #2211 — response headers (incl. Retry-After) structurally dropped on non-success errors; an open PR fixes it. The new error-envelope cassettes confirm bodies/status survive; headers need feat(core)!: preserve response headers on non-success HTTP errors #2211's new API surface.
• feat(openai): preserve Responses items for stateless replay #2269 / PR feat(openai): support server-side Responses compaction #2234 — OpenAI Responses phase field dropped on message items; compaction input items rejected. Overlaps the open compaction PR; fixing phase preservation independently risks conflicting with feat(openai): support server-side Responses compaction #2234's item-model changes.
• feat: support OpenAI prompt cache breakpoints (Responses API) #2170 / PR feat(openai): support manually placed prompt cache breakpoints in the Responses API #2171 — Responses prompt-cache options silently dropped; open PR ships the fix with its own cassette.
• feat: manually placed prompt-cache breakpoints for Anthropic #2168 / PR feat(anthropic): support manually placed prompt-cache breakpoints #2169 — Anthropic per-block cache_control discarded; open PR ships the fix with cassettes.
• PR fix(openai): accept Azure chat completion tool-call payloads #2255 (Azure tool-call deserialization), PR fix(openai): forward additional_params in image generation request #2264 (image-gen additional_params drop), PR feat(anthropic): add Claude 4.6 adaptive thinking support #1480 (adaptive thinking), PR fix(embeddings): respect token limits in EmbeddingsBuilder batching #1221 (embeddings batch token limit), Standardize Rig GenAI telemetry on OpenTelemetry semantic conventions #1593 (telemetry accounting) — each has an open PR or needs keys/surface outside this run's scope.

Adjacent (PRs touching these providers that would benefit from recorded regression tests, not added here to avoid churn against their in-flight diffs): #2243, #2234, #1905, #1867, #1185, #2121, #2190, #1279.

Out of scope: all remaining open items are other providers, features, docs, or infra — full classification list available in the sweep notes; representative: #2291/#2290 (Mistral), #2241/#2263 (Cohere), #2156 (Ollama), #2292 (CI), #2284/#2278 (agent surface), #2229/#2228 (refactors), etc.

Recording matrix

Scenario	Provider	Kind	Outcome
error_envelope unary	openai, anthropic, gemini	coverage gap	recorded, pass
error_envelope streaming	openai, anthropic, gemini	coverage gap	recorded, pass
additional_params tools clobber	openai (chat)	tracker #1890	bug reproduced live → fixed → re-recorded, pass

Cassettes: 7 newly recorded, 0 re-recorded (the fix changes the wire body only for requests carrying additional_params.tools, and no pre-existing cassette does), 0 touched otherwise. Hygiene review done file-by-file: keys/request_ids redacted by the scrubber, no secrets, no unrelated captures, placeholders redact rather than fabricate.

Verification

• cargo fmt — clean
• cargo clippy --all-targets --all-features — clean (no errors, no new warnings)
• cargo test -p rig-core --all-features — 1325 lib tests + doctests pass (includes a new unit test for the mixed function/native additional_params.tools split, which also keeps Groq's compound_custom native-tool folding intact)
• Replay suites (--test-threads=1): openai 108 passed / anthropic 79 passed / gemini 131 passed (baselines before any change: 105 / 77 / 129 passed, all green — no pre-existing failures)

Estimated API spend

Under $0.05 total: six error-envelope recordings are rejected requests (no tokens billed), and the two gpt-4o-mini tool-call recordings used ~100 tokens each.