fix: send reasoning_effort="none" so gpt-5.6 models accept function tools

[ivancrneto]

The problem

OpenAI rejects function tools on /v1/chat/completions for every gpt-5.6 variant while reasoning is on:

Function tools with reasoning_effort are not supported for gpt-5.6-luna in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'.

The extraction agent is a tool loop on chat/completions, so this makes the whole gpt-5.6 family unusable for extraction — and there is currently no way to reach reasoning_effort from configuration (no LLMConfig field, no extra_body, no kwargs passthrough).

Two things make it unpleasant to diagnose:

1. It looks partial. Non-tool calls to the same model keep succeeding, so generation and evaluation appear healthy while every extraction fails.
2. LiteLLM hides the cause. It masks the provider message behind a TypeError of its own — BadRequestError.__init__() missing 2 required positional arguments: 'model' and 'llm_provider' — so the logs never show the sentence above.

I hit this on a self-hosted deployment: pointing the org config at gpt-5.6-luna failed 9 extraction runs in ~9 minutes before I rolled back, with only that TypeError to go on.

The fix

Follows the existing TEMPERATURE_RESTRICTED_MODELS idiom rather than inventing a new one:

• TOOLS_REQUIRE_REASONING_DISABLED_MODELS = {"gpt-5.6"}, declared next to TEMPERATURE_RESTRICTED_MODELS with the API error quoted in a comment.
• _tools_require_reasoning_disabled(), mirroring _is_temperature_restricted_model() — same prefix-strip for provider routing (openrouter/openai/gpt-5.6-luna), same startswith matching.
• In _build_completion_params, when the request carries tools and the model matches, default reasoning_effort to "none".

Scope is deliberately narrow:

• setdefault, placed before params.update(kwargs), so an explicit caller-supplied reasoning_effort still wins.
• Non-tool calls to gpt-5.6 keep reasoning enabled — only the tools path is affected.
• No other model is touched.

Verification

Against the live OpenAI API through LiteLLM (1.89.3), using the same drop_params=True the client sets — that mattered, because if LiteLLM stripped the parameter the fix would be inert:

call	result
gpt-5.6-luna + tools + reasoning_effort="none"	tool call returned
gpt-5.6-luna + tools, no reasoning_effort	the 400 above
gpt-5.6-terra + tools, no reasoning_effort	the same 400
gpt-5.5 + tools	unaffected, works either way

Added 16 unit tests in TestToolsRequireReasoningDisabled, mirroring TestTemperatureRestriction: affected/unaffected model matching, provider-prefix stripping, and four param-level tests driving generate_chat_response through a patched litellm.completion — tools set it, no-tools does not, unaffected models are untouched, and an explicit caller value survives.

pytest tests/server/llm/ → 719 passed, all 16 new tests among them.

Two caveats on my local run, since I would rather state them than let them look like passes:

• test_litellm_client.py::test_installed_litellm_transport_round_trip fails on a clean checkout of main here too — pre-existing in my environment, unrelated.
• test_embedding_service_model_contract.py failures vary between runs in my environment (1 of 3 on clean main, 3 of 3 with this branch, 0 of 3 in isolation on both). They also collect before the file I touched, so they cannot be affected by these tests. I believe it is local flakiness around the HF model download, but I could not fully rule it out locally — worth a glance at CI.

Alternatives considered

• Routing gpt-5.6 tool calls to /v1/responses, as the error suggests. Correct long-term, but a much larger change than making the completions path work.
• Exposing reasoning_effort as an LLMConfig field. Useful on its own, but it would leave the default broken — callers would have to know to set it. Happy to add it as well, or instead, if you would prefer that shape.

Summary by CodeRabbit

• Bug Fixes
  • Improved compatibility with GPT-5.6 models when function tools are used.
  • Reasoning is automatically disabled in affected tool-enabled requests when no explicit setting is provided.
  • Explicit reasoning preferences remain unchanged, and unaffected models and requests without tools continue to behave as before.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Team

Run ID: 0d448e69-9f61-4e9a-9314-6cc5bbad0a9a

📥 Commits

Reviewing files that changed from the base of the PR and between 695070a and ed58d0d.

📒 Files selected for processing (2)

• reflexio/server/llm/_litellm_text_generation.py
• tests/server/llm/test_litellm_client_unit.py

Included review availability: Your plan provides up to 2 included reviews per hour; 1 remains after this review.

---

📝 Walkthrough

Walkthrough

The change detects GPT-5.6 model variants when function tools are used and injects reasoning_effort="none" unless the caller provides another value. Unit tests cover model matching, provider prefixes, tool presence, unaffected models, and overrides.

Changes

Tool reasoning compatibility

Layer / File(s)	Summary
Affected model detection reflexio/server/llm/_litellm_text_generation.py	Adds the affected-model set and detects matching model names after lowercasing and removing provider prefixes.
Completion parameter handling and validation reflexio/server/llm/_litellm_text_generation.py, tests/server/llm/test_litellm_client_unit.py	Sets reasoning_effort to "none" for affected tool calls while preserving explicit caller values. Tests cover tool and non-tool calls, model variants, prefixes, unaffected models, and overrides.

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: ⚪ Minimal · up to ed58d

Tool-enabled GPT-5.6 chat-completion requests now disable reasoning by default to avoid provider rejection, while explicit settings and other request types retain their existing behavior. The change is ready to merge.

Suggested reviewers: yyiilluu, guangyu-reflexio, yilu331

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 40.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 15 functions across 2 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: setting reasoning_effort="none" for gpt-5.6 models so they accept function tools.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

• Fix all pre-merge checks with AI

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

Warning

Your free Security trial is over. An organization admin can upgrade to Advanced for continuous pull request security review or dismiss this notice.

---

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