Route gpt-5.6 models through Responses API when function tools are bound

[adambalogh]

Problem

The chat-app agent page fails for the gpt-5.6 models (sol/terra/luna) with an OpenAI 400:

Error code: 400 - 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 gpt-5.6 family applies a default reasoning_effort server-side, and OpenAI's Chat Completions endpoint rejects that default when function tools are also present. The agent page always binds tools, so every gpt-5.6 agent request 400s.

The gateway never sends reasoning_effort itself, and langchain-openai only auto-prefers the Responses API for -pro/codex variants — so these models default to Chat Completions and hit the wall.

Fix

Route the gpt-5.6 family through OpenAI's Responses API whenever function tools are bound — the path already used for native web search (use_responses_api=True, output_version="responses/v1"). The Responses API supports reasoning and function tools together, so reasoning is preserved rather than disabled (as reasoning_effort='none' would).

• model_registry.py: new responses_api_for_tools flag on ModelConfig, set on gpt-5.6-sol/terra/luna.
• llm_backend.py: get_chat_model_cached gains a force_responses_api param; the OpenAI branch switches to the Responses API when it (or web_search) is set. Model cache key extended accordingly.
• chat_controller.py: both streaming and non-streaming paths build the tools list first, then set force_responses_api via _needs_responses_api_for_tools(provider, cfg, tools_list) before constructing the model.

Scoped precisely: only OpenAI gpt-5.6 models with tools bound switch APIs. Non-gpt-5.6 models, and tool-less gpt-5.6 requests, keep using Chat Completions unchanged — so structured-output / plain-chat paths are untouched.

Tests

Added three cases to test_tool_forwarding.py:

• gpt-5.6 + tools → force_responses_api=True
• gpt-5.6 without tools → force_responses_api=False
• gpt-4.1 (non-reasoning) + tools → force_responses_api=False

ruff format, ruff check, and mypy all pass on the changed files. Full existing suite shows no new failures (pre-existing unrelated loader/import failures remain).

🤖 Generated with Claude Code

---

Generated by Claude Code

[Copilot]

Pull request overview

This PR fixes OpenAI gpt-5.6 (sol/terra/luna) requests failing on the agent/chat path when function tools are bound by routing those requests through OpenAI’s Responses API (which supports reasoning + tools), while keeping non-gpt-5.6 and tool-less requests on Chat Completions.

Changes:

• Added a per-model registry flag to mark models that must use the Responses API when tools are present (gpt-5.6 family).
• Extended cached model construction to optionally force the Responses API for OpenAI, and updated chat controller logic to compute this before instantiation (streaming + non-streaming).
• Updated CI lint workflow to run ruff via uv run (locked toolchain), and added targeted unit tests for the gpt-5.6 tool-routing behavior.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tee_gateway/model_registry.py	Adds responses_api_for_tools to ModelConfig and enables it for gpt-5.6 sol/terra/luna.
tee_gateway/llm_backend.py	Adds force_responses_api to get_chat_model_cached and routes OpenAI construction to Responses API when requested.
tee_gateway/controllers/chat_controller.py	Computes tools list earlier and forces Responses API for flagged OpenAI models when tools are bound (streaming + non-streaming).
tee_gateway/test/test_tool_forwarding.py	Adds unit tests validating the new gpt-5.6 routing decision when tools are present/absent.
.github/workflows/lint.yml	Runs ruff via uv run instead of ruff-action to align CI linting with the locked toolchain.

---

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