Fix LiveKit reasoning effort default for gpt-5.4 models

[mstolarzblaxelai]

Fixes ENG-2508.

What changed:

• Default LiveKit reasoning_effort to none for Blaxel-resolved OpenAI gpt-5.4 models.
• Preserve caller-provided reasoning_effort exactly.
• Leave non-OpenAI provider behavior unchanged.

Why:

• Current LiveKit defaults gpt-5.4-mini to minimal.
• Blaxel rejects minimal for this model. Supported values include none, low, medium, high, and xhigh.

Verified:

• uv run --locked --python 3.12 --extra livekit --group test pytest tests/integration/livekit/test_model.py::test_gpt54_openai_models_default_reasoning_effort_to_none tests/integration/livekit/test_model.py::test_livekit_reasoning_effort_override_is_preserved tests/integration/livekit/test_model.py::test_livekit_non_openai_provider_does_not_default_reasoning_effort -q
• BL_WORKSPACE=main uv run --locked --python 3.12 --extra livekit --group test pytest tests/integration/livekit/test_model.py::TestBlModel::test_can_call_model -q
• BL_WORKSPACE=main uv run --locked --python 3.12 --extra livekit --group test pytest tests/integration/livekit -q
• uv run ruff check .
• make test
• git diff --check

---

Note

Adds a helper that defaults reasoning_effort to "none" for gpt-5.4 family models on the OpenAI provider in the LiveKit integration, preventing Blaxel from rejecting the LiveKit-default value of "minimal". Three focused unit tests cover the happy path, caller override, and non-OpenAI provider passthrough.

^{Written by Mendral for commit dd8df73.}

[mendral-app]

LGTM

Fix is correct and well-scoped. No security, correctness, or data-loss issues.

_{Tag @mendral-app with feedback or questions. View session}