fix(llms): send reasoning_effort to every openai reasoning model - #7187

• reasoning_effort now reaches every OpenAI reasoning model on the chat-completions path. It was gated behind is_o1_model = "o1" in model.lower() (completion.py:289), a literal substring test, so gpt-5, gpt-5-mini, o3, o3-mini and o4-mini all failed it and thought at the server default. The request still succeeded, so nothing surfaced — one measured extraction ran 6.2s with the setting applied against 149.7s with it silently dropped.
• The gate could not simply be widened. is_o1_model also drives supports_function_calling (:2656), supports_stop_words (:2663) and the system→user message rewrite (:2762), so marking gpt-5 as an o1 model would make CrewAI report that gpt-5 cannot call tools and would mangle its system prompts. A separate predicate, _supports_reasoning_effort, decides it instead.
• That predicate matches on model shape, not a list of names: the o-series (^o\d) and GPT generation 5 onwards (^gpt-(\d+) >= 5). A new member of an existing family therefore needs no release here — gpt-6 and o5 already classify correctly — while gpt-4o and gpt-4.1 parse to generation 4 and are excluded, so a non-reasoning model never pays a wasted round trip.
• The unsupported-parameter retry is kept as a safety net for the case the shape match is wrong for some future family: on a 400 that rejects the parameter, the call is retried once without it. It costs nothing when the match is right. The predicate is deliberately narrow — it matches code="unsupported_parameter" and the null-param "Unrecognized request argument" shape, and explicitly does not match the "Unsupported value" 400 that o1/o3 return for a bad value, since dropping the key there would silently restore the original bug. The existing GPT-5.6 tools retry is untouched and still recovers by sending "none".
• Adds "minimal" to LLM.reasoning_effort, which gpt-5 accepts and the Literal omitted. Widening a Literal is non-breaking; the native provider already accepted it (untyped str) while the litellm path rejected it, so the cheapest setting was unreachable on the typed surface.
• Worth knowing for review: tests/test_llm.py:298-327 already had three test_o3_mini_reasoning_effort_* tests. They pass either way — they assert "Paris" in result, and VCR's default matcher ignores the request body. Their cassettes do contain reasoning_effort, but were recorded under OpenAI/Python 1.61.0 in the litellm era, so they were asserting nothing. That is why this went unnoticed.
• 61 tests. They assert the parameter on the built request across six models; that non-reasoning models never receive it; the shape predicate across supported, unsupported and unreleased families; the retry across all four dispatch paths — sync, async, sync streaming, async streaming; that an unsupported value still raises; that unset stays off the wire; and that is_o1_model's three other behaviours are unchanged (gpt-5 still reports tool support, system messages are not rewritten).
• No docs change: docs/edge/en/concepts/llms.mdx shows reasoning_effort only in usage examples and never enumerates valid values or supported models, so nothing documented becomes false. It already describes o1/o3/o4 as reasoning models at :243.
• Keeps this intentionally small: is_o1_model's definition and its other uses are untouched, no reasoning-model allowlist, and no changes to the Azure, litellm or responses paths, which already forward correctly.
• Next: the two streaming handlers lack the failed-event suppression the non-streaming ones have (pre-existing); and the three fossil o3-mini cassettes want re-recording with a body matcher so they assert something.

🤖 Generated with Claude Code