fix(aux): title generation survives custom endpoints that reject reasoning_effort (#112781, supersedes #112789)

[teknium1]

Auxiliary title generation no longer dies with 400 Unrecognized request argument supplied: reasoning_effort on custom/OpenAI-compatible endpoints whose model does not accept the field — the call is retried once with every reasoning field omitted and the title lands.

Changes

• agent/auxiliary_client.py — new rung in the shared aux recovery ladder (_ladder_parameter_rungs; sync call_llm and async_call_llm drive the same generator, so one rung covers both): when a 400/422 names a reasoning field (reasoning_effort, reasoning, thinking/think) with an unsupported/unrecognized marker, strip top-level reasoning_effort, the adapter-private _reasoning_config and every extra_body reasoning key, then retry once. Same reactive shape as the existing temperature and response_format rungs; an unrelated 400 or a request that carried no reasoning field never retries.
• plugins/model-providers/custom/__init__.py — untouched: reasoning_effort: "none" remains the deliberate thinking-off encoding for Ollama /v1 ([Bug]: agent.reasoning_effort: none silently ignored on Ollama — main agent stuck in medium mode, bg-review fork can spiral (up to 65k tokens / 28 min) #25758), vLLM, SGLang and GLM/ARK routes, on the first request, for aux and main-agent calls alike.
• tests/agent/test_auxiliary_reasoning_field_rejection_retry.py — 2 invariants (sync/async retry with the reporter's verbatim 400; unrelated 400 does not strip). Red on origin/main (2 failed / 1 passed), green on this head.
• website/docs/user-guide/configuration.md — one paragraph under the auxiliary reasoning_effort knob describing the retry.

Validation

Live probe (monkeypatched openai Completions.create/AsyncCompletions.create acting as a relay that 400s on reasoning_effort, real call_llm/async_call_llm with provider=custom, model=gpt-4.1-mini, reasoning_config={"enabled": False}, temp HERMES_HOME):

	wire calls	outcome
before (origin/main 9796235), sync + async	1: {"model": "gpt-4.1-mini", "reasoning_effort": "none", "temperature": 0.3, "extra_body": {"response_format": …}}	BadRequestError 400 … reasoning_effort — title lost
after (this head), sync + async	1: same as above → 400; 2: {"model": "gpt-4.1-mini", "temperature": 0.3, "extra_body": {"response_format": …}}	{"title": "Relay Title Probe"}

Tests: scripts/run_tests.sh on the new file + test_structured_output_rejection_retry.py, test_unsupported_parameter_retry.py, test_unsupported_temperature_retry.py, test_auxiliary_client.py, test_injected_param_strip_retry_registry.py, test_auxiliary_auth_rung_fallthrough.py → 287 passed, 0 failed.

Root cause

title_generator.generate passes reasoning_config={"enabled": False} (#91927); CustomProfile.build_api_kwargs_extras projects that as top-level reasoning_effort: "none" for every custom endpoint, and because the profile handles reasoning there was no fallback or recovery when a chat-only model behind a relay rejects the field.

Fixes #112781

Supersedes #112789 (@KoNit-K) — that PR restricts the reasoning_effort: "none" encoding to Ollama URLs inside the shared custom profile, which would silently re-enable thinking for vLLM/SGLang/GLM users who set reasoning_effort: none (main agent included, not only the title lane). The recovery here is provider-agnostic and leaves the deliberate encoding in place; thanks to @KoNit-K for the report-to-test mapping.

Infographic

Review follow-up

• [MAJOR] _is_reasoning_field_rejection fired on route-gating 400s that merely name a thinking model (kimi-k2-thinking is not supported when using this account), spending a strip-retry and then re-raising before the fallback chain. Reproduced on head f4c7b4c (1 primary call → raise, _try_configured_fallback_chain never called; also 'max_tokens is not supported with reasoning models' matched). Fixed @ 2f2e059: the reasoning token must now be a standalone wire-field name (not a model-id segment, not "… with reasoning models"), and _param_rung_accepts accepts _is_model_incompatible_error so a gating 400 after any strip still reaches provider fallback. New test test_model_gating_400_naming_a_thinking_model_still_reaches_the_fallback_chain (red on f4c7b4c, green now).
• [MINOR] temperature-strip retry that 400s on reasoning_effort re-raised before the reasoning rung. Fixed @ 2f2e059: _param_rung_accepts also accepts _is_reasoning_field_rejection / _is_structured_output_rejection; probe now makes 3 calls, the third without both fields.
• [MINOR] zh-Hans configuration.md mirror lacked the retry paragraph. Fixed @ 2f2e059 (one-line mirror addition before the 后台审查 paragraph).

Known residuals

• None.

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 2f2e059 — fix(aux): only treat a standalone reasoning field name as a}

⚠️ Warnings

CI timings · View report · View job

Wall time 9m14s vs 6m8s (+50.5%). 6 job(s) slower, 6 faster, 1 unchanged.

• Python tests / e2e: +13.0s
• Python lints / Windows footguns (blocking): -7.0s
• Check no case-colliding filenames / check-case-collisions: +6.0s
• Python tests / Run tests: -5.0s
• OS-specific tests / Windows-only tests: -5.0s