Simon Ferquel's Clanker 3900ec89b6 fix(openai): degrade gracefully on Chat Completions for gpt-5.4+ tools+effort and warn on ignored thinking_budget ...

Chat Completions rejects an explicit reasoning_effort alongside function
tools starting at gpt-5.4 (live-verified against api.openai.com); gpt-5.6+
and gpt-6+ reject the combination outright even with the field omitted.
When a Responses-capable model is forced onto Chat Completions (an explicit
api_type or custom-provider override) with tools in the request, drop the
effort field and warn once instead of sending a value guaranteed to 400.

Also warn once per client when thinking_budget is configured on a model
that does not accept a reasoning-effort parameter at all (previously
silent — part of the #4162 symptom set before gpt-6 was even recognised).

Live-verify (per review from the architect session) that gpt-5.6 also
rejects reasoning.effort 'minimal', confirming openAIDropsMinimalEffort's
gpt-5.6+ scope is directly observed rather than inferred from gpt-6 alone.

Part of docker/docker-agent#4162.

2026-09-04 10:28:49 +00:00

..

modelinfo_test.go

feat(modelinfo): recognise the gpt-6 family via generation-based parsing

2026-09-04 10:24:35 +00:00

modelinfo.go

fix(openai): degrade gracefully on Chat Completions for gpt-5.4+ tools+effort and warn on ignored thinking_budget

2026-09-04 10:28:49 +00:00

resolve_caps_test.go

fix(providers): add per-model attachment capability override

2026-06-22 21:52:31 +02:00

thinking_levels_test.go

fix(openai): degrade gracefully on Chat Completions for gpt-5.4+ tools+effort and warn on ignored thinking_budget

2026-09-04 10:28:49 +00:00

thinking_levels.go

fix(openai): degrade gracefully on Chat Completions for gpt-5.4+ tools+effort and warn on ignored thinking_budget

2026-09-04 10:28:49 +00:00