[Rhiz3K]

Stacked on #9 — merge #9 first; until then this diff also includes its commits (Responses API for OpenAI).

What

Providers of type OpenAI and OpenAI compatible get a collapsible Advanced section in the provider form. Other provider types don't show it and never send it.

Field	Values	Responses API	Chat Completions
Endpoint	Auto (default) / Responses API / Chat Completions	picks the path; Auto keeps today's rule (api.openai.com → Responses)
Reasoning effort	Default / none / minimal / low / medium / high / xhigh / max	reasoning.effort	reasoning_effort
Reasoning summary	Default (auto) / auto / concise / detailed / off	reasoning.summary (off = not sent)	not applicable
Verbosity	Default / low / medium / high	text.verbosity	verbosity
Max output tokens	1–1,000,000	max_output_tokens	max_completion_tokens (OpenAI) / max_tokens (compatible)
Temperature	0–2	temperature	temperature

Leaving a field on Default sends nothing new, so the requests are the same as before. The values come from the OpenAI API reference (Responses create; Chat Completions create) and the latest-model guide: "When reasoning effort is not none, remove temperature, top_p, and top_logprobs."

Behaviour details

• Max output tokens: the user's limit never goes above a caller's own cap (Math.min). Callers include the engine's reply limit and the 2048 cap used by exposure explanations. With the field empty, the plugin keeps sending its own limit as before.
• Temperature: it is not sent to openai providers when the reasoning effort is set to anything other than none, and a warning is logged. A server that answers 400/422 about temperature is asked again without it, also with a warning. Compatible servers get it with any effort.
• Fallbacks:
  • The Chat reasoning_effort: "none" retry for tools still only runs when the user did not set an effort.
  • The Responses summary retry now drops only reasoning.summary and keeps a user-set effort.
  • If the model refuses the user's effort, the error is shown. Summaries are not turned off.
• Storage: one nullable JSON text column advanced on p_ai_providers, added by migration 0013_provider_advanced (sqlite/postgres text, mysql longtext). It was generated with termix-plugin migrations, and snapshot.json is updated.
• API:
  • POST /providers and PATCH /providers/:id accept advanced. PATCH replaces the whole object; null clears it.
  • Unknown keys, wrong types and out-of-range values get a 400 "Invalid advanced settings".
  • Every provider response returns advanced ({} when nothing is set).
  • The OpenAPI comments are updated.
• UI: AiProviderSettings.tsx, 27 new ai.advanced* keys in locales/en.json. Translations are not included.

Verification

• npx tsc -p .: OK. npx prettier --check .: OK.
• npx vitest run: 88 files, 937 passed, 1 skipped (base: 87 files, 823 passed).
• New and extended tests:
  • Request mapping for both endpoints: all fields set; empty advanced gives an unchanged body; summary off; summary refused keeps the effort; effort refused shows the error and summaries stay on; temperature rules and retry; endpoint override both ways; max token cap; a user effort is never replaced by the none retry.
  • Route validation for create, replace, clear and each invalid case.
  • Migration 0013 for every dialect.
  • Frontend form: hidden for other types, toggle, only set fields sent, null on edit, prefill, invalid temperature blocks Save.
• Mutation checks. Each of these broke the change on purpose, and tests caught every one:
  • summary retry dropping the effort
  • temperature skip removed
  • chat effort not sent
  • token cap removed
  • temperature range widened
  • endpoint override removed
• The plugin build (termix-plugin test/typecheck/build/validate/pack) passed.
• Manual: in the Termix test container (ghcr.io/termix-ssh/termix:latest, plugin mounted as bundled) the log shows Applied 1 migration(s) for plugin ai: 0013_provider_advanced and Activated plugin ai.
• Manual, in the browser: the Advanced section opens in the provider form for an OpenAI provider with gpt-6-luna (screenshot below) and the assistant still works with the settings left on Default.
• Not verified: requests against the live OpenAI API with non-default values in every field.

Screenshot

[termix-triage-bot]

AI review | Looks safe to merge

Adds validated advanced settings for OpenAI providers with proper migrations, UI, and fallback logic. No security or rule violations found.

• minor Plugin PRs should target main, not dev-1.3.0.

_{Reviewed 8aba74b. AI can be wrong, a maintainer still decides. Comment /termix review to run it again.}