fix: let models listing none reasoning run on native Codex (#198387) Root cause: src/config/model-provider-config.ts:129 (main 6445f7e13ff8) only counted `minimal|low|medium|high|xhigh|max|ultra` as native reasoning efforts, so a configured model whose `compat.supportedReasoningEfforts` also lists `none` (Codex's own `none` effort, which Codex `model/list` reports and the Codex plugin maps `/think off` to) was classified as an authored request override. Codex then refused the route and model selection reported that the runtime was not available. Related: #136292 (closed; same classifier, `off` instead of `none`) ## What Problem This Solves If you configure an OpenAI model with a reasoning-effort list that includes `none` (for example copied from what Codex lists for that model) and pick it on the native Codex runtime, the picker refuses it with `Runtime "codex" is not available for openai/<model> ...`, and turns that would have used Codex silently run on the OpenClaw runtime instead. ## User Impact Models whose effort list includes `none` next to other Codex efforts can be selected on the native Codex runtime again. Lists with custom effort labels, and a list of only `none` (reasoning disabled), still keep the OpenClaw adapter as before. No config change is needed. ## Why This Change Was Made Codex treats `none` as a regular reasoning effort: `ReasoningEffort::None` serializes as `none` (`codex-rs/protocol/src/openai_models.rs:74,93,160` at rust-v0.169.0, the bundled Codex version), and `model/list` passes the backend's `supported_reasoning_levels` through unchanged (`openai_models.rs:439,1097`, `codex-rs/app-server/src/models.rs:36,65`, `app-server-protocol/src/protocol/v2/model.rs:132,173`). OpenClaw's Codex plugin already publishes that list into `compat.supportedReasoningEfforts` (`extensions/codex/src/app-server/model-catalog.ts:52`) and sends `none` for `/think off` when the list has it (`extensions/codex/src/app-server/reasoning-effort.ts:31`). The classifier now accepts `none` as part of a native list. A list of only `none` stays request behavior, matching the existing "disabled effort" decision from #130752; the runtime docs say so. This keeps @dankarization's fix and credits them for it; the follow-up commit restores the `none`-only boundary and updates the docs. ## Evidence Built Gateway (dist) plus the real bundled Codex 0.169.0 app-server pointed at a local ChatGPT-backend fixture whose `/models` lists `none` for `gpt-6.1-sol`. The client picks `openai/gpt-6.1-sol` on runtime `codex` with `sessions.patch`, sets thinking `off`, and sends one turn. - main: `Runtime "codex" is not available for openai/gpt-6.1-sol: sign-in needed. ...` (2 of 2 runs) - this PR: selected on `codex`, turn answered by the fixture through Codex (2 of 2 runs) - controls: a list without `none` is selected on both; `["none","low","turbo"]` and `["none"]` are still refused on this PR - upgrade from v2026.11.5 with the same config and state: 11.5 refuses Codex and the session sits on the OpenClaw runtime; after `doctor --fix` on this PR the same session keeps its label, model and thinking level, resolves to Codex and completes a Codex turn Observed lines: https://github.com/openclaw/openclaw/pull/198387#issuecomment-7004208128 (fresh profiles), https://github.com/openclaw/openclaw/pull/198387#issuecomment-7004738151 (upgrade), https://github.com/openclaw/openclaw/pull/198387#issuecomment-7005059750 (upgrade without doctor, config hashes) <details> <summary>Merge danger</summary> Two-way door: a revert restores the old classification; nothing is written or migrated. Blast radius: only configured models whose `supportedReasoningEfforts` mixes `none` with other Codex efforts, which now stay eligible for the native Codex runtime instead of the OpenClaw adapter. </details> Co-authored-by: Ayaan Zaidi <hi@obviy.us>
Current
·
C2
·2026-11
fix: let models listing none reasoning run on native Codex (#198387)
Find Your Interest
- fix(knowledge): scribes stop after one item when a run can do more (#…
- Foxpit
- fix: stop rereading settled session accounting (#200497)
- fix(models): stop recommending superseded HY4 Preview
- fix: keep one child completion failure from dropping sibling results …
- fix: align Goal authority proof with admitted-run policy (#201567)
- fix: let remote Codex app-servers use library skills instead of an un…
- fix: keep a doubtful killed-update recovery on the repair path (#201361)
- fix(subagents): yielded parents lose subagent results when their run …
- fix: let Mac source builds package with the bundled Codex 0.169
Current
·
C2
·2026-11





