Codex: send the selected reasoning effort in thread/start (mismatch with the Codex default makes terra/luna end turns early) #33823

Follow-up to #32488, which was closed as upstream. This one is about a specific change in T3 that avoids the early stops. Background and full data: openai/codex#60133 and the last comments in #32488.

Problem

T3 sends the selected reasoning effort only in turn/start. thread/start carries the model but no effort, so the Codex thread starts with the user's own default (model_reasoning_effort in ~/.codex/config.toml, or the model default if unset).

When the two differ, Codex adds a configuration_update item ({"reasoning":{"effort":"..."}}) to the turn, and gpt-6.1-terra and gpt-6-luna then end the turn after a few tool calls with the task unfinished. The underlying behaviour is a Codex bug, but T3 can avoid triggering it.

Steps to reproduce

1. Set model_reasoning_effort = "medium" in ~/.codex/config.toml.
2. In T3, start a Codex thread (or a delegate_task child) with gpt-6.1-terra and reasoning effort high.
3. Give it a long multi-step task, for example processing 9 input files one at a time and running a check at the end.

The turn ends after 3 to 6 tool calls, in under about a minute, and the final message says the work is incomplete. With effort medium in step 2 the same task completes.

Data

Replays of T3's own thread/start + turn/start against a bare codex app-server (codex-cli 0.167.0, Windows 11, Nov 10). Only the effort placement changes; config.toml default is medium.

All 6 failing runs have the configuration_update item in the rollout. None of the 6 passing runs do.

Suggested change

In T3 Code (Nightly) 0.1.9-nightly.20261110.4465, codexThreadRuntimeParams (Codex adapter v2, next to CODEX_THREAD_CONFIG) receives input.modelSelection but forwards only the model:

...input.modelSelection === void 0 ? {} : { model: input.modelSelection.model },
config: {
  ...CODEX_THREAD_CONFIG,
  ...
}

Adding model_reasoning_effort: <selected effort> to that config object when an effort is selected should be enough (rows 3 and 6 above).

Related: when delegate_task is called without an effort, the child ran at low for gpt-6.1-terra while my Codex default was medium, so it hit the same mismatch and stopped after 56 seconds.

I only tested the first turn of a thread. Changing effort in the middle of an existing thread probably still needs the Codex fix.

Workaround

Set the Codex instance's "Launch arguments" to the effort you use:

-c model_reasoning_effort="high"

Tested as codex app-server -c model_reasoning_effort="high" with gpt-6.1-terra and turn effort high, while config.toml still says medium: 2/2 completed (195 s, 227 s), no configuration_update item. One instance covers one effort, so this does not replace the change above.

Environment

T3 Code (Nightly) desktop 0.1.9-nightly.20261110.4465, Windows 11 Pro (10.0.26200), codex-cli 0.167.0.