Codex app automations run at medium reasoning even when global model_reasoning_effort is xhigh

[johnsilvavlogs]

What version of Codex is running?

• Extension host: Cursor + OpenAI Codex extension 0.4.79 (Windows)
• Automation run reports cli_version = 0.108.0-alpha.8
• Local global CLI binary: codex-cli 0.110.0

Which model are you using?

gpt-5.3-codex

What is the issue?

Automations appear to ignore global reasoning config:

• Global config is set to model_reasoning_effort = "xhigh".
• Manual non-automation runs start with effort = "xhigh".
• Automation runs (Codex app automation, e.g. Daily bug scan) start with effort = "medium" and summary = "auto".

This makes automations less reliable for complex bug triage compared to manual runs under the same environment.

Steps to reproduce

1. Set global config:
   • model = "gpt-5.3-codex"
   • model_reasoning_effort = "xhigh"
2. Create an automation in Codex app (for example, a recurring bug scan).
3. Let the automation run.
4. Inspect the run JSONL turn_context.
5. Compare against a normal manual thread in the same environment.

Expected behavior

Either:

1. Automations inherit global model_reasoning_effort (xhigh), or
2. Automations expose an explicit reasoning-effort setting (low/medium/high/xhigh).

Actual behavior

Automation run turn_context shows:

• effort = "medium"
• collaboration_mode.settings.reasoning_effort = "medium"

while manual runs show xhigh.

Notes

• The automation definition file currently has schedule/prompt/cwd but no reasoning field.
• A prompt-level workaround helps but does not guarantee deterministic model effort.

[github-actions]

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• Ability to set reasoning effort for automations #12239

Powered by Codex Action