Feature Request: Add option to display reasoning summaries (thinking blocks) in Codex macOS app

@etraut-openai This seems like the same underlying request as #8204 (stream reasoning summaries incrementally so the UI doesn’t feel frozen during long thinking). Suggest consolidating into one canonical issue (e.g. add an app label to #8204 and close this as a dup) to keep discussion + prioritization in one place.

@swordfish444, I think these are unrelated. This is specifically about the Codex app and the presence of reasoning summaries. The other is about streaming reasoning summaries in the CLI.

@etraut-openai Good call — I was mapping this to the broader “UI feels frozen during long thinking” request, but agreed this is app-specific (displaying summaries at all) vs CLI streaming behavior. Thanks for the quick correction.

please implement this!!

Thanks — I dug into this on macOS and it looks like the app is currently dropping reasoning items at render time.

• The session JSONL still contains reasoning summaries (payload.type="reasoning" with non-empty payload.summary).
• In the desktop app bundle, the webview renderer explicitly filters them out: d.type==="reasoning"&&(l=null) (found in Codex.app/Contents/Resources/app.asar → webview/assets/index-*.js).
• I also see code that constructs {type:"reasoning", ...} items from payload.summary, so they’re being created and then nulled during rendering.

Tested on 26.212.1823 and 26.217.1959 — same behavior.

Request: please gate this behind a user-visible toggle / respect the existing config flags, and when enabled render collapsed-by-default with i18n’d labels + optional “Translate” action for the summary.

Adding a toggle to the macOS app to toggle show thinking output like the picture. Would be nice. The image is of the ollama app. I think the regular ChatGPT website even does this.

<img width="770" height="431" alt="Image" src="https://github.com/user-attachments/assets/fcec5d2e-c3dd-4412-add5-67432b259521" />

I also need this feature, and it includes the codex cli and the codex app. I often observe Codex in the process of thinking, but I have no idea what he is thinking about. Only when the "phase" stage appears can I see the brief response.

Agreed. It's often useful to read reasoning summary to evaluate model's performance and logic behind its behaviour.

Agree, reasoning summaries will be very helpful and useful.

Adding a current macOS reproduction with GPT-5.6 Sol Ultra and rollout-level evidence.

Environment

• ChatGPT/Codex Desktop: 26.727.51351 (build 6119)
• Bundled Codex core recorded in the rollout: 0.146.0-alpha.9.2
• macOS arm64
• Model: gpt-5.6-sol
• Reasoning effort: ultra

User-level ~/.codex/config.toml includes:

model_reasoning_summary = "detailed"
model = "gpt-5.6-sol"
model_reasoning_effort = "ultra"

hide_agent_reasoning is not enabled.

Observed behavior

During a long, tool-heavy Desktop task, the conversation showed visible commentary/progress messages and grouped action cards, but no reasoning-summary block at any point.

The local rollout proves that reasoning summaries were generated and persisted:

reasoning_items=71
nonempty_reasoning_items=39
summary_sections=73
agent_reasoning_events=73

None of those readable summary sections appeared in the Desktop transcript.

The rollout's effective turn context recorded:

model = gpt-5.6-sol
effort = ultra
summary = auto

That is despite the explicit global model_reasoning_summary = "detailed". This suggests two related Desktop-path problems:

1. The renderer omits persisted reasoning-summary items.
2. The Desktop/app-server turn override changes the configured summary mode from detailed to auto.

The first problem remains even under auto, because 73 readable summary sections were emitted and persisted but not shown.

Expected behavior

This is not a request for private/raw chain-of-thought. The app should expose the readable reasoning summaries that Codex already emits and persists, ideally behind an in-app display toggle, and it should respect model_reasoning_summary.

For long-running engineering work, this is operationally important: visible summaries let the user catch a wrong assumption and redirect the task before it spends substantial time or resources.

All task prompts, task/session identifiers, repository details, filenames, commands, infrastructure details, and actual summary text have been omitted.

Desktop supports live and retained reasoning summaries.