Simple Codex Pro tasks consistently take ~15 minutes with no visibility into queueing or backend delay

What issue are you seeing?

Simple Codex tasks are consistently taking around 15 minutes to finish even when the actual work is trivial.

A representative example is basic GitHub/repository administration: reading a small amount of existing context and making straightforward Issue metadata/organization changes. The task itself is not computationally complex and does not justify a ~15 minute wall-clock delay.

The main problem is not only raw latency, but the lack of visibility into where the time is going. From the UI it is impossible to distinguish among:

• backend queueing / model capacity;
• throttling or account-level scheduling;
• retries/timeouts;
• context loading or compaction;
• tool-call latency;
• model reasoning time;
• a stuck execution step.

Because the delay is repeatable even for simple tasks, it feels like the system may be spending substantial time waiting rather than doing useful work.

I am not claiming that Codex intentionally delays tasks. I am asking for the actual source of this latency to be investigated and made observable.

Steps to reproduce

1. Open Codex and use GPT Pro mode.
2. Give it a small, well-scoped task that requires little implementation work, for example straightforward GitHub Issue/metadata organization.
3. Let Codex execute the task normally.
4. Observe that completion can take roughly 15 minutes despite the small scope.
5. Repeat with other similarly small tasks; the long wall-clock latency recurs.

Expected behavior

For small, well-scoped tasks, wall-clock completion time should roughly reflect the amount of actual reasoning/tool work required.

If Codex is waiting on backend capacity, queueing, rate limits, retries, context processing, or another service dependency, the UI should expose that state instead of presenting the entire delay as undifferentiated task execution.

At minimum, it would be useful to expose separate timing/status for:

• queued / waiting for model capacity;
• model reasoning;
• context loading / compaction;
• tool execution;
• retry/backoff;
• final response generation.

That would make it possible to tell whether a 15-minute task is genuinely doing 15 minutes of work or spending most of that time waiting on infrastructure.

Additional context

This appears related in spirit to earlier latency/performance reports such as #21527, #24428, and #27576, but this report is specifically about repeatable ~15-minute completion times for very simple tasks and the absence of queue/backend-latency visibility.

Observed: 2026-08-18.

Exact Codex app version and platform can be added if maintainers need them for diagnosis.

Start by reproducing the delay in Codex GPT Pro mode with a small GitHub Issue or metadata task, recording the timing and the app version and platform when available. Investigate whether the elapsed time is queueing, model reasoning, context processing, tool execution, retries, or final response generation; done means the source of the delay is identified and those states are visible in the UI.