[Desktop App] Long-running tasks abort when the selected model reaches capacity

Open 💬 2 comments Opened Jul 17, 2026 by JWMatheo

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of the Codex App are you using?

Version 26.715.21425 (build 5488)

What platform is your computer?

macOS 26.3, Apple Silicon, build 25D125

What issue are you seeing?

While a long-running task is executing, Codex stops with:

"Selected model is at capacity. Please try a different model."

The task is terminated instead of being queued, retried, or safely resumed. I lose the active task context and must abandon it and create a new task to continue working.

Evidence

First recorded system error: task 019f70f4-26ca-7e22-95ba-15d9e3312737, stopped at 2026-07-17 18:46:52 CEST.
Reproduced again in task 019f7103-15f5-7d40-af95-9d594c323e89, stopped at 2026-07-17 19:35:57 CEST.
A continuation task had to be created afterward because the original task could not continue.

Steps to reproduce

Start a long-running task in the Codex Desktop App.
Let it perform multiple tool calls and background operations.
Wait until the selected model becomes unavailable.
Observe that the task terminates with the capacity error.
Try to continue the same task.

Expected behavior

The app should queue and retry the request, or offer a model switch while preserving the task context. Temporary model capacity should not permanently terminate the task.

Related issue

#33865 reports a similar capacity error for the CLI. This report concerns the Desktop App and loss of task continuity.

View original on GitHub ↗

2 Comments

github-actions[bot] contributor · 1 month ago

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

#33853
#33865
#33661
#33171

Powered by Codex Action

tylorsaling-source · 1 month ago

Additional Windows Desktop reproduction of this exact failure mode.

Environment

Codex Desktop: 26.721.4979.0
Platform: Windows
Model: gpt-5.6-sol, high reasoning
Turn: 019fa62c-0b6f-7bf0-ba34-e09aad4b8277

Measured event

Started: 2026-07-27 20:42:12 EDT
Stopped: 2026-07-27 21:07:13 EDT
Duration: 1,501,434 ms (~25 minutes)
Internal error: server_overloaded
Message: Selected model is at capacity. Please try a different model.
Terminal record: task_complete with last_agent_message: null

The important additional defect is that the error message was not shown to the user. The Desktop UI showed only a red failure icon beside the thread. There was no banner, transcript message, notification, retry control, or model-switch prompt. The user had to inspect the rollout record to discover why the agent stopped.

Expected behavior: preserve the turn, enter a visible paused/retrying state, and offer Retry, Switch model and resume, or Cancel. A capacity failure must not appear as unexplained completion or handback. A configurable automatic fallback to the closest available model would also be useful, with any model change disclosed.