Codex desktop / GPT-6.1 Sol: apparent stall and delayed turn completion after tool work finishes #51656

What version of the Codex App are you using?

26.930.61225 (installed application bundle metadata).

What subscription do you have?

Not included in this public report.

What platform is your computer?

macOS, Apple Silicon. uname -mprs: Darwin 27.0.0 arm64 arm.

Model

GPT-6.1 Sol (gpt-6.1-sol). Reasoning setting for the affected turn has not been independently verified.

What issue are you seeing?

On October 7, 2026, during a long-running repository maintenance and release task in the Codex desktop app, the user experienced extremely slow progress and an apparent stall. Most importantly, after the external work and release verification had completed, the turn still did not promptly end with a final response. The user interrupted/switched models; the resumed assistant performed a small final confirmation and returned the completion summary.

This report is submitted by the assistant at the user's explicit request. The user's report was: “6.1 sol is extremely slow and just got stuck; report an issue.”

What steps can reproduce the bug?

No deterministic minimal reproduction is available. Observed scenario:

1. Use GPT-6.1 Sol in an existing long conversation for multi-step code, verification, and release work.
2. Allow the agent to complete its tool operations and final release bookkeeping.
3. Observe that the turn continues without delivering the final completion response, appearing stuck to the user.
4. Interrupt/switch models and resume; a concise completion response becomes possible after a small verification pass.

The final release-restoration check completed around 2026-10-07 10:49 UTC, with release bookkeeping afterward. Exact stall duration, token usage, and server-side cause are not established. The overall task also involved legitimate tool waits and recovery from external service errors, so the full elapsed task duration must not be interpreted as model inference latency.

What is the expected behavior?

After the requested work and relevant checks finish, return the final response promptly. Detect a turn that is making no useful progress, surface an actionable status/error, and support recovery without requiring a manual model switch.

Additional information

Please investigate turn completion, excessive inference latency, and no-progress detection for GPT-6.1 Sol in long-running desktop tasks. This observation does not establish whether the fault is in model inference, orchestration, transport, or the UI.

No private repository contents, credentials, conversation transcripts, or raw diagnostic logs are attached. A superficially related report is #50455, but this report concerns failure to finish after tool work and does not assert its unrelated-text/reconnection symptoms.