[Windows] Error submitting message: turn/start fails with agent loop died unexpectedly

Codex app version

• Microsoft Store package: 26.721.4979.0
• Desktop log release: 26.721.41059

Platform

• Windows 11 Pro 64-bit
• Version 10.0.26100, build 26100

Issue

Sending a follow-up message to an existing completed local task intermittently fails immediately. The UI only shows:

Error submitting message

The desktop log contains the more specific error:

method=turn/start errorCode=-32603
failed to start turn: internal error; agent loop died unexpectedly

This occurred in two unrelated conversations during the same app session:

• 2026-07-29 12:40:15Z: turn/start failed in about 2 ms.
• 2026-07-29 12:46:27Z: turn/start failed in about 1 ms.

The failed user messages were not added to the rollout JSONL, so the failure occurs before a new turn is created.

Context-size comparison

The same error occurred for both:

1. A very large conversation:

• rollout JSONL approximately 120 MB
• approximately 79,840 events
• 77 context compactions
• recent model input approximately 119k tokens out of a 258.4k context window

1. A much smaller conversation:

• rollout JSONL approximately 5.7 MB
• 652 events
• one context compaction
• recent model input approximately 44k tokens out of a 258.4k context window

Neither rollout contains context_length_exceeded, server_overloaded, or a model error associated with these failed submissions.

Steps to reproduce

1. Keep multiple Codex tasks running in the desktop app.
2. Open an existing completed local task.
3. Enter a short follow-up prompt.
4. Submit it.
5. Intermittently, the UI shows Error submitting message and no new turn is recorded.

Expected behavior

A new turn should start, or the UI should surface an actionable backend error and recover without requiring the whole desktop app to restart.

Additional observations

• Just before the smaller conversation was used, thread/resume took approximately 10 seconds.
• The log also recorded Conversation state not found for the avatar-overlay renderer during resume, although the primary renderer subsequently reported maybe_resume_success.
• Starting a turn in a different conversation succeeded roughly 27 seconds after the second failure.
• The user could not restart Codex immediately because other long-running tasks were active.
• No session files or application state were modified while diagnosing this issue.

Relevant sanitized log excerpts

2026-07-29T12:40:15.204Z ... method=turn/start ... errorCode=-32603 ... durationMs=2
2026-07-29T12:40:15.223Z ... Request failed ... "failed to start turn: internal error; agent loop died unexpectedly"

2026-07-29T12:46:27.901Z ... method=turn/start ... errorCode=-32603 ... durationMs=1
2026-07-29T12:46:27.927Z ... Request failed ... "failed to start turn: internal error; agent loop died unexpectedly"