[macOS Desktop] app-server repeatedly SIGKILLed after update; crash feedback also fails

What version of the Codex App are you using?

ChatGPT/Codex desktop app 26.825.32147 on macOS.

What platform is your computer?

• Apple M4, arm64
• 16 GB RAM
• macOS 26.3 (build 25D125)

What issue are you seeing?

The macOS desktop app repeatedly displays “ChatGPT hit a snag” and asks for a restart. When attempting to stop or recover the affected task, it displays:

Failed to stop chat
Codex app-server is not available

The in-app Send feedback action also fails, so the crash cannot be reported from the affected app state.

This continues after installing the latest desktop update. The local app log confirms that the bundled Codex CLI/app-server is repeatedly terminated with SIGKILL, after which interactive and background requests fail with Codex app-server is not available.

Sanitized sequence from August 29, 2026:

05:19:20Z Codex CLI process exited classifiedAsExpected=false signal=SIGKILL
05:19:20Z fatal_error_broadcasted
05:19:21Z thread/resume requests failed: Codex app-server is not available
05:19:21Z composer submit failed: Codex app-server is not available
05:22:37Z app_server_connection.closed signal=SIGKILL
05:22:37Z Codex CLI process exited classifiedAsExpected=false signal=SIGKILL
05:22:37Z fatal_error_broadcasted
05:22:37Z turn/interrupt failed: Codex app-server is not available
05:22:38Z composer submit failed: Codex app-server is not available

The app repeatedly attempts automatic recovery and may briefly reconnect, but the app-server is terminated again minutes later.

Nearby warnings include:

skipping subagent thread analytics: missing inherited client metadata
ignoring interface.defaultPrompt[0]: prompt must be at most 128 characters

The second warning references a local ngs-analysis plugin manifest. These warnings may only be the last buffered messages rather than the cause of SIGKILL, but they are included because they recur near restart cycles.

Observed app telemetry near the affected session:

• App memory: approximately 2.3 GB
• Free system memory: approximately 447 MB

No prompt contents, account identifiers, local project paths, session transcripts, or full logs are attached.

What steps can reproduce the bug?

1. Install/update to desktop build 26.825.32147.
2. Open local Codex tasks on a 16 GB Apple Silicon Mac.
3. Run or resume tasks, including a task that may use plugins or subagents.
4. Leave the app running for several minutes.
5. Observe “ChatGPT hit a snag.”
6. Restart or attempt to stop the affected task.
7. Observe “Codex app-server is not available.”
8. Attempt to use Send feedback from the crash screen; submission fails.

What is the expected behavior?

• The bundled app-server should not be terminated during normal local task execution.
• Automatic recovery should restore active tasks without repeated restart loops.
• Stopping a task should remain possible after an app-server failure.
• The feedback uploader should remain usable when reporting a crash.

Additional information

Possibly related, but not exact duplicates:

• #40540 — macOS app-server memory growth and SIGKILL during thread hydration
• #40222 — app-server loss after workspace auth 401 and failed in-app feedback

Two screenshots are available showing the crash page and the “Failed to stop chat / Codex app-server is not available” messages. Sanitized logs can be provided privately if requested.

The payload names no source files or tests; start with the macOS desktop app's bundled Codex CLI/app-server process and its automatic recovery under the reported SIGKILL sequence. Check the Send feedback path from the failed state. Done means the app-server no longer enters repeated restart loops, task stopping or recovery remains possible, and crash feedback submission works.