Codex App on Windows crashes after legacy_notify hits os error 206

What version of the Codex App are you using?

26.623.11225.0

Observed installed package:

OpenAI.Codex_26.623.11225.0_x64__2p2nqsd0c76g0

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

The Codex App on Windows crashed to the full-screen "An error has occurred" page after an after_agent notification hook failed with Windows native error 206.

Exact error text from the crash screen:

An error has occurred
Codex crashed with the following error:

  (code=3221225786, signal=null).

Most recent error: {"timestamp":"2026-07-01T22:01:56.531371Z","level":"WARN","fields":{"message":"after_agent hook failed; continuing","turn_id":"e0e2450a-b919-449e-9490-2d5a1df228d7","hook_name":"legacy_notify","error":"The filename or extension is too long. (os error 206)"},"target":"codex_core::hook_runtime"}

This looks like the Desktop/App version of a Windows notify transport bug. A hook failure whose own log message says continuing should not be able to take down the Codex App control plane.

I searched before filing:

• Related but not exact: https://github.com/openai/codex/issues/18309
• Windows notify/turn-complete payloads hit native error 206 when a large notify payload is passed through Windows process creation.
• Related but not exact: https://github.com/openai/codex/issues/8929
• Notify hooks not triggering on Windows/WSL.
• Related error class but different surface: https://github.com/openai/codex/issues/9688
• Windows os error 206 when large generated tool payloads are applied/appended.

I did not find an exact existing issue for:

• after_agent hook failed
• hook_name="legacy_notify"
• codex_core::hook_runtime
• Codex App crash screen with code=3221225786

What steps can reproduce the bug?

The exact minimal repro is not fully isolated yet, but the observed path is:

1. Use Codex App on Windows in a long-running, high-context coding thread.
2. Let an agent turn complete after substantial prompt/context/history has accumulated.
3. Codex attempts the legacy notify hook after the agent turn.
4. The hook fails with Windows error 206:

``text
The filename or extension is too long. (os error 206)
``

1. Instead of treating the hook failure as non-fatal, the Codex App crashes to the error screen with:

``text
code=3221225786, signal=null
target="codex_core::hook_runtime"
message="after_agent hook failed; continuing"
hook_name="legacy_notify"
``

This is consistent with the Windows command-line/path-length failure mode described in #18309, except here it is surfacing through Codex App / legacy_notify and crashing the app UI.

What is the expected behavior?

Codex should not crash because a notification hook fails.

Expected behavior:

1. Do not pass large hook/notify payloads through Windows argv/path-sensitive process creation.
2. Use stdin, a temp file, a bounded digest payload, or another Windows-safe transport for hook payloads.
3. If a hook fails, report a non-fatal diagnostic and continue the session.
4. Preserve the active thread, terminal/process state, tool state, and queued work.
5. Keep the error visible enough for debugging without making the app unusable.

Additional information

This is part of a pattern of serious Windows Codex App stability failures during real coding work. In this same workstream, separate Codex crashes have cost significant time, tokens, and disrupted long-running build/proof work. These are not cosmetic issues: for a paid Pro workflow, losing agent continuity and having to spend time keeping the app alive is a major reliability problem.

I am disappointed by the Windows App stability compared with the macOS experience and compared with other desktop AI coding tools. The Windows app currently feels fragile under serious long-running engineering sessions, which makes it hard to use effectively on a Windows workstation.

The fix should be structural, not just better error wording: hook/notify execution should have a Windows-safe payload transport and hook failure isolation so a best-effort notification cannot crash the app control plane.