Repeated Codex Desktop systemError on Windows during connector-heavy local Outlook/Odoo threads

What version of the Codex App are you using (From “About Codex” dialog)?

Codex Desktop 26.623.19656.0

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Microsoft Windows 11 Business, version 10.0.26200, 64-bit.

What issue are you seeing?

Several local Codex Desktop threads have become unusable with systemError during my real office-assistant workflows. This happened 3 times on July 8, 2026, and 2 additional times in the days before.

The most recent affected thread:

• Thread title: E-Mail-Bestand aufräumen
• Feedback ID / thread ID: 019f406d-0bbc-7353-a35b-27056b4cef23
• Thread status as reported by Codex Desktop thread tools: systemError
• The latest visible agent turn had completed, but the overall thread was marked as systemError.

The affected workflow was a single mail-processing topic in a local Codex Desktop thread, using connected tools such as Microsoft Outlook Email search/fetch/list attachments and Odoo-related local context. Shortly before the failure, the agent had run multiple Outlook searches and many attachment-listing calls while checking whether relevant documents were already stored before archiving email messages.

The failure interrupts my practical business workflows because the thread can no longer be relied on as the working context. I can work around it with external checkpoint files and smaller task chunks, but repeated systemError crashes make long-running assisted email/task processing fragile.

What steps can reproduce the bug?

This is from real usage rather than a minimal synthetic repro:

1. Start a local Codex Desktop thread on Windows.
2. Use the thread for a mail-triage / office-assistant workflow with Microsoft Outlook Email connector calls.
3. Keep the task scoped to one mail topic, but allow several searches and attachment checks.
4. The agent performs multiple Outlook search_messages calls and then many list_attachments calls.
5. The latest visible turn may appear completed, but the thread is later marked as systemError and becomes unreliable/unrecoverable as a working thread.

What is the expected behavior?

Connector-heavy but scoped local Desktop threads should remain recoverable. If a connector call, compaction, renderer sync, or replay step fails, Codex Desktop should show a specific recoverable error and preserve the completed visible turn state. Ideally there should be a safe resume/rebuild path.

Additional information

I already uploaded feedback through Codex Desktop:

019f406d-0bbc-7353-a35b-27056b4cef23

Please use that feedback bundle for logs/details. I am intentionally avoiding private email contents, local business data, raw message IDs, or Odoo records beyond the thread/feedback ID in this public issue.

Potentially related areas:

• Codex Desktop app
• local thread session/replay state
• connector-heavy tool calls
• context management / compaction
• safe recovery after systemError