Codex Desktop should surface active service incidents before costly long-running agent work

Summary

During a long-running Codex Desktop repository session on July 23, 2026, the official OpenAI status page showed an active elevated-error-rate incident that had been monitored for hours. Codex Desktop did not display an incident banner, identify potentially affected components, or recommend pausing expensive autonomous work.

The same session later entered repeated context compaction and exhausted most of the included allowance. The public incident is included as correlation only; this report does not claim it caused the compaction defect.

Environment

• Product: Codex Desktop
• Platform: Windows 11
• Plan: ChatGPT Pro
• Workload: long-running, tool-heavy private repository coordination
• Related compaction report: #35032

Observed behavior

1. An official OpenAI incident was active and visible on status.openai.com.
2. Codex Desktop continued a costly long-running agent session without surfacing that incident.
3. The user had to leave the product and manually check the status page.
4. No in-product action explained whether to continue, pause, preserve a handoff, or collect diagnostics.
5. Usage continued to be consumed while the session was already behaving abnormally.

Expected behavior

When an official incident may affect Codex or an active dependency, Codex Desktop should show a first-party banner containing:

• incident title and current phase;
• affected product/components;
• start/update timestamp;
• a link to official status details;
• whether the local session has correlated errors;
• options to continue, pause after the current safe boundary, or create a bounded handoff;
• a reminder that correlation does not prove causation.

For expensive autonomous/goal runs, the app should warn before beginning or continuing while a relevant incident is active. It should not silently cancel user work, but it should make the risk and usage implications visible.

Impact

OpenAI can know about a service degradation in near real time while the affected paying user remains outside the loop and continues consuming allowance. This turns incident communication into something the user must discover manually or through third-party/social channels.

Suggested acceptance test

Inject a fixture Codex incident into the status feed while a long-running goal is active. Verify that Desktop displays a non-blocking banner, preserves current work, offers a bounded pause/handoff, and records the incident correlation in diagnostics without claiming causality.