Repeated HTTP 520 from Codex responses endpoint interrupts desktop tasks and reviewers

What version of the Codex App are you using?

26.908.61612 (build 9218), read from the installed desktop application's Info.plist.

What subscription do you have?

Not verified for this report.

What platform is your computer?

macOS 26.6.2 (build 25G83).

What issue are you seeing?

Repeated HTTP 520 responses from https://chatgpt.com/backend-api/codex/responses interrupt an ongoing local coding task and independent review agents. The user reports 19 occurrences. Multiple failures were observed during continuation attempts, and two review attempts also failed with HTTP 520.

The failing endpoint is the Codex responses service. Local shell commands and saved workspace artifacts remain accessible. The exact origin/edge cause is unknown.

What steps can reproduce the bug?

1. Work on an existing local repository repair and validation task in the desktop app.
2. Continue the task after an interruption.
3. The remote response request intermittently fails with HTTP 520.
4. Retrying and asking an independent review agent also produced HTTP 520 failures.

This is intermittent; no minimal deterministic reproducer is available. A smaller independent review with no inherited conversation history later succeeded once. That observation does not establish a context-size cause or a reliable workaround.

What is the expected behavior?

The task should continue reliably, or the app should provide an actionable error with a traceable support reference and a recovery path that preserves progress.

Additional information

Observed affected turn start times (UTC; these are turn start times, not exact HTTP response timestamps):

• 2026-09-14 23:41:56
• 2026-09-14 23:52:06
• 2026-09-15 00:01:24
• 2026-09-15 00:19:17

Cloudflare Ray IDs and task identifiers have been retained privately and were supplied through OpenAI's Help Center support chat. They are omitted here along with the account email, private prompts, source code, local paths, and raw logs.

The Help Center assistant acknowledged the report and contact email, stated that it could not create a case or provide a human handoff in that chat, and directed us to this issue tracker.

The OpenAI status page listed an incident titled “Elevated errors affecting Work Mode in ChatGPT”:
https://status.openai.com/incidents/01M2GA8XTS6VB3QCDEGZ0HNAQ5
A relationship between that incident and these specific HTTP 520 failures has not been confirmed.

App restart, sign-out/sign-in, VPN changes, and a different network have not been verified as tested. Please investigate repeated response-generation failures and advise on mitigation and a private channel for correlating the retained identifiers.

Start with the Codex responses endpoint at https://chatgpt.com/backend-api/codex/responses and correlate the reported failures with the privately retained Cloudflare Ray IDs and task identifiers. Compare the intermittent 520 failures with the linked Work Mode incident and determine whether the app preserves progress and provides an actionable recovery path.