TECH Signal 506
Codex desktop app reportedly experiences 404 errors breaking long-running AI sessions
OpenAI’s Codex desktop app is returning 404 errors that terminate multi-turn AI sessions mid-conversation.
Engineers who rely on long-running Codex sessions for orchestration or sub-agent workflows now face abrupt failures. The outage highlights a lack of resilient retry logic in the client, increasing operational friction for users on paid plans.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
A 404 error on the `/backend-api/codex/responses` endpoint is crashing Codex desktop sessions after steering or compaction events.
The issue affects specific GPT model configurations and session states, not all users or sessions simultaneously.
Users report the app retries only five times before giving up, with no exponential backoff or extended recovery window.
THE READ
What the cluster adds up to.
The Codex desktop app is currently returning 404 errors on its backend API endpoint, causing active AI sessions to terminate unexpectedly. This occurs during multi-turn interactions, particularly after steering commands or session compactions, and is isolated to certain GPT model versions and session states. The error does not appear universal; some users report unaffected sessions running different model configurations.
The failure mode suggests a backend routing or endpoint deprecation issue rather than a transient network problem. The app’s retry logic is limited to five attempts without exponential backoff, which is insufficient for recovering from what may be a backend service change. Users on paid plans are impacted, indicating the issue is not restricted to free-tier access or rate limits.
For engineers using Codex for orchestration or sub-agent workflows, this outage introduces operational risk. Long-running sessions with accumulated context are lost mid-task, requiring manual restart and context reconstruction. The lack of resilient error handling in the client means users must either avoid complex workflows or implement external monitoring and recovery scripts until the issue is resolved.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER