Intermittent AppServerBackendRequestError / UNKNOWN when sending follow-ups to cloud tasks

Codex Bugs
1

Environment

  • Surface: ChatGPT Web on macOS, using Atlas to coordinate saved GitHub-backed cloud tasks
  • Codex version: Not collected
  • OS/browser: macOS; exact OS and browser versions not collected
  • Model/authentication: Signed-in ChatGPT account; exact execution model and plan not confirmed for this report

Bug
Assistant-initiated follow-up messages to existing cloud tasks repeatedly failed with AppServerBackendRequestError / UNKNOWN. A fresh-task creation attempt also failed. Reading existing tasks and listing environments continued to work.

Expected: The instruction starts a task turn, or the error clearly explains why it failed and whether any work was admitted. An uncertain failure makes it difficult to retry safely without duplicating work.

Reproduction

  1. Use a saved GitHub-backed cloud environment with an existing completed task.
  2. Ask the assistant to send continuation instructions to that task.
  3. Observe AppServerBackendRequestError with error_code UNKNOWN.
  4. Read the task again: the previously completed turn is still the latest recorded turn.
  5. Try creating a fresh task using the saved environment. In our case, creation also failed and the subsequent task listing showed no new task.

Diagnostics
Exact error category: AppServerBackendRequestError
Error code: UNKNOWN

Observed on October 15, 2026, with repeated failed continuation attempts around 10:07–10:13 UTC and a failed fresh-task creation around 10:21 UTC. Task reads continued to succeed afterward without showing new execution.

Recovery: The original task completed a new turn around 15:41 UTC. An assistant-initiated continuation was then accepted and started around 15:42 UTC; later follow-ups also succeeded. We have not established what restored service or whether recovery is permanent.

A separate missing PHP/PDO runtime issue was resolved inside the task. That should not be conflated with the task-message admission failure.

Related report with closely matching symptoms:

Could OpenAI clarify:

  • How can we determine whether a failed request actually admitted a task or turn?
  • Is there a supported recovery procedure that preserves existing work?
  • Can UNKNOWN errors expose a more actionable cause and safe retry guidance?

Private task/environment identifiers and diagnostic request IDs can be supplied privately to Support.

1 Like