Codex Web: GitHub @codex mention tasks fail with "An unknown error occurred" while the identical prompt succeeds from the composer
Nobody has claimed this yet.
- Dominant language
- Rust
- Stars
- 125k
- Forks
- 19.4k
- PR merge metrics
- PR metrics pending
Description
What issue are you seeing?
Since 2026-08-31 ~16:00 ET, every Codex cloud task launched from an @codex mention on a GitHub issue fails. The task detail page shows only:
Failed after 1m 07s
An unknown error occurred
and the connector posts back to the issue thread:
Codex couldn't complete this request. Try again later.
What steps can reproduce the bug?
- Private repo with a Codex cloud environment configured (image universal, unrestricted internet, one secret, a one-line setup script).
- Post a comment on an issue in that repo beginning with @codex, giving the agent a task ("read AGENTS.md, add one file, push branch X").
- The task appears in chatgpt.com/codex/cloud as "GitHub Mention: ", runs for one to three minutes, then ends as Failed with "An unknown error occurred". The connector posts "Codex couldn't complete this request."
- Paste the exact same prompt text into the Codex composer, selecting the same repo and branch. It completes normally.
What is the expected behavior?
A mention-triggered task should execute the same way a composer-triggered task does against the same repo and environment. Failing that, the error surfaced on the task page and in the issue reply should say what actually went wrong, rather than "An unknown error occurred" — there is nothing in that message to act on.
Additional information
Plan: ChatGPT Pro. Component: Codex Web / GitHub integration.
Ruled out, with evidence:
- Usage or plan limits — composer tasks run fine in the same minutes that
mention tasks fail. - Broken environment or setup script — the composer task clones the repo,
runs the setup script, reads files, commits and pushes successfully. - GitHub App permissions / repo grant — the ChatGPT Codex Connector
installation has read on checks, commit statuses and metadata, and read+write
on actions, code, issues, pull requests and workflows, scoped to the repo in
question, with no pending approval requests. The connector demonstrably reads
the mention and writes a reply comment on every failed dispatch, so both
directions of the GitHub link are working. - Stale container cache — resetting the environment's post-setup cache moved
the failure point from 1m 07s to 2m 49s (it now rebuilds and clears the setup
scripts) but did not fix it. This does suggest the failure is not in the
environment build itself.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reproducing the GitHub @codex mention dispatch and compare its Codex cloud task path with the composer path using the same repository and environment. Trace the mention-triggered task and connector reply entry points; done means mention tasks execute successfully and failures expose an actionable error instead of "An unknown error occurred."
Written by the indexing model from the issue text.
Assessment
- Tech stack
- github, rust
- Domain
- api, backend, devtools
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 45/100