Skip to content

An Unattended Worker Backgrounded Its Review Wait and Stopped the Run on One Issue #1837

Description

@ptr727

In one unattended run, the worker on handoff #1835 broke two of the unattended-handoff skill's own worker rules in the same round.

What happened

  1. It waited in the background. Worker step 4, "Wait in the foreground", says each wait is one bounded command run in the worker's own turn, since a subagent receives no completion notification. The worker instead handed its local strict review pass to a background subagent and handed back before that pass reported.
  2. It returned STOP for a per-issue condition. The skill reserves STOP for "a state of the repository or the session rather than of one issue". An unfinished review pass on one branch is a state of one issue, so the right outcomes were to wait for the pass and continue, or to park per "Parking". STOP ended the whole run after three rounds with backlog still pickable.
  3. The handoff was left neither closed nor parked. It carries no blocked label and no decision issue, so a later picker reads it as a lane whose worker died, and the maintainer's attended session gets no parked signal for it.

The pass later returned clean, the worker woke and commented the state on #1835, then stopped again with the fix still unpushed.

Proposed direction

  • Make worker step 4 name the local strict review's subagent explicitly: its reviewer is dispatched and waited on in the foreground, never backgrounded.
  • State in "Return Lines" that a worker which cannot finish one issue parks it rather than returning STOP, and list what does qualify for STOP from a worker.

Both are rule text, so the wording is the maintainer's to settle in review.

Activity

  1. added
    bugSomething isn't working
    agentsAgents instructions
    on Sep 25, 2026
  2. ptr727 commented on Oct 9, 2026

    @ptr727
    OwnerAuthor

    Recurred in a later unattended run, round 10 of 10, on handoff #2643 (track auto-2590, a Sonnet-tier worker). The worker handed its local strict review pass to a background subagent and returned STOP naming the unfinished pass, with the fix committed and unpushed. This time the pass returned, the worker woke and finished the lane on its own: #2644 and promotion #2645 merged, #2590 closed, the handoff closed out, and the worktree removed. The run had already ended on the STOP, though, so 10 rounds of cap went unused, and the orchestrator's one-line record of that round was wrong until it was re-read. Both failure modes in the issue body held, plus a third: a STOP line the worker later contradicted by finishing.

  3. added a commit that references this issue on Oct 9, 2026
    01ffaa3
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    agentsAgents instructionsbugSomething isn't workingskillsAgent skill

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions