claude exits 1 twice on vps-showdown#41, issue abandoned as blocked #325

New issue

Closed

opened 2026-07-02 21:00:40 +00:00 by igor · 1 comment

igor commented

2026-07-02 21:00:40 +00:00

Collaborator

Diagnosis

The agent claimed joshtronic/vps-showdown#41 (Internal linking audit) on two consecutive ticks in this hour. Both invocations of claude terminated with a nonzero exit (exit 1) well before the 30m timeout (657s and 566s elapsed), producing a dirty-but-uncommitted worktree each time. Ship-safety correctly refused to commit the partial work and re-queued, but the second failure tripped the repeated-noop guard and the issue is now blocked with no successful run in the window. The task was abandoned, not completed. Memory peaks were unusually high (4.7G / 3.4G), which may point to the underlying cause of the crash.

This is not a retry-that-succeeded: both attempts failed, and the net result is a stuck issue.

Evidence

Jul 02 15:23:28 igor tick.sh[2013737]: [agent] claiming joshtronic/vps-showdown#41: Internal linking audit — connect the comparison pages to deepen the traffic funnel (Priority #3)
Jul 02 15:23:30 igor tick.sh[2013737]: [agent] invoking claude (timeout 30m)
Jul 02 15:34:27 igor tick.sh[2013737]: [agent] claude exited 1 (elapsed 657s)
Jul 02 15:34:27 igor tick.sh[2013737]: [agent] ship-safety: claude exited 1 mid-run -- NOT committing partial worktree (5 dirty file(s)); re-queuing
Jul 02 15:34:28 igor tick.sh[2013737]: [agent] outcome: no work produced (elapsed 657s)
Jul 02 15:34:29 igor systemd[785]: agent.service: Consumed 10min 6.933s CPU time, 4.7G memory peak.
...
Jul 02 15:36:16 igor tick.sh[2018051]: [agent] claiming joshtronic/vps-showdown#41: Internal linking audit ...
Jul 02 15:36:17 igor tick.sh[2018051]: [agent] invoking claude (timeout 30m)
Jul 02 15:45:43 igor tick.sh[2018051]: [agent] claude exited 1 (elapsed 566s)
Jul 02 15:45:43 igor tick.sh[2018051]: [agent] ship-safety: claude exited 1 mid-run -- NOT committing partial worktree (4 dirty file(s)); re-queuing
Jul 02 15:45:44 igor tick.sh[2018051]: [agent] outcome: blocked (repeated noop, prior count: 1)

What "fixed" looks like

A subsequent tick completes vps-showdown#41 and opens a PR (or the issue is deliberately deferred with a recorded reason), rather than claude crashing with exit 1 mid-run. Root-cause the exit 1 — capture the tail of claude's stderr on nonzero exit so the crash reason (OOM vs. tool error vs. content) is visible, given the 4.7G memory peak suggests a possible resource blowout on this particular issue.

service: agent.service
severity: medium
window: 2026-07-02 15:00-16:00 (filed by the hourly logwatch pass)

## Diagnosis The agent claimed `joshtronic/vps-showdown#41` (Internal linking audit) on two consecutive ticks in this hour. Both invocations of `claude` terminated with a nonzero exit (exit 1) well before the 30m timeout (657s and 566s elapsed), producing a dirty-but-uncommitted worktree each time. Ship-safety correctly refused to commit the partial work and re-queued, but the second failure tripped the repeated-noop guard and the issue is now `blocked` with no successful run in the window. The task was abandoned, not completed. Memory peaks were unusually high (4.7G / 3.4G), which may point to the underlying cause of the crash. This is not a retry-that-succeeded: both attempts failed, and the net result is a stuck issue. ## Evidence ``` Jul 02 15:23:28 igor tick.sh[2013737]: [agent] claiming joshtronic/vps-showdown#41: Internal linking audit — connect the comparison pages to deepen the traffic funnel (Priority #3) Jul 02 15:23:30 igor tick.sh[2013737]: [agent] invoking claude (timeout 30m) Jul 02 15:34:27 igor tick.sh[2013737]: [agent] claude exited 1 (elapsed 657s) Jul 02 15:34:27 igor tick.sh[2013737]: [agent] ship-safety: claude exited 1 mid-run -- NOT committing partial worktree (5 dirty file(s)); re-queuing Jul 02 15:34:28 igor tick.sh[2013737]: [agent] outcome: no work produced (elapsed 657s) Jul 02 15:34:29 igor systemd[785]: agent.service: Consumed 10min 6.933s CPU time, 4.7G memory peak. ... Jul 02 15:36:16 igor tick.sh[2018051]: [agent] claiming joshtronic/vps-showdown#41: Internal linking audit ... Jul 02 15:36:17 igor tick.sh[2018051]: [agent] invoking claude (timeout 30m) Jul 02 15:45:43 igor tick.sh[2018051]: [agent] claude exited 1 (elapsed 566s) Jul 02 15:45:43 igor tick.sh[2018051]: [agent] ship-safety: claude exited 1 mid-run -- NOT committing partial worktree (4 dirty file(s)); re-queuing Jul 02 15:45:44 igor tick.sh[2018051]: [agent] outcome: blocked (repeated noop, prior count: 1) ``` ## What "fixed" looks like A subsequent tick completes `vps-showdown#41` and opens a PR (or the issue is deliberately deferred with a recorded reason), rather than `claude` crashing with exit 1 mid-run. Root-cause the exit 1 — capture the tail of claude's stderr on nonzero exit so the crash reason (OOM vs. tool error vs. content) is visible, given the 4.7G memory peak suggests a possible resource blowout on this particular issue. --- service: agent.service severity: medium window: 2026-07-02 15:00-16:00 (filed by the hourly logwatch pass)

joshtronic was assigned by igor

2026-07-02 21:00:40 +00:00

igor added spent time

2026-07-02 21:00:40 +00:00

32 seconds

igor referenced this issue

2026-07-02 21:19:05 +00:00

Preserve claude's raw stream on a nonzero exit, so a claude-bomb is diagnosable #326

igor commented

2026-07-02 21:19:05 +00:00

Author

Collaborator

Closing (Igor / CoS) — this is a scope + visibility matter, not a harness bug. The harness behaved correctly: ship-safety refused the partial worktree and the 2-strike guard escalated to a human, exactly as designed. The two real gaps are both tracked:

Scope — the ticket was mis-scoped for a generated site (claude tried to build it and OOM'd at 4.7 GB). Re-scoped build-free as vps-showdown#46 (supersedes vps-showdown#41).
Visibility — this ticket's own 'what fixed looks like' (capture claude's output on nonzero exit so the crash reason is visible) is igor#326.
Closing in favor of those two; no separate harness fix here.

Closing (Igor / CoS) — this is a **scope + visibility** matter, not a harness bug. The harness behaved correctly: ship-safety refused the partial worktree and the 2-strike guard escalated to a human, exactly as designed. The two real gaps are both tracked: - **Scope** — the ticket was mis-scoped for a *generated* site (claude tried to build it and OOM'd at 4.7 GB). Re-scoped build-free as **vps-showdown#46** (supersedes vps-showdown#41). - **Visibility** — this ticket's own 'what fixed looks like' (capture claude's output on nonzero exit so the crash reason is visible) is **igor#326**. Closing in favor of those two; no separate harness fix here.

igor closed this issue

2026-07-02 21:19:05 +00:00