PR-review rework sessions cannot run bin/agent-ask.sh -- follow-up ticket silently dropped #430

New issue

Closed

opened 2026-07-26 03:00:37 +00:00 by igor · 0 comments

igor commented

2026-07-26 03:00:37 +00:00

Collaborator

Diagnosis

During the joshtronic/igor#429 rework tick (21:38-21:49), the reviewer asked the rework session to file a follow-up issue. The session attempted it and was denied by the permission profile the harness grants to PR-review invocations, so the drafted ticket was never filed. The rework completed and pushed (claude exited 0), so nothing surfaced as an error -- the work item just evaporated into the PR body text. This is a harness capability gap: the rework prompt is allowed to be told to file follow-ups but is not given the tool to do it, so any reviewer-requested follow-up in this path is silently lost.

Evidence

Jul 25 21:46:32 igor.sherver.org tick.sh[2664650]: The reviewer asked me to file the follow-up about the test suite's silent-skip trap. Checking whether the helper has what it needs:
Jul 25 21:47:01 igor.sherver.org tick.sh[2664650]: `agent-ask.sh` is blocked by this session's permission profile, so I can't file that follow-up myself. Final verification at the committed head:
Jul 25 21:47:35 igor.sherver.org tick.sh[2664650]: - **I could not file the `bin/test-*.sh` follow-up.** `bin/agent-ask.sh` is blocked by this session's permission profile. The ticket text is drafted and I'll file it as soon as that's permitted, or you can — the substance is that those suites run under `set -uo pipefail` without `-e` and verdict on a `FAIL` counter, so a command that never ran can't increment it, and a typo'd helper is a green skip.
Jul 25 21:47:35 igor.sherver.org tick.sh[2663596]: [agent] claude exited 0 after 536s

No retry, no fallback path, and no journal line from the harness itself noting the dropped follow-up -- the only record is prose inside the PR comment.

What "fixed" looks like

Either allow bin/agent-ask.sh (or an equivalent issue-filing entry point) in the PR-review/rework permission profile, or
Stop asking rework sessions to file follow-ups and have tick.sh harvest a structured "follow-up:" block from the session output and file it after the session exits.
Either way, a rework session that reports it could not complete a requested side-effect should emit an [agent] warning line so the drop is greppable in the journal instead of buried in PR prose.
The specific dropped item (bin/test-*.sh suites run without set -e and verdict on a FAIL counter, so a typo'd assertion is a green skip) should still be filed.

service: agent.service
severity: medium
window: 2026-07-25 21:00-22:00 (filed by the hourly logwatch pass)

## Diagnosis During the `joshtronic/igor#429` rework tick (21:38-21:49), the reviewer asked the rework session to file a follow-up issue. The session attempted it and was denied by the permission profile the harness grants to PR-review invocations, so the drafted ticket was never filed. The rework completed and pushed (`claude exited 0`), so nothing surfaced as an error -- the work item just evaporated into the PR body text. This is a harness capability gap: the rework prompt is allowed to be *told* to file follow-ups but is not given the tool to do it, so any reviewer-requested follow-up in this path is silently lost. ## Evidence ``` Jul 25 21:46:32 igor.sherver.org tick.sh[2664650]: The reviewer asked me to file the follow-up about the test suite's silent-skip trap. Checking whether the helper has what it needs: Jul 25 21:47:01 igor.sherver.org tick.sh[2664650]: `agent-ask.sh` is blocked by this session's permission profile, so I can't file that follow-up myself. Final verification at the committed head: Jul 25 21:47:35 igor.sherver.org tick.sh[2664650]: - **I could not file the `bin/test-*.sh` follow-up.** `bin/agent-ask.sh` is blocked by this session's permission profile. The ticket text is drafted and I'll file it as soon as that's permitted, or you can — the substance is that those suites run under `set -uo pipefail` without `-e` and verdict on a `FAIL` counter, so a command that never ran can't increment it, and a typo'd helper is a green skip. Jul 25 21:47:35 igor.sherver.org tick.sh[2663596]: [agent] claude exited 0 after 536s ``` No retry, no fallback path, and no journal line from the harness itself noting the dropped follow-up -- the only record is prose inside the PR comment. ## What "fixed" looks like - Either allow `bin/agent-ask.sh` (or an equivalent issue-filing entry point) in the PR-review/rework permission profile, or - Stop asking rework sessions to file follow-ups and have `tick.sh` harvest a structured "follow-up:" block from the session output and file it after the session exits. - Either way, a rework session that reports it could not complete a requested side-effect should emit an `[agent]` warning line so the drop is greppable in the journal instead of buried in PR prose. - The specific dropped item (bin/test-*.sh suites run without `set -e` and verdict on a FAIL counter, so a typo'd assertion is a green skip) should still be filed. --- service: agent.service severity: medium window: 2026-07-25 21:00-22:00 (filed by the hourly logwatch pass)