You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Bug 1 β Prompt/guard contradiction (3/4 failures) issue-triage.md instructed Claude to return {"skip": true} for issues with 2+ labels. However, aw.sh already enforces this as a pre-Claude label-count guard before calling the model β and then explicitly rejects any skip flag returned by the model as "possible prompt injection", failing the run. Internal issues with the initiative label (1 label, below the threshold) still triggered Claude to return skip, likely because the prompt's skip instruction created ambiguity in the model's decision.
Bug 2 β Non-existent label (1/4 failures) issue-triage.md included needs-triage in the allowed label set and instructed Claude to use it. The label needs-triage does not exist in the repo; gh issue edit --add-label fails with 'needs-triage' not found. Also fixed a latent version of the same bug: good-first-issue (hyphenated) was listed, but the repo label is good first issue (with space).
Plan Execution
Post implementation plan as issue comment
Write 5 failing regression tests (TDD, confirmed red before implementation)
Remove skip instruction from issue-triage.md prompt
Replace needs-triage β needs-human-review in allowed list and prompt
Replace good-first-issue β good first issue in allowed list and prompt
Update test 6 and scenarios.md to use corrected label names
Run full test suite β all pass
Test Results
PASS scenario-4: 2 existing labels β skip=true
PASS safe-output: 4 labels exceeds max-labels:3 β rejected
PASS safe-output: disallowed label β rejected
PASS safe-output: skip flag β no writes
PASS safe-output: empty comment β rejected
PASS safe-output: valid allowed labels β validation passed
PASS issue-triage.md: allowed list does not contain 'needs-triage'
PASS issue-triage.md: allowed list contains 'needs-human-review'
PASS issue-triage.md: allowed list does not contain 'good-first-issue'
PASS issue-triage.md: prompt does not contain skip instruction
PASS safe-output: 'needs-human-review' is a valid allowed label
Results: 11 passed, 0 failed
(runner yml: 7 passed, 0 failed β dev-lead-lint: all checks passed)
Files Changed
.github/workflows/issue-triage.md: removed skip instruction; replaced needs-triage β needs-human-review and good-first-issue β good first issue throughout
tests/aw/issue-triage/test_aw_run.sh: updated test 6 to use needs-human-review; added tests 7β11 as regression guards
tests/aw/issue-triage/scenarios.md: updated label names to match
Notes
The skip instruction removal is safe: aw.sh's pre-Claude guard (skip-if-labels-ge: 2 default) handles skip for over-labelled issues before Claude is ever called. The apply job is already gated by if: needs.triage.outputs.skip != 'true', so the skip flow continues to work correctly.
Activity
don-petry commentedon Jun 9, 2026
don-petry commentedon Jun 9, 2026
Dev-Lead Implementation Plan
Issue: #316 β [Fleet Monitor] petry-projects/.github-private β issue-triage-runner.yml
Root Cause Analysis
Two distinct failure modes were found across 4/12 failed runs:
Bug 1 β Prompt/guard contradiction (3 failures: runs 27168683186, 27168684938, 27168687742)
issue-triage.mdinstructs Claude to return{"skip": true}when an issue has 2+ labels.aw.shalready handles this as a pre-Claude label-count guard before calling the model.aw.shthen rejects any{"skip": true}from the model as "possible prompt injection" and exits 1.Bug 2 β Non-existent label (1 failure: run 27172272227)
issue-triage.mdallows and instructs Claude to use theneeds-triagelabel.needs-triagedoes not exist in the repo (repo hasneeds-human-review).gh issue edit --add-label "needs-triage"fails with'needs-triage' not found.Scope
Fix
.github/workflows/issue-triage.mdto align the prompt withaw.sh's runtime contract:aw.shalready handles 2+ labels).needs-triageβneeds-human-review(the label that actually exists).good-first-issueβgood first issue(latent label-name mismatch, same failure mode as bug 2).Implementation Checklist
tests/aw/issue-triage/test_aw_run.shβ add regression tests (TDD, write first).github/workflows/issue-triage.mdβ remove skip instruction; fix label namesTests to Write
issue-triage.mdallowed list does NOT containneeds-triageissue-triage.mdallowed list DOES containneeds-human-reviewissue-triage.mdallowed list does NOT containgood-first-issueissue-triage.mdprompt body does NOT contain the skip-for-labels instructionsafe-output applyacceptsneeds-human-reviewas a valid labelTest Command
bash tests/aw/issue-triage/test_aw_run.sh && bash tests/aw/issue-triage/test_runner_yml.shdon-petry commentedon Jun 9, 2026
Dev-Lead: Implementation Complete
Root Causes Found
Bug 1 β Prompt/guard contradiction (3/4 failures)
issue-triage.mdinstructed Claude to return{"skip": true}for issues with 2+ labels. However,aw.shalready enforces this as a pre-Claude label-count guard before calling the model β and then explicitly rejects any skip flag returned by the model as "possible prompt injection", failing the run. Internal issues with theinitiativelabel (1 label, below the threshold) still triggered Claude to return skip, likely because the prompt's skip instruction created ambiguity in the model's decision.Bug 2 β Non-existent label (1/4 failures)
issue-triage.mdincludedneeds-triagein the allowed label set and instructed Claude to use it. The labelneeds-triagedoes not exist in the repo;gh issue edit --add-labelfails with'needs-triage' not found. Also fixed a latent version of the same bug:good-first-issue(hyphenated) was listed, but the repo label isgood first issue(with space).Plan Execution
issue-triage.mdpromptneeds-triageβneeds-human-reviewin allowed list and promptgood-first-issueβgood first issuein allowed list and promptscenarios.mdto use corrected label namesTest Results
Files Changed
.github/workflows/issue-triage.md: removed skip instruction; replacedneeds-triageβneeds-human-reviewandgood-first-issueβgood first issuethroughouttests/aw/issue-triage/test_aw_run.sh: updated test 6 to useneeds-human-review; added tests 7β11 as regression guardstests/aw/issue-triage/scenarios.md: updated label names to matchNotes
The skip instruction removal is safe:
aw.sh's pre-Claude guard (skip-if-labels-ge: 2default) handles skip for over-labelled issues before Claude is ever called. Theapplyjob is already gated byif: needs.triage.outputs.skip != 'true', so the skip flow continues to work correctly.feat: implement issue #316 β [Fleet Monitor] petry-projects/.github-pβ¦
feat: implement issue #316 β [Fleet Monitor] petry-projects/.github-pβ¦
feat: implement issue #316 β [Fleet Monitor] petry-projects/.github-pβ¦
feat: implement issue #316 β [Fleet Monitor] petry-projects/.github-pβ¦
feat: implement issue #316 β [Fleet Monitor] petry-projects/.github-pβ¦