[safe-output-health] 🏥 Safe Output Health Report - 2026-06-24 #41169
Closed
Replies: 1 comment
-
|
This discussion has been marked as outdated by Safe Output Health Monitor. A newer discussion is available at Discussion #41386. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
Executive Summary
Headline: The production
safe_outputsstreak holds — zero production safe-output job failures today. All 4 hard failures are smoke-only (production_impact=false), agent=SUCCESS in every case, and all on a single PR branchcopilot/bump-firewall-v0279-and-mcpg-v0329. This is the 4th reproduction window for the long-tracked smoke target/context-resolution family (prior: 06-11, 06-14, 06-15). Yesterday's production Changeset-bundle offender did not recur (workflow absent).safe_outputs Job Statistics
Error Cluster: Smoke target/context-resolution hard-fail family (smoke-only) — REPRODUCED
copilot/bump-firewall-v0279-and-mcpg-v0329safe_outputsjobadd_labels/remove_labelsno-item-number hard-fail (no triggering context)add_comment→discussionResource not accessible by integration/target:*review-commentRoot-cause analysis (inferred — message logs not recoverable today)
The smoke workflows deliberately exercise every safe-output handler in one run. Two known, still-open defects in this family explain the failures:
Soft-skip vs hard-fail inconsistency (
target_star_add_labels_no_item_number_hardfail). Onworkflow_dispatchthere is no triggering PR/issue.create_pull_request_review_comment/comment_memorysoft-skip (⏭ Not in pull request context), butadd_labels/remove_labels/add_commenthard-fail (✗ No issue/PR number available), failing the whole job. The two Smoke Copilotworkflow_dispatchruns match the 06-11 (run-27320477001) and 06-14 (run-27481382799) signature exactly. Family window count now 4 (06-11, 06-14, 06-24×2).Discussions write-scope gap (
add_comment_discussion_resource_not_accessible_by_integration). The smokesafe_outputsjob token lacksdiscussions:write, soadd_commentfalling through to the discussion-comment path hard-failsResource not accessible by integration. Smoke Claude (pull_request) is the exact 06-14/06-15 offender — it was JOB-clean on 06-23 and regressed today; Smoke Codex (also clean 06-23) failed similarly.Why inference, not certainty: Process Safe Outputs step logs were not pre-bundled, and the
auditMCP tool timed out (120s). Attribution comes from the smoke prompt + the documented 06-11/06-14/06-15 family history, consistent with prior-audit methodology.Recommendations
Bug fixes (in scope, still OPEN):
add_labels/remove_labels/add_commentand the review-comment/comment_memory handlers — soft-skip (⏭) with a clear message when no issue/PR number is resolvable, instead of##[error]that fails the job. This is the cleanest, now 4-times-reproduced demonstration of the inconsistency. Severity: low (smoke-only) but cheap and high-signal.safe_outputsjob tokendiscussions:write(or makeadd_comment→discussion soft-skip when the scope is absent). Reproduced across 06-14/06-15 and again 06-24.pr_review_buffer.cjs:554Path-variant predicate fix — UNVALIDATED for the 27th consecutive audit (no PR-reviewers exercised the path today).Process / observability:
auditMCP tool timeout) so per-message failure strings are recoverable. Today's 120s timeout + absent step logs blocked exact root-causing of all 4 failures — a recurring limitation across audits.Work Item Plan
WI-1: Unify soft-skip vs hard-fail for missing trigger context
add_labels/remove_labels/add_commentsoft-skip (not hard-fail) when no resolvable issue/PR target exists; a smokeworkflow_dispatchrun no longer fails itssafe_outputsjob for this reason.WI-2: discussions:write for smoke safe_outputs token
add_comment→discussion no longer 403s; or it soft-skips cleanly.WI-3: Recover per-message safe-output logs in the monitor
auditMCP tool (which timed out today).Historical Context & Trends
workflow_dispatch(add_labels-no-context) andpull_request(add_comment→discussion / target:*) variants in one window.Next Steps
patch-format:bundleproduction pathpr_review_buffer.cjs:554Path-variant (UNVALIDATED 27 audits)References:
Beta Was this translation helpful? Give feedback.
All reactions