[safe-output-health] Safe Output Health Monitor - 2026-09-10: 98.1% success (206/210), 1 new root-caused update_pull_request failure

[github-actions[bot]]

Safe Output Health Monitor — 2026-09-10

Window: ~24h ending 2026-09-10 (audit run §34437063118)

Quiet, stable day. Of 210 workflow runs analyzed, the safe_outputs job executed in all 210, with 206 success / 3 skipped / 1 failure — a 98.1% success rate, consistent with the ~97–100% historical daily baseline.

Key finding: new root-caused failure pattern

The single failure this period is a brand-new signature for PR Sous Chef, distinct from both previously-tracked approve_workflow_run issues (fork-PR permission gap, protected-files decline):

• Run: §34416984733 (PR Sous Chef, schedule trigger)
• Safe output type: update_pull_request
• Target: PR Fix invalid empty GitHub server guard policy for enclave-only GitHub tools #59816
• Root cause: the update_pull_request update-branch call (PUT /repos/github/gh-aw/pulls/59816/update-branch) returned HTTP 422: expected head sha didn't match current head ref. The handler treated this as fatal (did not match any known benign update-branch condition) instead of re-fetching the current head SHA and retrying. This looks like a race between the agent reading PR Fix invalid empty GitHub server guard policy for enclave-only GitHub tools #59816's state and the safe_outputs job acting on it later (e.g. a new commit landed on the branch in between).
• Isolation held, job still failed: the other 2 items in the same batch (add_comment, create_issue) succeeded independently (Status: partial_success, Successful: 2, Failed: 1), yet the safe_outputs job's overall conclusion was still hard failure. This is the same partial-success-still-counts-as-failure shape flagged in prior audits for other declined/soft-fail conditions.
• Confidence: high — confirmed directly from the raw workflow-logs/3_safe_outputs.txt console log (not inferred from grader metrics).

No other actively-tracked recurring signature (Design Decision Gate allowed-files decline, hardcoded smoke-discussion temp ID, Smoke Project/Smoke Issues credential failures, submit_pull_request_review non-PR-context, concurrent same-PR write race) occurred today.

Methodology and full stats

• Scanned all 210 locally cached run_summary.json files; every run has exactly one safe_outputs job, classified by its top-level conclusion.
• Cross-checked the failing run against audit.json, then retrieved the raw Process Safe Outputs console log via the agenticworkflows logs tool (before_run_id/after_run_id bracketing + workflow_name pr-sous-chef + artifacts all) to get the exact error text rather than relying on grader-shape inference.
• Totals reconcile exactly: 206 + 3 + 1 = 210.

Trend vs. prior audit history

This is a new pattern — not a recurrence of either of PR Sous Chef's two known approve_workflow_run failure modes. The two known modes remain unresolved in the backlog as of the last audit (2026-09-06):

• grant-actions-write-fork-pr-scope (fork-PR token permission gap) — still proposed, not shipped.
• reclassify-protected-file-decline-as-non-failure (protected-files policy decline miscounted as a hard failure) — still proposed, not shipped.

Today's new pattern has only 1 data point so far; it needs 1–2 more occurrences before a specific fix should be committed to.

Recommendations

1. **New, needs more (redacted) track the update_pull_request stale-head-SHA condition (422 expected head sha didn't match current head ref). If it recurs, fix by re-fetching the PR's current head SHA immediately before the update-branch call and retrying once, rather than treating it as immediately fatal.
2. Still unresolved from prior audits (unrelated to today's finding, no new action needed today): the fork-PR permission gap and protected-files-decline miscategorization for approve_workflow_run remain open in the backlog — no new occurrences today, but still unshipped.
3. Process: continue generalizing the partial-success-still-hard-fails-the-job behavior — today's run had 2/3 items succeed yet the job still reported failure, the same shape already flagged for protected-file declines and non-PR-context submit_pull_request_review skips.

No action required beyond backlog tracking; overall safe-output health remains good (98.1%).

References:

• §34416984733 — PR Sous Chef, update_pull_request 422 failure (today's only failure)
• §34437063118 — this audit's run

Warning

Firewall blocked 1 domain

The following domain was blocked by the firewall during workflow execution:

• api.anthropic.com

To allow these domains, add them to the network.allowed list in your workflow frontmatter:

network:
  allowed:
    - defaults
    - "api.anthropic.com"

See Network Configuration for more information.

Generated by 🔒 Safe Output Health Monitor · claude · agent · 494.3 AIC · ⌖ 9.39 AIC · ⊞ 10.9K · ◷

• expires on Sep 10, 2026, 9:15 PM UTC-08:00

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-09-11T05:15:40.685Z.

Closed by Workflow