[safe-output-health] Safe Output Health Monitor - 2026-09-05

[github-actions[bot]]

Safe Output Health Monitor — 2026-09-05

Window: ~2026-09-04T04:39Z → 2026-09-05T04:39Z (~24h) · Repo: github/gh-aw

Summary

275 workflow runs analyzed (127 distinct workflows active). The unified safe_outputs job (this repo compiles every safe-output tool type — create_discussion, create_issue, add_comment, jira_create_issue, etc. — into one job with a single "Process Safe Outputs" step, not separate per-type jobs) executed 275 times: 263 success / 5 skipped (nothing to process) / 7 failure → 97.4% success rate. This breaks a 1-day clean streak (2026-09-04 was the first 0-failure audit day since tracking began 2026-08-22).

No raw GitHub Actions console logs were available in this audit's downloaded bundle (only structured job/step JSON) — root-cause confidence is noted per cluster below, and is lower than several prior audits that successfully retrieved raw safe-output-errors.json via the logs MCP tool's artifacts:["all"] option. That retrieval is the top recommended follow-up.

Error Clusters (7 failures → 4 clusters)

#	Cluster	Runs	Workflow	Status	Confidence
1	Allowed-files decline miscategorized as failure	1	Design Decision Gate 🏗️	Recurring (6th time), already root-caused since 2026-08-28, fix unshipped	High
2	Jira/Linear issue creation failing	3	Smoke Issues	New pattern, credential hypothesis	Medium
3	Project GraphQL "Bad credentials"	1	Smoke Project	Recurring (2nd time, confirmed since 2026-08-31)	Medium-high
4	Unresolved multi-tool scheduled-run failures	2	Smoke Claude, Smoke Copilot	Could not isolate failing item from available data	Low

Cluster 1 — Design Decision Gate: allowed-files decline (highest priority — affects real PRs)

§33921923497 failed on PR #58608 (add_comment + push_to_pull_request_branch), on run_attempt 2 — it failed on the retry too. This matches an already-confirmed recurring pattern (5 prior occurrences 2026-08-25 → 2026-08-30): push_to_pull_request_branch correctly declines when a PR's changeset touches files outside the workflow's allowed-files whitelist, but the decline is reported via the same hard E099 failure path as a genuine error, instead of a "declined by policy" outcome.

Why this matters now: all 5 prior occurrences were on internal/tooling PRs. This is the first confirmed occurrence on a real production PR, and it failed on retry — meaning a legitimate PR is being blocked by a false-failure signal, not just polluting monitoring metrics.

Fix (proposed since 2026-08-28, still not shipped): reclassify policy-driven declines (protected-files, allowed-files) as a distinct non-failure outcome across all safe-output types, not just approve_workflow_run.

Cluster 2 — Smoke Issues: Jira + Linear creation, 100% failure (new)

3/3 independent workflow_dispatch runs failed identically at "Process Safe Outputs": §33921898520, §33936697638, §33944666335 (latest, on current main). All exercise jira_create_issue + linear_create_issue together.

The compiled workflow (smoke-issues.lock.yml) shows the safe_outputs job depends on secrets.JIRA_API_TOKEN, secrets.JIRA_USER_EMAIL, vars.JIRA_BASE_URL, and secrets.LINEAR_API_KEY. Given 100% reproducibility and the same shape as the already-confirmed "Bad credentials" pattern in cluster 3, a missing/expired/invalid credential is the leading hypothesis — not yet confirmed, since no raw error text was retrieved this audit.

Recommended action: verify these 4 secrets/vars are present and unexpired; then fetch raw safe-output-errors.json for one of the 3 runs above to confirm the exact failure text before treating this as resolved-by-credential-rotation.

Cluster 3 — Smoke Project: GraphQL "Bad credentials" (recurring, 2nd occurrence)

§33932216650 (Copilot engine, scheduled) failed create_project_status_update + update_project. This matches the pattern first confirmed 2026-08-31 (run 33343991132, also Copilot engine, project github/projects/24068, HTTP 401 "Bad credentials" on all 5/5 items that run). Both data points use the Copilot engine, so this audit still cannot confirm whether the cause is Copilot-engine-specific or a project-wide token problem — that question remains open pending a non-Copilot-engine data point.

Cluster 4 — Smoke Claude / Smoke Copilot: unresolved (broad multi-tool, scheduled)

• §33932452774 (Smoke Claude, claude-sonnet-4-6, scheduled): configured with 13 safe-output tool types including submit_pull_request_review. Since this is a scheduled (non-PR-triggered) run, it plausibly matches a previously-confirmed pattern where submit_pull_request_review targeting "triggering" hard-fails outside PR context (2 prior occurrences, 2026-08-29 and 2026-08-31) — but this is not confirmed, since no per-item breakdown was available to isolate the actual failing tool.
• §33933690984 (Smoke Copilot, scheduled): 14 tool types configured; gateway usage summary shows 7 items processed (dispatch_workflow, upload_artifact, create_check_run, add_comment, create_issue, set_issue_type, send_slack_message) with 0 failed_calls recorded at the gateway level and no per-item outcome field — this failure does not clearly match any previously tracked pattern.

Both need raw artifact retrieval in a follow-up audit to identify the specific failing item.

Historical Context

This is audit day 15 since tracking began 2026-08-22. Two long-running patterns dominate the history: approve_workflow_run (PR Sous Chef, 11+6 occurrences, split into a confirmed permission-scope bug and a confirmed policy-miscategorization) and the push_to_pull_request_branch/allowed-files pattern now at 6 occurrences (Design Decision Gate). Both proposed fixes ("reclassify policy declines as non-failures") remain unshipped after 1+ week. Today adds one new pattern (Smoke Issues/Jira/Linear) and one recurrence (Smoke Project) to the tracked set. Full detail: /tmp/gh-aw/cache-memory/safe-output-health/ (index.json, 2026-09-05.json, error-patterns.json, recurring-failures.json, solutions.json).

Recommendations

1. Ship the policy-decline reclassification fix (proposed 2026-08-28, 6 occurrences across Design Decision Gate alone, now confirmed on a real production PR) — highest priority, longest-standing open item.
2. Verify Jira/Linear credentials for the Smoke Issues workflow (JIRA_API_TOKEN, JIRA_USER_EMAIL, JIRA_BASE_URL, LINEAR_API_KEY) and re-run with raw artifact retrieval to confirm root cause.
3. Retrieve raw safe-output-errors.json (agenticworkflows logs ... artifacts:["all"]) for at least the 4 lower-confidence runs above in the next audit — this audit relied only on pre-cached structured summaries and could not confirm exact error text for any of the 7 failures.
4. Smoke Project Bad-credentials: get one non-Copilot-engine data point to resolve whether the cause is engine-specific.

Next Steps

Next audit should prioritize raw log retrieval over structured-summary triage to restore higher-confidence root-causing, consistent with the methodology used successfully in the 2026-08-26 through 2026-09-02 audits.

References:

• §33921923497 — Design Decision Gate, PR Preserve explicit HTTP schemes in AWF API proxy targets #58608
• §33921898520 — Smoke Issues, Jira/Linear failure
• §33932216650 — Smoke Project, Bad credentials

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 · 525.1 AIC · ⌖ 6.72 AIC · ⊞ 10.8K · ◷

• expires on Sep 5, 2026, 8:57 PM UTC-08:00

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-09-06T04:57:44.021Z.

Closed by Workflow