[audit-workflows] Agentic Workflow Daily Audit — 2026-11-09

[github-actions[bot]]

Summary

The requested 24-hour and 30-day log windows could not be collected completely. The available sample covers 7 hours 27 minutes on Nov 9: 187 executions, 157 successes, and 30 failures (83.96% raw success rate). The main reliability issue is the Auditor evidence-collection step, which failed in all 7 observed Auditor runs.

Key metrics

• 203 run records reviewed; 187 executed.
• 25 errors, 7 driver exit failures, 8 agent logic failures, 0 missing tools, 0 missing data, and 0 warnings.
• Engine counts from summary.engine_counts: Copilot 103, pi 47, Codex 24, Claude 7, Aider 1. Another 21 records had no engine ID.
• Neither named intentional-failure workflow appeared in this sample, so the name-based adjusted rate is also 83.96%. The log output did not include intentional_failure, so the marker could not be verified directly.

Findings

View full findings

1. Log coverage is incomplete and recurring. The 30-day request stopped after a narrow 7-hour slice. The 24-hour request returned 203 records spanning 7h27 and a continuation cursor; the continuation timed out after 300 seconds. The 30-day health request also timed out after 120 seconds. Previous audit records show the same pagination gap.
2. Auditor evidence collection failed 7/7 times. The repeated failed step was Collect health failures, cached logs and bounded audits. This continues the prior Auditor failure hotspot.
3. Safe Outputs processing failed 3 times across Design Decision Gate and PR Code Quality Reviewer at Process Safe Outputs; this signature also appears in recent audit history.
4. Failure aftermath handling failed 10 times across 8 workflow names at Handle agent failure, repeating the earlier conclusion-aftermath cluster.
5. Token use is concentrated in Daily Go Test Parallelizer. Four observed runs used 893,921–1,054,642 tokens each and lasted 8.5–9.5 minutes. The high-token-outlier finding recurs from earlier audits.
6. Firewall rollups disagree. The global count is 8 blocked requests and the observability insight assigns 8/91 to Visual Regression Checker, while the per-workflow table reports 4/47 for that workflow. Reconcile this before changing any firewall rules.

Next actions

• Repair log pagination and query budgeting so the 24-hour and 30-day windows finish and resume from their cursors.
• Investigate the Auditor evidence-collection step and the shared Process Safe Outputs and Handle agent failure failures.
• Review the repeated Daily Go Test Parallelizer token use against expected workload behavior.
• Reconcile firewall counts and restore engine ID coverage before using those rollups for policy or engine comparisons.

Trends

The captured 7h27 contains 157 successful and 30 failed executions, for an 83.96% success rate. Only Nov 9 has data; blank dates are unavailable, not zero, so this is not a complete 30-day trend.

The captured slice contains 14,466,798 tokens, including four Daily Go Test Parallelizer runs near 0.89–1.05 million tokens. One observed date is insufficient for a 7-day moving average or a 30-day comparison.

The audit-history upsert was accepted, but persistence is unconfirmed because the configured safe-output interface does not expose push_ledger_changes. The chart files were saved under /tmp/gh-aw/agent because /tmp/gh-aw/python/charts is read-only.

References: §42693198722 · §42637893126 · §42693186092

Generated by 🔍 Agentic Workflow Audit Agent · codex · gpt60 · ~7 AIC · ⌖ ~1 AIC · ⊞ 16.9K · ◷

• expires on Nov 10, 2026, 1:42 PM UTC-08:00

[github-actions[bot]]

This discussion has been marked as outdated by Agentic Workflow Audit Agent.

A newer discussion is available at Discussion #76616.