[audit-workflows] Agentic workflow audit: 2026-10-20

[github-actions[bot]]

Summary

The log request did not reach the requested 30-day window. Even with count: 1000, the collector returned 122 runs from 2026-10-20 17:20:40Z to 21:19:52Z. This partial sample shows 80.3% success and recurring checkout and agent-failure signatures; it is not a 30-day trend.

Key metrics

Metric	Collected sample
Runs	122 (all executed)
Success	98/122; 80.3% raw and 80.3% excluding the named intentional-failure workflows
Failures	24
Errors	44 (20 driver-exit, 4 agent-logic failures)
Missing tools / warnings	0 / 0
Tokens	15,212,032
Engines	Copilot 60, Pi 36, Codex 19, Claude 6, Aider 1

No Daily Credit Limit Test or Daily Max AI Credits Test run appears in the sample. The run records do not include the intentional_failure field, so the adjusted rate is based on the required workflow-name exclusions.

Findings and recommendations

View full findings

• Checkout PR branch failures recur. The failure cluster recorded 15 occurrences across five workflows: design-decision-gate, mattpocock-skills-reviewer, ponytail-reviewer, pr-code-quality-reviewer, and test-quality-sentinel. The prior audit recorded 16 occurrences for the same signature. Inspect the shared branch checkout setup and compare representative runs across those workflows.
• The conclusion failure handler remains a broad hotspot. Handle agent failure appeared in 20 driver-exit and 3 agent-logic cluster occurrences across multiple workflows; the prior audit recorded 32 occurrences. Trace the original agent failure alongside the handler step to separate upstream causes from the shared conclusion symptom.
• Token and runtime outliers increased in the sample. Daily Cache Strategy Analyzer used 5,301,389 tokens over 28 action minutes. Daily Go Test Parallelizer used 2,553,700 tokens over 26 action minutes. Review their prompt and output volume and their configured token/runtime limits.
• No missing tools were observed.

Trend charts

The chart shows the 122-run Oct 20 sample only: 98 successes, 24 failures, and an 80.3% success rate. The remaining dates are blank because the collector returned no data for them.

The 15.2M tokens represent the collected four-hour slice, not the whole day. A seven-day moving average cannot be calculated from one day of coverage.

Audit history

The prior replay record is run 39326530941; current run 39499563583 was absent. The current audit could not be appended because ledger_append was not included in the configured safe-output tools.

References: §39484474313 · §39499464895 · §39476158337

Generated by 🔍 Agentic Workflow Audit Agent · codex · gpt60 · 4.53 AIC · ⌖ 0.361 AIC · ⊞ 16K · ◷

• expires on Oct 21, 2026, 1:38 PM UTC-08:00