[audit-workflows] Daily Agentic Workflow Audit — 2026-11-10

[github-actions[bot]]

Summary

The available log sample contains 200 runs from 2026-11-10 14:32:28 to 21:11:33 UTC. Of 197 completed runs, 161 succeeded and 36 failed: 81.7% raw success (81.7% excluding named intentional-failure workflows in this sample). This covers only part of the requested 24-hour window. The expanded continuation timed out after 300 seconds, so the 24-hour and 30-day rollups remain incomplete.

Collection and health

Measure	Sample result
Runs	200 (197 completed, 3 in progress)
Completed runs	161 success, 36 failure
Success rate	81.7% raw; 81.7% adjusted in sample
Engine counts	Copilot 103, Pi 56, Codex 31, Claude 8, Aider 2
Errors / missing tools / warnings	40 / 0 / 0
Completed-run token usage	28,929,227

The log response did not include the intentional_failure marker, and neither named guardrail test workflow appeared in the sample. The adjusted rate therefore equals the raw rate for these sampled runs; no full-window adjusted rate is available.

Findings

1. Log coverage is incomplete and recurring. The 100-run pages covered only 2026-11-10 14:32–21:11 UTC, leaving most of the requested 24-hour interval uncollected. A 1,000-run continuation for the older interval ended with context deadline exceeded at 300 seconds. The previous audit also recorded partial 24-hour and 30-day coverage and a continuation timeout.
2. Auditor evidence collection failed in 6 runs. Each failure was at collect_failure_evidence / Collect health failures, cached logs and bounded audits. The prior audit recorded 7 failures in this collector step, making this a recurring hotspot.
3. The shared conclusion handler failed in 27 distinct runs. The Handle agent failure step had 15 driver-exit failures and 12 agent-logic failures across the sample.
4. Firewall allowed-request rollups disagree. The global count is 11,870 allowed requests while summed workflow counts equal 4,428; blocked counts reconcile at 4 in each view. The prior audit reported the same mismatch.
5. Two additional signals need triage. Daily Cache Strategy Analyzer used 11,592,803 tokens in one run. One Visual Regression Checker run failed at Process Safe Outputs; the previous audit also recorded safe-output processing failures.

Charts

This single partial-day point shows 161 successes and 36 failures among completed runs, for an 81.7% sample success rate. Missing hours and days prevent interpreting it as a 24-hour or 30-day trend.

The captured completed runs total 28.9 million tokens, with one 11.59 million-token run from Daily Cache Strategy Analyzer. Only one date has data, so a seven-day moving average cannot be calculated and the requested 30-day trend is unavailable.

Recommended actions

• Fix log collection pagination and deadlines, resume cursors to full coverage, and mark rates and charts incomplete until the requested window is complete.
• Triage the recurring auditor evidence-collection failure and shared conclusion-handler failures; retain the originating failure details.
• Reconcile firewall telemetry rollups before changing allowlists.
• Review the Daily Cache Strategy Analyzer token outlier and the Visual Regression Checker safe-output processing failure.

References:

• §42895572114 — Auditor evidence collection failure
• §42874860033 — token outlier
• §42886362801 — safe-output processing failure

Generated by 🔍 Agentic Workflow Audit Agent · codex · gpt60 · ~9 AIC · ⌖ ~1 AIC · ⊞ 16.8K · ◷

• expires on Nov 11, 2026, 1:38 PM UTC-08:00