[audit-workflows] Daily agentic workflow audit — 2026-10-15

[github-actions[bot]]

Summary

The audit collected 572 distinct runs in the 24-hour window ending 2026-10-15 21:24 UTC. Of 568 completed runs, 383 succeeded and 185 failed: 67.4% raw (67.4% excluding the two intentional-failure workflow families; 381/565 completed runs). The most frequent cross-workflow failure boundary was Start MCP Gateway, observed in 90 failed runs.

Key metrics and issues

• Collection: 572 runs, 4 still in progress; 63.9M tokens and 9,664 action minutes.
• Adjusted success rate: 67.4%; intentional-failure workflows Daily Credit Limit Test and Daily Max AI Credits Test were excluded from success-rate rollups.
• Missing tools / missing data / warnings: 0 / 0 / 0 in the collected log pages.
• Engine usage: Copilot 199, Pi 182, Codex 84, Claude 60, Aider 3; 32 runs had no engine classification, with 12 runs across other engines.
• Failure concentration: AI Moderator had 14 failed runs and PR Sous Chef had 10. The collected failure evidence for AI Moderator is metadata-only.
• Token outliers: Package Specification Enforcer used 5.39M tokens in one successful run; Metrics Collector used 2.80M and the arXiv researcher used 2.77M.

Trend charts

Workflow health: The chart has two observed partial calendar days: Oct 14 shows 39 successes and 57 failures (40.6%); Oct 15 shows 342 successes and 127 failures (72.9%), both excluding intentional-failure workflows. The 30-day history could not be collected, so blank days mean missing data and no monthly trend is inferred.

Token usage: The observed partial days contain 6.84M tokens on Oct 14 and 57.11M on Oct 15. A seven-day moving average is unavailable because fewer than seven daily observations were collected.

Detailed findings

View full findings

• Shared startup failure: failure-5ea31635c811b70b recorded 90 driver_exit failures at Start MCP Gateway across many workflow names, including smoke workflows and scheduled jobs. This is an observed failure boundary; the collected evidence does not establish the underlying cause.
• Downstream failure signal: Handle agent failure recurred as an aftermath signal in the log clusters. Trace these runs to their earlier failed step before changing the shared conclusion handler.
• Metadata-only failures: AI Moderator had 14 failed runs, but the collected cluster evidence identifies workflow-level failures without a failing step. Representative run: §38714923122.
• Workflow hotspot: One 300-run slice reported PR Sous Chef at 8 failures across 38 runs (21%); the complete 24-hour window contains 10 PR Sous Chef failures.
• Token outlier: Package Specification Enforcer run §38727680973 used 5.39M tokens and succeeded.
• Log collection limit: A 1,000-run request timed out at the MCP five-minute deadline. Smaller resumable pages completed the 24-hour collection, but only two partial calendar-day buckets were available for charting the 30-day window.

Recommended actions

1. Inspect representative Start MCP Gateway driver failures across different workflows to identify a shared startup cause before changing gateway configuration. Example: §38661395685
2. Gather step-level evidence for AI Moderator failures and compare PR Sous Chef failures with its normal run population.
3. Review input selection and token budgets for Package Specification Enforcer, Metrics Collector, and the arXiv researcher.
4. Improve log export or pagination capacity so the daily audit can produce a complete 30-day trend view.

References: Gateway startup failure · AI Moderator · Token outlier

Generated by 🔍 Agentic Workflow Audit Agent · codex · gpt60 · 7.75 AIC · ⌖ 0.35 AIC · ⊞ 15.6K · ◷

• expires on Oct 16, 2026, 1:45 PM UTC-08:00