[audit-workflows] Agentic Workflow Audit — 2026-10-18 (partial log coverage)

[github-actions[bot]]

Summary

The workflow status check succeeded. The returned 24-hour logs were partial: 100 runs covered only 2026-10-18 16:49–21:20 UTC, so this is a bounded sample rather than a full-day fleet audit. In that sample, 73 runs succeeded, 25 failed, and 2 had no conclusion. That is 73% raw success (73/100) and 74.5% among runs with success or failure conclusions (73/98). The log objects did not contain the intentional_failure field, so the requested intentional-failure-adjusted fleet rate cannot be calculated.

The 30-day request for 1,000 runs exceeded the service’s 300-second deadline. The only 30-day response available was another 100-run page covering about five hours. I did not generate or upload trend charts because those samples cannot support daily 30-day health or token trends.

Collection findings

• The 100 returned runs report 0 missing tools, 8 errors, 0 warnings, 2 driver exit failures, and 5 agent logic failures.
• Engine counts returned by the log service: copilot 36, pi 34, codex 7, claude 4, aider 1. The per-run engine field was absent on 18 records.
• A recurring observed failure boundary was “Assert model-routing evidence” in the Copilot routing smoke workflow: 3 failures, including run 39147289566, run 39147795806, and run 39148386697. A separate routing and steering evidence assertion failed once.
• AI Moderator had 4 failed runs, but its cluster is marked metadata-only; the fetched summary does not establish a root cause. Other repeated metadata-only workflow failures included Matt Pocock Skills Reviewer, Ponytail Reviewer, PR Code Quality Reviewer, and Test Quality Sentinel (3 each).
• One Linter Miner failure ended at “Execute GitHub Copilot CLI”; one Daily Code Metrics failure ended at “Check daily workflow token guardrail.” These are observed boundaries, not confirmed causes.
• The largest token outlier in this sample was Daily Cache Strategy Analyzer, at 5.75M tokens and 27.4 minutes (31.4% of the 18.29M sampled tokens).

Recommendations

1. Inspect the model-routing smoke assertion and the evidence it receives. The same assertion failed three times within about 15 minutes, with a later passing routing run also present in the sample.
2. Retrieve step-level logs for the AI Moderator and other repeated metadata-only failures before assigning causes; the current cluster records identify affected workflows but not failing steps.
3. Review the Daily Cache Strategy Analyzer run’s token trace. Its 5.75M tokens account for nearly one-third of the sampled total; compare against a complete window before treating this as a sustained regression.
4. Re-run the fleet rollup with complete log pagination and intentional_failure markers, then generate the requested health and token charts from the complete 30-day data.

History and coverage limits

I could not query the audit-history SQLite projection because sqlite3 is not installed. Therefore, prior stable finding IDs, the current run’s duplicate status, and ledger diagnostics could not be checked, and I did not submit a ledger upsert without that required duplicate check. The recurring clusters above are identified within the returned sample only; they could not be compared with prior audits. The service also timed out on the higher-count 30-day request, preventing a valid 30-day chart and rollup.

Generated by 🔍 Agentic Workflow Audit Agent · codex · gpt60 · 2.35 AIC · ⌖ 0.352 AIC · ⊞ 16K · ◷

• expires on Oct 19, 2026, 1:35 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 #70378.