[audit-workflows] Agentic Workflow Audit — 2026-10-16 (partial coverage)

[github-actions[bot]]

Summary

The captured run sample has a 64.3% success rate among completed runs and two recurring failure signals: Redact secrets in logs and Handle agent failure. The logs MCP continuation timed out, so this is a partial audit covering 2026-10-16 13:41:10–21:17:05 UTC (283 runs), not a complete 24-hour or 30-day rollup.

Collection Phase

The sample includes 180 successes, 100 failures, and 3 runs still in progress. That is 64.3% raw success (180/280 completed) and 64.3% excluding intentional-failure workflows; neither named guardrail test appeared in the collected runs. The payload did not include the intentional_failure field, so the adjustment was checked by workflow name.

The log summaries report 14.44 million tokens across the 283 collected runs, with 0 missing tools and 0 missing-data events in those runs. Classified engines were Copilot 131, Pi 79, Codex 41, Claude 6, and Aider 2; 24 runs had no engine ID. The highest token run was Daily Go Test Parallelizer at 1,079,729 tokens in 12.4 minutes; a second run used 663,666 tokens in 11.9 minutes.

The sample shows 180 successful and 100 failed runs among completed records, with a 64.3% success rate; 3 in-progress runs are excluded. It covers only a 7.6-hour interval, so it cannot establish a daily or 30-day health trend.

The collected interval contains 14.44 million tokens, including two runs above 660,000 tokens. A 7-day moving average is unavailable because the logs did not provide seven complete days.

Clustering Phase

The agent / Redact secrets in logs step failed 53 times across 21 workflows in the returned pages. Handle agent failure appeared 16 times as a downstream failure signal; treat it as aftermath until the preceding error is identified. PR Code Quality Reviewer had 10 failures in 22 runs (45%) in the collected interval, including repeated Execute GitHub Copilot CLI failures.

Issue matching is unresolved: GitHub semantic search returned candidates, but the integrity filter withheld all five results. I cannot map these signals to a known issue from the available evidence.

Recommendation Phase

1. Compare representative Redact secrets in logs failures across engines and workflows, identify the shared failing condition, and then fix the common step or configuration if the evidence confirms one cause.
2. Prioritize PR Code Quality Reviewer: inspect the Copilot CLI failure path and separate it from the downstream conclusion-handler errors.
3. Review context size and output handling in Daily Go Test Parallelizer; its two sampled runs used 1.08M and 664K tokens.
4. Fetch the remaining 24-hour continuation and complete 30-day log history before publishing fleet trends or a 7-day moving average.

Synthesis Phase

The strongest current action is to investigate shared agent-step failures and the PR Code Quality Reviewer hotspot. These are provisional findings from 283 runs; the missing log coverage could change the fleet rate, engine mix, and recurrence counts.

References:

• §38944763413 — Release, Redact secrets in logs failure
• §38926451368 — PR Code Quality Reviewer failure cluster
• §38916921681 — Daily Go Test Parallelizer, 1.08M tokens

Generated by 🔍 Agentic Workflow Audit Agent · codex · gpt60 · 4.7 AIC · ⌖ 0.462 AIC · ⊞ 16K · ◷

• expires on Oct 17, 2026, 1:39 PM UTC-08:00