[audit-workflows] Daily Agentic Workflow Audit — 2026-08-11 (88.8%/89.1% ex-intentional; PR-review firewall cluster identified; 36-day audit gap)

[github-actions[bot]]

Overview

Full ~24h audit window (2026-08-10T21:30Z → 2026-08-11T21:16Z), 241 runs, pagination exhausted (not a count cap). This audit follows a 36-day gap — the previous run was 2026-07-06; no audit executed between 2026-07-07 and 2026-08-10. That gap is itself flagged below as a process anomaly.

Fleet health: 88.8% raw / 89.1% excluding intentional-failure guardrail workflows. prod-main 86.4% (108/125), non-main branches 92.1% (105/114+). Engines: copilot 132, pi 58, claude 36, codex 4, aider 2, goose 2, crush 1, unknown 4. 0 missing-tools, 0 missing-data, 0 MCP failures across all 241 runs — clean signal on tooling/config completeness.

Charts

Workflow Health — Daily Runs & Success Rate

Stacked daily success/failure counts (green/red) with success-rate line on the secondary axis across ~30 days. The shaded region marks the 36-day audit gap (07-07 → 08-10) — no data exists for that span, it is not interpolated. Today's rate (88.8%) sits in line with the last observed pre-gap trend (86–92%), suggesting the fleet's health did not materially drift during the unmonitored period, though that can't be confirmed without daily data.

Token Usage — Daily & 7-day Moving Average

Daily token bars with a true calendar-windowed 7-day moving average; the average line is deliberately broken (not drawn) across the audit gap and other missing-data spans so it never implies continuity where none exists. Today's ~8.9M tokens is consistent with historical daily volume once the fleet is fully sampled again.

Key findings

1. NEW — synchronized PR-review-cluster failure, candidate root cause identified (recommendation: rec-pr-review-firewall-allowlist-individual-copilot, MEDIUM)

6 distinct PR-review workflows (PR Code Quality Reviewer, Design Decision Gate, Test Quality Sentinel, Ponytail Reviewer, Matt Pocock Skills Reviewer, Impeccable Skills Reviewer) all failed with driver_exit at the identical timestamp 2026-08-11T13:44:42Z, triggered by the same PR commit d2e0811f.

Cross-referencing PR Code Quality Reviewer's firewall log against its workflow source (pr-code-quality-reviewer.md) shows network.allowed: [defaults, go] — this does not include api.individual.githubcopilot.com, and the firewall consistently blocked that domain in 4/6 (67%) of this workflow's runs across the full 24h sample, a stable rate rather than a one-off blip. This is a strong candidate trigger for the synchronized driver_exit cluster, since pi/copilot-backed engines depend on that domain.

Fix: add api.individual.githubcopilot.com to the network.allowed list for pi/copilot-engine PR-review workflows and re-observe the next PR-review burst for recurrence.

Single-day observation — logged as anomaly anom-pr-review-cluster-sync-fail-2026-08-11 with persistence=1, escalation=WATCH.

2. Chronic — Avenger driver_exit, unfixed 2 months (known-issue: avenger-err-config-no-structured-logs, recommendation: rec-avenger-escalate-2month-unfixed, HIGH)

2/9 Avenger runs (22%) failed with driver_exit this window. This issue was first seen 2026-06-13, has recurred 20+ times, and remains OPEN/unfixed after ~60 days. Given the persistence and recurrence count, recommending this be escalated beyond routine tracking — it has outlived the normal fix-or-triage window for a known, reproducible driver-level failure.

3. Signature evolution — Code Scanning Fixer (known-issue: code-scanning-fixer-0tok-agentjob-fail, recommendation: rec-code-scanning-fixer-new-signature, MEDIUM)

2/2 Code Scanning Fixer runs failed this window, but the failure signature has changed: originally (June) this was an instant 0-token/0-turn crash; today's failures involved ~26 minutes of real agent work before failing at the agent_logic level. This is not the same failure recurring unchanged — it indicates the underlying issue shifted from a startup/config crash to a downstream logic failure, and should be re-triaged rather than assumed already covered by the existing fix notes.

4. Already resolved — Linter Miner crash predates same-day fix

One Linter Miner agent_logic failure occurred at 17:40Z, before commit c993af6 ("Fix Linter Miner crash: unsupported model: inherited on inline sub-agents", #52133) merged at 20:19Z the same day. This is very likely already resolved — no action needed, but flagged as WATCH in workflow-trends.json in case of recurrence on the next run.

5. Process — 36-day audit cadence gap (recommendation: rec-audit-cadence-gap, HIGH)

No audit executed between 2026-07-07 and 2026-08-10 (36 days), despite this workflow presumably running on a daily schedule. Root cause is outside this audit's own visibility (no access to the audit workflow's own trigger/schedule history from within its own run). Recommend checking the Audit Agent's schedule trigger and Actions run history directly — a recurring gap like this defeats the purpose of daily trend tracking and delayed all of the above findings by over a month.

6. Isolated first-occurrence failures (no cluster, watch only)

4 workflows had single agent_logic failures with no shared signature: Daily Ambient Context Optimizer, Breaking Change Checker, Daily CLI Performance Agent, Daily Agent of the Day Blog Writer. No action recommended beyond continued monitoring; logged in workflow-trends.json for baseline tracking.

Note on Smoke CI

No push events to main occurred in this window, so Smoke CI's previously-tracked chronic push-failure signature was not observable — this is a coverage gap in this cycle's data, not a resolution.

Next actions

1. Add api.individual.githubcopilot.com to network.allowed for PR-review workflows using pi/copilot engines (finding 1).
2. Escalate the 2-month-unfixed Avenger driver_exit issue beyond routine tracking (finding 2).
3. Re-triage Code Scanning Fixer against its new failure signature rather than the original 0-tok crash (finding 3).
4. Investigate why the Audit Agent itself did not run for 36 days (finding 5).

References:

• §31497780681 — PR Code Quality Reviewer, synchronized cluster failure
• §31497780706 — Design Decision Gate, synchronized cluster failure

Generated by 🔍 Agentic Workflow Audit Agent · agent · 270.6 AIC · ⌖ 40.7 AIC · ⊞ 8.4K · ◷

• expires on Aug 12, 2026, 1:57 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 #52375.