[audit-workflows] Agentic Workflow Audit — 2026-07-29: Logs collection outage blocks fresh data

[github-actions[bot]]

Overview

This audit could not collect fresh workflow-run logs. The agenticworkflows logs MCP tool was queried four times with different parameters — none returned any data. That failure is the headline finding; everything else below is reconstructed from repo-memory history (last good data: 2026-07-06).

Critical finding: the last entry in audit-history.jsonl before today is dated 2026-07-06 — a 23-day gap to this run. Combined with today's total tool failure, this suggests the logs backend may have been degraded since early July, with prior audit cycles silently collecting nothing.

Key issue: logs collection outage (P1)

Attempt	Params	Result
1	start_date=-1d (full window)	0 runs, "Timeout reached before any runs could be downloaded"
2	start_date=-1d, count=150, timeout=280	0 runs, "No runs found matching the specified criteria"
3	start_date=-3d, count=30, timeout=200	0 runs (identical cache output to #2)
4	no date filter, count=10, timeout=200	0 runs (identical cache output to #2)

Each call took 200–300 seconds before giving up. Since this repo typically runs 100–400 agentic workflows/day (per historical memory), zero runs under every query shape points to a broken download path, not a real activity lull. Filed via missing_data safe-output this cycle, plus a new logs-mcp-tool-zero-runs-outage entry in known-issues.json and a P1 recommendation in recommendations.json.

Next action: whoever owns the agenticworkflows/aw-mcp bridge should check connectivity/auth on the logs-download path before the next scheduled audit — until fixed, this workflow (and any other audit relying on the same tool) will keep reporting nothing.

Historical trends (through 2026-07-06 — not current)

Success rate bounced in an 60–98% band with no sustained trend, generally clustering 85–95% outside of a couple of incident days (07-02, 07-03 troughs tied to a pi-engine/PR Sous Chef collapse that later recovered). Run volume varied 20x day-to-day (20 to ~400), mostly reflecting partial-window collection rather than real traffic swings.

Token reporting itself has a known gap (token-usage-reporting-gap): only 10 of the last 21 tracked days have any token figure at all, split into two disconnected clusters (06-13→06-20, then 07-03→07-05). The moving-average line is deliberately broken across that gap rather than interpolated, so it isn't mistaken for a continuous decline — it's two separate short observation windows.

Top chronic known issues (carried forward, unaffected by today's outage)

• copilot-sdk-driver-failures (recurrence 24) — dominant fleet fail family: copilot "Execute GitHub Copilot CLI" step exits 0-turn/0-tok, both instant (~3-5m, e.g. Smoke CI) and long-run (18-52m of real work then a driver fail) variants. ~71-78% of fleet failures on recent good-data days.
• smoke-ci-copilot-cli-100pct-fail-on-push (recurrence 6) — Smoke CI fails 100% of the time on every push to main; single largest reliability drag (was ~31% of all fleet fails as of 07-06).
• avenger-err-config-no-structured-logs (recurrence 18) — Avenger's follow-up engine invocation returns no structured logs after a successful pass, hard-failing the job; recommend disabling its schedule until the err-config path is fixed.
• chroot-node-not-available (recurrence 11) — Node.js unreachable inside the AWF chroot for Copilot-CLI-dependent workflows (e.g. Daily Issues Report Generator).
• codex-gh-aw-binary-not-found-for-mcp (recurrence 5) — Daily Cache Strategy Analyzer's codex MCP helper binary lookup path is stale; ~100% fail rate, unfixed 2+ weeks as of last data.
• token-usage-reporting-gap (recurrence 3) — fleet TokenUsage metric is null on most days despite runs happening; AIC (AI credits) is the reliable cost proxy in the meantime.

Zero missing-tools, zero missing-data (other than today's), and zero MCP failures were recorded on every prior good-data day — the fleet's problems are concentrated in a handful of chronic driver/engine-startup failure classes, not tooling gaps.

Next actions

1. P1 — Fix the logs MCP tool / backend. Blocks all fleet observability, not just this audit. See known-issues.json:logs-mcp-tool-zero-runs-outage and recommendations.json:rec-2026-07-29-fix-logs-tool-outage.
2. Confirm the 23-day audit gap. Determine whether workflows kept running normally 07-07→07-28 with only the audit broken, or whether the outage is broader. If run artifacts are still retrievable once the tool is fixed, backfill that window.
3. Chronic issues above remain open and unchanged by this cycle — they're carried forward from 2026-07-06, not re-verified today.

Repo memory (known-issues.json, anomalies.json, recommendations.json, metrics-summary.json, audit-history.jsonl) has been updated to record this outage; validated at 64 KB total / 24 KB patch (under the 60 KB patch limit).

References:

• This audit run (§30492378500)

Generated by 🔍 Agentic Workflow Audit Agent · age00 · 199.2 AIC · ⌖ 28.4 AIC · ⊞ 8.3K · ◷

• expires on Jul 30, 2026, 2:02 PM UTC-08:00

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-07-30T22:02:15.519Z.

Closed by Workflow