[detection-analysis] Detection Analysis Report — 2026-08-07 (Last 24h)

[github-actions[bot]]

Summary

• Window evaluated: last 24 full hours (UTC), ending 2026-08-07T23:20 UTC
• Total workflow runs analyzed: 43 (17 with complete log data, 26 without downloadable agent logs — see note below)
• Detection-enabled runs: 12 (70.6% of runs with complete data)
• Regular runs: 5 (29.4% of runs with complete data)
• Misconfigured workflows found: 0

Note

No misconfigured gh-aw-detection workflows were identified in this window. All findings below are based on the 17 runs that had complete, downloadable log data.

Comparison Chart

Metric	Regular Runs	Detection Runs
Total runs	5	12
Success rate	80.0%	91.7%
Failures	1	1
Avg. tokens/run	231,998	25,507

Regular runs show a much higher average token count, driven almost entirely by one outlier (Smoke Copilot - AOAI (Entra), 1,030,705 tokens, failed run). Excluding that outlier, regular-run average tokens would be ~32,322 — closer to the detection-run average.

Misconfigured Workflows

No misconfigured workflows detected in this window.

Checks performed against the 17 complete-data runs:

• Rule 1 (gh-aw-detection: false with >3 runs/7d): could not be fully evaluated — the pre-downloaded log set only covers the last 24h, and no workflow in that window repeated more than once, so there's no evidence of a violation, but a 7-day sample would be needed to confirm compliance.
• Rule 2 (name suggests audit/analyzer/report/detector/monitor/inspector without gh-aw-detection: true): no matches among observed workflow names. (The known documented exception, Daily Agentic Workflow AIC Usage Audit / agentic-token-audit.md, did not run in this window.)
• Rule 3 (detection enabled but detection steps failed): the one detection-enabled failure (Daily BYOK Ollama Test, run 31223940604) failed with 0 tokens and no detection-related errors in its logs — consistent with an LLM-backend connectivity failure, not a detection-step failure.
• Rule 4 (workflow alternates detection on/off within 24h): no workflow name appeared with mixed detection settings; PR Sous Chef ran 3 times, all with detection enabled.

View All Run Metrics

Workflow	Run ID	Detection	Status	Tokens	Engine
Smoke Copilot - AOAI (Entra)	31222505053	No	failure	1,030,705	copilot
Smoke Copilot MAI	31222506585	No	success	86,330	copilot
Smoke Copilot Small	31222507920	No	success	3,563	copilot
Smoke OTEL	31222509303	No	success	20,003	copilot
ESLint Monster	31222548390	No	success	19,391	pi
PR Sous Chef	31222749409	Yes	success	37,177	pi
Issue Monster	31223293352	Yes	success	14,731	pi
Daily AWF Spec Compiler Surfacing Review	31223464787	Yes	success	73,955	pi
Daily BYOK Ollama Test	31223940604	Yes	failure	0	copilot
PR Code Quality Reviewer	31223955453	Yes	success	0	pi
Test Quality Sentinel	31223955470	Yes	success	14,264	copilot
Design Decision Gate 🏗️	31223955476	Yes	success	14,275	claude
Matt Pocock Skills Reviewer	31223955505	Yes	success	12,353	copilot
Impeccable Skills Reviewer	31223955510	Yes	success	5,321	copilot
PR Sous Chef	31224364602	Yes	success	43,547	pi
PR Sous Chef	31225438989	Yes	success	80,509	pi
Daily Reliability Review	31225875315	Yes	success	9,956	opencode

Data limitation: 26 of the 43 runs in the window had no downloadable aw_info.json/run_summary.json (empty log directories) and are excluded from the table and metrics above. These are most likely activation-only runs where the agent job was skipped (e.g. skip-if-match conditions or concurrency deduplication), so no token/detection data exists for them — not a sign of misconfiguration.

View Historical Trend

This is the first recorded data point for detection-vs-regular comparison — a trend history file was created at cache-memory path trending/detection-comparison/history.jsonl, but there isn't enough history yet (1 data point) to render a meaningful trend chart. Future runs of this report will build up the series.

Recommendations

1. No immediate action required — no misconfigured gh-aw-detection workflows were found in this window.
2. To fully validate Rule 1 (>3 runs/7d without detection), a future run should widen the analysis window to 7 days, since the 24h sample is too short to catch low-frequency-but-non-compliant workflows.
3. Investigate the Smoke Copilot - AOAI (Entra) failure (run 31222505053) separately — its token usage (1,030,705) and failure status are unusual for a smoke test and are unrelated to detection configuration, but worth a look for cost/reliability reasons.
4. Consider adding gh-aw-detection: true visibility to the pre-downloaded log bundle itself (or documenting why ~60% of runs in a given window produce no agent logs) to reduce the number of "incomplete" runs future analyses have to exclude.

Generated by 🔎 Detection Analysis Report · agent · 109.9 AIC · ⊞ 7.6K · ◷

• expires on Aug 10, 2026, 3:27 PM UTC-08:00