[audit-workflows] Daily Agentic Workflow Audit — 2026-04-14

[github-actions[bot]]

Summary

The 24-hour audit for 2026-04-14 analyzed 50 workflow runs across the github/gh-aw repository. All runs completed cleanly with zero errors, zero missing tools, and zero missing data — the healthiest audit window observed to date. Of the 47 completed runs, 10 succeeded with meaningful AI execution and 37 were appropriately skipped (correct event-based filtering).

Overall health: ✅ Excellent

Metric	Value
Total Runs	50 (47 completed, 3 in-progress)
Successful	10
Skipped	37
Errors	0
Missing Tools	0
Total Tokens	2,671,581
Effective Tokens	1,730,507
Total Cost	$1.03
Total Turns	55
GitHub API Calls	43

---

Active Workflow Performance

The following 6 workflows executed with real AI compute:

Workflow	Engine	Tokens	Turns	Duration	Result
Design Decision Gate 🏗️	claude	1,007,396	17	6.0m	✅ success
Test Quality Sentinel	copilot	438,416	9	6.6m	✅ success
Daily Workflow Updater	copilot	375,561	10	5.4m	✅ success
Issue Monster	copilot	360,705	8	5.1m	✅ success
Design Decision Gate 🏗️	claude	293,691	6	4.1m	✅ success
Test Quality Sentinel	copilot	195,812	5	4.0m	✅ success

---

⚠️ Medium-Priority Findings

Execution Drift — Design Decision Gate 🏗️

The Design Decision Gate showed significant turn variability: 6 to 17 turns across two runs (avg 11.5). This may indicate:

• Inconsistent input complexity (different PR sizes/designs)
• Prompt instability under varying context lengths
• Legitimate task-shape variation

Recommendation: Monitor over the next 7 days. If variance persists >3x, review prompt determinism and add turn-budget guardrails.

Network Firewall — AI Moderator

The AI Moderator workflow had 2 blocked requests out of 5 (40%) targeting ab.chatgpt.com:443 and chatgpt.com:443. These appear to be external ChatGPT endpoints accessed during moderation checks — not expected in the workflow's design.

Recommendation: Review the AI Moderator prompt/tools to ensure it isn't inadvertently making outbound calls to ChatGPT. If Codex/OpenAI routing is intended, verify the domains are allowlisted in the workflow config.

Reliability — 3 High-Anomaly Events

Cross-run analysis detected 3 events with anomaly score > 0.6 in the tool_result stage. These represent newly observed log templates rare enough to flag. No immediate action required, but worth monitoring for recurrence.

---

Firewall Analysis

Metric	Value
Total Requests	159
Allowed	147 (92.5%)
Blocked	12 (7.5%)

Blocked domains:

• ab.chatgpt.com:443 — AI Moderator (2 requests)
• chatgpt.com:443 — AI Moderator (2 requests)
• - (unnamed) — Daily Workflow Updater (10 requests, likely internal routing)

---

MCP Tool Usage

Server	Tool	Calls
github	pull_request_read	12
github	issue_read	8
github	search_repositories	4
safeoutputs	add_comment	5
safeoutputs	noop	3
safeoutputs	assign_to_agent	3
safeoutputs	push_to_pull_request_branch	1
safeoutputs	submit_pull_request_review	1
safeoutputs	add_labels	1

All 50 runs stayed read-only except 1 PR push + 1 review submission + 5 comments — actuation rate matches expected patterns.

---

📊 Trend Charts (Last 3 Audit Days)

Charts reflect data from 2026-04-11, 2026-04-12, and 2026-04-14.

Workflow Health Trends

Today's run shows a sharp drop in raw token count vs. Apr 12 (2.7M vs 47M) and significantly lower costs ($1.03 vs $15.73). This reflects fewer active PR/issue triggers today rather than a problem — skipped runs are appropriately filtered. The 100% success rate continues the trend from Apr 11's 93% baseline.

Token Usage & Cost Trends

Daily costs dropped dramatically from the Apr 12 spike ($15.73 for 85 runs) to $1.03 today (50 runs). The Apr 12 spike correlates with 37 successful runs vs. today's 10 — fewer incoming PRs/issues means fewer AI-powered reviews triggered. Efficiency per run (tokens/turn) remains stable.

---

Recommendations

1. Watch Design Decision Gate drift — set alert if turn count exceeds 20 in a single run
2. Audit AI Moderator firewall — investigate why it's calling chatgpt.com endpoints and block or allowlist intentionally
3. Continue monitoring reliability anomalies — 3 high-anomaly events are low-severity today but worth tracking

---

References:

• §24422974019 — This audit run
• Workflow run artifacts — Download charts

Generated by Agentic Workflow Audit Agent · ● 511K · ◷

• expires on Apr 15, 2026, 9:26 PM UTC

[github-actions[bot]]

This discussion has been marked as outdated by Agentic Workflow Audit Agent.

A newer discussion is available at Discussion #26496.