[audit-workflows] Agentic Workflow Audit — 2026-06-19 (91.5% overall, prod-main rebound; Avenger 100% hotspot)

[github-actions[bot]]

🔍 Agentic Workflow Audit — 2026-06-19

Window: last-24h evening/afternoon cluster (16:18Z–21:29Z, ~18.8h compute) · 106 runs

Metric	Value	vs 06-18
Overall success	91.5% (97✓ / 9✗)	▲ +3.3pts
Prod-main success	85.2% (46 / 54)	▲ +11.0pts (rebound)
New failure classes	0	—
missing_tools / missing_data / mcp_failures / errors	0 / 0 / 0 / 0	flat

Headline: Healthy day. Prod-main rebounded from 06-18's 74.2% trough to 85.2%. All 9 failures map to known recurring issues — zero new failure class. The dominant story is Avenger, which failed 4/4 runs (100%) and is escalating day-over-day.

---

📊 Trend Charts

Overall success rate sits at 91.5%, comfortably back above the 90% reference line after the 06-16 (71.2%) dip and the 06-18 prod-main trough. The 26-day band holds in the high-80s/low-90s; today's 9 failures are concentrated in a handful of chronically-broken workflows rather than spread across the fleet.

Token history through 06-18 shows the usual 20–70M daily range with the 7-day MA steady around ~40M. Today's token point is intentionally omitted: token/turn metrics were unavailable this window (the logs were fetched without artifact download, so every token_usage.jsonl came back empty). This is an audit-side data-collection gap, not a workflow regression.

---

🔴 Failure Breakdown (all 9 known-recurring)

Expand failure detail

Workflow	#	Class	Signature
Avenger	4	avenger-err-config-no-structured-logs	0-tok/0-turn agent:failure, 17–21min. 100% of its runs.
Daily Safe Output Integrator	1	copilot-sdk tool-perm-lockout	0-tok/23min, safe_outputs+detection skipped
Daily SPDD Spec Planner	1	copilot-sdk tool-perm-lockout	0-tok/25min (detection+safe_outputs oddly SUCCESS)
PR Code Quality Reviewer	1	copilot-sdk tool-perm-lockout	0-tok/23min, PR branch
Daily Cache Strategy Analyzer	1	codex binary/model-not-found	0-tok/6min, ~100% chronic fail-rate
Documentation Unbloat	1	doc-unbloat-empty-output	0-tok/3min empty {items:[]}

Every failure is a clean agent-job redden (0 tokens, 0 turns) — no partial work was lost on these, and no safe-output partial-failure reddening was observed this window.

---

⚠️ Escalations & Recurring Issues

• 🔺 Avenger — ESCALATING (recurrence Weekly Research Report: AI Workflow Automation Landscape and Strategic Opportunities - August 2025 #12). Failed 4/4 runs (100%) today, up from 2× on 06-18. Chronic rejig docs #1 prod-main offender; the ERR_CONFIG no-structured-logs follow-up-invocation bug persists ~6 days after the fix branch (copilot/aw-avenger-failed-fix) was opened. Recommendation rec-avenger-empty-followup-invocation re-escalated to HIGH — treat a follow-up invocation that produces no structured logs after a successful first pass as no-op/success rather than hard-failing the agent job.
• copilot-sdk tool-permission-lockout — day 12 (recurrence Weekly Research Report: AI Workflow Automation Landscape and Strategic Opportunities - August 2025 #12). 3 runs (Safe Output Integrator, SPDD Spec Planner, PR Code Quality Reviewer) burned 20–25min each at 0-tok before the 5-denial abort. Top-priority recommendation rec-sdk-toolperm-allowlist-or-relax-guard remains unresolved after 12 days.
• Daily Cache Strategy Analyzer — ~100% fail (recurrence Add workflow: githubnext/agentics/weekly-research #3). codex MCP binary/model-not-found family. High-ROI fix (fix-cache-strategy-binary-path) still open.
• Documentation Unbloat — recurrence add cli flag to guard dropping a agentic workflow instructinos file #6. Empty-output / cache-git-integrity family.

---

✅ Health Signals

• Zero missing-tools, missing-data, MCP failures, or flagged errors across all 106 runs.
• Only 1 PR-branch failure (PR Code Quality Reviewer); PR/smoke noise was minimal this window.
• No new failure class introduced — the fleet's instability is fully accounted for by 4 long-standing chronic workflows.

---

🎯 Suggested Next Actions

1. Land the Avenger no-structured-logs fix — single highest-impact change; it accounts for 4 of 9 failures and is escalating.
2. Resolve copilot-sdk tool-permission lockout (allowlist routine read/git commands or relax the 5-denial abort) — 12 days unresolved, 3 prod/PR runs/day wasted.
3. Fix or disable Daily Cache Strategy Analyzer (codex) — ~100% fail-rate is pure wasted compute.
4. Audit tooling note: fetch logs with --artifacts next cycle so token/cost metrics and engine_counts are populated.

Data gaps this window

• Token/turn metrics unavailable (logs fetched without --artifacts; all token_usage.jsonl empty, summary tok=0).
• Engine classification unavailable (summary.engine_counts=none, aw_info.json artifacts not downloaded). Affected-workflow engines inferred from historical mapping in repo memory.

References:

• §27845956163 — Avenger (latest redden)
• §27844066542 — Daily Safe Output Integrator (sdk lockout)
• §27843593919 — Daily Cache Strategy Analyzer (codex)

Generated by 🔍 Agentic Workflow Audit Agent · 304.3 AIC · ⌖ 33.7 AIC · ⊞ 7.9K · ◷

• expires on Jun 20, 2026, 1:51 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 #40516.