[audit-workflows] Agentic Workflow Audit — 2026-07-09 (partial): broken feature branch dominates fails; chronic copilot 0-turn family persists

[github-actions[bot]]

🔍 Agentic Workflow Audit — 2026-07-09

Window: partial ~8.75h (12:34–21:20Z), 2 log pages deduped to 111 unique runs (logs bridge caps each fetch at 60–86s, so this is a mid/late-day sample, not the full 24h).

Headline

Metric	Value
Overall success	70/111 = 63.1% ⚠️
Main-branch only	61/80 = 76.25%
Ex broken-branch cluster	77.8%
AIC (partial sample)	6,459
Missing tools / data / MCP / safe-item errors	0 / 0 / 0 / 0 ✅
All failures	0-turn / pre-agent (zero agent-logic failures)

The low 63.1% headline is not a fleet-health collapse — it is dominated by a single broken feature branch. Strip that out and the picture is the familiar chronic copilot 0-turn family.

📊 Trends

Success rate swings between ~86–95% on full days and dips on partial/count-limited windows (07-02 60%, 07-03 68%). Today's 63% point is a partial sample skewed by the branch cluster below; main-only sits at 76%. The 06-20 and 07-06 tall bars are full-day (300–400 run) windows near baseline.

Cost is charted in AI credits (AIC) because fleet token totals (metrics.TokenUsage) have been null since ~06-19. AIC tracks window length (partial-day sampling), peaking on the full-day 07-06 window (~29.6K). Today's ~6.5K is proportional to the 8.75h sample.

🚩 Dominant driver — NEW: broken feature branch reds the trial-workflow fleet

copilot/feature-content-redaction — 21/21 pre-agent failures (51% of all fails)

Every push-triggered run on branch copilot/feature-content-redaction failed with null engine, null duration, 0 action-minutes — i.e. the workflow job died before the agent ever started (compile / .lock.yml validation / activation), across 9 distinct trial workflows:

Archie, Dev Hawk, Issue Monster, Auto-Triage Issues, Contribution Check, Approach Validator, CI Failure Doctor, Impeccable Skills Reviewer, Workflow Craft Agent

Several fired multiple times (Issue Monster ×3, Auto-Triage ×2, Dev Hawk ×2) — every push re-fans the whole fleet and all red. This is isolated to one in-development branch and is not a prod-main or engine regression; it simply pollutes fleet failure counts. Removing it lifts the sample from 63.1% → 77.8%.

Action: notify the branch owner to recompile the lock files (agenticworkflows compile) and fix the early-job/activation error, and/or scope these trial workflows so an arbitrary feature-branch push doesn't fan out across the fleet.

🔁 Recurring known issues (all unfixed, all 0-turn)

Chronic clusters confirmed again this window

Cluster	Count	Detail	Known issue (recurrence)
copilot-sdk longrun 0-turn	12	Agent works 13–34m then Execute Copilot CLI fails 0-turn. Daily Project Perf Summary MCP 29.3m, Daily Security Observability 34.5m, Linter Miner 26.6m, Daily Formal Spec Verifier 24.4m, The Daily Repository Chronicle 19.9m, +7 more	copilot-sdk-driver-failures (recur 25)
Smoke CI on push	1	5.0m 0-turn at Execute Copilot CLI; chronic red on main	smoke-ci-copilot-cli-100pct-fail-on-push (recur 7)
Daily Cache Strategy Analyzer	1	7.1m 0-turn, codex gh-aw binary/model-param; ~100% fail since 06-15	codex-gh-aw-binary-not-found-for-mcp (recur 6)
Agent Persona Explorer (pi)	1	7.2m 0-turn — NEW, observability-flagged 100% hotspot	new / watch
PR Sous Chef	2	pi 13.9m + null-agent 23.6m — pi no longer 100% this window (18/20)	0-turn agent-job fail

copilot-sdk-longrun-0turn remains the top real prod-main failure class once the branch cluster is excluded. claude was clean (13/13 = 100%).

📈 Data-quality gaps

• Token + turn reporting: metrics.TokenUsage null fleet-wide since ~06-19; additionally total_turns = 0 across all 111 runs including the 70 successes this window — turn-count reporting is also broken, not just tokens. AIC remains reliable. (token-usage-reporting-gap, recur 4)

✅ Recommended actions (evidence-linked)

1. [MEDIUM · NEW] Recompile lock files / fix pre-agent failure on copilot/feature-content-redaction, or scope trial workflows off feature-branch pushes — 21 fails, 9 workflows.
2. [HIGH · recur 25] Instrument the copilot Execute GitHub Copilot CLI step to emit stderr + exit code — the longrun 0-turn family (12 fails, 13–34m wasted compute each) remains opaque and is the rejig docs #1 real prod-main drag.
3. [HIGH · recur 7] Root-cause Smoke CI 100%-on-push — if it's the copilot-engine canary, its constant red signals a real startup regression.
4. [HIGH · recur 6] Fix Daily Cache Strategy Analyzer codex MCP-binary path (~100% fail since 06-15).
5. [LOW · recur 4] Restore metrics.TokenUsage aggregation or officially adopt AIC as the fleet cost metric; investigate zero turn-counts.

References:

• §29050289232 — Daily Project Perf Summary (MCP) longrun 0-turn 29.3m
• §29039518172 — Smoke CI 0-turn on push
• §29032591838 — Archie, feature-content-redaction pre-agent fail

Warning

Firewall blocked 1 domain

The following domain was blocked by the firewall during workflow execution:

• awmgmcpg

To allow these domains, add them to the network.allowed list in your workflow frontmatter:

network:
  allowed:
    - defaults
    - "awmgmcpg"

See Network Configuration for more information.

Generated by 🔍 Agentic Workflow Audit Agent · 302 AIC · ⌖ 41.6 AIC · ⊞ 7.4K · ◷

• expires on Jul 10, 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 #44841.