[audit-workflows] Daily Audit — 2026-06-01: 96.2% success (14-day high); both failures = one safe-output target=* class

[github-actions[bot]]

Overview

Audit of the last 24h of agentic workflow runs (window ending 2026-06-01 ~22:15 UTC): 53 completed runs, 51 success, 2 failure — 96.2% success, the best rate in the 14-day trend. Cost was flat and low ($15.62 claude-measured), and three previously-active failure classes (token-budget 429, threat-detection missing prompt, PR-branch-deleted race) all went quiet. The headline: both of today's failures are the same class — the recurring safe-output target="*" partial-failure intolerance — which is now the dominant and only repeating failure mode.

Summary

Metric	Value
Completed runs	53 (51 ✅ / 2 ❌) + 2 in-progress
Success rate	96.2% (14-day high)
Raw / effective tokens	38.1M / 324.5M
Cost (claude-measured1)	$15.62 (14-day low)
Turns / action-minutes	792 / 449
GitHub API calls	483
Safe-output items	26
Missing tools / data / MCP failures	2 (benign) / 0 / 0
Firewall blocked	700 / 3635 (19.3%)
Engines	copilot 36, claude 12, codex 4, antigravity 1, gemini 1, pi 1

1 Only the claude engine reports EstimatedCost; copilot/codex/gemini/pi/antigravity report 0, so total cost is claude-biased.

Critical Issues — both failures are one class ⚠️

Both red runs were the safe-output-partial-failure-intolerance class: one invalid target="*" item (with no resolvable issue/PR number) red-fails the entire safe_outputs job even though sibling items succeeded.

• Contribution Check — §26784313244 (copilot/sonnet-4.6). An add_comment item had target="*" with no item_number on a schedule event → Message 3 (add_comment) failed → whole job red despite create_issue + add_labels succeeding. This is a recurrence on the same workflow (also failed 05-30).
• Smoke Claude — §26784371743 (claude). Two create_pull_request_review_comment items had target="*" with no pull_request_number → Messages 8,9 failed → whole job red despite other smoke outputs succeeding. New workflow in this class. (The agent itself reported PARTIAL — all functional smoke tests passed.)

This class now spans 6 workflows since 05-26. Root defect is unchanged: Process Safe Outputs fails the whole job on any failed item instead of skip-with-warning when ≥1 item succeeded.

Good news — three classes went quiet 🟢

• Token-budget 429 ([aw-failures] Token-budget exhaustion (25M effective-tokens cap) recurring across 6+ scheduled workflows — 2026-05-29 02:00–07:32 UTC #35661) did NOT recur. Heaviest effective-token run was Linter Miner at 22.5M, under the 25M cap (Copilot Opt 20.2M, Smoke Copilot 20.2M also under). After two consecutive days of cap-blowouts (05-30 Linter Miner, 05-31 Daily Firewall Logs), this window stayed clear — severity downgraded high→medium.
• Threat-detection missing-prompt (05-31 Code Simplifier) did not recur.
• PR-branch-deleted race did not recur — Test Quality Sentinel (2×) and Design Decision Gate (3×) all passed.
• No cost outlier — Go Logger Enhancement (yesterday's $8.30 / 117-turn tail) was absent; today's most expensive run was Daily Code Metrics at $2.98.
• [aw] Failure Investigator completed cleanly ($2.45, 29 turns) — no CLI-step timeout/early-exit.

Trend charts (last ~14 days)

Workflow Health

Success rate climbed to 96.2%, the highest in the tracked window, recovering well above the high-80s/low-90s band and far from the 05-23 dip (41.6%). Failure count dropped to 2, both from a single class rather than scattered systemic regressions.

Token Volume & Cost

Raw token volume (38.1M) and claude-measured cost ($15.62) are both at the low end of the window, pulling the 7-day moving average down after the 05-31 spike ($31.63, driven by the Go Logger $8.30 outlier). No heavy-tail cost anomaly this window.

Capability & network details

Missing tools (2 — both benign smoke probes)

• mcpscripts-gh — Smoke Claude (test Add workflow: githubnext/agentics/weekly-research #2 probe; self-corrected with github_pr_query).
• web-fetch MCP — Smoke Codex (probe; unavailable by design).

Neither is a real capability gap; both are intentional smoke-test probes. 0 MCP failures, 0 missing-data signals.

Firewall (19.3% blocked, up from 16.7%)

The uptick is driven by new-engine smoke tests added this window, not workflow regressions. Hotspots: Smoke Antigravity 10/12 (83%), Smoke Copilot 118/325 (36%), Linter Miner 63/243 (26%). Blocks are predominantly Google telemetry (content-autofill, accounts.google, www.google), Playwright azureedge, localhost, and (unknown) — by-design noise. No firewall block caused either failure.

New-engine coverage 🆕

Three engines joined smoke coverage this window — Smoke Antigravity, Smoke Gemini, Smoke Pi — all passed.

Drift watch

PR Code Quality Reviewer turn count varied 3 → 22 (avg 13.8) across 3 successful runs — a 7× spread suggesting task-shape/prompt instability. No failures; monitoring.

Recommendations

1. [High] Make Process Safe Outputs partial-failure tolerant — treat an individual failed/invalid item as skipped-with-warning whenever ≥1 item in the batch succeeded, instead of red-failing the whole job. This alone would have turned both of today's failures green.
2. [High] Validate target="*" at the MCP emit boundary — reject an add_comment / create_pull_request_review_comment whose triggering context can't resolve to a concrete number (e.g. schedule events) so the agent self-corrects in-loop. Prompt-only guardrails have now failed across 6 workflows.
3. [Medium] Keep token-budget margin preventives ([aw-failures] Token-budget exhaustion (25M effective-tokens cap) recurring across 6+ scheduled workflows — 2026-05-29 02:00–07:32 UTC #35661) — heavy aggregation workflows still run at 20–22.5M effective tokens, near the 25M cliff; reduce footprint + add harness fail-fast on the 429 signature.
4. [Medium] Harden threat-detection prompt check — don't mask the cp with || true; verify the prompt exists before invoking the detection agent.
5. [Low] Watch PR Code Quality Reviewer turn drift; optional checkout skip-cleanly guard for deleted PR branches.

References:

• §26784313244 — Contribution Check (safe-output target=* failure)
• §26784371743 — Smoke Claude (safe-output target=* failure)
• §26785160929 — this audit run

Generated by 🔍 Agentic Workflow Audit Agent · opus48 3.3M · ◷

• expires on Jun 2, 2026, 10:28 PM UTC

[pelikhan]

/q remove estimated cost from report in this agentic workflow

[github-actions[bot]]

🔧 Pay attention, 007! Q is preparing your gadgets for this discussion comment...

[github-actions[bot]]

💥 WHOOSH! 🦸 The Smoke Test Agent ZOOMS in! 🌪️

KA-POW! All systems checked, gadgets tested, MCP servers BAM'd into shape! 💪 The Claude engine is running NOMINAL across the multiverse! 🚀

The Smoke Test Agent was here! ✨🦾 THWIP!

Warning

Firewall blocked 6 domains

The following domains were blocked by the firewall during workflow execution:

• accounts.google.com
• android.clients.google.com
• clients2.google.com
• contentautofill.googleapis.com
• safebrowsingohttpgateway.googleapis.com
• www.google.com

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

network:
  allowed:
    - defaults
    - "accounts.google.com"
    - "android.clients.google.com"
    - "clients2.google.com"
    - "contentautofill.googleapis.com"
    - "safebrowsingohttpgateway.googleapis.com"
    - "www.google.com"

See Network Configuration for more information.

💥 [THE END] — Illustrated by Smoke Claude · opus48 1.1M · ◷

[github-actions[bot]]

Smoke beast was here. Tiny sparks. Repo still stand. 🔥

Warning

Firewall blocked 5 domains

The following domains were blocked by the firewall during workflow execution:

• accounts.google.com
• clients2.google.com
• contentautofill.googleapis.com
• safebrowsingohttpgateway.googleapis.com
• www.google.com

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

network:
  allowed:
    - defaults
    - "accounts.google.com"
    - "clients2.google.com"
    - "contentautofill.googleapis.com"
    - "safebrowsingohttpgateway.googleapis.com"
    - "www.google.com"

See Network Configuration for more information.

📰 BREAKING: Report filed by Smoke Copilot · gpt54 25.2M · ◷

[github-actions[bot]]

Smoke beast was here. Tiny sparks. Repo still stand. 🔥

Warning

Firewall blocked 5 domains

The following domains were blocked by the firewall during workflow execution:

• accounts.google.com
• clients2.google.com
• contentautofill.googleapis.com
• safebrowsingohttpgateway.googleapis.com
• www.google.com

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

network:
  allowed:
    - defaults
    - "accounts.google.com"
    - "clients2.google.com"
    - "contentautofill.googleapis.com"
    - "safebrowsingohttpgateway.googleapis.com"
    - "www.google.com"

See Network Configuration for more information.

📰 BREAKING: Report filed by Smoke Copilot · gpt54 25.2M · ◷

[github-actions[bot]]

💥 KA-POW! 🦸 The Smoke Test Agent BURST through the firewall — WHOOSH! 🌪️ All systems checked, all gizmos GLEAMING! ⚡ The Claude engine roars to life... VROOOM! 🚀 "This repo is SAFE for another day!" 🛡️✨ THWIP! Until next time, citizens! 💨

Warning

Firewall blocked 6 domains

The following domains were blocked by the firewall during workflow execution:

• accounts.google.com
• android.clients.google.com
• clients2.google.com
• contentautofill.googleapis.com
• safebrowsingohttpgateway.googleapis.com
• www.google.com

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

network:
  allowed:
    - defaults
    - "accounts.google.com"
    - "android.clients.google.com"
    - "clients2.google.com"
    - "contentautofill.googleapis.com"
    - "safebrowsingohttpgateway.googleapis.com"
    - "www.google.com"

See Network Configuration for more information.

💥 [THE END] — Illustrated by Smoke Claude · opus48 1M · ◷

[github-actions[bot]]

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

A newer discussion is available at Discussion #36398.