[audit-workflows] Agentic Workflow Audit — 2026-07-15: 81.8% (90% ex-Smoke CI), 18 fails all known CLI-driver issues, 0 novel #45829
Replies: 2 comments
-
|
Smoke bot grunt.\nRun 29459778281 poke done.\nTiny fire still warm.
|
Beta Was this translation helpful? Give feedback.
0 replies
-
|
Smoke test: discussion interaction passed ✅
|
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
🔍 Agentic Workflow Audit — 2026-07-15
Window: partial ~5.6h (15:38–21:14Z) · 100 runs fetched (99 concluded + 1 in-progress). The full
-1dlogsfetch timed out at the 120s bridge cap; 225 run dirs were downloaded and 100 indexed intosummary.json— this is a representative partial slice, consistent with recent windows.Health at a glance
Every one of the 18 failures is a
Execute *-CLIdriver failure (0-turn or long-run-then-0-turn) that maps to an existing known issue. Zero novel anomalies, zero missing tools, zero MCP failures this window.Per-engine reliability
pi and claude are clean. All 18 failures are copilot (17) + codex (1) — the fleet's reliability drag is entirely the copilot/codex CLI-startup driver family.
Failure clusters (all known issues)
1. Smoke CI — 9 fails (50% of all failures) · DOMINANT chronic
copilot/claude-sonnet-4.6, fails on push (7) and pull_request (2) at Execute GitHub Copilot CLI, 0-turn / 0-tok, 4.3–5.4m each. 100% deterministic on every push tomain.smoke-ci-copilot-cli-100pct-fail-on-push— recurrence Weekly Research Report: AI Workflow Automation Landscape and Market Opportunities - August 2025 #7, chronic 06-30 → 07-15, unfixed.main.2. copilot-sdk-longrun — 8 fails · agent runs then Execute-CLI drops
Agent consumes tokens (40K–61K) over 6–29 min, then Execute GitHub Copilot CLI fails 0-turn. Affected this window:
copilot-sdk-driver-failures— recurrence Logs revamp. #25, chronic since 06-02.3. codex-gh-aw-binary — 1 fail
Daily Cache Strategy Analyzer (
codex), 7.0m, 0-turn at Execute Codex CLI.codex-gh-aw-binary-not-found-for-mcp— recurrence add cli flag to guard dropping a agentic workflow instructinos file #6, ~100% fail, unfixed since 06-15.Trend charts (30-day)
Fleet health is stable within its recent band (82–90%). Today's 81.8% sits at the low edge, but the ex–Smoke CI 90% is squarely on baseline — the headline dip is one chronically-broken workflow, not a fleet regression. The deep troughs on 06-20 (Skillet) and 07-02/07-03 (pi PR Sous Chef collapse) were prior one-off incidents that have since recovered.
The token chart shows the well-documented
metrics.TokenUsagereporting outage: daily fleet tokens crater from 40–55M/day (early June) to 1–4M/day from ~06-20 onward. Per-runtoken_usageis still present (e.g. 60,679 on Daily Formal Spec Verifier), and AIC remains reliable — so cost is trackable, but raw token trend lines are not. This is a reporting gap, not a real usage collapse.Recommendations (evidence-linked, all pre-existing)
rec-investigate-distributed-0tok-degradation) — 100% deterministic red onmain, 50% of all failures, 7 windows unfixed. Highest-leverage single fix: add a fast-fail diagnostic surfacing whyExecute GitHub Copilot CLIaborts 0-turn, and compare its copilot config vs. copilot workflows that succeed. Evidence: §29450871518.rec-sdk-toolperm-allowlist-or-relax-guard) — 8 workflows burn 15K–61K tokens then drop at Execute-CLI. Investigate the SDK headless-path session/tool-permission handling on long runs. Evidence: §29431830781.fix-cache-strategy-binary-path) — one chronically-100%-broken workflow since 06-15; ensure the gh-aw binary resolves for the codex MCP setup.metrics.TokenUsageaggregation (token-usage-reporting-gap) — or formally standardize on AIC as the fleet cost metric so token trend charts stop showing a phantom collapse.Bottom line
No new problems, no missing tools, no MCP failures. The fleet is healthy on baseline once the single chronic Smoke CI offender is excluded (90%). The failure surface is fully explained by three long-standing, already-tracked CLI-driver issues — the actionable move remains fixing Smoke CI, which alone accounts for half of all failures and a constant red on
main.References:
Warning
Firewall blocked 1 domain
The following domain was blocked by the firewall during workflow execution:
awmgmcpgSee Network Configuration for more information.
Beta Was this translation helpful? Give feedback.
All reactions