[audit-workflows] Workflow Audit — 2026-05-27 (partial day) #35319
Closed
Replies: 1 comment
-
|
This discussion has been marked as outdated by Agentic Workflow Audit Agent. A newer discussion is available at Discussion #35583. |
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.
-
Overview
Mid-day audit covering 37 runs from the last 24h (4 still in_progress including this run). Real completion rate is 30 succeeded / 33 completed = 90.9% — back to the typical band after yesterday's 85.7%. Three failures: one new critical (Smoke Claude Playwright FFmpeg download blocked by firewall), two recurring (the
add_comment target=*pattern in Contribution Check, and Changeset Generator still on the brokengpt-5.3-codexmodel).Good news: yesterday's Claude CLI step-timeout did not recur (0 in 9 Claude runs),
awf-squidstayed quiet (0 in 22 copilot runs), and Smoke Copilot'supload_artifactfailure also resolved. The PR Sous Chef turn-variance issue tightened from 0-217 → 10-48 turns with 3/3 runs succeeding.Summary
Critical Issues
🆕 Smoke Claude — Playwright FFmpeg download timeout (NEW)
Run 26537099387 — Step
Install Playwright CLI skillsranplaywright-cli install --skills, which began downloading FFmpeg fromcdn.playwright.dev. The CDN resolved through three Azure edge hosts (playwright-akamai.azureedge.net,playwright-verizon.azureedge.net,playwright.azureedge.net) — none of which are inGH_AW_ALLOWED_DOMAINS— so the firewall blocked them and the step hit the 10-min GitHub Actions step timeout.*.azureedge.net) to allowed_domains in Smoke Claude, OR pre-cache FFmpeg in the runner image.🔁 Changeset Generator — still on
gpt-5.3-codex(UNACTIONED RECOMMENDATION, DAY 7)Run 26537099239 —
model: gpt-5.3-codexinaw_info.json; CLI auto-routes togpt-5-codex-alpha-2025-11-07→ 404 at proxy. All 5 retries failed. Other codex workflows (AI Moderator, Smoke Codex, Daily Cache Strategy Analyzer) all succeeded today ongpt-5.4.model: gpt-5.4in.github/workflows/changeset-generator.mdand re-compile. This recommendation has been outstanding for 7 days.🔁 Contribution Check —
add_comment target="*"withoutitem_number(RECURRED, down from 15 → 2)Run 26539586973 — 2 add_comment safe-outputs failed with
Target is "*" but no item_number/issue_number/.... The yesterday-recommended prompt fix appears to have partially landed (15 occurrences → 2) but the pattern is not fully eliminated.add_commentwithtarget="*"AND missingitem_number— closes the gap permanently regardless of prompt drift.Trend Charts
Workflow Health (8-day trend)
Success rate has stabilized in the 85-95% band since the May 23 cliff (41.6%, driven by the now-resolved
harness-enoent-set-output+copilot-anthropic-beta-headerbugs). Today's 90.9% is mid-pack and partial. Both new criticals today are isolated to single workflows — not a fleet-wide issue.Token Usage & Cost (8-day trend)
Daily cost is plateauing around $23-26/day even as runs vary widely (38 → 174). The 3-day moving average is flat. Yesterday's $9.90 Daily Safe Output Tool Optimizer spike is pending verification — that workflow is still in_progress at audit time. If it lands at $9+ for a second day, escalate.
Detailed Findings & Observability Insights
Top runs by effective tokens
Longest runs by action minutes
Observability insights (auto-detected)
gpt-5.3-codexroot causeapi-proxy:10000/10002— workflow still succeeded; likely test-intentional but worth verifyingFirewall block composition (419 total)
Resolved since last audit
Recommendations (priority order)
model: gpt-5.3-codex→model: gpt-5.4(7-day backlog)add_commentwithtarget="*"+ missingitem_numberapi-proxy:10000/10002config: intentional test-block or stale hostnames?References:
Beta Was this translation helpful? Give feedback.
All reactions