[workflow-analysis] Weekly Workflow Analysis — week ending 2026-06-29

[github-actions[bot]]

Snapshot of agentic workflow activity in github/gh-aw. The logs tool times out (60s server deadline) above ~40 runs, so the fleet-wide figures below are a 40-run window (2026-06-29 09:13–10:31, ~8.1h of compute), and the reliability deep-dive is a true 7-day window for the only workflow that failed.

Key metrics (40-run window)

Metric	Value
Runs	40
Action-minutes	508
Avg duration	12.2 min/run
GitHub API calls	447
Failures	1 (2.5%)
Write-capable safe outputs	0 — all 40 read-only
Engines	copilot 30 · claude 7 · pi 2 · codex 1

Reliability — one failure hotspot

PR Description Updater: 2 failures / 20 runs over 7 days (10%). Both failures landed exclusively on signed/jsweep/* branches (unified-timeline, ai-credits-context); all 18 non-jsweep runs passed. This is a deterministic branch-class pattern, not random flake. Both failed runs were read-only (no safe outputs emitted) and burned ~10k tokens / ~7–10 min before erroring — wasted spend.

Failure detail

• §28363364071 — failure, 10.5m, signed/jsweep/unified-timeline
• §28318796220 — failure, 7.3m, signed/jsweep/ai-credits-context
• 18/20 other runs over the week succeeded. Root-cause error text wasn't retrievable (network-restricted), but the 100% correlation with signed/jsweep/ strongly suggests the workflow mishandles auto-generated/signed bot branches.

Performance / resource usage

Top action-minute consumers (review-class workflows dominate):

Workflow	Min	Runs
PR Code Quality Reviewer	65	4
Matt Pocock Skills Reviewer	63	4
Impeccable Skills Reviewer	51	4
Design Decision Gate	33	4

Three skills/quality reviewers consume ~180 action-min combined — the largest cost center. PR Description Updater alone averages ~9 min/run.

Optimization opportunities

1. Skip jsweep branches in PR Description Updater — add a branch filter (signed/jsweep/*) so it stops failing and wasting ~8 min + tokens per jsweep PR. Fixes the only recurring failure.
2. Model downgrade for read-only maintenance — audit flagged PR Description Updater (repo-maintenance, read-only, moderate profile) as not needing a frontier model; claude-haiku-4-5 / gpt-4.1-mini would cut per-run cost.
3. Reviewer consolidation — 3 skills reviewers run on every PR (~60+ min total); consider merging or gating to relevant changes.
4. Tooling: logs --count > ~40 exceeds the 60s deadline, blocking true week-wide fleet analysis — raise the timeout or add pagination.

References: §28363364071 · §28318796220 · §28365689771

Generated by 🔍 Weekly Workflow Analysis · 108 AIC · ⌖ 9.89 AIC · ⊞ 4K · ◷

• expires on Jun 30, 2026, 3:07 AM UTC-08:00