[workflow-analysis] Weekly Workflow Analysis — 2026-04-06

[github-actions[bot]]

Analysis of the 20 most recent workflow runs from the past week (all on 2026-04-06). Data covers 1.4h of wall time, 9.6M tokens (10.7M effective), 198 agent turns, and 58 GitHub API calls.

Overview

75% success rate — 15 of 20 runs succeeded (3 additional are still in-flight/baseline). Five failures cluster around three distinct root causes: infrastructure/dependency installation errors, a complete firewall misconfiguration, and a safe_outputs job crash after a successful agent run.

Summary

Metric	Value
Total runs	20
Successes	15
Failures	5
Total tokens	9.6M (10.7M effective)
Total wall time	1.4h
Total turns	198
Total errors	5

---

Failures

❌ Schema Feature Coverage Checker — Firewall Misconfiguration

Run: §24022436614 · Codex · 2.8m

This is the most severe failure: 100% of network requests were blocked (43/43). The workflow has no network.allowed list in its frontmatter, so the Codex agent cannot reach api.githubcopilot.com, api.github.com, json-schema.org, json.schemastore.org, or any other domain.

Fix: Add a network block to the workflow frontmatter:

network:
  allowed:
    - api.githubcopilot.com
    - api.github.com
    - github.com
    - json-schema.org
    - json.schemastore.org

---

❌ Daily News — Exit Code 127 (Command Not Found)

Run: §24025680379 · Copilot · 4.6m

Failed at agent/Write Safe Outputs Tools with exit code 127 (command not found). Baseline comparison shows a significant regression: the last successful run (2026-03-24) completed 10 turns with write_capable posture; this run completed 0 turns in read_only posture — the agent never started.

Likely cause: A missing tool or binary in the runner environment, possibly a recently removed or renamed dependency.

---

❌ CI Cleaner — Agent Exit Code 1

Run: §24022270848 · Claude Code · 3.8m

The agent job failed (exit code 1), also with 0 turns completed. Baseline (2026-04-04, successful) ran 22 turns with write_capable posture. Same pattern as Daily News — the agent never reached execution.

---

❌ Dev — SDK Installation Failure

Run: §24025857784 · Copilot · 1.5m

Failed at the indexing job in the Install @tobilu_qmd SDK step. This is a pre-agent setup step; the agent was never reached (all downstream jobs skipped). Likely a broken or missing package version.

---

❌ Contribution Check — Safe Outputs Job Failure

Run: §24020210365 · Copilot (claude-sonnet-4.6) · 6.3m

Unusual failure: the agent ran successfully (4m, 19 turns, 8 write actions), the detection job also passed, but the safe_outputs job itself failed. The agent produced valid outputs (2 comments, 1 label, 1 issue) but they were not applied.

This is a post-agent infrastructure failure unrelated to agent behavior. It may be the same underlying issue as the Daily News/CI Cleaner failures given the same date.

Contribution Check — Agentic Assessment

The audit also flagged this workflow for optimization (independent of the failure):

• Resource-heavy for triage domain: 783K tokens, 19 turns, 6.3m for a triage task
• Partially reducible: ~73% of turns are data-gathering (PR lookups) that could move to deterministic pre-agent steps
• Cache efficiency: 44% (620K cache-read / 776K total input tokens) — already benefiting from caching
• Recommendation: Move list_pull_requests / pull_request_read calls to a pre-agent shell step writing JSON to /tmp/gh-aw/agent/, reducing inference turns and cost

---

Performance Highlights (Successful Runs)

Workflow	Engine	Tokens	Turns	Duration
Code Simplifier	Copilot	2.2M	33	10.2m
Contribution Check (success)	Copilot	1.8M	47	9.6m
Daily CLI Tools Exploratory Tester	Copilot	1.7M	31	10.7m
Daily CLI Performance Agent	Copilot	819K	17	6.6m
PR Triage Agent	Copilot	669K	15	5.0m
GPL Dependency Cleaner	Copilot	472K	11	4.9m
Copilot PR Prompt Pattern Analysis	Copilot	453K	10	6.6m
Architecture Diagram Generator	Copilot	573K	13	6.2m
Bot Detection	Copilot	74K	2	2.8m

---

Failure Patterns

Three distinct root causes account for all 5 failures:

1. Infrastructure failures (Daily News, CI Cleaner, Dev) — Exit code 127/1 in pre-agent or agent setup steps. These likely share a common infrastructure or dependency cause (all occurred on 2026-04-06 within a 2-hour window). Worth checking if a runner image update or package removal happened around 07:00 UTC that day.

2. Firewall misconfiguration (Schema Feature Coverage Checker) — Codex workflow missing network.allowed entirely. 100% block rate. Simple fix: add required domains to workflow frontmatter.

3. Safe outputs job failure (Contribution Check) — Agent completed successfully; post-agent infrastructure failed. May be related to the same infrastructure event as pattern rejig docs #1.

---

Recommendations

1. Immediate: Fix Schema Feature Coverage Checker firewall config — add network.allowed domains including api.githubcopilot.com, api.github.com, json-schema.org, json.schemastore.org.

2. Investigate infrastructure event ~07:00 UTC 2026-04-06 — Daily News, CI Cleaner, Dev, and Contribution Check all failed within a 3-hour window. Check for runner image changes, package removals, or dependency version bumps that could explain exit code 127 and SDK install failures.

3. Optimize Contribution Check — Move PR data-fetching to a pre-agent deterministic step. Estimated savings: ~73% fewer agent turns, lower token cost.

4. Monitor Daily News baseline regression — The behavior shift from write_capable/10 turns → read_only/0 turns suggests the workflow isn't even starting its agent loop. Once the infrastructure issue is resolved, verify it recovers.

---

References:

• §24022436614 — Schema Feature Coverage Checker (firewall)
• §24025680379 — Daily News (exit 127)
• §24020210365 — Contribution Check (safe_outputs failure)

Generated by Weekly Workflow Analysis · ● 136.8K · ◷

• expires on Apr 7, 2026, 9:49 AM UTC

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-04-07T09:49:24.823Z.

Closed by Workflow