GitHub Actions — Autonomous Failure Fix Loop

Fetch failed workflow logs via gh, diagnose root causes, fix, verify locally, commit. All failures in one pass.

Input

User may provide:

• Nothing — find failures via gh run list --status failure --limit 5
• Run URL — e.g. https://github.com/org/repo/actions/runs/123 → extract run ID
• Run/Job ID — use directly with gh run view <id> --log-failed

Context Loading

Read .github/workflows/*.yml to understand the exact commands each job runs. These are your local verify commands.

Phase 1: Fetch and Triage

gh run view <run-id> --log-failed

Group failures by root cause:

Phase 2: Fix Loop

Process in dependency order: workflow config → lint → tests → build.

1. Diagnose — exact file(s) and line(s) from the log
2. Fix — minimal change
3. Verify — run the same command from the workflow YAML locally
4. If fails — revert, re-read error, try different approach (max 3 attempts)
5. Commit — one per logical fix, conventional commit format

Phase 3: Push and Confirm

1. Push all fix commits
2. gh run list --limit 1 --json status,conclusion,url
3. If new failures appear, loop back to Phase 1

Rules

• Batch all failures — don't fix one and stop
• Read the actual log — don't guess from job names
• Verify locally before committing
• One commit per fix
• Revert failed attempts — don't stack patches
• Don't diagnose CI from unpushed local code — push first

Common Pitfalls