Parallel CI Triage
Speed up fixing broken CI by splitting failing jobs (or independent failure clusters) across parallel subagents. Each subagent owns one vertical slice: logs, root cause, code fix, and local verification for that slice.
Prerequisites
- GitHub CLI (
gh) installed and authenticated (gh auth login), or use the GitHub web UI / API to copy logs manually. - Push access to the repo so fixes can be pushed and CI re-run.
Workflow
1. Identify the failing run
From the repo root:
bashgh run list --limit 5 gh run view <RUN_ID> --log-failed
Or open the Actions tab, open the failed workflow run, and note which jobs failed (not just which step — group by job name).
If gh is unavailable, download logs from the GitHub UI and paste them into the conversation.
2. Split by job (or by failure cluster)
- One subagent per failed job when jobs test different things (e.g.
lint,test-node-18,e2e). - One subagent per independent failure cluster when a single job logs multiple unrelated errors — but prefer one job per agent to avoid conflicting edits in the same files.
If two failures share the same root cause in the same file, assign one subagent to fix both.
3. Launch parallel subagents
For each failing job, launch a generalPurpose subagent in a single message so they run concurrently:
Task: Fix CI failure for job "<JOB_NAME>" Context: - Workflow run: <RUN_URL or RUN_ID> - Branch: <branch> - Relevant log excerpt (failed steps only): <paste gh run view --job <JOB_ID> --log or the failed section> Instructions: 1. Infer the root cause from the log (command, stack trace, file:line). 2. Open and edit only what this job requires. 3. Run the same commands locally that failed in CI (or the narrowest equivalent, e.g. `npm run lint`, `pytest tests/foo`, `pnpm test --filter pkg`). 4. Report: what failed, what you changed, and confirmation that the local command passes.
Include the exact failing command and error lines so the subagent does not guess.
4. Merge and verify
-
Collect each subagent’s changed files. Resolve overlaps manually if two agents touched the same file.
-
Run the full CI-equivalent locally when possible:
bash# Example: match your repo npm run lint && npm test -
Commit with a conventional message, push, and re-check the workflow:
bashgh run watch
When to use
- Multiple GitHub Actions jobs failed and the failures look independent.
- A long workflow log is easier to split by job than to fix sequentially.
When not to use
- A single job with one clear error — fix it in the main agent.
- Failures that are purely flaky infrastructure — retry or fix workflow config first.
Notes
- Redact secrets if pasting logs into chat.
- If agents conflict on shared files, merge sequentially after the parallel pass.
