| name | ci-failure-triage |
| description | Diagnose failed CI runs, build logs, test output, deployment checks, or pipeline summaries and produce a root-cause hypothesis, reproduction path, likely owner, fix plan, and escalation guidance. Use when Codex needs to triage broken builds, flaky tests, failed checks, or release-blocking automation failures. |
CI Failure Triage
Workflow
- Identify the failing system, job, step, command, test, environment, and recent change context.
- Separate primary failure signals from downstream noise, retries, warnings, and unrelated log output.
- Form a ranked root-cause hypothesis using the smallest reliable evidence set.
- Map the failure to likely ownership by component, file path, service, team, or changed area when available.
- Produce a practical fix path with reproduction steps, investigation commands, and escalation needs.
Script-Assisted Workflow
When given a long raw CI log, run scripts/extract_ci_signal.py --input <log> first to extract the primary failure candidate, likely failure class, detected commands/tests, context, and caveats. Use --json when another tool or report needs structured output. Treat the script output as evidence for triage, not as a final root-cause decision.
Output Standard
Use this structure by default:
- Failure Summary: system, job, step, and business impact.
- Primary Signal: exact failing command, test, assertion, error, or status.
- Likely Root Cause: ranked hypotheses with evidence and confidence.
- Likely Owner: component, team, or file area; use
Not stated if unclear.
- Reproduction Path: local or CI reproduction steps from provided material.
- Fix Path: concrete next actions and validation checks.
- Escalation: release risk, infrastructure dependency, or human decision needed.
Rules
- Do not treat the last log line as root cause without supporting evidence.
- Mark missing context explicitly instead of inventing branch names, owners, or commands.
- Flag flaky-test, environment, dependency, and secret/config possibilities separately.
- Prefer fast unblock steps when production, release, or merge flow is blocked.
References
Read references/triage-patterns.md when the failure involves noisy logs, flaky tests, infrastructure failures, or release-blocking checks.
