| name | ci-triage |
| description | Triage failed CI pipelines, GitHub Actions runs, GitLab pipelines, and local quality gates to identify root cause and the smallest safe fix. Use when a build, test, lint, deploy, PR check, MR pipeline, or workflow job fails. |
CI Triage
Find why automation failed and propose the smallest safe next step. This skill is platform-neutral: GitHub Actions, GitLab CI/CD, or local scripts are all just CI signals.
Hard rules
- Read logs selectively; do not paste huge logs into durable docs.
- Never print secrets. Redact tokens, env values, auth headers, private keys, and credentials.
- Distinguish root cause from downstream failures.
- Prefer a minimal fix over broad refactors.
- Do not rerun expensive or deployment jobs unless the user asks or the repo process requires it.
Platform commands
Use the platform available for the repo/request:
- GitHub:
gh run list, gh run view, gh pr checks, gh run rerun only when appropriate.
- GitLab:
glab ci list, glab ci view, glab mr view, glab pipeline ci view when available.
- Local: inspect
package.json, Makefiles, Gradle files, CI YAML, and documented verification commands.
If platform CLIs are unavailable, explain the missing tool and use local files/log excerpts provided by the user.
Workflow
- Identify failing system: provider, pipeline/run ID, branch, commit SHA, job name, and time.
- Fetch only relevant failure logs or use user-provided excerpts.
- Classify failure:
- code/test failure
- dependency/install/cache failure
- config/YAML failure
- credentials/permissions failure
- flaky or infrastructure failure
- timeout/resource limit
- deployment/environment failure
- Locate first meaningful error. Ignore cascading noise unless it reveals a separate issue.
- Map error to changed files, recent commits, or config drift.
- Recommend one of:
- fix code/config
- rerun because likely flaky/infrastructure
- restore cache/dependency state
- ask human/admin for credentials or permissions
- gather one more specific log/artifact
- If implementing a fix, require explicit implementation approval and follow repo workflow rules.
Output contract
Return:
- Failure summary: provider, job, branch/SHA, first failing signal
- Likely root cause: concise and evidence-backed
- Classification: one category from the workflow
- Smallest next action: fix/rerun/investigate/escalate
- Evidence: short log snippets or file references, redacted
- Commands used: or commands the user should run
