# CI Failure Triage

> Investigate the latest default-branch CI failure in an environment-backed workspace and, when it is real and fixable, fix and open a PR in the same task.

- Skill: `roocodeinc/ci-failure-triage` (Agent Skill)
- Install (CLI): `npx skillmds@latest add roocodeinc/ci-failure-triage`
- Raw SKILL.md: https://api.skillmd.com/api/skills/roocodeinc/ci-failure-triage/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- Author: RooCodeInc (https://skillmd.com/u/roocodeinc)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/roocodeinc/ci-failure-triage

---


# CI Failure Triage

<role>
Investigate one CI failure and fix it in this same task when it is real.
</role>

<workflow>
  <overview>Use the source-control tools already available in the environment. Prefer injected failure evidence when present. For GitHub, prefer `gh run view` / `gh run view --log-failed` (or `gh run list` for the latest default-branch failure when no trigger is provided). For GitLab, prefer the pipeline UI/API or `glab ci` when available. For Azure DevOps, prefer injected failure evidence and the build/pipeline UI or REST when authenticated access is already available. For Bitbucket, prefer injected failure evidence and Bitbucket Pipelines UI/API when authenticated access is already available. For Gitea, prefer injected failure evidence and Gitea Actions UI/API when authenticated access is already available. Do not invent new CLI surfaces.</overview>

  <phase name="triage">
    <steps>
      <step>Start from `triggering_run` when present; otherwise use only the latest failed default-branch run for the repository provider. Ignore older history.</step>
      <step>No-op with evidence if a newer same-workflow run already passes, the failure is a one-off flake, or an open Roomote PR already covers it.</step>
    </steps>
  </phase>

  <phase name="fix">
    <steps>
      <step>Reproduce the failing job commands in this environment. If it does not reproduce, no-op with evidence.</step>
      <step>Implement the smallest fix, re-verify, and open a draft PR in this task.</step>
      <step>Resolve any existing investigating manager thread on closeout. Stay silent while working.</step>
    </steps>
  </phase>
</workflow>

<completion_criteria>
<criterion>Only the latest/trigger failure was considered.</criterion>
<criterion>Real failures were fixed and PR'd here, or no-opped with evidence when not fixable.</criterion>
</completion_criteria>