| name | fix-ci |
| description | This skill should be used when the user asks to "fix CI", "fix failed checks", "PR is red", "diagnose CI failure". Also triggers when the user says in Japanese: "CI 直して", "CI が落ちてる", "PR の CI を直して". |
Fix CI (Cursor Skill)
Diagnose and fix CI failures on the current pull request in a single pass. Use the GitHub CLI to gather status and logs, dispatch an explore subagent for read-only investigation, then apply edits and push from the main agent. If the fix does not turn the PR green on the first attempt, hand off to the user with a clear report — do not enter a retry loop.
Persistence Target
Persist intermediate findings and the final summary to the active Cursor plan file:
- Look up the currently active plan via the agent's plan tools (e.g.
read_plan / update_plan).
- If no plan is active, create a new file at
./PLAN.md and use that as the persistence target.
- Append the
## CI Fix Summary section to the plan file once the fix is verified (or when handing off to the user).
Workflow (single pass)
Step 1. Verify GitHub CLI authentication
Run gh auth status. If unauthenticated:
- Surface the error to the user and stop.
- Do not attempt interactive
gh auth login from the agent — Cursor's shell is not guaranteed to be interactive.
- Recommend the user run
gh auth login (or set GH_TOKEN) and re-trigger the skill.
Step 2. Read CI status on the current PR
Run:
gh pr view --json statusCheckRollup --jq '.statusCheckRollup[]'
Parse the output to identify failing checks and capture their name, conclusion, and detailsUrl / run id.
Step 3. Pull failed logs
For each failing check:
gh run view <run-id>
gh run view <run-id> --log-failed
If a run id cannot be derived from the rollup, fall back to gh pr checks or gh run list --branch <branch> and pick the latest failing run.
Capture for each failure:
- The first actionable error message
- The failing command (e.g.
npm test, pytest, golangci-lint, gradle test)
- File paths and line numbers
- Whether the failure looks like code, configuration, or environment / permissions
Step 4. Dispatch an explore subagent for analysis
Dispatch one Task subagent of type explore (read-only). Pass it:
- The captured log excerpts and run ids
- The failing command(s)
- The file paths and line numbers from the logs
- The git diff range (e.g. base..HEAD on this branch) so it can correlate failures with recent changes
Ask the subagent to:
- Locate the failing code and any related tests / configs.
- Cross-reference recent commits on this branch.
- Classify the failure (build / test / lint / type / dep / config / env).
- Recommend the minimal fix.
The subagent must not make edits or run network commands beyond read-only exploration.
Step 5. Propose the fix in plain language
Before editing, write a short paragraph in chat that covers:
- Root cause in plain language
- Intended fix and why it should work
- Risks (behavior change, dependency bump, scope creep)
- Out-of-scope notes (anything observed but deliberately not fixed)
Step 6. Apply edits in the main agent
Apply the fix yourself (do not delegate edits to the subagent). Keep the change minimal:
- Fix only what is required to turn the failing check green.
- Preserve existing style and formatting.
- Do not refactor unrelated code.
Step 7. Commit and push
Run, in order:
git status
git add -A
git commit -m "fix(ci): <short summary of the fix>"
git push
Use a clear message that names the failing check when helpful.
Step 8. Watch CI (single attempt)
Run:
gh pr checks --watch
Wait for the run to complete. Do not start another retry pass — this skill is single-attempt by design.
Step 9. Report and persist
Two outcomes:
- All green: Append the
## CI Fix Summary section (template below) to the active plan file and post a short confirmation in chat.
- Still failing or fix not possible: Append a
## CI Fix Summary section with status NEEDS_HUMAN_INTERVENTION, summarize what was tried and what is blocking, and explicitly hand off to the user.
Output Template
Append this section to the active plan file:
## CI Fix Summary
### Status
[RESOLVED | PARTIALLY_RESOLVED | NEEDS_HUMAN_INTERVENTION]
### Failing Checks (before fix)
- <check name> — <one-line failure>
### Root Causes
- <root cause per failure>
### Fixes Applied
- <file>: <what changed and why>
### Verification
- `gh pr checks --watch` outcome: <pass / fail / partial>
- Local checks run (if any): <commands and results>
### Recommendations / Follow-ups
- <items the user should consider next>
Guidelines
- Single attempt only. Do not loop on failures — hand off instead.
- Auth gate first. Never attempt to fix CI without
gh auth status succeeding.
- Subagent is read-only. All edits, commits, and pushes happen in the main agent.
- Keep edits minimal. The goal is to turn checks green, not to refactor.
- Persist before exiting. The active plan file (or
./PLAN.md) must contain the summary.
- Surface unfixable cases. Permissions, secrets, infra, manual approvals, or sandbox-only blockers must be reported clearly with the user actions required.
