| name | squid-review-ci |
| description | Drive CI green on a pushed, review-clean feature PR — On-Call diagnoses failures and hands fix tasks to the SWE. Output: a CI-validated feature PR. Trigger after /squid-review passes, or when the user says "/squid-review-ci". |
| disable-model-invocation | false |
| argument-hint | <PR-number | branch> |
Review CI — validate the pushed PR's pipeline
Watch CI for the pushed feature PR and drive it to green. On failure, the On-Call engineer diagnoses and hands a fix task to the SWE (it does not fix the code itself); once CI is green, the PR is ready to merge.
You are the orchestrator — a MANAGER. You launch On-Call and the SWE, enforce the cap, and confirm green. You do NOT read the diff for review (that was /squid-review).
Read AGENTS.md first.
Input: a pushed, review-clean feature PR.
Output: a CI-validated "ready to merge" PR (or USER ACTION REQUIRED if CI can't be made green within the cap).
The On-Call → SWE → re-check loop (max 5)
Watch CI (On-Call)
Agent(
subagent_type="squid:oncall-engineer",
prompt="""Watch CI for the latest push on feat/{slug} (PR #{N}). {Working directory: {path}.} Read AGENTS.md first. Follow your role definition. CI/CD only — do not review the diff. If green, report success. If red, pull the failing logs (`gh run view --log-failed`), identify the failing job/step, trace it to the responsible task from the commit messages, and hand back a concrete FIX TASK (root cause + the exact failing command + affected files). Do NOT fix the code yourself — that's the SWE's job."""
)
- Green → done. The PR is validated and ready to merge.
- Red (infra / flake) → On-Call files a NEW infra task and reports it; surface to the human — do not burn fix attempts on infra.
- Red (real failure) → On-Call hands back a fix task → go to "SWE fixes".
SWE fixes
Agent(
subagent_type="squid:software-engineer",
prompt="""CI failed on PR #{N}. {Working directory: {path}.} Fix task from On-Call: {root cause + failing command + files}. Reproduce locally with the same command CI ran, fix the root cause (not the symptom; fix the bug, not the test), re-run the local suite, commit with `Refs #N` (the original task is already closed) via the commit-message generator (`/caveman-commit` if the caveman plugin is installed, else the `commit-commands` plugin), and push."""
)
Then re-run Watch CI (On-Call) to confirm the new run goes green.
Cap: 5 fix cycles per PR
If CI is still red after 5 On-Call→SWE→re-check cycles, stop, surface USER ACTION REQUIRED with the last failing log, and end.
Output
A CI-validated, ready-to-merge feature PR. Hand back to /squid-implement-night (which runs the squid-self-improve gate + final hand-off), or — if run standalone — report that CI is green and the PR is ready.
Notes
- On-Call diagnoses; the SWE fixes. On-Call owns pipeline health and traces failures to their root cause; the actual code fix is a task for the SWE (the team's role split).
- Self-improve is NOT here. It runs in
/squid-implement-night's tail after this skill returns. Running /squid-review-ci standalone just validates CI.
- Infra failures don't count against the 5-cycle cap — file a separate infra task and surface it.
