| name | multi-review |
| description | Review the current diff — first a cheap deterministic SonarQube-in-Docker pre-pass (fix + re-scan until clean), then fan out to three single-focused reviewer subagents in parallel, auto-apply the safe quality cleanups, and tell the human which areas to review first. Explicit invocation only — user types /multi-review. |
| disable-model-invocation | true |
/multi-review — review, fix, hand off
Run a quality review of the current uncommitted changes, apply the low-risk
fixes yourself, and end by pointing the human at what still needs their eyes.
Steps
-
Confirm there's something to review. Run git --no-pager diff --stat HEAD.
If empty, say "Nothing to review." and stop.
-
Cheap deterministic pre-pass FIRST — exhaust SonarQube before spending tokens on the LLM reviewers.
Run .claude/skills/multi-review/sonar-diff-review.sh (boots a reused SonarQube
container in Docker and scans only the changed backend Java files). Then loop on
its exit code:
- exit 0 (Sonar clean, or nothing for it to scan) → the cheap pre-pass is done;
go to step 3 (the LLM reviewers).
- exit 2 (Sonar found issues) → read its printed findings and
petclinic-backend/target/sonar-diff-review.json. Fix the clearly-safe /
mechanical ones (same auto-fix-vs-defer triage as step 4); leave genuine
judgment calls for the human. Then re-run the script to confirm. Repeat this
fix→re-scan loop until Sonar reports exit 0 (or only human-deferred issues remain
— cap at 3 iterations, then move on and list the rest for the human).
- exit 1 (Sonar could not run — e.g. Docker down / disk full) → note it briefly
and continue to step 3 anyway; never block the review on Sonar infrastructure.
Do not fix anything beyond what Sonar flagged during this pre-pass.
-
Fan out in parallel. In ONE message, make THREE Agent calls at once (so they
run concurrently — never one after another): reviewer-reuse,
reviewer-simplification, reviewer-efficiency. Give each the same task:
"Review the current working-tree change (git --no-pager diff HEAD) for your single
dimension. Return findings as a short list of file:line — issue → suggestion. If clean, say so."
Collect the three results into one set of findings grouped by dimension
(Reuse / Simplification / Efficiency), dropping duplicates.
-
Triage each finding:
- Auto-fix — mechanical, low-risk, clearly correct (dead code, obvious
duplication, a needless allocation, a trivially simpler form). Apply it.
- Defer to human — anything ambiguous, behavior-changing, or a judgment
call. Leave it untouched.
This is quality only — do not chase correctness bugs here; if you happen to
spot one, just note it for the human.
-
Apply the safe fixes and keep a short list of what you changed (file + what).
-
Hand off (the important part). End your response with this section, ordered
by risk — widest blast radius and deferred findings first:
## Review these first (human)
- <file:area> — <why it needs human judgment: deferred finding / risky change / high blast radius>
Be specific about files and why. If nothing needs human review, say so explicitly.
Output shape
- One-line summary:
Sonar: S issues fixed over R rounds · N LLM findings · auto-fixed X · deferred Y.
- What you auto-fixed (bullets with paths), separating the Sonar pre-pass fixes from the LLM-review fixes.
- The Review these first (human) section.
Do not commit or push — leave that to the human.
