Code review

ClosedLoop.ai's /code-review:start runs multi-agent, partitioned code review with deterministic hygiene, risk-based routing, and caching. It works on the current branch, on staged changes, on specific files, or on a GitHub PR.

Local branch review

claude /code-review:start

Reviews the current branch against main. Results land in .closedloop-ai/code-review/cr-<random>/.

Staged review

claude /code-review:start staged

Reviews only what is currently staged. Useful before committing.

Specific files

claude /code-review:start path/to/file1.ts path/to/file2.ts

PR review

claude /code-review:start 123

Reviews PR #123. With --github, posts inline comments and a summary comment via gh and resolves outdated threads.

Flags

• --github — CI handoff mode; posts to the PR.
• --hygiene-only — skip LLM review, run only deterministic checks.
• --base <ref> — override the base ref.
• --since-last-review — incremental review since the last cached review hash.
• --full-review — ignore the incremental cache.

What happens

1. Hygiene — deterministic checks run first (file size, trailing whitespace, TODO density, etc.).
2. Partitioning — the diff is split into partitions.
3. Routing — each partition is routed to a model based on risk signals.
4. Parallel review — code-review-worker subagents run concurrently.
5. Collection — findings aggregate into code-review-findings.json.
6. Validation — findings pass a Pydantic schema check.
7. Summary — code-review-summary.md is written.
8. Verdict — verdict.json reports approve, needs_attention, or changes_requested.

GitHub integration

With --github:

• Inline comments post at the correct line and file.
• Outdated review threads are resolved automatically.
• A summary comment lands on the PR.
• Verdict is surfaced so CI can gate merges.

Caching

code-review caches results by diff hash. Re-running on the same diff is near-instant. --since-last-review supports incremental review on long-lived branches.

CI pattern

# github/workflows/code-review.yml (sketch)
- run: claude /code-review:start ${{ github.event.pull_request.number }} --github

Couple this with ClosedLoop.ai's judges reports on the plan and code, and you have a full grading pipeline for every PR.

When to run

• Locally, before committing (staged or default branch review).
• In CI, on every PR (<pr-number> --github).
• Incrementally on large branches, to review only the new delta.
• In parallel with the code plugin's built-in code-reviewer agent — they complement each other; the built-in reviewer gates implementation phases, while /code-review:start is for PR-level triage.