| name | review |
| description | Six-dimension code review of the current branch's diff against main. Delegates to the code-reviewer agent for the actual review and reports findings. |
| allowed-tools | ["Read","Grep","Glob","Bash(git diff *)","Bash(git log *)","Bash(git show *)","Bash(uv run poe check)"] |
/review — Code Review
PURPOSE
Full review of branch changes for correctness, security, design, and Katana conventions.
CRITICAL
- Read-only review — propose fixes, don't apply them. Use
code-modernizer agent to actively rewrite.
- Trust the test suite — re-run
uv run poe check only to confirm green; don't second-guess linters.
STANDARD PATH
1. Classify the change
- Type: feature / bug fix / refactor / spec change / config / docs
- Risk: low (cosmetic) / medium (new feature) / high (spec, transport, auth, breaking)
- Affected packages: client / mcp / typescript / cross-cutting
2. Delegate to the code-reviewer agent
The code-reviewer agent (.claude/agents/code-reviewer.md) covers:
- Six review dimensions (Correctness, Design, Readability, Performance, Testing, Security)
- Project-specific checklist (generated files, transport-layer resilience, UNSET, response helpers, list
data envelope, spec/pydantic regen, help resource)
- Severity-tagged output (BLOCKING / SUGGESTION / NITPICK)
Spawn it via the Agent tool with subagent_type: code-reviewer.
3. Verify
uv run poe check
If it doesn't pass, surface the failures as [BLOCKING] issues.
4. Self-improvement
If the review reveals a pattern worth codifying (new anti-pattern, missing convention, pitfall), update CLAUDE.md so future work benefits.
EDGE CASES
- Caller impact — for every changed function signature, run
LSP findReferences (or git grep) and read each caller. Diff-only review misses ripple effects.
- Out-of-scope findings — recommend filing a GitHub issue; never let deferred work go untracked.
RELATED
code-reviewer agent — does the actual review
verifier agent — mechanical gate (passes/fails, no design opinions)
pr-preparer agent — process readiness (coverage, ADRs, help resource)
/open-pr skill — runs review automatically before opening PR