review-code

Perform a comprehensive review of all uncommitted changes.

Independent review — never self-verify. This review MUST be run by a separate verifier subagent (Agent) that did not write the code. The implementing agent does not review its own work. Spawn a judge with the diff + what to check; it returns only findings. (See the "Never self-verify" rule in run.md.) Apply the verify model preference from .claude/loobster.json when spawning the judge (see run.md "Preferred subagents"); no config → inherit the session model.

What this does

• Gathers and reviews all uncommitted changes in the current branch
• Produces a prioritized list of action items with file:line references
• Optionally posts findings to the relevant issue tracker

Steps

1. Collect change context:
   • git status --porcelain
   • git diff
   • git diff --cached
   • git diff HEAD
   • git log --oneline -n 5
2. If an issue ID is provided or can be inferred from the branch name, read the issue to understand the intent.
3. Analyze changes for: security, performance, style, consistency, missing edge cases, dependency impacts, and integration risks.
   • Token discipline (see .agents/skills/token-discipline/SKILL.md): for a wide diff, delegate per-area review to Agent subagents (e.g. one per dimension or per large file). Each subagent receives its slice of the diff plus what to check, and returns only a findings list (each: severity + file:line + one-line fix) — never the raw hunks. The main thread merges those findings into the single prioritized action list below without re-reading the hunks the subagents already saw.
4. If the diff touches frontend (.tsx/.jsx/.vue/.svelte/.html/.css/.scss or component/route dirs), run the verifiable frontend layer — follow .agents/skills/verify-frontend/SKILL.md to capture Playwright screenshots and attach them to the PR. A render failure (4xx/5xx or console errors) is a must-fix. (Independent: CI or a separate verifier subagent assesses — not the implementer.)
5. Output a summary and a single prioritized action list using indicators:
   • RED — must-fix
   • YELLOW — recommended
   • GREEN — consider
6. If Linear MCP tools are available and an issue was identified, post the review as a comment on the issue.
7. If there are no action items, suggest the issue is ready for the next phase.
8. If Claude Code Tasks track this work (TaskList), reflect the outcome: leave any task with open action items as in_progress, and only mark tasks completed once their action items are resolved. This keeps /resume accurate.

Output template

## Code Review

Summary: <1-2 sentences>

Action Items:
1. <indicator> <action> in `path:line`
2. <indicator> <action> in `path:start-end`