deep-review

name	deep-review
description	5-agent deep code review of TypeScript/TSX source code changes on the current branch. Only for reviewing code, NOT for reviewing plans, docs, or markdown files.

Deep Code Review Skill

Run a structured, multi-agent code review of all code changes on the current branch.

Scope Guard

This skill reviews .ts and .tsx source files only. It does NOT review:

• Plan files (.md in .claude/plans/ or docs/plans/)
• Documentation (.md, CLAUDE.md, SKILL.md)
• Config files (.json, .toml, .yaml)
• Skill definitions

If invoked and there are no .ts/.tsx changes, say "No code changes to review" and stop.

Steps

1. Identify scope: Run git diff main...HEAD --name-only to list all changed files. If no diff (on main), use git diff HEAD~1 --name-only for the last commit. Filter to only .ts and .tsx files. Exclude test files (*.test.ts) from the architectural review but include them in the correctness review. If no .ts/.tsx files remain after filtering, stop — nothing to review. Read all code files to build context.

2. Launch 5 review agents in parallel (single message, 5 tool calls):

   Agent 1 — Code Architect (subagent_type: feature-dev:code-architect): Review all changed files for architectural concerns: CLAUDE.md convention violations, Zustand selector usage, shared helper reuse, file organization rules, dollar basis consistency, no hardcoded SG values, no store-to-store imports, component structure, and separation of concerns.

   Agent 2 — Code Reviewer (subagent_type: feature-dev:code-reviewer): Review all changed .ts/.tsx files for correctness: TypeScript type errors, any usage, incorrect type narrowing, missing null checks, logic errors in calculations, off-by-one errors, missing test coverage for changed calculation/simulation code, and security issues.

   Agent 3 — Plan Compliance (subagent_type: superpowers:code-reviewer): Review the implementation against the original plan (if one exists in .claude/plans/ or docs/plans/) and project coding standards in CLAUDE.md. Check: does the code match what the plan specified? Are there deviations, missing steps, or extras that weren't planned? If no plan file exists, focus on CLAUDE.md coding standards only.

   Agent 4 — Codex MCP (use mcp__codex-cli__codex tool): Send the list of changed files and ask codex to review for bugs, logic errors, and improvements. Prompt: "Review these files for bugs, logic errors, type safety issues, and convention violations: [file list]. Read each file and report findings."

   Agent 5 — Gemini (use mcp__gemini-cli__ask-gemini tool, model: gemini-3-flash-preview): Send the list of changed files using @ syntax for file inclusion. Prompt: "Review these changed files for bugs, logic errors, missed edge cases, and potential regressions. Focus on issues the other reviewers might miss — subtle logic flaws, implicit assumptions, and cross-file interaction bugs: @file1.ts @file2.ts ..."

3. Consolidate findings: Collect all 5 agent results, deduplicate, and present as a severity-ranked list:

   • CRITICAL: Bugs, type errors, security issues, calculation errors
   • WARNING: Convention violations, missing tests, architectural concerns
   • INFO: Style suggestions, minor improvements

4. Ask user which findings to fix. Then apply fixes and run cd frontend && npm run type-check && npm run test to verify.