AgentSkillsCN

branch-review

仅对已验证的问题进行当前分支的代码审查。

SKILL.md
--- frontmatter
name: branch-review
description: Code review the current branch with verified issues only
argument-hint: "[base-branch]"
allowed-tools: Bash(git *), Bash(clip.exe *)

Code Review — Verified Issues Only

Review the current branch against $ARGUMENTS (default: main).

Process

Phase 1: Gather context

  1. •git log --oneline <base>..HEAD — list commits
  2. •git diff <base>..HEAD --stat — file summary
  3. •git diff <base>..HEAD — full diff

Phase 2: Identify candidate issues

Scan the diff for potential problems:

  • •Bugs, logic errors, race conditions
  • •Duplicated patterns that should be extracted
  • •Missing error handling at system boundaries
  • •Type safety holes (unsafe casts, as assertions)
  • •Stale dependencies in hooks/effects
  • •Security concerns (injection, XSS, secrets)
  • •API design issues (confusing interfaces, leaky abstractions)
  • •Performance implications (unnecessary re-renders, N+1 queries, missing memoization)
  • •Test coverage gaps (new code without tests, existing tests invalidated by changes)
  • •Project convention violations (naming, patterns, style inconsistent with surrounding code)

For each candidate, write a one-line summary and note which files/lines are involved.

Phase 3: Verify candidates (parallel)

Launch verification agents in parallel using the Task tool with subagent_type: "Explore".

Grouping strategy:

  • •Group candidates that share the same file(s) into a single agent
  • •Each agent handles one group of related candidates
  • •If there are 3 or fewer total candidates, verify them all in a single agent instead of parallelizing

Each agent prompt must include:

  1. •The candidate issue(s) to verify — summary, file paths, and line numbers
  2. •The relevant section of the diff for context
  3. •The verification checklist:
    • •Can the problematic state actually be reached? Trace callers and data flow.
    • •Does the UI or type system prevent the scenario? Check component props, select options, type constraints.
    • •Is there existing handling elsewhere that covers this case?
    • •Is the "missing" code actually unnecessary given the guarantees of the framework or surrounding code?
  4. •Instructions to return a verdict for each candidate: confirmed or false positive, with a one-line explanation

Launch all agents in a single message so they run concurrently. Collect all results before proceeding.

Phase 4: Report

Output a concise list of confirmed issues only. For each:

  • •One-line summary
  • •File and line reference
  • •Why it's a real problem (what you verified)

At the end, note how many candidates were dismissed as false positives (no need to list them individually unless the user asks).

Phase 5: Clipboard

Copy the confirmed issues list to clipboard using clip.exe. Keep it terse — one line per issue, no markdown formatting in the clipboard version.