chore: 添加 /review skill，多 agent 并行代码审查

[lishuceo]

Summary

• 新增 .claude/skills/review/SKILL.md，输入 /review 启动多 agent 并行代码审查
• 三个审查维度并行：安全 (🔒)、逻辑正确性 (🐛)、架构质量 (🏗️)
• 无参数审查本地未提交变更或最近 commit，传 PR 编号/URL 审查 PR
• 置信度 ≥75 才报告，内置误报过滤规则，参考项目 GitHub Action 标准

Test plan

• /review 审查本地变更
• /review <PR号> 审查指定 PR

🤖 Generated with Claude Code

[claude]

Claude finished @lishuceo's task in 2m 14s —— View job

---

Code Review: .claude/skills/review/SKILL.md

• Gather context and read changed files
• Review SKILL.md for issues
• Post review summary

---

Verdict: ⚠️ Issues Found

Findings

🟡 Warning

• .claude/skills/review/SKILL.md:35 — TeamCreate is not a real Claude Code tool or feature. The skill instructs Claude to "使用 TeamCreate 创建审查团队", but no such mechanism exists. The correct approach is to use the Task tool with subagent_type — when multiple Task calls are issued in a single message, they execute in parallel automatically. As written, this line will either be ignored or confuse the model during execution. (Confidence: 90) — See inline comment

Overall Assessment

The skill is well-designed in concept — the three review dimensions (security, logic, architecture) are practical, the confidence scoring system with a ≥50 reporting threshold is sound, and the false positive filters are a good addition. The frontmatter fields (disable-model-invocation, argument-hint) are all valid.

The one actionable issue is the TeamCreate reference on line 35, which should be replaced with a Task tool-based approach to ensure the parallel agent spawning actually works as intended.

[claude]

🟡 Warning (Confidence: 90)

TeamCreate is not a real Claude Code tool or feature. This reference will either be ignored or cause confusion during execution.

The correct mechanism for spawning parallel agents from within a skill is the Task tool with subagent_type. When multiple Task tool calls are made in a single message, they run in parallel automatically.

Suggested change

	使用 TeamCreate 创建审查团队，启动 **3 个并行 agent**，每个专注一个维度。
	使用 Task 工具启动 **3 个并行 agent**（在同一条消息中发起多个 Task 调用即可并行），每个专注一个维度。

[claude]

Claude finished @lishuceo's task in 1m 9s —— View job

---

Code Review: .claude/skills/review/SKILL.md

• Gather context and read changed files
• Review SKILL.md for issues
• Post review summary

---

Verdict: ⚠️ Issues Found

Findings

🟡 Warning

• .claude/skills/review/SKILL.md:44 — TeamCreate is not a real Claude Code tool or feature. This was flagged in the previous review and remains unfixed after commit f967fc6. The correct approach is to use the Task tool with subagent_type — multiple Task calls in a single message execute in parallel automatically. (Confidence: 90) — See inline comment

Changes Since Last Review

Commit f967fc6 ("fix: 完善 review skill 边界处理和参数校验") addressed several edge cases well:

• Added untracked file discovery via git ls-files --others --exclude-standard
• Added commit count check with git rev-list --count HEAD for repos with zero or one commit
• Added parameter validation for $ARGUMENTS with clear error messaging and usage examples
• Removed disable-model-invocation: true from frontmatter (correct, since the skill needs model invocation)

These are solid improvements. The only remaining issue is the TeamCreate reference on line 44.

[claude]

🟡 Warning (Confidence: 90)

TeamCreate is still not a real Claude Code tool. This was flagged in the previous review and remains unfixed after the latest commit.

The correct mechanism for parallel agent execution is the Task tool with subagent_type. When multiple Task tool calls are made in a single response message, they run in parallel automatically.

Suggested change

	使用 TeamCreate 创建审查团队，启动 **3 个并行 agent**，每个专注一个维度。
	使用 Task 工具启动 **3 个并行 agent**（在同一条消息中发起多个 Task 调用即可并行），每个专注一个维度。