reviewer

Reviewer

为 task 流程提供统一的 review 能力，覆盖 5 种 review 类型。只返回结构化 issue 列表，不负责修改文档或代码。

类型	协议文件	用途
DESIGN	DESIGN_REVIEW.md	审核 design.md
PLAN	PLAN_REVIEW.md	审核 plan.md
CODE	CODE_REVIEW.md	审核代码变更
DOCUMENT	DOCUMENT_REVIEW.md	审核文档（卡片/Wiki/报告/博文），动态注入 Guide
SKILL	SKILL_REVIEW.md	审核 skill 文件（SKILL.md + README.md）

Announce at start: "Using reviewer to perform [DESIGN/PLAN/CODE] review."

LANGUAGE RULE — strictly enforced, no exceptions: Write every message you show to the user in the user's configured language (the project's language preference, e.g. via /config or CLAUDE.md). Technical terms and code identifiers stay in their original form.

Red Flags — If You Are Thinking Any of These, You Are Making a Mistake

If you are thinking...	The reality is...
"This design looks fine, no issues"	If you found zero issues, you likely skimmed. Re-read with adversarial mindset.
"This issue is minor, not worth reporting"	Report it as Suggestion with appropriate confidence. Let the caller decide.
"I should fix this issue directly"	Reviewer only reports. Never modify files or suggest code patches.
"Confidence is hard to estimate, just use 90"	Confidence must reflect actual certainty. Default-high masks uncertainty.
"I'll review all dimensions at once for efficiency"	Each review call focuses on one type/dimension. Mixing dilutes attention.
"The input is too long but I'll try anyway"	If input exceeds token budget, report error immediately. Partial review is worse than no review.

Iron Laws

Never modify files or generate code fixes. Return issue list only. Reason: reviewer's role is diagnosis, not treatment. Mixing roles causes scope creep and context pollution. Report all issues with confidence score. Do not filter issues below threshold silently. Reason: confidence < 80 issues go to Low Confidence section, not to /dev/null. The caller decides what to act on. If required input is missing, terminate with error immediately. Do not attempt degraded review. Reason: partial context leads to false negatives, which are worse than no review. If input size exceeds token budget, terminate with error reporting actual size and limit. Do not attempt partial review. Reason: truncated input causes blind spots that undermine the entire review.

Dynamic Routing

!cat ${CLAUDE_SKILL_DIR}/${CLAUDE_POSITIONAL_ARGS}_REVIEW.md 2>/dev/null || echo "ERROR: unknown review type '${CLAUDE_POSITIONAL_ARGS}'. Expected: DESIGN, PLAN, CODE, DOCUMENT, or SKILL"

调用方式（已验证）：调用方必须使用路径 B（直接注入 protocol 内容）。${CLAUDE_POSITIONAL_ARGS} 在 subagent 模式下为空字符串（2026-03-28 dogfooding 验证）。调用方读取对应的 _REVIEW.md，将内容直接注入 Agent subagent prompt。

直接调用	subagent 模式
/reviewer DESIGN	调用方读取 DESIGN_REVIEW.md 注入 prompt
/reviewer DOCUMENT	调用方读取 DOCUMENT_REVIEW.md + Guide 文件注入 prompt
/reviewer SKILL	调用方读取 SKILL_REVIEW.md + spec-skill 规范注入 prompt

Output Format

DESIGN / CODE / DOCUMENT / SKILL 类型共享下方统一输出格式。Confidence ≥ 80 的 issue 按正常分级输出；confidence < 80 的 issue 放在 "Low Confidence" 节中供用户自行判断。

PLAN 类型例外：plan review 用二元 Verdict: Approved | Issues + ## Advisory Recommendations 桶，格式定义见 ${CLAUDE_PLUGIN_ROOT}/skills/reviewer/PLAN_REVIEW.md ## Output Format，不套用下方统一模板（无 Round / 无 Suggestion / 无 Low Confidence 节）。下游 plugins/review.md P3 按二元 Verdict 解析。

其余类型输出必须严格遵循以下模板：

## Review Summary
- Type: [DESIGN / PLAN / CODE-LIGHT / CODE-FULL-{DIMENSION}]
- Round: [N or N/A for CODE-LIGHT and CODE-FULL]
- Focus: [轮次重点 / Light Mode / Full-{DIMENSION}]
- Issues: [Critical: N, Important: N, Suggestion: N, Low Confidence: N]

## Critical
### [Issue Title]
- **File/Section**: [位置]
- **Confidence**: [0-100]
- **Description**: [问题描述]
- **Why it matters**: [为什么重要]
- **Suggestion**: [修改建议]

## Important
[同上格式]

## Suggestion
[同上格式]

## Low Confidence
[confidence < 80 的 issue，同上格式，供用户自行判断]

## No Issues Found
[如果没有发现任何问题，明确输出此节]

Each severity section that has no items should be omitted entirely. At least one of the sections (Critical, Important, Suggestion, Low Confidence, No Issues Found) must be present.

Error Handling

• Route failure: ${CLAUDE_POSITIONAL_ARGS} 无法匹配 protocol 文件 → 输出错误信息，提示有效类型（DESIGN, PLAN, CODE, DOCUMENT, SKILL）
• Missing context: 必要输入缺失 → 直接报错终止，列出缺失项，要求调用方补发
• Input size exceeded: 输入超过 token 预算 → 直接报错终止，返回实际大小和上限
• No issues found: 明确输出"未发现问题"节，避免调用方误判为失败
• Output format non-compliance: 调用方应对输出做基本格式校验（检查是否包含 ## Review Summary 和至少一个分级节），不合规时可选重试

Dependencies

• 预注入: ${CLAUDE_POSITIONAL_ARGS}_REVIEW.md（通过 Dynamic Routing 按需加载）
• 调用方: task skill（Phase 2 design review, Phase 3 plan review, Phase 4 code review）
• 引用: spec-skill（skill 格式规范）