review-skill-improver

审查技能改进器

用途

分析结构化反馈日志，以便：

1. 识别会产生误报的规则（高拒绝率）
2. 识别缺失的规则（本应发现但未发现的问题）
3. 提出具体的技能修改建议

输入

增强版架构格式的反馈日志（参见 review-feedback-schema 技能）。

硬性关卡

按顺序执行；在每个关卡都通过之前，不要输出最终的审查技能改进报告。

1. 记录输入 — 日志从仓库中指定的路径加载，或来自附加的文件，而非凭记忆或转述。**通过条件：**报告标题或摘要注明该路径，或说明“附加的反馈数据块”并注明字节数或行数。
2. 架构 / 结构 — 条目符合增强版架构（rule_source、verdict、rationale 等，详见 review-feedback-schema 技能）。**通过条件：**所有行均可解析，或者对跳过的格式错误行进行计数，并按行索引列出（不得静默丢弃）。
3. 先聚合，再应用阈值 — 完成步骤 1（对完整解析集按 rule_source 统计总数、ACCEPT 与 REJECT 数、拒绝率和拒绝理由），然后才能将任何规则标记为“高拒绝率”或撰写建议。**通过条件：**摘要中的“触发的唯一规则数”与聚合表一致。
4. 建议须有证据支撑 — 每条建议都必须在拟议修复之前包含至少一个具体证据指针（日志行，或文件:行 + 简短引述）。通过条件：每条建议的证据均不为空。

分析流程

步骤 1：按规则来源聚合

For each unique rule_source:
  - Count total issues flagged
  - Count ACCEPT vs REJECT
  - Calculate rejection rate
  - Extract rejection rationales

步骤 2：识别高拒绝率规则

拒绝率 >30% 的规则值得调查：

• 阅读拒绝理由
• 识别常见主题
• 判断规则是否需要细化或增加例外

步骤 3：模式分析

按理由主题对拒绝进行分组：

• “Linter 已经处理了此问题” -> 添加 Linter 验证步骤
• “框架支持此模式” -> 为技能添加例外
• “有意为之的设计决策” -> 添加代码库上下文检查
• “假设了错误的代码路径” -> 添加代码追踪步骤

步骤 4：生成改进建议

针对发现的每个问题，生成：

## Recommendation: [SHORT_TITLE]

**Affected Skill:** `skill-name/SKILL.md` or `skill-name/references/file.md`

**Problem:** [What's causing false positives]

**Evidence:**
- [X] rejections with rationale "[common theme]"
- Example: [file:line] - [issue] - [rationale]

**Proposed Fix:**
```markdown
[Exact text to add/modify in the skill]

**预期影响：**将 [rule] 的误报率从 X% 降至 Y%

## Output Format

```markdown
# Review Skill Improvement Report

## Summary
- Feedback entries analyzed: [N]
- Unique rules triggered: [N]
- High-rejection rules identified: [N]
- Recommendations generated: [N]

## High-Rejection Rules

| Rule Source | Total | Rejected | Rate | Theme |
|-------------|-------|----------|------|-------|
| ... | ... | ... | ... | ... |

## Recommendations

[Numbered list of recommendations in format above]

## Rules Performing Well

[Rules with <10% rejection rate - preserve these]

用法

调用 review-skill-improver 技能来分析反馈并生成改进报告，也可以选择传入输出路径：

review-skill-improver --output improvement-report.md

分析示例

给定以下反馈数据：

rule_source,verdict,rationale
python-code-review:line-length,REJECT,ruff check passes
python-code-review:line-length,REJECT,no E501 violation
python-code-review:line-length,REJECT,linter config allows 120
python-code-review:line-length,ACCEPT,fixed long line
pydantic-ai-common-pitfalls:tool-decorator,REJECT,docs support raw functions
python-code-review:type-safety,ACCEPT,added type annotation
python-code-review:type-safety,ACCEPT,fixed Any usage

分析输出：

# Review Skill Improvement Report

## Summary
- Feedback entries analyzed: 7
- Unique rules triggered: 3
- High-rejection rules identified: 2
- Recommendations generated: 2

## High-Rejection Rules

| Rule Source | Total | Rejected | Rate | Theme |
|-------------|-------|----------|------|-------|
| python-code-review:line-length | 4 | 3 | 75% | linter handles this |
| pydantic-ai-common-pitfalls:tool-decorator | 1 | 1 | 100% | framework supports pattern |

## Recommendations

### 1. Add Linter Verification for Line Length

**Affected Skill:** `commands/review-python.md`

**Problem:** Flagging line length issues that linters confirm don't exist

**Evidence:**
- 3 rejections with rationale "linter passes/handles this"
- Example: amelia/drivers/api/openai.py:102 - Line too long - ruff check passes

**Proposed Fix:**
Add step to run `ruff check` before manual review. If linter passes for line length, do not flag manually.

**Expected Impact:** Reduce false positive rate for line-length from 75% to <10%

### 2. Add Raw Function Tool Registration Exception

**Affected Skill:** `skills/pydantic-ai-common-pitfalls/SKILL.md`

**Problem:** Flagging valid pydantic-ai pattern as error

**Evidence:**
- 1 rejection with rationale "docs support raw functions"

**Proposed Fix:**
Add "Valid Patterns" section documenting that passing functions with RunContext to Agent(tools=[...]) is valid.

**Expected Impact:** Eliminate false positives for this pattern

## Rules Performing Well

| Rule Source | Total | Accepted | Rate |
|-------------|-------|----------|------|
| python-code-review:type-safety | 2 | 2 | 100% |

未来：自动化技能更新

当置信度足够高时，此技能可以：

1. 向 beagle 提交包含技能改进的 PR
2. 跟踪改进效果的变化
3. 对规则变体进行 A/B 测试

反馈循环

Review Code -> Log Outcomes -> Analyze Patterns -> Improve Skills -> Better Reviews
     ^                                                                    |
     +--------------------------------------------------------------------+

这样就能形成持续改进循环，让审查质量基于实证数据而非猜测不断提升。