nemoclaw-maintainer-cross-issue-sweep

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Cross-Issue Regression Sweep

跨Issue回归扫描

Surfaces the issues a single PR may also fix or accidentally break beyond the one it claims to address. Two outputs:

Adjacent fixes — "PR may also close #X" → bundling intel (ship one PR, close multiple issues)
Contradicting risks — "PR may break what #Y wants" → coordination needed before merge

揭示单个PR除了声称要解决的问题外，还可能修复或意外破坏的其他问题。输出分为两类：

相邻修复 — "PR可能同时关闭#X" → 合并情报（提交一个PR，关闭多个Issue）
冲突风险 — "PR可能破坏#Y想要实现的功能" → 合并前需要协调

Prerequisites

前提条件

```
gh
```
CLI authenticated
A target repository with open issues
An open PR to scan

已认证的
```
gh
```
CLI
包含未关闭Issue的目标仓库
待扫描的开放PR

Repo policy

仓库策略

Defaults assume NemoClaw conventions. Edit

repo-policy.md

to override per-repo (bot logins, candidate caps, language regex).

默认遵循NemoClaw约定。编辑

repo-policy.md

可覆盖仓库专属配置（机器人登录信息、候选数量上限、语言正则表达式）。

Workflow

工作流程

Copy this checklist into your response and check off each step:

text

Cross-issue sweep progress:
- [ ] Step 1: Extract fingerprint (files, symbols, error strings, primary issue)
- [ ] Step 2: Search candidate issues (capped at 30, primary excluded)
- [ ] Step 3: Classify each candidate (4-class with evidence)
- [ ] Step 4: Apply reverse-link boost
- [ ] Step 5: Filter (drop UNRELATED, SAME_ISSUE_DIFF, low-confidence)
- [ ] Step 6: Render report using templates/report.md

将以下检查清单复制到回复中，并勾选每个步骤：

text

跨Issue扫描进度:
- [ ] 步骤1：提取特征指纹（文件、符号、错误字符串、主Issue）
- [ ] 步骤2：搜索候选Issue（上限30个，排除主Issue）
- [ ] 步骤3：分类每个候选Issue（4类并附带证据）
- [ ] 步骤4：反向链接加权
- [ ] 步骤5：过滤（移除无关、同根问题、低置信度项）
- [ ] 步骤6：使用templates/report.md生成报告

Step 1: Extract fingerprint

步骤1：提取特征指纹

bash

scripts/extract-fingerprint.sh <pr-number>

Pulls four dimensions: touched files, touched symbols (per-language regex), error-string tokens, and the PR's primary linked issue (for exclusion). See

checks/fingerprint-extraction.md

bash

scripts/extract-fingerprint.sh <pr-number>

提取四个维度：修改的文件、修改的符号（基于各语言正则表达式）、错误字符串标记，以及PR关联的主要Issue（用于排除）。详见

checks/fingerprint-extraction.md

。

Step 2: Search candidate issues

步骤2：搜索候选Issue

bash

scripts/search-candidate-issues.sh <fingerprint-json>

Three search dimensions, capped at 30 total candidates:

Per symbol: top 10 by recency
Per file path: top 5 by recency
Per error string: top 5 by recency

Dedupes; excludes the PR's primary linked issue.

bash

scripts/search-candidate-issues.sh <fingerprint-json>

三个搜索维度，候选总数上限为30个：

按符号：最新的10个
按文件路径：最新的5个
按错误字符串：最新的5个

自动去重；排除PR关联的主要Issue。

Step 3: Classify each candidate

步骤3：分类每个候选Issue

For each candidate, the LLM classifies as one of four classes per

checks/relationship-judgment.md

ADJACENT_FIX — PR's changes likely also resolve this issue
CONTRADICTING — PR's approach blocks what this issue wants
SAME_ISSUE_DIFF — same root bug as PR's primary issue (dedup filter)
UNRELATED — no meaningful relationship

Required for ADJACENT_FIX or CONTRADICTING:

Cite specific PR diff line
Cite specific issue symptom
Confidence: high / medium / low

If no specific evidence can be cited, the LLM must answer UNRELATED. This floors hallucination.

根据

checks/relationship-judgment.md

，LLM将每个候选Issue分为以下四类之一：

ADJACENT_FIX — PR的修改可能同时解决该Issue
CONTRADICTING — PR的实现方式会阻碍该Issue的需求
SAME_ISSUE_DIFF — 与PR主Issue为同一根因bug（去重过滤项）
UNRELATED — 无实质性关联

标记为ADJACENT_FIX或CONTRADICTING时需满足：

引用PR差异的具体行号
引用Issue的具体症状
置信度：高/中/低

若无法引用具体证据，LLM必须标记为UNRELATED，以此减少幻觉。

Step 4: Reverse-link boost

步骤4：反向链接加权

If the candidate issue's body or comments already mention this PR's number, the relationship is already in someone's mental model. Boost confidence by one tier (low → medium, medium → high).

若候选Issue的正文或评论中已提及该PR编号，说明该关联已被关注。将置信度提升一级（低→中，中→高）。

Step 5: Filter

步骤5：过滤

Suppress UNRELATED + SAME_ISSUE_DIFF
Drop low-confidence judgments
Keep ADJACENT_FIX and CONTRADICTING with high or medium confidence

隐藏UNRELATED和SAME_ISSUE_DIFF项
移除低置信度判断结果
保留高或中置信度的ADJACENT_FIX和CONTRADICTING项

Step 6: Render report

步骤6：生成报告

bash

scripts/render-report.py < classifications.json

See

templates/report.md

for the format.

bash

scripts/render-report.py < classifications.json

格式详见

templates/report.md

。

Reference files

参考文件

```
repo-policy.md
```
— configurable per-repo defaults
```
relationship-rules.md
```
— 4-class definitions with worked examples
```
checks/fingerprint-extraction.md
```
— what to pull from the diff, per language
```
checks/relationship-judgment.md
```
— LLM judgment criteria + evidence requirement
```
templates/report.md
```
— output template
```
validation/backtest.md
```
— backtest the skill against historical PRs

```
repo-policy.md
```
— 可配置的仓库默认设置
```
relationship-rules.md
```
— 4类关系的定义及示例
```
checks/fingerprint-extraction.md
```
— 基于语言从差异中提取的内容说明
```
checks/relationship-judgment.md
```
— LLM判断标准及证据要求
```
templates/report.md
```
— 输出模板
```
validation/backtest.md
```
— 基于历史PR对该工具进行回测

Scripts (execute, do not read)

脚本（直接执行，无需阅读）

```
scripts/extract-fingerprint.sh
```
— symbols + paths + error strings, deterministic
```
scripts/search-candidate-issues.sh
```
— GitHub Search wrapper, dedupe, cap
```
scripts/render-report.py
```
— report renderer

```
scripts/extract-fingerprint.sh
```
— 符号+路径+错误字符串提取，结果可复现
```
scripts/search-candidate-issues.sh
```
— GitHub Search封装脚本，支持去重和数量上限
```
scripts/render-report.py
```
— 报告生成器

Composition with other skills

与其他工具组合使用

The pr-comparator (

nemoclaw-maintainer-pr-comparator

) calls this skill as a sub-step when comparing competing PRs. Adjacent-fix counts feed Tier 3 tiebreakers; contradicting hits factor into Tier 2 quality scoring.

pr-comparator（

nemoclaw-maintainer-pr-comparator

）在对比竞争PR时会调用本工具作为子步骤。相邻修复数量用于三级平局判定；冲突风险会纳入二级质量评分。

What this skill does NOT do (deferred)

本工具不支持的功能（待开发）

These would raise the ceiling but require infrastructure beyond GitHub API + LLM:

Run PR code against adversarial inputs (sandboxed)
Static-analyzer dataflow tracing (CodeQL, Semgrep)
ML-based symbol disambiguation across codebases

以下功能可提升上限，但需要GitHub API + LLM之外的基础设施支持：

在沙箱环境中运行PR代码对抗测试输入
静态分析器数据流追踪（CodeQL、Semgrep）
基于ML的跨代码库符号消歧义