A read-only, bounded code-review skill for recent diffs: find regressions and interface mistakes without turning every change into a repository-wide audit.
面向近期代码变更的只读、有界代码审查 Skill：优先发现回归和接口问题，不把每个 diff 扩大成全仓库审计。
Most review agents either inspect too little to find a real regression or expand every change into an expensive repository-wide investigation. Lightweight Diff Review stays between those extremes:
很多审查 Agent 要么只看表面，找不到真实回归；要么把每次修改都扩大成昂贵的全仓库调查。Lightweight Diff Review 保持在两者之间：
-
start from the changed diff, not the whole repository;
-
expand context only to answer a concrete unresolved question;
-
check correctness, interface consistency, relevant security, async/performance risks, and test coverage in risk order;
-
run at most one targeted validation by default;
-
report actionable findings first, with location, impact, recommendation, and coverage;
-
remain read-only unless the user explicitly requests a separate implementation task.
-
从变更 diff 开始，而不是扫描整个仓库；
-
只有为回答明确问题时才扩展上下文；
-
按风险顺序检查正确性、接口一致性、相关安全问题、异步/性能风险和测试覆盖；
-
默认最多运行一次定向验证；
-
优先输出带位置、影响、建议和覆盖范围的可执行发现；
-
默认保持只读，除非用户明确提出独立的修复任务。
This is a focused recent-change review skill, not a repository-wide audit, security red team, test-generation tool, or automatic fixer.
它专注于近期变更审查，不是全仓库架构审计、安全红队、测试生成器或自动修复器。
Inspect before installing:
安装前先检查：
hermes skills inspect linnnn89/lightweight-diff-review/skills/lightweight-diff-reviewInstall directly:
直接安装：
hermes skills install linnnn89/lightweight-diff-review/skills/lightweight-diff-reviewOr add the repository as a tap:
也可以将仓库添加为 tap：
hermes skills tap add linnnn89/lightweight-diff-review
hermes skills install linnnn89/lightweight-diff-review/lightweight-diff-review/lightweight-diff-review Review my recent changes.
/lightweight-diff-review 审核我刚刚修改的代码，只检查功能回归和接口一致性，不修改代码。
The more precisely you name the change range and risk, the less context the reviewer needs to load.
越明确地说明变更范围和风险重点，审查器就越不需要加载额外上下文。
The default workflow is deliberately conservative:
默认流程刻意保持保守：
| Default behavior / 默认行为 | Meaning / 含义 |
|---|---|
| Read-only / 只读 | No code, configuration, dependency, commit, branch, PR, or review-comment changes / 不修改代码、配置、依赖、提交、分支、PR 或审查评论 |
| Diff-first / 优先 diff | Inspect the requested change before surrounding files / 先看目标变更，再看必要上下文 |
| Evidence expansion / 证据扩展 | Read one direct caller or callee layer only when a concrete question requires it / 只有具体问题需要时才扩展一层调用关系 |
| Bounded validation / 有界验证 | Run at most one targeted check by default / 默认最多运行一次定向检查 |
| No network by default / 默认不联网 | Do not fetch remotes or external data unless explicitly requested / 未明确要求时不抓取远程或外部数据 |
| Stop on sufficiency / 证据足够即停止 | Do not continue for completeness when the conclusion is already supported / 结论有证据支持后不为“完整”而继续扩大范围 |
The skill does not install dependencies, rebuild environments, launch subagents, or create an automatic-fix loop.
本 Skill 不安装依赖、不重建环境、不启动子 Agent，也不创建自动修复循环。
Unless the user explicitly asks for a broader review, the default limits are:
除非用户明确要求扩大范围，默认限制为：
- no more than 8 changed source files / 最多 8 个变更源文件；
- no more than 1,200 changed lines / 最多 1,200 行变更；
- no more than one direct caller or callee layer outside the diff / diff 外最多扩展一层直接调用关系；
- no more than one targeted test, check, or command after static review / 静态审查后最多运行一次定向验证。
When a change exceeds the budget, the reviewer starts with the highest-risk portion, states current coverage, and asks before expanding the audit.
如果变更超过预算，审查器先检查最高风险部分，说明当前覆盖范围，并在扩大审计前征求确认。
1. Establish target
→ status, diff stat, complete relevant diff
2. Identify concrete questions
→ expand only to the closest required context
3. Review in risk order
→ behavior, interfaces, security, async/performance, tests
4. Run one proportionate validation
→ one test, syntax check, type check, or dry run
5. Stop and report
→ findings first, then validation and coverage
1. 建立目标：状态、diff 统计、完整相关变更
2. 提取具体问题：只扩展必要的最近上下文
3. 按风险顺序审查：行为、接口、安全、异步/性能、测试
4. 做一次相称验证：测试、语法、类型或 dry run
5. 停止并报告：先发现，再验证与覆盖范围
An item is actionable only when the review has all four parts:
一个发现只有同时具备以下四项，才会作为可执行问题输出：
-
a concrete file and line or code location;
-
a plausible execution path or affected behavior;
-
an observable consequence;
-
a proportionate correction or next check.
-
明确的文件和代码位置；
-
可信的执行路径或受影响行为；
-
可观察的后果；
-
相称的修复建议或下一步检查。
Findings are classified by evidence level:
发现按证据强度分类：
CONFIRMED— demonstrated by the diff or executed validation / 已由代码或验证直接证明；HIGH-PROBABILITY— strongly supported but not executed end to end / 证据很强但尚未端到端执行；NEEDS-CONTEXT— cannot be decided within the bounded scope / 在当前范围内无法判断；OPTIONAL— non-blocking maintainability improvement / 非阻塞的可维护性建议。
Low-confidence speculation is not presented as a defect.
低置信度推测不会被包装成缺陷。
The default report is compact and findings-first:
默认报告保持紧凑，并优先列出发现：
Lightweight Diff Review: <N> actionable finding(s)
[CONFIRMED | HIGH-PROBABILITY] <severity> — <file:line>
Problem: <specific defect and triggering path>
Impact: <observable consequence>
Recommendation: <minimal correction or verification>
Validation:
- <command or static-only review result>
Coverage:
- Reviewed: <diff/files/context>
- Not reviewed: <explicit exclusions>
If no substantive issue is found, the report says so directly and still states validation and exact coverage limits.
如果没有实质问题，报告会直接说明，并仍然列出验证方式和准确的覆盖边界。
Recommend a broader review only when the diff gives a concrete reason, such as:
只有在 diff 提供明确理由时，才建议扩大审查，例如：
- public API or schema changes / 公共 API 或 Schema 变化；
- authentication, authorization, or sensitive-data paths / 认证、授权或敏感数据路径；
- data migrations or destructive persistence changes / 数据迁移或破坏性持久化变化；
- concurrency or shared infrastructure changes / 并发或共享基础设施变化；
- a change that clearly exceeds the default budget / 明显超过默认预算的变更。
The reviewer explains the required scope and waits for approval before expanding.
审查器会说明所需范围，并在扩大审查前等待批准。
skills/
└── lightweight-diff-review/
└── SKILL.md
The runtime skill is intentionally small; the repository README is documentation for people and installers, not a file loaded into every review context.
运行时 Skill 刻意保持小型；本 README 面向人类和安装器，不会在每次审查时加载进上下文。
-
Diff before repository — start with the smallest evidence surface that can answer the question.
-
Bounded context — spend tokens on unresolved questions, not completeness theater.
-
Independent review — inspect the recent change without silently fixing it.
-
Evidence-calibrated language — separate confirmed defects from context gaps.
-
A clear stop — stop when the evidence supports the conclusion.
-
先 diff 后仓库：从足以回答问题的最小证据面开始；
-
上下文有界：Token 用于解决未决问题，而不是制造“完整感”；
-
独立审查：检查近期变更，不静默修复；
-
证据匹配措辞：区分已确认缺陷与上下文不足；
-
明确停止：证据足够支持结论后停止。
The workflow is independently written for low-token, low-side-effect diff review. It is conceptually informed by useful review categories in garrytan/gstack, but it does not copy its prompts, implementation, CLI, or detector corpus.
本工作流针对低 Token、低副作用的 diff 审查独立编写。它在概念上参考了 garrytan/gstack 的部分审查分类，但不复制其 Prompt、实现、CLI 或 Detector 数据集。