代码一致性审计

安全 社区

技能解读

当前原文的自动摘译,尚未人工审校。

目标

合并更改后审核存储库的一致性。在团队中,这项工作通常分为合并后验证、集成和回归测试、发布准备情况审查、文档漂移检查和架构适应性功能。使用此技能来运行横切一致性,传递那些仅部分覆盖的标签。

要捕获的输入

正在审核的存储库或子树以及任务是仅审核还是审核并修复。最近的重构、分支、PR 或代理线程输出最有可能发生交互。预期的真实来源,例如自述文件、文档、示例、配置、

技能特定规则

将存储库一致性视为目标。不要将任务变成通用样式或代码质量审查。在判断实施细节之前,先从声明的行为和事实来源开始。将两个公开的事实来源之间的冲突视为调查结果,即使

边界

当存储库在合并、重构、迁移或并行代理工作后可能出现语义漂移并且用户想要一致性审核而不是样式反馈时,请使用此技能。如果用户只想对有限补丁内的错误或回归进行代码审查,请使用正常的审查流程而不是整个存储库审核。

工作流程

  1. 首先定义预期行为 从信号最强的本地源中识别声明的行为:自述文件、文档、示例、测试、配置、清单、自动化、控制文件、命名和合同。对于技能库或代理库,请处理SKILL.md,agents/openai.yaml,自述文件
流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
安全
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
88 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@Ceratops-Code · 未声明许可证
上下文用量评级
低消耗
接入复杂程度
需简单配置
是否需要外部接口密钥
不需要
兼容的系统
未声明(默认跨平台)
运行环境要求
Python
检测到的文件与系统行为
  • 只读
  • 命令执行
网络行为
仅限本地
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

# Ceratops Code Consistency Audit

## Goal

Audit repository coherence after combined changes. Across teams this work is
usually split across post-merge validation, integration and regression testing,
release-readiness review, documentation-drift checks, and architectural fitness
functions. Use this skill to run the cross-cutting consistency pass those labels
only partially cover.

## Context

### Inputs To Capture

- Repo or subtree under audit and whether the task is audit-only or
  audit-and-fix.
- Recent refactors, branches, PRs, or agent-thread outputs most likely to have
  interacted.
- Expected sources of truth such as README files, docs, examples, configs,
  tests, manifests, generated metadata, automation, or control files.
- Local comment standards such as
  `skills/ceratops-gh-repo-lifecycle/references/code-comment-nondeterministic-contract.md`,
  README guidance, or language-specific conventions already present in the repo.
- High-risk surfaces such as public APIs, migrations, feature flags, rename
  waves, packaging, install flows, or generated artifacts.

Infer missing inputs from repo state before asking.

## Constraints

### Skill-Specific Rules

- Treat repository coherence as the goal. Do not turn the task into a generic
  style or code-quality review.
- Start from declared behavior and sources of truth before judging
  implementation details.
- Treat conflicts between two declared sources of truth as findings even when
  the code still happens to work.
- Prefer concrete contradictions, stale follow-through, and merged-only
  interaction bugs over speculative architecture advice.
- Treat comment sufficiency as a non-deterministic maintainability consistency
  issue when scripts, public APIs, automation helpers, contract checkers, or
  non-obvious safety logic are in scope.
- Use local project conventions and
  `skills/ceratops-gh-repo-lifecycle/references/code-comment-nondeterministic-contract.md`
  when present; do not research current language comment standards unless the
  user explicitly asks for a standards refresh.
- When git history or recent merge context is available, judge the current
  merged result against the latest target-branch state, not each branch in
  isolation.
- Do not ask for credentials for normal local repo audits.
- If external systems are genuinely needed, first exhaust local repo state,
  local git history, and no-auth metadata.

### Boundaries

- Use this skill when a repo may have semantic drift after merges, refactors,
  migrations, or parallel agent work and the user wants a consistency audit
  rather than style feedback.
- If the user wants only a code review for bugs or regressions inside a bounded
  patch, use normal review flow instead of this whole-repo audit.
- If the task is only diagnosing and fixing one current breakage, stop and use
  `$ceratops-task-lifecycle` with the `execute-in-stages` action.
- If the task is only documentation maintenance without broader repository
  coherence concerns, narrow the work to documentation drift rather than running
  the full audit.

### Workflow

#### 1. Define intended behavior first

- Identify declared behavior from the highest-signal local sources: README
  files, docs, examples, tests, configs, manifests, automation, control files,
  naming, and contracts.
- For skill repos or agent repos, treat `SKILL.md`, `agents/openai.yaml`, README
  tables, bundled resources, and any install metadata as first-class sources of
  truth.
- If multiple artifacts disagree about the intended behavior, record that
  conflict explicitly instead of guessing which one is correct.

#### 2. Map merged interaction surfaces

- Inspect adjacent modules, flags, configs, scripts, templates, tests, and
  examples that were likely touched by separate refactors or threads.
- Use git history, diff context, or merge context when available to find where
  independently reasonable changes now interact.
- Prioritize hidden coupling, duplicated logic that drifted, partial renames,
  stale options, orphaned files, and follow-through gaps.

#### 3. Run consistency passes

Check as many of these as the repo justifies:

- implementation vs README and docs
- implementation vs examples, scripts, and run instructions
- file names and stated purpose vs actual behavior
- tests vs current intended behavior
- configs, manifests, automation, and control files vs implementation
- public interfaces vs internal assumptions
- comments vs non-obvious behavior, safety boundaries, external side effects,
  script usage, and README-maintained workflow expectations
- merged-only edge cases that appear when features or refactors combine
- stale artifacts, dead references, partial migrations, and unused compatibility
  shims
- documented validation commands only for touched surfaces: Markdown or docs,
  YAML or workflows, Python helpers or validators, and skill source or
  governance files. Treat lint or type output as consistency evidence, not as
  generic style findings.

#### 4. Validate findings before reporting

- Prefer findings that point to exact conflicting artifacts, exact
  inconsistency, actual risk, and the smallest credible fix.
- Avoid style nits, naming preferences, or refactor suggestions unless they are
  needed to resolve a real contradiction.
- Distinguish confirmed findings from plausible but unverified risk areas.

#### 5. Fix only when asked or clearly justified

- Default to audit and report.
- If the user asked to repair the repo or the smallest safe fix is already in
  scope, apply the narrowest credible correction and rerun the relevant checks.
- Escalate before risky deletions, broad rewrites, or behavior-changing
  interpretation choices.

#### 6. Close with explicit coverage

- Classify important checked surfaces as consistent, conflicting, blocked, or
  not verified.
- If no findings remain, say so explicitly and note the residual coverage limits
  instead of implying exhaustive certainty.

## Done When

### Completion Gate

- Verify the audit inspected the repo's declared sources of truth, not only the
  implementation files.
- Verify merged interaction surfaces and likely cross-thread integration seams
  were checked where evidence existed.
- Verify comments and README coverage are sufficient for important scripts,
  public interfaces, automation helpers, and non-obvious safety or contract
  logic when those surfaces were part of the audit.
- Verify every reported finding ties back to concrete conflicting artifacts or
  an explicitly stated verification gap.
- Verify the final answer reports findings, important consistent areas, and
  important verification limits without drifting into generic review commentary.
- Verify relevant documented validation commands were run or explicitly scoped
  out for touched Markdown, YAML, Python, and skill-source surfaces.

### Output Contract

Report only:

- findings ordered by severity
- for each finding: conflicting artifacts, exact inconsistency, risk, and the
  smallest credible fix
- important areas checked and found consistent
- anything important not verified

If no findings remain, say so explicitly and mention the main residual risk
areas or coverage limits.

### Example Invocation

`Use $ceratops-code-consistency-audit on this repo. Ignore style. Find
post-merge contradictions, docs-to-code drift, stale assumptions, and
merged-only edge cases.`

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。