跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
开发与测试

dx-audit

Audits libraries, CLIs, and SDKs using 38 rules for public contracts, package exports, piped output, errors, and configuration. Use when asked to "audit my CLI", "review my SDK", "make this agent-friendly", or diagnose package type resolution. For agentic product trust use ax-audit; for docs use docs-writing.

mblode · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add mblode/agent-skills --skill dx-audit

技能说明


name: dx-audit description: Audits libraries, CLIs, and SDKs using 38 rules for public contracts, package exports, piped output, errors, and configuration. Use when asked to "audit my CLI", "review my SDK", "make this agent-friendly", or diagnose package type resolution. For agentic product trust use ax-audit; for docs use docs-writing.

DX Audit

Audit or improve what developers import, run, configure, or read when something fails.

  • IS: a bounded review of public APIs, developer-facing errors, CLI commands, exported types, install and first-run behavior, and config, with fixes only when asked.
  • IS NOT: a repo-wide quality sweep (pr-reviewer), end-user UI (ui-design Audit mode), agent trust review (ax-audit), docs prose (docs-writing), README (readme-creator), repository architecture (codebase-architecture), or building a new CLI (scaffold-cli).

Modes

Pick the narrowest mode the request supports, and write a one-line scope receipt before reading code:

ModeWhenOutput
Targeted (default)a named or changed public surfacefindings report, read-only
Fixthe user says fix, improve, simplify, or implementlocalized edits inside the receipt, then verification
Exhaustivethe user explicitly asks for the whole package or every public surfaceevery material finding, partitioned by surface
Scope: <mode>; surfaces: <commands/exports/config>; prefixes: <err-, cli->; excludes: <UI, docs, architecture, private internals>

"DX", "gold standard", and "review holistically" do not by themselves widen a targeted audit into exhaustive. When several skills are invoked together, this one owns only the surfaces above.

Audit progress

DX audit progress:
- [ ] 1. Lock the public surface and write the scope receipt
- [ ] 2. Gather local evidence and run the safe probes
- [ ] 3. Select prefixes, then open candidate rules
- [ ] 4. Rank root causes
- [ ] 5. Report, or fix when asked
- [ ] 6. Verify on the same scope

1. Lock the public surface

Start from git diff against the normal base and keep only changed files reachable from a public entry point: package.json exports or bin, a command registry, an exported type, a documented config loader, or an observed error path. With no useful diff, use the command, export, error, or config the user named. A private helper enters scope only through a public caller.

2. Gather evidence, then stop

  1. Local instructions, the manifest, and the diff or named entry point.
  2. Direct public dependencies and the nearest tests that pin behavior.
  3. Safe probes against the local build:
    • CLI: --help, --version, one success path, one invalid-input path, and the same command with stdout piped (| cat) to see non-TTY behavior. Never trigger a real mutation to test DX; use --dry-run where it exists.
    • Package: npx publint and npx @arethetypeswrong/cli --pack . after a build. These inspect packaging, but packing may invoke lifecycle scripts. Inspect those scripts first or use a disposable checkout before calling --pack.
  4. The prior release contract, only when the diff changes a public export, signature, or return shape.

Stop when the behavior is proven, disproven, private, or out of scope. External research is for an explicit comparison request or a named uncertainty local evidence cannot resolve; references/standards-map.md carries the standards this skill already leans on.

3. Dispatch rules candidate-first

Read rules/_sections.md, then select prefixes by surface:

PriorityPrefixCategoryDefault impactRules
1api-Public API and SDKCRITICAL7
2err-Developer-facing errorsCRITICAL5
3cli-CLI UX for humans and agentsHIGH13
4types-Exported type ergonomicsHIGH5
5onboard-Install and first runHIGH5
6config-Config ergonomicsMEDIUM3
SurfacePrefixes
Public API entry pointapi-, types-, reached err- paths
CLI commandcli-, reached err- paths
Exported declarationstypes-, plus api- when behavior changed
Install, package.json, first runonboard-
Config loaderconfig-, reached err- paths

Applicability outranks priority: a CLI-only audit never loads api- because API rules rank higher.

Targeted mode: list the filenames for the selected prefixes (the names are the checklist), look for concrete evidence, then open only the rule files a finding needs. Exhaustive mode: read every rule in the selected prefixes.

Capability gates, applied inside a selected prefix:

  • Structured JSON input, schema introspection, and compact polling snapshots apply when automation or agent use is promised, requested, or already supported.
  • Dry-run and confirmation apply to destructive, expensive, or hard-to-reverse mutations.
  • Progress and resume apply to operations that can block, outlive one command, or be retried after ambiguous output.
  • stdin applies when the command semantically accepts file or stream data.
  • Stable-contract comparison applies only when a public contract changed.

4. Rank root causes, not instances

  • Order CRITICAL, HIGH, MEDIUM by each cited rule's frontmatter impact, copied exactly. Impact is the rule's declared consequence, not a confidence score.
  • Merge repeated instances of one root cause into one finding with up to three representative locations. Missing JSDoc, error codes, or --json across a surface is one finding, not one per symbol.
  • A missing feature with no current consumer path is not a defect.
  • Targeted mode: every CRITICAL finding, then the highest-value remainder up to five total; summarize the rest by category.

5. Report or fix

Audit requests are read-only. A fix request authorizes localized changes inside the receipt, not a redesign of a

阅读完整来源 ↗