Agent Skills › szw321127/repowise › pk-preflight

pk-preflight

GitHub

在项目实现前检查项目知识库，检索匹配的最佳实践和本地证据提示。根据任务描述解析脚本并生成中文摘要，包括推荐依据、意图推断及证据路径，支持只读模式与命中记录选项。

plugins/pk/skills/pk-preflight/SKILL.md szw321127/repowise

Trigger Scenarios

用户希望在开始任务前了解项目相关知识需要检索与当前任务匹配的实践或本地证据

Install

npx skills add szw321127/repowise --skill pk-preflight -g -y

More Options

Non-standard path

npx skills add https://github.com/szw321127/repowise/tree/main/plugins/pk/skills/pk-preflight -g -y

Use without installing

npx skills use szw321127/repowise@pk-preflight

指定 Agent (Claude Code)

npx skills add szw321127/repowise --skill pk-preflight -a claude-code -g -y

安装 repo 全部 skill

npx skills add szw321127/repowise --all -g -y

预览 repo 内 skill

npx skills add szw321127/repowise --list

SKILL.md

Frontmatter

{
    "name": "pk-preflight",
    "description": "Use when the user wants to inspect project knowledge before starting a task and retrieve matching practices or local evidence hints."
}

PK Preflight

Treat this skill as the project-knowledge entrypoint before implementation work starts.

Required Behavior

Accept a project root or .project-knowledge/ directory plus a task description.
Resolve ../../scripts/pk-preflight.mjs relative to this SKILL.md file.
The wrapper delegates to preflight-session.mjs.
Report the result in Chinese, including whether existing knowledge was hit, which practices/options were recommended, why they matched via matchReasons, what task intent was inferred, or which local evidence paths should be inspected.
If .project-knowledge/project-profile.md is missing, report mode: no-knowledge and do not scan local code or evidence; suggest pk-init only if the user wants this project to opt in.
Treat the output as a context-budgeted summary: by default only the top matching practices are returned, source_evidence is a preview, and source_evidence_count / source_evidence_truncated indicate whether more evidence exists.
source_evidence previews and evidenceHints exclude missing evidence paths; if expected evidence is absent, ask the user to run pk-lint to inspect node-missing-evidence-path findings.
If evidenceHintsTruncated is true, inspect the referenced Markdown node or source files only when the current task needs deeper verification.
Matching uses node keywords plus deterministic task intent (taskKinds, technologies, pathHints, operationHints). Mention applies_when / does_not_apply_when hits or exclusions when relevant.
Keep preflight read-only by default. Only record preflight_hits and last_hit_at in state/usage-index.json when the caller explicitly opts in with --record-hits or recordHits: true.