same-set-reusable-flow-for-project-file-inventor - SKILL.md Agent Skill

name: same-set-reusable-flow-for-project-file-inventor version: 0.1.0 author: Meta_Kim user-invocable: false candidateType: skill sourceGapId: gap-create-skill-same-set-reusable-flow--c89068df05 approvalEvidence: warden-approved-file-inventory-skill-2026-06-05 trigger: "project file inventory|file classification|repo cleanup|项目文件整理|文件用途整理|清理项目" tools: - shell - filesystem description: | Reusable Meta_Kim file inventory classification flow. It helps separate durable sources, generated evidence, runtime mirrors, temporary state, and risky unknowns before cleanup or commit.

Project File Inventory Classification

Use this skill when a Meta_Kim workspace has too many files and the user needs to know which are useful, generated, risky, or safe to ignore.

Goal

Produce a source-backed inventory, not a cleanup guess. The output should help the user decide what to keep, commit, ignore, or inspect again.

Required Evidence

Start with read-only evidence:

git status --short --branch
git ls-files
git ls-files --others --exclude-standard
git check-ignore -v <path> for suspicious ignored paths
rg -n "<filename-or-command>" package.json docs scripts tests canonical config for tracked candidates
npm run script inventory from package.json

Do not classify a tracked file as removable only because it has few references. Check package scripts, tests, docs, sync scripts, setup scripts, and contracts first.

Classification Buckets

Core source: canonical/, config/contracts/, config/capability-index/, durable scripts, package scripts, and tests that define behavior.
Runtime projection: .claude/, .agents/, .cursor/, openclaw/, .codex/ mirrors generated from canonical sources or sync scripts.
Generated evidence: .meta-kim/state/default/, graph outputs, reports, SQLite run state, smoke artifacts, and temporary proof files.
Planning continuity: task_plan.md, findings.md, progress.md; keep during active work even if ignored.
Reference docs: tracked docs and plans that explain product, architecture, contracts, or release decisions.
Suspicious or cleanup candidate: files with no package, test, docs, sync, install, contract, or runtime evidence.

Output Format

Return a compact table with:

path or pattern
bucket
why it exists
evidence command
recommendation: keep, commit, ignore, regenerate, inspect again, or cleanup candidate
risk if deleted

End with a short "do not delete yet" list for any file whose role is uncertain.

Boundaries

Do not auto-write canonical state without Warden approval.
Keep one-run task details in workerTaskPackets, not durable identity.
Do not delete files as part of inventory classification unless the user separately approves a cleanup action.
Do not treat runtime mirrors or generated reports as source of truth when canonical or config sources exist.

Prompt Acceptance

This skill binds the execution-tools-and-commands, capability-discovery-and-retrieval, planning-continuity, runtime-native-surfaces, and verification-eval-and-release abstract capability families. It may inspect filesystem, Git, scripts, config, docs, and runtime projections, but it does not mutate the workspace by itself.

Pass

Inventory output separates source, runtime projection, generated evidence, planning continuity, reference docs, and cleanup candidates.
Each recommendation names the evidence command that supports it.
Active task_plan.md, findings.md, and progress.md are preserved as update-only continuity state.
Unknown or weakly evidenced files are marked inspect again rather than deleted or ignored.

Fail

A tracked file is called removable without checking scripts, tests, docs, setup, sync, contracts, and runtime projections.
Runtime mirrors, generated reports, or one-run artifacts are treated as the canonical source of truth.
The skill recommends deletion, untracking, or canonical writeback without explicit user approval and Warden approval where required.

Block

Block cleanup or write actions when evidence is missing, Git state is unclear, the file is active planning state, or the user has not explicitly approved mutation.

Return to stage

Return to Fetch when file purpose or source-of-truth status is unclear. Return to Thinking when the inventory exposes a capability gap or a cleanup path with competing trade-offs. Return to Verification when the classification depends on generated evidence that needs a fresh command.

Verification

Use read-only commands such as git status --short --branch, git ls-files, git ls-files --others --exclude-standard, git check-ignore -v <path>, targeted rg, package script inspection, and the most specific validator for the affected file class. Run npm run meta:prompt:validate after editing this skill.

Preserve

Preserve Skills, WebSearch/browser/research, filesystem, shell, apply_patch, MCP, memory, Graphify, graph, hooks, commands, rules, agents, subagents, approval, sandbox, runtime tools, package scripts, setup, sync, install, uninstall, status, doctor, validators, runtime projections, and active planning files.