Skill Cleaner

Skill Cleaner

Use this when trimming skill prompt budget, finding duplicate skills, auditing enabled/disabled skill roots, or deciding which skills/plugins to remove.

Workflow

1. Resolve this skill's directory from the loaded SKILL.md, then run its bundled analyzer:

node ./scripts/skill-cleaner.mjs --no-logs

Useful variants:

node ./scripts/skill-cleaner.mjs --months 3
node ./scripts/skill-cleaner.mjs --months 6 --max-log-mb 800 --deep-logs
node ./scripts/skill-cleaner.mjs --context-tokens 200000 --budget-percent 2 --no-logs
node ./scripts/skill-cleaner.mjs --root /path/to/another/skills-directory --no-logs
node ./scripts/skill-cleaner.mjs --json --no-logs

2. Read the report in this order:

• Skill Budget: GPT-5.5 context size, 2% skills budget, Codex-budgeted usage, and pre-budget full-list pressure.
• Description candidates: long descriptions where relaxed grammar saves prompt budget.
• Duplicates: same skill name or near-identical description/body across Codex, plugin cache, repo siblings, and personal skill roots.
• Unused candidates: no recent $skill mention, SKILL.md read, or explicit skill-use trace in recent Codex/OpenClaw logs.
• Root summary: where skills came from and whether config marks them disabled.

3. Before deleting or editing:

• Verify the kept copy exists and is loaded.
• Treat cross-runtime duplicates as informational: separate runtimes may legitimately need separate copies.
• Never delete managed plugin-cache files directly; use the runtime's plugin manager.
• Delete a suggested duplicate only after verifying the kept path exists and the same runtime loads it.
• Keep repo-local OpenClaw maintainer skills when they encode repo policy or live operations.
• Preserve trigger nouns in descriptions: product, tool, action, object.

Analyzer Notes

• The script mirrors Codex's model-visible line shape: - name: description (file: path).
• It applies Codex-like frontmatter rules: YAML frontmatter only, default name from parent dir, single-line sanitized name and description.
• It follows Codex core-skills/src/render.rs: 2% of raw context_window, token cost ceil(utf8_bytes / 4), then full descriptions -> equal description truncation -> omitted minimum lines.
• It reads ~/.codex/models_cache.json for the selected model's context_window; fallback is 200,000 tokens and 2%. Pass --context-tokens for an exact budget.
• It scans common global roots for .agents, Pi, Codex, Claude Code, OpenCode, and OpenClaw, plus skill roots in the current project and its ancestors.
• Extra folders are included only with repeatable --root <path> flags.
• It realpath-dedupes roots, so agent-specific symlinks to a shared skill directory do not create false duplicates.
• For duplicate names, it reports description/body similarity and suggests deletion candidates only when bodies are near copies. Keep priority defaults to direct Codex system skills, then direct Codex skills, then plugin skills, then personal/repo copies.
• It scans ~/.codex/history.jsonl and recent ~/.codex/sessions/**/*.jsonl by default. Add --deep-logs for archived sessions and common OpenClaw/Clawd log folders.
• Usage evidence is heuristic: $skill, Use $skill, and paths like skills/<name>/SKILL.md.

Privacy and Output Policy

• Scanning stays local. The report emits paths and usage counts, not session contents.
• Use --no-logs unless the user wants usage-history analysis.
• Reports can contain usernames and local paths; review them before sharing.
• Suggest first; edit only when the user asks.
• If asked to apply cleanup, make small grouped commits: descriptions, deletes, config disables.
• Do not delete ignored/untracked skill dirs without naming the destination or confirming they are disposable.