skill-manager

Skill Manager

Add community skills to this agent, list what's installed, or remove what you no longer want. A "skill" here is a directory with a SKILL.md — the same format Claude Code natively uses — so the exact same artifact works in either place.

This is a CORE feature. See docs/skill-manager.md for the full reference.

Dispatch

Parse the user's command and route to the correct MCP tool:

User says	Action
/agent:skill install <source>	Call skill_install({ source })
/agent:skill install <source> --scope=<plugin|project|user>	Call skill_install({ source, scope })
/agent:skill install <source> --force	Call skill_install({ source, force: true }) — only after the user confirms an overwrite
/agent:skill list	Call skill_list() and print the card
/agent:skill remove <name>	Call skill_remove({ name }) (dry run) then ask the user to confirm, then call again with confirm: true
/agent:skill remove <name> --force	Call skill_remove({ name, confirm: true }) directly — no confirmation prompt

Default scope is plugin (./skills/). Users can promote a skill later.

Install flow

1. Validate the source (GitHub shorthand, full URL, or local path — the tool parses all three)
2. Call skill_install({ source }) — the tool clones, detects format, checks requirements
3. Print the returned card verbatim
4. If format: openclaw was detected (rejected), DO NOT retry with force. Suggest /agent:import-skill and stop.
5. If the skill installed with warnings, remind the user what needs to be set up (missing env var, binary, etc.)

List flow

1. Call skill_list()
2. Print the card. On messaging channels, compress to one line per skill.

Remove flow

1. Call skill_remove({ name }) first — this is a dry run; it reports what WOULD be removed
2. Ask the user: "Remove from ? [y/N]" — short and clear
3. If they confirm, call skill_remove({ name, confirm: true })
4. If they say no, just acknowledge and stop

If the user passed --force, skip step 2 and call directly with confirm: true.

Response style

• Terse. No preamble.
• Install: show the install card (already formatted) then one line of what to do next if warnings.
• List: the card as-is on CLI/WebChat; one line per skill on WhatsApp/Telegram.
• Remove: confirmation prompt → result line.

Never

• Do NOT run install scripts from cloned repos (no postinstall, no install.sh). The tool does not run them and neither do you.
• Do NOT try to auto-convert OpenClaw-flavored skills. If the tool rejects one as OpenClaw, direct the user to /agent:import-skill.
• Do NOT install without showing the install card — the user needs to see what they're getting.

References

• docs/skill-manager.md — full docs
• lib/skill-manager.ts — implementation
• docs/INDEX.md — master feature list