Skill Installer

Skill Installer

Overview

Installs URL-supplied third-party skills into the openscience CLI. The skill itself is a thin wrapper around the bundled openscience skill bash subcommand — when the user expresses install/uninstall intent, invoke openscience skill add|list|remove via the bash tool. The CLI handles fetching, the safety gate, confirmation, on-disk write, and cross-machine sync.

When to Use This Skill

Trigger on user messages like:

• "Add this skill: "
• "Install " / "Install the skill at "
• "Set up skills from "
• "Remove the X skill" / "Uninstall "
• "List my installed skills"

Do NOT use this skill for:

• Generating new skills (different workflow — see openscience learn)
• Installing system packages (use bash directly)
• Anything that's not a openscience skill source

How to Use

Add a skill

openscience skill add <git-url>

Accepted URL forms:

• https://github.com/<owner>/<repo> — default branch, all skills in skills/**/SKILL.md
• https://github.com/<owner>/<repo>/tree/<ref> — specific ref/tag
• gh:<owner>/<repo> — shorthand
• gh:<owner>/<repo>@<ref>[/<path>] — shorthand with ref and optional path scope
• Generic git+ssh://... / https://gitlab.com/... URLs

What happens when the user runs add:

1. CLI clones the repo into a temp dir, pins the commit SHA.
2. Enumerates skills/<name>/SKILL.md files.
3. Runs Layers 1, 2, 4 (local regex passes) on every SKILL.md + companion script.
4. POSTs the surviving manifest to /api/cli/skill-review for the Layer-3 Haiku classifier (server-side, sandboxed input with canary integrity check).
5. Renders a confirm screen showing the manifest + warnings + classifier reasoning + first lines of each SKILL.md.
6. On y, writes everything to ~/.openscience/installed-skills/<namespace>/<name>/ and uploads to the dashboard.

The <namespace> is derived from the repo name (last URL segment, lowercased).

List installed skills

openscience skill list

Prints <namespace>/<name> with a ⚠ marker for any skill the safety gate warned on.

Remove a skill

openscience skill remove <namespace>            # whole namespace
openscience skill remove <namespace>/<name>     # single skill within a namespace

Soft-deletes the cloud record (archived_at set) and removes the on-disk directory.

Examples

User: "Install the brainstorming skill from Anthropic's superpowers repo."

You: invoke bash with openscience skill add gh:anthropics/superpowers/skills/brainstorming. Wait for the spinner + confirm prompt. Relay the manifest to the user. Pass through their y/N response.

User: "Uninstall everything in superpowers."

You: invoke bash with openscience skill remove superpowers. Report the count of archived skills.

User: "What's installed?"

You: invoke bash with openscience skill list. Print the output.

Safety Notes

The user-facing openscience skill add command runs every install through 6 layers (regex catastrophic-reject, regex classifier-injection-reject, server-side LLM classifier, regex suspicious-warn, user-confirm screen, deferred allowed-tools sandboxing). Don't try to bypass any of these by editing ~/.openscience/installed-skills/ directly — that breaks the cloud-sync invariant. Always go through the bash subcommand.

If openscience skill add reports classifier unreachable, the dashboard backend is down — surface that to the user and stop. Do not retry with a flag that disables the classifier without the user's explicit consent.

Reference

• Spec: thesis/docs/superpowers/specs/2026-05-14-cli-add-skill-from-url-design.md
• Plan: thesis/docs/superpowers/plans/2026-05-14-cli-add-skill-from-url.md
• On-disk layout: ~/.openscience/installed-skills/<namespace>/<name>/SKILL.md (mirror of ~/.openscience/learned-skills/)
• Backend endpoints under /api/cli/installed-skills + /api/cli/skill-review