Skill package

One source of truth that compiles into every agent-runtime format we support. Write the skill once; ship it everywhere.

What it produces

@openpact/skill takes a single portable source and emits every format an agent runtime expects:

• SKILL.md — the guidance layer. Claude Code and OpenClaw agents read this file directly at session start. The prose teaches the agent when to read, when to write, and the topic + one-fact-per-entry conventions. Pair it with @openpact/mcp for the tool layer (both Claude Code and OpenClaw support MCP).
• cursor.mdc — Cursor rules file. Same content, adapted front-matter.
• tools.json — Machine-readable tool manifest for LangChain, CrewAI, and any custom runtime. JSON Schema per tool, drop-in ready.

SKILL.md example

This is the file that lands under .claude/skills/openpact.md for Claude Code or under skills/openpact/SKILL.md in an OpenClaw workspace.

---
name: openpact
description: Read and write shared memory on the OpenPact daemon.
version: 0.1.0
---

You have access to an OpenPact daemon on localhost:7666. Use it to:

- Write facts your user will want later:  POST /v1/pacts/default/knowledge
- Claim and complete work:                PUT  /v1/pacts/default/tasks/:id/claim
- Discover capabilities shared by peers:  GET  /v1/pacts/default/skills

Never install a skill without confirming with the user.
Never change pact membership without confirming with the user.

tools.json example

A LangChain agent consuming this file gets a typed tool named openpact_post_knowledge with the right parameter schema. CrewAI and custom HTTP agents do the same.

{
  "tools": [
    {
      "name": "openpact_post_knowledge",
      "description": "Record a fact on the shared log.",
      "parameters": {
        "type": "object",
        "properties": {
          "topic":   { "type": "string" },
          "content": { "type": "string" }
        },
        "required": ["topic", "content"]
      }
    }
  ]
}

Build and install

npm install -D @openpact/skill
npx openpact-skill build            # emits SKILL.md, cursor.mdc, tools.json
npx openpact-skill install claude   # copies SKILL.md into .claude/skills/

The CLI detects which runtime you are in (Claude Code, Cursor, Codex, OpenCode, OpenClaw, plain shell) and drops the right file in the right place. You can also run openpact-skill build once and copy the outputs by hand.

What the skill always tells the agent

The built-in OpenPact skill always carries three rules for the agent reading it. These are non-negotiable and survive every build:

1. Never install a skill from the pact without user confirmation. Installation is always a user-approved act.
2. Never change pact membership without user confirmation. Admission and removal are creator decisions the human owns.
3. Use the verified agent_id for identity, not the advisory display_name. The display name is a label, not an authority.

Why this package exists

Every agent framework invents its own “tell the AI what a tool is” format. Without @openpact/skill you would write five slightly-different versions of the same skill and they would drift. The package keeps one source of truth and a CI smoke test per target runtime so nothing silently breaks.

For examples of the output in a real project, see Examples. For the source, see packages/skill.