ROSClaw Skills

Official ROSClaw Skills Catalog — versioned, verifiable physical skills for embodied agents.

ROSClaw Skills are reusable physical capabilities for robots and embodied agents. A Skill is not just a prompt or a policy checkpoint. A ROSClaw Skill is a versioned physical asset containing behavior recipes, safety constraints, provider routes, compatible embodiment profiles, evaluation evidence, and lineage.

What is a ROSClaw Skill?

A Skill may contain:

• SKILL.md — agent-readable skill instruction
• skill.yaml — skill metadata and runtime contract
• behavior_tree.xml — executable behavior structure
• prompts/ — planner, verifier, recovery, and safety prompts
• policies/ — policy adapters, parameters, or checkpoints
• providers.yaml — model and tool provider routing
• e-urdf-compat.yaml — compatible robot embodiments
• safety.yaml — runtime safety constraints
• dojo.yaml — practice / mining / replay configuration
• darwin_eval.yaml — evaluation and promotion gates
• evidence/ — evaluation summaries and reports
• lineage.yaml — origin, version, and promotion history

Install a Skill

rosclaw skill install ros-claw/ros_install

Install a specific version:

rosclaw skill install ros-claw/ros_install@0.1.0

No ROSClaw API key is required to install public official skills.

Available Skills

The generated registry is located at:

registry/skills.json

Each skill lives under:

skills/<skill_name>/

Contribute a Skill

Submit a pull request adding your skill under:

skills/<your_skill_name>/

Your PR must pass:

• schema validation
• required file checks
• safety checks
• secret scanning
• local path scanning
• installation dry-run
• registry generation

See CONTRIBUTING.md.

Official Verification

A Skill becomes official only after:

1. It is added to this repository.
2. GitHub Actions validation passes.
3. A maintainer reviews and merges the pull request.
4. The registry is rebuilt.
5. The Skill is synchronized to the ROSClaw Skill Hub.

Safety Notice

ROSClaw Skills may control or influence physical robots. Every physical Skill must include safety.yaml, compatible embodiment constraints, and evaluation evidence. Unsafe, unverifiable, or sandbox-bypassing Skills will be rejected or revoked.

License

Apache-2.0 unless otherwise specified inside each Skill directory.