Skip to content
Wheelhouse Wheelhouse

Skills

A skill is a versioned recipe — a set of markdown files and steps stored in git — that an agent can invoke by publishing a SkillInvocation object into a stream.

Skills are deliberately decoupled from any LLM’s native tool-calling format. The stream object is the invocation mechanism.

Structure

Section titled “Structure”
skills/
  web-search/
    skill.md        # description, inputs, outputs
    steps.md        # execution steps
    requirements.txt  # optional dependencies

Invocation

Section titled “Invocation”
await stream.publish(SkillInvocation(
    skill="web-search",
    input={"query": "latest ZMQ Rust bindings"},
    reply_to="main"
))

The agent picks up the SkillInvocation, executes the skill, and publishes a SkillResult — or an error object if it fails. Silent timeouts are forbidden.

For long-running skills, the agent publishes SkillProgress objects at regular intervals so surfaces can show progress to the user.

Declaring skills in .wh

Section titled “Declaring skills in .wh”

Skill references must be pinned to a specific git commit hash — not a branch name. This ensures reproducibility and prevents supply chain drift.

agents:
  - name: donna
    skills:
      - repo: github.com/wheelhouse-paris/skills
        name: web-search
        ref: a3f9c2d   # pinned commit hash, not a branch

Skills are lazy-loaded on first invocation. An agent can only invoke skills declared in its .wh — undeclared skill invocations are rejected.