Skip to content

agent-skills.md

Latest commit

 

History

History
63 lines (42 loc) · 1.82 KB

agent-skills.md

File metadata and controls

63 lines (42 loc) · 1.82 KB

Agent skills

ASMP integrates with the Open Agent Skill spec. When a State Frame includes active_skill, the client should load that skill only in the current state, keeping context minimal (just-in-time).

State Frame field

active_skill is an optional object on the State Frame:

Field Required Description
name Yes Skill slug (e.g. audit-skill).
url Yes URL to fetch the skill (e.g. a SKILL.md or skill manifest).
context_summary No Brief recap of previous steps to avoid context gaps.
version No Optional skill version.

Client behavior

When the frame includes active_skill, the client should:

  1. Fetch the skill from active_skill.url (e.g. GET that URL).
  2. Inject its content (and optionally context_summary) into the LLM system message or message history.

Only the skill for the current state is loaded. When the run transitions to another state, the frame may expose a different active_skill or none.

Server: registering a skill per state

Python

workflow.skill(
    "REVIEW",
    "approval-skill",
    "approval-skill/SKILL.md",
    context_summary="Document uploaded and analyzed.",
)

The server must serve the skill file (e.g. at GET /skills/approval-skill/SKILL.md) so that active_skill.url (e.g. {base_url}/skills/approval-skill/SKILL.md) is fetchable by the client.

TypeScript

workflow.skill(
  "REVIEW",
  "approval-skill",
  "approval-skill/SKILL.md",
  "Document uploaded and analyzed."
);

Same idea: ensure the skill URL is reachable (e.g. mount a route that serves files from a skills/ directory).

Spec

See spec/SKILL_INTEGRATION.md for full integration details.