@tuttiai/skills

Agent-callable skills synthesised from observed trajectories.

A trajectory is the recorded tool-call sequence of one agent run. When the same trajectory shape recurs across runs, a future synthesiser proposes a skill candidate — a name, description, signature, and the list of constituent tools. An operator approves or rejects the candidate; approved candidates become skills that the agent can call directly.

This v0.1 package ships only the storage and review primitives:

• Trajectory / TrajectoryToolCall — recorded run shapes
• SkillCandidate / Skill — proposal and approval states
• SkillStore — interface for trajectory + candidate + skill persistence
• InMemorySkillStore — reference implementation, suitable for tests and single-process deployments

The candidate synthesiser, the runtime adapter that exposes approved skills as tools, and persistent backends are out of scope for v0.1.

Event emission

If constructed with an EventBus, InMemorySkillStore emits:

• skill:candidate_proposed — on proposeCandidate
• skill:approved — on approveCandidate
• skill:rejected — on rejectCandidate

recordTrajectory does not emit — trajectories are append-only observability.

Quick start

import { InMemorySkillStore } from "@tuttiai/skills";
import { EventBus } from "@tuttiai/core";

const bus = new EventBus();
const store = new InMemorySkillStore({ events: bus });

bus.on("skill:candidate_proposed", (e) => {
  console.log(`new candidate: ${e.name_suggestion} (${e.evidence_count} runs)`);
});

await store.recordTrajectory(trajectory);
await store.proposeCandidate(candidate);
await store.approveCandidate(candidate.id, { reviewed_by: "alice" });