@eidentic/skills

Skill substrate for Eidentic — parse SKILL.md prompt skill manifests, manage skill banks (approve / list / register), run test-gated executable skills signed with ed25519 keypairs, and evolve skills with LLM-driven feedback loops. The skill system enables agents that improve their own behavior over time.

Install

pnpm add @eidentic/skills

Usage

import { SkillBank, SkillSet, evolveSkill } from "@eidentic/skills";
import type { ExecutableSkillDef } from "@eidentic/skills";
import { AIModel } from "@eidentic/model";
import { anthropic } from "@ai-sdk/anthropic";

// Load a prompt skill from SKILL.md.
const skillMd = `
---
name: summarize
description: Summarize long documents concisely.
allowed-tools: [read_document]
---
When asked to summarize, produce a bullet-point list of key points...
`;
const promptSkills = SkillSet.fromManifests([
  { content: skillMd, source: "inline:summarize" },
]);

// Agent core enforces this capability after skill_use activates the prompt skill.

// Executable skills use a separate, test-gated bank.
const executableSkill: ExecutableSkillDef = {
  name: "summarize-text",
  description: "Summarize text concisely.",
  tests: [{ name: "returns text", input: "hello", check: out => typeof out === "string" }],
  run: async input => String(input),
};
const bank = new SkillBank();
await bank.register(executableSkill);

// Approve for use in agents (agent-authored skills are quarantined until approved)
bank.approve("summarize-text");

// Evolve an executable skill using LLM-driven refinement
const result = await evolveSkill(executableSkill, {
  model: new AIModel(anthropic("claude-sonnet-4-5")),
  maxRounds: 3,
});

When skill_use activates a prompt skill, core enforces its allowed-tools in both the next model schema and runtime dispatch. Omitted/empty lists deny every non-skill_* tool; * is an explicit full-capability grant. The active capability is restored from session history on continuation or resume, so reconnecting cannot silently widen the skill's authority.

Signed production banks

requireSigned: true is fail-closed: it accepts only serialized code skills executed through a SandboxPort. In-process run functions remain supported by the default unsigned/trusted development bank, but cannot be registered in a signed bank because JavaScript closures and their captured state cannot be represented honestly by a portable content signature.

To migrate a signed deployment, move the executable body to code, provide a real sandbox, approve agent-authored skills, then sign the approved lock and attach it with setSignature. SkillBank snapshots definitions and locks at registration, and verifies the stored content digest again before approval, signature attachment, and execution. Test inputs must be structured-cloneable so the asynchronous test gate cannot observe caller mutations after registration begins.

Links

• GitHub
• Issue tracker
• Root README

Apache-2.0