runskein

One TypeScript API for coding-agent engines — process lifecycle, sessions, permissions, and a single persistent transcript format — with engines integrated as auto-discovered adapters over the Agent Client Protocol (which is never exposed to you).

npm install runskein

Node.js 22 or newer, ESM only. This installs no engine: runskein drives OpenCode, Kimi Code, Claude Code, Codex or pi already on your PATH, so install at least one and log in first.

Quickstart

import { createHub, policies } from 'runskein';

const hub = createHub(); // bundles the opencode / kimi / claude-code / codex / pi adapters

// What's available on this machine? (cheap; never spawns)
console.table(await hub.engines());

// One conversation on one engine (spawns on demand, engines share processes)
const session = await hub.session({
  engine: 'opencode',
  cwd: process.cwd(),
  permissionPolicy: policies.allowAll, // the default; denyAll / rules([...]) / your own
});

session.on('update', (e) => {
  if (e.update.sessionUpdate === 'agent_message_chunk' && e.update.content.type === 'text') {
    process.stdout.write(e.update.content.text);
  }
});

const result = await session.prompt('Summarize this repository.');
console.log('\nturn ended:', result.stopReason, `${result.durationMs}ms`);

// The transcript persisted itself (.transcripts/ by default; sqliteStore available)
for await (const event of session.transcript()) {
  // {seq, ts, sessionId, engineId, update} — ACP vocabulary, runskein envelope
}

// Resume later — native when the engine supports it, transcript-digest
// rebuild when it doesn't; the session id stays the same either way:
const resumed = await hub.session({ engine: 'opencode', cwd: process.cwd(), resume: session.id });
console.log(resumed.resumeTier); // 'native' | 'load' | 'rebuilt'

await session.close();
await hub.quit();

The default policy is allowAll, so the agent may read and write files under the cwd you pass. Point a first experiment at a directory you can throw away.

Adding an engine

Drop a directory with a runskein.adapter marker in <cwd>/adapters/<id>/ (or publish a runskein-adapter-<id> package) whose default export follows the adapter spec — no core or consumer changes. See the adapter guide and the bundled adapters/* packages for reference.

Docs

• README — the three-minute start, and why this exists
• API specification — the frozen v1 surface
• Capability matrix — capability tiers and what each engine measured
• Architecture — how the pieces fit together
• Adapter guide — writing an adapter
• Measured matrix — the probe output the tables are drawn from

Questions, or an engine behaving differently from the matrix? Open an issue.