Linear Agent Bridge

linear-agent-bridge

Run your Claude Code setup as a Linear agent. Delegate an issue to it or message it in an agent session, and a Claude Agent SDK session runs on your machine, in a working directory you choose, with everything that directory carries: CLAUDE.md instructions, MCP servers, skills. Replies land in the issue's agent-session thread.

This is a compact reference implementation (no framework, tested). It is not a coding agent; for assign-an-issue-get-a-PR flows, see [Cyrus](https://github.com/ceedaragents/cyrus). This bridge is for talking to an agent that knows your context: a knowledge base, an ops repo, a project directory.

Runs on Claude Code subscription auth. No Anthropic API key.

Architecture

Linear (mention / delegate / follow-up prompt)
  -> webhook + reconciliation: AgentSessionEvent / AgentSession activities
  -> src/server.ts        verify HMAC, persist ingress, ack < 5s
  -> src/state/store.ts   durable receipt + semantic execution claim
  -> src/queue.ts         per-session FIFO lanes, concurrent across sessions
  -> src/runtime/claude.ts  Agent SDK query(), cwd = KB_PATH, resume
  -> src/linear/client.ts   agentActivityCreate (thought/action/response)

Session mapping (Linear session id -> SDK session id), bounded webhook state, and Linear's rotating OAuth token pair persist in JSON files. Follow-up prompts resume the same conversation, webhook retries do not dispatch the same turn twice, accepted work survives a process exit before dispatch, missed prompt and stop webhooks recover through reconciliation, and access refreshes without another browser authorization.

For each valid agent event, the bridge durably persists a receipt keyed by Linear's `webhookId` and a semantic execution claim before returning 200. Created turns use `created:` as the execution identity; prompted turns use `agentActivity.id`. A claim left active by a process crash is reclaimed by a replacement process only when dispatch never st