Add webhookUrl terminal callback to /repo/agent and make request state durable

[Evanfeenstra]

What

Long-running /repo/agent consumers (e.g. Hive's workflow_explorer_agent) currently pull results via GET /progress from inside a duration-capped lambda — runs longer than the caller's poll budget lose their result even though the swarm finished fine. This flips delivery from pull to push and hardens the request-state layer.

webhookUrl terminal callback

POST /repo/agent (non-streaming path) accepts an optional webhookUrl in the body. When the run reaches a terminal state, the swarm POSTs:

• success: { request_id, status: "completed", result: { success, final_answer, content, tool_use, usage, sessionId } }
• failure: { request_id, status: "failed", error } (aborts arrive as error: "aborted")

Same shape as GET /progress, with request_id added for correlation. Delivery retries 3× (0s / +5s / +30s, 15s timeout each) and is strictly best-effort — it can never affect the run's recorded status, and /progress polling still works as a fallback. The URL is opaque to the swarm; callers embed their own auth (per-run token / JWT) in it.

Completion-race fix

bus.emit on both terminal paths is now isolated in try/catch, so a post-completion side effect throwing can no longer fall into the promise .catch() and overwrite a just-written completed status with failed.

Durable request state (reqs.ts)

• checkReq reads the .reqs/<id>.json file directly instead of gating on the in-memory META map — a process restart no longer 404s a completed run whose file survived on disk.
• finishReq/failReq write to disk even when the entry was evicted past MAX_REQS, so a long run outliving eviction still records its terminal state.

Testing

• tsc --noEmit clean.
• Two-process smoke test of the restart scenario: process 1 startReq + finishReq, a fresh process (empty META) checkReqs the id — previously undefined (404), now returns the completed result.

🤖 Generated with Claude Code