queue-agents

queue-agents

The same agent, offloaded to background workers through a Valkey queue. The web tier returns instantly; the review runs out-of-band.

Guided walkthrough: docs/02-queue-agents.md

browser ─POST /api/reviews─▶ web (producer)
                               └─ create review, XADD job to Valkey, return 202
Valkey stream ──▶ worker (consumer)   [run N of these to scale out]
                    └─ prepareDiff → filterDiff → [reviewers…] → judge (same building blocks as naive-agent)
                    └─ publish progress (pub/sub) ──▶ web ──SSE──▶ browser
                    └─ write telemetry to Postgres

• Render primitives: Web Service + Background Worker + Valkey + Postgres.
• What this unlocks: the web tier returns instantly; long reviews run out-of-band; many reviews run concurrently; scale by adding workers (numInstances); a web redeploy doesn't kill in-flight runs; failed jobs stay un-acked and retry.
• What you now own: the queue, the consumer group, acks, retry semantics, and the progress plumbing. That hand-rolled coordination is exactly what workflow-agents removes.

Run locally

Needs local Postgres and Valkey.

# from the repo root
npm install
createdb agents_workshop
valkey-server &                            # or: docker run -p 6379:6379 valkey/valkey
cp ../../.env.example .env                 # set DATABASE_URL + VALKEY_URL

npm run worker:web                         # terminal A — http://localhost:3000
npm run worker:worker                      # terminal B — start one worker
npm run worker:worker                      # terminal C — add another (scale out)

Submit a PR URL in the UI; watch a worker pick it up. Kill/restart the web service mid-run to see the review survive.

Deploy

render.yaml provisions web + worker + Valkey + Postgres. Bump the worker's numInstances to scale out.