Stream Chat's developer experience, your infrastructure.
Your backend mints user tokens, the browser talks straight to the chat backend, the backend authorises every operation. Channels, members, reactions, read receipts, typing, threads, flags - the same model, the same method names where they were free to keep.
No Redis, no Kafka, no broker to babysit. The event log, the webhook queue and the metrics rollups all live in Postgres, migrated by Flyway on boot.
@firemoot/test starts the real image plus Postgres through Testcontainers, seeds a fixture and hands back a driver. Your integration tests run against the actual server, not a mock.
Subscribe with the last seq you saw and the server replays everything after it, then splices in live events by seq. No gap, no duplicate - proven by a chaos test that severs the TCP connection mid-stream.
@firemoot/core is generated from the tapir endpoints and CI fails on drift, so the types cannot lie. @firemoot/client adds optimistic send, a channel-state reducer and read receipts on top.
Active users, messages per day by channel type, concurrent connections and storage growth, plus API key rotation and one-click replay of dead-lettered webhooks. Served from the same binary at /admin.
Every persisted channel event is fanned out to your endpoints, HMAC-signed and retried on a 1m/5m/30m/2h ladder. What still fails is dead-lettered rather than dropped, and you can requeue it from the dashboard.
Presigned uploads against any S3-compatible store - MinIO, Tigris, Garage, SeaweedFS, AWS. Thumbnails are generated in the background and patched onto the attachment via message.updated.
The nightly soak holds the container to a 1GiB cap and gates on both latency and memory. Latest run - p95 delivery latency 48ms at 290MiB peak RSS, with zero failed checks.
