ADR 0001 — Decouple the agent runtime from the sync fabric

- **Status:** Accepted - **Date:** 2026-07-23

Context

Multiple humans need to watch and steer a running agent in real time. The naive approach — the agent process pushing to browsers directly — turns "live updates for N users" into a bespoke distributed-systems problem (websocket servers, backpressure, late-joiner replay).

Decision

The agent runtime (orchestrator) only ever **writes events to a shared, append-only log**. Clients only ever **read that log**. The sync fabric (Supabase Postgres + Realtime) fans each insert out to connected clients, and a late-joiner simply queries the log for history, then subscribes.

Consequences

- Fan-out to N humans is free — the transcript is just rows. - Late-joiner backfill is free — replay the ordered log. - The orchestrator must be an always-on worker (it cannot run in a serverless function that times out); see [ADR 0002](./0002-worktree-per-agent.md). - Persistence is behind a `PersistenceAdapter` interface, so the fabric is swappable (`PERSISTENCE_PROVIDER`).