One Radiant
Thing.
A persistent supervisor for coding agents: every project gets a durable agent that remembers its work.
Models, harnesses and hosts plug in and ORT holds the state. You steer a tree of chats instead of a grid of terminals.
Presence is computed on every read, never stored — idle · queued · working · waiting · needs_input · blocked · failed · offline.
One drop, one light — the glow rolls to the bead that serves. (Illustrative fleet.)
One agent per project
§1ORT is one tree of chats. The ORT chat at its root delegates to each project's main chat, and the project's saved sessions sit beneath it as child chats. An agent is bound to one project folder (agent:<project_id>), and a session's question climbs the tree as a thread card until it reaches a human.
An optional owner agent answers each chat inside an autonomy bound, while an autonomous loop works every onboarded project's goal tree until it is exhausted.
Models, not endpoints
§3.0You think in models, ORT thinks in routes. A canonical model gathers every provider's route to the same model into one chain rail in your order: workbuddy/deepseek-v4.1-flash-1m and the gateway's deepseek-v4.1-flash are two routes, one model. A run's harness, model and host are decided before anything spawns, from one frozen view of the fleet.
A rate limit files a provider hold and demotes the pair. The chain walks on to the next route, then to your fallback models. A switched-off model admits nothing.
Capacity as pools
§3.17Where work may run at once is decided by pools: global, harness, provider, model, host and project. One admission decision is shared by the planner, the dispatch commit and the session guard; occupancy is the sessions creating or busy, and a pool no row bounds is unlimited.
A refusal names the full pool (capacity: provider workbuddy 1/1), and the task stays READY for the next cycle.
Remote nodes, one walk
§3.19A node is an ssh login and Bun 1.3 or later, and nothing of ORT is installed on it. Registering a host starts the walk (probe, provision, handshake, smoke), each step idempotent and recorded, halting on the first failure and resuming where it stopped. Runs ship per session over one ssh channel, or stays resident on an agent host (ort host serve), whose runs survive a control-plane restart.
A node never holds a provider credential; the walk mints it a revocable gateway key of its own, delivered to a 0600 file on stdin. Then it can drain and retire.
Sessions that survive
§3.9Sessions are durable and resumable. Every dispatch writes its runs row before the first event; every event is persisted before it is acted on, so the record outlives the process that made it. Steer a live run, cancel it or close it, and each control call is recorded with its source.
A retried run resumes its harness session rather than restarting. One run, one row, settled once.
One source, three mounts
§3.16 · §1.1The console exists once — Chats, Board, Agents, Hosts, Models, Projects, Settings, Memory, Gateway. The web shell serves it on ORT's own origin; the desktop app, a thin Electron shell under a single-instance lock, loads the same page and starts or reuses the supervisor.
The Harness ui-ort panel mounts a generated mirror pinned by a SHA256SUMS manifest; a drift spec fails when any file differs. No mount keeps a second implementation or a second copy of a fact.
Web console
Desktop app
Harness panel
How it runs
one process, one file of truthdata/supervisor.db, moved by ORT_DB. Every write goes through ORT.:7777. serve ticks on its own interval; tick runs one cycle by hand.AgentAdapter/1, a frozen nine-method contract. Credentials are references only, never stored, never shown.Still on the bench
from TODO.md, marked coming- An ORT SDK: a published client for driving projects, tasks, board moves and runs instead of raw HTTP.coming
- A JSONL API: newline-delimited requests and events for streaming and scripting.coming
- A JSON-RPC 2.0 interface exposing ORT's operations as methods.coming
- The
canonicalpool scope: one cap across every route of a model, in admission and on the lanes.coming
Put a drop on your desk
The supervisor is one Bun process over one SQLite file. Clone, install, point it at a folder and serve — the console is up on loopback with its own token.
Loopback only by default, and a non-loopback bind without a secret refuses to start. State lives in data/supervisor.db — never point two supervisors at the same file. Every call but GET /status is bearer-authenticated.