ORT supervisor
Persistent supervisor · Bun + SQLite · zero cloud lock-in

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.

$ bun run src/main.ts serve 7777
~1,600 routes discovered live
6 admission pool scopes
1 file SQLite state of truth
3 mounts one shared console
fleet observatory · click any bead
ORT · owner chat
live
root supervisor · one drop, one light (illustrative fleet)
Presence is computed from live work on every read — never stored. working: radiant light — the supervisor spends its glow only on active execution.
§1Chat tree & escalation

One agent per project

ORT 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.

Optional owner agent Answers non-command chat inside a strict autonomy bound so routine clarifications never stall a fleet.
Autonomous goal-tree loop On by default per onboarded project — works the repository's goal tree tick by tick until it is exhausted.
chat-tree · directives ↓ · escalations ↑
thread card · s-41 → invoices → ORT needs_input

“The old fixtures name a field the schema dropped — keep legacy alias or rename?”

awaiting decision · click “answer via owner agent” above to resolve inside autonomy bound

click any node to inspect binding (illustrative) agent:<project_id>
§3.0 · §3.4Canonical routing

Models, not endpoints

You 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.

~1,600 routes discovered live 1 chain per canonical model route before spawn
route-chain · owner priority order
gpt-5.2-codex via orca-router · serving
claude-sonnet via llm-gateway · ready
deepseek-v4.1-flash provider hold · rate-limited
local-7b provider off
handoff owner's fallback models
frozen tick plan → harness: pi-durable · route: orca-router/gpt-5.2-codex · host: vps1 admit: READY → SPAWN
a switched-off model admits nothing · rate limits demote the provider/model pair inv 58 · frozen tick view
§3.17Shared admission

Capacity as pools

Where work may run at once is decided by six pool scopes. One admit(run) path is shared by the planner, the dispatch commit, and the session guard.

Occupancy is sessions creating or busy; any pool no row bounds is unlimited. A refusal names the full pool and leaves the task READY.

admit(run) · pool sockets
global fleet 5 / 8 occupied
decision: admit — 3 open sockets in global fleet (click “+ admit run” or switch to provider)
inv 4 · one admission decision occupancy = creating + busy
§3.9Persist before act

Sessions that survive

Every dispatch writes its runs row before the first event, and every event is persisted before it is acted on — 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.

session s-41 · run ledger
#01 you · directive — port the invoice parser to the new schema, keep tests green src: human
#02 session s-41 · working — 3 files touched, test suite running… runs: r_41a
#03 thread · needs_input — old fixtures name a dropped field: keep or rename? escalated
one run, one row, settled once · timeline reads the same record the console shows
resumeSession(), never restart status: working
§3.19Three-plane execution

Remote nodes, one walk

A node is an ssh login and Bun 1.3 or later — nothing of ORT is installed on it. Registering a host starts the idempotent walk: probe, provision, handshake, smoke, halting on the first failure and resuming where it stopped.

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.

Profile A · Shipped

Per-session SSH channel

Runs ship per session over a single SSH channel. Zero resident daemon; clean teardown when the run settles.

Profile B · Resident

ort host serve

Stays resident on the remote agent host — in-flight runs survive a control-plane restart untouched.

host:vps1 · onboarding walk
stage 6/6 · enabled — smoke-passed (16-token run OK); host vps1 admits work with its minted 0600 gateway key. lifecycle: active
each step idempotent & recorded · click any stage or simulate a halt to test resume 0 provider creds on node
§3.16 · §1.1Zero-drift console

One source, three mounts

The console exists once — nine pages served from one canonical source. No mount keeps a second implementation or a second copy of a fact.

Select any page below to see all three mounts follow the same route in lockstep.

localhost:7777/app/chats

Web console

The agent shell served directly by ort serve on loopback with its write token embedded.

origin: 127.0.0.1:7777
ORT.app · /chats

Desktop app

A thin Electron shell under a single-instance lock that starts or reuses the local supervisor, tray included.

Electron · Linux AppImage/deb
dsh · ui-ort/chats

Harness panel

The DeepSeek Harness ui-ort panel mounts a generated mirror pinned by SHA256SUMS; a drift spec fails on any diff.

inv 10 · manifest verified
RuntimeOne process, one file of truth

How it runs

Built to sit quietly on your workstation or home server — zero container sprawl, zero hidden credential vaults.

Runtime

Single Bun process

Bun ≥ 1.3 and TypeScript, one process. No Node.js or Python needed for the supervisor itself.

State of truth

One SQLite database

Everything lives in data/supervisor.db, moved with ORT_DB. Every write goes through ORT.

API & Loop

Serve or tick by hand

HTTP API and console on :7777. serve auto-ticks on its interval; tick runs one cycle manually.

Harness boundary · AgentAdapter/1

Frozen nine-method contract

Harnesses plug in behind a frozen contract. pi-durable is the default under the launcher and desktop app, strictly opt-in — ORT core imports no substrate module.

identity health discoverModels discoverUsage createSession resumeSession run steer cancel
Memory, Credentials & Dogfood

Shared memory, reference-only secrets

Owner memory is its own canonical store serving every agent and chat. Credentials are references only (an env var, file, or login) — never stored, never shown. ORT is its own dogfood project, working this repository's TODO.md.

QuickstartSource install

Put a drop on your desk

Clone, install, point it at a folder and serve — the console comes up on loopback with its own 0600 bearer token.

# needs Bun ≥ 1.3 and nothing else
git clone git@github.com:KinoThe-Kafkaesque/ORT.git
cd ORT && bun install
# onboard a folder, queue a READY task, serve API + console
bun run src/main.ts add-project .
bun run src/main.ts task ORT "My first task"
bun run src/main.ts serve 7777
# open http://localhost:7777
# Electron is pinned as its own dev dependency inside apps/desktop
cd ORT/apps/desktop && bun install
# starts or reuses the local supervisor under a single-instance lock
cd .. && bun run app
# package Linux x64 AppImage + deb: bun run app:package
# Profile B: serve this machine's harnesses to another ORT on 127.0.0.1:7790
bun run src/main.ts host serve
# runs stay resident and survive a control-plane restart
# Profile A nodes need only SSH + Bun ≥ 1.3 (zero ORT install)
Loopback by default

serve binds 127.0.0.1. A non-loopback ORT_BIND without an auth secret refuses to start.

Bearer-authenticated

Every call except GET /status requires a bearer token generated at 0600 beside the database.

One supervisor per DB

State lives in data/supervisor.db. Never point two supervisors at the same file — use ORT_DB for isolation.

TODO.mdOpen items

Still on the bench

Planned surfaces and pool scopes currently tracked in the repository's own roadmap.

Phase 28coming

ORT SDK

A published client library for driving projects, tasks, Board moves, runs and events instead of raw HTTP.

Phase 28coming

JSONL streaming API

Newline-delimited requests and event streams for headless scripting and live pipelines against ORT.

Phase 28coming

JSON-RPC 2.0 interface

Exposing ORT's supervisor operations as typed RPC methods for editor and tool integrations.

Phase 29coming

canonical pool scope

One capacity cap across every provider route of a canonical model, enforced in admission and on the lanes.