captoys/toybox/README.md

# toybox

Vats whose code is installed at runtime: each vat is a Durable Object whose app runs as a [facet](https://developers.cloudflare.com/dynamic-workers/usage/durable-object-facets/) on code loaded through workerd's Worker Loader. Think `vatAdmin.createVat(bundle)`, with a supervisor per vat. Vats have no names: creating one returns its id, which addresses it only on the admin socket. Everything else a vat can do or be reached by is granted explicitly: egress by labels it may fetch within, ingress by labels it is mounted at.

- [supervisor.js](supervisor.js): `Vat` stores code by SHA-256 and runs the app on it; `Directory` holds the mounts; `Public` (the ingress) and `Admin` are the two entrypoints; `Egress` is an egress grant.
- [switchboard.js](switchboard.js): a separate workerd process, the vats' only way out, after [Flower](https://github.com/NuxiNL/flower)'s switchboard: each of its sockets serves a Cap'n Web `Handle` with fixed labels, and `constrain(labels)` returns a narrower one.
- [labels.js](labels.js): Flower's label rule, shared by both: labels only accumulate, a missing key matches anything, and a key both sides set must agree. Ingress lookup is stricter (below).
- [config.capnp](config.capnp): the sockets have no addresses; whoever starts workerd hands them in, so holding a socket is the authority it grants.

| socket | reaches | what its holder can do |
|---|---|---|
| `http` | `Public` | reach the vat mounted at labels the request carries |
| `admin` | `Admin` | `POST /` with an ES module body: create a vat and start it; the reply is `{"vat": <id>, "hash": <code hash>}`. `PUT /<id>` with a body: upgrade it; its storage carries over. `PUT /by-name/<name>`: the same, for a vat from before ids. |

Every start states the vat's grants, as JSON labels in query parameters; absent means none:

- `?egress=<labels>`: what the app may fetch (below).
- `?mount=<labels>`: where the ingress finds it. A mount must conflict with every other one (some shared key, different values), so no request can carry two mounts' labels; an overlapping mount is a 409.

## Ingress

A request's labels are its query parameters; a key given twice is a 400. Routing carries no secrets: a URL ends up in logs and `Referer`s. A secret travels in-band, as on a webkey page: the page keeps it in its fragment, reaches the vat by public labels, and sends it in the body of a call, which the vat checks like a swissnum.

The request reaches the vat whose mount labels it carries, every one with the same value. That is stricter than Flower's match (no conflicting key): an ingress client holds no constrained handle that supplies labels for it, so it reaches a vat only by naming every label of its mount. The path is the app's own. The app sees the mount's and the request's labels together in `Toybox-Connection-Labels` (JSON), as a Flower server gets its connection labels. Responses carry `Referrer-Policy: no-referrer`.

## Egress

An app must export `class App extends DurableObject` (see [apps/](apps/)). Its global `fetch` always throws: nothing in a vat is ambient. A grant arrives as `env.egress`, an `Egress` with the labels in `ctx.props`. Its `fetch`, WebSocket upgrades included, goes through a switchboard handle constrained to those labels; the switchboard refuses the rest, which comes back as a 403. `env.egress.constrain(labels)` returns a narrower `Egress`. It is an object, so the app can pass it to whichever of its parts should have it.

| label | constrains |
|---|---|
| `scheme` | `http` or `https` |
| `host` | host name or address, no port |
| `port` | explicit, or the scheme's default |
| `network` | set by the switchboard socket: `public` or `local` |

The switchboard has two sockets, so two root handles: `public` (no loopback or private addresses) and `local`. toybox holds one, as an external service (`-e switchboard=unix:…`): `pnpm start` and the systemd units give it `public`; the tests give theirs `local` to reach an echo server.

## Run it

```sh
pnpm install
pnpm test         # real workerds: createVat, upgrade, mounts, egress grants and constrain, fetch and WebSocket out, crash and restart
pnpm switchboard  # systemd-socket-activate on ./switchboard-public.sock and ./switchboard-local.sock
pnpm start        # likewise on 127.0.0.1:8787 and ./admin.sock; state in ./state
```

Then, in a third terminal:

```sh
admin() { curl -sg --unix-socket admin.sock "$@"; }
alice=$(admin -X POST --data-binary @apps/counter-v1.js 'http://toybox/?mount={"vat":"alice"}' | jq -r .vat)
curl 'localhost:8787/?vat=alice'          # v1 count 1
admin -X PUT --data-binary @apps/counter-v2.js "http://toybox/$alice?mount={\"vat\":\"alice\"}"
curl 'localhost:8787/?vat=alice'          # v2 count 2

admin -X POST --data-binary @apps/egress-probe.js 'http://toybox/?mount={"vat":"erin"}&egress={"host":"example.com"}'
curl 'localhost:8787/?vat=erin&url=https://example.com/'   # got: <!doctype html>…
curl 'localhost:8787/?vat=erin&url=https://example.org/'   # got: egress refused: egress to example.org not granted
```

## Under systemd

[systemd/](systemd/) has socket and service pairs for toybox and its switchboard, assuming this directory is installed at `/opt/toybox`. toybox gets no network of its own (`PrivateNetwork=yes`): only the two sockets it is handed and the switchboard's `public` socket, which only members of the `toybox` group can open. The switchboard binds its `local` socket in its own 0700 runtime directory, so no other unit holds that handle.

Neither unit has been run under the system manager yet.

`PrivateNetwork=` needs the system manager; as a `systemd-run --user` unit, systemd logs "lack privileges for network namespace, proceeding without".