marifold 0.67.0 → 0.70.5

package/dist/bridge-template/.env.example

@@ -0,0 +1,6 @@
+# Use a TCP Redis endpoint. A REST-only Redis URL is not compatible.
+# Use TLS (rediss://) for a remote Redis service. Do not commit real values.
+MARIFOLD_BRIDGE_REDIS_URL=rediss://default:REPLACE_ME@redis.example.com:6379
+# Generate a random secret with at least 32 characters. This registers hosts;
+# joining a workspace uses a separate, single-use host invitation.
+MARIFOLD_BRIDGE_REGISTRATION_TOKEN=REPLACE_WITH_A_RANDOM_SECRET_OF_AT_LEAST_32_CHARACTERS

package/dist/bridge-template/HOSTING.md

@@ -0,0 +1,477 @@
+# Bridge hosting guide
+
+Updated September 11, 2026. These are manual deployment recipes for the existing
+bridge. The standalone package has passed local smoke tests; Vercel, Cloudflare,
+AWS, and Aliyun live deployments remain unverified. Review current provider plans,
+limits, regions, and costs before creating resources.
+
+Only deploy the bridge package. Never upload the parent `.marifold` directory,
+profiles, model credentials, or the local Marifold application service. The bridge
+is an authenticated encrypted relay; the personal application service remains on
+the owner's device.
+
+## Choose a deployment
+
+For a Linux VM, start with the [guided installer](#guided-linux-installer).
+The manual systemd recipe below remains available for installations without Docker.
+
+| Provider | Recipe for the current code | Additional infrastructure |
+| --- | --- | --- |
+| Vercel | Prepared Node function project | TLS TCP Redis |
+| Cloudflare | Named Tunnel to a persistent Node service | Always-on Linux server, domain, TCP Redis |
+| AWS | EC2 Linux instance with systemd and Caddy HTTPS | Domain, persistent TCP Redis |
+| Aliyun / Alibaba Cloud | ECS Linux instance with systemd and Caddy HTTPS | Domain, persistent TCP Redis |
+
+Cloudflare Tunnel provides the public HTTPS endpoint; it does not host the Node
+process. The current package contains no Workers or Containers deployment adapter.
+Those products would need separate integration and validation. Similarly, these
+AWS and Aliyun recipes use virtual machines, not Lambda or Function Compute.
+
+Choose a region and hostname reachable from the owner's actual devices. A provider
+name does not guarantee reachability from mainland China. Test both HTTPS and
+paired WebSocket traffic from home and office before relying on a deployment.
+
+## Guided Linux installer
+
+On Aliyun ECS, AWS EC2, or another Linux server with systemd, install Docker Engine
+and its Compose plugin using the [official distribution instructions](https://docs.docker.com/engine/install/).
+Python 3 is also required. The installer checks these prerequisites; it does not
+replace existing Docker packages, change cloud security groups, or configure DNS.
+The server needs access to container/package registries for the first build.
+
+If you pulled the Marifold repository onto the server:
+
+```sh
+pnpm install --frozen-lockfile
+pnpm -r build
+node packages/cli/dist/index.js workspace bridge install
+```
+
+If the `marifold` command already points to that build, the final command is simply:
+
+```sh
+marifold workspace bridge install
+```
+
+The CLI requests sudo when necessary and runs the installer on **this server**.
+You do not need to start the personal Marifold application service on the ECS.
+Alternatively, upload just a prepared bridge package (excluding `.env`, `.vercel`
+and `node_modules`), enter that directory on the server, and run:
+
+```sh
+sudo bash setup.sh
+```
+
+The standalone route does not require Node or pnpm on the VM: the bridge runtime
+and dependency installation run inside the Node container. Its image build uses
+npm independently of your development package-manager choice.
+
+The wizard asks for:
+
+1. **Redis:** create a dedicated persistent Redis on loopback port 32144 (default),
+   or enter an existing Redis URL through a hidden prompt. Existing Redis remains
+   unchanged; its credentials and permissions must already be configured. Startup
+   checks connectivity, not the full persistence/ACL policy. If using the server's
+   existing Redis, `127.0.0.1:6379` works because these Linux containers use host
+   networking.
+2. **HTTPS:** create Caddy configuration for a domain you control, or keep an
+   existing HTTPS proxy/Cloudflare Tunnel. For Caddy, point DNS at this server and
+   allow TCP 80/443 in the cloud security group. Existing ingress should forward
+   to `http://127.0.0.1:32143` on the same server.
+3. **Review and start:** confirm the destination and selected services. The wizard
+   generates secrets, builds/starts containers, waits for bridge HTTP and Redis
+   connectivity checks, and prints the next steps. It does not print secrets.
+
+The installation lives at `/opt/marifold-bridge` with owner-only access:
+
+| File/resource | Purpose |
+| --- | --- |
+| `compose.json` | Named project and restart policies |
+| `bridge.env` | Protected bridge connection URL and generated registration token |
+| `registration-token` | Protected copy of the token for the host creation prompt |
+| `package/` | Compiled bridge inputs and container build files; no personal state |
+| `redis.conf` | Dedicated Redis settings, only when selected |
+| Named Redis volume | Persistent AOF/snapshot data; never automatically deleted |
+| Named Caddy volumes | Certificate/configuration state, only when selected |
+
+Read the generated registration token **locally**, then enter it on your host Mac:
+
+```sh
+sudo cat /opt/marifold-bridge/registration-token
+```
+
+Docker is enabled at boot and containers use `restart: unless-stopped`. They return
+after reboot unless deliberately stopped. Dedicated Redis uses AOF with every-second
+fsync and a no-eviction policy with a 128 MiB data limit; provision monitoring and
+backups before relying on it. Persistence is not a backup. Other projects' Redis
+processes and configuration are left alone.
+
+Management commands:
+
+```sh
+sudo docker compose -f /opt/marifold-bridge/compose.json ps
+sudo docker compose -f /opt/marifold-bridge/compose.json logs --tail 50 bridge
+sudo docker compose -f /opt/marifold-bridge/compose.json restart bridge
+```
+
+After editing `bridge.env`, use `up -d` to recreate containers with changed values;
+`restart` alone does not reload container environment configuration. Review logs
+locally before sharing them; never print `docker compose config` without `--quiet`
+because expanded configuration may contain secrets.
+
+If startup fails, the installer exits with failure and keeps generated files and
+data. Fix the connection, registry access, or port problem and retry:
+
+```sh
+marifold workspace bridge install --start
+# Or, from the prepared package:
+sudo bash setup.sh --start
+```
+
+This reuses the original configuration/token. A normal install refuses an existing
+directory, containers, or volumes instead of overwriting them. It does not migrate
+an existing manual systemd deployment or upgrade an older installed package.
+Never use `docker compose down -v` unless you intend to destroy Redis and certificate
+volumes. Public HTTPS and real pairing remain separate acceptance checks after
+local health succeeds.
+
+Installer validation covers generated configuration, permissions, input rejection,
+secret exclusion, collision protection and startup command sequencing in disposable
+fixtures. Docker execution, reboot recovery and live certificates still require
+the first Linux server trial; they have not been verified on this development Mac.
+Developers can run the installer fixtures with
+`python3 -B -m unittest discover -s apps/bridge/setup -p 'test_*.py'` from the repository.
+
+## Common package and Redis prerequisites
+
+Prepare once on the development Mac, or use your existing prepared directory:
+
+```sh
+marifold workspace bridge prepare ~/.marifold/bridge
+cd ~/.marifold/bridge
+npm install --ignore-scripts
+```
+
+Preparation refuses an existing destination; do not rerun it over a prepared
+package. It includes compiled `dist/`, `vendor/`, a manifest, and Vercel files.
+Linux deployments use `node dist/serve.js` and ignore `vercel.json` and the
+Vercel-specific API entry point. Install dependencies on the destination OS;
+do not transfer your Mac's `node_modules`.
+
+Use a Redis service that supports the current `ioredis` client's TCP commands,
+Lua scripts, and pub/sub. A REST-only endpoint or an unvalidated cluster endpoint
+is insufficient. Start with a dedicated single-primary Redis-compatible endpoint;
+the bridge does not configure Redis Cluster discovery. Enable persistence,
+non-evicting storage, and backups: host identities and revocations must survive
+restarts. Use TLS (`rediss://`) for remote Redis and allow access only from the
+bridge's network where possible. Never expose an unauthenticated Redis port.
+
+All deployments require:
+
+```dotenv
+MARIFOLD_BRIDGE_REDIS_URL=rediss://default:REPLACE_ME@redis.example.com:6379
+MARIFOLD_BRIDGE_REGISTRATION_TOKEN=REPLACE_WITH_A_RANDOM_SECRET
+```
+
+Generate a separate registration secret using `openssl rand -hex 32`. Store it
+securely; it registers hosts and is not the guest invitation. Percent-encode
+reserved characters in Redis URL credentials. Do not commit real environment
+files or copy secrets into support messages.
+
+## Vercel
+
+### Using pnpm instead of npm
+
+The prepared package can use pnpm. Choose one package manager for this deployment
+directory. If switching from npm, move its `package-lock.json` and `node_modules`
+aside before reinstalling; do not deploy two competing lockfiles.
+
+```sh
+cd ~/.marifold/bridge
+pnpm install --ignore-scripts
+pnpm --version
+```
+
+Keep the resulting `pnpm-lock.yaml` and use the same exact pnpm version in
+deployment. For example, with pnpm **11.17.0**, replace only `installCommand` in
+the prepared `vercel.json` with:
+
+```json
+"installCommand": "npx --yes pnpm@11.17.0 install --frozen-lockfile --ignore-scripts"
+```
+
+This is a property to edit in the existing JSON object, not a replacement for the
+whole configuration. Match the version to the one used to create your lockfile.
+The explicit version avoids relying on Vercel's preinstalled pnpm default; the
+build needs registry access to fetch that version. Preserve all routing/function
+settings. On a Linux VM with that pnpm version installed, replace `npm ci` with
+`pnpm install --frozen-lockfile --ignore-scripts` and transfer `pnpm-lock.yaml`
+instead of `package-lock.json`. The runtime command remains `node dist/serve.js`.
+
+See [pnpm install](https://pnpm.io/cli/install) and
+[Vercel package-manager selection](https://vercel.com/docs/package-managers).
+
+### Deploy the project
+
+Use the [complete project configuration in README](README.md#standalone-project-configuration).
+The short sequence, after preparing the package, is:
+
+1. Run `npx vercel link` inside the prepared directory. Select the account and
+   create/select the bridge project; no Git repository is required.
+2. Provision TLS TCP Redis and add the two variables above to the project's
+   Production environment settings.
+3. Select Other as the framework, Node 24.x, and enable Fluid compute. Keep the
+   generated empty build command, install command, function duration, and rewrites.
+4. Ensure Deployment Protection permits machine access to `/health`, `/v1/hosts`,
+   and `/v1/connect` without an interactive login.
+5. Run `npx vercel --prod`, then use the stable production project URL for pairing.
+   Redeploy after changing deployment environment variables.
+
+WebSocket connections can end at function duration limits and reconnect to another
+instance, so external Redis coordination is required. See
+[Vercel WebSockets](https://vercel.com/docs/functions/websockets).
+
+## Shared Linux service for AWS, Aliyun, and Cloudflare origins
+
+Use a dedicated Ubuntu/Debian-style Linux VM with systemd. Install Node.js **24.x**
+and npm using a supported [Node installation method](https://nodejs.org/en/download).
+Use a system-wide executable, not an interactive user's shell-only version-manager
+setup. Check `node --version` and `command -v node`.
+
+Transfer only the prepared package to a new deployment directory, such as
+`/opt/marifold-bridge`. Exclude `.env`, `.env.*`, `.vercel`, and `node_modules`;
+keep `.env.example` as a template if wanted. Retain `package-lock.json`, then run
+`npm ci --ignore-scripts` on the VM as the deployment user. The deployment user
+owns installation; the runtime account only needs read access to the package.
+
+On a fresh VM, create a dedicated runtime account and environment file:
+
+```sh
+sudo useradd --system --user-group --no-create-home --shell /usr/sbin/nologin marifold-bridge
+sudo install -m 600 /dev/null /etc/marifold-bridge.env
+sudoedit /etc/marifold-bridge.env
+```
+
+These creation commands are for a fresh installation; preserve existing accounts
+and environment files during updates. Fill the file using your actual values:
+
+```dotenv
+MARIFOLD_BRIDGE_REDIS_URL=rediss://default:REPLACE_ME@redis.example.com:6379
+MARIFOLD_BRIDGE_REGISTRATION_TOKEN=REPLACE_WITH_YOUR_GENERATED_SECRET
+HOST=127.0.0.1
+PORT=32143
+```
+
+The Node entry point does not automatically read `.env`. This service unit loads
+the protected environment file. Create
+`/etc/systemd/system/marifold-bridge.service`:
+
+```ini
+[Unit]
+Description=Marifold personal workspace bridge
+Wants=network-online.target
+After=network-online.target
+
+[Service]
+Type=simple
+User=marifold-bridge
+Group=marifold-bridge
+WorkingDirectory=/opt/marifold-bridge
+EnvironmentFile=/etc/marifold-bridge.env
+ExecStart=/usr/bin/node /opt/marifold-bridge/dist/serve.js
+Restart=on-failure
+RestartSec=5
+TimeoutStopSec=20
+NoNewPrivileges=true
+PrivateTmp=true
+ProtectHome=true
+ProtectSystem=strict
+
+[Install]
+WantedBy=multi-user.target
+```
+
+Replace `/usr/bin/node` with the system-wide path verified above if different.
+Ensure the runtime user can traverse/read the deployment directory. Then:
+
+```sh
+sudo systemctl daemon-reload
+sudo systemctl enable --now marifold-bridge
+sudo systemctl status marifold-bridge
+curl --fail http://127.0.0.1:32143/health
+```
+
+Health identifies the process; pairing is still needed to verify Redis delivery.
+Check service logs locally when diagnosing errors and redact credentials before
+sharing them. Keep port 32143 private. Choose **one** ingress method below:
+Caddy for direct HTTPS, or Cloudflare Tunnel on the same VM.
+
+### Direct HTTPS with Caddy
+
+Point a domain such as `bridge.example.com` at the VM's stable public IP. Install
+Caddy with its official system service. Configure `/etc/caddy/Caddyfile`:
+
+```caddyfile
+bridge.example.com {
+    reverse_proxy 127.0.0.1:32143
+}
+```
+
+Replace the example hostname. Allow inbound TCP 80 and 443 for this recipe, and
+ensure DNS resolves correctly for certificate issuance. Leave 32143 and Redis
+closed to public inbound traffic. Caddy handles HTTPS and reverse proxying; see
+the [reverse proxy guide](https://caddyserver.com/docs/quick-starts/reverse-proxy)
+and [service setup](https://caddyserver.com/docs/running#using-the-service).
+
+```sh
+sudo caddy validate --config /etc/caddy/Caddyfile
+sudo systemctl reload caddy
+curl --fail https://bridge.example.com/health
+```
+
+Use the whole hostname for this relay, without stripping `/v1/connect` or adding
+a path prefix. Avoid browser login gates, interactive bot challenges, and caching
+on bridge endpoints. Exercise WebSocket pairing after configuring the proxy.
+
+## AWS EC2
+
+1. Create a Linux EC2 instance in the chosen region, with persistent disk and a
+   stable public address for direct HTTPS. Use a dedicated SSH key or your existing
+   approved administrative access. Review instance, disk, IP, and transfer charges.
+2. Configure its security group: admin access only from your management network;
+   TCP 80/443 for Caddy ingress; no public 32143 or Redis port. Keep outbound access
+   needed for DNS, package installation, certificates, and the Redis endpoint.
+   [EC2 security groups](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-security-groups.html)
+3. Provision a dedicated compatible Redis endpoint in the same region, preferably
+   on a private network reachable from the instance. Enable TLS/authentication,
+   persistence and backups; verify service command/endpoint compatibility before
+   selecting a managed Redis offering. Do not assume all serverless cache products
+   are interchangeable with this client's single-primary connection.
+4. Install the shared Linux service, point your DNS hostname at the stable address,
+   and configure Caddy as above. Keep the environment file outside the package.
+5. Perform the acceptance checklist below, including an instance/service restart.
+
+An AWS load balancer is optional and not required for this one-owner initial
+recipe. Adding one requires separate TLS, WebSocket timeout, and cost review.
+
+## Aliyun / Alibaba Cloud ECS
+
+1. Select a Linux ECS instance, persistent disk, region, public bandwidth, and
+   stable public IP/EIP appropriate to your devices. Review the region's domain
+   and hosting prerequisites before purchase. For mainland hosting, consult
+   [Alibaba Cloud's ICP filing guidance](https://www.alibabacloud.com/help/en/icp-filing/basic-icp-service/product-overview/what-is-an-icp-filing)
+   and confirm the requirements for your intended endpoint with the provider.
+2. Configure the ECS security group for restricted administrative access and
+   Caddy TCP 80/443. Keep the Node port and Redis private. See
+   [ECS security group rules](https://www.alibabacloud.com/help/en/ecs/user-guide/security-group-rules).
+3. Provision persistent compatible TCP Redis in the same region/VPC. Check TLS,
+   authentication, non-eviction, backups, Lua/pub/sub support, and endpoint mode.
+   A managed Redis-compatible product still needs this compatibility check.
+4. Install the shared Linux service and Caddy. Configure DNS for the endpoint and
+   complete applicable provider domain setup before certificate issuance/testing.
+5. Test from the actual mainland networks and devices, then complete acceptance.
+   Direct is the intended connection mode when this endpoint is reachable.
+
+This recipe does not use Aliyun Function Compute or require Cloudflare in front
+of Aliyun. The same Node service layout can be adapted to another Linux VM
+provider, with that provider's firewall, DNS, and storage setup.
+
+## Cloudflare Tunnel
+
+This recipe needs an always-on server running the shared Linux service. It can
+be an AWS/Aliyun VM or another server. Cloudflare supplies the public endpoint and
+outbound tunnel, while that server supplies compute and reaches Redis.
+
+1. Add/control the intended domain in Cloudflare and install the bridge service
+   on the origin VM. Confirm local health on `127.0.0.1:32143`.
+2. Create a named Cloudflare Tunnel in the dashboard. Install `cloudflared` on the
+   **same VM** using the dashboard's OS-specific connector instructions; run it
+   as a persistent service. Treat its connector token as a secret.
+3. Add a published application route for `bridge.example.com`, service type HTTP,
+   service URL `localhost:32143`, with no path restriction. Cloudflare terminates
+   public TLS; the origin HTTP hop is local to the VM. Do not use the private
+   Marifold application service as the tunnel origin.
+4. Keep inbound Node/Redis ports closed. Allow the connector's documented outbound
+   traffic. An origin using only Tunnel does not need public inbound 80/443 or a
+   Caddy installation. Retain your separate administrative access.
+5. Ensure the bridge hostname allows noninteractive HTTP/WebSocket clients. The
+   current client does not supply Cloudflare Access service-token headers and
+   cannot complete browser challenges. Scope any exception to this relay hostname;
+   do not remove protections from unrelated applications.
+6. Verify public health, pair two devices, and restart the connector to exercise
+   reconnect. Use the stable hostname rather than a temporary Quick Tunnel URL.
+
+Follow the [official named-tunnel setup](https://developers.cloudflare.com/cloudflare-one/networks/connectors/cloudflare-tunnel/get-started/create-remote-tunnel/).
+Cloudflare proxies WebSockets but connections can close during network updates;
+see [WebSocket behavior](https://developers.cloudflare.com/network/websockets/).
+This route does not guarantee reachability from every mainland network.
+
+## Local device proxies
+
+Bridge proxy support is **planned, not implemented**. The intended setting is
+local to each device's workspace connection: Direct by default, with an optional
+explicit proxy applied consistently to registration, pairing and WebSocket
+reconnects. Home and office may use different values for the same endpoint.
+
+This is independent of hosting provider and model-provider proxies. Never put a
+Mac's proxy address in the server environment file: `127.0.0.1` there means the
+server. Do not synchronize local proxy credentials. CLI/Web create/join and later
+connection editing should expose the setting when implemented. For now a VPN/TUN
+that routes the service process may supply connectivity; provider proxy settings
+do not configure workspace bridge traffic.
+
+## Acceptance, updates, and recovery
+
+For an installer-managed Linux deployment, pull and build Marifold on the server,
+then run the updater:
+
+```sh
+cd ~/repos/marifold
+git pull --ff-only
+pnpm install --frozen-lockfile
+pnpm -r build
+node packages/cli/dist/index.js workspace bridge update
+```
+
+With a linked CLI, the last command is `marifold workspace bridge update`. For a
+freshly prepared standalone package, use `sudo bash setup.sh --update` instead.
+`workspace bridge install --start` remains a restart/recovery command, not an upgrade.
+
+The updater targets the existing `/opt/marifold-bridge` installation. It builds a
+new release while the current container runs, preserves the installed Dockerfile
+(including a China registry mirror), and replaces only the bridge container. It
+preserves the environment file, registration token, Redis container/data and Caddy
+configuration. Connections briefly reconnect. Failed deployment or health checks
+trigger a rollback to the retained previous image. Build failures leave the running
+container and active configuration intact. Release directories and rollback
+configuration remain under `/opt/marifold-bridge`; the updater prints their location.
+This updater does not manage manual systemd or Vercel deployments.
+
+The v0.70.3 relay prevents repeated delivery of unacknowledged packets within a
+connection. Updated devices enable concurrent chunks only when both relay
+connections support this behavior. Upgrade the ECS bridge as well as both Macs
+to enable the throughput improvement; older relays retain sequential transfers.
+After updating, `/health` includes `"deliveryReplay":"on-reconnect"`.
+
+For every provider, verify before relying on the bridge:
+
+- HTTPS health responds without a login/challenge; then authenticated registration
+  and pairing succeed. Health alone is not a Redis or end-to-end execution test.
+- The guest reads shared host data and receives a host-authenticated model answer.
+- An approved file write reaches the selected device; its artifact downloads.
+- Separate sessions run concurrently; the same session rejects competing runs.
+- Connection/service/ingress restart recovers without duplicate side effects.
+- Host offline behavior, guest executor opt-out, and device revocation work.
+- Redis metadata survives a planned restart and has a recovery procedure. Test
+  backup/restore on disposable data before applying it to a real workspace.
+
+Retain the prior package for rollback. For a Linux update, stage dependencies in
+a separate release directory, stop/switch/restart the service during a planned
+interruption, and preserve the environment file and Redis data. For Vercel, retain
+the prior deployment and environment configuration. Recheck pairing and reconnect
+after a change. Do not flush Redis as a troubleshooting shortcut: losing host and
+revocation metadata can break existing workspaces.
+
+The remaining live acceptance work is shared with the repository's
+`docs/workspaces.md`. No provider resources are created by preparing this package
+or by saving this guide.

package/dist/bridge-template/README.md

@@ -0,0 +1,176 @@
+# Personal workspace bridge
+
+This package routes encrypted messages between one owner's Marifold devices. It
+contains no provider credentials, profile database, model endpoint, or agent tools.
+The host service must stay online. Redis stores host public identities, revocations,
+connection generations and encrypted delivery queues; it is required for production.
+
+See [Hosting options and setup](HOSTING.md) for Vercel, Cloudflare Tunnel, AWS EC2,
+and Aliyun ECS recipes, local device proxy design, and deployment acceptance.
+
+For a guided Linux server install, run `marifold workspace bridge install`, or
+`sudo bash setup.sh` inside a prepared package on that server. The
+[installer guide](HOSTING.md#guided-linux-installer) covers Docker prerequisites,
+Redis/HTTPS choices, automatic startup, and token retrieval.
+
+## Vercel test deployment
+
+1. Use `marifold workspace bridge prepare ./marifold-bridge` to create a standalone
+   deployment directory. Review its README, environment template and `vercel.json`.
+   It contains compiled bridge/protocol code and pinned direct dependencies, with
+no personal Marifold data. Run `npm install --ignore-scripts` inside it to create
+   the dependency lockfile before deploying.
+2. Provision a **TCP Redis** service with TLS, persistent storage, and enough
+   connections for the relay. A REST-only Redis endpoint is insufficient. Keep the
+   Redis service and Vercel region close together. Retain host/revocation metadata
+   across restarts; use a non-evicting database and provider backups.
+3. In Vercel, select Node.js 24 and enable **Fluid compute**. Add the two variables
+   from `.env.example` as encrypted project environment variables. Use a random
+   registration token (for example `openssl rand -hex 32`), and copy it securely
+   to the host's workspace creation prompt. It is not a guest invitation.
+4. From the prepared directory, use the Vercel dashboard or `vercel` CLI to review
+   the account, project, region and deployment. Review Vercel and Redis plan limits
+   and spending controls before confirming creation. This package does not create
+   paid resources automatically. Deployment Protection must permit `/health`,
+   `/v1/hosts` and `/v1/connect`; Marifold authenticates these endpoints itself.
+5. Check `https://<deployment>/health` returns `marifold-bridge`. On the host, run
+   `marifold workspace create Home --bridge https://<deployment>` and enter the
+   registration token. On the other Mac, run `marifold workspace add <url>` and
+   enter the single-use invitation. Add `--executor` to enable remote tools.
+
+The source checkout can also be deployed with `apps/bridge` as the project root;
+include files outside that root for the workspace protocol dependency and use the
+checked-in build command. The prepared standalone directory needs no monorepo.
+
+Vercel WebSockets are currently beta. Its Node function exports an `http.Server`;
+connections reconnect before/after function rotation and may reach another
+instance. Redis, not process memory, coordinates delivery. Test rotation and a
+host–MacBook round trip before relying on the deployment.
+
+Official references: [Vercel WebSockets](https://vercel.com/docs/functions/websockets),
+[Vercel function limits](https://vercel.com/docs/functions/limitations).
+
+### Standalone project configuration
+
+For a package already prepared at `~/.marifold/bridge`, start with:
+
+```sh
+cd ~/.marifold/bridge
+npm install --ignore-scripts
+npx vercel link
+```
+
+Sign in if prompted, select your account/team, and create or select a project such
+as `marifold-bridge`. Link this prepared directory, not the whole `.marifold`
+directory. A Git repository is not required for CLI deployment. Review the chosen
+account and plan before creating resources.
+
+Use these settings for the **standalone prepared package**:
+
+| Setting | Value |
+| --- | --- |
+| Framework preset | Other (`framework: null` in the generated configuration) |
+| Project root | The prepared bridge directory |
+| Node.js | 24.x |
+| Fluid compute | Enabled |
+| Install command | `npm install --ignore-scripts` |
+| Build command | Empty; the prepared runtime is already compiled |
+| Function | `api/bridge.ts`, with `maxDuration: 300` |
+| Routing | Keep the generated rewrite from `/(.*)` to `/api/bridge` |
+| Function region | Choose near your TCP Redis service |
+| Deployment Protection | The bridge URL must accept service requests without an interactive Vercel login |
+
+Keep the generated `vercel.json`; do not replace it with settings for a static
+website. Source-checkout builds use the different build command described above.
+
+Provision TLS TCP Redis with persistent, non-evicting metadata storage and backups.
+In the Vercel project's environment settings, add both variables for **Production**:
+
+| Variable | Value |
+| --- | --- |
+| `MARIFOLD_BRIDGE_REDIS_URL` | Your provider's full `rediss://…` TCP connection URL, including credentials |
+| `MARIFOLD_BRIDGE_REGISTRATION_TOKEN` | A separate random secret used to register hosts |
+
+Generate the registration token locally with `openssl rand -hex 32`. Store it
+securely for the host's workspace creation prompt; do not paste it into shared
+documentation. The `.env.example` file is a template, and a local `.env` does not
+replace Vercel project environment configuration. Add Preview variables separately
+only if you intend to test preview deployments.
+
+After reviewing configuration and costs, deploy from the prepared directory:
+
+```sh
+npx vercel --prod
+curl --fail 'https://<your-project>.vercel.app/health'
+```
+
+The health response should identify `marifold-bridge`. Use the stable project
+production URL for workspace creation and joining. Health alone does not validate
+pairing or WebSocket reconnects; finish the two-device acceptance test above.
+Redeploy after changing production environment variables for them to take effect.
+
+CLI references: [link a project](https://vercel.com/docs/cli/link),
+[deploy](https://vercel.com/docs/cli/deploy), and
+[environment variables](https://vercel.com/docs/environment-variables).
+
+Prefer pnpm? Follow [Using pnpm instead of npm](HOSTING.md#using-pnpm-instead-of-npm)
+to create its lockfile and replace the generated npm install command with a
+version-pinned pnpm command.
+
+## Device connectivity and proxies
+
+Proxy support for workspace bridge connections is **planned, not implemented**.
+The current bridge client does not use Marifold's model-provider proxy settings.
+Do not add a bridge proxy key to `.env` or use a `--proxy` workspace option yet.
+
+The planned setting belongs to each device's local workspace connection. Default
+to **Direct**; allow an explicit proxy for a connection that needs one. Apply the
+same choice to host registration, pairing, WebSocket traffic and reconnects. Make
+it available in CLI and Web UI before the first create/join connection, with later
+editing supported. Keep proxy addresses and any credentials local and protected;
+never synchronize them through the host or expose credentials in status output.
+
+Choose based on reachability from that device, not the hosting provider's name:
+
+| Deployment and device situation | Connection choice |
+| --- | --- |
+| A Vercel bridge is unreachable directly from this Mac | Configure a local proxy once support exists |
+| An Aliyun or another bridge is reachable directly | Direct |
+| Home and office have different network requirements | Configure each device independently |
+
+An Aliyun deployment still needs a compatible HTTPS/WebSocket relay and persistent
+Redis; the [Aliyun ECS recipe](HOSTING.md#aliyun--alibaba-cloud-ecs) has not been
+live-tested. The prepared Vercel configuration is not a generic deployment
+configuration for every provider.
+
+The server's `.env` configures the bridge runtime, including Redis and host
+registration. It does not configure the Macs connecting to it. A proxy address
+such as `127.0.0.1:7890` on Vercel refers to the Vercel runtime, not your Mac.
+Provider proxies remain independent from bridge proxies. A VPN/TUN that routes
+the Marifold service's traffic may already provide connectivity; a successful
+browser page load alone does not prove that background WebSocket traffic works.
+
+## Local smoke test
+
+After building, set the two environment variables and run `node dist/serve.js`.
+It binds loopback port 32143 by default; `PORT` and `HOST` override that for a
+reviewed standalone deployment. Local `http://127.0.0.1:32143` is accepted by
+Marifold. Every non-loopback bridge URL must use HTTPS.
+
+`MARIFOLD_TEST_REDIS_BIN=/path/to/redis-server pnpm --filter @marifold/bridge test`
+starts a disposable loopback Redis with persistence disabled for the integration
+test. It never connects to a configured production Redis database.
+
+## Retention and recovery
+
+Delivery queues retain up to 128 frames per recipient for five minutes. Endpoint
+messages expire after one minute; a reconnect retransmits the same operation ID
+in a fresh encrypted frame. The authoritative endpoint journal prevents repeated
+effects. Redis loss can make existing workspaces unreachable; restore its metadata
+from backup or create and pair a new workspace explicitly. Rotating the registration
+token affects future host registrations; existing membership is revoked by the host.
+
+TLS protects each connection. HPKE and endpoint signatures additionally protect
+message contents from the bridge. The bridge can observe device/workspace IDs,
+routing, packet sizes, timing and public keys. Logs must never include request
+bodies, authorization headers, invitation tokens or decrypted application data.

package/dist/bridge-template/api/bridge.ts

@@ -0,0 +1,5 @@
+import { createBridge, RedisRelayStore } from '../dist/index.js';
+export default createBridge(
+  new RedisRelayStore(process.env.MARIFOLD_BRIDGE_REDIS_URL!),
+  process.env.MARIFOLD_BRIDGE_REGISTRATION_TOKEN!,
+);