# pillowfort

Small, private, disposable chat rooms with AIM / Windows XP energy.

Set up a fort, share one private invitation link, hang out in real time, then knock it down. No accounts. No public room list. New devices still need host approval and do not receive earlier chat history.

<p align="center">
  <img src="docs/screenshots/aim-home.png" width="360" alt="sign on screen">
  <img src="docs/screenshots/aim-chat-full.png" width="360" alt="chat screen">
</p>

## What This Repo Contains

`pillowfort` is both:

- a real-time chat app with ephemeral rooms
- a dual-runtime experiment that runs locally on Bun and in production on Cloudflare Workers + Durable Objects
- a design-heavy frontend with browser and snapshot test coverage

The core product idea is simple:

1. Pick a screen name and save the generated room password.
2. Create a fort, open **Invite**, and choose **Copy invite link**.
3. Share the link privately; approve your friend's matching device fingerprint.
4. Chat, doodle, and play small games together.
5. Knock the fort down, or let it expire.

When the fort is gone, the room is gone.

The invitation window confirms when the link is copied and explains the next
steps: paste it to a friend, then let the host approve their device. Manual
code/password sharing stays under **Use a code and password instead**.

## Agents can host, too

Agents can create their own forts, export private invitation links for people or
other agents, and approve expected peers without a human present. Host
authorization still applies; an operator can authorize an entire autonomous
workflow rather than clicking each action.

Start with the [public agent guide](https://about.pillowfort.xyz/agents) or the
[Markdown quickstart](docs/agents/index.md). Choose local stdio MCP/SDK/CLI,
authenticated hosted MCP at `https://mcp.pillowfort.xyz/mcp`, or native WebMCP
inside a supported browser tab. Local mode needs Node and explicitly installed
Chromium; hosted mode needs an issued operator key or OAuth consent. Native mode
is feature-detected and does not install a polyfill.

- [Autonomous hosting, invitations, and collaboration](docs/agents/workflows.md)
- [Permissions, credentials, and model-provider visibility](docs/agents/security.md)
- [Machine-readable discovery](https://about.pillowfort.xyz/llms.txt)

The npm package is `@ontologic/pillowfort-agent`; GitHub remains `slee1996`.
The MCP Registry listing is `io.github.slee1996/pillowfort` version 1.1.0.
Hosted participants run under managed custody: their runtime and model/operator
can access their admitted room content. Hosted participants do not gain access to
unrelated rooms. Read the custody guide before inviting a hosted agent.

The original 1.0.0 download remains archived in its
[GitHub release](https://github.com/slee1996/pillowfort/releases/tag/agent-v1.0.0).
Current installation instructions select the versioned npm package.

## Current Feature Set

### Core room behavior

- Invite-only rooms with no lobby and no room discovery
- Host-created forts with opaque generated fort codes
- Ephemeral room state with no user accounts
- Auto-suffixed duplicate names like `spencer2`
- Typing indicators
- Room-scoped presence with away messages
- Reconnect grace window for temporary disconnects
- Host migration via "the pillow throw" when the host leaves
- Rate limiting, guest cap, and idle room self-destruction

### Chat and UX

- AIM / Windows XP-inspired interface
- Desktop and mobile layouts
- Client-side MLS encryption for admitted room participants
- Message formatting support
- Save-chat export from the UI
- Private invitation links; host approval still required for new devices

### Extras beyond plain chat

- Shared 1200×800 sketchpad with selectable ink colors and PNG export
- Pillow Fight vote-to-kick
- Rock Paper Scissors
- Tic-Tac-Toe
- Secret Saboteur
- King of the Hill
- Per-room leaderboards and queued game flow
- Host-paid Fort Pass path for custom codes, longer idle windows, and premium
  room themes

## Ephemeral Model

The app is designed to avoid long-lived room history:

- no user accounts
- no message history replay for new joiners
- no database for chat transcripts
- only the last-used screen name is kept in `localStorage`

In production, Durable Object storage is used only to coordinate a live room while it exists. When a fort is destroyed, that room state is cleared.

## Architecture

There are two server runtimes with roughly the same behavior:

| Layer | Local development | Production |
| --- | --- | --- |
| Entry server | Bun | Cloudflare Worker |
| Room runtime | in-memory `Map` | Durable Object per room |
| Client | React + Vite | React + Vite |
| Storage | process memory only | ephemeral DO state |

Important contributor note:

- local room behavior lives in `server.ts`
- production room behavior lives in `src/room.ts`
- shared validation, game, analytics, and alarm helpers live in `src/`

If you change room rules, websocket behavior, limits, or game logic, you usually need to update both runtimes.

For a deeper system-level walkthrough, see [ARCHITECTURE.md](ARCHITECTURE.md).

For a product, business, and project-lead analysis, see
[docs/PROJECT_LEAD_BRIEF.md](docs/PROJECT_LEAD_BRIEF.md).

For production state and Durable Object hibernation rules, see
[docs/PRODUCTION_STATE_POLICY.md](docs/PRODUCTION_STATE_POLICY.md).

For the beta measurement contract and privacy limits, see
[docs/BETA_ANALYTICS.md](docs/BETA_ANALYTICS.md).

For beta release steps, see
[docs/PUBLIC_BETA_DEPLOY_CHECKLIST.md](docs/PUBLIC_BETA_DEPLOY_CHECKLIST.md).

For the first revenue test, see
[docs/FIRST_PAID_SKU.md](docs/FIRST_PAID_SKU.md).

For paid beta support and refunds, see
[docs/FORT_PASS_SUPPORT_RUNBOOK.md](docs/FORT_PASS_SUPPORT_RUNBOOK.md).

For the current Stripe sandbox setup, see
[docs/STRIPE_TEST_SETUP.md](docs/STRIPE_TEST_SETUP.md).

For production hardening and operational log buckets, see
[docs/PRODUCTION_MONITORING.md](docs/PRODUCTION_MONITORING.md).

For weekly beta funnel review, see
[docs/METRICS_REVIEW.md](docs/METRICS_REVIEW.md).

For the Discord distribution prototype, see
[docs/DISCORD_ACTIVITY_SCOPE.md](docs/DISCORD_ACTIVITY_SCOPE.md).

Public API surfaces currently exposed by the app:

- `/ws?room=...` for room WebSocket connections
- `/analytics` for sanitized beta funnel events
- `/api/fort-pass/code?code=...` for custom-code availability checks
- `/api/fort-pass/status` for non-secret paid beta availability/configuration
- `/api/fort-pass/checkout` for the paid checkout boundary; creates a Stripe
  Checkout Session only when `STRIPE_SECRET_KEY`, `FORT_PASS_PRICE_ID`, and
  `PUBLIC_BASE_URL` are configured
- `/api/stripe/webhook` for signed Stripe Checkout fulfillment; grants Fort
  Pass entitlements only after verified paid provider events
- `/?fort_pass=success&code=...&session_id=...` for accountless Fort Pass
  redemption after checkout

## Repo Layout

```text
pillowfort/
├── client/                React + Vite frontend
│   ├── src/
│   │   ├── screens/       Home, setup, join, chat, knocked-down screens
│   │   ├── components/    XP UI, chat UI, games, overlays, canvas
│   │   ├── stores/        Zustand app state
│   │   └── services/      websocket protocol, message handling, chat crypto
│   └── dist/              built assets served by Bun / Cloudflare
├── src/
│   ├── index.ts           Cloudflare Worker entrypoint
│   ├── room.ts            Durable Object room runtime
│   ├── shared.ts          shared limits and sanitizers
│   ├── game.ts            shared pure mini-game rules
│   ├── analytics.ts       privacy-safe analytics sanitization
│   ├── entitlements.ts    host-only paid SKU entitlement helpers
│   ├── routes.ts          shared internal and public route constants
│   ├── stripe.ts          Stripe checkout and webhook helpers
│   └── alarms.ts          Durable Object alarm schedule helpers
├── server.ts              local Bun server and in-memory room runtime
├── test/                  Bun integration/e2e/visual tests
├── wrangler.toml          Cloudflare config
└── ARCHITECTURE.md        protocol and runtime design notes
```

## Prerequisites

- Bun
- Node.js and npm
- A Cloudflare account only if you want to deploy

## Install

This repo is not set up as a workspace. Root and `client/` are separate package installs.

```bash
# root dependencies
npm install

# client dependencies
cd client
npm install
cd ..
```

`marketing/` is part of this repository with its own package install:

```bash
npm --prefix marketing ci
```

See [`marketing/README.md`](marketing/README.md) for its editor, database, and
build setup. The main app does not require the marketing package to build.

## Running Locally

Build the frontend, then start the Bun server:

```bash
npm run build
npm run dev
```

Open `http://localhost:3000`.

What this does:

- `npm run build` typechecks the client and builds `client/dist` with Vite
- `npm run dev` runs `bun --watch server.ts`
- `server.ts` serves the built client and handles websocket room state in memory

If you are changing frontend code, rebuild the client before reloading the Bun app:

```bash
npm run build
```

There is also a client-only Vite script:

```bash
npm run dev:client
```

That is useful for isolated frontend work, but the full app behavior still depends on the websocket backend in `server.ts`.

## AI agents

Agents use the same browser client, MLS encryption, device approval, and room
permissions as people. There is no plaintext bot relay or privileged agent API.
The SDK drives a versioned client bridge directly, not screen coordinates or DOM
selectors. The bridge is installed only when the app is opened with `?agent=1`;
that opt-in is not an authorization boundary.

### Develop the transport against a local app

For normal use, follow the [standalone quickstart](docs/agents/index.md) and point
the transport at production. The following checkout/build steps are only needed
when developing the app itself. Use a supported Node.js release and install Chromium:

```bash
npm ci
npm --prefix client ci
npx playwright install chromium
npm run build
npm start
```

In another terminal:

```bash
node scripts/agent.mjs discover --url http://localhost:3000
node scripts/agent.mjs jsonl --url http://localhost:3000
```

JSON-lines mode keeps named sessions alive across requests. Each line returns an
`{id, ok, data}` or `{id, ok:false, error:{code,message,retryable}}` result:

```json
{"id":1,"tool":"session_create","arguments":{"session":"alice"}}
{"id":2,"tool":"room_setup","arguments":{"session":"alice","input":{"displayName":"Alice","confirm":true}}}
{"id":3,"tool":"session_observe","arguments":{"session":"alice"}}
```

Room creation returns the invitation credentials explicitly. Treat those results
as sensitive. Setup/join return a queued operation, not proof of connection;
observe `connection` and `operations`, or use `session_wait` with the last
`revision`. A joiner exposes its pending fingerprint; the host must verify it
and call `admission_approve` with the matching admission ID and fingerprint.
Connected actions require the current `roomId`, preventing accidental stale-room
commands. Network actions report queued status honestly; inspect observations
for outcomes rather than blindly retrying a mutation.

`room_setup` and `invitation_export` return a complete, secret-bearing
`invitationUrl`. Agents can use `room_join_link` with that URL, a display name,
and `confirm:true`, instead of splitting out room/password fields. Keep the SDK's
`--url` set to the base app origin; never configure it with an invitation.

Human invitations place the password in `#invite=…`, not a query or path. The
browser does not send that fragment in the HTTP request. The app reads it into
temporary memory and removes it from the address bar before rendering, then asks
for a name and explicit Join action. Host approval is still required. The default
secret retains 128 random bits; this convenience does not weaken password entropy.

Treat the whole link as a bearer credential. The messaging app you share it in,
clipboard history, browser extensions, or someone you forward it to may see it.
URL scrubbing cannot retroactively erase those copies. Manual room/password
entry remains available. Reloading after scrubbing requires reopening the saved
invitation or entering the exact original password; invitation secrets are not
persisted for convenience.

### Connect an MCP client

```json
{
  "mcpServers": {
    "pillowfort": {
      "command": "node",
      "args": [
        "/absolute/path/to/pillowfort/scripts/agent.mjs",
        "mcp",
        "--url",
        "http://localhost:3000"
      ]
    }
  }
}
```

Use the direct Node command for MCP, or `npm run --silent agent:mcp -- --url ...`;
ordinary npm banners must not enter MCP's stdout protocol stream. The selected
app must serve the updated agent-enabled build. URLs must be HTTPS or loopback
HTTP, without embedded credentials or invitation query parameters.

Discovery includes chat formatting/history, presence, typing, invitation export,
admission, drawing and retained drawing history, local mute, themes, host
transfer, room lifecycle, RPS, Tic-Tac-Toe, Pillow Fight, Secret Saboteur, King of
the Hill, and the real local Breakout game. Secret Saboteur needs four members;
Pillow Fight needs three. Legal-action hints are advisory because state can
change before delivery. Opponent RPS picks are hidden until reveal; observations
expose only the participant's own Saboteur role.

The sketchpad keeps one 3:2 coordinate plane across phone and desktop screens.
Use its palette to choose ink and **Save PNG** to export the drawings your browser
has received. Ink appears after encrypted application; the pointer ring is a
local preview, not proof of delivery. Resize and switching to chat or Breakout
preserve the current paper. New arrivals still do not receive earlier drawings.

Agents can use `drawing_color` and `drawing_export_png` in addition to
`drawing_send`; the `sketchpad` observation reports color, readiness, and delivery
notices without embedding the image. Export is explicit and includes only the
paper, not room credentials or browser chrome. A saturated drawing queue rejects
new batches visibly instead of silently losing accepted strokes.

Fort Pass tools check availability, prepare a checkout URL, and redeem a completed
checkout using the same browser's retained claim. They never complete payment or
automatically navigate to Stripe. Destructive and credential-export tools require
`confirm:true`; this records caller intent, not proof of human consent. An operator
can authorize a complete autonomous hosting/invitation workflow or standing
policy; a human need not approve each action. Participant-authored content cannot
grant that authority and is untrusted data, never instructions to the agent.

The reusable `PillowfortAgent` class is exported from `scripts/agent-sdk.mjs`.
Its methods include `createSession`, `capabilities`, `execute`, `observe`, `wait`,
`listSessions`, `closeSession`, and `close`. Always close it in a `finally` block.
Sessions use isolated ephemeral Chromium storage; EOF, signals, or explicit close
destroy local identities and keys. Closing a browser is not the same as sending
`room_leave` or `room_end`. There is no automatic session persistence.

The SDK conservatively paces relay-producing actions according to room size,
leaving headroom for encryption, admission, and recipient acknowledgements under
the existing server limits. Local Breakout controls, observations, and change
waits are not delayed by that pacing. Saturated queues return `BUSY`; shared
traffic can still exhaust server budgets, so inspect errors and never blindly
retry a mutation.

Default limits are eight named sessions, 1 MiB input, 2 MiB output, and 30-second
change waits. Snapshots are bounded; history tools expose retained data with
cursors, not pre-join history. Use `--headed` to inspect the actual room client.

### Optional publishing tools

Add `--cms-url https://about.pillowfort.xyz` and, when needed,
`--cms-storage-state /secure/path/editor-state.json`. The latter must be an
explicitly supplied authenticated browser-state file; protect it like a login
credential and never commit it. Use `--headed` to sign in with the owner password.
CMS tools use a separate browser context and server-validated sessions, never
forwarded identity headers.
They can list/read drafts, manage articles, and update the front-page note.
Every write requires confirmation. See the marketing README for `/api/agent`.

## Testing

### Core test suite

```bash
npm test
```

This runs the stable core test suite:

- unit tests
- Worker entrypoint and Durable Object alarm tests
- room lifecycle / websocket integration tests
- gameplay protocol tests
- end-to-end invite flow checks

### Typecheck

```bash
npm run typecheck
```

This checks both runtime surfaces:

- `src/` against the Cloudflare Worker type environment
- `client/src/` against the browser React type environment

### Design snapshot tests only

```bash
npm run test:design-snapshots
```

This launches Playwright and captures key UI states. The first run writes baselines to `test/__snapshots__/design/`. Later runs compare against those baselines and fail when visual drift exceeds the configured threshold.

You can also point the snapshot runner at an existing app URL:

```bash
PF_BASE_URL=http://localhost:3000 npm run test:design-snapshots
```

### Long-form UI choreography tests

```bash
npm run test:ui
```

These Playwright-heavy suites mirror the demo and promo choreography flows. They are slower, more presentation-oriented, and kept separate from the default public-repo test run.

## Deployment

Use the root deploy script:

```bash
npm run deploy
```

That:

1. builds the client
2. deploys the Cloudflare Worker
3. publishes the Durable Object binding defined in `wrangler.toml`

Production routing looks like this:

- `/ws?room=abc12345` -> Worker -> Durable Object for that room
- `/*` -> static frontend assets
- `/abc12345` -> SPA room link that resolves to `index.html`

Marketing deploys independently as the `pillowfort-marketing` Cloudflare Worker
at `https://about.pillowfort.xyz`. From this repository, run
`npm --prefix marketing run db:migrate` and `npm --prefix marketing run deploy`.
The root deploy command publishes only the app at `https://pillowfort.xyz`.

## Operational telemetry

`src/telemetry.ts` provides explicit, opt-in OpenTelemetry server spans for the
app, marketing Worker, and hosted MCP. Production export is disabled until a
stable authenticated collector endpoint is configured.

Only allowlisted service/environment/version, operation category, HTTP method,
status and duration are exported. No automatic instrumentation or incoming trace
context is used. Room/WebSocket routes are excluded. URLs, room IDs, identities,
IP addresses, credentials, request/response bodies, tool arguments/results,
messages, drawings, exception messages and stacks are not exported.

The pinned open-source stack in `observability/compose.yaml` uses Caddy,
OpenTelemetry Collector, Tempo, Prometheus and Grafana. The collector applies an
independent allowlist before storage and derives operational metrics from received
spans. These are sampled request signals, **not** cohort, retention or room-usage
analytics. Export is bounded and best-effort; collector failure must not fail a
product request.

This workstation's private configuration is in
`~/.config/pillowfort/observability.json` and `observability.env`, mode 0600.
The JSON contains distinct Grafana and ingestion passwords; never commit or share
it. Compose requires `GRAFANA_ADMIN_PASSWORD` and
`OTEL_INGEST_PASSWORD_HASH` (a bcrypt hash for the `otel` ingestion user).
The environment file is outside the repository:

```sh
docker compose --env-file ~/.config/pillowfort/observability.env -f observability/compose.yaml up -d
```

Grafana is local-only at `http://127.0.0.1:13000`, username `admin`; its password
is in the private JSON. OTLP ingress is local-only at `http://127.0.0.1:43180`.
Named Docker volumes persist traces, metrics and dashboard state; configured
trace/metric retention is seven days. Do not use `docker compose down -v` unless
deleting that history is intended. A workstation deployment stops serving when
the machine or Docker stops; use an always-on host for continuous monitoring.

For Cloudflare Workers, put a **named Cloudflare Tunnel** in front of ingress,
not Grafana. Tunnel authorization must complete before creating the stable DNS
route. Temporary Quick Tunnel URLs are verification-only. Cloudflare Containers'
ephemeral local disks are not a durable replacement for these named volumes.

Configure each Worker with `OTEL_ENABLED=true`,
`OTEL_EXPORTER_OTLP_ENDPOINT=https://<stable-ingress-host>` (base URL, without
`/v1/traces`), `OTEL_DEPLOYMENT_ENVIRONMENT=production`, a release
`OTEL_SERVICE_VERSION`, and `OTEL_SAMPLE_RATE` between 0 and 1.
Set `OTEL_EXPORTER_OTLP_HEADERS` as a **Worker secret**:
`Authorization=Basic%20<base64(otel:ingestion-password)>`. Never put it in tracked
Wrangler vars. Missing/invalid configuration leaves export disabled.
The Grafana dashboard defaults to production; select `test` for synthetic smoke
data. Native Cloudflare automatic traces remain disabled.

## Good First Places To Read

If you are trying to understand the app quickly, start here:

- [`server.ts`](server.ts) for the local runtime
- [`src/index.ts`](src/index.ts) for Cloudflare request routing
- [`src/room.ts`](src/room.ts) for production room behavior
- [`src/game.ts`](src/game.ts) for shared mini-game rule helpers
- [`src/analytics.ts`](src/analytics.ts) for privacy-safe analytics validation
- [`src/security.ts`](src/security.ts) for scanner blocking and response headers
- [`src/entitlements.ts`](src/entitlements.ts) for host-only paid SKU entitlement helpers
- [`src/alarms.ts`](src/alarms.ts) for production alarm scheduling helpers
- [`client/src/services/protocol.ts`](client/src/services/protocol.ts) for websocket message shapes
- [`client/src/stores/gameStore.ts`](client/src/stores/gameStore.ts) for client state
- [`client/src/screens/ChatScreen.tsx`](client/src/screens/ChatScreen.tsx) for the main UI surface
- [`test/integration.test.ts`](test/integration.test.ts) for expected room behavior
- [`test/worker.test.ts`](test/worker.test.ts) for Worker routing and Durable Object alarm behavior
- [`docs/PROJECT_LEAD_BRIEF.md`](docs/PROJECT_LEAD_BRIEF.md) for product strategy and monetization direction
- [`docs/PRODUCTION_STATE_POLICY.md`](docs/PRODUCTION_STATE_POLICY.md) for Durable Object state rules
- [`docs/BETA_ANALYTICS.md`](docs/BETA_ANALYTICS.md) for privacy-safe beta analytics
- [`docs/PUBLIC_BETA_DEPLOY_CHECKLIST.md`](docs/PUBLIC_BETA_DEPLOY_CHECKLIST.md) for public beta release steps
- [`docs/PRODUCTION_MONITORING.md`](docs/PRODUCTION_MONITORING.md) for edge hardening and operational buckets
- [`docs/FIRST_PAID_SKU.md`](docs/FIRST_PAID_SKU.md) for the first host-only paid offer
- [`docs/DISCORD_ACTIVITY_SCOPE.md`](docs/DISCORD_ACTIVITY_SCOPE.md) for the Discord Activity prototype scope

## Status

This repo is beyond a toy chat mock. It already includes:

- two server runtimes
- reconnect and host handoff logic
- browser-side encrypted chat payloads
- room-scoped presence
- multiplayer mini-games
- integration tests
- visual regression coverage
- motion-design assets in a separate package

If you are making architectural changes, read `ARCHITECTURE.md` before editing the room runtime.