feat(relay): rendezvous-relay service — 7 packages + plans (contracts/transport/agent/control-plane/e2e/auth/web)

Multi-tenant reverse-tunnel service ("ngrok for Claude Code" with E2E): a
host-agent dials OUT to an operator-run relay; external devices reach the host
THROUGH the relay, routed by per-tenant subdomain, forwarding ciphertext only
(the relay never sees plaintext). Lets a customer reach their own self-hosted
web-terminal from anywhere with zero networking setup.

Packages — all tsc-strict + vitest green (656 tests), cross-package integration verified:
- relay-contracts: frozen shared contracts (mux frame codec, data model,
  capability token, E2E envelope, pairing) — the src/types.ts analog
- term-relay:   native WS mux + stateless data plane (subdomain routing, ciphertext forward)
- agent:        host-agent (pairing, per-host Ed25519 + mTLS dial-out, forwards to 127.0.0.1:3000)
- control-plane: accounts/hosts registry, pairing-code flow, routing table, provisioning
- relay-e2e:    browser<->agent E2E (X25519 ECDH through relay, AEAD, anti-replay, recoverable replay key)
- relay-auth:   Passkey/WebAuthn, capability tokens, per-host certs, deny-by-default tenant isolation
- relay-web:    browser login + Web Crypto E2E + client-side preview rendering

Security invariants INV1-15 enforced; cross-tenant isolation CI tripwire live
(.github/workflows/relay-tripwire.yml). Design + implementation-level plans in
docs/PLAN_RELAY_*.md and docs/EXPLORE_RELAY_SERVICE.md.

NOTE: generated autonomously per the reviewed plans. The security-critical
packages (relay-e2e, relay-auth) REQUIRE expert security audit before any real
deployment — passing tests prove self-consistency, not resistance to attackers.
Base app (src/, public/) unchanged; concurrent desktop work left uncommitted.
This commit is contained in:
Yaojia Wang
2026-07-02 06:10:16 +02:00
parent e4c327e25e
commit 2af57e6686
326 changed files with 40877 additions and 0 deletions

View File

@@ -0,0 +1,53 @@
# frp scaffold — v0.8 MVP transport (café-demo runbook)
The **v0.8 stepping-stone** transport for the rendezvous-relay: a single-node `frps` on a VPS with
wildcard-subdomain vhost routing, so a phone browser can open `alice.term.<domain>` and land in the
laptop's shell with **no router/VPN/static-IP** configured. This is deliberately temporary — v0.9
replaces it with the native §4.1 WS mux (see **Retirement** below).
## Pieces
| Piece | File | Role |
|---|---|---|
| `frps` config | `frps.toml` | wildcard `*.term.<domain>` vhost + edge auth plugin |
| server-plugin shim | `plugin-hook.ts` | delegates `Login`/`NewProxy`/`NewUserConn` to P3 (no tenancy logic) |
| agent (`frpc`) wrap | *P2 `agent/`* | dials out, wraps `ws://127.0.0.1:3000` |
| control-plane authz | *P3 `control-plane/`* | owns the deny-by-default decision |
## Runbook (one VPS)
1. **Wildcard DNS**`*.term.<domain>` → the VPS public IP (an `A`/`AAAA` record on the wildcard).
2. **Wildcard TLS** — LetsEncrypt **DNS-01** for `*.term.<domain>` (or put Cloudflare in front and
terminate TLS there — M6 scheme-following is preserved end-to-end either way).
3. **`frps`** — run with `frps.toml`; set `subdomainHost = <BASE_DOMAIN>`, `vhostHTTPSPort = 443`.
4. **Edge auth** — start the `plugin-hook.ts` HTTP shim on `127.0.0.1:9001`; point it at P3's authz
endpoint. Every `Login`/`NewProxy`/`NewUserConn` is **deny-by-default** (INDEX v0.8 shared-token
gate — a KNOWN temporary shortcut).
5. **Agent** — on the laptop, run the P2 `frpc`-wrapped agent; it registers subdomain `alice` and
proxies to the local web-terminal `ws://127.0.0.1:3000`.
6. **Demo** — from a phone browser open `alice.term.<domain>`, pass the `clientToken` gate, and you
are in the laptop shell.
### Confirm (acceptance)
- A **relay-node restart does NOT kill the running shell** — the PTY survives on the laptop
(base-app PTY≠WS decoupling, INV7).
- A **foreign-Origin page cannot open the WS** (CSWSH retained at the edge).
- **All 212 base-app tests still pass** (`npm test` in the repo root) — the relay is a layer
*underneath* the existing WebSocket; it does not modify `src/`.
## Retirement → native §4.1 mux (v0.9)
This scaffold is a **swap, not a rewrite**. Each frp piece maps 1:1 onto a native-mux task:
| v0.8 frp piece | v0.9 native replacement |
|---|---|
| frp **yamux** stream multiplexing | §4.1 **mux frame codec + `MuxSession`** (T2/T6) |
| frp **subdomain vhost** routing | **`extractSubdomain` + `RouteResolver`** (T7) |
| frp **login/NewUserConn** plugin | **`authorizeUpgrade`** thin adapter → P5 `onUpgrade` (T8) |
| frp shared `clientToken`/`agentToken` | **capability tokens** (P5) + **mTLS** agent listener (T9) |
| frp opaque TCP/WS forwarding | **opaque byte splice** in `relay-node` (T10, INV2/INV11) |
Because both speak subdomain routing + opaque byte forwarding, ciphertext (P4's `E2EEnvelope`) drops
into DATA frames without reshaping anything (INV2). Even in v0.8, **no terminal bytes are parsed**
— frp forwards opaque streams (INV11).

View File

@@ -0,0 +1,26 @@
# T12 · frp server (frps) — v0.8 MVP transport stepping-stone.
#
# Single-node frps on the VPS with WILDCARD vhost so `*.term.<domain>` routes by subdomain to
# each agent's proxy (EXPLORE §5.4). Contract-compatible with the native §4.1 mux so the v0.9 swap
# reshapes nothing (frp subdomain vhost → T7 router; frp login plugin → T8 gate; yamux → §4.1 mux).
#
# TLS: terminate here OR at Cloudflare in front (M6 scheme-following preserved end-to-end). Values
# below reference environment via your process manager; DO NOT commit real secrets (INV5/INV9).
bindPort = 7000 # agent (frpc) control connection
# Wildcard subdomain vhost — the tenant routing substrate.
vhostHTTPSPort = 443 # browser-facing wss:// (or set vhostHTTPPort behind Cloudflare TLS)
subdomainHost = "term.example.com" # = BASE_DOMAIN; *.term.example.com routes by leftmost label
# Deny-by-default auth at the edge (the auth the base app never had, EXPLORE §5.5).
# The server-plugin delegates EVERY Login/NewProxy/NewUserConn decision to P3's control plane
# via plugin-hook.ts — no tenancy logic lives in frps (v0.8 shared-token gate is a KNOWN,
# temporary shortcut, replaced by capability tokens + mTLS in v0.9).
[[httpPlugins]]
name = "control-plane-authz"
addr = "127.0.0.1:9001" # plugin-hook.ts HTTP shim → P3 authz endpoint
path = "/handler"
ops = ["Login", "NewProxy", "NewUserConn"]
# Opaque forwarding only — frps forwards TCP/WS streams and parses NO terminal bytes (INV11).

View File

@@ -0,0 +1,49 @@
/**
* T12 · frp server-plugin HTTP shim (v0.8 MVP stepping-stone). A THIN adapter implementing frp's
* `Login` / `NewProxy` / `NewUserConn` server-plugin hooks by DELEGATING the decision to P3's
* control-plane authz endpoint. NO tenancy logic lives here — it forwards and enforces the
* deny-by-default answer. Malformed hook bodies are rejected at the Zod boundary.
*
* Retirement (v0.9): frp login plugin → T8 upgrade gate; frp subdomain vhost → T7 router;
* frp yamux → the native §4.1 mux. This shim is deleted when the native mux lands.
*/
import { z } from 'zod'
export type FrpOp = 'Login' | 'NewProxy' | 'NewUserConn'
/** frp server-plugin request envelope: `{ version, op, content }` (content shape varies by op). */
const FrpHookRequestSchema = z
.object({
version: z.string().min(1),
op: z.enum(['Login', 'NewProxy', 'NewUserConn']),
content: z.record(z.unknown()),
})
.strict()
export type FrpHookRequest = z.infer<typeof FrpHookRequestSchema>
/** frp plugin response: reject (deny) OR pass-through unchanged. */
export type FrpHookResponse =
| { readonly reject: true; readonly reject_reason: string }
| { readonly reject: false; readonly unchange: true }
/** The control-plane (P3) decision the shim forwards to — it owns ALL tenancy logic. */
export interface ControlPlaneAuthz {
authorize(op: FrpOp, content: Readonly<Record<string, unknown>>): Promise<{ allow: boolean }>
}
const REJECT_MALFORMED: FrpHookResponse = { reject: true, reject_reason: 'malformed hook request' }
const REJECT_DENIED: FrpHookResponse = { reject: true, reject_reason: 'denied by control plane' }
const ALLOW: FrpHookResponse = { reject: false, unchange: true }
/**
* Handle one frp server-plugin hook call. Validates the body (deny-by-default on malformed),
* forwards the decision to P3, and maps allow→pass-through / deny→reject. Never decides tenancy.
*/
export async function handleFrpHook(raw: unknown, authz: ControlPlaneAuthz): Promise<FrpHookResponse> {
const parsed = FrpHookRequestSchema.safeParse(raw)
if (!parsed.success) return REJECT_MALFORMED
const { op, content } = parsed.data
const decision = await authz.authorize(op, content)
return decision.allow ? ALLOW : REJECT_DENIED
}