feat: make each plugin an npm workspace built by a termix-plugin CLI

plugins/web-endpoint/src/backend/index.ts

@@ -0,0 +1,11 @@
+import type { PluginContext } from "@termix/plugin-sdk/backend";
+import { startWebEndpointService, stopWebEndpointService } from "./routes.js";
+
+export async function activate(ctx: PluginContext) {
+  startWebEndpointService();
+  ctx.log.info("Web Endpoint router registered at /ssh/tunnel/web-endpoint");
+}
+
+export async function deactivate() {
+  stopWebEndpointService();
+}

plugins/web-endpoint/src/backend/routes.ts

@@ -0,0 +1,323 @@
+import express from "express";
+import { Client } from "ssh2";
+import { AuthManager } from "../../../../src/backend/utils/auth-manager.js";
+import { tunnelLogger } from "../../../../src/backend/utils/logger.js";
+import { parseWebUiConfig } from "../../../../src/backend/database/routes/host-web-endpoints.js";
+import {
+  activeTunnelRuntimes,
+  cleanupTunnelResources,
+  connectSSHTunnel,
+  connectionStatus,
+  tunnelConnecting,
+} from "../../../../src/backend/hosts/tunnel/manager.js";
+import { forwardOut } from "../../../../src/backend/hosts/tunnel/ssh-primitives.js";
+import { buildWebEndpointTunnelName } from "../../../../src/backend/hosts/tunnel/utils.js";
+import {
+  registerWebEndpointRouter,
+  unregisterWebEndpointRouter,
+} from "../../../../src/backend/hosts/tunnel/web-endpoint-dispatch.js";
+import type { TunnelConfig, WebEndpoint } from "../../../../src/types/index.js";
+
+/** Matches the spec's ten minutes. */
+const WEB_ENDPOINT_IDLE_TIMEOUT_MS = 10 * 60 * 1000;
+
+/**
+ * Long enough for a slow SSH handshake, short enough that a wedged connection
+ * does not hold the HTTP request open indefinitely. One second past the
+ * manager's own connection timeout so the manager reports first.
+ */
+const TUNNEL_READY_TIMEOUT_MS = 61_000;
+const TUNNEL_READY_POLL_MS = 100;
+const TARGET_PROBE_TIMEOUT_MS = 10_000;
+
+/**
+ * What the live forward was built for. Compared on reuse so an endpoint edited
+ * under a live tunnel is reopened rather than silently forwarding to the old
+ * target.
+ *
+ * JSON rather than a joined string: no separator can be smuggled in by a field
+ * value, and the result stays readable in a log. The SSH identity is included
+ * because a host repointed at a different server must also invalidate the
+ * forward, not just a changed endpoint port.
+ */
+function targetFingerprint(
+  endpoint: Pick<
+    WebEndpoint,
+    "scheme" | "port" | "access" | "bindHost" | "localPort"
+  >,
+  host: { ip?: string; port?: number; username?: string },
+): string {
+  return JSON.stringify({
+    scheme: endpoint.scheme,
+    port: endpoint.port,
+    access: endpoint.access,
+    bindHost: endpoint.bindHost ?? null,
+    localPort: endpoint.localPort ?? null,
+    ip: host.ip ?? null,
+    sshPort: host.port ?? null,
+    username: host.username ?? null,
+  });
+}
+
+const openedTargets = new Map<string, string>();
+
+function sleep(ms: number): Promise<void> {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
+
+/**
+ * connectSSHTunnel(config) returns Promise<void> and NEVER REJECTS: every
+ * failure path inside it reports over broadcastTunnelStatus and returns
+ * normally, and its own promise resolves right after conn.connect(), long
+ * before the SSH "ready" event that would populate activeTunnelRuntimes.
+ *
+ * So "awaited" is not "connected". This is what actually waits.
+ *
+ * Polling the maps the manager already exports, rather than adding a
+ * promise-returning variant to manager.ts: every existing caller of
+ * connectSSHTunnel is fire-and-forget by design (the connect route responds
+ * "Connection request received" immediately and reports asynchronously over
+ * the status broadcast), so changing that contract to serve one new caller is
+ * a far larger blast radius than polling here.
+ */
+async function waitForTunnelSettled(tunnelName: string): Promise<void> {
+  const deadline = Date.now() + TUNNEL_READY_TIMEOUT_MS;
+
+  while (Date.now() < deadline) {
+    if (activeTunnelRuntimes.has(tunnelName)) return;
+
+    const status = connectionStatus.get(tunnelName);
+    if (!tunnelConnecting.has(tunnelName) && status && !status.connected) {
+      throw new Error(status.reason || "The endpoint tunnel failed to connect");
+    }
+    await sleep(TUNNEL_READY_POLL_MS);
+  }
+
+  throw new Error("Timed out establishing the endpoint tunnel");
+}
+
+/**
+ * Opens one channel to the target and closes it immediately.
+ *
+ * Without this the route returns 200 the moment tcpServer.listen succeeds, and
+ * forwardOut is only attempted per inbound socket -- where a failure is
+ * swallowed by the socket's own error handling. The user then gets a blank
+ * frame and no message for the single likeliest error: nothing is actually
+ * listening on that port. Costs one round trip per open.
+ */
+async function probeEndpointTarget(
+  sourceClient: Client,
+  targetPort: number,
+  tunnelName: string,
+): Promise<void> {
+  const channel = await Promise.race([
+    forwardOut(sourceClient, "127.0.0.1", targetPort, tunnelName),
+    new Promise<never>((_resolve, reject) =>
+      setTimeout(
+        () => reject(new Error("Timed out reaching the endpoint port")),
+        TARGET_PROBE_TIMEOUT_MS,
+      ),
+    ),
+  ]);
+  try {
+    channel.end();
+  } catch {
+    // Already gone; the probe succeeded either way.
+  }
+}
+
+export async function handleWebEndpointOpen(
+  req: express.Request,
+  res: express.Response,
+): Promise<express.Response | void> {
+  const userId = (req as { userId?: string }).userId;
+  if (!userId) {
+    return res.status(401).json({ error: "Authentication required" });
+  }
+
+  const { hostId, endpointId } = req.body ?? {};
+  if (
+    !Number.isInteger(hostId) ||
+    hostId < 1 ||
+    typeof endpointId !== "string"
+  ) {
+    return res.status(400).json({ error: "Invalid web endpoint request" });
+  }
+
+  // Deliberately NOT gated to the desktop. The forward binds wherever this
+  // backend runs, exactly as the server tunnels feature does, and the
+  // endpoint's own bindHost decides whether that is reachable from the
+  // browser -- loopback for the desktop, an address the server answers on for
+  // a web deployment.
+
+  const { resolveHostById } =
+    await import("../../../../src/backend/hosts/host-resolver.js");
+  const host = await resolveHostById(hostId, userId);
+  if (!host) {
+    return res.status(403).json({ error: "Host not found or access denied" });
+  }
+
+  // All three write paths null the config on disable, so a configured-but-
+  // disabled host is not normally reachable -- except a bulk update that sends
+  // only { webUiConfig } while the stored enable_web_ui is still 0. The UI
+  // reads as off everywhere in that state, so the endpoint merely being
+  // present in the config is not on its own a licence to open a tunnel.
+  if (!host.enableWebUi) {
+    return res
+      .status(400)
+      .json({ error: "Web endpoints are not enabled for this host" });
+  }
+
+  // Re-normalized rather than trusted: the stored value predates any later
+  // tightening of the rules, and this is the value a forward is built from.
+  const endpoint = parseWebUiConfig(host.webUiConfig).endpoints.find(
+    (candidate) => candidate.id === endpointId,
+  );
+  if (!endpoint) {
+    return res.status(400).json({ error: "Web endpoint not found" });
+  }
+  if (endpoint.access !== "tunnel") {
+    return res
+      .status(400)
+      .json({ error: "This endpoint does not use a tunnel" });
+  }
+
+  const tunnelName = buildWebEndpointTunnelName(hostId, endpointId);
+  const fingerprint = targetFingerprint(endpoint, host);
+  const existing = activeTunnelRuntimes.get(tunnelName);
+
+  if (existing) {
+    const recorded = openedTargets.get(tunnelName);
+    // The runtime is the source of truth and openedTargets is derived from it,
+    // so an ABSENT fingerprint means reuse, not staleness. A reserved-prefixed
+    // runtime can only have been created by this route.
+    if (recorded === undefined || recorded === fingerprint) {
+      openedTargets.set(tunnelName, fingerprint);
+      return res.status(200).json({ port: existing.bindPort });
+    }
+
+    tunnelLogger.info("Reopening web endpoint tunnel after config change", {
+      operation: "web_endpoint_reopen",
+      tunnelName,
+    });
+    // Forced, for the same reason handleDisconnect's reserved-name branch
+    // forces: cleanupTunnelResources no-ops while tunnelConnecting holds the
+    // name, and a web tunnel has no retry pass to self-heal a runtime left
+    // behind. An unforced call could leave a dead listener registered.
+    await cleanupTunnelResources(tunnelName, true);
+    openedTargets.delete(tunnelName);
+  }
+
+  const tunnelConfig: TunnelConfig = {
+    name: tunnelName,
+    scope: "s2s",
+    mode: "local",
+    tunnelType: "local",
+    // Loopback unless the endpoint asks otherwise. A non-loopback bind
+    // publishes the target's web UI to anyone who can reach the port, with no
+    // authentication in front of it -- the same trade the server tunnels
+    // feature exposes, and only ever an explicit per-endpoint choice. The
+    // value is validated to a bare host literal by normalizeBindHost.
+    bindHost: endpoint.bindHost || "127.0.0.1",
+    sourceHostId: host.id ?? hostId,
+    tunnelIndex: 0,
+    requestingUserId: userId,
+    hostName: host.name || `${host.username}@${host.ip}`,
+    sourceIP: host.ip,
+    sourceSSHPort: host.port,
+    sourceUsername: host.username,
+    sourcePassword: host.password,
+    sourceAuthMethod: host.authType,
+    sourceSSHKey: host.key,
+    sourceKeyPassword: host.keyPassword,
+    sourceKeyType: host.keyType,
+    sourceCredentialId: host.credentialId,
+    // connectSSHTunnel reads sourceUserId when building its client.
+    sourceUserId: host.userId,
+    endpointIP: host.ip,
+    endpointSSHPort: host.port,
+    endpointUsername: host.username,
+    // Load-bearing: connectSSHTunnel picks its strategy via
+    // shouldEstablishDirectTunnel -> isSingleHostTunnel, which returns true
+    // when endpointHost is "127.0.0.1". That is what selects
+    // establishDirectTunnel's local branch -- the bind-locally-and-forwardOut
+    // path this feature needs. Setting only targetHost would NOT select it.
+    endpointHost: "127.0.0.1",
+    endpointAuthMethod: host.authType,
+    endpointSSHKey: host.key,
+    endpointKeyPassword: host.keyPassword,
+    endpointKeyType: host.keyType,
+    endpointCredentialId: host.credentialId,
+    endpointUserId: host.userId,
+    // A fixed port when the endpoint names one, as the server tunnels
+    // feature's Source Port does -- a container can only publish ports it
+    // knows in advance. Otherwise 0 and the kernel picks; the manager reports
+    // back what it actually bound either way.
+    sourcePort: endpoint.localPort ?? 0,
+    endpointPort: endpoint.port,
+    // Documentation only: maxRetries 0 does NOT disable retries, because the
+    // manager reads `maxRetries || 3`. Retry is actually suppressed by the
+    // reserved-name early return in handleDisconnect.
+    maxRetries: 0,
+    retryInterval: 0,
+    autoStart: false,
+    isPinned: Boolean(host.pin),
+    idleTimeoutMs: WEB_ENDPOINT_IDLE_TIMEOUT_MS,
+    useSocks5: Boolean(host.useSocks5),
+    socks5Host: host.socks5Host,
+    socks5Port: host.socks5Port,
+    socks5Username: host.socks5Username,
+    socks5Password: host.socks5Password,
+    socks5ProxyChain: host.socks5ProxyChain,
+  } as TunnelConfig;
+
+  try {
+    await connectSSHTunnel(tunnelConfig);
+    await waitForTunnelSettled(tunnelName);
+
+    const runtime = activeTunnelRuntimes.get(tunnelName);
+    if (!runtime) {
+      throw new Error("The endpoint tunnel closed before it could be used");
+    }
+
+    await probeEndpointTarget(runtime.sourceClient, endpoint.port, tunnelName);
+
+    openedTargets.set(tunnelName, fingerprint);
+    return res.status(200).json({ port: runtime.bindPort });
+  } catch (error) {
+    const reason = error instanceof Error ? error.message : String(error);
+    tunnelLogger.error("Failed to open web endpoint tunnel", error, {
+      operation: "web_endpoint_open_failed",
+      tunnelName,
+    });
+    // Leave nothing half-built: a listener with no working forward behind it
+    // would be handed out on the next open as a live tunnel.
+    await cleanupTunnelResources(tunnelName, true).catch(() => undefined);
+    openedTargets.delete(tunnelName);
+    return res.status(502).json({ error: reason });
+  }
+}
+
+export const router = express.Router();
+
+let started = false;
+
+export function startWebEndpointService(): void {
+  if (!started) {
+    // Built on first start rather than at module scope so importing this
+    // module for a handler unit test does not construct the auth singleton.
+    const authenticateJWT = AuthManager.getInstance().createAuthMiddleware();
+    // Only "/open" here: the dispatcher this router is registered with is
+    // mounted at /ssh/tunnel/web-endpoint, so the full path stays
+    // /ssh/tunnel/web-endpoint/open exactly as before. The "/ssh" prefix is
+    // part of the path the tunnel service serves -- nginx proxies /ssh/tunnel/
+    // through with the path intact.
+    router.post("/open", authenticateJWT, handleWebEndpointOpen);
+    started = true;
+  }
+  registerWebEndpointRouter(router);
+}
+
+export function stopWebEndpointService(): void {
+  unregisterWebEndpointRouter();
+}