@hienlh/ppm 0.18.14 → 0.18.16

package/src/services/edge-forwarder.ts

@@ -1,191 +1,191 @@
-/**
- * Edge forwarder — owns the public port and pipes raw TCP to the server's
- * current loopback port.
- *
- * ─── CRITICAL INVARIANT: THIS PROCESS MUST NEVER SPAWN A CHILD ───
- *
- * On Windows a child spawned with fd stdio inherits every inheritable handle,
- * including a listening socket. When such a descendant is orphaned it keeps the
- * socket open, so the port stays in LISTEN under a dead PID and can never be
- * rebound — a "zombie port". The server hits this constantly because it spawns
- * chat/tool/MCP subprocesses; those inherit its listener and wedge its port.
- *
- * The edge exists precisely because it spawns nothing, so its socket cannot be
- * inherited and its port cannot zombie. cloudflared can therefore stay pinned
- * to it forever and the public URL stops rotating. Adding any child-process
- * call to this file reintroduces the exact bug the edge was built to remove.
- *
- * Raw TCP, never HTTP: WebSocket upgrades and SSE streams pass through
- * untouched, and forwarded requests keep the headers cloudflared set.
- */
-import net from "node:net";
-import { resolveTargetPort, _resetTargetCache } from "./edge-target-resolver.ts";
-
-/** Delay between attempts while waiting for the server to come back. */
-const CONNECT_RETRY_MS = 250;
-/**
- * How long a client connection waits for a reachable server before giving up.
- * During an upgrade the server is briefly absent; refusing instantly would
- * surface as a broken page rather than a slow one.
- */
-const CONNECT_WINDOW_MS = 5000;
-
-const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));
-
-/** Resolve once the socket is connected; reject on the first connect error. */
-function connectUpstream(port: number): Promise<net.Socket> {
-  return new Promise((resolvePromise, reject) => {
-    const upstream = net.connect({
-      port,
-      host: "127.0.0.1", // never taken from status.json — loopback only
-      allowHalfOpen: true,
-    });
-    const onError = (err: Error) => {
-      upstream.destroy();
-      reject(err);
-    };
-    upstream.once("error", onError);
-    upstream.once("connect", () => {
-      upstream.removeListener("error", onError);
-      resolvePromise(upstream);
-    });
-  });
-}
-
-/**
- * Largest amount of client data held while waiting for the server. Bounds the
- * memory a slow upload can pin during an upgrade window.
- */
-const MAX_PENDING_BYTES = 8 * 1024 * 1024;
-
-/**
- * Pipe one client connection to the server.
- *
- * Resolving the upstream is async, so the first bytes usually arrive before
- * there is anywhere to send them. They must be captured by a `data` listener
- * attached synchronously on this very tick and replayed once the upstream
- * exists.
- *
- * `socket.pause()` is NOT a substitute — verified against Bun 1.3.13: bytes
- * that arrive while paused are lost, and a later `pipe()`/`resume()` does not
- * bring them back. A socket with no `data` listener drops them outright too.
- * Both variants silently swallowed the first request in testing.
- */
-async function handleConnection(
-  client: net.Socket,
-  connectWindowMs: number,
-): Promise<void> {
-  const pending: Buffer[] = [];
-  let pendingBytes = 0;
-  let clientEnded = false;
-  const onEarlyData = (chunk: Buffer) => {
-    pendingBytes += chunk.length;
-    if (pendingBytes > MAX_PENDING_BYTES) {
-      client.destroy();
-      return;
-    }
-    pending.push(chunk);
-  };
-  client.on("data", onEarlyData);
-  client.once("end", () => { clientEnded = true; });
-
-  let upstream: net.Socket | null = null;
-  let clientGone = false;
-
-  // A socket error with no listener is an unhandled 'error' event, which would
-  // take the whole edge process — and the public port — down with it.
-  const teardown = () => {
-    upstream?.destroy();
-    client.destroy();
-  };
-  client.on("error", teardown);
-  client.once("close", () => { clientGone = true; });
-
-  const deadline = Date.now() + connectWindowMs;
-  while (!clientGone && Date.now() < deadline) {
-    const port = resolveTargetPort();
-    if (port !== null) {
-      try {
-        upstream = await connectUpstream(port);
-        break;
-      } catch {
-        // Server not up yet (or moved mid-restart) — re-resolve and retry.
-        _resetTargetCache();
-      }
-    }
-    await sleep(CONNECT_RETRY_MS);
-  }
-
-  if (clientGone) return;
-  if (!upstream) {
-    client.destroy();
-    return;
-  }
-
-  upstream.on("error", teardown);
-
-  // Replay what arrived during the connect, then hand both directions to the
-  // stream machinery. Removing the listener and piping happen in the same tick,
-  // so no chunk can slip through the gap.
-  client.off("data", onEarlyData);
-  for (const chunk of pending) upstream.write(chunk);
-  pending.length = 0;
-
-  client.pipe(upstream);
-  upstream.pipe(client);
-
-  // The client may have half-closed while we were still connecting.
-  if (clientEnded) upstream.end();
-}
-
-export interface EdgeForwarderOptions {
-  publicPort: number;
-  host?: string;
-  /** How long a client waits for a reachable server. Defaults to 5s. */
-  connectWindowMs?: number;
-}
-
-/**
- * Bind the public port and start forwarding. Resolves once listening so the
- * caller can fail fast when the port is unavailable.
- */
-export function startEdgeForwarder(
-  opts: EdgeForwarderOptions,
-): Promise<net.Server> {
-  const host = opts.host ?? "0.0.0.0";
-  const connectWindowMs = opts.connectWindowMs ?? CONNECT_WINDOW_MS;
-  return new Promise((resolvePromise, reject) => {
-    // allowHalfOpen keeps the other direction alive when one side sends FIN —
-    // long-lived chat WS and SSE streams depend on it.
-    const server = net.createServer({ allowHalfOpen: true }, (client) => {
-      void handleConnection(client, connectWindowMs);
-    });
-    server.once("error", reject);
-    server.listen(opts.publicPort, host, () => {
-      server.removeListener("error", reject);
-      resolvePromise(server);
-    });
-  });
-}
-
-// ─── Process entry ─────────────────────────────────────────────────────
-// Mirrors the `__serve__` guard in src/server/index.ts: the supervisor
-// re-invokes the binary as `<bin> __edge__ <publicPort> <host>`.
-if (process.argv.includes("__edge__")) {
-  const idx = process.argv.indexOf("__edge__");
-  const publicPort = parseInt(process.argv[idx + 1] ?? "", 10);
-  const host = process.argv[idx + 2] ?? "0.0.0.0";
-
-  if (!Number.isInteger(publicPort) || publicPort <= 0 || publicPort > 65535) {
-    process.stderr.write(`[edge] Invalid public port: ${process.argv[idx + 1]}\n`);
-    process.exit(2);
-  }
-
-  try {
-    await startEdgeForwarder({ publicPort, host });
-    process.stderr.write(`[edge] Listening on ${host}:${publicPort}\n`);
-  } catch (err) {
-    process.stderr.write(`[edge] Failed to bind ${host}:${publicPort}: ${err}\n`);
-    process.exit(1);
-  }
-}
+/**
+ * Edge forwarder — owns the public port and pipes raw TCP to the server's
+ * current loopback port.
+ *
+ * ─── CRITICAL INVARIANT: THIS PROCESS MUST NEVER SPAWN A CHILD ───
+ *
+ * On Windows a child spawned with fd stdio inherits every inheritable handle,
+ * including a listening socket. When such a descendant is orphaned it keeps the
+ * socket open, so the port stays in LISTEN under a dead PID and can never be
+ * rebound — a "zombie port". The server hits this constantly because it spawns
+ * chat/tool/MCP subprocesses; those inherit its listener and wedge its port.
+ *
+ * The edge exists precisely because it spawns nothing, so its socket cannot be
+ * inherited and its port cannot zombie. cloudflared can therefore stay pinned
+ * to it forever and the public URL stops rotating. Adding any child-process
+ * call to this file reintroduces the exact bug the edge was built to remove.
+ *
+ * Raw TCP, never HTTP: WebSocket upgrades and SSE streams pass through
+ * untouched, and forwarded requests keep the headers cloudflared set.
+ */
+import net from "node:net";
+import { resolveTargetPort, _resetTargetCache } from "./edge-target-resolver.ts";
+
+/** Delay between attempts while waiting for the server to come back. */
+const CONNECT_RETRY_MS = 250;
+/**
+ * How long a client connection waits for a reachable server before giving up.
+ * During an upgrade the server is briefly absent; refusing instantly would
+ * surface as a broken page rather than a slow one.
+ */
+const CONNECT_WINDOW_MS = 5000;
+
+const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));
+
+/** Resolve once the socket is connected; reject on the first connect error. */
+function connectUpstream(port: number): Promise<net.Socket> {
+  return new Promise((resolvePromise, reject) => {
+    const upstream = net.connect({
+      port,
+      host: "127.0.0.1", // never taken from status.json — loopback only
+      allowHalfOpen: true,
+    });
+    const onError = (err: Error) => {
+      upstream.destroy();
+      reject(err);
+    };
+    upstream.once("error", onError);
+    upstream.once("connect", () => {
+      upstream.removeListener("error", onError);
+      resolvePromise(upstream);
+    });
+  });
+}
+
+/**
+ * Largest amount of client data held while waiting for the server. Bounds the
+ * memory a slow upload can pin during an upgrade window.
+ */
+const MAX_PENDING_BYTES = 8 * 1024 * 1024;
+
+/**
+ * Pipe one client connection to the server.
+ *
+ * Resolving the upstream is async, so the first bytes usually arrive before
+ * there is anywhere to send them. They must be captured by a `data` listener
+ * attached synchronously on this very tick and replayed once the upstream
+ * exists.
+ *
+ * `socket.pause()` is NOT a substitute — verified against Bun 1.3.13: bytes
+ * that arrive while paused are lost, and a later `pipe()`/`resume()` does not
+ * bring them back. A socket with no `data` listener drops them outright too.
+ * Both variants silently swallowed the first request in testing.
+ */
+async function handleConnection(
+  client: net.Socket,
+  connectWindowMs: number,
+): Promise<void> {
+  const pending: Buffer[] = [];
+  let pendingBytes = 0;
+  let clientEnded = false;
+  const onEarlyData = (chunk: Buffer) => {
+    pendingBytes += chunk.length;
+    if (pendingBytes > MAX_PENDING_BYTES) {
+      client.destroy();
+      return;
+    }
+    pending.push(chunk);
+  };
+  client.on("data", onEarlyData);
+  client.once("end", () => { clientEnded = true; });
+
+  let upstream: net.Socket | null = null;
+  let clientGone = false;
+
+  // A socket error with no listener is an unhandled 'error' event, which would
+  // take the whole edge process — and the public port — down with it.
+  const teardown = () => {
+    upstream?.destroy();
+    client.destroy();
+  };
+  client.on("error", teardown);
+  client.once("close", () => { clientGone = true; });
+
+  const deadline = Date.now() + connectWindowMs;
+  while (!clientGone && Date.now() < deadline) {
+    const port = resolveTargetPort();
+    if (port !== null) {
+      try {
+        upstream = await connectUpstream(port);
+        break;
+      } catch {
+        // Server not up yet (or moved mid-restart) — re-resolve and retry.
+        _resetTargetCache();
+      }
+    }
+    await sleep(CONNECT_RETRY_MS);
+  }
+
+  if (clientGone) return;
+  if (!upstream) {
+    client.destroy();
+    return;
+  }
+
+  upstream.on("error", teardown);
+
+  // Replay what arrived during the connect, then hand both directions to the
+  // stream machinery. Removing the listener and piping happen in the same tick,
+  // so no chunk can slip through the gap.
+  client.off("data", onEarlyData);
+  for (const chunk of pending) upstream.write(chunk);
+  pending.length = 0;
+
+  client.pipe(upstream);
+  upstream.pipe(client);
+
+  // The client may have half-closed while we were still connecting.
+  if (clientEnded) upstream.end();
+}
+
+export interface EdgeForwarderOptions {
+  publicPort: number;
+  host?: string;
+  /** How long a client waits for a reachable server. Defaults to 5s. */
+  connectWindowMs?: number;
+}
+
+/**
+ * Bind the public port and start forwarding. Resolves once listening so the
+ * caller can fail fast when the port is unavailable.
+ */
+export function startEdgeForwarder(
+  opts: EdgeForwarderOptions,
+): Promise<net.Server> {
+  const host = opts.host ?? "0.0.0.0";
+  const connectWindowMs = opts.connectWindowMs ?? CONNECT_WINDOW_MS;
+  return new Promise((resolvePromise, reject) => {
+    // allowHalfOpen keeps the other direction alive when one side sends FIN —
+    // long-lived chat WS and SSE streams depend on it.
+    const server = net.createServer({ allowHalfOpen: true }, (client) => {
+      void handleConnection(client, connectWindowMs);
+    });
+    server.once("error", reject);
+    server.listen(opts.publicPort, host, () => {
+      server.removeListener("error", reject);
+      resolvePromise(server);
+    });
+  });
+}
+
+// ─── Process entry ─────────────────────────────────────────────────────
+// Mirrors the `__serve__` guard in src/server/index.ts: the supervisor
+// re-invokes the binary as `<bin> __edge__ <publicPort> <host>`.
+if (process.argv.includes("__edge__")) {
+  const idx = process.argv.indexOf("__edge__");
+  const publicPort = parseInt(process.argv[idx + 1] ?? "", 10);
+  const host = process.argv[idx + 2] ?? "0.0.0.0";
+
+  if (!Number.isInteger(publicPort) || publicPort <= 0 || publicPort > 65535) {
+    process.stderr.write(`[edge] Invalid public port: ${process.argv[idx + 1]}\n`);
+    process.exit(2);
+  }
+
+  try {
+    await startEdgeForwarder({ publicPort, host });
+    process.stderr.write(`[edge] Listening on ${host}:${publicPort}\n`);
+  } catch (err) {
+    process.stderr.write(`[edge] Failed to bind ${host}:${publicPort}: ${err}\n`);
+    process.exit(1);
+  }
+}

package/src/services/fs-credential-path-guard.ts

@@ -2,6 +2,7 @@ import { resolve, sep } from "node:path";
 import { homedir } from "node:os";
 import { getPpmDir } from "./ppm-dir.ts";
 import { getUploadsDir } from "./chat-upload-storage.service.ts";
+import { getBackupsDir } from "./db-backup/db-backup-paths.ts";
 import { realPathOrSelf } from "./fs-ops/fs-real-path.ts";
 
 /**
@@ -11,6 +12,9 @@ import { realPathOrSelf } from "./fs-ops/fs-real-path.ts";
  * cert). Split out of `fs-path-guard.service.ts` so this one seam — "is this
  * a credential path" — stays a single file instead of growing alongside the
  * platform-allowlist and protected-root logic that lives there.
+ *
+ * Three roots, not two: the PPM config DB, `~/.cloudflared`, and the database
+ * snapshot directory, which holds full copies of the first one.
  */
 
 /** Case-insensitive prefix test on Windows/macOS-style paths. */
@@ -50,9 +54,23 @@ export function isCloudflaredDirPath(resolved: string): boolean {
   return isInside(resolved, resolve(homedir(), ".cloudflared"));
 }
 
+/**
+ * True when the path is the database-snapshot directory or anything inside it.
+ *
+ * A snapshot is a byte-for-byte copy of the config database, so it carries the
+ * same provider keys, encrypted accounts, and auth token. In production that
+ * directory deliberately sits OUTSIDE `~/.ppm` (so it survives a wipe of the
+ * PPM directory), which means `isPpmDirPath` does not cover it — without this
+ * branch, moving snapshots out of the PPM directory would have quietly opened
+ * a credential-read hole through the generic file routes.
+ */
+export function isDbBackupsDirPath(resolved: string): boolean {
+  return isInside(resolved, getBackupsDir());
+}
+
 /** True when the path holds credential material a generic file route must never serve or relocate. */
 export function isCredentialPath(resolved: string): boolean {
-  return isPpmDirPath(resolved) || isCloudflaredDirPath(resolved);
+  return isPpmDirPath(resolved) || isCloudflaredDirPath(resolved) || isDbBackupsDirPath(resolved);
 }
 
 /**

package/src/services/named-tunnel/cloudflare-dns-api.ts

@@ -31,10 +31,26 @@ export async function fetchDnsRecords(zoneID: string, apiToken: string, name: st
   return Array.isArray(json.result) ? json.result : [];
 }
 
-/** UUID of an existing, non-deleted tunnel named `name` — null if none exists yet. */
-export async function fetchTunnelIdByName(accountID: string, apiToken: string, name: string): Promise<string | null> {
+export interface ExistingTunnel {
+  id: string;
+  /** Connectors currently registered on it — non-zero means something is serving it right now. */
+  activeConnections: number;
+}
+
+/** An existing, non-deleted tunnel named `name` — null if none exists yet. */
+export async function fetchTunnelByName(
+  accountID: string,
+  apiToken: string,
+  name: string,
+): Promise<ExistingTunnel | null> {
   const url = `https://api.cloudflare.com/client/v4/accounts/${encodeURIComponent(accountID)}/cfd_tunnel?name=${encodeURIComponent(name)}&is_deleted=false`;
-  const json = await cfGet<{ success: true; result?: Array<{ id?: string }> }>(url, apiToken);
+  const json = await cfGet<{ success: true; result?: Array<{ id?: string; connections?: unknown[] }> }>(url, apiToken);
   const found = (json.result ?? []).find((t) => typeof t.id === "string" && t.id);
-  return found?.id ?? null;
+  if (!found?.id) return null;
+  return { id: found.id, activeConnections: Array.isArray(found.connections) ? found.connections.length : 0 };
+}
+
+/** UUID of an existing, non-deleted tunnel named `name` — null if none exists yet. */
+export async function fetchTunnelIdByName(accountID: string, apiToken: string, name: string): Promise<string | null> {
+  return (await fetchTunnelByName(accountID, apiToken, name))?.id ?? null;
 }

package/src/services/named-tunnel/cloudflared-login.service.ts

@@ -98,7 +98,7 @@ export async function startLogin(opts: { relogin?: boolean } = {}): Promise<Logi
       // itself failed to run, so cert.pem is left untouched and the user
       // gets a clear, retryable error instead of an unwanted fresh login.
       if (liveCheck === "unreachable") {
-        setTerminal("error", "không kết nối được Cloudflare để kiểm tra đăng nhập — thử lại");
+        setTerminal("error", "could not reach Cloudflare to check the sign-in — try again");
         return getLoginSnapshot();
       }
 
@@ -117,7 +117,7 @@ export async function startLogin(opts: { relogin?: boolean } = {}): Promise<Logi
       renameCertAside();
     }
   } else {
-    // Explicit ?relogin=1 — the user deliberately pressed "Đăng nhập lại",
+    // Explicit ?relogin=1 — the user deliberately pressed "Sign in again",
     // so this renames unconditionally regardless of network reachability.
     renameCertAside();
   }

package/src/services/named-tunnel/named-tunnel-args.ts

@@ -8,10 +8,10 @@
  * passed via `--token-file`, so it can't leak through `ppm status`, a process
  * list, or a crash dump that captures cmdlines.
  */
-import { resolve } from "node:path";
+import { resolve, basename } from "node:path";
 import { hostname as osHostname } from "node:os";
 import { existsSync, mkdirSync, writeFileSync, chmodSync } from "node:fs";
-import { getPpmDir } from "../ppm-dir.ts";
+import { getPpmDir, isIsolatedPpmHome } from "../ppm-dir.ts";
 import { getOriginCertPath } from "./cloudflared-cert.ts";
 
 const namedTunnelConfigPath = () => resolve(getPpmDir(), "cloudflared-named.yml");
@@ -106,5 +106,19 @@ export function namedRunArgs(token: string, port: number): string[] {
 /** Stable, DNS-safe tunnel name derived from this machine's hostname. */
 export function tunnelNameForHost(): string {
   const sanitized = osHostname().toLowerCase().replace(/[^a-z0-9-]/g, "-");
-  return `ppm-${sanitized}`.slice(0, 32);
+  // A second PPM on the same machine (a dev instance under its own PPM_HOME)
+  // would otherwise derive the identical name, "reuse" the production tunnel,
+  // and register a second connector on it — Cloudflare then load-balances the
+  // production hostname across both, so the domain intermittently serves the
+  // wrong instance. Observed live. The PPM_HOME directory is what actually
+  // distinguishes the two, so it belongs in the name.
+  const suffix = ppmHomeSuffix();
+  return `ppm-${sanitized}${suffix}`.slice(0, 32);
+}
+
+/** `-<dir>` for a non-default PPM_HOME, empty for the real `~/.ppm`. */
+function ppmHomeSuffix(): string {
+  if (!isIsolatedPpmHome()) return "";
+  const dir = basename(getPpmDir()).toLowerCase().replace(/^\.+/, "").replace(/[^a-z0-9-]/g, "-");
+  return dir ? `-${dir}` : "-alt";
 }

package/src/services/named-tunnel/named-tunnel-setup.service.ts

@@ -8,7 +8,7 @@
  */
 import { readOriginCertState } from "./cloudflared-cert.ts";
 import { fetchZoneName } from "./cloudflare-zone-api.ts";
-import { fetchDnsRecords, fetchTunnelIdByName } from "./cloudflare-dns-api.ts";
+import { fetchDnsRecords, fetchTunnelByName } from "./cloudflare-dns-api.ts";
 import { proposeHostname, validateHostname } from "./hostname-rules.ts";
 import { createTunnelArgs, routeDnsArgs, tunnelTokenArgs, tunnelNameForHost } from "./named-tunnel-args.ts";
 import { runCloudflared } from "./cloudflared-exec.ts";
@@ -110,7 +110,25 @@ async function runSetupInner(hostname: string): Promise<SetupOutcome> {
   const tunnelName = tunnelNameForHost();
 
   broadcastGlobalEvent({ type: "tunnel:setup_step", step: "precheck", message: "checking for a DNS collision" });
-  const existingId = await fetchTunnelIdByName(accountID, apiToken, tunnelName);
+  const existing = await fetchTunnelByName(accountID, apiToken, tunnelName);
+  const existingId = existing?.id ?? null;
+
+  // Reusing a same-named tunnel is normal (re-running setup, changing the
+  // prefix). Reusing one that ANOTHER PPM is currently serving is not: both
+  // connectors register on the same tunnel and Cloudflare spreads requests
+  // across them, so each hostname on that tunnel intermittently answers from
+  // the wrong instance. Only the installation that already owns this tunnel in
+  // its own config may take it over.
+  // Read the raw row, not `resolveTunnelConfig`: that degrades to "quick" when
+  // named mode is off, which would hide the fact that this installation still
+  // owns the tunnel and would wrongly block its own re-setup.
+  const ownsTunnelAlready = configService.get("tunnel")?.namedTunnelName === tunnelName;
+  if (existing && existing.activeConnections > 0 && !ownsTunnelAlready) {
+    throw new SetupError(
+      409,
+      `a tunnel named "${tunnelName}" is already running elsewhere — stop that PPM instance first, or run this one with its own PPM_HOME`,
+    );
+  }
   const records = await fetchDnsRecords(zoneID, apiToken, hostname);
   let overwrite = false;
   if (records.length > 0) {