centralu 0.1.0-beta.12 → 0.1.0-beta.14

package/bin/centralu.mjs

@@ -25,6 +25,7 @@
 import { execFileSync, spawn } from 'node:child_process'
 import { copyDiffers, isNewer } from './semver.mjs'
 import {
+  asideCopies,
   busyMessage,
   earlyExitMessage,
   executableIn,
@@ -41,8 +42,18 @@ import {
   webview2Status,
   windowsInstall,
 } from './platform.mjs'
-import { parseServeArgs, printConnection, rotateToken, runServe, SERVE_HELP } from './serve.mjs'
-import { chmodSync, cpSync, existsSync, mkdirSync, readFileSync, realpathSync, renameSync, rmSync, writeFileSync } from 'node:fs'
+import {
+  decodeChildSpec,
+  parseServeArgs,
+  printConnection,
+  rotateToken,
+  runDetach,
+  runDetachedChild,
+  runServe,
+  runStop,
+  SERVE_HELP,
+} from './serve.mjs'
+import { chmodSync, cpSync, existsSync, mkdirSync, readdirSync, readFileSync, realpathSync, renameSync, rmSync, writeFileSync } from 'node:fs'
 import { createRequire } from 'node:module'
 import { homedir } from 'node:os'
 import { dirname, join } from 'node:path'
@@ -322,6 +333,7 @@ function installDesktopEntry(app) {
 function installWindows(app) {
   const staging = `${WIN.dir}.new`
   const aside = `${WIN.dir}.old-${Date.now()}`
+  sweepAsideCopies()
   rmSync(staging, { recursive: true, force: true })
   mkdirSync(dirname(WIN.dir), { recursive: true })
   cpSync(app, staging, { recursive: true })
@@ -358,6 +370,32 @@ function installWindows(app) {
   console.log('It can now be found in Start and Windows search, and `centralu` starts this copy too.')
 }
 
+/**
+ * Removes the copies earlier installs renamed aside (`Centralu.old-<time>`), whichever can go.
+ *
+ * An update from inside the app renames the running copy aside, and Windows will not delete a
+ * running exe, so that copy stays until the app quits. Nothing else would ever remove it, and
+ * each in-app update would leave one more (about 70 files each, measured 2026-10-08). One still
+ * in use is left for the next install; deleting what can be deleted from it does not disturb the
+ * app running from it, which reaches its files by the folder's original path, now the new copy.
+ */
+function sweepAsideCopies() {
+  const parent = dirname(WIN.dir)
+  let names
+  try {
+    names = readdirSync(parent)
+  } catch {
+    return
+  }
+  for (const name of asideCopies(names)) {
+    try {
+      rmSync(join(parent, name), { recursive: true, force: true })
+    } catch {
+      // Still in use: the next install tries again
+    }
+  }
+}
+
 function uninstall() {
   const paths = installedPaths(PLATFORM, process.env, HOME)
   const present = paths.filter((p) => existsSync(p))
@@ -551,9 +589,21 @@ switch (cmd) {
         process.exitCode = rotateToken({ env: process.env, home: HOME })
         break
       }
+      // A launcher `--detach` started: what it carried (the data folder, a host entry) applies before
+      // anything below reads the environment
+      const spec = args.child !== undefined ? decodeChildSpec(args.child) : null
+      if (args.child !== undefined && !spec) {
+        console.error('centralu serve: --detached-child is how --detach starts serve; this one could not be read')
+        process.exit(2)
+      }
+      if (spec) Object.assign(process.env, spec.env)
       // realpath: npm's bin is a symlink to this file, and the launcher must not depend on the link
       const opts = { env: process.env, home: HOME, entry: requireHostEntry(), version: pkg.version, cliPath: realpathSync(process.argv[1]) }
-      process.exitCode = args.mode === 'connection' ? await printConnection(opts) : await runServe({ ...opts, port: args.port })
+      if (args.mode === 'connection') process.exitCode = await printConnection(opts)
+      else if (args.mode === 'detach') process.exitCode = await runDetach({ ...opts, port: args.port })
+      else if (args.mode === 'stop') process.exitCode = await runStop(opts)
+      else if (spec) process.exitCode = await runDetachedChild({ ...opts, spec, port: args.port })
+      else process.exitCode = await runServe({ ...opts, port: args.port })
     } catch (e) {
       // A state file that cannot be read or written: one sentence, not a stack, and nothing on stdout
       console.error(`centralu serve: ${e?.message ?? e}`)

package/bin/platform.mjs

@@ -137,6 +137,16 @@ export function windowsInstall(env, home) {
   }
 }
 
+/**
+ * Which of the names in `%LOCALAPPDATA%\Programs` are copies an earlier `centralu install` renamed
+ * aside (`Centralu.old-<time>`). Only those: the folder is shared with every other per-user
+ * program, and `Centralu` itself and a `Centralu.new` being assembled are not leftovers.
+ */
+export function asideCopies(names) {
+  const prefix = `${APP_NAME}.old-`.toLowerCase()
+  return names.filter((n) => n.toLowerCase().startsWith(prefix) && /^\d+$/.test(n.slice(prefix.length)))
+}
+
 /**
  * What `centralu install` creates, per platform. `uninstall` removes exactly these and
  * `update` refreshes them when the first one exists.

package/bin/serve.mjs

@@ -17,8 +17,9 @@
  */
 import { spawn } from 'node:child_process'
 import { randomBytes } from 'node:crypto'
-import { chmodSync, existsSync, mkdirSync, openSync, readFileSync, renameSync, statSync, writeSync, closeSync } from 'node:fs'
-import { connect } from 'node:net'
+import { chmodSync, existsSync, mkdirSync, openSync, readFileSync, renameSync, rmSync, statSync, writeFileSync, writeSync, closeSync } from 'node:fs'
+import { connect, createServer } from 'node:net'
+import { userInfo } from 'node:os'
 import { dirname, join } from 'node:path'
 import { createInterface } from 'node:readline'
 
@@ -33,6 +34,70 @@ import { createInterface } from 'node:readline'
  */
 export const DEFAULT_SERVE_PORT = 17175
 
+/** How many ports from the recorded one `serve` tries when it is taken (#482); none at or above 20000 */
+export const SERVE_PORT_TRIES = 20
+
+/**
+ * The port `serve` starts the host on (#482): `{ port, from }`, `from` being the port it moved away
+ * from (null when it did not), or null when every candidate is taken.
+ *
+ * A port the person gave (`--port`) is used as given: they asked for that one. Otherwise the recorded
+ * port (else the default) when it is free, and when another program listens there, the next free
+ * one above it, below 20000 (the ranges `DEFAULT_SERVE_PORT` keeps out of). The host's ready line
+ * then records the port it got, so `--connection` reports it and the hub forwards to it.
+ *
+ * The usual other program is the other half of one Windows laptop: WSL2 relays a distro's
+ * 127.0.0.1 ports to Windows (`wslrelay.exe`), so a Windows install and a WSL install linked as two
+ * machines both wanted 17175, and whichever started second could not start at all.
+ * `taken(port)` answers whether something listens there.
+ */
+export async function chooseServePort({ asked, recorded, taken }) {
+  if (asked !== null && asked !== undefined) return { port: asked, from: null }
+  const first = recorded ?? DEFAULT_SERVE_PORT
+  for (let p = first; p < first + SERVE_PORT_TRIES && p <= 65535; p++) {
+    if (p !== first && p >= 20000) break
+    if (!(await taken(p))) return { port: p, from: p === first ? null : first }
+  }
+  return null
+}
+
+/** Whether something listens on 127.0.0.1:`port` here: a listen that fails with EADDRINUSE (or EACCES) */
+export function listensHere(port) {
+  return new Promise((resolve) => {
+    const srv = createServer()
+    srv.once('error', (e) => resolve(e?.code === 'EADDRINUSE' || e?.code === 'EACCES'))
+    srv.listen({ host: '127.0.0.1', port, exclusive: true }, () => srv.close(() => resolve(false)))
+  })
+}
+
+/**
+ * The TCP ports Windows listens on, from `netstat.exe -an -p TCP` output: a listening socket is the
+ * row whose foreign address is `0.0.0.0:0` (or `[::]:0`). Read that way rather than by the state
+ * column, which Windows translates (a Korean Windows does not print LISTENING everywhere).
+ */
+export function parseNetstatListening(text) {
+  const ports = new Set()
+  for (const line of String(text).split(/\r?\n/)) {
+    const m = /^\s*TCP\s+(?:127\.0\.0\.1|0\.0\.0\.0|\[::1?\]):(\d+)\s+(?:0\.0\.0\.0:0|\[::\]:0)\s/i.exec(line)
+    if (m) ports.add(Number(m[1]))
+  }
+  return ports
+}
+
+/**
+ * The ports the Windows side listens on, seen from inside a WSL distro (empty elsewhere, or when
+ * Windows cannot be asked). Inside the distro a port Windows holds looks free, yet a host there
+ * would be unreachable from the hub: the hub's forward ends on Windows' 127.0.0.1, where the
+ * Windows program answers instead (#482).
+ */
+async function windowsListening(env, platform) {
+  if (detachHow(platform, env) !== 'wsl') return new Set()
+  const netstat = windowsTool('wsl', env, ['System32', 'netstat.exe'])
+  if (!netstat) return new Set()
+  const r = await runCaptured(netstat, ['-an', '-p', 'TCP'], env)
+  return r.code === 0 ? parseNetstatListening(r.stdout) : new Set()
+}
+
 /**
  * The file in the data folder that holds the token and the port, mode 0600.
  *
@@ -43,36 +108,62 @@ export const DEFAULT_SERVE_PORT = 17175
  */
 export const SERVE_STATE_FILE = 'serve.json'
 
+/** What a detached serve writes: its own lines, and the host's until it is ready (host.log has the rest) */
+export const SERVE_LOG_FILE = 'serve.log'
+
+/** The running launcher's pid and port, so `--stop` can find it. Removed when it exits */
+export const SERVE_PID_FILE = 'serve.pid'
+
 export const SERVE_HELP = `centralu serve: run the Centralu host on this machine, without a window
 
   centralu serve                  Start the host in the foreground, on 127.0.0.1 only
   centralu serve --port <n>       ... on that port (kept for the next start; default ${DEFAULT_SERVE_PORT})
+  centralu serve --detach         Start it in the background, outside this session; return once it answers
+  centralu serve --stop           Ask the running host to stop, and wait until it has
   centralu serve --connection     Print how to reach it as one JSON line, then exit
   centralu serve --rotate-token   Replace the token (a running serve keeps the old one until restarted)
 
 The data folder is ~/.centralu (CC_DATA_DIR overrides it). The app on another computer reaches
-this host through an SSH local forward. Stop it with Ctrl+C or SIGTERM.
+this host through an SSH local forward, and starts it with --detach when it is not running.
+Stop a foreground one with Ctrl+C or SIGTERM, a detached one with --stop. A detached one writes
+how its start went to serve.log in the data folder; the host's own log is host.log.
 
-Both serve and --connection keep ~/.centralu/bin/centralu up to date: a launcher with absolute
-paths, for SSH sessions whose PATH has neither npm's global folder nor node.`
+serve, --detach and --connection keep ~/.centralu/bin/centralu up to date: a launcher with absolute
+paths, for SSH sessions whose PATH has neither npm's global folder nor node. Not when they run
+from an install the app made over SSH, which has a launcher of its own.`
+
+/** The flags that each make serve a different command */
+const MODE_FLAGS = { '--connection': 'connection', '--rotate-token': 'rotate', '--detach': 'detach', '--stop': 'stop' }
+const FLAG_OF = Object.fromEntries(Object.entries(MODE_FLAGS).map(([f, m]) => [m, f]))
 
 /**
- * What the arguments after `serve` ask for: `{ mode, port }` or `{ error }`.
+ * The argument a detached launcher is started with: what it needs from the `--detach` that started
+ * it, which a process created through WMI does not inherit (it gets the account's default
+ * environment, not this one). Never a secret: the token stays in serve.json. Internal, not in the help.
+ */
+const CHILD_FLAG = '--detached-child='
+
+/**
+ * What the arguments after `serve` ask for: `{ mode, port }` (plus `child` for a detached launcher)
+ * or `{ error }`.
  *
- * `mode` is `serve`, `connection`, `rotate` or `help`; `port` is null when not given. `--port` does not go
- * with `--connection`: the answer has to be the port `serve` uses, not one the caller names.
+ * `mode` is `serve`, `detach`, `stop`, `connection`, `rotate` or `help`; `port` is null when not
+ * given. `--port` goes with `serve` and `--detach` only: `--connection` has to report the port
+ * `serve` uses, not one the caller names, and `--stop` stops the one that runs.
  */
 export function parseServeArgs(argv) {
   let mode = 'serve'
   let port = null
+  let help = false
+  let child = null
   for (let i = 0; i < argv.length; i++) {
     const a = argv[i]
-    if (a === '--connection' || a === '--rotate-token') {
-      const m = a === '--connection' ? 'connection' : 'rotate'
-      if (mode !== 'serve' && mode !== 'help' && mode !== m) return { error: '--connection and --rotate-token are separate commands; pass one' }
-      if (mode !== 'help') mode = m
-    }
-    else if (a === '--help' || a === '-h') mode = 'help'
+    if (Object.hasOwn(MODE_FLAGS, a)) {
+      const m = MODE_FLAGS[a]
+      if (mode !== 'serve' && mode !== m) return { error: `${FLAG_OF[mode]} and ${a} are separate commands; pass one` }
+      mode = m
+    } else if (a === '--help' || a === '-h') help = true
+    else if (a.startsWith(CHILD_FLAG)) child = a.slice(CHILD_FLAG.length)
     else if (a === '--port' || a.startsWith('--port=')) {
       const raw = a === '--port' ? argv[++i] : a.slice('--port='.length)
       const n = parsePort(raw)
@@ -80,9 +171,15 @@ export function parseServeArgs(argv) {
       port = n
     } else return { error: `unknown option for serve: ${a}\n\n${SERVE_HELP}` }
   }
+  if (help) return { mode: 'help', port }
+  if (child !== null) {
+    if (mode !== 'serve') return { error: `${CHILD_FLAG.slice(0, -1)} is how --detach starts serve, not an option of ${FLAG_OF[mode]}` }
+    return { mode, port, child }
+  }
   if ((mode === 'connection' || mode === 'rotate') && port !== null) {
     return { error: '--port goes with `centralu serve`; `--connection` reports the port serve listens on' }
   }
+  if (mode === 'stop' && port !== null) return { error: '--stop stops the serve that is running, on the port it recorded; it takes no --port' }
   return { mode, port }
 }
 
@@ -214,9 +311,16 @@ export function ensureLauncher(dataDir, { platform = process.platform, execPath
   return file
 }
 
-/** ensureLauncher, with a failure said on stderr instead of thrown: the launcher is a convenience */
-function keepLauncher(dataDir, cliPath) {
-  if (!cliPath) return
+/**
+ * ensureLauncher, with a failure said on stderr instead of thrown: the launcher is a convenience.
+ *
+ * Not under the managed launcher (`CENTRALU_MANAGED=1`, docs/plans/remote-hub.md §10.1): an install
+ * the hub made over ssh has its own launcher in `<data>/remote/bin/`, which reads `current`. Writing
+ * `<data>/bin/centralu` from there would point an npm install's launcher at one managed version, and
+ * the two would rewrite each other's file on every start.
+ */
+function keepLauncher(dataDir, cliPath, env) {
+  if (!cliPath || env?.CENTRALU_MANAGED === '1') return
   try {
     ensureLauncher(dataDir, { cliPath })
   } catch (e) {
@@ -269,9 +373,121 @@ export function protocolVersionOf(entry) {
   return null
 }
 
-/** The `--connection` answer. The shape is the client's interface (docs/agent-host.md §4.7) */
-export function connectionLine({ port, token, version, protocolVersion, dataDir, hostRunning }) {
-  return JSON.stringify({ v: 1, port, token, version, protocolVersion, dataDir, hostRunning })
+/**
+ * The `--connection` answer. The shape is the client's interface (docs/agent-host.md §4.7).
+ * `install` (additive, the line stays `v: 1`) says what runs this command (`installInfo`); `detach`
+ * is added by `--detach` only.
+ */
+export function connectionLine({ port, token, version, protocolVersion, dataDir, hostRunning, install, detach }) {
+  return JSON.stringify({
+    v: 1,
+    port,
+    token,
+    version,
+    protocolVersion,
+    dataDir,
+    hostRunning,
+    ...(install ? { install } : {}),
+    ...(detach ? { detach } : {}),
+  })
+}
+
+/**
+ * The install the hub manages on this machine (docs/plans/remote-hub.md §10.1): side-by-side
+ * versions, the pinned Node they run on, and two pointer files naming what runs and what to roll
+ * back to. The hub's installer writes it; `serve` only reads it.
+ */
+export function remoteLayout(dataDir) {
+  const root = join(dataDir, 'remote')
+  return {
+    root,
+    current: join(root, 'current'),
+    previous: join(root, 'previous'),
+    versions: join(root, 'versions'),
+    node: join(root, 'node'),
+    bin: join(root, 'bin'),
+  }
+}
+
+/** A word a pointer file may hold: a folder name, nothing a path could escape with */
+const POINTER_WORD = /^[0-9A-Za-z][0-9A-Za-z.+-]{0,63}$/
+
+/** `current` / `previous`: one line, "<centralu version> <node version>". null for anything else */
+export function parsePointer(text) {
+  const m = /^\s*(\S+)[ \t]+(\S+)\s*$/.exec(String(text ?? ''))
+  if (!m || !POINTER_WORD.test(m[1]) || !POINTER_WORD.test(m[2])) return null
+  return { version: m[1], node: m[2] }
+}
+
+/**
+ * The connection line's `install` field (docs/plans/remote-hub.md §10.5): whether this command was
+ * started by the managed launcher, the managed install's `current` and `previous`, and the Node
+ * running this command. Read from the machine at every ask, never kept by the hub (S12).
+ */
+export function installInfo(dataDir, env, nodeVersion = process.versions.node) {
+  const l = remoteLayout(dataDir)
+  const read = (file) => {
+    try {
+      return parsePointer(readFileSync(file, 'utf8'))
+    } catch {
+      return null
+    }
+  }
+  return { managed: env.CENTRALU_MANAGED === '1', current: read(l.current), previous: read(l.previous), node: nodeVersion }
+}
+
+/**
+ * The managed launcher, `<data>/remote/bin/centralu` (`centralu.cmd` on Windows), as the installer
+ * writes it (docs/plans/remote-hub.md §10.1, S8).
+ *
+ * **Its text never changes.** It reads `current` at every start and runs that version's
+ * `centralu.mjs` on that version's Node, so switching versions is one file replaced by a rename,
+ * atomic on both systems. A launcher rewritten per version would be replaced while it may be
+ * running, and cmd.exe reads a batch file by offset as it runs. It sets `CENTRALU_MANAGED=1`, so
+ * the `serve` it starts leaves `<data>/bin/` to an npm install on the same machine.
+ *
+ * Both refuse a pointer that holds anything but plain version words, so an edited `current` cannot
+ * point it outside the install.
+ */
+export function managedLauncherScript(platform) {
+  if (platform === 'win32') {
+    return (
+      [
+        '@echo off',
+        'rem The Centralu launcher for the install the app manages over SSH. It never changes: it reads',
+        'rem "current" at every start, so switching versions replaces that file and never this one.',
+        'setlocal',
+        'set "CENTRALU_REMOTE=%~dp0.."',
+        'set "CENTRALU_LINE="',
+        'set "CENTRALU_V="',
+        'set "CENTRALU_N="',
+        'set /p CENTRALU_LINE=<"%CENTRALU_REMOTE%\\current"',
+        'for /f "tokens=1,2" %%a in ("%CENTRALU_LINE%") do (set "CENTRALU_V=%%a" & set "CENTRALU_N=%%b")',
+        'if not defined CENTRALU_N goto none',
+        'echo %CENTRALU_V% %CENTRALU_N%| findstr /r /c:"^[0-9A-Za-z][0-9A-Za-z.+-]* [0-9A-Za-z][0-9A-Za-z.+-]*$" >nul || goto none',
+        'set CENTRALU_MANAGED=1',
+        '"%CENTRALU_REMOTE%\\node\\v%CENTRALU_N%\\node.exe" "%CENTRALU_REMOTE%\\versions\\%CENTRALU_V%\\node_modules\\centralu\\bin\\centralu.mjs" %*',
+        'exit /b %ERRORLEVEL%',
+        ':none',
+        'echo Centralu: no version is installed in %CENTRALU_REMOTE% 1>&2',
+        'exit /b 1',
+      ].join('\r\n') + '\r\n'
+    )
+  }
+  return [
+    '#!/bin/sh',
+    '# The Centralu launcher for the install the app manages over SSH. It never changes: it reads',
+    '# `current` at every start, so switching versions replaces that file and never this one.',
+    'remote=$(CDPATH= cd -- "$(dirname -- "$0")/.." && pwd -P) || exit 1',
+    'v= n=',
+    'read -r v n < "$remote/current" 2>/dev/null',
+    'case "$v" in ""|[!0-9A-Za-z]*|*[!0-9A-Za-z.+-]*) echo "Centralu: no version is installed in $remote" >&2; exit 1 ;; esac',
+    'case "$n" in ""|[!0-9A-Za-z]*|*[!0-9A-Za-z.+-]*) echo "Centralu: no version is installed in $remote" >&2; exit 1 ;; esac',
+    'CENTRALU_MANAGED=1',
+    'export CENTRALU_MANAGED',
+    'exec "$remote/node/v$n/bin/node" "$remote/versions/$v/node_modules/centralu/bin/centralu.mjs" "$@"',
+    '',
+  ].join('\n')
 }
 
 /**
@@ -285,12 +501,16 @@ export function connectionLine({ port, token, version, protocolVersion, dataDir,
  * - Headless: no `DISPLAY` / `WAYLAND_DISPLAY`. Nothing the host or an agent starts gets a screen
  *   to open a window or a keyring prompt on, so a tool that would ask falls back to its file store
  *   or fails with a message in the log instead of waiting on a dialog nobody can see.
- * - The keeper's and dev mode's variables are dropped: this host is neither.
+ * - The keeper's and dev mode's variables are dropped: this host is neither. So is
+ *   `CENTRALU_MANAGED`, which is the launcher's business and would otherwise reach every agent.
+ * - `CC_SERVE=1` lets the host take `host.stop` (`--stop`); a host the app runs refuses it.
  */
 export function hostEnv(env, { token, dataDir, version, hostDir }) {
   const out = { ...env }
-  for (const k of ['DISPLAY', 'WAYLAND_DISPLAY', 'CC_KEEPER', 'CC_FRONT_DOOR', 'CC_DEV', 'CC_HOST_TOKEN', 'CC_HOST_SOURCE']) delete out[k]
+  for (const k of ['DISPLAY', 'WAYLAND_DISPLAY', 'CC_KEEPER', 'CC_FRONT_DOOR', 'CC_DEV', 'CC_HOST_TOKEN', 'CC_HOST_SOURCE', 'CENTRALU_MANAGED']) delete out[k]
   out.CC_DATA_DIR = dataDir
+  // The host `host.stop` may end (`centralu serve --stop`): nothing else stops it in order on Windows
+  out.CC_SERVE = '1'
   out.CC_HOST_TOKEN = token
   out.CC_HOST_SOURCE = JSON.stringify({ version, bundlePath: hostDir })
   return out
@@ -432,9 +652,33 @@ function isAlive(pid) {
   }
 }
 
+/**
+ * Where `say` writes: stderr, except in a detached launcher, which nobody reads the stderr of and
+ * which writes serve.log instead (`runDetachedChild`).
+ */
+let sink = (text) => process.stderr.write(text)
+
 /** Writes a line to stderr, prefixed so it reads apart from the host's own log lines */
 function say(line) {
-  process.stderr.write(`[centralu serve] ${line}\n`)
+  sink(`[centralu serve] ${line}\n`)
+}
+
+/** What `--connection` answers, as fields: the running host's versions when one answers, else the installed ones */
+async function describe({ env, dataDir, state, entry, version }) {
+  const port = state.port ?? DEFAULT_SERVE_PORT
+  const installed = protocolVersionOf(entry)
+  const probe = await probeHost({ port, token: state.token, protocolVersion: installed })
+  const protocolVersion = (probe.running ? probe.protocolVersion : undefined) ?? installed
+  if (protocolVersion === null) return null
+  return {
+    port,
+    token: state.token,
+    version: (probe.running ? probe.version : undefined) ?? version,
+    protocolVersion,
+    dataDir,
+    hostRunning: probe.running,
+    install: installInfo(dataDir, env),
+  }
 }
 
 /**
@@ -444,40 +688,51 @@ function say(line) {
 export async function printConnection({ env, home, entry, version, cliPath }) {
   const dataDir = serveDataDir(env, home)
   const state = ensureServeState(dataDir)
-  keepLauncher(dataDir, cliPath)
-  const port = state.port ?? DEFAULT_SERVE_PORT
-  const installed = protocolVersionOf(entry)
-  const probe = await probeHost({ port, token: state.token, protocolVersion: installed })
-  const protocolVersion = (probe.running ? probe.protocolVersion : undefined) ?? installed
-  if (protocolVersion === null) {
+  keepLauncher(dataDir, cliPath, env)
+  const d = await describe({ env, dataDir, state, entry, version })
+  if (!d) {
     say(`cannot read the protocol version of the host at ${entry}`)
     return 1
   }
-  process.stdout.write(
-    `${connectionLine({
-      port,
-      token: state.token,
-      version: (probe.running ? probe.version : undefined) ?? version,
-      protocolVersion,
-      dataDir,
-      hostRunning: probe.running,
-    })}\n`,
-  )
+  process.stdout.write(`${connectionLine(d)}\n`)
   return 0
 }
 
+/** serve.pid: `{ pid, hostPid, port }` of the launcher that runs, or null */
+export function readServePid(dataDir) {
+  try {
+    const v = JSON.parse(readFileSync(join(dataDir, SERVE_PID_FILE), 'utf8'))
+    return Number.isInteger(v?.pid) && v.pid > 0 ? { pid: v.pid, hostPid: Number.isInteger(v.hostPid) ? v.hostPid : null, port: v.port ?? null } : null
+  } catch {
+    return null
+  }
+}
+
+function removeServePid(dataDir, pid) {
+  if (readServePid(dataDir)?.pid !== pid) return
+  try {
+    rmSync(join(dataDir, SERVE_PID_FILE), { force: true })
+  } catch {
+    // another serve's by now, or gone
+  }
+}
+
 /**
  * `centralu serve`. Resolves with the exit code once the host has exited.
  *
  * The host's stdout is read here and never passed on: its ready line carries the token, which is
  * printed nowhere but `--connection`. Every other line the host writes to stdout (a lock conflict,
  * a store from a newer build) it also writes to stderr, which is passed through untouched.
+ *
+ * `logFd` is serve.log of a detached launcher (`runDetachedChild`): the host's stderr goes there
+ * until it is ready, so a host that dies while starting (a native module that does not load, a store
+ * from a newer build) says why in a file `--detach` can quote. From then on host.log has every line,
+ * and serve.log stops growing.
  */
-export async function runServe({ env, home, entry, version, cliPath, port: askedPort, platform = process.platform, execPath = process.execPath }) {
+export async function runServe({ env, home, entry, version, cliPath, port: askedPort, platform = process.platform, execPath = process.execPath, logFd = null }) {
   const dataDir = serveDataDir(env, home)
   const state = ensureServeState(dataDir, platform)
-  keepLauncher(dataDir, cliPath)
-  const port = askedPort ?? state.port ?? DEFAULT_SERVE_PORT
+  keepLauncher(dataDir, cliPath, env)
   const protocolVersion = protocolVersionOf(entry)
 
   // A serve that is already up answers on the port it recorded; say so before starting a host that
@@ -490,15 +745,45 @@ export async function runServe({ env, home, entry, version, cliPath, port: asked
     return 1
   }
 
+  const onWindows = await windowsListening(env, platform)
+  const chosen = await chooseServePort({ asked: askedPort, recorded: state.port, taken: async (p) => onWindows.has(p) || (await listensHere(p)) })
+  if (chosen === null) {
+    say(`another program listens on 127.0.0.1:${recorded} and on the ${SERVE_PORT_TRIES - 1} ports above it. Give a free one with --port <n>.`)
+    return 1
+  }
+  const port = chosen.port
+  if (chosen.from !== null) {
+    const who = onWindows.has(chosen.from) ? 'Windows' : 'another program'
+    say(`127.0.0.1:${chosen.from} is in use by ${who}; starting on ${port} instead (kept for the next start)`)
+  }
+
   const plan = hostCommand(execPath, entry, { port, dataDir, platform })
   const child = spawn(plan.command, plan.args, {
-    stdio: ['pipe', 'pipe', 'inherit'],
+    stdio: ['pipe', 'pipe', logFd === null ? 'inherit' : 'pipe'],
     env: hostEnv(env, { token: state.token, dataDir, version, hostDir: dirname(entry) }),
     detached: plan.detached,
     windowsHide: true,
   })
 
   let ready = false
+  if (logFd !== null) {
+    // Read to the end in any case: a pipe nobody drains stops the host once its buffer is full
+    child.stderr.on('data', (b) => {
+      if (ready) return
+      try {
+        writeSync(logFd, b)
+      } catch {
+        // the log is a convenience
+      }
+    })
+  }
+  if (child.pid) {
+    try {
+      writeFileSync(join(dataDir, SERVE_PID_FILE), `${JSON.stringify({ pid: process.pid, hostPid: child.pid, port })}\n`)
+    } catch (e) {
+      say(`could not write ${SERVE_PID_FILE}: ${e.message}`)
+    }
+  }
   createInterface({ input: child.stdout }).on('line', (line) => {
     const p = readyPort(line)
     if (p === null || ready) return
@@ -509,7 +794,8 @@ export async function runServe({ env, home, entry, version, cliPath, port: asked
       say(`could not record the port in ${join(dataDir, SERVE_STATE_FILE)}: ${e.message}`)
     }
     say(`listening on 127.0.0.1:${p} (Centralu ${version}, protocol ${protocolVersion ?? 'unknown'}, data ${dataDir})`)
-    say(`from your computer: ssh -N -L ${p}:127.0.0.1:${p} <this machine>, or let the app connect over SSH`)
+    if (logFd === null) say(`from your computer: ssh -N -L ${p}:127.0.0.1:${p} <this machine>, or let the app connect over SSH`)
+    else say(`the host's own log continues in ${join(dataDir, 'host.log')}`)
   })
 
   /*
@@ -532,11 +818,13 @@ export async function runServe({ env, home, entry, version, cliPath, port: asked
 
   return await new Promise((resolve) => {
     child.on('error', (e) => {
+      removeServePid(dataDir, process.pid)
       say(`could not start the host (${plan.command}): ${e.message}`)
       resolve(1)
     })
     child.on('exit', (code, signal) => {
       for (const s of signals) process.off(s, forward)
+      removeServePid(dataDir, process.pid)
       if (!ready && code !== 0) {
         const pid = lockHolderPid(dataDir)
         if (pid && pid !== child.pid && isAlive(pid)) {
@@ -546,6 +834,7 @@ export async function runServe({ env, home, entry, version, cliPath, port: asked
           )
         }
       }
+      if (logFd !== null) say(`the host exited (${signal ?? code})`)
       resolve(code ?? (signal ? 1 : 0))
     })
   })
@@ -559,3 +848,473 @@ export function rotateToken({ env, home }) {
   say('A serve that is running keeps the old one until it restarts; clients read the new one with --connection.')
   return 0
 }
+
+// ── --detach ────────────────────────────────────────────────────────────────────────────────
+
+/**
+ * What a detached launcher needs from the `--detach` that started it. A process created through WMI
+ * does not inherit this environment: it gets the account's default one. Everything else it needs
+ * (the token, the port) is in the data folder. Never a secret.
+ */
+export const CARRIED_ENV = ['CC_DATA_DIR', 'CENTRALU_HOST_ENTRY', 'CENTRALU_MANAGED']
+
+export function encodeChildSpec({ env, log }) {
+  const carried = {}
+  for (const k of CARRIED_ENV) if (typeof env[k] === 'string' && env[k]) carried[k] = env[k]
+  return Buffer.from(JSON.stringify({ log, env: carried }), 'utf8').toString('base64url')
+}
+
+/** `{ log, env }` from `--detached-child=`, only the carried variables; null when it is not one */
+export function decodeChildSpec(raw) {
+  try {
+    const v = JSON.parse(Buffer.from(String(raw), 'base64url').toString('utf8'))
+    if (typeof v?.log !== 'string' || !v.log) return null
+    const env = {}
+    for (const k of CARRIED_ENV) if (typeof v.env?.[k] === 'string') env[k] = v.env[k]
+    return { log: v.log, env }
+  } catch {
+    return null
+  }
+}
+
+/**
+ * How `--detach` starts the launcher so that it outlives the session that asked (measured,
+ * docs/plans/remote-hub.md §10.4):
+ *
+ *   setsid  posix: a new session, no terminal, nothing from the ssh session's process group. Once
+ *           `--detach` returns, the launcher is reparented to init (the second fork of the usual
+ *           double fork); it opens no terminal, so it can never get a controlling one back
+ *   wmi     Windows: `Win32_Process.Create`. Windows' OpenSSH ends every process of a session with
+ *           it, `Start-Process` included; a process WMI creates is not part of the session
+ *   wsl     inside WSL: the same WMI call from the Windows side, creating `wsl.exe -d <distro>
+ *           --exec <launcher>`. A `setsid` launcher inside the distro survives the session but not the
+ *           distro, which WSL stops about 15 s after its last `wsl.exe` client exits; this
+ *           `wsl.exe` is a client for as long as the host runs, and ending it ends the host
+ */
+export function detachHow(platform, env, interop = wslInterop) {
+  if (platform === 'win32') return 'wmi'
+  if (platform === 'linux' && env.WSL_DISTRO_NAME && interop()) return 'wsl'
+  return 'setsid'
+}
+
+function wslInterop() {
+  return existsSync('/proc/sys/fs/binfmt_misc/WSLInterop') || existsSync('/proc/sys/fs/binfmt_misc/WSLInterop-late')
+}
+
+/**
+ * One argument for a Windows command line, as `CommandLineToArgvW` and the C runtime read it: always
+ * quoted, backslashes before the closing quote doubled. A double quote or a line break cannot be a
+ * path on Windows and is refused rather than escaped.
+ */
+export function windowsArg(arg) {
+  const s = String(arg)
+  if (/["\r\n\0]/.test(s)) throw new Error(`cannot pass ${JSON.stringify(s)} on a Windows command line`)
+  return `"${s.replace(/(\\+)$/, '$1$1')}"`
+}
+
+/** The PowerShell that asks WMI to create `commandLine`. Values cross as base64, so no quoting can break */
+export function wmiScript({ commandLine, cwd = null }) {
+  const text = (s) => `[Text.Encoding]::UTF8.GetString([Convert]::FromBase64String('${Buffer.from(s, 'utf8').toString('base64')}'))`
+  return [
+    "$ErrorActionPreference = 'Stop'",
+    '[Console]::OutputEncoding = [Text.Encoding]::UTF8',
+    `$a = @{ CommandLine = ${text(commandLine)} }`,
+    ...(cwd ? [`$a.CurrentDirectory = ${text(cwd)}`] : []),
+    // No console window when this runs from a desktop session; left out where it cannot be built
+    'try { $a.ProcessStartupInformation = New-CimInstance -ClassName Win32_ProcessStartup -ClientOnly -Property @{ ShowWindow = [uint16]0 } } catch { }',
+    'try {',
+    '  $r = Invoke-CimMethod -ClassName Win32_Process -MethodName Create -Arguments $a',
+    "  if ($r.ReturnValue -eq 21 -and $a.ContainsKey('ProcessStartupInformation')) { $a.Remove('ProcessStartupInformation'); $r = Invoke-CimMethod -ClassName Win32_Process -MethodName Create -Arguments $a }",
+    '  "CENTRALU-WMI $($r.ReturnValue) $($r.ProcessId)"',
+    '} catch { "CENTRALU-WMI-ERROR $($_.Exception.Message)" }',
+  ].join('\n')
+}
+
+/** `Win32_Process.Create`'s return values (Microsoft's documentation of the method) */
+const WMI_CODES = { 2: 'access denied', 3: 'insufficient privilege', 8: 'unknown failure', 9: 'path not found', 21: 'invalid parameter' }
+
+/** What the WMI script said: `{ ok: true, pid }` or `{ ok: false, message }` */
+export function parseWmiAnswer(stdout) {
+  const text = String(stdout ?? '')
+  const m = /CENTRALU-WMI (\d+) (\d*)/.exec(text)
+  if (m) {
+    const rv = Number(m[1])
+    if (rv === 0 && Number(m[2]) > 0) return { ok: true, pid: Number(m[2]) }
+    return { ok: false, message: `WMI's Win32_Process.Create returned ${rv}${WMI_CODES[rv] ? ` (${WMI_CODES[rv]})` : ''}` }
+  }
+  const e = /CENTRALU-WMI-ERROR (.*)/.exec(text)
+  if (e) return { ok: false, message: `WMI refused to create the process: ${e[1].trim()}` }
+  return { ok: false, message: 'PowerShell did not answer the WMI request' }
+}
+
+/**
+ * The sentence for a start WMI did not make. Some managed machines block process creation through
+ * WMI (Microsoft Defender's attack surface reduction rule "Block process creations originating from
+ * PSExec and WMI commands", off by default); the hub then runs the host bound to its link
+ * (docs/plans/remote-hub.md §10.9, decision 3).
+ */
+export function wmiBlockedMessage(detail) {
+  return (
+    `Windows did not start Centralu through WMI (${detail}). Some managed Windows machines block it ` +
+    '(Microsoft Defender\'s attack surface reduction rule "Block process creations originating from PSExec and WMI commands"). ' +
+    'Centralu can run there while this computer is linked instead, and stops when the link does.'
+  )
+}
+
+/** Windows PowerShell 5.1, by full path: under %SystemRoot% on Windows, through the drive WSL mounts inside a distro */
+export function windowsPowershell(how, env, mounts) {
+  return windowsTool(how, env, ['System32', 'WindowsPowerShell', 'v1.0', 'powershell.exe'], mounts)
+}
+
+/** A program under the Windows folder, by full path, from Windows or from inside a WSL distro; null when unreachable */
+function windowsTool(how, env, rel, mounts = () => readFileSync('/proc/mounts', 'utf8')) {
+  if (how === 'wmi') return [env.SystemRoot || env.SYSTEMROOT || 'C:\\Windows', ...rel].join('\\')
+  let text = ''
+  try {
+    text = mounts()
+  } catch {
+    // no /proc: not WSL after all
+  }
+  for (const line of text.split('\n')) {
+    const [, at, type] = line.split(' ')
+    if (!at || !(type === '9p' || type === 'drvfs')) continue
+    const candidate = [at, 'Windows', ...rel].join('/')
+    if (existsSync(candidate)) return candidate
+  }
+  return null
+}
+
+function runCaptured(command, args, env) {
+  return new Promise((resolve) => {
+    let stdout = ''
+    let stderr = ''
+    let child
+    try {
+      child = spawn(command, args, { env, stdio: ['ignore', 'pipe', 'pipe'], windowsHide: true })
+    } catch (e) {
+      return resolve({ code: null, stdout, stderr: e.message })
+    }
+    child.stdout.on('data', (b) => (stdout += String(b)))
+    child.stderr.on('data', (b) => (stderr += String(b)))
+    child.once('error', (e) => resolve({ code: null, stdout, stderr: e.message }))
+    child.once('close', (code) => resolve({ code, stdout, stderr }))
+  })
+}
+
+/**
+ * Starts the launcher outside this session (`detachHow`). `{ ok: true, pid, exited?, alive? }` once
+ * something was created, `{ ok: false, blocked, message }` when nothing was: `blocked` is the WMI
+ * refusal the hub falls back from.
+ */
+export async function startDetached({ how, execPath, args, env, cwd }) {
+  if (how === 'setsid') {
+    return await new Promise((resolve) => {
+      let child
+      try {
+        child = spawn(execPath, args, { detached: true, stdio: 'ignore', cwd, env, windowsHide: true })
+      } catch (e) {
+        return resolve({ ok: false, blocked: false, message: `could not start ${execPath}: ${e.message}` })
+      }
+      child.once('error', (e) => resolve({ ok: false, blocked: false, message: `could not start ${execPath}: ${e.message}` }))
+      child.once('spawn', () => {
+        const exited = new Promise((r) => child.once('exit', (code, signal) => r(signal ?? code)))
+        child.unref()
+        resolve({ ok: true, pid: child.pid, exited })
+      })
+    })
+  }
+  const ps = windowsPowershell(how, env)
+  if (!ps) return { ok: false, blocked: true, message: wmiBlockedMessage('Windows PowerShell is not reachable from this WSL distro') }
+  let commandLine
+  try {
+    if (how === 'wmi') commandLine = [execPath, ...args].map(windowsArg).join(' ')
+    else {
+      const distro = env.WSL_DISTRO_NAME ?? ''
+      const user = userInfo().username
+      if (!/^[A-Za-z0-9._-]{1,64}$/.test(distro)) throw new Error(`not a WSL distro name: ${distro}`)
+      if (!/^[A-Za-z_][A-Za-z0-9_.-]{0,31}$/.test(user)) throw new Error(`cannot name the user ${user} to wsl.exe`)
+      commandLine = ['wsl.exe', '-d', distro, '-u', user, '--cd', cwd, '--exec', execPath, ...args].map(windowsArg).join(' ')
+    }
+  } catch (e) {
+    return { ok: false, blocked: false, message: e.message }
+  }
+  const script = wmiScript({ commandLine, cwd: how === 'wmi' ? cwd : null })
+  const r = await runCaptured(ps, ['-NoProfile', '-NonInteractive', '-EncodedCommand', Buffer.from(script, 'utf16le').toString('base64')], env)
+  const answer = parseWmiAnswer(r.stdout)
+  if (!answer.ok) return { ok: false, blocked: true, message: wmiBlockedMessage(r.stdout.trim() || !r.stderr.trim() ? answer.message : r.stderr.trim().split(/\r?\n/).at(-1)) }
+  // A Windows pid can be watched from Windows; from inside WSL it names nothing here
+  /*
+   * Measured on a Windows 11 laptop (2026-10-08): a `wsl.exe` that WMI creates can hang without
+   * running anything in the distro (it starts a second wsl.exe and waits), while `wsl.exe --list`
+   * from the same context answers. `--detach` reports that as a start that never ran, and ends what
+   * it created rather than leave it waiting: this pid and its children, nothing else.
+   */
+  const taskkill = windowsTool(how, env, ['System32', 'taskkill.exe'])
+  const stop = taskkill ? () => runCaptured(taskkill, ['/PID', String(answer.pid), '/T', '/F'], env) : null
+  return { ok: true, pid: answer.pid, alive: how === 'wmi' ? () => isAlive(answer.pid) : null, stop }
+}
+
+/** The last lines of a file, for a reason; empty when it cannot be read */
+function tail(file, n = 15) {
+  try {
+    return readFileSync(file, 'utf8').trimEnd().split(/\r?\n/).slice(-n).join('\n')
+  } catch {
+    return ''
+  }
+}
+
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms))
+
+/**
+ * `centralu serve --detach` (docs/plans/remote-hub.md §10.4, S9): starts `serve` outside this
+ * session (`detachHow`) and returns once its host answers a hello, so the caller, typically the
+ * hub over ssh, can connect at once.
+ *
+ * stdout carries one JSON line, read by the hub (`links/tunnel.ts`): the connection line with
+ * `hostRunning: true` and `detach: { ok: true, how, already }`, or `{ v: 1, detach: { ok: false,
+ * how, reason, message } }`. `reason` is `wmi_blocked` when WMI created nothing or created something
+ * that never ran, which the hub answers with a link-bound start; `exited` and `timeout` quote the
+ * end of serve.log. Exit codes do not survive Windows' ssh (agent-host.md §4.8), so the line is
+ * what counts; the exit code is 0 or 1 for a person at a prompt.
+ *
+ * `start` and the two waits are parameters so tests can stand in for WMI.
+ */
+export async function runDetach({
+  env,
+  home,
+  entry,
+  version,
+  cliPath,
+  port = null,
+  platform = process.platform,
+  execPath = process.execPath,
+  cwd = process.cwd(),
+  start = startDetached,
+  how = detachHow(platform, env),
+  timeoutMs = 30_000,
+  noStartMs = 10_000,
+  write = (line) => process.stdout.write(line),
+}) {
+  const dataDir = serveDataDir(env, home)
+  // Created here, so the token exists before the launcher reads it and `answered` reads it again
+  ensureServeState(dataDir, platform)
+  keepLauncher(dataDir, cliPath, env)
+  const fail = (reason, message) => {
+    say(message)
+    write(`${JSON.stringify({ v: 1, detach: { ok: false, how, reason, message } })}\n`)
+    return 1
+  }
+  const answered = async (already) => {
+    const d = await describe({ env, dataDir, state: ensureServeState(dataDir, platform), entry, version })
+    if (!d?.hostRunning) return false
+    write(`${connectionLine({ ...d, detach: { ok: true, how, already } })}\n`)
+    return true
+  }
+  if (await answered(true)) return 0
+
+  const log = join(dataDir, SERVE_LOG_FILE)
+  // One start per log: the last one's is kept beside it, and its absence is how a start that never ran shows
+  try {
+    renameSync(log, `${log}.1`)
+  } catch {
+    // there was none
+  }
+  const args = [cliPath, 'serve', ...(port !== null ? ['--port', String(port)] : []), `${CHILD_FLAG}${encodeChildSpec({ env, log })}`]
+  let started
+  try {
+    started = await start({ how, execPath, args, env, cwd })
+  } catch (e) {
+    return fail('failed', `could not start centralu serve in the background: ${e.message}`)
+  }
+  if (!started.ok) return fail(started.blocked ? 'wmi_blocked' : 'failed', started.message)
+  say(`started centralu serve in the background (${how}, pid ${started.pid ?? 'unknown'}); waiting for its host`)
+
+  let exited
+  started.exited?.then((c) => (exited = c ?? 'unknown'))
+  const t0 = Date.now()
+  while (Date.now() - t0 < timeoutMs) {
+    if (await answered(false)) return 0
+    const why = tail(log)
+    if (exited !== undefined) return fail('exited', `centralu serve exited (${exited}) before its host answered${why ? `:\n${why}` : ''}`)
+    const gone = started.alive?.() === false
+    if (how !== 'setsid' && !existsSync(log) && (gone || Date.now() - t0 > noStartMs)) {
+      // WMI said yes, and nothing ran: the process was stopped before Node started, or hangs (`startDetached`)
+      if (!gone) await started.stop?.()
+      return fail('wmi_blocked', wmiBlockedMessage(gone ? `process ${started.pid} ended before it started` : `nothing started within ${Math.round(noStartMs / 1000)} s`))
+    }
+    if (gone) return fail('exited', `centralu serve exited before its host answered${why ? `:\n${why}` : ''}`)
+    await sleep(250)
+  }
+  const why = tail(log)
+  return fail('timeout', `the host did not answer within ${Math.round(timeoutMs / 1000)} s${why ? `:\n${why}` : ''}`)
+}
+
+/**
+ * The launcher `--detach` started: `runServe` writing to serve.log. The carried variables are
+ * already in `env` (centralu.mjs merges them before it looks for the host).
+ */
+export async function runDetachedChild({ spec, ...opts }) {
+  let fd = null
+  try {
+    fd = openSync(spec.log, 'a')
+    sink = (text) => {
+      try {
+        writeSync(fd, text)
+      } catch {
+        // nowhere left to say it
+      }
+    }
+  } catch {
+    // no log: say() keeps stderr, which nobody reads; --detach then reports the start as never run
+  }
+  say(`started in the background (pid ${process.pid}, Node ${process.version})`)
+  try {
+    return await runServe({ ...opts, logFd: fd })
+  } catch (e) {
+    say(e?.message ?? String(e))
+    return 1
+  }
+}
+
+// ── --stop ──────────────────────────────────────────────────────────────────────────────────
+
+/**
+ * Asks a running host to stop (`host.stop`, an additive RPC): a hello with the token, then the call.
+ * `{ ok: true }` once the host took it, `{ ok: false, message, protocolVersion? }` otherwise; the
+ * last names the host's protocol when the hello was refused for it, so the caller can ask again.
+ */
+export function askHostToStop({ port, token, protocolVersion, timeoutMs = 5000, WebSocketImpl = globalThis.WebSocket }) {
+  if (typeof WebSocketImpl !== 'function') return Promise.resolve({ ok: false, message: `Node ${process.version} has no WebSocket to ask with` })
+  return new Promise((resolve) => {
+    let ws = null
+    let settled = false
+    const done = (result) => {
+      if (settled) return
+      settled = true
+      clearTimeout(timer)
+      try {
+        ws?.close()
+      } catch {
+        // already closed
+      }
+      resolve(result)
+    }
+    const timer = setTimeout(() => done({ ok: false, message: 'no answer in time' }), timeoutMs)
+    try {
+      ws = new WebSocketImpl(`ws://127.0.0.1:${port}`)
+    } catch (e) {
+      done({ ok: false, message: e.message })
+      return
+    }
+    ws.onopen = () => ws.send(JSON.stringify({ kind: 'hello', token, protocolVersion: protocolVersion ?? 0 }))
+    ws.onmessage = (ev) => {
+      let m
+      try {
+        m = JSON.parse(String(ev.data))
+      } catch {
+        return
+      }
+      if (m?.kind === 'hello_ok') ws.send(JSON.stringify({ kind: 'rpc', id: 'stop', method: 'host.stop', params: {} }))
+      else if (m?.kind === 'res' && m.id === 'stop') done(m.ok ? { ok: true } : { ok: false, message: m.error?.message ?? 'refused' })
+      else if (m?.kind === 'res' && m.ok === false) {
+        const pv = m.error?.code === 'version_mismatch' ? (m.error?.data?.protocolVersion ?? mismatchServerVersion(m.error.message)) : null
+        done({ ok: false, message: m.error?.message ?? 'refused', ...(Number.isInteger(pv) ? { protocolVersion: pv } : {}) })
+      }
+    }
+    ws.onerror = () => done({ ok: false, message: 'could not connect' })
+    ws.onclose = () => done({ ok: false, message: 'the connection closed' })
+  })
+}
+
+async function waitUntil(pred, ms) {
+  const t0 = Date.now()
+  while (Date.now() - t0 < ms) {
+    if (await pred()) return true
+    await sleep(200)
+  }
+  return pred()
+}
+
+/**
+ * `centralu serve --stop` (docs/plans/remote-hub.md §10.4, S10): asks the running host to stop over
+ * its own socket and waits until it has, so its shutdown runs (its agents, terminals and app
+ * processes stopped in order, the store closed). Measured on Windows: killing the WMI-started
+ * launcher ended the host 143 ms later with no "shutting down" line, because the host sits in
+ * libuv's kill-on-close job object; so on Windows killing is the last resort, not the way.
+ *
+ * Then, only for a host that did not go: posix SIGTERM to the launcher (which passes it to the host,
+ * the same clean stop), and on Windows ending the host. A host from before `host.stop` gets that
+ * too. The launcher ends with its host; one still there afterwards is ended.
+ *
+ * Only the processes serve.pid and host.lock name, and only while they agree with each other and
+ * the host answered with this folder's token: a stale pid is never signalled.
+ *
+ * stdout: one line, `{ v: 1, stop: { ok, wasRunning, how?, message? } }`.
+ */
+export async function runStop({
+  env,
+  home,
+  entry,
+  platform = process.platform,
+  timeoutMs = 30_000,
+  graceMs = 15_000,
+  kill = (pid, sig) => process.kill(pid, sig),
+  write = (line) => process.stdout.write(line),
+}) {
+  const dataDir = serveDataDir(env, home)
+  const report = (stop) => {
+    if (stop.message) say(stop.message)
+    write(`${JSON.stringify({ v: 1, stop })}\n`)
+    return stop.ok ? 0 : 1
+  }
+  const state = readServeState(dataDir)
+  if (!state) return report({ ok: true, wasRunning: false, message: `no centralu serve has run for ${dataDir}` })
+  const port = state.port ?? DEFAULT_SERVE_PORT
+  const installed = protocolVersionOf(entry) ?? 0
+  const probe = () => probeHost({ port, token: state.token, protocolVersion: installed })
+  const first = await probe()
+  if (!first.running) return report({ ok: true, wasRunning: false, message: `centralu serve is not running for ${dataDir}` })
+
+  const hostPid = lockHolderPid(dataDir)
+  const launcher = readServePid(dataDir)
+  // serve.pid is only believed when it names the host that holds the folder now
+  const launcherPid = launcher && hostPid && launcher.hostPid === hostPid ? launcher.pid : null
+  const down = async () => !(await probe()).running && !(hostPid && isAlive(hostPid))
+
+  let asked = await askHostToStop({ port, token: state.token, protocolVersion: first.protocolVersion ?? installed })
+  if (!asked.ok && asked.protocolVersion !== undefined) asked = await askHostToStop({ port, token: state.token, protocolVersion: asked.protocolVersion })
+  let how = 'asked'
+  let stopped = asked.ok && (await waitUntil(down, timeoutMs))
+  let note = asked.ok ? null : `the host did not take the request to stop (${asked.message})`
+  if (!stopped) {
+    const target = platform !== 'win32' && launcherPid ? launcherPid : hostPid
+    if (target) {
+      how = platform === 'win32' ? 'ended' : 'signal'
+      try {
+        kill(target, 'SIGTERM')
+      } catch {
+        // gone in the meantime
+      }
+      stopped = await waitUntil(down, platform === 'win32' ? 5_000 : graceMs)
+      if (!stopped && hostPid && isAlive(hostPid)) {
+        how = 'ended'
+        try {
+          kill(hostPid, 'SIGKILL')
+        } catch {
+          // gone in the meantime
+        }
+        stopped = await waitUntil(down, 5_000)
+      }
+    }
+    if (how === 'ended') note = `${note ? `${note}; ` : ''}ended the host without its shutdown`
+  }
+  if (launcherPid && launcherPid !== process.pid && !(await waitUntil(() => !isAlive(launcherPid), 5_000))) {
+    try {
+      kill(launcherPid, 'SIGKILL')
+    } catch {
+      // gone in the meantime
+    }
+  }
+  if (!stopped) return report({ ok: false, wasRunning: true, how, message: `the host on 127.0.0.1:${port} is still running${note ? ` (${note})` : ''}` })
+  return report({ ok: true, wasRunning: true, how, ...(note ? { message: note } : {}) })
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "centralu",
-  "version": "0.1.0-beta.12",
+  "version": "0.1.0-beta.14",
   "description": "Run, watch and steer several Claude Code and Codex CLI sessions from one window",
   "license": "MIT",
   "author": "ijun17",
@@ -37,9 +37,9 @@
     "win32"
   ],
   "optionalDependencies": {
-    "@centralu/darwin-arm64": "0.1.0-beta.12",
-    "@centralu/linux-arm64": "0.1.0-beta.12",
-    "@centralu/linux-x64": "0.1.0-beta.12",
-    "@centralu/win32-x64": "0.1.0-beta.12"
+    "@centralu/darwin-arm64": "0.1.0-beta.14",
+    "@centralu/linux-arm64": "0.1.0-beta.14",
+    "@centralu/linux-x64": "0.1.0-beta.14",
+    "@centralu/win32-x64": "0.1.0-beta.14"
   }
 }