@twinsearth/w2m-dsh-plugin 0.5.4 → 0.5.6

package/src/plugin/mdns-discover.mjs

@@ -0,0 +1,577 @@
+/**
+ * LAN discovery for the W2M plugin: mDNS candidates, offered *next to* the relay roster.
+ *
+ * # THE RULE, and why it is written here in these words
+ *
+ * **Discovery is a hint, never an authorisation.** A discovered device is a *candidate to ask*. It
+ * never skips, weakens or short-circuits any existing gate — the device token, the per-machine
+ * allow-list, the git anchors, the verdict model. If you find yourself writing code where a discovered
+ * device is trusted because it was discovered, stop: that is a defect, not a shortcut.
+ *
+ * This module is where that rule has to be *executable* rather than aspirational, so it is spelled out
+ * as code in {@link dispatchRefusal}:
+ *
+ *   1. mDNS carries no proof of anything. Every candidate is reported `verified: false`, and no field
+ *      of a discovered record is ever copied into a token, a URL credential, or a config value.
+ *   2. A **service name is not identity.** A matching name is reported as `name_match`, a hint a human
+ *      can read, and can never on its own promote a candidate. Identity is a `machine_id` the *relay*
+ *      assigned at pairing time, which is the only value that can be matched against the roster.
+ *   3. A candidate the relay does not list is `dispatchable: false` with a refusal that names what
+ *      would change it. It is visible — an operator should see the machine that is on the LAN but
+ *      unpaired — and it is not selectable.
+ *   4. Even a paired, relay-listed candidate is only a *dispatch target* in the sense the relay
+ *      already allows: this module does not add authority, it removes ambiguity.
+ *
+ * What this buys, concretely: a device that shows itself on the LAN without pairing is listed as
+ * *unpaired* rather than silently appearing in the fleet, and no code path here can turn "I saw it" into
+ * "I may run there".
+ *
+ * # Fail soft, but never silently
+ *
+ * Multicast is unavailable on plenty of real networks — a locked-down corporate WLAN, a container with
+ * no multicast route, a host where every interface refuses the membership. That is not an error
+ * condition for the plugin: `createDiscoverySource` logs **one** line naming the reason (repeat
+ * failures with the same reason are not logged again) and every call still returns a structure with
+ * `available: false` and `error` set. A caller can therefore always tell "nothing was discovered"
+ * apart from "nothing could be discovered", which is the distinction a silent success would destroy.
+ */
+
+import {
+  DEFAULT_DISCOVERY_TIMEOUT_MS,
+  DEFAULT_SERVICE_TYPE,
+  createResponder,
+  discoverServices,
+} from '../agent/mdns.mjs';
+import { isUsableUnicastIpv4 } from '../agent/mdns-wire.mjs';
+// The liveness rule is the relay control's own (`fleet-target.mjs`), imported rather than re-derived:
+// "is this machine online" must mean one thing across the plugin, and a second copy of the rule here
+// would be a second chance to disagree about who is reachable.
+import { isSelectableDevice } from './fleet-target.mjs';
+
+/** The `source` value every entry produced by this module carries. */
+export const DISCOVERY_SOURCE = 'mdns';
+
+/** TXT key carrying the relay's machine id. The one field that can be matched against the roster. */
+export const TXT_MACHINE_ID = 'mid';
+
+/** TXT key carrying the machine's display name. */
+export const TXT_MACHINE_NAME = 'name';
+
+/**
+ * The sentence every discovered candidate carries, so a reader of the JSON cannot miss the posture.
+ *
+ * It is a constant rather than a literal per site because a model reading `w2m_devices` output sees the
+ * same rule whether it looked at the block, the candidate, or only the counts.
+ */
+export const DISCOVERY_NOTE =
+  'discovery is a hint, never an authorisation: a discovered device is unverified (mDNS authenticates ' +
+  'nothing) and is only offered as a dispatch target when the relay already lists it as a live machine - ' +
+  'pairing, the device token, the allow-list and the anchors all still apply';
+
+/** Why a candidate may not be dispatched to. Each message says what would change it. */
+export const DISPATCH_REFUSAL = Object.freeze({
+  NOT_DISCOVERED: 'MDNS_NOT_DISCOVERED',
+  IDENTITY_UNKNOWN: 'MDNS_IDENTITY_UNKNOWN',
+  UNPAIRED: 'MDNS_UNPAIRED',
+  RELAY_STALE: 'MDNS_RELAY_STALE',
+  LOCAL_INSTANCE: 'MDNS_LOCAL_INSTANCE',
+  NO_ENDPOINT: 'MDNS_NO_ENDPOINT',
+});
+
+/** Lower-case a name for comparison. Names arrive from a human, so case and padding are noise. */
+function fold(value) {
+  return typeof value === 'string' ? value.trim().toLowerCase() : '';
+}
+
+/** A trimmed non-empty string, or null. */
+function text(value) {
+  return typeof value === 'string' && value.trim() !== '' ? value.trim() : null;
+}
+
+/**
+ * Index the relay roster for the two lookups this module needs.
+ *
+ * `byId` is the only index that can establish identity: a machine id the relay minted at pairing time.
+ * `byName` exists so a *hint* can be shown — "this looks like the machine the relay calls
+ * `mac-studio`" — and is deliberately a separate structure so no future edit can accidentally use it
+ * where identity is required.
+ *
+ * @param {object[]} relayRecords `GET /v1/devices`'s `devices`, verbatim.
+ * @returns {{byId: Map<string, object>, byName: Map<string, object[]>, size: number}}
+ */
+export function buildRelayIndex(relayRecords) {
+  const byId = new Map();
+  const byName = new Map();
+  let size = 0;
+  for (const record of Array.isArray(relayRecords) ? relayRecords : []) {
+    if (record === null || typeof record !== 'object') continue;
+    size += 1;
+    const machineId = text(record.machine_id);
+    const name = text(record.machine_name) ?? text(record.name);
+    const entry = {
+      machine_id: machineId,
+      name,
+      online: isSelectableDevice(record),
+      state: text(record.state) ?? (isSelectableDevice(record) ? 'online' : 'unknown'),
+    };
+    if (machineId !== null && !byId.has(machineId)) byId.set(machineId, entry);
+    const key = fold(name);
+    if (key !== '') {
+      if (!byName.has(key)) byName.set(key, []);
+      byName.get(key).push(entry);
+    }
+  }
+  return { byId, byName, size };
+}
+
+/**
+ * Why a discovered candidate may not be dispatched to, or `null` when it may.
+ *
+ * This function **is** the rule. Read it as the list of things a discovered device would have to fake to
+ * become a dispatch target: an identity the relay assigned, a live listing in the relay's own roster,
+ * and an endpoint. A name it chose for itself is in none of those.
+ *
+ * @param {object|null|undefined} candidate A candidate from {@link describeDiscoveredCandidate}.
+ * @returns {string|null}
+ */
+export function dispatchRefusal(candidate) {
+  if (candidate === null || typeof candidate !== 'object') {
+    return `${DISPATCH_REFUSAL.NOT_DISCOVERED}: not a discovered candidate at all`;
+  }
+  if (candidate.source !== DISCOVERY_SOURCE) {
+    return `${DISPATCH_REFUSAL.NOT_DISCOVERED}: this is not a LAN-discovered candidate, so this rule does not speak for it`;
+  }
+  if (text(candidate.machine_id) === null) {
+    return (
+      `${DISPATCH_REFUSAL.IDENTITY_UNKNOWN}: the advertisement carries no \`${TXT_MACHINE_ID}\` key, so ` +
+      'there is no relay machine id to look up; a service name is not identity and is never used as one. ' +
+      'Pair the machine (its agent then advertises the machine id the relay gave it)'
+    );
+  }
+  if (candidate.known_to_relay !== true) {
+    return (
+      `${DISPATCH_REFUSAL.UNPAIRED}: the relay does not list machine \`${candidate.machine_id}\`, so nothing ` +
+      'has authorised this machine; a discovered device is a candidate to ask, never an authorisation. ' +
+      'Pair it, or start its agent so the relay holds a lease, and it becomes an ordinary fleet member'
+    );
+  }
+  if (candidate.relay_online !== true) {
+    return (
+      `${DISPATCH_REFUSAL.RELAY_STALE}: the relay lists machine \`${candidate.machine_id}\` but its lease is not ` +
+      `live (state: ${candidate.relay_state ?? 'unknown'}), so a dispatch narrowed to it would have no target; ` +
+      'wait for the machine to come online'
+    );
+  }
+  if (candidate.local === true) {
+    return `${DISPATCH_REFUSAL.LOCAL_INSTANCE}: this is this machine's own advertisement, not a peer`;
+  }
+  if (!Number.isInteger(candidate.port) || candidate.port < 1 || candidate.port > 65535) {
+    return `${DISPATCH_REFUSAL.NO_ENDPOINT}: the advertisement named no usable port (SRV port ${JSON.stringify(candidate.port)})`;
+  }
+  if (text(candidate.address) === null) {
+    return `${DISPATCH_REFUSAL.NO_ENDPOINT}: the advertisement named no usable address (A record or observed source)`;
+  }
+  return null;
+}
+
+/**
+ * May this candidate be offered as a dispatch target?
+ *
+ * @param {object|null|undefined} candidate
+ * @returns {boolean}
+ */
+export function mayDispatchTo(candidate) {
+  return dispatchRefusal(candidate) === null;
+}
+
+/**
+ * The candidates a dispatch surface may offer.
+ *
+ * A caller that builds a target list from anything else has bypassed the rule; this exists so the
+ * correct list is the easy one to build.
+ *
+ * @param {object[]} candidates
+ * @returns {object[]}
+ */
+export function dispatchableCandidates(candidates) {
+  return (Array.isArray(candidates) ? candidates : []).filter((candidate) => mayDispatchTo(candidate));
+}
+
+/**
+ * Turn one reassembled mDNS instance into a candidate with its provenance and its verdict.
+ *
+ * @param {object} instance One entry of {@link discoverServices}'s `instances`.
+ * @param {object} context
+ * @param {{byId: Map<string, object>, byName: Map<string, object[]>}} context.relayIndex
+ * @param {string|null} [context.localMachineId] This machine's relay id, for the "this is me" mark.
+ * @param {number|null} [context.observedAt]
+ * @returns {object}
+ */
+export function describeDiscoveredCandidate(instance, context) {
+  const relayIndex = context?.relayIndex ?? { byId: new Map(), byName: new Map() };
+  const localMachineId = text(context?.localMachineId);
+  const txt = instance?.txt?.map ?? {};
+  const machineId = text(txt[TXT_MACHINE_ID]);
+  const relay = machineId !== null ? (relayIndex.byId.get(machineId) ?? null) : null;
+  const nameKey = fold(instance?.name);
+  const nameMatches = nameKey === '' ? [] : (relayIndex.byName.get(nameKey) ?? []);
+
+  const advertised = Array.isArray(instance?.addresses)
+    ? instance.addresses.filter((value) => typeof value === 'string')
+    : [];
+  const usableAdvertised = advertised.find((value) => isUsableUnicastIpv4(value)) ?? null;
+  const observed = text(instance?.observed_address);
+  const usableObserved = observed !== null && isUsableUnicastIpv4(observed) ? observed : null;
+  const address = usableAdvertised ?? usableObserved;
+
+  const candidate = {
+    source: DISCOVERY_SOURCE,
+    // Never true, for any candidate, ever: mDNS cannot verify anything. The field exists so a reader
+    // does not have to infer the posture from a missing key.
+    verified: false,
+    name: text(instance?.name) ?? text(instance?.instance) ?? null,
+    instance: text(instance?.instance) ?? null,
+    service_type: text(instance?.service_type) ?? null,
+    host: text(instance?.host) ?? null,
+    address,
+    address_source: usableAdvertised !== null ? 'a' : usableObserved !== null ? 'observed' : null,
+    addresses: advertised,
+    observed_address: observed,
+    port: Number.isInteger(instance?.port) ? instance.port : null,
+    // The relay's machine id as *claimed* by the advertisement. Used to look the machine up in the
+    // roster, never trusted on its own: the roster entry is what makes it mean anything.
+    machine_id: machineId,
+    machine_name: text(txt[TXT_MACHINE_NAME]),
+    known_to_relay: relay !== null,
+    relay_state: relay?.state ?? null,
+    relay_online: relay === null ? null : relay.online === true,
+    // A name match is a *hint for a human*, never a promotion. Reported separately so it cannot be
+    // mistaken for `known_to_relay`, which is decided by machine id alone.
+    name_match:
+      machineId === null && nameMatches.length === 1
+        ? { machine_id: nameMatches[0].machine_id, name: nameMatches[0].name, online: nameMatches[0].online }
+        : null,
+    local: machineId !== null && localMachineId !== null && machineId === localMachineId,
+    ttl: instance?.ttl ?? null,
+    superseded_srv: Number.isInteger(instance?.superseded_srv) ? instance.superseded_srv : 0,
+    seen_at: Number.isFinite(context?.observedAt) ? context.observedAt : null,
+    dispatchable: false,
+    refusal: null,
+    note: DISCOVERY_NOTE,
+  };
+  candidate.refusal = dispatchRefusal(candidate);
+  candidate.dispatchable = candidate.refusal === null;
+  return candidate;
+}
+
+/**
+ * Assemble the block `w2m_devices` (and a status surface) reports.
+ *
+ * @param {object} input
+ * @param {object[]} [input.instances] From {@link discoverServices}.
+ * @param {object[]} [input.relayRecords] `GET /v1/devices`'s `devices`, for the "does the relay list it" question.
+ * @param {string|null} [input.localMachineId]
+ * @param {boolean} [input.available] Whether a scan could actually be sent.
+ * @param {string|null} [input.error] Why not, when it could not.
+ * @param {number|null} [input.scannedAt]
+ * @param {number|null} [input.elapsedMs]
+ * @param {boolean} [input.aborted]
+ * @param {object|null} [input.packets]
+ * @param {number} [input.truncated]
+ * @returns {object}
+ */
+export function describeDiscovered(input = {}) {
+  const relayRecords = Array.isArray(input.relayRecords) ? input.relayRecords : [];
+  const relayIndex = buildRelayIndex(relayRecords);
+  const observedAt = Number.isFinite(input.scannedAt) ? input.scannedAt : null;
+  const instances = Array.isArray(input.instances) ? input.instances : [];
+  const candidates = instances.map((instance) =>
+    describeDiscoveredCandidate(instance, {
+      relayIndex,
+      localMachineId: input.localMachineId ?? null,
+      observedAt,
+    }),
+  );
+  const counts = {
+    instances: candidates.length,
+    relay_listed: candidates.filter((candidate) => candidate.known_to_relay).length,
+    // Every candidate we cannot even ask the relay about, plus every one it does not know: both are
+    // "not a dispatch target", and both are reported so the number is never quietly zero.
+    unpaired: candidates.filter((candidate) => candidate.machine_id === null || !candidate.known_to_relay)
+      .length,
+    dispatchable: candidates.filter((candidate) => candidate.dispatchable).length,
+    local: candidates.filter((candidate) => candidate.local).length,
+    verified: 0,
+  };
+  return {
+    source: DISCOVERY_SOURCE,
+    service_type: text(input.serviceType) ?? DEFAULT_SERVICE_TYPE,
+    // Three states, kept apart on purpose: true (a scan ran and queries went out), false (a scan ran
+    // and multicast was unavailable), null (nobody has asked yet). Collapsing the last two would turn
+    // "not asked" into a claim about the network.
+    available: input.available === true ? true : input.available === false ? false : null,
+    error: input.error ?? null,
+    scanned_at: observedAt,
+    elapsed_ms: Number.isFinite(input.elapsedMs) ? input.elapsedMs : null,
+    aborted: input.aborted === true,
+    truncated: Number.isInteger(input.truncated) ? input.truncated : 0,
+    packets: input.packets ?? null,
+    counts,
+    candidates,
+    relay_index_size: relayIndex.size,
+    note: DISCOVERY_NOTE,
+    ...(input.available === true
+      ? {}
+      : {
+          unavailable_note:
+            input.available === false
+              ? 'no scan was possible, so an empty candidate list here means "unknown", not "no devices on the LAN"'
+              : 'no scan has been asked for yet; ask for one and an empty candidate list then means the same thing',
+        }),
+  };
+}
+
+/**
+ * One list showing both sources, with provenance, for a UI or a status surface.
+ *
+ * Dispatchability is reported per source and is deliberately *not* uniform:
+ *
+ *   * a **relay** entry keeps the relay's own liveness verdict — nothing in this module narrows or
+ *     widens what the relay path already allows;
+ *   * a **discovered** entry is additionally required to be listed by the relay (see
+ *     {@link dispatchRefusal}), which is the only reason it can ever be `dispatchable`.
+ *
+ * @param {object} input
+ * @param {object[]} [input.relayRecords]
+ * @param {object[]} [input.candidates] Already described by {@link describeDiscovered}.
+ * @param {string|null} [input.localMachineId]
+ * @returns {{roster: object[], counts: object, note: string}}
+ */
+export function mergeRosterView(input = {}) {
+  const localMachineId = text(input.localMachineId);
+  const relayed = (Array.isArray(input.relayRecords) ? input.relayRecords : [])
+    .filter((record) => record !== null && typeof record === 'object')
+    .map((record) => {
+      const machineId = text(record.machine_id);
+      return {
+        source: 'relay',
+        verified: true,
+        machine_id: machineId,
+        name: text(record.machine_name) ?? text(record.name),
+        address: null,
+        port: null,
+        online: isSelectableDevice(record),
+        state: text(record.state) ?? null,
+        local: machineId !== null && localMachineId !== null && machineId === localMachineId,
+        relay_listed: true,
+        dispatchable: isSelectableDevice(record),
+        refusal: isSelectableDevice(record)
+          ? null
+          : `RELAY_STALE: the relay lists machine \`${machineId ?? 'unknown'}\` but its lease is not live`,
+      };
+    });
+  const discovered = (Array.isArray(input.candidates) ? input.candidates : []).map((candidate) => ({
+    source: DISCOVERY_SOURCE,
+    verified: false,
+    machine_id: candidate.machine_id ?? null,
+    name: candidate.name ?? null,
+    address: candidate.address ?? null,
+    port: candidate.port ?? null,
+    online: candidate.relay_online === true,
+    state: candidate.relay_state ?? null,
+    local: candidate.local === true,
+    relay_listed: candidate.known_to_relay === true,
+    dispatchable: candidate.dispatchable === true,
+    refusal: candidate.refusal ?? null,
+  }));
+  return {
+    roster: [...relayed, ...discovered],
+    counts: {
+      relay: relayed.length,
+      discovered: discovered.length,
+      dispatchable_relay: relayed.filter((entry) => entry.dispatchable).length,
+      dispatchable_discovered: discovered.filter((entry) => entry.dispatchable).length,
+      unpaired_discovered: discovered.filter((entry) => entry.relay_listed !== true).length,
+    },
+    note: DISCOVERY_NOTE,
+  };
+}
+
+/**
+ * The plugin's discovery source: one object owning the scan, the status, and the opt-in responder.
+ *
+ * @param {object} [options]
+ * @param {(options: object) => Promise<object>} [options.scan] Injected scanner (tests use this).
+ * @param {string} [options.serviceType]
+ * @param {number} [options.timeoutMs]
+ * @param {{warn?: Function, info?: Function}|null} [options.logger]
+ * @param {() => number} [options.now]
+ * @param {object|null} [options.advertise] Opt-in responder settings; `null` (the default) means this
+ *   machine never binds 5353 and never announces. See {@link createDiscoverySource}'s `start()`.
+ * @returns {object}
+ */
+export function createDiscoverySource(options = {}) {
+  const scanner = typeof options.scan === 'function' ? options.scan : discoverServices;
+  const serviceType = options.serviceType ?? DEFAULT_SERVICE_TYPE;
+  const timeoutMs = Number.isFinite(options.timeoutMs)
+    ? Number(options.timeoutMs)
+    : DEFAULT_DISCOVERY_TIMEOUT_MS;
+  const logger = options.logger ?? null;
+  const now = typeof options.now === 'function' ? options.now : () => Date.now();
+  const advertise = options.advertise ?? null;
+
+  const state = {
+    scans: 0,
+    last_scan_at: null,
+    last_error: null,
+    available: null,
+    instances: [],
+    logged: null,
+    advertising: false,
+  };
+  /** @type {{start: Function, close: Function, status: Function}|null} */
+  let responder = null;
+
+  /** One line per distinct failure: a locked-down network must not fill a log with the same sentence. */
+  function logOnce(line) {
+    if (state.logged === line) return;
+    state.logged = line;
+    try {
+      if (typeof logger?.warn === 'function') logger.warn(line);
+      else if (typeof logger?.info === 'function') logger.info(line);
+    } catch {
+      // A logger that throws is not allowed to break discovery.
+    }
+  }
+
+  /** The current view, without scanning: recomputed against the roster it is given. */
+  function describe(context = {}) {
+    return describeDiscovered({
+      instances: state.instances,
+      relayRecords: context.relayRecords ?? [],
+      localMachineId: context.localMachineId ?? null,
+      available: state.available === true ? true : state.available === false ? false : null,
+      error: state.last_error,
+      scannedAt: state.last_scan_at,
+      elapsedMs: context.elapsedMs ?? null,
+      aborted: context.aborted === true,
+      packets: context.packets ?? null,
+      truncated: context.truncated ?? 0,
+      serviceType,
+    });
+  }
+
+  /**
+   * Run one bounded scan and describe the result against the relay roster.
+   *
+   * Never throws and never rejects: a scanner that fails is a scan that is `available: false`, which the
+   * caller reports alongside the relay roster it already has.
+   *
+   * @param {{signal?: AbortSignal, relayRecords?: object[], localMachineId?: string|null}} [context]
+   * @returns {Promise<object>}
+   */
+  async function scan(context = {}) {
+    state.scans += 1;
+    let result = null;
+    try {
+      result = await scanner({ serviceType, timeoutMs, signal: context.signal });
+    } catch (error) {
+      result = {
+        ok: false,
+        available: false,
+        error: `MDNS_SCAN_FAILED: ${error instanceof Error ? error.message : String(error)}`,
+        instances: [],
+      };
+    }
+    const available = result?.available === true;
+    state.last_scan_at = now();
+    state.available = available;
+    state.last_error = available
+      ? null
+      : (result?.error ?? 'MDNS_UNAVAILABLE: the scan reported neither success nor a reason');
+    state.instances = Array.isArray(result?.instances) ? result.instances : [];
+    if (!available) {
+      logOnce(
+        `w2m: LAN discovery is unavailable (${state.last_error}); continuing without discovered candidates`,
+      );
+    }
+    return describe({
+      relayRecords: context.relayRecords ?? [],
+      localMachineId: context.localMachineId ?? null,
+      elapsedMs: result?.elapsed_ms ?? null,
+      aborted: result?.aborted === true,
+      packets: result?.packets ?? null,
+      truncated: result?.truncated ?? 0,
+    });
+  }
+
+  /**
+   * Start the opt-in responder.
+   *
+   * Opt-in, not automatic: binding a multicast port and announcing on the LAN is a side effect on the
+   * host's network that nobody asked for by loading a plugin, and `w2m_devices` works without it. When
+   * it is asked for, a failure is soft — one log line, `advertising: false` in {@link status}, and
+   * everything else keeps working.
+   *
+   * @returns {Promise<object>} {@link status}.
+   */
+  async function start() {
+    if (advertise === null || responder !== null) return status();
+    try {
+      responder = createResponder({
+        ...advertise,
+        serviceType,
+        logger: (line) => logOnce(line),
+      });
+      const started = await responder.start();
+      state.advertising = started?.running === true;
+      if (!state.advertising) {
+        logOnce(`w2m: mDNS advertising is off (${started?.last_error ?? 'the responder did not start'})`);
+      }
+    } catch (error) {
+      responder = null;
+      state.advertising = false;
+      logOnce(
+        `w2m: mDNS advertising failed (${error instanceof Error ? error.message : String(error)}); continuing`,
+      );
+    }
+    return status();
+  }
+
+  /** Stop the responder, sending the TTL=0 goodbye. Never throws. */
+  async function stop() {
+    if (responder === null) return status();
+    const closing = responder;
+    responder = null;
+    state.advertising = false;
+    try {
+      await closing.close();
+    } catch (error) {
+      logOnce(`w2m: mDNS shutdown failed (${error instanceof Error ? error.message : String(error)})`);
+    }
+    return status();
+  }
+
+  /**
+   * The status surface: what has been scanned, whether it worked, and whether we are advertising.
+   *
+   * @returns {object}
+   */
+  function status() {
+    return {
+      source: DISCOVERY_SOURCE,
+      service_type: serviceType,
+      timeout_ms: timeoutMs,
+      scans: state.scans,
+      available: state.available,
+      last_error: state.last_error,
+      last_scan_at: state.last_scan_at,
+      candidates: state.instances.length,
+      advertising: state.advertising,
+      advertising_requested: advertise !== null,
+      note: DISCOVERY_NOTE,
+    };
+  }
+
+  return { scan, describe, status, start, stop };
+}

package/src/plugin/tools.mjs

@@ -97,6 +97,9 @@ import {
   sessionKeyFor,
   TARGET_ROUTE_PATH,
 } from './fleet-target.mjs';
+// v0.7.0: the LAN hint layer. Imported here and used in exactly one place (`w2m_devices`); see the
+// module header for why a discovered device is a candidate to *ask* and never an authorisation.
+import { createDiscoverySource } from './mdns-discover.mjs';
 import { loadSharedConfig, describeSharedConfig } from './shared-config.mjs';
 import { DEFAULT_DAILY_TIMES, DEFAULT_TIME_ZONE } from './schedule.mjs';
 
@@ -131,7 +134,7 @@ const MAX_STAGES = 16;
  * `package.json` during `scripts/pack.mjs` rather than hand-maintained, and `W2M_PLUGIN_VERSION`
  * overrides it at runtime for tests and packagers.
  */
-const PLUGIN_VERSION = '0.5.4';
+const PLUGIN_VERSION = '0.5.6';
 
 /** How long any single HTTP call may take before it is aborted, in milliseconds. */
 const DEFAULT_REQUEST_TIMEOUT_MS = 15_000;
@@ -2872,6 +2875,12 @@ export async function apply(ctx, config = {}) {
   // host without `connection.fetch` still gets every tool, and `w2m_status` reports the absence —
   // which is the honest place for it, since a host that has no Connection at all is simply not a web
   // composition and does not need a warning on every headless run.
+  //
+  // Nothing here reads the service: `registerTargetRoutes` declares the optional dependency with
+  // `ctx.inject` and attaches from that callback (the comment there records why a plain
+  // `ctx.connection` read fails the whole entry, and why the read cannot be synchronous anyway). The
+  // one mismatch worth a log line — a Connection that carries no fetch registry — is logged where it
+  // is observed, inside that callback, so this call returns a status object that stays live.
   const targetRoutes = registerTargetRoutes(ctx, {
     cfg: () => cfg,
     store: defaultSelectionStore,
@@ -2879,9 +2888,11 @@ export async function apply(ctx, config = {}) {
     request,
     resolveToken,
   });
-  if (targetRoutes.unexpected) {
-    ctx.logger?.warn?.(`w2m: the composer target control is inactive — ${targetRoutes.reason}`);
-  }
+
+  // The LAN discovery source (v0.7.0). Constructing it binds nothing: a socket is opened only when a
+  // scan is asked for, and only `w2m_devices` asks. A network where multicast is unavailable produces
+  // one log line and an `available: false` block in the answer - never a failure of the tool.
+  const discovery = createDiscoverySource({ logger: ctx.logger });
 
   /** Text output: rendered as JSON, because every one of these tools answers with a structure. */
   const jsonOutput = {
@@ -2898,12 +2909,22 @@ export async function apply(ctx, config = {}) {
       'List the machines the Rabbit currently holds leases for: machine_id, name, platform, ' +
       'capabilities, and when each was last seen. Read-only. Answers "who is online" without ' +
       'running anything; a machine that is listed may still refuse the next command for a reason ' +
-      'the capability gate will report.',
+      'the capability gate will report. With include_discovered it also reports what mDNS sees on ' +
+      'the LAN in `discovery.candidates` - unverified hints, each marked with whether the relay ' +
+      'lists it, and never a dispatch target on their own.',
     parameters: {
       include_stale: {
         type: 'boolean',
         description: 'Include machines whose lease looks expired rather than only the live ones. Defaults to false.',
       },
+      include_discovered: {
+        type: 'boolean',
+        default: false,
+        description:
+          'Also scan the LAN over mDNS for about 1.5 s and report discovered candidates under ' +
+          '`discovery`. Each candidate is marked unverified and carries `known_to_relay`; only a ' +
+          'candidate the relay also lists (by machine_id) may be dispatched to. Defaults to false.',
+      },
     },
     output: jsonOutput,
     async execute(args, exec) {
@@ -2947,6 +2968,22 @@ export async function apply(ctx, config = {}) {
       const includeStale = args?.include_stale === true;
       const listed = fleet.records.filter((machine) => includeStale || String(machine?.state ?? 'online') !== 'expired');
 
+      // The LAN hint layer. `include_discovered` runs one bounded scan; without it the block reports
+      // the *last* scan's candidates against this call's roster, so a model that already paid for a
+      // scan does not have to pay for it twice. Either way the candidates are described by
+      // `mdns-discover.mjs`, which is where "discovered but not relay-listed is not a target" lives.
+      const discoveryRequested = args?.include_discovered === true;
+      const discoveryBlock = discoveryRequested
+        ? await discovery.scan({
+            signal: exec?.signal,
+            relayRecords: fleet.records,
+            localMachineId: identity.machine_id,
+          })
+        : discovery.describe({ relayRecords: fleet.records, localMachineId: identity.machine_id });
+      // A cancel during the scan window is a cancel, not a partial answer: the signal is our caller's
+      // withdrawal, and the same typed abort every other wait in this file produces is owed to them.
+      if (discoveryRequested) throwIfAborted(exec?.signal);
+
       return JSON.stringify(
         {
           ok: true,
@@ -2965,6 +3002,7 @@ export async function apply(ctx, config = {}) {
             state: machine?.state ?? 'online',
           })),
           excluded_stale: includeStale ? 0 : fleet.records.length - listed.length,
+          discovery: discoveryBlock,
           note:
             'a listed device has paired and is being tracked; it can still refuse the next command ' +
             'at the capability gate, which w2m_wait reports per machine as `refused`',