@quolu/lattice 0.53.1 → 0.54.0

package/bin/lattice-bridge.mjs

@@ -2,14 +2,26 @@
 
 import { readBridgeConfig } from '../src/bridge-config.mjs';
 import {
-  readBridgeStopRequest, removeBridgeDaemonActiveMarker, removeBridgeDaemonDescriptor,
-  writeBridgeDaemonDescriptor, writeBridgeStopReceipt,
+  bridgeDaemonVersionDrifted, readBridgeStopRequest, removeBridgeDaemonActiveMarker,
+  removeBridgeDaemonDescriptor, writeBridgeDaemonDescriptor, writeBridgeStopReceipt,
 } from '../src/bridge-daemon.mjs';
 import { createBridgeHubHeartbeatController } from '../src/bridge-hub-heartbeat.mjs';
+import { migrateBridgeToHub, retireBridgeTunnelLaunchAgent } from '../src/bridge-hub-migration.mjs';
+import { bridgeRegistrarSettings } from '../src/bridge-registrar.mjs';
 import { bridgeRuntimeController } from '../src/bridge-server.mjs';
 
+// Throttles bridgeDaemonVersionDrifted's disk read and the migration/tunnel-
+// retirement checks' subprocess calls (ssh, launchctl) — the 250ms reconcile
+// tick exists for local responsiveness, not for polling external processes
+// 4x/sec. Migration and tunnel-retirement share this interval: once migrated,
+// the migration check itself becomes a single cheap config-field read
+// (`current.hub !== null`), so there is no cost to leaving both armed forever.
+const BACKGROUND_CHECK_INTERVAL_MS = 60_000;
+
 const env = process.env;
 const hubHeartbeat = createBridgeHubHeartbeatController({ env });
+let lastVersionCheckAt = 0;
+let lastMigrationCheckAt = 0;
 const instanceToken = env.LATTICE_BRIDGE_INSTANCE_TOKEN;
 if (typeof instanceToken !== 'string' || !/^[0-9a-f]{64}$/u.test(instanceToken)) {
   process.stderr.write(`${JSON.stringify({ schema: 'lattice.bridge_daemon_error.v1',
@@ -61,6 +73,51 @@ timer = setInterval(async () => {
       await removeBridgeDaemonActiveMarker({ env });
       process.exit(0);
     }
+    // A stale-version exit is a clean stop, not a failure: whatever supervises
+    // this process (launchd KeepAlive, the Windows supervisor loop) relaunches
+    // it immediately, and the fresh process imports whatever is on disk now —
+    // this is the mechanism that makes "npm update, done" actually true rather
+    // than leaving an already-running daemon serving replaced code forever.
+    if (Date.now() - lastVersionCheckAt >= BACKGROUND_CHECK_INTERVAL_MS) {
+      lastVersionCheckAt = Date.now();
+      if (await bridgeDaemonVersionDrifted({})) {
+        await close();
+        await removeBridgeDaemonDescriptor({ env });
+        await removeBridgeDaemonActiveMarker({ env });
+        process.exit(0);
+      }
+    }
+    // bh5 auto-migration: a terminal still carrying the pre-hub registrar env
+    // (LaunchAgent-baked, so it outlives any single process) upgrades itself
+    // to hub registration with no operator action — see bridge-hub-migration.mjs's
+    // module doc for why this is the whole point of the owner's "update it,
+    // done" acceptance test. Runs on the same throttle as the version check;
+    // once migrated it is a single cheap config-field read, so leaving it
+    // armed forever costs nothing. Tunnel retirement is attempted alongside
+    // it (not gated to the migration transition alone) so a retirement that
+    // failed once keeps getting retried rather than being a one-shot.
+    if (Date.now() - lastMigrationCheckAt >= BACKGROUND_CHECK_INTERVAL_MS) {
+      lastMigrationCheckAt = Date.now();
+      if (bridgeRegistrarSettings(env) !== null) {
+        await migrateBridgeToHub({ env }).catch((error) => {
+          process.stderr.write(`${JSON.stringify({ schema: 'lattice.bridge_daemon_error.v1',
+            code: error?.code ?? 'BRIDGE_HUB_MIGRATION_FAILED',
+            message: error?.message ?? 'bridge hub migration failed' })}\n`);
+        });
+        const migratedConfig = await readBridgeConfig({ env });
+        if (migratedConfig?.hub !== null && migratedConfig?.hub !== undefined) {
+          // retireBridgeTunnelLaunchAgent's own contract never throws for any
+          // expected outcome (not loaded, bootout failure, launchctl absent —
+          // all typed returns); this catch is only for a genuinely unexpected
+          // bug in that function, and it is still logged, not swallowed.
+          await retireBridgeTunnelLaunchAgent({ env }).catch((error) => {
+            process.stderr.write(`${JSON.stringify({ schema: 'lattice.bridge_daemon_error.v1',
+              code: error?.code ?? 'BRIDGE_TUNNEL_RETIREMENT_FAILED',
+              message: error?.message ?? 'bridge tunnel retirement failed' })}\n`);
+          });
+        }
+      }
+    }
     const config = await readBridgeConfig({ env });
     if (config === null || !config.enabled) {
       await close();

package/bin/lattice-dashboard.mjs

@@ -1,10 +1,7 @@
 #!/usr/bin/env node
 
-import { stat } from 'node:fs/promises';
-import path from 'node:path';
-
-import { readTodoStoreStable } from '../src/todo-store.mjs';
 import { TODO_STATUS_DISPATCH_ONLY, projectTodoStatus } from '../src/todo-status.mjs';
+import { createTodoStoreCache } from '../src/todo-store-cache.mjs';
 import { ganttLiveHeadDigest, renderTodoGanttForProject } from '../src/todo-cli.mjs';
 import { readProjectExternalPane } from '../src/project-identity.mjs';
 import {
@@ -35,27 +32,12 @@ const port = typeof configured === 'string' && /^(?:0|[1-9][0-9]{0,4})$/u.test(c
 const registry = createTodoGanttProjectRegistry();
 const roots = new Map();
 const reportedStoreReadFailures = new Set();
-const storeCache = new Map();
-
-function manifestFingerprint(value) {
-  return `${value.dev}:${value.ino}:${value.size}:${value.mtimeMs}:${value.ctimeMs}`;
-}
-
-async function readCachedStore(repoRoot) {
-  const manifestRef = path.join(repoRoot, '.lattice', 'todo', 'manifest.json');
-  for (let attempt = 0; attempt < 3; attempt += 1) {
-    const beforeFingerprint = manifestFingerprint(await stat(manifestRef));
-    const cached = storeCache.get(repoRoot);
-    if (cached?.fingerprint === beforeFingerprint) return cached.store;
-    const store = await readTodoStoreStable({ repoRoot });
-    const afterFingerprint = manifestFingerprint(await stat(manifestRef));
-    if (beforeFingerprint === afterFingerprint) {
-      storeCache.set(repoRoot, { fingerprint: afterFingerprint, store });
-      return store;
-    }
-  }
-  return readTodoStoreStable({ repoRoot });
-}
+// room 2488's "gantt serve固着" symptom (a dashboard stuck on stale/broken state that
+// only a process restart cleared, with the store and git both already fixed) traced to
+// this cache — see src/todo-store-cache.mjs for why it is content-digest keyed rather
+// than stat()-fingerprint keyed.
+const storeCache = createTodoStoreCache();
+const readCachedStore = (repoRoot) => storeCache.read(repoRoot);
 
 async function synchronize() {
   const active = await readVisibleTodoDashboardProjects({ env,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@quolu/lattice",
-  "version": "0.53.1",
+  "version": "0.54.0",
   "description": "Schedulability compiler for multi-agent development: observe real code boundaries, refactor the conflicting seam, recompile the plan for parallel execution",
   "author": {
     "name": "Quo / クオ at kitepon.dev",

package/src/bridge-address.mjs

@@ -105,3 +105,22 @@ export function resolveBridgeListenAddress({ configured, interfaces = {} } = {})
   return { state: 'rebindable', effective: candidates[0], configured: wanted, candidates,
     reason: 'configured_address_absent_rebound_within_subnet' };
 }
+
+/**
+ * Pick a LAN-facing address for a terminal that has none yet — bh5's Mac
+ * auto-migration (loopback + ssh tunnel → direct hub registration) and any
+ * first-time setup with no address preference. Deliberately simpler than
+ * `resolveBridgeListenAddress`: that function preserves "same intent" across
+ * a DHCP move by requiring a same-subnet match against an already-configured
+ * address, but there is no prior intent to preserve when the terminal had no
+ * LAN presence before. The first non-internal address, sorted for
+ * determinism, is a reasonable default; ambiguous hosts (more than one
+ * candidate) are still reported so a caller can choose to surface that rather
+ * than silently pick.
+ */
+export function pickBridgeLanAddress({ interfaces = {}, family = null } = {}) {
+  const candidates = bridgeHostAddresses(interfaces)
+    .filter((entry) => !entry.internal && (family === null || entry.family === family))
+    .map((entry) => entry.address);
+  return { address: candidates[0] ?? null, candidates };
+}

package/src/bridge-daemon.mjs

@@ -1,7 +1,7 @@
 import { spawn } from 'node:child_process';
 import { randomBytes } from 'node:crypto';
 import { constants as fsConstants } from 'node:fs';
-import { chmod, lstat, open, rename, rm, writeFile } from 'node:fs/promises';
+import { chmod, lstat, open, readFile, rename, rm, writeFile } from 'node:fs/promises';
 import { isIP } from 'node:net';
 import path from 'node:path';
 import { parseTree } from 'jsonc-parser';
@@ -9,6 +9,7 @@ import { parseTree } from 'jsonc-parser';
 import {
   BRIDGE_PORT_MAX, BRIDGE_PORT_MIN, BridgeConfigError, bridgeConfigPaths, readBridgeConfig,
 } from './bridge-config.mjs';
+import packageJson from '../package.json' with { type: 'json' };
 
 const DESCRIPTOR_SCHEMA = 'lattice.bridge_daemon.v1';
 const START_TIMEOUT_MS = 5_000;
@@ -379,3 +380,25 @@ export async function stopBridgeDaemon({ env = process.env } = {}) {
   }
   throw new BridgeConfigError('BRIDGE_DAEMON_STOP_FAILED', 'bridge daemon did not stop');
 }
+
+const PACKAGE_JSON_PATH = path.resolve(import.meta.dirname, '../package.json');
+
+/**
+ * Whether the on-disk package.json now reports a different version than the
+ * one this running process loaded at start — i.e. `npm install`/`update`
+ * replaced the files under a still-running daemon (the "daemon の版持ち"
+ * trap, AGENTS.md: a daemon keeps serving whatever module it imported at
+ * startup no matter what gets installed afterward). A long-running process
+ * cannot hot-swap its own already-imported modules; the only fix is to exit
+ * and let whatever supervises it (launchd's KeepAlive, the Windows
+ * supervisor's restart loop) relaunch a fresh process that imports the new
+ * code. Read failures return `false` — an unrelated fs hiccup must not force
+ * a restart loop.
+ */
+export async function bridgeDaemonVersionDrifted({ packageJsonPath = PACKAGE_JSON_PATH } = {}) {
+  let onDisk;
+  try {
+    onDisk = JSON.parse(await readFile(packageJsonPath, 'utf8'));
+  } catch { return false; }
+  return typeof onDisk?.version === 'string' && onDisk.version !== packageJson.version;
+}

package/src/bridge-hub-migration.mjs

@@ -0,0 +1,118 @@
+/**
+ * Mac auto-migration (bh5): a terminal running the old single-slot topology
+ * (loopback bridge + `LATTICE_BRIDGE_REGISTRAR_*` ssh registrar) upgrades to
+ * hub registration with zero manual commands. The owner's stated acceptance
+ * test is literal: an agent who knows nothing about hub/port/flags/migration
+ * runs a normal package update, and the bridge finds its own way onto the
+ * public page (room 2446, 2461) — no `--hub` flag, no LaunchAgent surgery.
+ *
+ * The trigger is the registrar call the daemon already makes on every new
+ * binding (`bridge-registrar.mjs`'s `registerBridgeUpstream`, used by
+ * `bridge-launch-agent.mjs`'s plist today only to keep the reverse-proxy
+ * literal current). The v2 registrar script (room 2452) additionally returns
+ * `hub_url` in that same response — this module is what turns "a hub_url
+ * showed up in a registration reply" into "reconfigure this bridge to use
+ * it and retire the ssh tunnel", entirely from information the terminal
+ * already had a reason to ask for.
+ */
+
+import { execFile } from 'node:child_process';
+import { rm } from 'node:fs/promises';
+import { networkInterfaces } from 'node:os';
+import path from 'node:path';
+import { promisify } from 'node:util';
+
+import { pickBridgeLanAddress } from './bridge-address.mjs';
+import { configureBridge, readBridgeConfig } from './bridge-config.mjs';
+import {
+  bridgeRegistrarSettings, deriveBridgeHubUrlFromRegistration, registerBridgeUpstream,
+} from './bridge-registrar.mjs';
+
+const execFileAsync = promisify(execFile);
+
+/** The ssh reverse-tunnel LaunchAgent from the pre-hub topology
+  * (docs/operations/lattice-kitepon-deployment.md) — distinct from
+  * `dev.kitepon.lattice.bridge`, which `bridge-launch-agent.mjs` owns and
+  * this migration never touches. */
+export const BRIDGE_TUNNEL_LAUNCH_AGENT_LABEL = 'dev.kitepon.lattice.bridge-tunnel';
+
+/**
+ * Attempt one migration step. Called from the daemon's reconcile loop, so it
+ * must be cheap to call when there is nothing to do and must never throw for
+ * a condition the caller should just keep running through (no registrar
+ * configured, already migrated, hub unreachable this cycle) — only a
+ * genuinely invalid registrar env (`bridgeRegistrarSettings`'s own
+ * half-configured-pair failure) propagates, matching every other registrar
+ * caller's behavior.
+ */
+export async function migrateBridgeToHub({
+  env = process.env, interfaces = networkInterfaces(), readConfig = readBridgeConfig,
+  configure = configureBridge, register = registerBridgeUpstream,
+} = {}) {
+  const registrar = bridgeRegistrarSettings(env);
+  if (registrar === null) return { migrated: false, reason: 'registrar_not_configured' };
+  const current = await readConfig({ env });
+  if (current === null || !current.enabled) return { migrated: false, reason: 'bridge_not_enabled' };
+  if (current.hub !== null) return { migrated: false, reason: 'already_migrated' };
+
+  const registration = await register({ port: current.listen.port, env });
+  const hubUrl = deriveBridgeHubUrlFromRegistration(registration);
+  if (hubUrl === null) return { migrated: false, reason: 'no_hub_url_available', registration };
+
+  const picked = pickBridgeLanAddress({ interfaces });
+  if (picked.address === null) return { migrated: false, reason: 'no_lan_address_available' };
+
+  const updated = await configure({
+    address: picked.address, port: null, reuseCurrentPort: false,
+    upstream: current.upstream, hub: { url: hubUrl },
+    allowedHosts: current.allowed_hosts.filter((host) => host !== current.listen.address),
+    env,
+  });
+  return { migrated: true, config: updated, hubUrl };
+}
+
+function launchAgentPlistPath(label, env) {
+  const home = env.HOME;
+  if (typeof home !== 'string' || !path.isAbsolute(home)) return null;
+  return path.join(home, 'Library', 'LaunchAgents', `${label}.plist`);
+}
+
+/**
+ * Retire the pre-hub ssh reverse-tunnel LaunchAgent, once migration has
+ * actually landed a hub URL — never speculatively, so a bridge that never
+ * reaches `migrateBridgeToHub`'s success path never touches this agent.
+ * Idempotent and non-fatal: a tunnel that is not loaded (already retired, or
+ * this deployment never had one) is success, not an error, and any
+ * `launchctl` failure here must not crash a daemon whose primary job — hub
+ * registration — has already succeeded by the time this runs.
+ */
+export async function retireBridgeTunnelLaunchAgent({
+  env = process.env, uid = process.getuid?.(), runner = defaultTunnelLaunchctlRunner,
+  label = BRIDGE_TUNNEL_LAUNCH_AGENT_LABEL,
+} = {}) {
+  if (!Number.isSafeInteger(uid) || uid < 0) return { retired: false, reason: 'uid_unavailable' };
+  const service = `gui/${uid}/${label}`;
+  let probe;
+  try { probe = await runner(['print', service]); } catch { return { retired: false, reason: 'launchctl_unavailable' }; }
+  if (probe.code !== 0) return { retired: false, reason: 'not_loaded' };
+  try {
+    const bootout = await runner(['bootout', service]);
+    if (bootout.code !== 0) return { retired: false, reason: 'bootout_failed' };
+  } catch { return { retired: false, reason: 'bootout_failed' }; }
+  const plistPath = launchAgentPlistPath(label, env);
+  if (plistPath !== null) await rm(plistPath, { force: true }).catch(() => {});
+  return { retired: true };
+}
+
+export async function defaultTunnelLaunchctlRunner(args) {
+  try {
+    const result = await execFileAsync('/bin/launchctl', args, { encoding: 'utf8' });
+    return { code: 0, stdout: result.stdout ?? '', stderr: result.stderr ?? '' };
+  } catch (error) {
+    if (Number.isInteger(error?.code)) {
+      return { code: error.code, stdout: error.stdout ?? '', stderr: error.stderr ?? '' };
+    }
+    throw error;
+  }
+}
+

package/src/bridge-hub-server.mjs

@@ -523,6 +523,15 @@ export async function startBridgeHubServer({
       return;
     }
     const rawPath = requestUrl.split('?', 1)[0];
+    if (rawPath === '/') {
+      // The old single-terminal bridge served the project index at root; hub only ever
+      // routed `/projects/*`, so the public entrance 404'd (room 2488 — a functional
+      // regression, not a design one: the front door itself was gone, independent of how
+      // it looks). Redirect rather than duplicate handleProjectsIndex's logic at a second path.
+      response.writeHead(301, { location: '/projects/', 'cache-control': 'no-store' });
+      response.end();
+      return;
+    }
     if (rawPath === '/__lattice/hub/register') { await handleRegister(incoming, response); return; }
     if (rawPath === '/projects/') { await handleProjectsIndex(incoming, response); return; }
     const match = PROJECT_ROUTE.exec(rawPath);

package/src/todo-store-cache.mjs

@@ -0,0 +1,45 @@
+/**
+ * A content-addressed cache in front of `readTodoStoreStable`, for long-running
+ * processes (the dashboard daemon) that re-render on every request/poll and would
+ * otherwise re-validate the whole merged store every time.
+ *
+ * An earlier version keyed the cache on a stat()-derived fingerprint (dev/ino/size/
+ * mtimeMs/ctimeMs). On filesystems with coarse mtime granularity (observed on
+ * WSL/DrvFs), two different manifest contents written close together can land on the
+ * same fingerprint, so the cache would serve a stale store as current — and since the
+ * mismatch never surfaces as an error, nothing invalidates it until the process
+ * restarts. Hashing the manifest's actual bytes costs one small file read and removes
+ * that failure mode entirely.
+ */
+
+import { createHash } from 'node:crypto';
+import { readFile } from 'node:fs/promises';
+import path from 'node:path';
+
+import { readTodoStoreStable } from './todo-store.mjs';
+
+async function manifestContentDigest(manifestRef) {
+  return createHash('sha256').update(await readFile(manifestRef)).digest('hex');
+}
+
+/**
+ * @param {object} [options]
+ * @param {(options: object) => Promise<object>} [options.readStable] injection point for tests
+ */
+export function createTodoStoreCache({ readStable = readTodoStoreStable } = {}) {
+  const cache = new Map();
+  return {
+    async read(repoRoot) {
+      const manifestRef = path.join(repoRoot, '.lattice', 'todo', 'manifest.json');
+      const digest = await manifestContentDigest(manifestRef);
+      const cached = cache.get(repoRoot);
+      if (cached?.digest === digest) return cached.store;
+      // Read before caching: a failure here (including the store's own inconsistency
+      // detection) must not populate the cache, so the very next call re-reads instead
+      // of serving a poisoned entry.
+      const store = await readStable({ repoRoot });
+      cache.set(repoRoot, { digest, store });
+      return store;
+    },
+  };
+}

package/src/todo-store.mjs

@@ -1392,6 +1392,18 @@ export async function readTodoStoreStable(options = {}) {
   if (!Number.isSafeInteger(maximumAttempts) || maximumAttempts < 1 || maximumAttempts > 16) {
     throw new TypeError('maximumAttempts must be 1..16');
   }
+  // `manifest_journal_head_mismatch`/`manifest_plan_binding_mismatch` are treated as a
+  // transient in-flight write and retried. That is correct while a concurrent writer is
+  // mid-commit, but a crashed writer can leave the SAME mismatch permanently — retrying
+  // forever against a manifest that never changes just burns attempts and then reports
+  // a content-free STORE_BUSY, hiding the real STORE_INCONSISTENT reason the caller needs
+  // to actually recover (2026-08-10 P0: a crashed `todo start` left exactly this behind).
+  // Track the manifest digest seen at the START of the previous attempt: if it is
+  // unchanged going into this attempt too, no writer completed anything in between, so
+  // the "transient" classification no longer has evidence behind it — surface the real
+  // error instead of exhausting the budget on a window that was never closing.
+  let previousAttemptManifestDigest = null;
+  let lastError = null;
   for (let attempt = 1; attempt <= maximumAttempts; attempt += 1) {
     const before = await readArtifact(repoRoot, MANIFEST_REF, {
       code: 'STORE_INCONSISTENT', maxBytes: TODO_LIMITS.snapshotBytes, validate: validateTodoManifest,
@@ -1411,12 +1423,21 @@ export async function readTodoStoreStable(options = {}) {
       const transientWriteWindow = error.code === 'STORE_INCONSISTENT'
         && ['manifest_journal_head_mismatch', 'manifest_plan_binding_mismatch']
           .includes(error.detail.reason);
-      if (before.manifest_digest === after.manifest_digest && !transientWriteWindow) throw error;
+      const stableAcrossAttempts = previousAttemptManifestDigest === before.manifest_digest;
+      if (before.manifest_digest === after.manifest_digest
+        && (!transientWriteWindow || stableAcrossAttempts)) throw error;
+      lastError = error;
     }
+    previousAttemptManifestDigest = before.manifest_digest;
     if (attempt < maximumAttempts) {
       await new Promise((resolve) => setTimeout(resolve, Math.min(16, 2 ** attempt)));
     }
   }
+  // Exhausted without ever observing a genuinely closing write. Surface the last typed
+  // STORE_INCONSISTENT reason rather than a bare STORE_BUSY — the caller (and a human
+  // reading the error) needs to know which store artifact actually disagrees, not just
+  // that reads kept failing.
+  if (lastError !== null) throw lastError;
   fail('STORE_BUSY', 'stable_read_exhausted', { attempts: maximumAttempts });
 }