@gotcos/glasses-server 6.37.2 → 6.37.3

package/.env.example

@@ -218,7 +218,26 @@ BIND_HOST=0.0.0.0
 # 72 covers a long weekend away from the Mac.
 # COS_UNSAVED_AUDIO_RETENTION_HOURS=72
 
-# ── LIVE CUES (optional — requires the FULL COS PIPELINE above) ──────────
+# ── LIVE CUES (optional, OFF unless you set it — costs model calls) ───────
+#
+# The closest thing COS has to conversing with you DURING a meeting rather than
+# after it: it reads the running transcript and puts short prompts on the lens —
+# the question you would have wanted to ask, while it is still useful.
+#
+# WHAT IT COSTS, PLAINLY. Every cue is a model call on YOUR agent quota, not a
+# free local pass. It is bounded, but it is not free:
+#
+#     at most 8 cue pipelines per meeting     MAX_PIPELINES_PER_MEETING
+#     at most one start per 60 seconds        FLOOR_BETWEEN_STARTS_MS
+#     30s cooldown after a cue fires          COOLDOWN_AFTER_CUE_MS
+#
+# So a long meeting can spend up to 8 model calls it would not otherwise make,
+# and each one runs the full planner -> memory -> insight chain below. On a
+# metered plan, leave this off. It IS off unless you set it.
+#
+# Runs on Cursor Composer. That is the only supported model in v1 — any other
+# value fails closed rather than quietly billing a model you did not choose.
+#
 # Live meeting coaching cues on the lens: transcript window -> Composer
 # planner -> Qdrant -> LightRAG -> Composer insight -> coaching_nudge.
 # Requires COS_SCRIPTS_DIR (memory hops) plus the Cursor Agent CLI.

package/CHANGELOG.md

@@ -1,3 +1,40 @@
+## 6.37.3
+
+Speaker ID works out of the box. A silent capture stops retrying forever. Live
+Cues says what it costs.
+
+SPEAKER MODEL IN NORMAL SETUP. 6.37.2 added `--setup-speaker-model` as a command
+you had to know about. Now the voiceprint model is fetched during ordinary setup,
+with `SKIP_SPEAKER_MODEL_DOWNLOAD=1` as the escape hatch — the same shape the
+whisper models already use, and those are 1.5 GB, sixty times larger.
+
+Deliberately NOT a lazy fetch on first use. Lazy would fire the download at the
+START OF A MEETING: 26 MB on hotel wifi degrading the exact session it exists to
+improve, and a GitHub release asset promoted from an install-time dependency to a
+runtime one. A failed fetch is non-fatal — diarization is opt-in by design, so
+the server falls back to wearer/Ext rather than refusing to start.
+
+A CAPTURE WITH NO SPEECH IS RECOVERED, NOT FAILED. The recover route threw
+`recovery produced an empty transcript` when whisper returned nothing, which left
+the capture in the unsaved list and retried it on every boot and every button
+press. Measured on this machine: one 33-second capture failed that way 1,131
+times, alternating in the log with the auto-recover path claiming success for the
+same session, while the panel showed only "1 recoverable" and the error went to
+stderr where nobody looks. The user presses Recover and nothing happens, because
+nothing can.
+
+Silence now writes a receipt with `outcome: 'no_speech'` and clears. The audio is
+NOT deleted — it leaves on the ordinary retention clock, so a capture wrongly
+judged silent by a bad decode is still on disk for its full window.
+
+LIVE CUES NOW STATES ITS COST. The flag was already documented as a master
+switch; what was missing was that every cue is a model call on your own agent
+quota. Now stated with the real bounds — at most 8 pipelines per meeting, one
+start per 60s, 30s cooldown — and the real model, Cursor Composer, which is the
+only supported value in v1.
+
+223 files, 3117 tests.
+
 ## 6.37.2
 
 `--setup-speaker-model` — one command to install the voiceprint model.

package/bin/cli.cjs

@@ -706,6 +706,38 @@ if (whisperCliPath && SETUP_TRANSCRIPTION) {
   }
 }
 
+// Step 5b: the voiceprint model — named per-speaker diarization.
+//
+// Folded into normal setup rather than fetched lazily on first use. Lazy would
+// fire the download at the START OF A MEETING, which is the worst possible
+// moment: 26 MB on hotel wifi degrades the exact session it exists to improve,
+// and it makes a GitHub release asset a runtime dependency instead of an
+// install-time one.
+//
+// This follows the precedent already set by the whisper models, which are 1.5 GB
+// -- sixty times larger -- downloaded here with SKIP_WHISPER_DOWNLOAD as the
+// escape hatch. Same shape, same opt-out, so nobody on a metered or restricted
+// network is forced into it.
+if (!SETUP_SPEAKER_MODEL && process.env.SKIP_SPEAKER_MODEL_DOWNLOAD !== '1') {
+  const spkDest = join(CONFIG_DIR, 'models', SPEAKER_MODEL.filename)
+  if (existsSync(spkDest) && sha256File(spkDest) === SPEAKER_MODEL.sha256) {
+    console.log(green('  ✓') + ' Voiceprint model ready ' + dim('— named speakers available'))
+  } else {
+    console.log('    ' + dim('Fetching the voiceprint model (~26 MB) for named speakers.'))
+    console.log('    ' + dim('Skip: SKIP_SPEAKER_MODEL_DOWNLOAD=1 npx --yes @gotcos/glasses-server@latest'))
+    const spkCode = setupSpeakerModel()
+    if (spkCode !== 0) {
+      // NOT fatal. Diarization is opt-in by design: without the model the server
+      // stays on wearer/Ext fallback rather than failing, so a failed fetch must
+      // not block a working install.
+      console.log(yellow('  ⚠') + ' Voiceprint model unavailable ' + dim('— speakers stay wearer/Ext'))
+      console.log('    ' + dim('Retry later: npx --yes @gotcos/glasses-server@latest --setup-speaker-model'))
+    }
+  }
+} else if (process.env.SKIP_SPEAKER_MODEL_DOWNLOAD === '1') {
+  console.log(yellow('  ⚠') + ' SKIP_SPEAKER_MODEL_DOWNLOAD=1 — named speakers unavailable')
+}
+
 if (PREPARE_ONLY && SETUP_TRANSCRIPTION) {
   console.log('')
   if (transcriptionSetupFailures.length > 0) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gotcos/glasses-server",
-  "version": "6.37.2",
+  "version": "6.37.3",
   "description": "COS Glasses — self-hosted AI heads-up-display server for Even G2 smart glasses, powered by Claude Code, Codex, or Cursor Agent CLI",
   "type": "module",
   "bin": {

package/server/lib/unsaved-audio-quarantine.ts

@@ -303,6 +303,41 @@ export function isRecoveryActive(dirName: string): boolean {
   return activeRecoveries.has(dirName)
 }
 
+/**
+ * A capture that held no speech is RECOVERED, not failed.
+ *
+ * The recover route used to throw `recovery produced an empty transcript` when
+ * whisper returned nothing, which left the capture in the unsaved list and
+ * retried it on every boot and every button press. Measured 2026-08-25: one
+ * 33-second capture failed this way 1,131 times, alternating with the
+ * auto-recover path claiming success, while the panel said only "1 recoverable"
+ * and the error went to stderr where nobody looks.
+ *
+ * Silence is a legitimate outcome. There is nothing to save and nothing to
+ * retry, so the capture is receipted and clears -- with `outcome: 'no_speech'`
+ * so the distinction stays auditable rather than looking like a normal save.
+ *
+ * The audio is NOT deleted here. It leaves on the ordinary retention clock, so a
+ * capture wrongly judged silent (a bad decode, a broken model) is still on disk
+ * for its full window.
+ */
+export function markRecoveredNoSpeech(dirPath: string, chunkFiles: number): void {
+  try {
+    writeFileSync(
+      resolve(dirPath, RECOVERED_RECEIPT),
+      `${JSON.stringify({
+        schemaVersion: 1,
+        recoveredAt: new Date().toISOString(),
+        outcome: 'no_speech',
+        chunkFiles,
+        words: 0,
+        note: 'Transcription produced no words. Audio retained until the retention clock clears it.',
+      })}\n`,
+      { encoding: 'utf8', mode: 0o600 },
+    )
+  } catch { /* receipt is best-effort; findBySessionId remains the true guard */ }
+}
+
 export function markRecovered(dirPath: string, savedFilename: string): void {
   try {
     writeFileSync(

package/server/routes/meeting.ts

@@ -83,6 +83,7 @@ import {
   findQuarantineDir,
   listUnsavedCaptures,
   markRecovered,
+  markRecoveredNoSpeech,
   registerActiveRecovery,
 } from '../lib/unsaved-audio-quarantine.js'
 import {
@@ -1915,7 +1916,24 @@ export function createMeetingRouter(deps: MeetingRouteDependencies = {}): Router
       clearSessionHallucinationState(sessionId)
       const transcript = cleanFinalTranscript(results.map(item => item.text).join(' '))
       if (!transcript.trim()) {
-        throw new Error('recovery produced an empty transcript')
+        // SILENCE IS AN OUTCOME, NOT A FAILURE.
+        //
+        // This used to throw, which left the capture in the unsaved list and
+        // retried it forever -- 1,131 times for one 33-second capture on
+        // 2026-08-25, alternating with the auto-recover path claiming success,
+        // while the panel showed only "1 recoverable" and the error went to
+        // stderr. The user sees a Recover button that does nothing, because it
+        // literally cannot succeed on audio with no speech in it.
+        //
+        // Receipt it and let it clear. The audio stays on disk for its full
+        // retention window, so a capture wrongly judged silent is not lost.
+        markRecoveredNoSpeech(quarantineDir, entries.length)
+        console.log(
+          `[meeting/orphans] ${sessionId} held no speech `
+          + `(${entries.length} chunks, ${Math.round((Date.now() - startedAt) / 1000)}s) `
+          + '— receipted as no_speech; audio retained until retention expires',
+        )
+        return
       }
       const recoveredChunks = results.map(item => ({
         text: item.text,