@gotcos/glasses-server 6.21.4 → 6.21.6

package/.env.example

@@ -28,13 +28,13 @@ BIND_HOST=0.0.0.0
 # server after Wi-Fi/Tailscale changes and process restarts.
 # COS_SERVER_INSTANCE_ID_PATH=/path/to/server-instance-id
 
-# Build 204+ can opt into server-owned durable query jobs. The server fsyncs an
+# Build 204+ uses server-owned durable query jobs by default. The server fsyncs an
 # accepted prompt before returning 202, then keeps provider work running if the
 # phone backgrounds, reloads, or changes networks. Reopening COS reattaches to
-# the same job. Set to 0 (or remove) to send NEW prompts through the legacy
+# the same job. Set to 0 to send NEW prompts through the legacy
 # streaming route; already accepted durable jobs remain readable/cancellable
 # and drain safely.
-# COS_DURABLE_QUERY_JOBS=1
+# COS_DURABLE_QUERY_JOBS=0
 
 # ── THE LLM (chat) ──────────────────────────────────────────────────────
 # Chat runs through your LOCAL agent CLI — NOT an API key:

package/CHANGELOG.md

@@ -1,3 +1,23 @@
+## 6.21.6
+
+- Make server-owned durable query jobs the default so accepted replies keep
+  running through phone backgrounding, WebView reloads, and network handoffs.
+- Preserve `COS_DURABLE_QUERY_JOBS=0` as the machine-wide rollback. Existing
+  accepted jobs remain recoverable and cancellable while new prompts use the
+  legacy streaming path.
+- Publish the capability truth through authenticated health/model surfaces so
+  COS Control and the companion can independently enforce machine and device
+  preferences without changing the proven two-step cancel gesture.
+
+## 6.21.5
+
+- Drop the narrow Large-v3 stock-caption artifact family beginning with
+  “closed captioning provided by” from prompt dictation.
+- When that exact HQ artifact occurs, retain the already-validated Fast
+  transcript for the same audio and report the result as degraded instead of
+  saving either the caption credit or an empty gap. All other HQ and no-speech
+  paths are unchanged.
+
 ## 6.21.4
 
 - Preserve validated phone-photo references on Claude and Codex conversation

package/README.md

@@ -102,8 +102,8 @@ range is the exact Tailscale/CGNAT allocation (`100.64.0.0/10`), not all of
 ## What it does
 
 - Ask anything, get a streamed answer on the lens (`/api/query`, `/v1/chat/completions`)
-- With COS Glasses build 204+, opt into server-owned durable queries with
-  `COS_DURABLE_QUERY_JOBS=1`: accepted work survives phone backgrounding,
+- With COS Glasses build 204+, server-owned durable queries are on by default:
+  accepted work survives phone backgrounding,
   WebView reloads, and network handoffs, then reattaches without duplicate work
   or duplicate replies
 - Choose Opus, Fable, Sonnet, GPT Frontier, GPT Balanced, Composer 2.5 Fast, or
@@ -174,8 +174,8 @@ in the managed CLI working directory),
 `COS_CURSOR_PERSIST_SESSIONS=0` (disable Cursor session resume),
 `COS_WEATHER_DEFAULT_LAT` / `COS_WEATHER_DEFAULT_LON` /
 `COS_WEATHER_DEFAULT_CITY` (optional home fallback when phone GPS is denied),
-`COS_SCRIPTS_DIR` (full pipeline), `COS_DURABLE_QUERY_JOBS=1` (build 204+
-server-owned query recovery), and `COS_MEDIA_ROOT` (optional image-store
+`COS_SCRIPTS_DIR` (full pipeline), `COS_DURABLE_QUERY_JOBS=0` (optional
+machine-wide rollback for build 204+ server-owned query recovery), and `COS_MEDIA_ROOT` (optional image-store
 location; default `~/.cos-glasses/data/media`). Your name + transcription vocabulary live in
 `~/.cos-glasses/.cos-profile.json` (see `.cos-profile.example.json`).
 Factory example values are ignored; add the real names, companies, acronyms,
@@ -338,10 +338,10 @@ BIND_HOST=0.0.0.0 npm run start:server
   only when OpenAI playback is intentionally configured.
 - *Photos unavailable?* — install `ffmpeg`, restart the server, and confirm `/api/health` reports `features.mediaProcessingReady: true`.
 - *Prompt recovery unavailable?* — update with `npx --yes @gotcos/glasses-server@latest`, then confirm `/api/health` reports `features.promptRecovery: true`.
-- *Durable query recovery unavailable?* — build 204+ requires server 6.10.0+ and
-  `COS_DURABLE_QUERY_JOBS=1`. Restart once, then confirm `/api/health` reports
+- *Durable query recovery unavailable?* — build 204+ requires server 6.10.0+.
+  Restart once, then confirm `/api/health` reports
   `features.durableQueryJobs: true`, protocol `1`, and state `ready`. To roll
-  back, remove the flag; accepted jobs still drain while new prompts use legacy streaming.
+  back, set `COS_DURABLE_QUERY_JOBS=0`; accepted jobs still drain while new prompts use legacy streaming.
 - *Offline meeting recovery unavailable?* — build 209+ requires server 6.11.0+.
   Restart once, then confirm `/api/health` reports
   `features.localFirstMeetings: true` and

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gotcos/glasses-server",
-  "version": "6.21.4",
+  "version": "6.21.6",
   "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/index.ts

@@ -432,23 +432,22 @@ listenRequiredServers(listeners).then(() => {
     console.log(`[startup] Silero VAD: ${vadOk ? 'active' : 'disabled (model not found)'}`)
   }
 
-  // Durable recovery and state-mutating background work remain closed until an
-  // authenticated controller release. The lifecycle callback is delivered for
-  // both an already-open fresh boot and a later cross-boot release.
+  // Durable recovery starts with the admitted runtime by default. A literal
+  // COS_DURABLE_QUERY_JOBS=0 is the machine-wide rollback.
   let admittedRuntimeStarted = false
   const startAdmittedRuntime = () => {
     if (admittedRuntimeStarted) return
     admittedRuntimeStarted = true
 
     void initQueryJobRuntime().then(health => {
-      if (process.env.COS_DURABLE_QUERY_JOBS === '1') {
+      if (process.env.COS_DURABLE_QUERY_JOBS !== '0') {
         console.log(`[COS API] Durable query jobs: ${health.store.state} · ${health.store.retainedIdentities} retained`)
       } else {
-        console.log('[COS API] Durable query jobs: disabled (set COS_DURABLE_QUERY_JOBS=1 to enable)')
+        console.log('[COS API] Durable query jobs: disabled by COS_DURABLE_QUERY_JOBS=0')
       }
     }).catch(error => {
       // The store remains degraded and rejects admission. Legacy /api/query is
-      // still mounted, so disabling the feature flag is an immediate rollback.
+      // still mounted, so the kill switch is an immediate rollback.
       console.error('[COS API] Durable query-job store unavailable:', error)
     })
 

package/server/lib/hallucination-filter.ts

@@ -14,8 +14,11 @@
 import { getNegativeRules, getVocabulary, getOwnerName, getWhisperCorrections } from './profile.js'
 
 // ── Whole-chunk silence hallucinations ─────────────────────────────────────
+const CAPTION_CREDIT_HALLUCINATION = /^\s*(?:closed\s+)?caption(?:ing|s)?\s+(?:is\s+)?provided\s+by\b/i
+
 const KNOWN_HALLUCINATIONS = [
   /^subtitles?\s+by\b/i,
+  CAPTION_CREDIT_HALLUCINATION,
   /\blike\s+and\s+subscribe\b/i,
   /\bthanks?\s+for\s+watching\b/i,
   /\bplease\s+subscribe\b/i,
@@ -27,6 +30,13 @@ const KNOWN_HALLUCINATIONS = [
   /^\s*♪+\s*[^♪\n]{0,40}\s*♪+\s*$/,
 ]
 
+/** Narrow stock-caption detector used to retain a valid Fast draft when an HQ
+ * decode leaks a caption credit. It intentionally requires the phrase to begin
+ * the chunk so normal questions about closed captioning remain untouched. */
+export function isCaptionCreditHallucination(text: string): boolean {
+  return CAPTION_CREDIT_HALLUCINATION.test(text)
+}
+
 // ── Inline "Name:" detector (session-aware in streaming, static in one-shot) ──
 const INLINE_HALLUCINATION_THRESHOLD = 3
 // Seed list of known Whisper "Name:" training artifacts. Empty by default — the

package/server/lib/query-job-feature.ts

@@ -1,12 +1,12 @@
 import { QUERY_JOB_PROTOCOL_VERSION } from './query-job-types.js'
 
 /**
- * Durable jobs ship dark. The private canary enables them explicitly with
- * COS_DURABLE_QUERY_JOBS=1; removing the flag is an immediate server-side
- * rollback that leaves the legacy /api/query path untouched.
+ * Durable jobs are the public default. COS_DURABLE_QUERY_JOBS=0 is the
+ * machine-wide rollback; every other value (including unset) keeps the
+ * capability available. COS Control owns the user-facing policy switch.
  */
 export function durableQueryJobsEnabled(env: NodeJS.ProcessEnv = process.env): boolean {
-  return env.COS_DURABLE_QUERY_JOBS === '1'
+  return env.COS_DURABLE_QUERY_JOBS !== '0'
 }
 
 export function durableQueryJobsCapability(): {

package/server/routes/prompt-drafts.ts

@@ -29,6 +29,7 @@ import {
   stripInlineHallucinations,
   stripPromptDictationArtifacts,
   isFullHallucination,
+  isCaptionCreditHallucination,
   isBrandUrlOnly,
   clearSessionHallucinationState,
   applyNegativeRules,
@@ -250,6 +251,35 @@ async function transcribeChunk(draftId: string, chunkIndex: number, audio: Buffe
       return text
     } catch (err) {
       if (err instanceof NoSpeechDetectedError) {
+        // Large-v3 can rarely replace a clipped but valid phrase with a stock
+        // caption credit. Keep the already-decoded Fast text for this one
+        // recognized failure shape instead of persisting either the artifact
+        // or an empty gap. Other no-speech results retain existing behavior.
+        if (mode === 'hq' && isCaptionCreditHallucination(err.rawText)) {
+          const fastJob = modeQualityJobs.get(modeQualityKey(draftId, chunkIndex, 'fast', hash))
+          if (fastJob) {
+            try { await fastJob } catch { /* normal no-speech path below */ }
+          }
+          const fast = loadPromptDraftMeta(draftId)?.warmTranscripts?.[String(chunkIndex)]
+          if (
+            fast
+            && fast.hash === hash
+            && fast.requestedMode === 'fast'
+            && fast.actualQuality === 'fast'
+            && fast.text.trim()
+          ) {
+            if (purpose === 'final') {
+              await markPromptDraftChunkTranscript(draftId, chunkIndex, {
+                ...fast,
+                requestedMode: 'hq',
+                degraded: true,
+                degradationReason: 'hq_caption_credit_hallucination',
+              }, 'final')
+            }
+            console.warn(`[prompt-draft] dropped HQ caption-credit hallucination ${draftId}/${chunkIndex}; retained Fast text`)
+            return fast.text
+          }
+        }
         await markPromptDraftChunkTranscript(draftId, chunkIndex, {
           text: '', hash, requestedMode: mode, actualQuality: mode, backend: 'no-speech', degraded: false,
         }, purpose)