@gotcos/glasses-server 6.17.0 → 6.18.0

package/server/lib/live-cues-memory.ts

@@ -0,0 +1,202 @@
+// Live Cues memory hops.
+//
+// callPython() cannot be used here: it shells exactly one script,
+// cos_api_bridge.py, whose dispatch has no semantic_search or lightrag_search
+// verb — an unknown verb returns nothing useful, indistinguishable from an
+// empty result. Both hops spawn python directly against the exported
+// PYTHON_BIN / COS_SCRIPTS_DIR with their own preflight, timeout, and parse.
+
+import { execFile, spawn, type ChildProcess } from 'node:child_process'
+import { existsSync, readFileSync } from 'node:fs'
+import { resolve } from 'node:path'
+import { COS_SCRIPTS_DIR, PYTHON_BIN } from './python-bridge.js'
+import { terminateProviderProcess } from './provider-process-lifecycle.js'
+import { logTokenAudit } from './token-audit.js'
+
+export interface SemanticHit {
+  title: string
+  date: string
+  domain: string
+  score: number
+  summary: string
+}
+
+export interface Hop1Result {
+  ok: boolean
+  snippets: string[]
+  reason?: string
+}
+
+export interface Hop2Result {
+  ok: boolean
+  text: string | null
+  reason?: string
+  /** False when the python tree (which may hold claude -p grandchildren)
+   *  could not be proven dead after a timeout kill. */
+  treeClosed: boolean
+}
+
+const HOP1_TIMEOUT_MS = 15_000
+const HOP2_TIMEOUT_MS = 40_000
+
+// The LightRAG query pool is 200/day SHARED with /why, --explore, and prep,
+// and it increments per LLM call (~2 per query), not per invocation. The
+// reserve leaves headroom for interactive queries; it does NOT give cues an
+// isolated pool (that would need a third pool in lightrag_adapter.py).
+const LIGHTRAG_DAILY_CAP = 200
+
+function lightragReserve(): number {
+  const raw = Number(process.env.COS_LIVE_CUES_LIGHTRAG_RESERVE ?? 120)
+  return Number.isFinite(raw) && raw >= 0 ? Math.floor(raw) : 120
+}
+
+export function lightragBudgetAllows(): boolean {
+  if (!COS_SCRIPTS_DIR) return false
+  try {
+    const counterPath = resolve(COS_SCRIPTS_DIR, '.lightrag_daily_calls.json')
+    if (!existsSync(counterPath)) return true
+    const parsed = JSON.parse(readFileSync(counterPath, 'utf-8')) as { date?: string; query?: number }
+    const today = new Date().toISOString().slice(0, 10)
+    if (parsed.date !== today) return true
+    const spent = Number.isFinite(parsed.query) ? Number(parsed.query) : 0
+    return (LIGHTRAG_DAILY_CAP - spent) > lightragReserve()
+  } catch {
+    // An unreadable counter fails toward NOT spending: the reserve exists to
+    // protect the interactive pool, and blind spending defeats it.
+    return false
+  }
+}
+
+/** hop1 — Qdrant semantic search. No LLM spend; plain execFile timeout is fine. */
+export function semanticSearchHop(query: string): Promise<Hop1Result> {
+  const pythonBin = PYTHON_BIN
+  const scriptsDir = COS_SCRIPTS_DIR
+  if (!pythonBin || !scriptsDir) {
+    return Promise.resolve({ ok: false, snippets: [], reason: 'no_cos_pipeline' })
+  }
+  const script = resolve(scriptsDir, 'semantic_search.py')
+  if (!existsSync(script)) {
+    return Promise.resolve({ ok: false, snippets: [], reason: 'no_memory_scripts' })
+  }
+  return new Promise(resolveHop => {
+    execFile(
+      pythonBin,
+      // --min-score pinned explicitly: the CLI default is 0.25 but the
+      // programmatic default is 0.35, and inheriting the wrong one silently
+      // cuts recall.
+      [script, query, '--json', '--limit', '5', '--min-score', '0.25'],
+      { cwd: scriptsDir, timeout: HOP1_TIMEOUT_MS, maxBuffer: 1024 * 1024 },
+      (error: Error | null, stdout: string | Buffer) => {
+        if (error) {
+          resolveHop({ ok: false, snippets: [], reason: 'qdrant_unreachable' })
+          return
+        }
+        try {
+          const hits = JSON.parse(String(stdout)) as SemanticHit[]
+          const snippets = (Array.isArray(hits) ? hits : [])
+            .slice(0, 5)
+            .map(hit => `${hit.date ?? '?'} ${hit.title ?? 'Untitled'}: ${String(hit.summary ?? '').slice(0, 220)}`)
+          resolveHop({ ok: true, snippets })
+        } catch {
+          resolveHop({ ok: false, snippets: [], reason: 'qdrant_parse_error' })
+        }
+      },
+    )
+  })
+}
+
+/** Strip the human chrome lightrag_search.py --explore prints around the
+ *  answer. --explore returns before the --json branch — it can NEVER emit
+ *  JSON, so this is a plain-text contract by construction. */
+export function parseExploreOutput(raw: string): string {
+  return raw
+    .split('\n')
+    .filter(line => {
+      const trimmed = line.trim()
+      if (/^Exploring entity:/i.test(trimmed)) return false
+      if (/^-{10,}$/.test(trimmed)) return false
+      if (/^\[\d+(?:\.\d+)?s \| mode: /.test(trimmed)) return false
+      if (/^(Query|Mode):/i.test(trimmed)) return false
+      return true
+    })
+    .join('\n')
+    .trim()
+}
+
+/** hop2 — LightRAG entity exploration. Spends claude -p from the shared query
+ *  pool; spawned detached because the python child holds claude grandchildren
+ *  a plain timeout kill would orphan. */
+export function lightragExploreHop(
+  entity: string,
+  onProcess?: (proc: ChildProcess) => void,
+): Promise<Hop2Result> {
+  const pythonBin = PYTHON_BIN
+  const scriptsDir = COS_SCRIPTS_DIR
+  if (!pythonBin || !scriptsDir) {
+    return Promise.resolve({ ok: false, text: null, reason: 'no_cos_pipeline', treeClosed: true })
+  }
+  const script = resolve(scriptsDir, 'lightrag_search.py')
+  if (!existsSync(script)) {
+    return Promise.resolve({ ok: false, text: null, reason: 'no_memory_scripts', treeClosed: true })
+  }
+  if (!lightragBudgetAllows()) {
+    return Promise.resolve({ ok: false, text: null, reason: 'daily_graph_cap', treeClosed: true })
+  }
+  const startedAt = Date.now()
+  return new Promise(resolveHop => {
+    let settled = false
+    let stdout = ''
+    let timedOut = false
+
+    const proc = spawn(pythonBin, [script, '--explore', entity], {
+      stdio: ['ignore', 'pipe', 'pipe'],
+      cwd: scriptsDir,
+      env: { ...process.env },
+      detached: true,
+    })
+    onProcess?.(proc)
+
+    const finish = (result: Hop2Result) => {
+      if (settled) return
+      settled = true
+      clearTimeout(timer)
+      resolveHop(result)
+    }
+
+    const timer = setTimeout(() => {
+      timedOut = true
+      void terminateProviderProcess(proc).then(termination => {
+        finish({ ok: false, text: null, reason: 'graph_timeout', treeClosed: termination.closed })
+      })
+    }, HOP2_TIMEOUT_MS)
+
+    proc.on('error', () => finish({ ok: false, text: null, reason: 'graph_spawn_failed', treeClosed: !proc.pid }))
+    proc.stdout?.on('data', chunk => {
+      stdout += String(chunk)
+      if (stdout.length > 1024 * 1024) stdout = stdout.slice(-1024 * 1024)
+    })
+
+    proc.on('close', code => {
+      if (timedOut) return
+      const durationMs = Date.now() - startedAt
+      if (code !== 0) {
+        finish({ ok: false, text: null, reason: 'graph_failed', treeClosed: true })
+        return
+      }
+      const text = parseExploreOutput(stdout)
+      if (!text || /^LightRAG query failed/im.test(stdout)) {
+        finish({ ok: false, text: null, reason: 'graph_empty', treeClosed: true })
+        return
+      }
+      logTokenAudit({
+        source: 'live-cues',
+        model: 'lightrag',
+        inputChars: entity.length,
+        outputChars: text.length,
+        durationMs,
+        caller: 'live-cues-lightrag',
+      })
+      finish({ ok: true, text: text.slice(0, 4_000), treeClosed: true })
+    })
+  })
+}

package/server/lib/live-cues-prompt.ts

@@ -0,0 +1,98 @@
+// Live Cues prompts — planner and insight, both Composer asks via the
+// dedicated spawn (live-cues-cursor.ts). The bridge's system prompt is gone,
+// so these prompts own EVERY constraint themselves, including the anti-markdown
+// rule (the old bridge prompt forbade structured output, which is why the
+// bridge cannot be used) and the attribution rule (with sparse voiceprints,
+// diarization collapses to owner/Ext and a model will invent named speakers).
+
+export interface PlannerResult {
+  query: string
+  entity: string | null
+}
+
+export interface InsightResult {
+  nudge: string
+  type: string
+  priority: number
+}
+
+const JSON_ONLY_RULES = [
+  'Output ONLY minified JSON on a single line. No markdown, no code fences, no prose before or after.',
+  'If you have nothing sharp to contribute, output the literal token null.',
+].join(' ')
+
+export function buildPlannerPrompt(transcriptWindow: string): string {
+  return [
+    'You plan memory lookups for a live-meeting coaching system.',
+    'Given the transcript window below, produce the single most valuable memory query.',
+    JSON_ONLY_RULES,
+    'Schema: {"query": string, "entity": string | null}',
+    '- "query": a semantic search phrase for past-meeting retrieval (topics, decisions, commitments). Under 12 words.',
+    '- "entity": ONE named person, company, or project from the transcript worth exploring in the knowledge graph, or null.',
+    'Output null when the window is small talk, logistics, or filler with no retrievable substance.',
+    '',
+    'TRANSCRIPT WINDOW:',
+    transcriptWindow,
+  ].join('\n')
+}
+
+export function buildInsightPrompt(input: {
+  transcriptWindow: string
+  memorySnippets: string[]
+  graphContext: string | null
+}): string {
+  const memory = input.memorySnippets.length
+    ? input.memorySnippets.map((snippet, index) => `${index + 1}. ${snippet}`).join('\n')
+    : '(no related past meetings found)'
+  return [
+    'You produce ONE live coaching cue for smart-glasses during a meeting.',
+    'The wearer sees at most 85 characters, so the cue must be a single sharp line.',
+    JSON_ONLY_RULES,
+    'Schema: {"nudge": string, "type": string, "priority": number}',
+    '- "nudge": max 85 characters, max 14 words. Plain text. No markdown, no asterisks, no em dashes.',
+    '- "type": a short snake_case category, e.g. commitment_check, past_context, open_question, risk_flag. NEVER the literal string coaching_nudge.',
+    '- "priority": 1 (glance-worthy) to 3 (say this now).',
+    'Rules:',
+    '- The cue must use the MEMORY or GRAPH context to surface an edge the room cannot see. Do not replay what was just said.',
+    '- Never attribute a statement to a named person unless the transcript line is labeled with that name. Prefer "someone raised X".',
+    '- A sharp question is a valid cue when context is thin.',
+    '- Output null unless the cue is genuinely worth interrupting a meeting for.',
+    '',
+    'TRANSCRIPT WINDOW:',
+    input.transcriptWindow,
+    '',
+    'MEMORY (related past meetings):',
+    memory,
+    '',
+    'GRAPH CONTEXT:',
+    input.graphContext ?? '(unavailable this cycle)',
+  ].join('\n')
+}
+
+/** Parse a Composer reply that was instructed to emit minified JSON or null.
+ *  Tolerates accidental code fences; anything else unparseable returns null. */
+export function parseJsonReply<T>(raw: string, validate: (value: unknown) => value is T): T | null {
+  const trimmed = raw.trim().replace(/^```(?:json)?\s*/i, '').replace(/\s*```$/, '').trim()
+  if (!trimmed || trimmed === 'null') return null
+  try {
+    const parsed: unknown = JSON.parse(trimmed)
+    return validate(parsed) ? parsed : null
+  } catch {
+    return null
+  }
+}
+
+export function isPlannerResult(value: unknown): value is PlannerResult {
+  if (!value || typeof value !== 'object') return false
+  const record = value as Record<string, unknown>
+  return typeof record.query === 'string' && record.query.trim().length > 0
+    && (record.entity === null || typeof record.entity === 'string')
+}
+
+export function isInsightResult(value: unknown): value is InsightResult {
+  if (!value || typeof value !== 'object') return false
+  const record = value as Record<string, unknown>
+  return typeof record.nudge === 'string' && record.nudge.trim().length > 0
+    && typeof record.type === 'string' && record.type.trim().length > 0
+    && typeof record.priority === 'number' && Number.isFinite(record.priority)
+}

package/server/lib/maintenance-lifecycle.ts

@@ -47,6 +47,12 @@ export type MaintenanceWorkKind =
   | 'prompt_draft_write'
   | 'prompt_draft_warm'
   | 'prompt_draft_finalize'
+  // Live Cues pipeline (planner + memory hops + insight). Held for the
+  // pipeline's in-flight duration so an Update Server drain cannot unload the
+  // server while a Cursor CLI or python child is live. The pipeline wall
+  // budget must stay under COS Control's 90s drain timeout (main.swift:1853)
+  // or every drain that catches a cue in flight hard-fails to Repair.
+  | 'live_cue_pipeline'
 
 export type MaintenanceWorkPhase = 'queued' | 'active'
 export type MaintenanceOperationScope = 'same_boot' | 'cross_boot'

package/server/lib/meeting-store.ts

@@ -409,7 +409,16 @@ export class MeetingStore {
       '',
       '## Summary',
       '',
-      '*Standalone recording — canonical transcript shown in meeting detail.*',
+      // When COS ops is configured, use the private-app pipeline markers so
+      // sync_meetings.py --g2-file will enrich (and reclassify domain). Plain
+      // "Standalone recording" summaries are treated as already-final and skipped.
+      ...(process.env.COS_SCRIPTS_DIR || process.env.COS_OPERATIONS_DIR || process.env.COS_MEETINGS_ROOT
+        ? [
+            '<!-- g2-needs-domain-review -->',
+            '',
+            '*G2 recording — summary pending pipeline processing.*',
+          ]
+        : ['*Standalone recording — canonical transcript shown in meeting detail.*']),
       '',
       '## Transcript',
       '',

package/server/routes/health.ts

@@ -36,6 +36,7 @@ import { managedRuntimeCapability, managedServerVersion } from '../lib/managed-r
 import { getServerGenerationId } from '../lib/managed-runtime.js'
 import { maintenanceLifecycle } from '../lib/maintenance-lifecycle.js'
 import { getLocalTtsHealth, refreshLocalTtsHealth } from '../lib/tts-local.js'
+import { liveCuesCapability } from '../lib/live-cues-capability.js'
 
 export const healthRouter = Router()
 
@@ -203,6 +204,9 @@ healthRouter.get('/health', async (_req, res) => {
   const recovery = managedRuntimeCapability()
   const maintenance = maintenanceLifecycle.snapshot()
   const tts_local = getLocalTtsHealth()
+  // Computed once per request; the same value feeds features.liveCues and
+  // capabilities.liveCues so the two surfaces can never disagree.
+  const liveCues = liveCuesCapability()
   const features = {
     claude: claudeAvailable,
     codex: codexAvailable,
@@ -219,6 +223,7 @@ healthRouter.get('/health', async (_req, res) => {
     durableQueryJobsProtocol: durableJobs.protocolVersion,
     localFirstMeetings: localFirstMeetings !== null,
     transcriptionPolicy: transcription.mode,
+    liveCues: liveCues.available,
   }
   const voice = {
     available: keyStatus.hasKey || tts_local.ready,
@@ -277,6 +282,7 @@ healthRouter.get('/health', async (_req, res) => {
         carriedAcrossBoot: maintenance.operation?.carriedAcrossBoot ?? false,
       },
       cliDebug: CLI_DEBUG_CAPABILITY,
+      liveCues,
       ...(localFirstMeetings ? { localFirstMeetings } : {}),
     },
     // /api/health is intentionally unauthenticated for setup diagnostics.
@@ -319,6 +325,10 @@ healthRouter.get('/models', async (req, res) => {
       transcription: { ...transcription, hq: transcriptionHq },
       cliDebug: CLI_DEBUG_CAPABILITY,
       recovery: managedRuntimeCapability(),
+      // Same helper as /api/health — the companion's 15s liveness poll reads
+      // THIS surface, so a value present only on /api/health leaves the
+      // live-cues indicator blind.
+      liveCues: liveCuesCapability(),
       ...(localFirstMeetings ? { localFirstMeetings } : {}),
     },
   })

package/server/routes/live-cues.ts

@@ -0,0 +1,58 @@
+// Live Cues routes. Authenticated by the global /api X-Cos-Token middleware
+// (index.ts:148) like every other /api route — no route-local auth needed.
+//
+// POST /live-cues/start { sessionId }  -> { armed, sessionId, pipelinesUsed }
+// POST /live-cues/stop  { sessionId }  -> { nudgesGenerated, nudges }
+// GET  /live-cues/status               -> engine snapshot + capability
+
+import { Router } from 'express'
+import { errMsg } from '../lib/utils.js'
+import {
+  armLiveCues,
+  disarmLiveCues,
+  getLiveCuesStatus,
+  LiveCuesArmError,
+} from '../lib/live-cues-engine.js'
+import { liveCuesCapability } from '../lib/live-cues-capability.js'
+
+export const liveCuesRouter = Router()
+
+liveCuesRouter.post('/live-cues/start', async (req, res) => {
+  const sessionId = typeof (req.body as { sessionId?: unknown })?.sessionId === 'string'
+    ? String((req.body as { sessionId: string }).sessionId).trim()
+    : ''
+  // A start with no sessionId is refused, never silently armed on a global
+  // counter — the per-meeting cap is only real when it is session-scoped.
+  if (!sessionId || !/^[A-Za-z0-9:_-]{3,96}$/.test(sessionId)) {
+    return res.status(400).json({ error: 'sessionId required', reason: 'missing_session_id' })
+  }
+  try {
+    const result = await armLiveCues(sessionId)
+    return res.json(result)
+  } catch (error) {
+    if (error instanceof LiveCuesArmError) {
+      return res.status(error.status).json({ error: error.message, reason: error.code })
+    }
+    return res.status(500).json({ error: errMsg(error) })
+  }
+})
+
+liveCuesRouter.post('/live-cues/stop', (req, res) => {
+  const sessionId = typeof (req.body as { sessionId?: unknown })?.sessionId === 'string'
+    ? String((req.body as { sessionId: string }).sessionId).trim()
+    : ''
+  if (!sessionId) {
+    return res.status(400).json({ error: 'sessionId required', reason: 'missing_session_id' })
+  }
+  // Shape matches the existing client stop handler: it reads nudgesGenerated
+  // for its log line and hydrates meetingNudgeHistory from nudges when SSE
+  // missed events (the 200-event replay buffer is flooded by transcript_chunk).
+  return res.json(disarmLiveCues(sessionId))
+})
+
+liveCuesRouter.get('/live-cues/status', (_req, res) => {
+  res.json({
+    capability: liveCuesCapability(),
+    sessions: getLiveCuesStatus(),
+  })
+})

package/server/routes/meeting.ts

@@ -41,6 +41,13 @@ import {
 } from './transcribe-stream.js'
 import { getServerInstanceId } from '../lib/server-instance-id.js'
 import { acquireMaintenanceWork, type MaintenanceWorkLease } from '../lib/maintenance-lifecycle.js'
+import { handoffMeetingToOperations } from '../lib/g2-ops-handoff.js'
+
+function cosOpsPipelineConfigured(): boolean {
+  // Read env live (not the module-load COS_SCRIPTS_DIR const) so unit tests that
+  // clear ops env stay standalone, and Control-updated env is visible.
+  return Boolean(process.env.COS_SCRIPTS_DIR?.trim())
+}
 
 interface MeetingSessionSource {
   getTranscript(sessionId: string): string | null
@@ -313,24 +320,48 @@ export function createMeetingRouter(deps: MeetingRouteDependencies = {}): Router
           allowDuringDrain: true,
           phase: 'queued',
         })
-        const task = Promise.resolve().then(() => {
+        const task = Promise.resolve().then(async () => {
           batchLease.setPhase('active')
-          return finalizeBatch({
-            audioDir: pendingAudioDir,
-            entries: chunkEntries.map(entry => ({ ...entry, chunk: { ...entry.chunk } })),
-            streamingWordCount: countWords(transcript),
-            meetingPath: saved.filepath,
-            sidecarPath: saved.sidecarPath,
-            runBatch,
-          })
+          try {
+            await finalizeBatch({
+              audioDir: pendingAudioDir,
+              entries: chunkEntries.map(entry => ({ ...entry, chunk: { ...entry.chunk } })),
+              streamingWordCount: countWords(transcript),
+              meetingPath: saved.filepath,
+              sidecarPath: saved.sidecarPath,
+              runBatch,
+            })
+          } catch (error) {
+            // Raw audio deliberately remains for bounded cleanup / retry.
+            console.error(
+              `[meeting/save] Batch finalization failed for ${sessionId}: `
+              + `${error instanceof Error ? error.message : String(error)}`,
+            )
+          }
+          // Always hand off the durable local scribe (streaming or batch) into
+          // operations/ when COS pipeline is configured — this is what was
+          // missing on managed public server and left today's G2 files unsynced.
+          if (cosOpsPipelineConfigured()) {
+            await handoffMeetingToOperations(saved.filepath)
+          }
         }).catch(error => {
-          // Raw audio deliberately remains for the existing two-hour cleanup.
           console.error(
-            `[meeting/save] Batch finalization failed for ${sessionId}: `
+            `[meeting/save] G2 ops handoff failed for ${sessionId}: `
             + `${error instanceof Error ? error.message : String(error)}`,
           )
         }).finally(() => batchLease.release())
         scheduleBackground(task)
+      } else if (cosOpsPipelineConfigured()) {
+        // No HQ batch (no audio / incomplete writes) — still hand off the
+        // streaming scribe into operations when COS pipeline is configured.
+        scheduleBackground(
+          handoffMeetingToOperations(saved.filepath).catch(error => {
+            console.error(
+              `[meeting/save] G2 ops handoff failed for ${sessionId}: `
+              + `${error instanceof Error ? error.message : String(error)}`,
+            )
+          }),
+        )
       }
     } catch (error) {
       if (error instanceof MeetingStoreError) {
@@ -399,7 +430,7 @@ async function finalizeBatch(options: {
   if (canDeletePendingBatchAudio(transcriptApplied, metadataPersisted)) {
     rmSync(options.audioDir, { recursive: true, force: true })
   } else {
-    console.warn('[meeting/save] Pending raw audio retained for bounded two-hour cleanup')
+    console.warn('[meeting/save] Pending raw audio retained for bounded cleanup')
   }
 }
 

package/server/routes/transcribe-stream.ts

@@ -46,8 +46,10 @@ import { getServerInstanceId } from '../lib/server-instance-id.js'
 import {
   acquireMaintenanceWork,
   maintenanceAdmissionsOpen,
+  maintenanceLifecycle,
   type MaintenanceWorkLease,
 } from '../lib/maintenance-lifecycle.js'
+import { feedLiveCueTranscript } from '../lib/live-cues-engine.js'
 
 function ensurePrivateDirectory(path: string): void {
   if (!existsSync(path)) mkdirSync(path, { recursive: true, mode: 0o700 })
@@ -636,11 +638,15 @@ setInterval(() => {
       }
     }
   } catch {}
-  // Purge stale pending-batch dirs older than 2 hours after move (batch should complete in minutes).
+  // Purge stale pending-batch dirs. HQ large-v3 on long meetings + Control
+  // restart can exceed 2h (2026-07-27: two sessions purged before batch).
+  // Use 12h, and never purge while a meeting_batch_finalization lease is held.
   // Age measured by _batch_pending.marker mtime (set by moveSessionAudioToPending), NOT the first
   // chunk's mtime — chunk files keep their original write-time across the atomic rename, so for
   // meetings > 1 hour, chunk mtimes would always look stale. Fallback: dir mtime for older marker-less dirs.
   try {
+    const batchBusy = ((maintenanceLifecycle.snapshot().activeByKind as Record<string, number>)
+      .meeting_batch_finalization ?? 0) > 0
     for (const dir of readdirSync(PENDING_BATCH_DIR)) {
       const dirPath = resolve(PENDING_BATCH_DIR, dir)
       try {
@@ -655,7 +661,11 @@ setInterval(() => {
           // Fallback for pre-v5.4.3 dirs: use directory ctime (changes on rename)
           ageSource = statSync(dirPath).ctimeMs
         }
-        if (Date.now() - ageSource > 2 * 60 * 60 * 1000) {
+        if (Date.now() - ageSource > 12 * 60 * 60 * 1000) {
+          if (batchBusy) {
+            console.warn(`[cleanup] Retaining stale pending-batch while batch lease held: ${dir}`)
+            continue
+          }
           rmSync(dirPath, { recursive: true, force: true })
           console.log(`[cleanup] Purged stale pending-batch: ${dir}`)
         }
@@ -1538,6 +1548,11 @@ async function processStreamChunk(opts: {
   console.log(`[perf] persistSession: ${(performance.now() - tPersist).toFixed(1)}ms (${session.chunks.filter(c => c).length} chunks)`)
 
   emitDisplay({ type: 'transcript_chunk', data: { text: trimmedText, speaker, chunkIndex, elapsed, sessionId } })
+  // Live Cues feed — fire-and-forget, NEVER awaited: transcription must not
+  // block on a cue, and no LLM runs on this path. .catch() is required — a
+  // bare `void` on a rejecting promise is an unhandled rejection, which Node
+  // throws on by default. All gates live inside the feed.
+  void feedLiveCueTranscript(sessionId, trimmedText, Boolean(fallbackReason)).catch(() => {})
 
   console.log(`[perf] TOTAL request: ${(performance.now() - tReq).toFixed(1)}ms | chunk #${chunkIndex} | ${audioBuffer.length}b | rms=${Math.round(rms)} q=${isQuiet ? 1 : 0} | ${asrProvider} | "${trimmedText.slice(0, 50)}"`)
   return canonicalChunkResponse(chunk, sessionId, chunkIndex)