@gotcos/glasses-server 6.38.0 → 6.39.0

package/server/routes/health.ts

@@ -32,6 +32,11 @@ import {
   getCursorModelCatalogSnapshot,
   isCursorProviderReady,
 } from '../lib/cursor-model-catalog.js'
+import {
+  getOllamaCatalog,
+  getOllamaCatalogSnapshot,
+  isOllamaProviderReady,
+} from '../lib/ollama-catalog.js'
 import { isMediaProcessingReady } from '../lib/image-safety.js'
 import {
   MAX_OTHER_MEDIA_BYTES,
@@ -117,6 +122,7 @@ healthRouter.get('/health', async (_req, res) => {
     claude: staticProbes.claude,
     codex: staticProbes.codex,
     cursor: staticProbes.cursor,
+    ollama: staticProbes.ollama,
     uptime_seconds: Math.floor((Date.now() - serverMetrics.startedAt) / 1000),
     request_count: serverMetrics.requestCount,
   }
@@ -125,6 +131,7 @@ healthRouter.get('/health', async (_req, res) => {
   const claudeAvailable = staticProbes.claudeAvailable
   const codexAvailable = staticProbes.codexAvailable
   const cursorAvailable = staticProbes.cursorAvailable
+  const ollamaAvailable = staticProbes.ollamaAvailable
 
   // Check session cache freshness (COS mode only)
   if (COS_SCRIPTS_DIR) {
@@ -195,6 +202,7 @@ healthRouter.get('/health', async (_req, res) => {
     claude: claudeAvailable,
     codex: codexAvailable,
     cursor: cursorAvailable,
+    ollama: ollamaAvailable,
     voice: keyStatus.hasKey || tts_local.ready,
     cos_pipeline: COS_MODE,
     whisper: isWhisperLocalAvailable(),
@@ -267,6 +275,7 @@ healthRouter.get('/health', async (_req, res) => {
   // agent binary paths stay on the authenticated /api/models surface.
   const cursorSnapshot = getCursorModelCatalogSnapshot()
   const { agentBinary: _cursorAgentBinary, ...cursor_models } = cursorSnapshot
+  const ollama_models = getOllamaCatalogSnapshot()
   const meeting_sync = getMeetingSyncSnapshot()
   const progressiveHq = getProgressiveHqSnapshot()
   // Quarantined unsaved captures (6.19.0). Compact on this unauthenticated
@@ -327,6 +336,10 @@ healthRouter.get('/health', async (_req, res) => {
     tts_local,
     codex_models,
     cursor_models,
+    ollama_models: {
+      ready: ollama_models.ready,
+      model: ollama_models.model,
+    },
     meeting_sync,
     meeting_library: {
       layout: meetingLibrary.layout,
@@ -397,6 +410,7 @@ healthRouter.get('/models', async (req, res) => {
   const forceRefresh = req.query.refresh === '1'
   const catalog = await getCodexModelCatalog(forceRefresh)
   const cursorCatalog = await getCursorModelCatalog(forceRefresh)
+  const ollamaCatalog = await getOllamaCatalog(forceRefresh)
   const durableJobs = durableQueryJobStatus()
   const localFirstMeetings = localFirstMeetingsCapability(getServerInstanceId())
   const transcription = getTranscriptionPolicySnapshot()
@@ -407,6 +421,9 @@ healthRouter.get('/models', async (req, res) => {
   const richMedia = await getRichMediaProcessingCapabilities()
   const videoUploadV2 = videoUploadV2Capability(richMedia.video)
   const cursorOptions = cursorCatalog.options.filter(option => !!option.id)
+  const ollamaOptions = isOllamaProviderReady() && ollamaCatalog.model
+    ? [{ preference: 'ollama' as const, id: ollamaCatalog.model, displayName: ollamaCatalog.model }]
+    : []
   // Same helper and the same three key names as /api/health, for the reason
   // liveCues carries three lines below: the companion's 15s liveness poll reads
   // THIS surface and Main.ts states outright that /api/health alone is not used,
@@ -418,9 +435,17 @@ healthRouter.get('/models', async (req, res) => {
     options: [
       ...(catalog.options ?? []),
       ...cursorOptions,
+      ...ollamaOptions,
     ],
     cursor: cursorCatalog,
     cursorReady: isCursorProviderReady(),
+    ollama: {
+      origin: ollamaCatalog.origin,
+      model: ollamaCatalog.model,
+      models: ollamaCatalog.models,
+      refreshedAt: ollamaCatalog.refreshedAt,
+    },
+    ollamaReady: isOllamaProviderReady(),
     serverInstanceId: getServerInstanceId(),
     capabilities: {
       durableQueryJobs: {

package/server/routes/openai-compat.ts

@@ -114,6 +114,7 @@ export function resolveModel(model?: string, _query?: string): ModelPreference {
   if (model === 'cos-haiku') return 'haiku'
   if (model === 'cos-gpt-frontier' || model === 'cos-codex-high' || model === 'cos-codex') return 'codex-frontier'
   if (model === 'cos-gpt-balanced') return 'codex-balanced'
+  if (model === 'cos-ollama' || model === 'ollama') return 'ollama'
   return normalizeModelPreference(process.env.COS_G2_DEFAULT_MODEL) ?? DEFAULT_MODEL
 }
 
@@ -126,6 +127,7 @@ const MODEL_NAMES: Record<ModelPreference, string> = {
   'codex-balanced': 'cos-gpt-balanced',
   'cursor-grok': 'cursor-grok',
   'cursor-composer': 'cursor-composer',
+  ollama: 'cos-ollama',
 }
 // Extract the user's latest message from the OpenAI messages array
 function extractUserQuery(messages: Array<{ role: string; content: string }>): string {

package/server/routes/transcribe-stream.ts

@@ -61,6 +61,14 @@ import {
   appendChunkEmbedding,
   sweepExpiredChunkEmbeddings,
 } from '../lib/chunk-embedding-store.js'
+import {
+  evenSpeakerRoleMode,
+  formatEvenRoleAgreement,
+  parseEvenHubSpeakerRoleBody,
+  parseEvenHubSpeakerRoleQuery,
+  warnEvenSpeakerRoleApplyNotImplemented,
+  type EvenSpeakerRoleHistogram,
+} from '../lib/even-hub-speaker-role.js'
 import { archiveSessionAudio, runMeetingAudioRetention } from '../lib/meeting-audio-archive.js'
 import {
   countChunkWavs,
@@ -276,6 +284,7 @@ export interface TranscriptChunk {
   latencyMs?: number
   audioSha256?: string
   canonical?: boolean
+  evenHubSpeakerRole?: EvenSpeakerRoleHistogram
 }
 
 export interface ProviderCandidateRecord {
@@ -1945,8 +1954,11 @@ async function processStreamChunk(opts: {
   clientElapsed?: number
   /** Original client recording start, applied only before canonical chunks. */
   startTimeOverride?: number
+  evenHubSpeakerRole?: EvenSpeakerRoleHistogram
 }): Promise<StreamChunkCompletionResponse> {
   const { sessionId, chunkIndex, clientSpeaker, audioBuffer, candidate } = opts
+  const evenHubSpeakerRole = evenSpeakerRoleMode() === 'off' ? undefined : opts.evenHubSpeakerRole
+  if (evenHubSpeakerRole) warnEvenSpeakerRoleApplyNotImplemented()
   const tReq = performance.now()
   validateSessionId(sessionId)
   validateChunkIndex(chunkIndex)
@@ -2066,6 +2078,15 @@ async function processStreamChunk(opts: {
   }
 
   const { speaker, similarity } = await speakerPromise
+  if (evenHubSpeakerRole) {
+    console.log(formatEvenRoleAgreement({
+      chunkIndex,
+      even: evenHubSpeakerRole,
+      amp: clientSpeaker,
+      emb: speaker,
+      similarity,
+    }))
+  }
   // Client time is authoritative for live network jitter and deferred replay.
   const elapsed = Number.isFinite(opts.clientElapsed) && (opts.clientElapsed as number) >= 0
     ? Math.round(opts.clientElapsed as number)
@@ -2110,6 +2131,7 @@ async function processStreamChunk(opts: {
     latencyMs,
     audioSha256,
     canonical: true,
+    evenHubSpeakerRole,
   }
   const finalExisting = session.chunks[chunkIndex]
   if (finalExisting?.text) {
@@ -2263,6 +2285,7 @@ transcribeStreamRouter.post('/transcribe-stream', async (req, res) => {
       audioBuffer,
       clientElapsed,
       startTimeOverride,
+      evenHubSpeakerRole: parseEvenHubSpeakerRoleQuery(req.query.eh),
     }))
   } catch (err: unknown) {
     sendStreamError(res, err)
@@ -2327,6 +2350,7 @@ transcribeStreamRouter.post('/transcribe-stream/offline-sessions/:sessionId/chun
       clientSpeaker: String(body.clientSpeaker ?? 'Unknown'),
       audioBuffer,
       clientElapsed: typeof body.elapsed === 'number' ? body.elapsed : Number(body.elapsed ?? 0),
+      evenHubSpeakerRole: parseEvenHubSpeakerRoleBody(body.evenHubSpeakerRole),
       candidate: {
         provider: 'iphone-whisperkit-beta',
         text: normalizeCandidateText(candidate.text),
@@ -2405,6 +2429,7 @@ transcribeStreamRouter.post('/transcribe-stream/candidates', async (req, res) =>
       clientSpeaker: String(body.clientSpeaker ?? 'Unknown'),
       audioBuffer,
       clientElapsed: typeof body.elapsed === 'number' ? body.elapsed : Number(body.elapsed ?? 0),
+      evenHubSpeakerRole: parseEvenHubSpeakerRoleBody(body.evenHubSpeakerRole),
       candidate: {
         provider: 'iphone-whisperkit-beta',
         text: normalizeCandidateText(candidate.text),

package/server/scripts/repair-archive-duplicates.ts

@@ -0,0 +1,181 @@
+#!/usr/bin/env tsx
+// Repair day files written by the pre-upsert archive merge.
+//
+//   npx tsx server/scripts/repair-archive-duplicates.ts            # dry run, writes nothing
+//   npx tsx server/scripts/repair-archive-duplicates.ts --apply    # rewrites, after backing up
+//
+// WHAT WENT WRONG. `runDailyArchiveMirror` re-archives every session still
+// resident in memory, skipping only today's, at boot and every 24h -- without
+// evicting it. `appendToArchive` merged with a blind `existing.chats.push(...)`.
+// So a session that stayed resident gained one more copy of itself in its day
+// file on every restart. Measured before the fix: 1.28 GB across 176 day files,
+// ~1.26 GB of it duplicates. One 69 MB file held ONE conversation 2,388 times.
+//
+// The upsert in archive.ts fixes new writes AND self-heals a file the next time
+// it is touched -- so most affected days repair themselves once the mirror
+// revisits them. This script exists for the remainder: days whose sessions have
+// since been evicted, which nothing will ever touch again.
+//
+// THIS SCRIPT IMPORTS NOTHING FROM THE SERVER. `archive.ts` runs
+// checkYesterdayArchive() at module scope, so importing it would start archive
+// work while we are rewriting the archive. The atomic write below is inlined for
+// the same reason. Nothing here has an effect until --apply.
+
+import { execFileSync } from 'node:child_process'
+import { copyFileSync, existsSync, readFileSync, readdirSync, renameSync, statSync, writeFileSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { join, resolve } from 'node:path'
+import { pathToFileURL } from 'node:url'
+
+const APPLY = process.argv.includes('--apply')
+const DATE_RE = /^\d{4}-\d{2}-\d{2}$/
+
+interface Chat {
+  id: number
+  sessionId: string
+  startedAt: number
+  exchangeCount: number
+  [k: string]: unknown
+}
+interface Day { date: string; summary: string; chats: Chat[]; archivedAt: string; [k: string]: unknown }
+
+function archiveDirPath(): string {
+  const base = process.env.COS_DATA_DIR ?? join(homedir(), '.cos-glasses', 'data')
+  return resolve(base, 'archive')
+}
+
+/** Identity of a chat. `startedAt` is its first exchange's timestamp, so it
+ *  survives re-archiving. `id` does NOT -- it is renumbered on every merge,
+ *  which is exactly why the old code could never see a duplicate. */
+const keyOf = (c: Chat): string => `${c.sessionId}:${c.startedAt}`
+
+/** Collapse duplicates, keeping the most complete copy of each chat. Pure. */
+export function dedupeChats(chats: Chat[]): { kept: Chat[]; removed: number } {
+  const byKey = new Map<string, Chat>()
+  for (const chat of chats) {
+    const prior = byKey.get(keyOf(chat))
+    if (!prior || (chat.exchangeCount ?? 0) > (prior.exchangeCount ?? 0)) byKey.set(keyOf(chat), chat)
+  }
+  const kept = [...byKey.values()].sort((a, b) => a.startedAt - b.startedAt)
+  kept.forEach((c, i) => { c.id = i })
+  return { kept, removed: chats.length - kept.length }
+}
+
+function atomicWrite(path: string, data: string): void {
+  // Inlined rather than imported: see the header note about module-scope effects.
+  const tmp = `${path}.repair-tmp`
+  writeFileSync(tmp, data, { encoding: 'utf8', mode: 0o600 })
+  renameSync(tmp, path)
+}
+
+function main(): void {
+  const dir = archiveDirPath()
+  if (!existsSync(dir)) {
+    console.error(`No archive directory at ${dir}`)
+    process.exit(2)
+  }
+
+  // A concurrent appendToArchive would race this rewrite. The in-process archive
+  // lock cannot be taken from outside the server, so the only safe answer is to
+  // refuse while it is up rather than to hope the window is small.
+  // Only the DEFAULT data dir is at risk: that is the one the running server writes
+  // to. Pointed at a scratch copy, there is nothing to race, and refusing there
+  // would block the very rehearsal this script deserves before it touches real data.
+  const isLiveDataDir = process.env.COS_DATA_DIR === undefined
+  if (APPLY && isLiveDataDir && serverIsUp()) {
+    console.error('The COS server is listening on 127.0.0.1:3141.')
+    console.error('Stop it through COS Control before repairing, then re-run.')
+    console.error('Refusing to rewrite archive files while the server may write to them.')
+    process.exit(3)
+  }
+
+  const files = readdirSync(dir)
+    .filter(f => f.endsWith('.json') && DATE_RE.test(f.slice(0, -5)))
+    .sort()
+
+  let affected = 0
+  let chatsBefore = 0
+  let chatsAfter = 0
+  let bytesBefore = 0
+  let bytesAfter = 0
+
+  for (const file of files) {
+    const path = join(dir, file)
+    const size = statSync(path).size
+
+    let day: Day
+    try {
+      day = JSON.parse(readFileSync(path, 'utf8')) as Day
+    } catch (err) {
+      console.error(`  SKIP ${file} — unreadable: ${(err as Error).message.slice(0, 80)}`)
+      continue
+    }
+    if (!Array.isArray(day.chats) || day.chats.length === 0) continue
+
+    const { kept, removed } = dedupeChats(day.chats)
+    if (removed === 0) continue
+
+    affected++
+    chatsBefore += day.chats.length
+    chatsAfter += kept.length
+    bytesBefore += size
+
+    const before = day.chats.length
+    day.chats = kept
+    const serialised = `${JSON.stringify(day, null, 2)}\n`
+    bytesAfter += Buffer.byteLength(serialised, 'utf8')
+
+    console.log(
+      `  ${APPLY ? 'REPAIR' : 'would repair'} ${file}  ` +
+      `${(size / 1e6).toFixed(1)} MB → ${(Buffer.byteLength(serialised, 'utf8') / 1e6).toFixed(1)} MB  ` +
+      `chats ${before} → ${kept.length}  (-${removed})`,
+    )
+
+    if (APPLY) {
+      // Back up BEFORE writing. This is user conversation history; a bad rewrite
+      // with no copy is unrecoverable.
+      const backup = `${path}.bak-${Date.now()}`
+      copyFileSync(path, backup)
+      atomicWrite(path, serialised)
+
+      // Verify by UNIQUE CHAT COUNT, never by file size -- size is the metric the
+      // bug distorted, so shrinkage proves nothing about correctness.
+      const reread = JSON.parse(readFileSync(path, 'utf8')) as Day
+      const uniq = new Set(reread.chats.map(keyOf)).size
+      if (reread.chats.length !== kept.length || uniq !== kept.length) {
+        console.error(`  FAILED verification on ${file}; original preserved at ${backup}`)
+        process.exit(4)
+      }
+    }
+  }
+
+  const summary = {
+    mode: APPLY ? 'applied' : 'dry-run',
+    filesScanned: files.length,
+    filesAffected: affected,
+    chats: { before: chatsBefore, after: chatsAfter, removed: chatsBefore - chatsAfter },
+    bytes: { before: bytesBefore, after: bytesAfter, reclaimed: bytesBefore - bytesAfter },
+  }
+  console.log('')
+  console.log(JSON.stringify(summary, null, 2))
+  if (!APPLY && affected > 0) {
+    console.log('')
+    console.log('Nothing was written. Re-run with --apply to repair (each file is backed up first).')
+  }
+}
+
+function serverIsUp(): boolean {
+  try {
+    const out = execFileSync('/usr/sbin/lsof', ['-ti', ':3141'], { encoding: 'utf8', timeout: 5000 })
+    return out.trim().length > 0
+  } catch {
+    return false // lsof missing or nothing listening — do not block on an inconclusive probe
+  }
+}
+
+// Run ONLY when invoked directly. Importing this file (a test, or any tooling)
+// must not execute a repair or call process.exit -- the same module-scope hazard
+// this script refuses to inherit from archive.ts.
+const invokedDirectly = process.argv[1] !== undefined
+  && import.meta.url === pathToFileURL(process.argv[1]).href
+if (invokedDirectly) main()

package/shared/model-preference.ts

@@ -8,9 +8,11 @@ export type CodexModelPreference = 'codex-frontier' | 'codex-balanced'
 // even when features.cursor is false so version skew fail-closes instead of
 // silently remapping to Claude.
 export type CursorModelPreference = 'cursor-grok' | 'cursor-composer'
+/** Local Ollama chat slot. Hidden unless the daemon answers on loopback. */
+export type OllamaModelPreference = 'ollama'
 /** Cursor Agent CLI execution posture for glasses queries. */
 export type CursorExecutionMode = 'ask' | 'agent'
-export type ModelPreference = ClaudeModelPreference | CodexModelPreference | CursorModelPreference
+export type ModelPreference = ClaudeModelPreference | CodexModelPreference | CursorModelPreference | OllamaModelPreference
 
 /** Invalid/omitted → ask (safe for old clients that don't send a mode). */
 export function normalizeCursorExecutionMode(value: unknown): CursorExecutionMode {
@@ -25,6 +27,7 @@ export const CODEX_BALANCED_MODEL: CodexModelPreference = 'codex-balanced'
 export const CODEX_HIGH_MODEL: CodexModelPreference = CODEX_FRONTIER_MODEL
 export const CURSOR_GROK_MODEL: CursorModelPreference = 'cursor-grok'
 export const CURSOR_COMPOSER_MODEL: CursorModelPreference = 'cursor-composer'
+export const OLLAMA_MODEL: OllamaModelPreference = 'ollama'
 // Existing 6.1–6.3 installs may pin the legacy codex-high slot. Frontier is its
 // migration target; Balanced remains auto-catalog even when this override is set.
 export const CODEX_MODEL_ID = process.env.COS_CODEX_MODEL?.trim() ?? ''
@@ -49,6 +52,7 @@ export const MODEL_OPTIONS: ModelPreference[] = [
   CODEX_BALANCED_MODEL,
   CURSOR_GROK_MODEL,
   CURSOR_COMPOSER_MODEL,
+  OLLAMA_MODEL,
 ]
 
 const MODEL_SET = new Set<ModelPreference>([
@@ -60,6 +64,7 @@ const MODEL_SET = new Set<ModelPreference>([
   CODEX_BALANCED_MODEL,
   CURSOR_GROK_MODEL,
   CURSOR_COMPOSER_MODEL,
+  OLLAMA_MODEL,
 ])
 
 // Bare Claude tier aliases resolve to the newest model in that tier at spawn.
@@ -149,6 +154,21 @@ export function isCursorModel(model: ModelPreference): model is CursorModelPrefe
   return model === CURSOR_GROK_MODEL || model === CURSOR_COMPOSER_MODEL
 }
 
+export function isOllamaModel(model: ModelPreference): model is OllamaModelPreference {
+  return model === OLLAMA_MODEL
+}
+
+/** Picker families. Cursor and Ollama stay hidden until their local probe is ready. */
+export function visibleModelOptions(
+  cursorAvailable: boolean,
+  ollamaAvailable: boolean,
+): ModelPreference[] {
+  return MODEL_OPTIONS.filter(model =>
+    (!isCursorModel(model) || cursorAvailable) &&
+    (!isOllamaModel(model) || ollamaAvailable),
+  )
+}
+
 export interface RuntimeCodexModelLabel {
   preference: CodexModelPreference
   displayName: string
@@ -197,6 +217,7 @@ export function modelLabel(model: ModelPreference): string {
     case 'codex-balanced': return runtimeCodexLabels[model] ?? 'GPT Balanced'
     case 'cursor-grok': return runtimeCursorLabels[model] ?? 'Grok Fast'
     case 'cursor-composer': return runtimeCursorLabels[model] ?? 'Composer 2.5 Fast'
+    case 'ollama': return 'Ollama'
     case 'opus':
     default:
       return 'Opus'
@@ -212,6 +233,7 @@ export function modelShortLabel(model: ModelPreference): string {
     case 'codex-balanced': return 'GPT Bal'
     case 'cursor-grok': return 'Grok'
     case 'cursor-composer': return 'Composer'
+    case 'ollama': return 'Ollama'
     case 'opus':
     default:
       return 'Opus'
@@ -227,6 +249,7 @@ export function modelButtonLabel(model: ModelPreference): string {
     case 'codex-balanced': return 'GPT BAL'
     case 'cursor-grok': return 'GROK'
     case 'cursor-composer': return 'CMP'
+    case 'ollama': return 'OLLAMA'
     case 'opus':
     default:
       return 'OPUS'
@@ -242,6 +265,7 @@ export function modelTag(model: ModelPreference): string {
     case 'codex-balanced': return 'GB'
     case 'cursor-grok': return 'GK'
     case 'cursor-composer': return 'C2'
+    case 'ollama': return 'OL'
     case 'opus':
     default:
       return 'O'