@gotcos/glasses-server 6.38.1 → 6.39.1

package/server/lib/ollama-run-ledger.ts

@@ -0,0 +1,178 @@
+import crypto from 'node:crypto'
+import { appendFileSync, existsSync, mkdirSync, readFileSync } from 'node:fs'
+import { dirname, resolve } from 'node:path'
+import { OLLAMA_MODEL, type OllamaModelPreference } from '../../shared/model-preference.js'
+
+const DEFAULT_MAX_RUNS = 100
+const DEFAULT_TTL_MS = 7 * 24 * 60 * 60_000
+const ERROR_PREVIEW_CHARS = 160
+const RUNNING_STALE_MS = 30 * 60_000
+
+export type OllamaRunStatus = 'running' | 'completed' | 'failed' | 'cancelled' | 'client_disconnected'
+
+export interface OllamaRunRecord {
+  runId: string
+  turnId?: string
+  clientJobId?: string
+  cosSessionId: string
+  status: OllamaRunStatus
+  createdAt: string
+  updatedAt: string
+  model: OllamaModelPreference
+  ollamaModel: string
+  origin: string
+  queryPreview?: string
+  outputPreview?: string
+  errorCode?: string
+  errorPreview?: string
+  durationMs?: number
+}
+
+interface OllamaRunEvent {
+  runId: string
+  ts: string
+  patch: Partial<OllamaRunRecord>
+}
+
+function getProcessStartedAtMs(): number {
+  return Date.now() - Math.floor(process.uptime() * 1000)
+}
+
+export function getOllamaLedgerPath(): string {
+  return resolve(process.env.COS_OLLAMA_RUN_LEDGER_FILE || resolve(import.meta.dirname, '..', 'data', 'ollama-runs.jsonl'))
+}
+
+function getMaxRuns(): number {
+  const raw = Number(process.env.COS_OLLAMA_RUN_LEDGER_MAX ?? DEFAULT_MAX_RUNS)
+  return Number.isFinite(raw) && raw > 0 ? Math.floor(raw) : DEFAULT_MAX_RUNS
+}
+
+function getTtlMs(): number {
+  const rawDays = Number(process.env.COS_OLLAMA_RUN_LEDGER_TTL_DAYS ?? 7)
+  return Number.isFinite(rawDays) && rawDays > 0 ? rawDays * 24 * 60 * 60_000 : DEFAULT_TTL_MS
+}
+
+function appendEvent(event: OllamaRunEvent): void {
+  try {
+    const path = getOllamaLedgerPath()
+    mkdirSync(dirname(path), { recursive: true })
+    appendFileSync(path, JSON.stringify(event) + '\n')
+  } catch (err) {
+    console.warn('[ollama-run-ledger] write skipped:', err)
+  }
+}
+
+function readEvents(): OllamaRunEvent[] {
+  const path = getOllamaLedgerPath()
+  if (!existsSync(path)) return []
+  try {
+    const events: OllamaRunEvent[] = []
+    for (const line of readFileSync(path, 'utf-8').split('\n').map(row => row.trim()).filter(Boolean)) {
+      try {
+        const event = JSON.parse(line) as OllamaRunEvent
+        if (typeof event.runId === 'string' && typeof event.ts === 'string' && typeof event.patch === 'object') {
+          events.push(event)
+        }
+      } catch {
+        // Skip torn JSONL rows.
+      }
+    }
+    return events
+  } catch {
+    return []
+  }
+}
+
+export function redactForOllamaLedger(value: string, maxChars = ERROR_PREVIEW_CHARS): string {
+  return value.replace(/\s+/g, ' ').trim().slice(0, maxChars)
+}
+
+export function classifyOllamaError(message: string): string {
+  const text = message.toLowerCase()
+  if (/unreachable|econnrefused|fetch failed|enotfound/.test(text)) return 'ollama.unavailable'
+  if (/no models|not ready/.test(text)) return 'ollama.no_model'
+  if (/text-only|photo|image/.test(text)) return 'ollama.text_only'
+  if (/timeout|timed out|aborted/.test(text)) return 'ollama.timeout'
+  return 'ollama.error'
+}
+
+export function startOllamaRun(input: {
+  turnId?: string
+  clientJobId?: string
+  cosSessionId: string
+  ollamaModel: string
+  origin: string
+  query: string
+}): OllamaRunRecord {
+  const now = new Date().toISOString()
+  const run: OllamaRunRecord = {
+    runId: `ollama-${crypto.randomUUID().slice(0, 8)}`,
+    turnId: input.turnId,
+    clientJobId: input.clientJobId,
+    cosSessionId: input.cosSessionId,
+    status: 'running',
+    createdAt: now,
+    updatedAt: now,
+    model: OLLAMA_MODEL,
+    ollamaModel: input.ollamaModel,
+    origin: input.origin,
+    queryPreview: redactForOllamaLedger(input.query),
+  }
+  appendEvent({ runId: run.runId, ts: now, patch: run })
+  return run
+}
+
+export function finishOllamaRun(runId: string, input: {
+  status: Exclude<OllamaRunStatus, 'running'>
+  startedAtMs: number
+  output?: string
+  error?: string
+}): OllamaRunRecord | null {
+  const patch: Partial<OllamaRunRecord> = {
+    status: input.status,
+    durationMs: Math.max(0, Date.now() - input.startedAtMs),
+  }
+  if (input.output) patch.outputPreview = redactForOllamaLedger(input.output)
+  if (input.error) {
+    patch.errorCode = classifyOllamaError(input.error)
+    patch.errorPreview = redactForOllamaLedger(input.error)
+  }
+  const ts = new Date().toISOString()
+  appendEvent({ runId, ts, patch: { ...patch, updatedAt: ts } })
+  return getOllamaRun(runId)
+}
+
+export function getOllamaRun(runId: string): OllamaRunRecord | null {
+  return listOllamaRuns(getMaxRuns()).find(run => run.runId === runId) ?? null
+}
+
+export function listOllamaRuns(limit = 20, sessionId?: string): OllamaRunRecord[] {
+  const runs = new Map<string, OllamaRunRecord>()
+  const order = new Map<string, number>()
+  let eventIndex = 0
+  for (const event of readEvents()) {
+    eventIndex += 1
+    const existing = runs.get(event.runId)
+    const next = { ...(existing ?? {}), ...event.patch, runId: event.runId } as OllamaRunRecord
+    runs.set(event.runId, next)
+    order.set(event.runId, eventIndex)
+  }
+  const cutoff = Date.now() - getTtlMs()
+  return Array.from(runs.values())
+    .filter(run => run.createdAt && Date.parse(run.updatedAt || run.createdAt) >= cutoff)
+    .filter(run => !sessionId || run.cosSessionId === sessionId)
+    .map(run => {
+      const updatedMs = Date.parse(run.updatedAt || run.createdAt)
+      const predatesCurrentProcess = updatedMs < getProcessStartedAtMs() - 1000
+      if (run.status === 'running' && (predatesCurrentProcess || Date.now() - updatedMs > RUNNING_STALE_MS)) {
+        return {
+          ...run,
+          status: 'client_disconnected' as const,
+          errorCode: run.errorCode ?? 'ollama.timeout',
+        }
+      }
+      return run
+    })
+    .sort((a, b) => Date.parse(b.createdAt) - Date.parse(a.createdAt))
+    .slice(0, limit)
+}

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

@@ -13,6 +13,7 @@ import { QueryJobStore } from './query-job-store.js'
 import {
   isCodexModel,
   isCursorModel,
+  isOllamaModel,
   normalizeEffortPreference,
   normalizeModelPreference,
   type CursorExecutionMode,
@@ -101,7 +102,8 @@ export async function preparePublicDurableQueryAdmission(raw: unknown): Promise<
   }
 }
 
-function providerFor(model: ModelPreference): 'claude' | 'codex' | 'cursor' {
+function providerFor(model: ModelPreference): 'claude' | 'codex' | 'cursor' | 'ollama' {
+  if (isOllamaModel(model)) return 'ollama'
   if (isCursorModel(model)) return 'cursor'
   return isCodexModel(model) ? 'codex' : 'claude'
 }
@@ -183,6 +185,7 @@ const runner: QueryJobRunner = async ({ jobId, turnId, request, signal, callback
           codexRunId: metadata?.codexRunId,
           codexThreadId: metadata?.codexThreadId,
           cursorRunId: metadata?.cursorRunId,
+          ollamaRunId: metadata?.ollamaRunId,
         } as const
         await callbacks.onStart({ sessionId, ...linkage })
         emitDisplay({ type: 'start', data: {
@@ -205,7 +208,9 @@ const runner: QueryJobRunner = async ({ jobId, turnId, request, signal, callback
           ? { claudeRunId: metadata.runId }
           : metadata.provider === 'cursor'
             ? { cursorRunId: metadata.runId }
-            : { codexRunId: metadata.runId }),
+            : metadata.provider === 'ollama'
+              ? { ollamaRunId: metadata.runId }
+              : { codexRunId: metadata.runId }),
       }),
       onChunk: text => { callbacks.onChunk(text) },
       onToolStatus: toolName => {
@@ -239,6 +244,7 @@ const runner: QueryJobRunner = async ({ jobId, turnId, request, signal, callback
           codexRunId: metadata?.codexRunId,
           codexThreadId: metadata?.codexThreadId,
           cursorRunId: metadata?.cursorRunId,
+          ollamaRunId: metadata?.ollamaRunId,
         } as const
         // Publish compatibility completion only after the durable terminal is
         // fsynced. Display subscribers can disappear without owning this job.

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

@@ -550,11 +550,11 @@ export class QueryJobStore {
   }
 
   private applyLinkage(snapshot: QueryJobSnapshot, raw: Record<string, unknown>): void {
-    const provider = raw.provider === 'claude' || raw.provider === 'codex' || raw.provider === 'cursor'
+    const provider = raw.provider === 'claude' || raw.provider === 'codex' || raw.provider === 'cursor' || raw.provider === 'ollama'
       ? raw.provider
       : undefined
     if (provider) snapshot.provider = provider
-    const fields = ['resolvedModel', 'cliSessionId', 'claudeRunId', 'codexRunId', 'codexThreadId', 'cursorRunId'] as const
+    const fields = ['resolvedModel', 'cliSessionId', 'claudeRunId', 'codexRunId', 'codexThreadId', 'cursorRunId', 'ollamaRunId'] as const
     for (const field of fields) {
       const value = safeOptional(raw[field])
       if (value) snapshot[field] = value
@@ -886,7 +886,11 @@ export class QueryJobStore {
 
   private safeLinkage(linkage: QueryJobProviderLinkage): QueryJobProviderLinkage {
     return {
-      ...(linkage.provider === 'claude' || linkage.provider === 'codex' || linkage.provider === 'cursor'
+      // 'ollama' joined this allowlist in 6.39.1. The 6.39.0 types allowed it and
+      // query-job-runtime stamped it, but this sanitizer silently dropped it on
+      // persist -- so phone acknowledgement of a live or replayed Ollama job could
+      // never see its provider. Unknown providers still strip.
+      ...(linkage.provider === 'claude' || linkage.provider === 'codex' || linkage.provider === 'cursor' || linkage.provider === 'ollama'
         ? { provider: linkage.provider }
         : {}),
       ...(safeOptional(linkage.resolvedModel, 64) ? { resolvedModel: safeOptional(linkage.resolvedModel, 64) } : {}),
@@ -895,6 +899,7 @@ export class QueryJobStore {
       ...(safeOptional(linkage.codexRunId) ? { codexRunId: safeOptional(linkage.codexRunId) } : {}),
       ...(safeOptional(linkage.codexThreadId) ? { codexThreadId: safeOptional(linkage.codexThreadId) } : {}),
       ...(safeOptional(linkage.cursorRunId) ? { cursorRunId: safeOptional(linkage.cursorRunId) } : {}),
+      ...(safeOptional(linkage.ollamaRunId) ? { ollamaRunId: safeOptional(linkage.ollamaRunId) } : {}),
     }
   }
 

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

@@ -75,13 +75,14 @@ export interface QueryJobRequest {
 }
 
 export interface QueryJobProviderLinkage {
-  provider?: 'claude' | 'codex' | 'cursor'
+  provider?: 'claude' | 'codex' | 'cursor' | 'ollama'
   resolvedModel?: string
   cliSessionId?: string
   claudeRunId?: string
   codexRunId?: string
   codexThreadId?: string
   cursorRunId?: string
+  ollamaRunId?: string
 }
 
 /** Path/id-free aggregate from output-image finalization. Values are bounded

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/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'