osborn 0.9.193 → 0.9.195

package/dist/config.d.ts

@@ -2,8 +2,8 @@ import type { McpServerConfig } from './claude-handler.js';
 export type VoiceMode = 'pipeline';
 export type EditMode = 'read-only' | 'edit';
 export type AgentMode = 'plan' | 'execute' | 'research';
-export type STTProvider = 'soniox' | 'deepgram' | 'groq-whisper' | 'openai-whisper';
-export type TTSProvider = 'soniox' | 'openai' | 'deepgram';
+export type STTProvider = 'soniox' | 'deepgram' | 'deepgram-flux' | 'groq-whisper' | 'openai-whisper';
+export type TTSProvider = 'soniox' | 'openai' | 'deepgram' | 'groq-orpheus' | 'fishaudio' | 'rime';
 export interface DirectConfig {
     stt?: {
         provider?: STTProvider;
@@ -16,11 +16,16 @@ export interface DirectConfig {
         voice?: string;
     };
 }
+export interface InferenceConfig {
+    fastBrainModel?: string;
+    fastBrainProvider?: 'openrouter';
+}
 export interface OsbornConfig {
     workingDirectory?: string;
     mcpServers?: Record<string, McpServerConfigYaml>;
     voiceMode?: VoiceMode;
     direct?: DirectConfig;
+    inference?: InferenceConfig;
 }
 interface McpServerConfigYaml {
     enabled?: boolean;
@@ -76,7 +81,8 @@ export declare function getVoiceMode(_config: OsbornConfig): VoiceMode;
 /**
  * Get pipeline mode config with defaults merged
  */
-export declare function getDirectConfig(config: OsbornConfig): Required<DirectConfig>;
+export declare function getDirectConfig(_config: OsbornConfig): Required<DirectConfig>;
+export declare function getInferenceConfig(_config: OsbornConfig): Required<InferenceConfig>;
 /**
  * Save config to file
  */

package/dist/config.js

@@ -43,35 +43,74 @@ export const MCP_CATALOG = [
 // (`/usr/local/nvm/.../osborn`) and gets persisted to ~/.osborn/config.yaml forever.
 // Leaving it undefined lets the runtime self-heal in index.ts resolve it on every boot
 // from OSBORN_CWD → process.cwd() at the actual time the agent starts.
+// ============================================================
+// PROVIDER DEFAULTS — Single source of truth for STT, TTS, and inference.
+// To switch provider: change `provider` (and model/voice) here.
+// All alternatives are listed as comments. Env vars override without redeploy:
+//   STT:       OSBORN_STT_PROVIDER, OSBORN_STT_MODEL, OSBORN_STT_LANGUAGE
+//   TTS:       OSBORN_TTS_PROVIDER, OSBORN_TTS_MODEL, OSBORN_TTS_VOICE
+//   Inference: OSBORN_FAST_BRAIN_MODEL
+// These are admin-owned defaults — NOT written to config.yaml on disk.
+// ============================================================
 const DEFAULT_CONFIG = {
     voiceMode: 'pipeline',
     direct: {
         stt: {
-            // Soniox stt-rt-v4 — semantic endpointing (ML, not silence), word timestamps,
-            // custom vocabulary via context.terms. ~60% cheaper than nova-3 ($0.0017 vs $0.0043/min).
-            // Needs SONIOX_API_KEY.
+            // ── Active ─────────────────────────────────────────────────────────────
             provider: 'soniox',
             model: 'stt-rt-v4',
-            // Previous: Deepgram nova-3, silence-based endpointing (550ms configured in voice-io.ts).
-            // Switch back: provider: 'deepgram', model: 'nova-3'
+            // Semantic endpointing — ML model holds on incomplete thoughts, commits on
+            // natural sentence ends. ~60% cheaper than nova-3 ($0.0017 vs $0.0043/min).
+            // Needs SONIOX_API_KEY. See voice-io.ts createSTT for tuning params.
+            //
+            // ── Alternatives ───────────────────────────────────────────────────────
+            // provider: 'deepgram', model: 'nova-3'
+            //   Silence-based endpointing (550ms). Fast, reliable. $0.0043/min.
+            // provider: 'deepgram-flux', model: 'flux-general-en'
+            //   ML turn detection (server-side). Flux V2 has 30s silent keepalive bug.
+            // provider: 'groq-whisper', model: 'whisper-large-v3-turbo'
+            //   Batch STT via Groq (very fast). Requires VAD for turn detection.
+            // provider: 'openai-whisper', model: 'whisper-1'
+            //   Batch STT via OpenAI. Requires VAD. $0.006/min.
         },
         tts: {
-            // Soniox tts-rt-v1 — real-time WebSocket streaming, speed control (0.7–1.3x),
-            // clean abort on interruption. Estimated ~$4–16/M chars vs OpenAI tts-1-hd $30/M.
-            // Pricing: $0.70/hr of generated speech (preview). Needs SONIOX_API_KEY.
+            // ── Active ─────────────────────────────────────────────────────────────
             provider: 'soniox',
             model: 'tts-rt-v1',
             voice: 'Victoria',
-            // Previous: OpenAI tts-1-hd, voice fable — high quality, $30/M chars, ~500ms TTFB.
-            // Switch back: provider: 'openai', model: 'tts-1-hd', voice: 'fable'
-            // Other options already wired in voice-io.ts:
-            //   Rime Mist v3:     provider: 'rime',      voice: 'cove'  — 37ms TTFB, $30/M, WebSocket
-            //   Fish Audio s2-pro: provider: 'fishaudio', voice: '<id>'  — $15/M, voice cloning
-            //   Groq Orpheus:     provider: 'groq-orpheus', voice: 'autumn' — fast Groq chips, $22/M
-            //   Deepgram Aura-2:  provider: 'deepgram',  model: 'aura-2-asteria-en' — $15/M, ~100ms TTFB
-            //   OpenAI tts-1:     provider: 'openai',    model: 'tts-1', voice: 'fable' — $15/M
+            // WebSocket streaming — clean abort on interruption. Speed control 0.7–1.3x.
+            // Est. ~$4–16/M chars vs OpenAI $30/M. Victoria = en-GB female, refined.
+            // Also: Isla (en-GB, lively), Maya (en-US female). Needs SONIOX_API_KEY.
+            //
+            // ── Alternatives ───────────────────────────────────────────────────────
+            // provider: 'openai', model: 'tts-1-hd', voice: 'fable'
+            //   $30/M chars, ~500ms TTFB, HTTP streaming. 6 voices: alloy echo fable onyx nova shimmer.
+            // provider: 'openai', model: 'tts-1', voice: 'fable'
+            //   $15/M chars, slightly lower quality, same voices.
+            // provider: 'deepgram', model: 'aura-2-asteria-en'
+            //   $15/M chars, ~100ms TTFB, WebSocket. Voices: asteria luna stella hera orion arcas perseus angus orpheus.
+            // provider: 'rime', model: 'mistv3', voice: 'cove'
+            //   $30/M chars, 37ms TTFB, WebSocket (clean abort). Voices: aurora ember cove. Needs RIME_API_KEY.
+            // provider: 'fishaudio', model: 's2-pro', voice: '<voice-id>'
+            //   $15/M chars, voice cloning, HTTP streaming. Needs FISH_AUDIO_API_KEY.
+            // provider: 'groq-orpheus', model: 'canopylabs/orpheus-v1-english', voice: 'autumn'
+            //   $22/M chars, fast (Groq chips). Voices: autumn diana hannah austin daniel troy.
         },
     },
+    inference: {
+        // ── Active ───────────────────────────────────────────────────────────────
+        fastBrainModel: 'deepseek/deepseek-chat',
+        fastBrainProvider: 'openrouter',
+        // Parallel fast brain: answers greetings/follow-ups instantly while main agent
+        // processes. Uses OpenRouter (OPENROUTER_API_KEY). ~20ms TTFB on deepseek-chat.
+        //
+        // ── Alternatives ─────────────────────────────────────────────────────────
+        // fastBrainModel: 'deepseek/deepseek-reasoner'            — slower, deeper CoT
+        // fastBrainModel: 'google/gemini-flash-1.5'               — Google via OpenRouter
+        // fastBrainModel: 'openai/gpt-4o-mini'                    — OpenAI via OpenRouter
+        // fastBrainModel: 'meta-llama/llama-3.1-8b-instruct:free' — free tier via OpenRouter
+        // fastBrainModel: 'anthropic/claude-haiku-4-5-20251001'   — Haiku via OpenRouter
+    },
     mcpServers: {
     // ─────────────────────────────────────────────────────────────────────────
     // MCP Servers for Read-Only Plan Mode
@@ -139,10 +178,12 @@ export function loadConfig() {
         mkdirSync(CONFIG_DIR, { recursive: true });
         console.log(`📁 Created config directory: ${CONFIG_DIR}`);
     }
-    // Create default config if it doesn't exist
+    // Create default config if it doesn't exist.
+    // Exclude `direct` (STT/TTS) and `inference` — admin-owned, not user config on disk.
+    // Writing them to disk causes the file to permanently override npm package updates.
     if (!existsSync(CONFIG_FILE)) {
-        const defaultYaml = stringify(DEFAULT_CONFIG);
-        writeFileSync(CONFIG_FILE, defaultYaml, 'utf-8');
+        const { direct: _d, inference: _i, ...persistable } = DEFAULT_CONFIG;
+        writeFileSync(CONFIG_FILE, stringify(persistable), 'utf-8');
         console.log(`📝 Created default config: ${CONFIG_FILE}`);
         return DEFAULT_CONFIG;
     }
@@ -207,22 +248,33 @@ export function getVoiceMode(_config) {
 /**
  * Get pipeline mode config with defaults merged
  */
-export function getDirectConfig(config) {
+export function getDirectConfig(_config) {
     const defaults = DEFAULT_CONFIG.direct;
-    const userConfig = config.direct || {};
+    // STT/TTS settings are admin-owned defaults — they live in DEFAULT_CONFIG only.
+    // config.yaml is intentionally NOT consulted here: writing provider settings
+    // to the volume creates a stale file that permanently overrides npm updates.
+    // To override without a redeploy, set OSBORN_STT_PROVIDER / OSBORN_TTS_PROVIDER
+    // / OSBORN_TTS_VOICE / OSBORN_STT_MODEL / OSBORN_TTS_MODEL on the machine.
     return {
         stt: {
-            provider: userConfig.stt?.provider || defaults.stt.provider,
-            model: userConfig.stt?.model || defaults.stt.model,
-            language: userConfig.stt?.language || 'en',
+            provider: (process.env.OSBORN_STT_PROVIDER || defaults.stt.provider),
+            model: process.env.OSBORN_STT_MODEL || defaults.stt.model,
+            language: process.env.OSBORN_STT_LANGUAGE || 'en',
         },
         tts: {
-            provider: userConfig.tts?.provider || defaults.tts.provider,
-            model: userConfig.tts?.model || defaults.tts.model,
-            voice: userConfig.tts?.voice || defaults.tts.voice,
+            provider: (process.env.OSBORN_TTS_PROVIDER || defaults.tts.provider),
+            model: process.env.OSBORN_TTS_MODEL || defaults.tts.model,
+            voice: process.env.OSBORN_TTS_VOICE || defaults.tts.voice,
         },
     };
 }
+export function getInferenceConfig(_config) {
+    const defaults = DEFAULT_CONFIG.inference;
+    return {
+        fastBrainModel: process.env.OSBORN_FAST_BRAIN_MODEL || defaults.fastBrainModel,
+        fastBrainProvider: 'openrouter',
+    };
+}
 /**
  * Save config to file
  */

package/dist/index.js

@@ -30,7 +30,7 @@ const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
 import { createPatch } from 'diff';
 import { loadConfig, getMcpServers, getEnabledMcpServerNames, getVoiceMode, getDirectConfig, listSessions, listAllClaudeSessions, invalidateSessionListCache, getMostRecentSessionId, sessionExists, getSessionSummary, getConversationHistory, ensureSessionWorkspace, getMcpServerStatusList, buildMcpServersForKeys, listWorkspaceArtifacts } from './config.js';
-import { createSTT, createTTS, DIRECT_MODE_STT, DIRECT_MODE_TTS } from './voice-io.js';
+import { createSTT, createTTS } from './voice-io.js';
 import { createClaudeLLM, NAMED_AGENTS, applyTurbo } from './claude-llm.js';
 import { clearPipelineFastBrainSession, prewarmBM25Index } from './pipeline-fastbrain.js';
 import { getIndexPath, buildSummaryIndex } from './summary-index.js';
@@ -417,13 +417,11 @@ async function synthMp3(text) {
         console.warn('⚠️ synthMp3: no OPENAI_API_KEY — meeting has no voice');
         return null;
     }
-    // Meeting voice = the SAME OpenAI model/voice as the website's regular TTS
-    // (DIRECT_MODE_TTS), so the bot sounds IDENTICAL on both fronts (user directive
-    // 2026-08-04: Deepgram aura sounded "cheap and inconsistent"). Deepgram removed
-    // from the meeting path entirely. Pulls model/voice from DIRECT_MODE_TTS when
-    // it's an OpenAI config so the two never drift.
-    const model = DIRECT_MODE_TTS.provider === 'openai' ? (DIRECT_MODE_TTS.model || 'tts-1-hd') : 'tts-1-hd';
-    const voice = DIRECT_MODE_TTS.provider === 'openai' ? (DIRECT_MODE_TTS.voice || 'fable') : 'fable';
+    // Meeting audio is always OpenAI TTS (HTTP MP3 required for Recall output_audio).
+    // Reads OSBORN_TTS_* env vars — same source as getDirectConfig — so voice is
+    // consistent between the pipeline and meeting canvas when TTS provider is openai.
+    const model = process.env.OSBORN_TTS_PROVIDER === 'openai' ? (process.env.OSBORN_TTS_MODEL || 'tts-1-hd') : 'tts-1-hd';
+    const voice = process.env.OSBORN_TTS_PROVIDER === 'openai' ? (process.env.OSBORN_TTS_VOICE || 'fable') : 'fable';
     try {
         const r = await fetch('https://api.openai.com/v1/audio/speech', {
             method: 'POST',
@@ -859,18 +857,16 @@ function startApiServer(workingDir, port) {
                 return;
             }
             const t0 = Date.now();
-            // Meeting voice = the SAME OpenAI model/voice as the website's regular TTS
-            // (DIRECT_MODE_TTS) — user directive 2026-08-04: Deepgram aura removed, it
-            // sounded cheap/inconsistent. Consistency over the ~2-4s latency Deepgram
-            // saved. mp3 out (the canvas <audio> element plays it into the meeting).
+            // Meeting canvas TTS — always OpenAI HTTP (mp3 for <audio> element).
+            // Reads OSBORN_TTS_* env vars so voice stays in sync with pipeline when provider is openai.
             const key = process.env.OPENAI_API_KEY;
             if (!key) {
                 res.writeHead(400, { 'Content-Type': 'application/json' });
                 res.end(JSON.stringify({ error: 'no OPENAI_API_KEY' }));
                 return;
             }
-            const model = DIRECT_MODE_TTS.provider === 'openai' ? (DIRECT_MODE_TTS.model || 'tts-1-hd') : 'tts-1-hd';
-            const voice = url.searchParams.get('voice') || (DIRECT_MODE_TTS.provider === 'openai' ? (DIRECT_MODE_TTS.voice || 'fable') : 'fable');
+            const model = process.env.OSBORN_TTS_PROVIDER === 'openai' ? (process.env.OSBORN_TTS_MODEL || 'tts-1-hd') : 'tts-1-hd';
+            const voice = url.searchParams.get('voice') || (process.env.OSBORN_TTS_PROVIDER === 'openai' ? (process.env.OSBORN_TTS_VOICE || 'fable') : 'fable');
             try {
                 const tts = await fetch('https://api.openai.com/v1/audio/speech', {
                     method: 'POST',
@@ -3173,8 +3169,8 @@ async function main() {
     // Create DIRECT session (STT + Claude Agent SDK + TTS)
     async function createDirectSession(resumeSessionId, llmOverride) {
         console.log('🎯 Creating direct session...');
-        const stt = createSTT(DIRECT_MODE_STT);
-        const tts = createTTS(DIRECT_MODE_TTS);
+        const stt = createSTT(directConfig.stt);
+        const tts = createTTS(directConfig.tts);
         // Create Claude LLM wrapper — direct mode uses speech-optimized system prompt
         // skipTTSQueue: bypass LiveKit's BufferedTokenStream, use session.say() instead
         // llmOverride: pipeline mode passes PipelineDirectLLM which wraps its own ClaudeLLM

package/dist/pipeline-fastbrain.js

@@ -15,7 +15,9 @@
 // ============================================================
 // CONSTANTS
 // ============================================================
-const OPENROUTER_MODEL = 'deepseek/deepseek-chat';
+// Model reads from env var (set by main process from getInferenceConfig) or DEFAULT_CONFIG default.
+// To switch model: update DEFAULT_CONFIG.inference.fastBrainModel in config.ts, or set OSBORN_FAST_BRAIN_MODEL env var.
+const OPENROUTER_MODEL = process.env.OSBORN_FAST_BRAIN_MODEL || 'deepseek/deepseek-chat';
 const OPENROUTER_URL = 'https://openrouter.ai/api/v1/chat/completions';
 const TIMEOUT_MS = 20_000;
 const MAX_TOOL_ROUNDS = 4;

package/dist/voice-io.d.ts

@@ -42,9 +42,3 @@ export declare function createTTS(config: TTSConfig): any;
  * - False triggers from ambient noise
  */
 export declare function createVAD(): Promise<silero.VAD>;
-/**
- * Pipeline mode voice config — centralized here for easy provider swapping.
- * To switch providers: comment out the active line, uncomment the alternative.
- */
-export declare const DIRECT_MODE_STT: STTConfig;
-export declare const DIRECT_MODE_TTS: TTSConfig;

package/dist/voice-io.js

@@ -154,21 +154,3 @@ export async function createVAD() {
         activationThreshold: 0.85, // default — balanced for interruptions only
     });
 }
-/**
- * Pipeline mode voice config — centralized here for easy provider swapping.
- * To switch providers: comment out the active line, uncomment the alternative.
- */
-export const DIRECT_MODE_STT = {
-    // provider: 'groq-whisper', model: 'whisper-large-v3-turbo',          // Batch — needs VAD
-    // provider: 'openai-whisper', model: 'whisper-1',                     // Batch — needs VAD
-    // provider: 'deepgram-flux', model: 'flux-general-en', language: 'en',  // Streaming, ML-based turn detection — Flux V2 has silent keepalive timeout bug (~30s silence kills connection)
-    provider: 'deepgram', model: 'nova-3', language: 'en', // Streaming, silence-based endpointing (550ms)
-};
-export const DIRECT_MODE_TTS = {
-    // provider: 'deepgram', model: 'aura-2-asteria-en',  // WebSocket-based: handles TTS abort cleanly — but quality rejected (run-on sentences)
-    // provider: 'openai', model: 'tts-1', voice: 'fable',  // HTTP streaming: throws APIUserAbortError on interrupt → unrecoverable session crash
-    provider: 'openai', model: 'tts-1-hd', voice: 'fable', // 0.9.70: test tts-1-hd — tts-1 had chronic per-sentence HTTP hangs (40s SDK watchdog → APIUserAbortError mid-message)
-    // provider: 'groq-orpheus', model: 'canopylabs/orpheus-v1-english', voice: 'autumn',  // $22/M chars — voices: autumn, diana, hannah, austin, daniel, troy
-    // provider: 'fishaudio', model: 's2-pro', voice: '<voice-id>',  // $15/M chars — blind test winner, HTTP streaming (test abort behavior)
-    // provider: 'rime', model: 'mistv3', voice: 'cove',  // $30/M chars, 37ms TTFB — WebSocket (safe on interruption); voices: aurora, ember, cove
-};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.193",
+  "version": "0.9.195",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {