osborn 0.9.177 → 0.9.178

package/dist/claude-llm.js

@@ -167,7 +167,7 @@ ensureCompactionSettings();
 // (1M) window; without it opus runs at its 200K default. NOTE: 1M activation
 // also depends on account entitlement (auto on Team seats; else usage credits).
 process.env.ENABLE_1M_CONTEXT = '1';
-process.env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE = '92';
+process.env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE = '60';
 // Research mode tools — full research capabilities
 // Named sub-agents — the orchestrator delegates to these specialists. Each has
 // a specific role, model, and tool set. Module-level + exported so the HTTP

package/dist/voice-io.js

@@ -131,8 +131,8 @@ export const DEFAULT_VOICE_IO_CONFIG = {
 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', model: 'nova-3', language: 'en', // Streaming, silence-based endpointing
-    // 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
+    provider: 'deepgram-flux', model: 'flux-general-en', language: 'en', // Streaming, ML-based turn detection (requires Deepgram V2 access)
 };
 export const DIRECT_MODE_TTS = {
     // provider: 'deepgram', model: 'aura-2-asteria-en',  // WebSocket-based: handles TTS abort cleanly (no unrecoverable crash on interruption)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.177",
+  "version": "0.9.178",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {

package/tests/voice-io-stt-config.test.ts

@@ -0,0 +1,177 @@
+/**
+ * Regression tests for voice-io.ts STT configuration
+ *
+ * Scope: covers the 0.9.175 → 0.9.177 change that switches DIRECT_MODE_STT
+ * from deepgram-flux (ML turn detection) back to deepgram (silence-based
+ * endpointing), due to Flux V2 keepalive timeout bug (~30s of silence kills
+ * the connection).
+ *
+ * Derived from: requirements in CLAUDE.md (Three Voice Modes), voice-io.ts
+ * interface definitions, and the documented bug. NOT derived from the diff.
+ */
+
+import { strict as assert } from 'node:assert'
+
+// ──────────────────────────────────────────────────────────────────────────────
+// Inline the minimal types — avoid importing from the module which requires
+// live LK plugins (can't instantiate in a unit test without credentials).
+// ──────────────────────────────────────────────────────────────────────────────
+interface STTConfig {
+  provider: 'deepgram' | 'deepgram-flux' | 'groq-whisper' | 'openai-whisper'
+  model?: string
+  language?: string
+  eotThreshold?: number
+  eotTimeoutMs?: number
+}
+
+// ──────────────────────────────────────────────────────────────────────────────
+// Read the compiled output so we get the real runtime values, not just TS types
+// ──────────────────────────────────────────────────────────────────────────────
+// We import from the transpiled dist to avoid plugin side-effects at import time.
+// If dist is missing, fall back to a direct static-analysis of the source file.
+
+let DIRECT_MODE_STT_PROVIDER: string | undefined
+let DIRECT_MODE_STT_MODEL: string | undefined
+let DIRECT_MODE_STT_LANG: string | undefined
+
+try {
+  // dist/voice-io.js is the compiled output of the build step
+  const mod = await import('../dist/voice-io.js')
+  const cfg: STTConfig = mod.DIRECT_MODE_STT
+  DIRECT_MODE_STT_PROVIDER = cfg.provider
+  DIRECT_MODE_STT_MODEL = cfg.model
+  DIRECT_MODE_STT_LANG = cfg.language
+} catch (e) {
+  // Fallback: parse source file statically to extract the active provider line
+  const { readFileSync } = await import('node:fs')
+  const { fileURLToPath } = await import('node:url')
+  const { dirname, join } = await import('node:path')
+  const here = dirname(fileURLToPath(import.meta.url))
+  const src = readFileSync(join(here, '../src/voice-io.ts'), 'utf8')
+
+  // Find the non-commented active provider line in DIRECT_MODE_STT
+  const providerMatch = src.match(
+    /export const DIRECT_MODE_STT[\s\S]*?^\s+provider:\s*'([^']+)'/m
+  )
+  const modelMatch = src.match(
+    /export const DIRECT_MODE_STT[\s\S]*?model:\s*'([^']+)'/m
+  )
+  const langMatch = src.match(
+    /export const DIRECT_MODE_STT[\s\S]*?language:\s*'([^']+)'/m
+  )
+  DIRECT_MODE_STT_PROVIDER = providerMatch?.[1]
+  DIRECT_MODE_STT_MODEL = modelMatch?.[1]
+  DIRECT_MODE_STT_LANG = langMatch?.[1]
+}
+
+// ──────────────────────────────────────────────────────────────────────────────
+// Tests
+// ──────────────────────────────────────────────────────────────────────────────
+
+let passed = 0
+let failed = 0
+
+function test(name: string, fn: () => void) {
+  try {
+    fn()
+    console.log(`  ✅ ${name}`)
+    passed++
+  } catch (err: any) {
+    console.error(`  ❌ ${name}`)
+    console.error(`     ${err.message}`)
+    failed++
+  }
+}
+
+console.log('\n=== voice-io STT config regression tests ===\n')
+
+// ── 1. Provider must be deepgram (not deepgram-flux) ──────────────────────────
+test('DIRECT_MODE_STT.provider is "deepgram" (not deepgram-flux)', () => {
+  assert.equal(
+    DIRECT_MODE_STT_PROVIDER,
+    'deepgram',
+    `Expected provider "deepgram" but got "${DIRECT_MODE_STT_PROVIDER}". ` +
+    'deepgram-flux has a known keepalive timeout bug (silent ~30s kills the connection).'
+  )
+})
+
+test('DIRECT_MODE_STT.provider is NOT "deepgram-flux"', () => {
+  assert.notEqual(
+    DIRECT_MODE_STT_PROVIDER,
+    'deepgram-flux',
+    'deepgram-flux must not be the active provider — ' +
+    'Flux V2 has silent keepalive timeout bug (~30s silence kills connection).'
+  )
+})
+
+// ── 2. Model and language ──────────────────────────────────────────────────────
+test('DIRECT_MODE_STT.model is "nova-3"', () => {
+  assert.equal(
+    DIRECT_MODE_STT_MODEL,
+    'nova-3',
+    `Expected model "nova-3" but got "${DIRECT_MODE_STT_MODEL}".`
+  )
+})
+
+test('DIRECT_MODE_STT.language is "en"', () => {
+  assert.equal(
+    DIRECT_MODE_STT_LANG,
+    'en',
+    `Expected language "en" but got "${DIRECT_MODE_STT_LANG}".`
+  )
+})
+
+// ── 3. STTConfig type-level: provider values are the expected set ─────────────
+// (These are compile-time guarantees verified at build time, but we assert
+//  them at runtime too so future changes surface here.)
+test('DIRECT_MODE_STT provider is one of the valid union members', () => {
+  const validProviders = ['deepgram', 'deepgram-flux', 'groq-whisper', 'openai-whisper']
+  assert.ok(
+    validProviders.includes(DIRECT_MODE_STT_PROVIDER!),
+    `provider "${DIRECT_MODE_STT_PROVIDER}" is not in valid set: ${validProviders.join(', ')}`
+  )
+})
+
+// ── 4. Backward-compat: createSTT deepgram path shape ────────────────────────
+// We verify the source still has the deepgram case in createSTT so the switch
+// won't hit the `default: throw` path at runtime.
+test('createSTT source still contains deepgram case (backward compat)', async () => {
+  const { readFileSync } = await import('node:fs')
+  const { fileURLToPath } = await import('node:url')
+  const { dirname, join } = await import('node:path')
+  const here = dirname(fileURLToPath(import.meta.url))
+  const src = readFileSync(join(here, '../src/voice-io.ts'), 'utf8')
+  assert.ok(
+    src.includes("case 'deepgram':"),
+    "createSTT source must still contain case 'deepgram' — removing it would throw at runtime"
+  )
+})
+
+test('createSTT source still contains deepgram-flux case (backward compat)', async () => {
+  const { readFileSync } = await import('node:fs')
+  const { fileURLToPath } = await import('node:url')
+  const { dirname, join } = await import('node:path')
+  const here = dirname(fileURLToPath(import.meta.url))
+  const src = readFileSync(join(here, '../src/voice-io.ts'), 'utf8')
+  assert.ok(
+    src.includes("case 'deepgram-flux':"),
+    "createSTT source must still contain case 'deepgram-flux' — it must remain selectable via config"
+  )
+})
+
+// ── 5. deepgram endpointing value in createSTT ────────────────────────────────
+test('createSTT deepgram case uses 550ms endpointing (mid-sentence fragment prevention)', async () => {
+  const { readFileSync } = await import('node:fs')
+  const { fileURLToPath } = await import('node:url')
+  const { dirname, join } = await import('node:path')
+  const here = dirname(fileURLToPath(import.meta.url))
+  const src = readFileSync(join(here, '../src/voice-io.ts'), 'utf8')
+  assert.ok(
+    src.includes('endpointing: 550'),
+    'createSTT deepgram case must use endpointing: 550ms to prevent mid-sentence transcript fragments'
+  )
+})
+
+// ── Summary ───────────────────────────────────────────────────────────────────
+console.log(`\n--- ${passed + failed} tests: ${passed} passed, ${failed} failed ---\n`)
+if (failed > 0) process.exit(1)