feat: switch voice from Groq to Claude Haiku via ai-proxy

app/api/voice/chat/route.ts

@@ -2,8 +2,6 @@ export const dynamic = 'force-dynamic'
 export const runtime = 'nodejs'
 
 import { NextResponse } from 'next/server'
-import Groq from 'groq-sdk'
-import { HttpsProxyAgent } from 'https-proxy-agent'
 
 import { voiceBus } from '@/lib/voice-bus'
 import { systemPrompt } from '@/lib/voice-prompts'
@@ -14,12 +12,15 @@ import {
   HistoryMessage,
 } from '@/lib/voice-history'
 
-const MODEL = process.env.GROQ_MODEL || 'llama-3.3-70b-versatile'
+const MODEL = process.env.CLAUDE_MODEL || 'claude-haiku-4-5'
 const MAX_TOKENS = parseInt(process.env.VOICE_MAX_TOKENS || '300', 10)
 const MAX_TOOL_ROUNDS = 4
 const RATE_LIMIT_PER_MINUTE = parseInt(process.env.VOICE_RATE_LIMIT || '20', 10)
 
-// In-memory rate-limit per IP / cookie (host один — Docker контейнер).
+const AI_PROXY_URL = process.env.AI_PROXY_URL || 'http://192.168.31.103:3301'
+const AI_PROXY_KEY = process.env.AI_PROXY_KEY || 'review-bot-proxy-d3ff719d7c87e529909c09fadcbf2748'
+
+// Rate limit
 const rateBuckets = new Map<string, { count: number; resetAt: number }>()
 function rateLimit(key: string): boolean {
   const now = Date.now()
@@ -40,28 +41,92 @@ function sweep() {
   for (const [k, v] of rateBuckets) if (v.resetAt <= now) rateBuckets.delete(k)
 }
 
-let _client: Groq | null = null
+// Convert OpenAI-style tool schemas to Anthropic format
-function client(): Groq {
+function toAnthropicTools(tools: any[]): any[] {
-  if (_client) return _client
+  return tools.map(t => ({
-  const apiKey = process.env.GROQ_API_KEY
+    name: t.function.name,
-  if (!apiKey) throw new Error('GROQ_API_KEY not set')
+    description: t.function.description || '',
-  const proxyUrl = process.env.GROQ_PROXY
+    input_schema: t.function.parameters || { type: 'object', properties: {} },
-  const httpAgent = proxyUrl ? new HttpsProxyAgent(proxyUrl) : undefined
+  }))
-  _client = new Groq({ apiKey, httpAgent } as any)
-  return _client
 }
 
-function emitVoice(event: string, agent: 'cosmo' | 'lusya', text?: string) {
+// Convert history (OpenAI format) to Anthropic messages format
-  voiceBus.emit('voice', {
+// History may contain tool calls — need to convert
-    event,
+function historyToAnthropicMessages(history: HistoryMessage[]): any[] {
-    agent,
+  const result: any[] = []
-    text,
+  for (const msg of history) {
-    timestamp: new Date().toISOString(),
+    if (msg.role === 'system') continue // skip, goes in system field
+    if (msg.role === 'user') {
+      result.push({ role: 'user', content: msg.content || '' })
+    } else if (msg.role === 'assistant') {
+      if (msg.tool_calls && msg.tool_calls.length > 0) {
+        // Assistant with tool calls
+        const content: any[] = []
+        if (msg.content) content.push({ type: 'text', text: msg.content })
+        for (const tc of msg.tool_calls) {
+          let input: any = {}
+          try { input = JSON.parse(tc.function.arguments) } catch {}
+          content.push({ type: 'tool_use', id: tc.id, name: tc.function.name, input })
+        }
+        result.push({ role: 'assistant', content })
+      } else {
+        result.push({ role: 'assistant', content: msg.content || '' })
+      }
+    } else if (msg.role === 'tool') {
+      // Tool result goes as user message
+      const last = result[result.length - 1]
+      const toolResultBlock = {
+        type: 'tool_result',
+        tool_use_id: msg.tool_call_id,
+        content: msg.content || '',
+      }
+      if (last && last.role === 'user' && Array.isArray(last.content)) {
+        last.content.push(toolResultBlock)
+      } else {
+        result.push({ role: 'user', content: [toolResultBlock] })
+      }
+    }
+  }
+  return result
+}
+
+async function claudeRequest(model: string, system: string, messages: any[], tools?: any[]): Promise<any> {
+  const body: any = {
+    model,
+    max_tokens: MAX_TOKENS,
+    system,
+    messages,
+  }
+  if (tools && tools.length > 0) {
+    body.tools = tools
+  }
+
+  const res = await fetch(`${AI_PROXY_URL}/v1/messages`, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      'X-Proxy-Key': AI_PROXY_KEY,
+      'anthropic-version': '2023-06-01',
+    },
+    body: JSON.stringify(body),
   })
+
+  if (!res.ok) {
+    const err = await res.text()
+    throw new Error(`claude_proxy_${res.status}: ${err}`)
+  }
+  return res.json()
 }
 
 type AgentId = 'cosmo' | 'lusya'
 
+function emitVoice(event: string, agent: AgentId, text?: string) {
+  voiceBus.emit('voice', {
+    event, agent, text,
+    timestamp: new Date().toISOString(),
+  })
+}
+
 export async function POST(req: Request) {
   const cookie = req.headers.get('cookie') || ''
   const tokenMatch = cookie.match(/auth_token=([a-f0-9]{32,})/i)
@@ -89,104 +154,84 @@ export async function POST(req: Request) {
     return NextResponse.json({ text: msg, reset: true })
   }
 
-  // Загружаем историю и строим messages для Groq (OpenAI-compatible format)
   const history = await loadHistory(agent)
+  const sysPrompt = systemPrompt(agent)
+  const anthropicTools = toAnthropicTools(TOOL_SCHEMAS as any[])
 
-  // Системный prompt + история + новый user message
+  // Build messages array
-  const apiMessages: any[] = [
+  const messages: any[] = [
-    { role: 'system', content: systemPrompt(agent) },
+    ...historyToAnthropicMessages(history),
-    ...history,
     { role: 'user', content: userText },
   ]
 
   let finalText = ''
-  const historyStartLen = apiMessages.length // позиция после которой добавляем новые turns
+  // Track new turns for history saving
-
+  const newTurns: HistoryMessage[] = [{ role: 'user', content: userText }]
-  // Защита от tool-cycling
-  let lastToolSig = ''
 
   try {
-    const c = client()
     for (let round = 0; round < MAX_TOOL_ROUNDS; round++) {
       const t0 = Date.now()
-      const resp = await c.chat.completions.create({
+      const resp = await claudeRequest(MODEL, sysPrompt, messages, anthropicTools)
-        model: MODEL,
-        max_tokens: MAX_TOKENS,
-        messages: apiMessages,
-        tools: TOOL_SCHEMAS as any,
-        tool_choice: 'auto',
-      })
-
-      const choice = resp.choices[0]
-      const msg = choice.message
-      const usage = resp.usage as any
 
       console.log(
         `[voice/chat] ${agent} round ${round + 1} ${Date.now() - t0}ms · ` +
-        `stop=${choice.finish_reason} · in=${usage?.prompt_tokens} out=${usage?.completion_tokens}`
+        `stop=${resp.stop_reason} · in=${resp.usage?.input_tokens} out=${resp.usage?.output_tokens}`
       )
 
-      // Добавляем assistant message в messages
+      const content: any[] = resp.content || []
-      apiMessages.push(msg)
+      const stopReason: string = resp.stop_reason || 'end_turn'
 
-      if (choice.finish_reason === 'tool_calls' && msg.tool_calls?.length) {
+      if (stopReason === 'tool_use') {
-        const toolCalls = msg.tool_calls
+        const toolUseBlocks = content.filter((b: any) => b.type === 'tool_use')
+        const textBlocks = content.filter((b: any) => b.type === 'text')
+        const partialText = textBlocks.map((b: any) => b.text).join('')
 
-        // Loop-guard: сигнатура текущего раунда
+        // Add assistant message to messages
-        const sig = toolCalls
+        messages.push({ role: 'assistant', content })
-          .map((tc: any) => `${tc.function.name}:${tc.function.arguments}`)
+        newTurns.push({
-          .sort()
+          role: 'assistant',
-          .join('|')
+          content: partialText || null,
-        if (sig === lastToolSig) {
+          tool_calls: toolUseBlocks.map((b: any) => ({
-          console.warn('[voice/chat] tool cycle detected, breaking loop')
+            id: b.id,
-          finalText += '\nНе получилось выполнить запрос.'
+            function: { name: b.name, arguments: JSON.stringify(b.input) },
-          break
+          })),
-        }
+        })
-        lastToolSig = sig
 
-        // Выполняем все tool calls и добавляем результаты
+        // Execute tools and collect results
-        for (const tc of toolCalls) {
+        const toolResults: any[] = []
-          console.log(`[voice/chat] tool ${tc.function.name}(${tc.function.arguments.slice(0, 200)})`)
+        for (const tb of toolUseBlocks) {
-          let args: any = {}
+          console.log(`[voice/chat] tool ${tb.name}(${JSON.stringify(tb.input).slice(0, 200)})`)
-          try { args = JSON.parse(tc.function.arguments) } catch (_) {}
+          const result = await executeTool(tb.name, tb.input || {}, agent)
-          const result = await executeTool(tc.function.name, args, agent)
+          toolResults.push({
-          apiMessages.push({
+            type: 'tool_result',
-            role: 'tool',
+            tool_use_id: tb.id,
-            tool_call_id: tc.id,
             content: JSON.stringify(result),
           })
+          newTurns.push({
+            role: 'tool',
+            content: JSON.stringify(result),
+            tool_call_id: tb.id,
+          })
         }
+
+        // Add tool results as user message
+        messages.push({ role: 'user', content: toolResults })
         continue
       }
 
-      // Финальный ответ
+      // end_turn — final response
-      finalText = msg.content || ''
+      const textBlocks = content.filter((b: any) => b.type === 'text')
+      finalText = textBlocks.map((b: any) => b.text).join('')
+
+      newTurns.push({ role: 'assistant', content: finalText })
       break
     }
   } catch (e: any) {
     const errStr = String(e?.message || e)
-    console.error('[voice/chat] groq error:', errStr)
+    console.error('[voice/chat] claude error:', errStr)
-
+    const msg = 'Что-то сломалось.'
-    // tool_use_failed: модель неправильно сформировала tool call — повторить без tools
+    emitVoice('error', agent, msg)
-    if (errStr.includes('tool_use_failed') || errStr.includes('Failed to call a function')) {
+    return NextResponse.json({ error: 'llm_failed', detail: errStr, text: msg }, { status: 502 })
-      try {
-        const c2 = client()
-        const fallback = await c2.chat.completions.create({
-          model: MODEL,
-          max_tokens: MAX_TOKENS,
-          messages: apiMessages.slice(0, historyStartLen + 1),
-        })
-        finalText = fallback.choices[0]?.message?.content || ''
-        console.log('[voice/chat] tool_use_failed fallback ok')
-      } catch (e2) {
-        console.error('[voice/chat] fallback failed:', e2)
-        finalText = 'Не удалось выполнить запрос.'
-      }
-    } else {
-      const msg = 'Что-то сломалось.'
-      emitVoice('error', agent, msg)
-      return NextResponse.json({ error: 'llm_failed', detail: errStr, text: msg }, { status: 502 })
-    }
   }
 
   if (!finalText.trim()) {
@@ -195,22 +240,20 @@ export async function POST(req: Request) {
     return NextResponse.json({ text: msg }, { status: 200 })
   }
 
-  // Сохраняем новые turns в историю (без системного prompt'а)
+  // Save history
-  const newTurns = apiMessages.slice(historyStartLen)
+  const updatedHistory: HistoryMessage[] = [...history, ...newTurns]
-  const updatedHistory: HistoryMessage[] = [
-    ...history,
-    ...newTurns.map((m: any) => ({ role: m.role, content: m.content ?? null, tool_calls: m.tool_calls, tool_call_id: m.tool_call_id } as HistoryMessage)),
-  ]
   await saveHistory(agent, updatedHistory)
 
-  // Убрать строки вида "get_weather ..." или "<function=...>" которые иногда генерирует LLM
+  // Filter any tool call artifacts from text
   const filteredText = finalText
     .split('\n')
     .filter(line => {
       const l = line.trim()
-      return !(/^(get_|set_|control_|create_|update_|delete_|cancel_)[a-z_]+\s/.test(l) ||
+      return !(
-               l.startsWith('<function') ||
+        /^(get_|set_|control_|create_|update_|delete_|cancel_)[a-z_]+\s/.test(l) ||
-               l.startsWith('function='))
+        l.startsWith('<function') ||
+        l.startsWith('function=')
+      )
     })
     .join('\n')
     .trim()

components/VoiceController.tsx

@@ -121,7 +121,6 @@ export default function VoiceController() {
         model: 'v5',
         baseAssetPath: '/vad/',
         onnxWASMBasePath: '/vad/',
-        logLevel: 'error',
         ortConfig: (ort: any) => {
           ort.env.wasm.numThreads = 1
           ort.env.wasm.simd = true