Attached-file path to agent, more cloud providers, 16kHz voice

- Attached files now pass their REAL server path into the prompt so the
  agent reads them with read_file instead of inventing /path/to/file.
- Added cloud providers: xAI (Grok), Together AI, DeepSeek, Perplexity
  (OpenAI-compatible) with preset models — alongside OpenAI/Anthropic/
  Gemini/Groq/Mistral/OpenRouter.
- Voice: resample the recording to 16 kHz mono (what Whisper expects) and
  warn on empty/silent capture — fixes 'records but transcribes nothing'.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: metiu1

vortelio/internal/cloud/cloud.go

@@ -96,6 +96,46 @@ var Providers = []Provider{
 		Format:       FormatOpenAI,
 		KeyHint:      "https://openrouter.ai/keys",
 	},
+	{
+		ID:           "xai",
+		Name:         "xAI (Grok)",
+		DefaultModel: "grok-2-latest",
+		BaseURL:      "https://api.x.ai/v1/chat/completions",
+		AuthHeader:   "Authorization",
+		AuthPrefix:   "Bearer ",
+		Format:       FormatOpenAI,
+		KeyHint:      "https://console.x.ai",
+	},
+	{
+		ID:           "together",
+		Name:         "Together AI",
+		DefaultModel: "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+		BaseURL:      "https://api.together.xyz/v1/chat/completions",
+		AuthHeader:   "Authorization",
+		AuthPrefix:   "Bearer ",
+		Format:       FormatOpenAI,
+		KeyHint:      "https://api.together.xyz/settings/api-keys",
+	},
+	{
+		ID:           "deepseek",
+		Name:         "DeepSeek",
+		DefaultModel: "deepseek-chat",
+		BaseURL:      "https://api.deepseek.com/v1/chat/completions",
+		AuthHeader:   "Authorization",
+		AuthPrefix:   "Bearer ",
+		Format:       FormatOpenAI,
+		KeyHint:      "https://platform.deepseek.com/api_keys",
+	},
+	{
+		ID:           "perplexity",
+		Name:         "Perplexity",
+		DefaultModel: "sonar",
+		BaseURL:      "https://api.perplexity.ai/chat/completions",
+		AuthHeader:   "Authorization",
+		AuthPrefix:   "Bearer ",
+		Format:       FormatOpenAI,
+		KeyHint:      "https://www.perplexity.ai/settings/api",
+	},
 }
 
 func FindProvider(id string) (Provider, bool) {

vortelio/internal/server/server_cloud.go

@@ -48,6 +48,25 @@ var cloudModelChoices = map[string][][2]string{
 		{"openai/gpt-4o", "GPT-4o"},
 		{"deepseek/deepseek-r1", "DeepSeek R1"},
 	},
+	"xai": {
+		{"grok-2-latest", "Grok 2"},
+		{"grok-2-vision-latest", "Grok 2 Vision"},
+		{"grok-beta", "Grok Beta"},
+	},
+	"together": {
+		{"meta-llama/Llama-3.3-70B-Instruct-Turbo", "Llama 3.3 70B Turbo"},
+		{"Qwen/Qwen2.5-72B-Instruct-Turbo", "Qwen2.5 72B Turbo"},
+		{"mistralai/Mixtral-8x7B-Instruct-v0.1", "Mixtral 8x7B"},
+	},
+	"deepseek": {
+		{"deepseek-chat", "DeepSeek V3 (chat)"},
+		{"deepseek-reasoner", "DeepSeek R1 (reasoner)"},
+	},
+	"perplexity": {
+		{"sonar", "Sonar"},
+		{"sonar-pro", "Sonar Pro"},
+		{"sonar-reasoning", "Sonar Reasoning"},
+	},
 }
 
 // GET /api/cloud/providers

vortelio/internal/server/ui.html

@@ -4219,10 +4219,19 @@ <h3 style="margin:0">☁️ Add cloud model</h3>
       toast('⏳ Trascrivo…');
       try {
         const blob = new Blob(_recChunks, { type: _recMedia.mimeType || 'audio/webm' });
+        if (!blob.size) { toast('🎤 Nessun audio registrato — controlla il microfono e riprova.', 'err'); return; }
         const ab = await blob.arrayBuffer();
         const ctx = new (window.AudioContext || window.webkitAudioContext)();
         const audio = await ctx.decodeAudioData(ab);
-        const wav = _encodeWAV(audio);
+        // Resample to 16 kHz mono — exactly what Whisper expects (avoids empty results).
+        const off = new OfflineAudioContext(1, Math.max(1, Math.ceil(audio.duration * 16000)), 16000);
+        const node = off.createBufferSource(); node.buffer = audio; node.connect(off.destination); node.start();
+        const rendered = await off.startRendering();
+        // Bail out if the mic captured basically silence.
+        const pcm = rendered.getChannelData(0); let peak = 0;
+        for (let i = 0; i < pcm.length; i += 64) { const a = Math.abs(pcm[i]); if (a > peak) peak = a; }
+        if (peak < 0.005) { toast('🎤 Audio troppo silenzioso — parla più vicino al microfono.', 'err'); return; }
+        const wav = _encodeWAV(rendered);
         const fd = new FormData();
         fd.append('file', wav, 'rec.wav');
         fd.append('model', audioModel.type + '/' + audioModel.name + ':' + audioModel.tag);
@@ -4609,9 +4618,16 @@ <h3 style="margin:0">☁️ Add cloud model</h3>
   }
 
   try {
+    // If files are attached, tell the model their REAL paths so its file tools
+    // can read them (otherwise it invents a placeholder like /path/to/file).
+    let promptText = text;
+    const attachPaths = (files || []).map(f => f.serverPath).filter(Boolean);
+    if (attachPaths.length) {
+      promptText += '\n\n[File allegati dall\'utente — leggine il contenuto con lo strumento read_file usando ESATTAMENTE questo percorso: ' + attachPaths.join('  |  ') + ']';
+    }
     // Send full history so the LLM has conversation context
     const payload = {
-      model: selectedModel, prompt: text, stream: true,
+      model: selectedModel, prompt: promptText, stream: true,
       force_cpu: useCPU, context_size: contextSize,
       think: thinkOn, // enable chain-of-thought for reasoning models
       tools_enabled: toolsOn, // enable function calling / tool use
@@ -4620,7 +4636,7 @@ <h3 style="margin:0">☁️ Add cloud model</h3>
     };
     const ag = buildAgenticPayload();
     if (ag) payload.agentic = ag;
-    if (files.length) payload.input_file = files[0].serverPath;
+    if (attachPaths.length) payload.input_file = attachPaths[0];
     const r = await fetchRetry('/api/generate', {method:'POST', headers:{'Content-Type':'application/json'}, body:JSON.stringify(payload), signal});
 
     pulse.remove();