fix(tts): support OpenRouter voice overrides and fallback

docs/guides/configuration.md

@@ -455,7 +455,7 @@ Use `mode = lines` when:
   "tools": {
     "read_file": {
       "enabled": true,
-       "mode": "lines",
+      "mode": "lines",
       "max_read_file_size": 65536
     }
   }
@@ -838,6 +838,9 @@ Legacy Telegram environment variables remain compatible: `PICOCLAW_CHANNELS_TELE
 
 Failure behavior is intentionally conservative: if streaming fails before any visible chunk is sent, PicoClaw retries once through the normal `Chat()` path. If a chunk has already been shown to the user, PicoClaw does not send a second non-streaming answer, because that would duplicate visible output.
 
+For model-specific TTS request fields such as custom speech `voice` names or
+`response_format: "mp3"`, use `model_list[].extra_body`.
+
 #### Vendor-Specific Examples
 
 > **Tip**: You can omit `api_key` fields and store them in `.security.yml` for better security. See [Security Configuration](#-security-configuration-recommended).
@@ -856,6 +859,35 @@ Failure behavior is intentionally conservative: if streaming fails before any vi
 
 </details>
 
+<details>
+<summary><b>OpenRouter TTS (MAI Voice 2)</b></summary>
+
+```json
+{
+  "model_name": "mai-voice-2",
+  "provider": "openrouter",
+  "model": "microsoft/mai-voice-2",
+  "api_base": "https://openrouter.ai/api/v1",
+  "extra_body": {
+    "voice": "en-US-Harper:MAI-Voice-2",
+    "response_format": "mp3"
+  }
+  // api_key: set in .security.yml
+}
+```
+
+Pair this with:
+
+```json
+{
+  "voice": {
+    "tts_model_name": "mai-voice-2"
+  }
+}
+```
+
+</details>
+
 <details>
 <summary><b>VolcEngine (Doubao)</b></summary>
 

docs/guides/providers.md

@@ -138,6 +138,9 @@ This design also enables **multi-agent support** with flexible provider selectio
 
 When streaming is disabled, omit the `streaming` block. Writing `"streaming": {"enabled": false}` is optional and not needed in generated or hand-written config.
 
+`extra_body` is especially useful for model-specific TTS fields on OpenAI-compatible
+speech routes, for example custom `voice` names or `response_format: "mp3"`.
+
 #### Tool Schema Compatibility
 
 By default, PicoClaw now forwards tool JSON Schemas unchanged.
@@ -205,6 +208,41 @@ If `voice.model_name` is not configured, PicoClaw will continue to fall back to
 }
 ```
 
+#### Voice Synthesis
+
+You can configure a dedicated text-to-speech model with `voice.tts_model_name`.
+When the provider needs model-specific TTS request fields, put them in
+`model_list[].extra_body`.
+
+Example with OpenRouter `microsoft/mai-voice-2`:
+
+```json
+{
+  "model_list": [
+    {
+      "model_name": "mai-voice-2",
+      "provider": "openrouter",
+      "model": "microsoft/mai-voice-2",
+      "api_base": "https://openrouter.ai/api/v1",
+      "extra_body": {
+        "voice": "en-US-Harper:MAI-Voice-2",
+        "response_format": "mp3"
+      },
+      "api_keys": ["sk-or-your-openrouter-key"]
+    }
+  ],
+  "voice": {
+    "tts_model_name": "mai-voice-2"
+  }
+}
+```
+
+Notes that matter:
+
+- PicoClaw still uses the OpenAI-compatible `/audio/speech` route for this setup.
+- The default TTS request uses `voice: alloy` and `response_format: opus`.
+- Override those defaults with `extra_body` when your selected TTS model requires different values.
+
 #### Vendor-Specific Examples
 
 **OpenAI**

pkg/audio/tts/README.md

@@ -25,6 +25,8 @@ Instead:
 
 - `voice.tts_model_name` selects a named entry from `model_list`.
 - That `model_list` entry provides the provider, model ID, API base, and proxy settings.
+- For providers that need model-specific TTS parameters, use `model_list[].extra_body`
+  to pass fields such as `voice` and `response_format`.
 - `.security.yml` stores the API key for the same named model entry.
 
 This is the recommended and supported configuration pattern.
@@ -87,6 +89,43 @@ model_list:
 
 If you use a custom MiMo endpoint, you can also set `api_base` explicitly. Otherwise PicoClaw will use the provider default.
 
+### Option C: OpenRouter MAI Voice 2
+
+Some OpenAI-compatible TTS routes require provider-specific request fields.
+OpenRouter's `microsoft/mai-voice-2` is one example: it needs a model-specific
+voice name and works best with `response_format: "mp3"`.
+
+`config.json`
+
+```json
+{
+  "voice": {
+    "tts_model_name": "mai-voice-2"
+  },
+  "model_list": [
+    {
+      "model_name": "mai-voice-2",
+      "provider": "openrouter",
+      "model": "microsoft/mai-voice-2",
+      "api_base": "https://openrouter.ai/api/v1",
+      "extra_body": {
+        "voice": "en-US-Harper:MAI-Voice-2",
+        "response_format": "mp3"
+      }
+    }
+  ]
+}
+```
+
+`.security.yml`
+
+```yaml
+model_list:
+  mai-voice-2:
+    api_keys:
+      - "sk-or-your-openrouter-key"
+```
+
 ## What PicoClaw Sends Today
 
 The current TTS runtime uses an OpenAI-compatible speech request with these defaults:
@@ -96,11 +135,14 @@ The current TTS runtime uses an OpenAI-compatible speech request with these defa
 - Voice: `alloy`
 - Model: taken from the selected `model_list` entry
 
+These defaults can now be overridden per model through `model_list[].extra_body`.
+
 That means:
 
 - `openai/tts-1` works naturally.
 - Other OpenAI-compatible providers can work if they accept the same request format.
-- PicoClaw currently does not expose a user-facing config field for changing the TTS voice from `alloy`.
+- Provider-specific TTS models may need their own `voice` and `response_format` values.
+- If a provider rejects `response_format`, PicoClaw retries once without that field.
 
 ## How PicoClaw Chooses a TTS Provider
 
@@ -124,7 +166,8 @@ PicoClaw normalizes the configured base URL for TTS:
 
 - Setting `voice.tts_model_name` to a name that does not exist in `model_list`.
 - Adding a TTS model but forgetting to put its API key in `.security.yml`.
-- Assuming PicoClaw will automatically use provider-specific custom voices.
+- Assuming PicoClaw will automatically infer provider-specific custom voices.
+- Forgetting to set `model_list[].extra_body.voice` or `model_list[].extra_body.response_format` for TTS models that require them.
 - Using a provider endpoint that is not compatible with the OpenAI `/audio/speech` request format.
 
 ## Minimal Checklist

pkg/audio/tts/openai_tts.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
@@ -16,14 +17,49 @@ import (
 )
 
 type OpenAITTSProvider struct {
-	apiKey     string
+	apiKey         string
-	apiBase    string
+	apiBase        string
-	voice      string
+	voice          string
-	model      string
+	model          string
-	httpClient *http.Client
+	responseFormat string
+	httpClient     *http.Client
+}
+
+type OpenAITTSOptions struct {
+	Voice          string
+	ResponseFormat string
+}
+
+type openAITTSAudioStream struct {
+	io.ReadCloser
+	fileExt     string
+	contentType string
+}
+
+func (s *openAITTSAudioStream) AudioFileMeta() (string, string) {
+	return s.fileExt, s.contentType
+}
+
+type openAITTSAPIError struct {
+	statusCode int
+	body       string
+}
+
+func (e *openAITTSAPIError) Error() string {
+	return fmt.Sprintf("API error (status %d): %s", e.statusCode, e.body)
 }
 
 func NewOpenAITTSProvider(apiKey string, apiBase string, proxyURL string, model string) *OpenAITTSProvider {
+	return NewOpenAITTSProviderWithOptions(apiKey, apiBase, proxyURL, model, OpenAITTSOptions{})
+}
+
+func NewOpenAITTSProviderWithOptions(
+	apiKey string,
+	apiBase string,
+	proxyURL string,
+	model string,
+	options OpenAITTSOptions,
+) *OpenAITTSProvider {
 	// Normalize apiBase to avoid malformed endpoints like
 	// "https://api.openai.com/audio/speech" when "/v1" is required.
 	if apiBase == "" {
@@ -75,12 +111,23 @@ func NewOpenAITTSProvider(apiKey string, apiBase string, proxyURL string, model
 		model = "tts-1"
 	}
 
+	voice := strings.TrimSpace(options.Voice)
+	if voice == "" {
+		voice = "alloy"
+	}
+
+	responseFormat := strings.TrimSpace(options.ResponseFormat)
+	if responseFormat == "" {
+		responseFormat = "opus"
+	}
+
 	return &OpenAITTSProvider{
-		apiKey:     apiKey,
+		apiKey:         apiKey,
-		apiBase:    apiBase,
+		apiBase:        apiBase,
-		voice:      "alloy",
+		voice:          voice,
-		model:      model,
+		model:          model,
-		httpClient: client,
+		responseFormat: responseFormat,
+		httpClient:     client,
 	}
 }
 
@@ -91,11 +138,42 @@ func (t *OpenAITTSProvider) Name() string {
 func (t *OpenAITTSProvider) Synthesize(ctx context.Context, text string) (io.ReadCloser, error) {
 	logger.DebugCF("voice-tts", "Starting TTS synthesis", map[string]any{"text_len": len(text)})
 
+	responseFormat := t.responseFormat
+	resp, err := t.doSpeechRequest(ctx, text, responseFormat)
+	if err != nil {
+		var apiErr *openAITTSAPIError
+		if errors.As(err, &apiErr) && shouldRetryWithoutResponseFormat(apiErr.body) {
+			logger.InfoCF("voice-tts", "Retrying TTS without response_format after provider rejection", map[string]any{
+				"model": t.model,
+			})
+			responseFormat = ""
+			resp, err = t.doSpeechRequest(ctx, text, responseFormat)
+		}
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	fileExt, contentType := audioFileMetaForResponseFormat(responseFormat)
+	return &openAITTSAudioStream{
+		ReadCloser:  resp.Body,
+		fileExt:     fileExt,
+		contentType: contentType,
+	}, nil
+}
+
+func (t *OpenAITTSProvider) doSpeechRequest(
+	ctx context.Context,
+	text string,
+	responseFormat string,
+) (*http.Response, error) {
 	reqBody := map[string]any{
-		"model":           t.model,
+		"model": t.model,
-		"input":           text,
+		"input": text,
-		"voice":           t.voice,
+		"voice": t.voice,
-		"response_format": "opus",
+	}
+	if responseFormat != "" {
+		reqBody["response_format"] = responseFormat
 	}
 
 	jsonData, err := json.Marshal(reqBody)
@@ -119,8 +197,30 @@ func (t *OpenAITTSProvider) Synthesize(ctx context.Context, text string) (io.Rea
 	if resp.StatusCode != http.StatusOK {
 		defer resp.Body.Close()
 		body, _ := io.ReadAll(resp.Body)
-		return nil, fmt.Errorf("API error (status %d): %s", resp.StatusCode, string(body))
+		return nil, &openAITTSAPIError{
+			statusCode: resp.StatusCode,
+			body:       string(body),
+		}
 	}
 
-	return resp.Body, nil
+	return resp, nil
+}
+
+func shouldRetryWithoutResponseFormat(body string) bool {
+	lower := strings.ToLower(body)
+	if !strings.Contains(lower, "response_format") {
+		return false
+	}
+	return strings.Contains(lower, "invalid") || strings.Contains(lower, "unsupported")
+}
+
+func audioFileMetaForResponseFormat(responseFormat string) (string, string) {
+	switch strings.ToLower(strings.TrimSpace(responseFormat)) {
+	case "", "mp3":
+		return ".mp3", "audio/mpeg"
+	case "wav":
+		return ".wav", "audio/wav"
+	default:
+		return ".ogg", "audio/ogg"
+	}
 }

pkg/audio/tts/tts.go

@@ -19,6 +19,10 @@ type TTSProvider interface {
 	Synthesize(ctx context.Context, text string) (io.ReadCloser, error)
 }
 
+type ttsAudioMetaProvider interface {
+	AudioFileMeta() (fileExt string, contentType string)
+}
+
 func providerFromModelConfig(mc *config.ModelConfig) TTSProvider {
 	if mc == nil || mc.APIKey() == "" {
 		return nil
@@ -33,10 +37,31 @@ func providerFromModelConfig(mc *config.ModelConfig) TTSProvider {
 	case "mimo":
 		return NewMimoTTSProvider(mc.APIKey(), providers.ResolveAPIBase(mc), modelID, mc.Proxy)
 	default:
-		return NewOpenAITTSProvider(mc.APIKey(), providers.ResolveAPIBase(mc), mc.Proxy, modelID)
+		return NewOpenAITTSProviderWithOptions(
+			mc.APIKey(),
+			providers.ResolveAPIBase(mc),
+			mc.Proxy,
+			modelID,
+			openAITTSOptionsFromModelConfig(mc),
+		)
 	}
 }
 
+func openAITTSOptionsFromModelConfig(mc *config.ModelConfig) OpenAITTSOptions {
+	options := OpenAITTSOptions{}
+	if mc == nil || mc.ExtraBody == nil {
+		return options
+	}
+
+	if voice, ok := mc.ExtraBody["voice"].(string); ok {
+		options.Voice = strings.TrimSpace(voice)
+	}
+	if responseFormat, ok := mc.ExtraBody["response_format"].(string); ok {
+		options.ResponseFormat = strings.TrimSpace(responseFormat)
+	}
+	return options
+}
+
 func DetectTTS(cfg *config.Config) TTSProvider {
 	if cfg == nil {
 		return nil
@@ -100,6 +125,12 @@ func SynthesizeAndStore(
 		fileExt = ".mp3"
 		contentType = "audio/mpeg"
 	}
+	if metaProvider, ok := stream.(ttsAudioMetaProvider); ok {
+		if ext, ct := metaProvider.AudioFileMeta(); ext != "" && ct != "" {
+			fileExt = ext
+			contentType = ct
+		}
+	}
 
 	file, err := os.CreateTemp(media.TempDir(), "tts-*"+fileExt)
 	if err != nil {

pkg/audio/tts/tts_test.go

@@ -132,6 +132,64 @@ func TestOpenAITTSProvider_SynthesizeNon200(t *testing.T) {
 	}
 }
 
+func TestOpenAITTSProvider_SynthesizeRetriesWithoutResponseFormat(t *testing.T) {
+	var requestBodies []map[string]any
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		bodyBytes, _ := io.ReadAll(r.Body)
+		_ = r.Body.Close()
+
+		var body map[string]any
+		_ = json.Unmarshal(bodyBytes, &body)
+		requestBodies = append(requestBodies, body)
+
+		if len(requestBodies) == 1 {
+			w.WriteHeader(http.StatusBadRequest)
+			_, _ = w.Write([]byte(`{"error":"response_format is invalid"}`))
+			return
+		}
+
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte("audio-bytes"))
+	}))
+	defer server.Close()
+
+	provider := NewOpenAITTSProvider("k123", server.URL, "", "x-ai/grok-voice-tts-1.0")
+	stream, err := provider.Synthesize(context.Background(), "hello")
+	if err != nil {
+		t.Fatalf("Synthesize failed: %v", err)
+	}
+	defer stream.Close()
+
+	data, err := io.ReadAll(stream)
+	if err != nil {
+		t.Fatalf("read stream failed: %v", err)
+	}
+	if string(data) != "audio-bytes" {
+		t.Fatalf("response body mismatch: got %q", string(data))
+	}
+
+	if len(requestBodies) != 2 {
+		t.Fatalf("request count mismatch: got %d, want 2", len(requestBodies))
+	}
+	if requestBodies[0]["response_format"] != "opus" {
+		t.Fatalf("first request should include opus response_format, got %#v", requestBodies[0]["response_format"])
+	}
+	if _, ok := requestBodies[1]["response_format"]; ok {
+		t.Fatalf("second request should omit response_format, got %#v", requestBodies[1]["response_format"])
+	}
+
+	metaStream, ok := stream.(interface {
+		AudioFileMeta() (string, string)
+	})
+	if !ok {
+		t.Fatal("stream does not expose audio metadata")
+	}
+	fileExt, contentType := metaStream.AudioFileMeta()
+	if fileExt != ".mp3" || contentType != "audio/mpeg" {
+		t.Fatalf("audio metadata mismatch: got (%q, %q)", fileExt, contentType)
+	}
+}
+
 func TestNewOpenAITTSProvider_UsesConfiguredModel(t *testing.T) {
 	t.Parallel()
 
@@ -144,6 +202,27 @@ func TestNewOpenAITTSProvider_UsesConfiguredModel(t *testing.T) {
 	}
 }
 
+func TestNewOpenAITTSProvider_UsesConfiguredVoiceAndResponseFormat(t *testing.T) {
+	t.Parallel()
+
+	provider := NewOpenAITTSProviderWithOptions(
+		"key",
+		"https://openrouter.ai/api/v1",
+		"",
+		"microsoft/mai-voice-2",
+		OpenAITTSOptions{
+			Voice:          "en-US-Harper:MAI-Voice-2",
+			ResponseFormat: "mp3",
+		},
+	)
+	if provider.voice != "en-US-Harper:MAI-Voice-2" {
+		t.Fatalf("voice mismatch: got %q", provider.voice)
+	}
+	if provider.responseFormat != "mp3" {
+		t.Fatalf("responseFormat mismatch: got %q", provider.responseFormat)
+	}
+}
+
 func TestDetectTTS_UsesMimoProviderForMimoModels(t *testing.T) {
 	t.Parallel()
 
@@ -170,6 +249,36 @@ func TestDetectTTS_UsesMimoProviderForMimoModels(t *testing.T) {
 	}
 }
 
+func TestDetectTTS_UsesOpenAIExtraBodyVoiceAndResponseFormat(t *testing.T) {
+	t.Parallel()
+
+	provider := DetectTTS(&config.Config{
+		Voice: config.VoiceConfig{TTSModelName: "mai-voice"},
+		ModelList: []*config.ModelConfig{
+			{
+				ModelName: "mai-voice",
+				Model:     "openrouter/microsoft/mai-voice-2",
+				APIKeys:   config.SimpleSecureStrings("sk-openrouter"),
+				ExtraBody: map[string]any{
+					"voice":           "en-US-Harper:MAI-Voice-2",
+					"response_format": "mp3",
+				},
+			},
+		},
+	})
+
+	ttsProvider, ok := provider.(*OpenAITTSProvider)
+	if !ok {
+		t.Fatalf("DetectTTS() type = %T, want *OpenAITTSProvider", provider)
+	}
+	if ttsProvider.voice != "en-US-Harper:MAI-Voice-2" {
+		t.Fatalf("voice mismatch: got %q", ttsProvider.voice)
+	}
+	if ttsProvider.responseFormat != "mp3" {
+		t.Fatalf("responseFormat mismatch: got %q", ttsProvider.responseFormat)
+	}
+}
+
 type stubTTSProvider struct {
 	name string
 }
@@ -245,3 +354,51 @@ func TestSynthesizeAndStore_UsesMp3MetadataForMimo(t *testing.T) {
 		t.Fatalf("filename extension = %q, want %q", filepath.Ext(meta.Filename), ".mp3")
 	}
 }
+
+func TestSynthesizeAndStore_UsesStreamProvidedAudioMetadata(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		bodyBytes, _ := io.ReadAll(r.Body)
+		_ = r.Body.Close()
+
+		var body map[string]any
+		_ = json.Unmarshal(bodyBytes, &body)
+		if body["response_format"] == "opus" {
+			w.WriteHeader(http.StatusBadRequest)
+			_, _ = w.Write([]byte(`{"error":"response_format is invalid"}`))
+			return
+		}
+
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte("mp3-audio"))
+	}))
+	defer server.Close()
+
+	store := media.NewFileMediaStore()
+	provider := NewOpenAITTSProvider("k123", server.URL, "", "x-ai/grok-voice-tts-1.0")
+	ref, err := SynthesizeAndStore(
+		context.Background(),
+		provider,
+		store,
+		"hello",
+		"",
+		"telegram",
+		"chat123",
+	)
+	if err != nil {
+		t.Fatalf("SynthesizeAndStore failed: %v", err)
+	}
+
+	path, meta, err := store.ResolveWithMeta(ref)
+	if err != nil {
+		t.Fatalf("ResolveWithMeta failed: %v", err)
+	}
+	if meta.ContentType != "audio/mpeg" {
+		t.Fatalf("ContentType = %q, want %q", meta.ContentType, "audio/mpeg")
+	}
+	if filepath.Ext(path) != ".mp3" {
+		t.Fatalf("stored file extension = %q, want %q", filepath.Ext(path), ".mp3")
+	}
+	if filepath.Ext(meta.Filename) != ".mp3" {
+		t.Fatalf("filename extension = %q, want %q", filepath.Ext(meta.Filename), ".mp3")
+	}
+}

workspace/skills/picoclaw-agent/SKILL.md

@@ -468,7 +468,8 @@ Recommended setup:
 1. add a TTS-capable model entry to `model_list`
 2. set `voice.tts_model_name` to that entry's `model_name`
 3. store the API key in `.security.yml`
-4. enable `send_tts` in the tool configuration if you want the agent to emit speech files
+4. if the provider needs model-specific TTS fields, add them under `model_list[].extra_body`
+5. enable `send_tts` in the tool configuration if you want the agent to emit speech files
 
 Example:
 
@@ -493,6 +494,35 @@ model_list:
       - "sk-openai-your-key"
 ```
 
+Example with OpenRouter MAI Voice 2:
+
+```json
+{
+  "model_list": [
+    {
+      "model_name": "mai-voice-2",
+      "provider": "openrouter",
+      "model": "microsoft/mai-voice-2",
+      "api_base": "https://openrouter.ai/api/v1",
+      "extra_body": {
+        "voice": "en-US-Harper:MAI-Voice-2",
+        "response_format": "mp3"
+      }
+    }
+  ],
+  "voice": {
+    "tts_model_name": "mai-voice-2"
+  }
+}
+```
+
+```yaml
+model_list:
+  mai-voice-2:
+    api_keys:
+      - "sk-or-your-openrouter-key"
+```
+
 ### Current TTS Provider Paths
 
 | Provider path | Example model | Notes |
@@ -505,6 +535,8 @@ Operational notes:
 - the preferred selection path is `voice.tts_model_name`
 - if that is missing, PicoClaw can still scan `model_list` for the first API-backed model whose ID contains `tts`
 - the current OpenAI-style TTS request defaults to `voice: alloy` and `response_format: opus`
+- you can override `voice` and `response_format` for a specific TTS model through `model_list[].extra_body`
+- if a provider rejects `response_format`, PicoClaw retries once without that field
 - `send_tts` is only registered when TTS detection succeeds