fix: Gemini default model and AAC/AMR audio

docs/providers.md

@@ -7,7 +7,7 @@ not statically link the other vendors' SDKs:
 |---|---|---|---|
 | OpenAI | `pkg/llm/openai` | `openai.New(key, model)` | gpt-4o, gpt-4o-mini, o1, ... |
 | Anthropic | `pkg/llm/anthropic` | `anthropic.New(key, model)` | claude-sonnet, claude-opus, ... |
-| Google Gemini | `pkg/llm/gemini` | `gemini.New(key, model)` | gemini-2.5-flash, gemini-2.5-pro, ... |
+| Google Gemini | `pkg/llm/gemini` | `gemini.New(key, model)` | gemini-3.8-flash (default), gemini-2.5-pro, ... |
 | Vertex AI (Gemini) | `pkg/llm/gemini` | `gemini.NewVertex(project, location, model)` | Vertex-hosted Gemini via ADC |
 | OpenAI-compatible | `pkg/llm/openai` | `openai.NewCompat(key, model, baseURL)` | OpenRouter, Ollama, Groq, vLLM, Together, ... |
 

examples/demo/.env.example

@@ -10,7 +10,7 @@ ANTHROPIC_API_KEY=sk-ant-...
 ANTHROPIC_MODEL=claude-sonnet-4-5
 
 GEMINI_API_KEY=AIza...
-GEMINI_MODEL=gemini-2.5-flash
+GEMINI_MODEL=gemini-3.8-flash
 
 # ── Tools ────────────────────────────────────────────────────────────────────
 TAVILY_API_KEY=tvly-...

pkg/audio/audio.go

@@ -140,6 +140,9 @@ var extByMIME = map[string]string{
 	"audio/opus":      "ogg",
 	"audio/flac":      "flac",
 	"audio/x-flac":    "flac",
+	"audio/aac":       "aac",
+	"audio/x-aac":     "aac",
+	"audio/amr":       "amr",
 }
 
 // Ext returns the file extension for a media type, or "" when the type is not

pkg/audio/audio_test.go

@@ -20,6 +20,9 @@ func TestExtStripsParametersAndCase(t *testing.T) {
 		{"audio/x-m4a", "m4a"},
 		{"audio/flac", "flac"},
 		{"audio/opus", "ogg"},
+		{"audio/aac", "aac"},
+		{"audio/x-aac", "aac"},
+		{"audio/amr", "amr"},
 		{"text/plain", ""},
 		{"", ""},
 	} {

pkg/llm/gemini/default_model_test.go

@@ -0,0 +1,28 @@
+package gemini
+
+import "testing"
+
+// A retired default fails every fresh signup on its first call while old keys
+// keep passing, so pin the value and check every constructor honors it.
+func TestConstructorsUseDefaultModel(t *testing.T) {
+	if defaultModel != "gemini-3.8-flash" {
+		t.Fatalf("defaultModel = %q, want gemini-3.8-flash", defaultModel)
+	}
+	p, err := New("test-key", "")
+	if err != nil {
+		t.Fatalf("New: %v", err)
+	}
+	tr, err := NewTranscriber("test-key", "")
+	if err != nil {
+		t.Fatalf("NewTranscriber: %v", err)
+	}
+	ma, err := NewMediaAnalyzer("test-key", "")
+	if err != nil {
+		t.Fatalf("NewMediaAnalyzer: %v", err)
+	}
+	for name, got := range map[string]string{"New": p.model, "NewTranscriber": tr.model, "NewMediaAnalyzer": ma.model} {
+		if got != defaultModel {
+			t.Fatalf("%s model = %q, want %q", name, got, defaultModel)
+		}
+	}
+}

pkg/llm/gemini/gemini.go

@@ -56,7 +56,13 @@ func WithSeed(n int64) Option {
 	return func(p *Provider) { p.seed = &n }
 }
 
-// New creates a wrapper over Google Gemini's API.
+// defaultModel is the model every constructor in this package uses when
+// given "". gemini-2.5-flash is rejected with a 404 for API keys created after
+// its retirement, so a fresh signup would fail on its first call.
+const defaultModel = "gemini-3.8-flash"
+
+// New creates a wrapper over Google Gemini's API. Empty model defaults to
+// "gemini-3.8-flash".
 func New(apiKey string, model string, opts ...Option) (*Provider, error) {
 	if apiKey == "" {
 		apiKey = os.Getenv("GEMINI_API_KEY")
@@ -65,7 +71,7 @@ func New(apiKey string, model string, opts ...Option) (*Provider, error) {
 		return nil, errors.New("GEMINI_API_KEY is not set in environment")
 	}
 	if model == "" {
-		model = "gemini-2.5-flash"
+		model = defaultModel
 	}
 
 	client, err := genai.NewClient(context.Background(), &genai.ClientConfig{

pkg/llm/gemini/gemini_vision.go

@@ -20,7 +20,7 @@ type MediaAnalyzer struct {
 }
 
 // NewMediaAnalyzer builds an analyzer. apiKey defaults to GEMINI_API_KEY;
-// model defaults to "gemini-2.5-flash".
+// model defaults to "gemini-3.8-flash".
 func NewMediaAnalyzer(apiKey, model string) (*MediaAnalyzer, error) {
 	if apiKey == "" {
 		apiKey = os.Getenv("GEMINI_API_KEY")
@@ -29,7 +29,7 @@ func NewMediaAnalyzer(apiKey, model string) (*MediaAnalyzer, error) {
 		return nil, fmt.Errorf("gemini: GEMINI_API_KEY not set")
 	}
 	if model == "" {
-		model = "gemini-2.5-flash"
+		model = defaultModel
 	}
 	client, err := genai.NewClient(context.Background(), &genai.ClientConfig{
 		APIKey: apiKey,

pkg/llm/gemini/transcriber.go

@@ -47,7 +47,7 @@ type Transcriber struct {
 var _ audio.Transcriber = (*Transcriber)(nil)
 
 // NewTranscriber builds a transcriber. apiKey defaults to GEMINI_API_KEY;
-// model defaults to "gemini-2.5-flash".
+// model defaults to "gemini-3.8-flash".
 func NewTranscriber(apiKey, model string) (*Transcriber, error) {
 	if apiKey == "" {
 		apiKey = os.Getenv("GEMINI_API_KEY")
@@ -56,7 +56,7 @@ func NewTranscriber(apiKey, model string) (*Transcriber, error) {
 		return nil, fmt.Errorf("gemini: NewTranscriber: GEMINI_API_KEY not set")
 	}
 	if model == "" {
-		model = "gemini-2.5-flash"
+		model = defaultModel
 	}
 	client, err := genai.NewClient(context.Background(), &genai.ClientConfig{APIKey: apiKey})
 	if err != nil {

pkg/llm/gemini/vertex.go

@@ -17,7 +17,7 @@ import (
 //
 // Empty projectID / location fall back to GOOGLE_CLOUD_PROJECT and
 // GOOGLE_CLOUD_LOCATION respectively. Empty model defaults to
-// "gemini-2.5-flash".
+// "gemini-3.8-flash".
 //
 // The returned *Provider reuses the same GenerateStream implementation
 // as the public-API backend — only the transport differs.
@@ -35,7 +35,7 @@ func NewVertex(projectID, location, model string, opts ...Option) (*Provider, er
 		location = "us-central1"
 	}
 	if model == "" {
-		model = "gemini-2.5-flash"
+		model = defaultModel
 	}
 
 	client, err := genai.NewClient(context.Background(), &genai.ClientConfig{

pkg/llm/openai/transcriber.go

@@ -16,6 +16,11 @@ import (
 // instead of after transferring the payload.
 const maxClipBytes = 25 << 20
 
+// unsupportedExts are containers pkg/audio accepts (Gemini decodes them) but
+// the OpenAI audio endpoint does not list. Rejecting them here fails fast with
+// ErrUnsupportedFormat instead of after a full upload.
+var unsupportedExts = map[string]bool{"aac": true, "amr": true}
+
 // Transcriber implements audio.Transcriber against OpenAI's audio
 // transcription endpoint.
 //
@@ -52,6 +57,10 @@ func (t *Transcriber) Transcribe(ctx context.Context, clip audio.Clip, opts audi
 	if err := clip.Validate(); err != nil {
 		return audio.Transcript{}, fmt.Errorf("openai: Transcriber: %w", err)
 	}
+	if ext := audio.Ext(clip.MIME); unsupportedExts[ext] {
+		return audio.Transcript{}, fmt.Errorf("openai: Transcriber: %w: %s is not accepted by the OpenAI audio endpoint",
+			audio.ErrUnsupportedFormat, ext)
+	}
 	if len(clip.Data) > maxClipBytes {
 		return audio.Transcript{}, fmt.Errorf("openai: Transcriber: %w: %d bytes exceeds %d",
 			audio.ErrTooLarge, len(clip.Data), maxClipBytes)

pkg/llm/openai/transcriber_test.go

@@ -151,6 +151,8 @@ func TestTranscribeRejectsInvalidClip(t *testing.T) {
 	}{
 		{"empty", audio.Clip{MIME: "audio/wav"}, audio.ErrNoAudio},
 		{"unsupported", audio.Clip{MIME: "application/pdf", Data: []byte("x")}, audio.ErrUnsupportedFormat},
+		{"aac not accepted by endpoint", audio.Clip{MIME: "audio/aac", Data: []byte("x")}, audio.ErrUnsupportedFormat},
+		{"amr not accepted by endpoint", audio.Clip{MIME: "audio/amr", Data: []byte("x")}, audio.ErrUnsupportedFormat},
 	} {
 		t.Run(tc.name, func(t *testing.T) {
 			if _, err := tr.Transcribe(context.Background(), tc.clip, audio.Options{}); !errors.Is(err, tc.want) {