feat(audio): add smart routing for native vs transcribed audio

- Create capabilities.ts with supportsAudioInput() detection
- Gemini, OpenAI, and GitHub Models get native audio passthrough
- Anthropic, Bedrock, Ollama, llama.cpp fall back to Whisper transcription
- routing.ts now checks model capability before deciding to transcribe
- Audio attachments are stripped for non-native models (only transcript text passed)
- Remove deprecated audioConfig from createMessageRouter deps (read from config.audio)

src/models/capabilities.ts

@@ -0,0 +1,46 @@
+/**
+ * Model capability detection for native audio input support.
+ * 
+ * Models that support native audio will receive raw audio data directly.
+ * Models that don't will receive a Whisper transcript as text instead.
+ */
+
+export type ModelProvider = 'anthropic' | 'openai' | 'gemini' | 'bedrock' | 'github' | 'ollama' | 'llamacpp' | 'openrouter' | 'zhipuai' | 'xai';
+
+/** Providers that support native audio input in their API. */
+const AUDIO_CAPABLE_PROVIDERS = new Set<string>([
+  'gemini',
+  'openai',
+  'github',  // GitHub Models uses OpenAI-compatible API
+]);
+
+/**
+ * Models known NOT to support audio despite their provider supporting it.
+ * For example, older OpenAI models or specialized models.
+ */
+const AUDIO_INCAPABLE_MODELS = new Set<string>([
+  // Older OpenAI models that predate audio input support
+  'gpt-3.5-turbo',
+  'gpt-4',
+  'gpt-4-turbo',
+]);
+
+/**
+ * Check whether a provider+model combination supports native audio input.
+ * 
+ * Returns true if the model can receive raw audio data directly via its API,
+ * false if audio must be transcribed to text before sending.
+ */
+export function supportsAudioInput(provider: string, model: string): boolean {
+  // Provider must be in the capable set
+  if (!AUDIO_CAPABLE_PROVIDERS.has(provider)) {
+    return false;
+  }
+
+  // Check model-specific exclusions
+  if (AUDIO_INCAPABLE_MODELS.has(model)) {
+    return false;
+  }
+
+  return true;
+}

src/models/index.ts

@@ -8,6 +8,7 @@ export { LlamaCppClient, type LlamaCppClientConfig } from './local/index.js';
 export { ModelRouter, type ModelRouterConfig, type ModelTier } from './router.js';
 export { withRetry, isRetryable, DEFAULT_RETRY_CONFIG, type RetryConfig } from './retry.js';
 export { estimateCost, MODEL_COSTS_PER_MILLION } from './costs.js';
+export { supportsAudioInput } from './capabilities.js';
 export {
   isSupportedImage,
   isSupportedAudio,