@animalabs/membrane 0.5.55 → 0.5.63

package/src/types/provider.ts

@@ -215,6 +215,9 @@ export interface ProviderRequest {
   /** Frequency penalty */
   frequencyPenalty?: number;
 
+  /** Repetition penalty (multiplicative, vLLM/HuggingFace style) */
+  repetitionPenalty?: number;
+
   /** Stop sequences */
   stopSequences?: string[];
   
@@ -232,6 +235,13 @@ export interface ProviderRequestOptions {
   idleTimeoutMs?: number;
   /** Called with the raw API request body right before fetch */
   onRequest?: (rawRequest: unknown) => void;
+  /**
+   * Wrap native thinking deltas in <thinking>...</thinking> tags on the
+   * onChunk stream. Used by the XML formatter path so its tag-based parser
+   * tracks thinking blocks; without this, native thinking content streams
+   * indistinguishably from visible text.
+   */
+  wrapThinkingTags?: boolean;
 }
 
 export interface ProviderResponse {

package/src/types/request.ts

@@ -30,11 +30,22 @@ export interface GenerationConfig {
   
   /** Frequency penalty (provider-specific) */
   frequencyPenalty?: number;
-  
+
+  /** Repetition penalty — multiplicative (vLLM/HuggingFace style, typically 1.0-1.2) */
+  repetitionPenalty?: number;
+
   /** Enable thinking/reasoning mode */
   thinking?: {
     enabled: boolean;
     budgetTokens?: number;
+    /** Thinking type for the API: 'enabled' (default, explicit budget) or 'adaptive' (model-managed) */
+    type?: 'enabled' | 'adaptive';
+    /**
+     * Controls how thinking content is returned: 'summarized' (readable summary)
+     * or 'omitted' (empty thinking field, signature only). Models like Fable 5 /
+     * Opus 4.7+ default to 'omitted' — set 'summarized' to receive thinking text.
+     */
+    display?: 'summarized' | 'omitted';
   };
   
   /** Image generation config (Gemini) */