@animalabs/membrane 0.5.19 → 0.5.21

package/src/membrane.ts

@@ -233,7 +233,7 @@ export class Membrane {
     // Initialize parser from formatter for format-specific tracking
     const parser = formatter.createStreamParser();
     let toolDepth = 0;
-    let totalUsage: BasicUsage = { inputTokens: 0, outputTokens: 0 };
+    let totalUsage: DetailedUsage = { inputTokens: 0, outputTokens: 0 };
     const contentBlocks: ContentBlock[] = [];
     let lastStopReason: StopReason = 'end_turn';
     let rawRequest: unknown;
@@ -352,9 +352,15 @@ export class Membrane {
 
         lastStopReason = this.mapStopReason(streamResult.stopReason);
 
-        // Accumulate usage
+        // Accumulate usage (including cache metrics)
         totalUsage.inputTokens += streamResult.usage.inputTokens;
         totalUsage.outputTokens += streamResult.usage.outputTokens;
+        if (streamResult.usage.cacheCreationTokens) {
+          totalUsage.cacheCreationTokens = (totalUsage.cacheCreationTokens ?? 0) + streamResult.usage.cacheCreationTokens;
+        }
+        if (streamResult.usage.cacheReadTokens) {
+          totalUsage.cacheReadTokens = (totalUsage.cacheReadTokens ?? 0) + streamResult.usage.cacheReadTokens;
+        }
         onUsage?.(totalUsage);
 
         // Flush the parser to complete any in-progress streaming block
@@ -649,7 +655,7 @@ export class Membrane {
     } = options;
 
     let toolDepth = 0;
-    let totalUsage: BasicUsage = { inputTokens: 0, outputTokens: 0 };
+    let totalUsage: DetailedUsage = { inputTokens: 0, outputTokens: 0 };
     let lastStopReason: StopReason = 'end_turn';
     let rawRequest: unknown;
     let rawResponse: unknown;
@@ -709,9 +715,15 @@ export class Membrane {
 
         lastStopReason = this.mapStopReason(streamResult.stopReason);
 
-        // Accumulate usage
+        // Accumulate usage (including cache metrics)
         totalUsage.inputTokens += streamResult.usage.inputTokens;
         totalUsage.outputTokens += streamResult.usage.outputTokens;
+        if (streamResult.usage.cacheCreationTokens) {
+          totalUsage.cacheCreationTokens = (totalUsage.cacheCreationTokens ?? 0) + streamResult.usage.cacheCreationTokens;
+        }
+        if (streamResult.usage.cacheReadTokens) {
+          totalUsage.cacheReadTokens = (totalUsage.cacheReadTokens ?? 0) + streamResult.usage.cacheReadTokens;
+        }
         onUsage?.(totalUsage);
 
         // Parse content blocks from response
@@ -822,9 +834,9 @@ export class Membrane {
           },
           cache: {
             markersInRequest: 0,
-            tokensCreated: 0,
-            tokensRead: 0,
-            hitRatio: 0,
+            tokensCreated: totalUsage.cacheCreationTokens ?? 0,
+            tokensRead: totalUsage.cacheReadTokens ?? 0,
+            hitRatio: this.calculateCacheHitRatio(totalUsage),
           },
         },
         raw: {
@@ -1003,12 +1015,17 @@ export class Membrane {
       cacheTtl: request.cacheTtl,
       additionalStopSequences,
       maxParticipantsForStop,
+      contextPrefix: request.contextPrefix,
     });
 
     const providerRequest = {
       model: request.config.model,
       maxTokens: request.config.maxTokens,
       temperature: request.config.temperature,
+      topP: request.config.topP,
+      topK: request.config.topK,
+      presencePenalty: request.config.presencePenalty,
+      frequencyPenalty: request.config.frequencyPenalty,
       messages: buildResult.messages,
       system: buildResult.systemContent,
       stopSequences: buildResult.stopSequences,
@@ -1272,7 +1289,7 @@ export class Membrane {
     accumulated: string,
     contentBlocks: ContentBlock[],
     stopReason: StopReason,
-    usage: BasicUsage,
+    usage: DetailedUsage,
     request: NormalizedRequest,
     prefillResult: {
       cacheMarkersApplied?: number;
@@ -1334,10 +1351,10 @@ export class Membrane {
           provider: this.adapter.name,
         },
         cache: {
-          markersInRequest: 0,
-          tokensCreated: 0,
-          tokensRead: 0,
-          hitRatio: 0,
+          markersInRequest: prefillResult.cacheMarkersApplied ?? 0,
+          tokensCreated: usage.cacheCreationTokens ?? 0,
+          tokensRead: usage.cacheReadTokens ?? 0,
+          hitRatio: this.calculateCacheHitRatio(usage),
         },
       },
       raw: {

package/src/providers/anthropic.ts

@@ -141,7 +141,15 @@ export class AnthropicAdapter implements ProviderAdapter {
     if (request.temperature !== undefined) {
       params.temperature = request.temperature;
     }
-    
+
+    if (request.topP !== undefined) {
+      params.top_p = request.topP;
+    }
+
+    if (request.topK !== undefined) {
+      params.top_k = request.topK;
+    }
+
     if (request.stopSequences && request.stopSequences.length > 0) {
       params.stop_sequences = request.stopSequences;
     }

package/src/providers/bedrock.ts

@@ -58,6 +58,8 @@ interface BedrockMessageRequest {
   }>;
   system?: string | Array<{ type: 'text'; text: string; cache_control?: { type: 'ephemeral' } }>;
   temperature?: number;
+  top_p?: number;
+  top_k?: number;
   stop_sequences?: string[];
   tools?: unknown[];
   thinking?: { type: 'enabled'; budget_tokens: number };
@@ -343,6 +345,14 @@ export class BedrockAdapter implements ProviderAdapter {
       params.temperature = request.temperature;
     }
 
+    if (request.topP !== undefined) {
+      params.top_p = request.topP;
+    }
+
+    if (request.topK !== undefined) {
+      params.top_k = request.topK;
+    }
+
     if (request.stopSequences && request.stopSequences.length > 0) {
       params.stop_sequences = request.stopSequences;
     }

package/src/providers/gemini.ts

@@ -52,6 +52,7 @@ interface GeminiRequest {
     maxOutputTokens?: number;
     temperature?: number;
     topP?: number;
+    topK?: number;
     stopSequences?: string[];
   };
   tools?: { functionDeclarations: GeminiFunctionDeclaration[] }[];
@@ -336,6 +337,14 @@ export class GeminiAdapter implements ProviderAdapter {
       geminiRequest.generationConfig.temperature = request.temperature;
     }
 
+    if (request.topP !== undefined) {
+      geminiRequest.generationConfig.topP = request.topP;
+    }
+
+    if (request.topK !== undefined) {
+      geminiRequest.generationConfig.topK = request.topK;
+    }
+
     if (request.stopSequences && request.stopSequences.length > 0) {
       // Gemini API limits stop sequences to 5
       geminiRequest.generationConfig.stopSequences = request.stopSequences.slice(0, 5);

package/src/providers/openai-compatible.ts

@@ -278,7 +278,19 @@ export class OpenAICompatibleAdapter implements ProviderAdapter {
     if (request.temperature !== undefined) {
       params.temperature = request.temperature;
     }
-    
+
+    if (request.topP !== undefined) {
+      params.top_p = request.topP;
+    }
+
+    if (request.presencePenalty !== undefined) {
+      params.presence_penalty = request.presencePenalty;
+    }
+
+    if (request.frequencyPenalty !== undefined) {
+      params.frequency_penalty = request.frequencyPenalty;
+    }
+
     if (request.stopSequences && request.stopSequences.length > 0) {
       params.stop = request.stopSequences;
     }

package/src/providers/openai-completions.ts

@@ -37,6 +37,9 @@ interface CompletionsRequest {
   prompt: string;
   max_tokens?: number;
   temperature?: number;
+  top_p?: number;
+  presence_penalty?: number;
+  frequency_penalty?: number;
   stop?: string[];
   stream?: boolean;
 }
@@ -344,6 +347,18 @@ export class OpenAICompletionsAdapter implements ProviderAdapter {
       params.temperature = request.temperature;
     }
 
+    if (request.topP !== undefined) {
+      params.top_p = request.topP;
+    }
+
+    if (request.presencePenalty !== undefined) {
+      params.presence_penalty = request.presencePenalty;
+    }
+
+    if (request.frequencyPenalty !== undefined) {
+      params.frequency_penalty = request.frequencyPenalty;
+    }
+
     // Generate stop sequences from participant names + EOT token + any extras
     const stopSequences = [
       ...this.generateStopSequences(participants),

package/src/providers/openai.ts

@@ -376,7 +376,19 @@ export class OpenAIAdapter implements ProviderAdapter {
     if (request.temperature !== undefined && !noTemperatureSupport(model)) {
       params.temperature = request.temperature;
     }
-    
+
+    if (request.topP !== undefined && !noTemperatureSupport(model)) {
+      params.top_p = request.topP;
+    }
+
+    if (request.presencePenalty !== undefined) {
+      params.presence_penalty = request.presencePenalty;
+    }
+
+    if (request.frequencyPenalty !== undefined) {
+      params.frequency_penalty = request.frequencyPenalty;
+    }
+
     // Reasoning models (o1, o3, o4) don't support stop sequences
     if (request.stopSequences && request.stopSequences.length > 0 && !noStopSupport(model)) {
       params.stop = request.stopSequences;

package/src/providers/openrouter.ts

@@ -298,7 +298,19 @@ export class OpenRouterAdapter implements ProviderAdapter {
     if (request.temperature !== undefined) {
       params.temperature = request.temperature;
     }
-    
+
+    if (request.topP !== undefined) {
+      params.top_p = request.topP;
+    }
+
+    if (request.presencePenalty !== undefined) {
+      params.presence_penalty = request.presencePenalty;
+    }
+
+    if (request.frequencyPenalty !== undefined) {
+      params.frequency_penalty = request.frequencyPenalty;
+    }
+
     if (request.stopSequences && request.stopSequences.length > 0) {
       params.stop = request.stopSequences;
     }

package/src/types/provider.ts

@@ -202,7 +202,19 @@ export interface ProviderRequest {
   
   /** Temperature */
   temperature?: number;
-  
+
+  /** Top P nucleus sampling */
+  topP?: number;
+
+  /** Top K sampling */
+  topK?: number;
+
+  /** Presence penalty */
+  presencePenalty?: number;
+
+  /** Frequency penalty */
+  frequencyPenalty?: number;
+
   /** Stop sequences */
   stopSequences?: string[];
   

package/src/types/request.ts

@@ -139,7 +139,14 @@ export interface NormalizedRequest {
    * '1h' = 1 hour TTL (extended caching)
    */
   cacheTtl?: '5m' | '1h';
-  
+
+  /**
+   * Context prefix for simulacrum seeding.
+   * Injected as first assistant message (before conversation history).
+   * Cached when promptCaching is enabled.
+   */
+  contextPrefix?: string;
+
   /** Provider-specific parameters (pass-through) */
   providerParams?: Record<string, unknown>;
 }