@flutchai/flutch-sdk 0.6.5 → 0.6.6

package/dist/index.cjs

@@ -7113,6 +7113,48 @@ function modelAcceptsSamplingParams(modelIdentifier) {
   return !atOrAboveThreshold;
 }
 __name(modelAcceptsSamplingParams, "modelAcceptsSamplingParams");
+var PROMPT_CACHE_MIN_VERSIONS = {
+  opus: {
+    major: 4,
+    minor: 0
+  },
+  sonnet: {
+    major: 4,
+    minor: 0
+  },
+  haiku: {
+    major: 4,
+    minor: 5
+  }
+};
+function modelSupportsPromptCache(modelIdentifier) {
+  if (!modelIdentifier) return false;
+  const match = CLAUDE_FAMILY_VERSION.exec(modelIdentifier.toLowerCase());
+  if (!match) return false;
+  const [, family, majorRaw, minorRaw] = match;
+  const threshold = PROMPT_CACHE_MIN_VERSIONS[family];
+  if (!threshold) return false;
+  const major = Number(majorRaw);
+  const minor = minorRaw === void 0 ? 0 : Number(minorRaw);
+  return major > threshold.major || major === threshold.major && minor >= threshold.minor;
+}
+__name(modelSupportsPromptCache, "modelSupportsPromptCache");
+function resolvePromptCacheControl(modelIdentifiers, promptCache, ttl) {
+  const enabled = promptCache ?? modelIdentifiers.some((id) => modelSupportsPromptCache(id));
+  return enabled ? {
+    type: "ephemeral",
+    ttl: ttl ?? "5m"
+  } : void 0;
+}
+__name(resolvePromptCacheControl, "resolvePromptCacheControl");
+function buildEffortRequestFields(effort) {
+  return effort ? {
+    output_config: {
+      effort
+    }
+  } : void 0;
+}
+__name(buildEffortRequestFields, "buildEffortRequestFields");
 function resolveSamplingParams(modelIdentifiers, requested, supportsSamplingParams) {
   const accepted = supportsSamplingParams ?? modelIdentifiers.every((id) => modelAcceptsSamplingParams(id));
   if (accepted) {
@@ -7443,7 +7485,9 @@ var ModelInitializer = class _ModelInitializer {
   generateModelCacheKey(config) {
     const base = generateModelCacheKey(config.modelId, config.temperature, config.maxTokens, config.toolsConfig, config.baseURL);
     const withMcp = config.mcpServers && config.mcpServers.length > 0 ? `${base}:mcp:${JSON.stringify(config.mcpServers)}` : base;
-    return config.supportsSamplingParams === void 0 ? withMcp : `${withMcp}:sp:${config.supportsSamplingParams}`;
+    const withSampling = config.supportsSamplingParams === void 0 ? withMcp : `${withMcp}:sp:${config.supportsSamplingParams}`;
+    const withCache = config.promptCache === void 0 && config.promptCacheTtl === void 0 ? withSampling : `${withSampling}:pc:${config.promptCache ?? "auto"}:${config.promptCacheTtl ?? "default"}`;
+    return config.effort === void 0 ? withCache : `${withCache}:eff:${config.effort}`;
   }
   // Chat model creators
   chatModelCreators = {
@@ -7458,7 +7502,7 @@ var ModelInitializer = class _ModelInitializer {
       }
       return new openai.ChatOpenAI(config);
     },
-    [ModelProvider.ANTHROPIC]: ({ modelName, defaultTemperature, defaultMaxTokens, apiToken, baseURL }) => {
+    [ModelProvider.ANTHROPIC]: ({ modelName, defaultTemperature, defaultMaxTokens, apiToken, baseURL }, promptCacheControl) => {
       const routerURL = resolveRouterURL(baseURL);
       return new anthropic.ChatAnthropic({
         modelName,
@@ -7468,6 +7512,13 @@ var ModelInitializer = class _ModelInitializer {
           temperature: defaultTemperature
         },
         maxTokens: defaultMaxTokens,
+        // Top-level prompt caching: one breakpoint on the last cacheable block,
+        // advanced automatically as the conversation grows. Survives the router
+        // (which forwards the Anthropic body untouched) and is accepted by
+        // Bedrock's InvokeModel schema behind it.
+        ...promptCacheControl && {
+          cache_control: promptCacheControl
+        },
         anthropicApiKey: apiToken || this.resolveApiKey(ModelProvider.ANTHROPIC),
         ...routerURL && {
           anthropicApiUrl: routerURL,
@@ -7672,9 +7723,17 @@ var ModelInitializer = class _ModelInitializer {
       defaultMaxTokens: toOptionalNumber(config.maxTokens) ?? toOptionalNumber(modelConfig.defaultMaxTokens),
       baseURL: config.baseURL ?? modelConfig.baseURL
     };
+    const promptCacheControl = resolvePromptCacheControl([
+      finalConfig.bedrockModelId,
+      finalConfig.modelName
+    ], config.promptCache ?? modelConfig.promptCache, config.promptCacheTtl ?? modelConfig.promptCacheTtl);
+    if (promptCacheControl) {
+      this.logger.debug(`Prompt caching enabled for ${finalConfig.bedrockModelId ?? finalConfig.modelName} (ttl=${promptCacheControl.ttl})`);
+    }
     this.logger.debug(`Creating new chat model instance: ${cacheKey}`);
     let model;
     if (finalConfig.useBedrock && finalConfig.bedrockModelId) {
+      const effortFields = buildEffortRequestFields(config.effort ?? modelConfig.defaultEffort);
       model = new FlutchChatBedrockConverse({
         model: finalConfig.bedrockModelId,
         region: this.resolveBedrockRegion(),
@@ -7682,6 +7741,9 @@ var ModelInitializer = class _ModelInitializer {
         ...finalConfig.defaultTemperature !== void 0 && {
           temperature: finalConfig.defaultTemperature
         },
+        ...effortFields && {
+          additionalModelRequestFields: effortFields
+        },
         maxTokens: finalConfig.defaultMaxTokens,
         streaming: true
       });
@@ -7690,14 +7752,24 @@ var ModelInitializer = class _ModelInitializer {
       if (!creator) {
         throw new Error(`Chat models not supported for provider: ${modelConfig.provider}`);
       }
-      model = creator(finalConfig);
+      model = creator(finalConfig, promptCacheControl);
     }
     model.metadata = {
       ...model.metadata,
       modelId: config.modelId
     };
     if (config.toolsConfig || config.customTools || config.mcpServers && config.mcpServers.length > 0) {
-      const boundModel = await this.bindToolsToModel(model, config.toolsConfig, config.customTools, config.mcpServers, config.mcpContext);
+      const boundModel = await this.bindToolsToModel(
+        model,
+        config.toolsConfig,
+        config.customTools,
+        config.mcpServers,
+        config.mcpContext,
+        // Converse takes cache control as a per-call option; the Anthropic
+        // client already got it as a constructor field, and handing it the
+        // same value twice would put an unknown key in its call options.
+        finalConfig.useBedrock ? promptCacheControl : void 0
+      );
       this.modelInstanceCache.set(cacheKey, boundModel);
       return boundModel;
     }
@@ -7709,11 +7781,20 @@ var ModelInitializer = class _ModelInitializer {
   * For toolsConfig: fetch tool executors from MCP Runtime
   * For customTools: use as-is (already prepared DynamicStructuredTool)
   *
+  * `promptCacheControl` is the **Bedrock-only** half of prompt caching: it is
+  * bound alongside the tools as a default call option, and `ChatBedrockConverse`
+  * turns it into `cachePoint` blocks after the tool schemas, the system prompt
+  * and the last message. Tool schemas dominate the request, so this is the one
+  * place worth caching — a model with no tools keeps its plain `BaseChatModel`
+  * type and no cache points (callers there rely on `withStructuredOutput` and
+  * friends). The Anthropic client takes its cache settings in the constructor
+  * instead and must not be handed them again here.
+  *
   * Returns:
   * - Runnable when tools are bound (model.bindTools returns Runnable)
   * - BaseChatModel when no tools
   */
-  async bindToolsToModel(model, toolsConfig, customTools, mcpServers, mcpContext) {
+  async bindToolsToModel(model, toolsConfig, customTools, mcpServers, mcpContext, promptCacheControl) {
     const allTools = [];
     const enabledToolsConfig = (toolsConfig || []).filter((tc) => tc.enabled !== false);
     const hasInlineServers = !!(mcpServers && mcpServers.length > 0);
@@ -7734,7 +7815,10 @@ var ModelInitializer = class _ModelInitializer {
     }
     if (allTools.length > 0) {
       this.logger.debug(`Binding ${allTools.length} tools to model`);
-      const modelWithTools = model.bindTools(allTools);
+      const boundOptions = {
+        cache_control: promptCacheControl
+      };
+      const modelWithTools = promptCacheControl ? model.bindTools(allTools, boundOptions) : model.bindTools(allTools);
       return modelWithTools;
     }
     return model;
@@ -8553,6 +8637,7 @@ exports.ModelProvider = ModelProvider;
 exports.ModelType = ModelType;
 exports.MongoTokenStore = MongoTokenStore;
 exports.OAuthTokenManager = OAuthTokenManager;
+exports.PROMPT_CACHE_MIN_VERSIONS = PROMPT_CACHE_MIN_VERSIONS;
 exports.RetrieverSearchType = RetrieverSearchType;
 exports.SAMPLING_PARAM_FREE_THRESHOLDS = SAMPLING_PARAM_FREE_THRESHOLDS;
 exports.StreamChannel = StreamChannel;
@@ -8564,6 +8649,7 @@ exports.WithEndpoints = WithEndpoints;
 exports.WithUIEndpoints = WithUIEndpoints;
 exports._internals = _internals;
 exports.bootstrap = bootstrap;
+exports.buildEffortRequestFields = buildEffortRequestFields;
 exports.buildOAuthAuthorizationUrl = buildOAuthAuthorizationUrl;
 exports.buildOpenAIModelConfig = buildOpenAIModelConfig;
 exports.clearAttachmentDataStore = clearAttachmentDataStore;
@@ -8599,10 +8685,12 @@ exports.isReasoningModel = isReasoningModel;
 exports.isSendableReasoningBlock = isSendableReasoningBlock;
 exports.loadOAuthProviders = loadOAuthProviders;
 exports.modelAcceptsSamplingParams = modelAcceptsSamplingParams;
+exports.modelSupportsPromptCache = modelSupportsPromptCache;
 exports.normalizeToolConfigs = normalizeToolConfigs;
 exports.registerFinanceExampleCallback = registerFinanceExampleCallback;
 exports.registerUIEndpointsFromClass = registerUIEndpointsFromClass;
 exports.resolveOAuthProviderConfig = resolveOAuthProviderConfig;
+exports.resolvePromptCacheControl = resolvePromptCacheControl;
 exports.resolveRouterURL = resolveRouterURL;
 exports.resolveSamplingParams = resolveSamplingParams;
 exports.sanitizeMessageReasoning = sanitizeMessageReasoning;