@dexto/core 1.4.0 → 1.5.1

package/dist/llm/errors.js

@@ -24,6 +24,15 @@ class LLMError {
       { provider }
     );
   }
+  static missingConfig(provider, configName) {
+    return new DextoRuntimeError(
+      LLMErrorCode.CONFIG_MISSING,
+      ErrorScope.LLM,
+      ErrorType.USER,
+      `Provider '${provider}' requires ${configName}`,
+      { provider, configName }
+    );
+  }
   static unsupportedProvider(provider) {
     const availableProviders = getSupportedProviders();
     return new DextoRuntimeError(
@@ -34,6 +43,21 @@ class LLMError {
       { provider, availableProviders }
     );
   }
+  /**
+   * Runtime error when API key is missing for a provider that requires it.
+   * This occurs when relaxed validation allowed the app to start without an API key,
+   * and the user then tries to use the LLM functionality.
+   */
+  static apiKeyMissing(provider, envVar) {
+    return new DextoRuntimeError(
+      LLMErrorCode.API_KEY_MISSING,
+      ErrorScope.LLM,
+      ErrorType.USER,
+      `API key required for provider '${provider}'`,
+      { provider, envVar },
+      `Set the ${envVar} environment variable or configure it in Settings`
+    );
+  }
   static modelProviderUnknown(model) {
     const availableProviders = getSupportedProviders();
     return new DextoRuntimeError(

package/dist/llm/executor/provider-options.cjs

@@ -0,0 +1,87 @@
+"use strict";
+var __defProp = Object.defineProperty;
+var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
+var __getOwnPropNames = Object.getOwnPropertyNames;
+var __hasOwnProp = Object.prototype.hasOwnProperty;
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+var __copyProps = (to, from, except, desc) => {
+  if (from && typeof from === "object" || typeof from === "function") {
+    for (let key of __getOwnPropNames(from))
+      if (!__hasOwnProp.call(to, key) && key !== except)
+        __defProp(to, key, { get: () => from[key], enumerable: !(desc = __getOwnPropDesc(from, key)) || desc.enumerable });
+  }
+  return to;
+};
+var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
+var provider_options_exports = {};
+__export(provider_options_exports, {
+  buildProviderOptions: () => buildProviderOptions,
+  getDefaultReasoningEffort: () => getDefaultReasoningEffort
+});
+module.exports = __toCommonJS(provider_options_exports);
+var import_registry = require("../registry.js");
+function buildProviderOptions(config) {
+  const { provider, model, reasoningEffort } = config;
+  const modelLower = model.toLowerCase();
+  if (provider === "anthropic") {
+    return {
+      anthropic: {
+        // Enable prompt caching - saves money and improves latency
+        cacheControl: { type: "ephemeral" },
+        // Stream reasoning/thinking content when model supports it
+        sendReasoning: true
+      }
+    };
+  }
+  if (provider === "bedrock" && modelLower.includes("claude")) {
+    return {
+      bedrock: {
+        cacheControl: { type: "ephemeral" },
+        sendReasoning: true
+      }
+    };
+  }
+  if (provider === "vertex" && modelLower.includes("claude")) {
+    return {
+      "vertex-anthropic": {
+        cacheControl: { type: "ephemeral" },
+        sendReasoning: true
+      }
+    };
+  }
+  if (provider === "google" || provider === "vertex" && !modelLower.includes("claude")) {
+    return {
+      google: {
+        thinkingConfig: {
+          // Include thoughts in the response for transparency
+          includeThoughts: true
+        }
+      }
+    };
+  }
+  if (provider === "openai") {
+    const effectiveEffort = reasoningEffort ?? getDefaultReasoningEffort(model);
+    if (effectiveEffort) {
+      return {
+        openai: {
+          reasoningEffort: effectiveEffort
+        }
+      };
+    }
+  }
+  return void 0;
+}
+function getDefaultReasoningEffort(model) {
+  if ((0, import_registry.isReasoningCapableModel)(model)) {
+    return "medium";
+  }
+  return void 0;
+}
+// Annotate the CommonJS export names for ESM import in node:
+0 && (module.exports = {
+  buildProviderOptions,
+  getDefaultReasoningEffort
+});

package/dist/llm/executor/provider-options.d.ts

@@ -0,0 +1,49 @@
+/**
+ * Provider-specific options builder for Vercel AI SDK's streamText/generateText.
+ *
+ * Centralizes provider-specific configuration that requires explicit opt-in:
+ * - Anthropic: cacheControl for prompt caching, sendReasoning for extended thinking
+ * - Bedrock/Vertex Claude: Same as Anthropic (Claude models on these platforms)
+ * - Google: thinkingConfig for Gemini thinking models
+ * - OpenAI: reasoningEffort for o1/o3/codex/gpt-5 models
+ *
+ * Caching notes:
+ * - Anthropic: Requires explicit cacheControl option (we enable it)
+ * - OpenAI: Automatic for prompts ≥1024 tokens (no config needed)
+ * - Google: Implicit caching automatic for Gemini 2.5+ (≥1024 tokens for Flash,
+ *   ≥2048 for Pro). Explicit caching requires pre-created cachedContent IDs.
+ * All providers return cached token counts in the response (cachedInputTokens).
+ */
+import type { LLMProvider } from '../types.js';
+export type ReasoningEffort = 'none' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh';
+export interface ProviderOptionsConfig {
+    provider: LLMProvider;
+    model: string;
+    reasoningEffort?: ReasoningEffort | undefined;
+}
+/**
+ * Build provider-specific options for streamText/generateText.
+ *
+ * @param config Provider, model, and optional reasoning effort configuration
+ * @returns Provider options object or undefined if no special options needed
+ */
+export declare function buildProviderOptions(config: ProviderOptionsConfig): Record<string, Record<string, unknown>> | undefined;
+/**
+ * Determine the default reasoning effort for OpenAI models.
+ *
+ * OpenAI reasoning effort levels (from lowest to highest):
+ * - 'none': No reasoning, fastest responses
+ * - 'low': Minimal reasoning, fast responses
+ * - 'medium': Balanced reasoning (OpenAI's recommended daily driver)
+ * - 'high': Thorough reasoning for complex tasks
+ * - 'xhigh': Extra high reasoning for quality-critical, non-latency-sensitive tasks
+ *
+ * Default strategy:
+ * - Reasoning-capable models (codex, o1, o3, gpt-5): 'medium' - OpenAI's recommended default
+ * - Other models: undefined (no reasoning effort needed)
+ *
+ * @param model The model name
+ * @returns Reasoning effort level or undefined if not applicable
+ */
+export declare function getDefaultReasoningEffort(model: string): Exclude<ReasoningEffort, 'none'> | undefined;
+//# sourceMappingURL=provider-options.d.ts.map

package/dist/llm/executor/provider-options.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"provider-options.d.ts","sourceRoot":"","sources":["../../../src/llm/executor/provider-options.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AAG/C,MAAM,MAAM,eAAe,GAAG,MAAM,GAAG,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,CAAC;AAEvF,MAAM,WAAW,qBAAqB;IAClC,QAAQ,EAAE,WAAW,CAAC;IACtB,KAAK,EAAE,MAAM,CAAC;IACd,eAAe,CAAC,EAAE,eAAe,GAAG,SAAS,CAAC;CACjD;AAED;;;;;GAKG;AACH,wBAAgB,oBAAoB,CAChC,MAAM,EAAE,qBAAqB,GAC9B,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,GAAG,SAAS,CAgErD;AAED;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,yBAAyB,CACrC,KAAK,EAAE,MAAM,GACd,OAAO,CAAC,eAAe,EAAE,MAAM,CAAC,GAAG,SAAS,CAS9C"}

package/dist/llm/executor/provider-options.js

@@ -0,0 +1,63 @@
+import "../../chunk-PTJYTZNU.js";
+import { isReasoningCapableModel } from "../registry.js";
+function buildProviderOptions(config) {
+  const { provider, model, reasoningEffort } = config;
+  const modelLower = model.toLowerCase();
+  if (provider === "anthropic") {
+    return {
+      anthropic: {
+        // Enable prompt caching - saves money and improves latency
+        cacheControl: { type: "ephemeral" },
+        // Stream reasoning/thinking content when model supports it
+        sendReasoning: true
+      }
+    };
+  }
+  if (provider === "bedrock" && modelLower.includes("claude")) {
+    return {
+      bedrock: {
+        cacheControl: { type: "ephemeral" },
+        sendReasoning: true
+      }
+    };
+  }
+  if (provider === "vertex" && modelLower.includes("claude")) {
+    return {
+      "vertex-anthropic": {
+        cacheControl: { type: "ephemeral" },
+        sendReasoning: true
+      }
+    };
+  }
+  if (provider === "google" || provider === "vertex" && !modelLower.includes("claude")) {
+    return {
+      google: {
+        thinkingConfig: {
+          // Include thoughts in the response for transparency
+          includeThoughts: true
+        }
+      }
+    };
+  }
+  if (provider === "openai") {
+    const effectiveEffort = reasoningEffort ?? getDefaultReasoningEffort(model);
+    if (effectiveEffort) {
+      return {
+        openai: {
+          reasoningEffort: effectiveEffort
+        }
+      };
+    }
+  }
+  return void 0;
+}
+function getDefaultReasoningEffort(model) {
+  if (isReasoningCapableModel(model)) {
+    return "medium";
+  }
+  return void 0;
+}
+export {
+  buildProviderOptions,
+  getDefaultReasoningEffort
+};

package/dist/llm/executor/stream-processor.cjs

@@ -88,22 +88,30 @@ class StreamProcessor {
               });
             }
             break;
-          case "tool-call":
+          case "tool-call": {
             if (!this.assistantMessageId) {
               this.assistantMessageId = await this.createAssistantMessage();
             }
-            await this.contextManager.addToolCall(this.assistantMessageId, {
+            const toolCall = {
               id: event.toolCallId,
               type: "function",
               function: {
                 name: event.toolName,
                 arguments: JSON.stringify(event.input)
               }
-            });
+            };
+            const shouldPersistProviderMetadata = this.config.provider === "google" || this.config.provider === "vertex";
+            if (shouldPersistProviderMetadata && event.providerMetadata) {
+              toolCall.providerOptions = {
+                ...event.providerMetadata
+              };
+            }
+            await this.contextManager.addToolCall(this.assistantMessageId, toolCall);
             this.pendingToolCalls.set(event.toolCallId, {
               toolName: event.toolName
             });
             break;
+          }
           case "tool-result": {
             const rawResult = event.output;
             this.logger.debug("Tool result received", {
@@ -150,8 +158,10 @@ class StreamProcessor {
           }
           case "finish-step":
             if (event.usage) {
-              const cacheWriteTokens = event.providerMetadata?.["anthropic"]?.["cacheCreationInputTokens"] ?? // @ts-expect-error - Bedrock metadata typing not in Vercel SDK
-              event.providerMetadata?.["bedrock"]?.["usage"]?.["cacheWriteInputTokens"] ?? 0;
+              const anthropicMeta = event.providerMetadata?.["anthropic"];
+              const bedrockMeta = event.providerMetadata?.["bedrock"];
+              const cacheWriteTokens = anthropicMeta?.["cacheCreationInputTokens"] ?? bedrockMeta?.usage?.["cacheWriteInputTokens"] ?? 0;
+              const cacheReadTokens = anthropicMeta?.["cacheReadInputTokens"] ?? bedrockMeta?.usage?.["cacheReadInputTokens"] ?? event.usage.cachedInputTokens ?? 0;
               this.actualTokens = {
                 inputTokens: (this.actualTokens.inputTokens ?? 0) + (event.usage.inputTokens ?? 0),
                 outputTokens: (this.actualTokens.outputTokens ?? 0) + (event.usage.outputTokens ?? 0),
@@ -160,16 +170,17 @@ class StreamProcessor {
                   reasoningTokens: (this.actualTokens.reasoningTokens ?? 0) + event.usage.reasoningTokens
                 },
                 // Cache tokens
-                cacheReadTokens: (this.actualTokens.cacheReadTokens ?? 0) + (event.usage.cachedInputTokens ?? 0),
+                cacheReadTokens: (this.actualTokens.cacheReadTokens ?? 0) + cacheReadTokens,
                 cacheWriteTokens: (this.actualTokens.cacheWriteTokens ?? 0) + cacheWriteTokens
               };
             }
             break;
           case "finish": {
             this.finishReason = event.finishReason;
-            const cachedInputTokens = event.totalUsage.cachedInputTokens ?? 0;
+            const cacheReadTokens = this.actualTokens.cacheReadTokens ?? event.totalUsage.cachedInputTokens ?? 0;
+            const cacheWriteTokens = this.actualTokens.cacheWriteTokens ?? 0;
             const providerExcludesCached = this.config.provider === "anthropic" || this.config.provider === "bedrock";
-            const adjustedInputTokens = providerExcludesCached ? event.totalUsage.inputTokens ?? 0 : (event.totalUsage.inputTokens ?? 0) - cachedInputTokens;
+            const adjustedInputTokens = providerExcludesCached ? event.totalUsage.inputTokens ?? 0 : (event.totalUsage.inputTokens ?? 0) - cacheReadTokens;
             const usage = {
               inputTokens: adjustedInputTokens,
               outputTokens: event.totalUsage.outputTokens ?? 0,
@@ -178,9 +189,9 @@ class StreamProcessor {
               ...event.totalUsage.reasoningTokens !== void 0 && {
                 reasoningTokens: event.totalUsage.reasoningTokens
               },
-              // Cache tokens - read from totalUsage, write from accumulated finish-step events
-              cacheReadTokens: cachedInputTokens,
-              cacheWriteTokens: this.actualTokens.cacheWriteTokens ?? 0
+              // Cache tokens from accumulated finish-step events or totalUsage fallback
+              cacheReadTokens,
+              cacheWriteTokens
             };
             this.actualTokens = usage;
             this.logger.info("LLM response complete", {

package/dist/llm/executor/stream-processor.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stream-processor.d.ts","sourceRoot":"","sources":["../../../src/llm/executor/stream-processor.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,gBAAgB,EAAE,OAAO,IAAI,aAAa,EAAE,MAAM,IAAI,CAAC;AAChE,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AAC1D,OAAO,EAAE,eAAe,EAAmB,MAAM,uBAAuB,CAAC;AACzE,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAC;AAE3D,OAAO,EAAE,qBAAqB,EAAE,MAAM,YAAY,CAAC;AAGnD,OAAO,EAAE,YAAY,EAAE,MAAM,0BAA0B,CAAC;AAExD,OAAO,EAAE,WAAW,EAAc,MAAM,aAAa,CAAC;AAEtD,MAAM,WAAW,qBAAqB;IAClC,QAAQ,EAAE,WAAW,CAAC;IACtB,KAAK,EAAE,MAAM,CAAC;CACjB;AAED,qBAAa,eAAe;IAwBpB,OAAO,CAAC,cAAc;IACtB,OAAO,CAAC,QAAQ;IAChB,OAAO,CAAC,eAAe;IACvB,OAAO,CAAC,WAAW;IACnB,OAAO,CAAC,MAAM;IAEd,OAAO,CAAC,SAAS;IACjB,OAAO,CAAC,gBAAgB,CAAC;IA9B7B,OAAO,CAAC,kBAAkB,CAAuB;IACjD,OAAO,CAAC,YAAY,CAAmE;IACvF,OAAO,CAAC,YAAY,CAA8B;IAClD,OAAO,CAAC,aAAa,CAAc;IACnC,OAAO,CAAC,eAAe,CAAc;IACrC,OAAO,CAAC,MAAM,CAAe;IAC7B;;;OAGG;IACH,OAAO,CAAC,gBAAgB,CAAgD;IAExE;;;;;;;;;OASG;gBAES,cAAc,EAAE,cAAc,EAC9B,QAAQ,EAAE,eAAe,EACzB,eAAe,EAAE,eAAe,EAChC,WAAW,EAAE,WAAW,EACxB,MAAM,EAAE,qBAAqB,EACrC,MAAM,EAAE,YAAY,EACZ,SAAS,GAAE,OAAc,EACzB,gBAAgB,CAAC,EAAE,GAAG,CAC1B,MAAM,EACN;QAAE,eAAe,EAAE,OAAO,CAAC;QAAC,cAAc,CAAC,EAAE,UAAU,GAAG,UAAU,CAAA;KAAE,CACzE,YAAA;IAKC,OAAO,CACT,QAAQ,EAAE,MAAM,gBAAgB,CAAC,aAAa,EAAE,OAAO,CAAC,GACzD,OAAO,CAAC,qBAAqB,CAAC;YAmYnB,sBAAsB;YAKtB,gBAAgB;IAO9B;;;;OAIG;YACW,2BAA2B;CAmC5C"}
+{"version":3,"file":"stream-processor.d.ts","sourceRoot":"","sources":["../../../src/llm/executor/stream-processor.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,gBAAgB,EAAE,OAAO,IAAI,aAAa,EAAE,MAAM,IAAI,CAAC;AAChE,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AAC1D,OAAO,EAAE,eAAe,EAAmB,MAAM,uBAAuB,CAAC;AACzE,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAC;AAE3D,OAAO,EAAE,qBAAqB,EAAE,MAAM,YAAY,CAAC;AAGnD,OAAO,EAAE,YAAY,EAAE,MAAM,0BAA0B,CAAC;AAExD,OAAO,EAAE,WAAW,EAAc,MAAM,aAAa,CAAC;AAEtD,MAAM,WAAW,qBAAqB;IAClC,QAAQ,EAAE,WAAW,CAAC;IACtB,KAAK,EAAE,MAAM,CAAC;CACjB;AAED,qBAAa,eAAe;IAwBpB,OAAO,CAAC,cAAc;IACtB,OAAO,CAAC,QAAQ;IAChB,OAAO,CAAC,eAAe;IACvB,OAAO,CAAC,WAAW;IACnB,OAAO,CAAC,MAAM;IAEd,OAAO,CAAC,SAAS;IACjB,OAAO,CAAC,gBAAgB,CAAC;IA9B7B,OAAO,CAAC,kBAAkB,CAAuB;IACjD,OAAO,CAAC,YAAY,CAAmE;IACvF,OAAO,CAAC,YAAY,CAA8B;IAClD,OAAO,CAAC,aAAa,CAAc;IACnC,OAAO,CAAC,eAAe,CAAc;IACrC,OAAO,CAAC,MAAM,CAAe;IAC7B;;;OAGG;IACH,OAAO,CAAC,gBAAgB,CAAgD;IAExE;;;;;;;;;OASG;gBAES,cAAc,EAAE,cAAc,EAC9B,QAAQ,EAAE,eAAe,EACzB,eAAe,EAAE,eAAe,EAChC,WAAW,EAAE,WAAW,EACxB,MAAM,EAAE,qBAAqB,EACrC,MAAM,EAAE,YAAY,EACZ,SAAS,GAAE,OAAc,EACzB,gBAAgB,CAAC,EAAE,GAAG,CAC1B,MAAM,EACN;QAAE,eAAe,EAAE,OAAO,CAAC;QAAC,cAAc,CAAC,EAAE,UAAU,GAAG,UAAU,CAAA;KAAE,CACzE,YAAA;IAKC,OAAO,CACT,QAAQ,EAAE,MAAM,gBAAgB,CAAC,aAAa,EAAE,OAAO,CAAC,GACzD,OAAO,CAAC,qBAAqB,CAAC;YAmanB,sBAAsB;YAKtB,gBAAgB;IAO9B;;;;OAIG;YACW,2BAA2B;CAmC5C"}

package/dist/llm/executor/stream-processor.js

@@ -66,22 +66,30 @@ class StreamProcessor {
               });
             }
             break;
-          case "tool-call":
+          case "tool-call": {
             if (!this.assistantMessageId) {
               this.assistantMessageId = await this.createAssistantMessage();
             }
-            await this.contextManager.addToolCall(this.assistantMessageId, {
+            const toolCall = {
               id: event.toolCallId,
               type: "function",
               function: {
                 name: event.toolName,
                 arguments: JSON.stringify(event.input)
               }
-            });
+            };
+            const shouldPersistProviderMetadata = this.config.provider === "google" || this.config.provider === "vertex";
+            if (shouldPersistProviderMetadata && event.providerMetadata) {
+              toolCall.providerOptions = {
+                ...event.providerMetadata
+              };
+            }
+            await this.contextManager.addToolCall(this.assistantMessageId, toolCall);
             this.pendingToolCalls.set(event.toolCallId, {
               toolName: event.toolName
             });
             break;
+          }
           case "tool-result": {
             const rawResult = event.output;
             this.logger.debug("Tool result received", {
@@ -128,8 +136,10 @@ class StreamProcessor {
           }
           case "finish-step":
             if (event.usage) {
-              const cacheWriteTokens = event.providerMetadata?.["anthropic"]?.["cacheCreationInputTokens"] ?? // @ts-expect-error - Bedrock metadata typing not in Vercel SDK
-              event.providerMetadata?.["bedrock"]?.["usage"]?.["cacheWriteInputTokens"] ?? 0;
+              const anthropicMeta = event.providerMetadata?.["anthropic"];
+              const bedrockMeta = event.providerMetadata?.["bedrock"];
+              const cacheWriteTokens = anthropicMeta?.["cacheCreationInputTokens"] ?? bedrockMeta?.usage?.["cacheWriteInputTokens"] ?? 0;
+              const cacheReadTokens = anthropicMeta?.["cacheReadInputTokens"] ?? bedrockMeta?.usage?.["cacheReadInputTokens"] ?? event.usage.cachedInputTokens ?? 0;
               this.actualTokens = {
                 inputTokens: (this.actualTokens.inputTokens ?? 0) + (event.usage.inputTokens ?? 0),
                 outputTokens: (this.actualTokens.outputTokens ?? 0) + (event.usage.outputTokens ?? 0),
@@ -138,16 +148,17 @@ class StreamProcessor {
                   reasoningTokens: (this.actualTokens.reasoningTokens ?? 0) + event.usage.reasoningTokens
                 },
                 // Cache tokens
-                cacheReadTokens: (this.actualTokens.cacheReadTokens ?? 0) + (event.usage.cachedInputTokens ?? 0),
+                cacheReadTokens: (this.actualTokens.cacheReadTokens ?? 0) + cacheReadTokens,
                 cacheWriteTokens: (this.actualTokens.cacheWriteTokens ?? 0) + cacheWriteTokens
               };
             }
             break;
           case "finish": {
             this.finishReason = event.finishReason;
-            const cachedInputTokens = event.totalUsage.cachedInputTokens ?? 0;
+            const cacheReadTokens = this.actualTokens.cacheReadTokens ?? event.totalUsage.cachedInputTokens ?? 0;
+            const cacheWriteTokens = this.actualTokens.cacheWriteTokens ?? 0;
             const providerExcludesCached = this.config.provider === "anthropic" || this.config.provider === "bedrock";
-            const adjustedInputTokens = providerExcludesCached ? event.totalUsage.inputTokens ?? 0 : (event.totalUsage.inputTokens ?? 0) - cachedInputTokens;
+            const adjustedInputTokens = providerExcludesCached ? event.totalUsage.inputTokens ?? 0 : (event.totalUsage.inputTokens ?? 0) - cacheReadTokens;
             const usage = {
               inputTokens: adjustedInputTokens,
               outputTokens: event.totalUsage.outputTokens ?? 0,
@@ -156,9 +167,9 @@ class StreamProcessor {
               ...event.totalUsage.reasoningTokens !== void 0 && {
                 reasoningTokens: event.totalUsage.reasoningTokens
               },
-              // Cache tokens - read from totalUsage, write from accumulated finish-step events
-              cacheReadTokens: cachedInputTokens,
-              cacheWriteTokens: this.actualTokens.cacheWriteTokens ?? 0
+              // Cache tokens from accumulated finish-step events or totalUsage fallback
+              cacheReadTokens,
+              cacheWriteTokens
             };
             this.actualTokens = usage;
             this.logger.info("LLM response complete", {

package/dist/llm/executor/turn-executor.cjs

@@ -79,17 +79,19 @@ module.exports = __toCommonJS(turn_executor_exports);
 var import_ai = require("ai");
 var import_api = require("@opentelemetry/api");
 var import_stream_processor = require("./stream-processor.js");
+var import_provider_options = require("./provider-options.js");
 var import_types4 = require("../../logger/v2/types.js");
 var import_defer = require("../../utils/defer.js");
 var import_DextoRuntimeError = require("../../errors/DextoRuntimeError.js");
 var import_types7 = require("../../errors/types.js");
 var import_error_codes = require("../error-codes.js");
 var import_error_conversion = require("../../utils/error-conversion.js");
-var import_overflow = require("../../context/compression/overflow.js");
-var import_reactive_overflow = require("../../context/compression/reactive-overflow.js");
+var import_overflow = require("../../context/compaction/overflow.js");
+var import_reactive_overflow = require("../../context/compaction/strategies/reactive-overflow.js");
 const toolSupportCache = /* @__PURE__ */ new Map();
+const LOCAL_PROVIDERS = ["ollama", "local"];
 class TurnExecutor {
-  constructor(model, toolManager, contextManager, eventBus, resourceManager, sessionId, config, llmContext, logger, messageQueue, modelLimits, externalSignal) {
+  constructor(model, toolManager, contextManager, eventBus, resourceManager, sessionId, config, llmContext, logger, messageQueue, modelLimits, externalSignal, compactionStrategy) {
     this.model = model;
     this.toolManager = toolManager;
     this.contextManager = contextManager;
@@ -103,8 +105,10 @@ class TurnExecutor {
     this.externalSignal = externalSignal;
     this.logger = logger.createChild(import_types4.DextoLogComponent.EXECUTOR);
     this.stepAbortController = new AbortController();
-    if (modelLimits) {
-      this.compressionStrategy = new import_reactive_overflow.ReactiveOverflowStrategy(model, {}, this.logger);
+    if (compactionStrategy !== void 0) {
+      this.compactionStrategy = compactionStrategy;
+    } else if (modelLimits) {
+      this.compactionStrategy = new import_reactive_overflow.ReactiveOverflowStrategy(model, {}, this.logger);
     }
   }
   logger;
@@ -113,8 +117,7 @@ class TurnExecutor {
    * This allows soft cancel (abort current step) while still continuing with queued messages.
    */
   stepAbortController;
-  // TODO: improve compression configurability
-  compressionStrategy = null;
+  compactionStrategy = null;
   /**
    * Map to track approval metadata by toolCallId.
    * Used to pass approval info from tool execution to result persistence.
@@ -147,6 +150,24 @@ class TurnExecutor {
       let lastText = "";
       this.eventBus.emit("llm:thinking");
       const supportsTools = await this.validateToolSupport();
+      if (!supportsTools) {
+        const modelKey = `${this.llmContext.provider}:${this.llmContext.model}`;
+        this.eventBus.emit("llm:unsupported-input", {
+          errors: [
+            `Model '${modelKey}' does not support tool calling.`,
+            "You can still chat, but the model will not be able to use tools or execute commands."
+          ],
+          provider: this.llmContext.provider,
+          model: this.llmContext.model,
+          details: {
+            feature: "tool-calling",
+            supported: false
+          }
+        });
+        this.logger.warn(
+          `Model ${modelKey} does not support tools - continuing without tool calling`
+        );
+      }
       let currentAbortHandler = null;
       try {
         while (true) {
@@ -183,6 +204,11 @@ class TurnExecutor {
             streaming,
             this.approvalMetadata
           );
+          const providerOptions = (0, import_provider_options.buildProviderOptions)({
+            provider: this.llmContext.provider,
+            model: this.llmContext.model,
+            reasoningEffort: this.config.reasoningEffort
+          });
           const result = await streamProcessor.process(
             () => (0, import_ai.streamText)({
               model: this.model,
@@ -196,6 +222,10 @@ class TurnExecutor {
               ...this.config.temperature !== void 0 && {
                 temperature: this.config.temperature
               },
+              // Provider-specific options (caching, reasoning, etc.)
+              ...providerOptions !== void 0 && {
+                providerOptions
+              },
               // Log stream-level errors (tool errors, API errors during streaming)
               onError: (error) => {
                 this.logger.error("Stream error", { error });
@@ -299,20 +329,25 @@ class TurnExecutor {
    * Validates if the current model supports tools.
    * Uses a static cache to avoid repeated validation calls.
    *
-   * For models using custom baseURL endpoints, makes a test call to verify tool support.
-   * Built-in providers without baseURL are assumed to support tools.
+   * For local providers (Ollama, local) and custom baseURL endpoints, makes a test call to verify tool support.
+   * Known cloud providers without baseURL are assumed to support tools.
    */
   async validateToolSupport() {
     const modelKey = `${this.llmContext.provider}:${this.llmContext.model}:${this.config.baseURL ?? ""}`;
     if (toolSupportCache.has(modelKey)) {
       return toolSupportCache.get(modelKey);
     }
-    if (!this.config.baseURL) {
-      this.logger.debug(`Skipping tool validation for ${modelKey} - no custom baseURL`);
+    const isLocalProvider = LOCAL_PROVIDERS.includes(this.llmContext.provider);
+    if (!this.config.baseURL && !isLocalProvider) {
+      this.logger.debug(
+        `Skipping tool validation for ${modelKey} - known cloud provider without custom baseURL`
+      );
       toolSupportCache.set(modelKey, true);
       return true;
     }
-    this.logger.debug(`Testing tool support for custom endpoint model: ${modelKey}`);
+    this.logger.debug(
+      `Testing tool support for ${isLocalProvider ? "local provider" : "custom endpoint"} model: ${modelKey}`
+    );
     const testTool = {
       test_tool: {
         inputSchema: (0, import_ai.jsonSchema)({
@@ -342,7 +377,9 @@ class TurnExecutor {
       const errorMessage = error instanceof Error ? error.message : String(error);
       if (errorMessage.includes("does not support tools")) {
         toolSupportCache.set(modelKey, false);
-        this.logger.debug(`Model ${modelKey} does not support tools`);
+        this.logger.debug(
+          `Detected that model ${modelKey} does not support tool calling - tool functionality will be disabled`
+        );
         return false;
       }
       this.logger.debug(
@@ -623,7 +660,7 @@ class TurnExecutor {
    * Check if context has overflowed based on actual token usage from API.
    */
   checkAndHandleOverflow(tokens) {
-    if (!this.modelLimits || !this.compressionStrategy) {
+    if (!this.modelLimits || !this.compactionStrategy) {
       return false;
     }
     return (0, import_overflow.isOverflow)(tokens, this.modelLimits);
@@ -638,16 +675,16 @@ class TurnExecutor {
    * @param originalTokens The actual input token count from API that triggered overflow
    */
   async compress(originalTokens) {
-    if (!this.compressionStrategy) {
+    if (!this.compactionStrategy) {
       return;
     }
     this.logger.info(
       `Context overflow detected (${originalTokens} tokens), running compression`
     );
     const history = await this.contextManager.getHistory();
-    const summaryMessages = await this.compressionStrategy.compress(history);
+    const summaryMessages = await this.compactionStrategy.compact(history);
     if (summaryMessages.length === 0) {
-      this.logger.debug("Compression returned no summary (history too short)");
+      this.logger.debug("Compaction returned no summary (history too short)");
       return;
     }
     for (const summary of summaryMessages) {
@@ -656,17 +693,17 @@ class TurnExecutor {
     const { filterCompacted, estimateMessagesTokens } = await import("../../context/utils.js");
     const updatedHistory = await this.contextManager.getHistory();
     const filteredHistory = filterCompacted(updatedHistory);
-    const compressedTokens = estimateMessagesTokens(filteredHistory);
-    this.eventBus.emit("context:compressed", {
+    const compactedTokens = estimateMessagesTokens(filteredHistory);
+    this.eventBus.emit("context:compacted", {
       originalTokens,
-      compressedTokens,
+      compactedTokens,
       originalMessages: history.length,
-      compressedMessages: filteredHistory.length,
-      strategy: this.compressionStrategy.name,
+      compactedMessages: filteredHistory.length,
+      strategy: this.compactionStrategy.name,
       reason: "overflow"
     });
     this.logger.info(
-      `Compression complete: ${originalTokens} \u2192 ~${compressedTokens} tokens (${history.length} \u2192 ${filteredHistory.length} messages after filtering)`
+      `Compaction complete: ${originalTokens} \u2192 ~${compactedTokens} tokens (${history.length} \u2192 ${filteredHistory.length} messages after filtering)`
     );
   }
   /**

package/dist/llm/executor/turn-executor.d.ts

@@ -8,7 +8,8 @@ import type { ResourceManager } from '../../resources/index.js';
 import { DynamicContributorContext } from '../../systemPrompt/types.js';
 import { LLMContext } from '../types.js';
 import type { MessageQueueService } from '../../session/message-queue.js';
-import { type ModelLimits } from '../../context/compression/overflow.js';
+import { type ModelLimits } from '../../context/compaction/overflow.js';
+import type { ICompactionStrategy } from '../../context/compaction/types.js';
 /**
  * TurnExecutor orchestrates the agent loop using `stopWhen: stepCountIs(1)`.
  *
@@ -39,7 +40,7 @@ export declare class TurnExecutor {
      * This allows soft cancel (abort current step) while still continuing with queued messages.
      */
     private stepAbortController;
-    private compressionStrategy;
+    private compactionStrategy;
     /**
      * Map to track approval metadata by toolCallId.
      * Used to pass approval info from tool execution to result persistence.
@@ -50,7 +51,8 @@ export declare class TurnExecutor {
         maxOutputTokens?: number | undefined;
         temperature?: number | undefined;
         baseURL?: string | undefined;
-    }, llmContext: LLMContext, logger: IDextoLogger, messageQueue: MessageQueueService, modelLimits?: ModelLimits | undefined, externalSignal?: AbortSignal | undefined);
+        reasoningEffort?: 'none' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | undefined;
+    }, llmContext: LLMContext, logger: IDextoLogger, messageQueue: MessageQueueService, modelLimits?: ModelLimits | undefined, externalSignal?: AbortSignal | undefined, compactionStrategy?: ICompactionStrategy | null);
     /**
      * Get StreamProcessor config from TurnExecutor state.
      */
@@ -77,8 +79,8 @@ export declare class TurnExecutor {
      * Validates if the current model supports tools.
      * Uses a static cache to avoid repeated validation calls.
      *
-     * For models using custom baseURL endpoints, makes a test call to verify tool support.
-     * Built-in providers without baseURL are assumed to support tools.
+     * For local providers (Ollama, local) and custom baseURL endpoints, makes a test call to verify tool support.
+     * Known cloud providers without baseURL are assumed to support tools.
      */
     private validateToolSupport;
     /**

package/dist/llm/executor/turn-executor.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"turn-executor.d.ts","sourceRoot":"","sources":["../../../src/llm/executor/turn-executor.ts"],"names":[],"mappings":"AAAA,OAAO,EACH,aAAa,EAMb,KAAK,YAAY,EAEpB,MAAM,IAAI,CAAC;AAEZ,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AAE1D,OAAO,EAAE,WAAW,EAAE,MAAM,6BAA6B,CAAC;AAG1D,OAAO,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAE5C,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,0BAA0B,CAAC;AAE7D,OAAO,KAAK,EAAE,eAAe,EAAmB,MAAM,uBAAuB,CAAC;AAC9E,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAC;AAChE,OAAO,EAAE,yBAAyB,EAAE,MAAM,6BAA6B,CAAC;AACxE,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AACzC,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,gCAAgC,CAAC;AAQ1E,OAAO,EAAc,KAAK,WAAW,EAAE,MAAM,uCAAuC,CAAC;AAUrF;;;;;;;;;;;GAWG;AACH,qBAAa,YAAY;IAmBjB,OAAO,CAAC,KAAK;IACb,OAAO,CAAC,WAAW;IACnB,OAAO,CAAC,cAAc;IACtB,OAAO,CAAC,QAAQ;IAChB,OAAO,CAAC,eAAe;IACvB,OAAO,CAAC,SAAS;IACjB,OAAO,CAAC,MAAM;IAMd,OAAO,CAAC,UAAU;IAElB,OAAO,CAAC,YAAY;IACpB,OAAO,CAAC,WAAW,CAAC;IACpB,OAAO,CAAC,cAAc,CAAC;IAlC3B,OAAO,CAAC,MAAM,CAAe;IAC7B;;;OAGG;IACH,OAAO,CAAC,mBAAmB,CAAkB;IAE7C,OAAO,CAAC,mBAAmB,CAAyC;IACpE;;;OAGG;IACH,OAAO,CAAC,gBAAgB,CAGpB;gBAGQ,KAAK,EAAE,aAAa,EACpB,WAAW,EAAE,WAAW,EACxB,cAAc,EAAE,cAAc,CAAC,YAAY,CAAC,EAC5C,QAAQ,EAAE,eAAe,EACzB,eAAe,EAAE,eAAe,EAChC,SAAS,EAAE,MAAM,EACjB,MAAM,EAAE;QACZ,QAAQ,EAAE,MAAM,CAAC;QACjB,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QACrC,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QACjC,OAAO,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;KAChC,EACO,UAAU,EAAE,UAAU,EAC9B,MAAM,EAAE,YAAY,EACZ,YAAY,EAAE,mBAAmB,EACjC,WAAW,CAAC,EAAE,WAAW,YAAA,EACzB,cAAc,CAAC,EAAE,WAAW,YAAA;IAiBxC;;OAEG;IACH,OAAO,CAAC,wBAAwB;IAOhC;;;;;;OAMG;IACG,OAAO,CACT,kBAAkB,EAAE,yBAAyB,EAC7C,SAAS,GAAE,OAAc,GAC1B,OAAO,CAAC,cAAc,CAAC;IAsL1B;;;OAGG;IACH,KAAK,IAAI,IAAI;IAIb;;;OAGG;YACW,oBAAoB;IAmBlC;;;;;;OAMG;YACW,mBAAmB;IAkEjC;;;;;;;OAOG;YACW,WAAW;IA0GzB;;;;;OAKG;IACH,OAAO,CAAC,sBAAsB;IAuF9B;;OAEG;IACH,OAAO,CAAC,gBAAgB;IAyBxB;;OAEG;IACH,OAAO,CAAC,eAAe;IAavB;;OAEG;IACH,OAAO,CAAC,oBAAoB;IAS5B;;OAEG;IACH,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,aAAa,CAAU;IAC/C,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,aAAa,CAAU;IAE/C;;;;;;;;;;;OAWG;YACW,mBAAmB;IAkDjC;;;;;;OAMG;IACH,OAAO,CAAC,kBAAkB;IAgB1B;;;OAGG;IACH,OAAO,CAAC,OAAO;IAYf;;OAEG;IACH,OAAO,CAAC,sBAAsB;IAO9B;;;;;;;;OAQG;YACW,QAAQ;IA8CtB;;OAEG;IACH,OAAO,CAAC,sBAAsB;IAoB9B;;OAEG;IACH,OAAO,CAAC,gBAAgB;CA0D3B"}
+{"version":3,"file":"turn-executor.d.ts","sourceRoot":"","sources":["../../../src/llm/executor/turn-executor.ts"],"names":[],"mappings":"AAAA,OAAO,EACH,aAAa,EAMb,KAAK,YAAY,EAEpB,MAAM,IAAI,CAAC;AAEZ,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AAE1D,OAAO,EAAE,WAAW,EAAE,MAAM,6BAA6B,CAAC;AAG1D,OAAO,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAG5C,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,0BAA0B,CAAC;AAE7D,OAAO,KAAK,EAAE,eAAe,EAAmB,MAAM,uBAAuB,CAAC;AAC9E,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAC;AAChE,OAAO,EAAE,yBAAyB,EAAE,MAAM,6BAA6B,CAAC;AACxE,OAAO,EAAE,UAAU,EAAoB,MAAM,aAAa,CAAC;AAC3D,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,gCAAgC,CAAC;AAQ1E,OAAO,EAAc,KAAK,WAAW,EAAE,MAAM,sCAAsC,CAAC;AAEpF,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mCAAmC,CAAC;AAc7E;;;;;;;;;;;GAWG;AACH,qBAAa,YAAY;IAkBjB,OAAO,CAAC,KAAK;IACb,OAAO,CAAC,WAAW;IACnB,OAAO,CAAC,cAAc;IACtB,OAAO,CAAC,QAAQ;IAChB,OAAO,CAAC,eAAe;IACvB,OAAO,CAAC,SAAS;IACjB,OAAO,CAAC,MAAM;IAQd,OAAO,CAAC,UAAU;IAElB,OAAO,CAAC,YAAY;IACpB,OAAO,CAAC,WAAW,CAAC;IACpB,OAAO,CAAC,cAAc,CAAC;IAnC3B,OAAO,CAAC,MAAM,CAAe;IAC7B;;;OAGG;IACH,OAAO,CAAC,mBAAmB,CAAkB;IAC7C,OAAO,CAAC,kBAAkB,CAAoC;IAC9D;;;OAGG;IACH,OAAO,CAAC,gBAAgB,CAGpB;gBAGQ,KAAK,EAAE,aAAa,EACpB,WAAW,EAAE,WAAW,EACxB,cAAc,EAAE,cAAc,CAAC,YAAY,CAAC,EAC5C,QAAQ,EAAE,eAAe,EACzB,eAAe,EAAE,eAAe,EAChC,SAAS,EAAE,MAAM,EACjB,MAAM,EAAE;QACZ,QAAQ,EAAE,MAAM,CAAC;QACjB,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QACrC,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QACjC,OAAO,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QAE7B,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,SAAS,CAAC;KAC1F,EACO,UAAU,EAAE,UAAU,EAC9B,MAAM,EAAE,YAAY,EACZ,YAAY,EAAE,mBAAmB,EACjC,WAAW,CAAC,EAAE,WAAW,YAAA,EACzB,cAAc,CAAC,EAAE,WAAW,YAAA,EACpC,kBAAkB,CAAC,EAAE,mBAAmB,GAAG,IAAI;IAqBnD;;OAEG;IACH,OAAO,CAAC,wBAAwB;IAOhC;;;;;;OAMG;IACG,OAAO,CACT,kBAAkB,EAAE,yBAAyB,EAC7C,SAAS,GAAE,OAAc,GAC1B,OAAO,CAAC,cAAc,CAAC;IAsN1B;;;OAGG;IACH,KAAK,IAAI,IAAI;IAIb;;;OAGG;YACW,oBAAoB;IAmBlC;;;;;;OAMG;YACW,mBAAmB;IA0EjC;;;;;;;OAOG;YACW,WAAW;IA0GzB;;;;;OAKG;IACH,OAAO,CAAC,sBAAsB;IAuF9B;;OAEG;IACH,OAAO,CAAC,gBAAgB;IAyBxB;;OAEG;IACH,OAAO,CAAC,eAAe;IAavB;;OAEG;IACH,OAAO,CAAC,oBAAoB;IAS5B;;OAEG;IACH,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,aAAa,CAAU;IAC/C,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,aAAa,CAAU;IAE/C;;;;;;;;;;;OAWG;YACW,mBAAmB;IAkDjC;;;;;;OAMG;IACH,OAAO,CAAC,kBAAkB;IAgB1B;;;OAGG;IACH,OAAO,CAAC,OAAO;IAYf;;OAEG;IACH,OAAO,CAAC,sBAAsB;IAO9B;;;;;;;;OAQG;YACW,QAAQ;IA8CtB;;OAEG;IACH,OAAO,CAAC,sBAAsB;IAoB9B;;OAEG;IACH,OAAO,CAAC,gBAAgB;CA0D3B"}