@axlsdk/axl 0.5.0 → 0.6.0

package/dist/index.cjs

@@ -140,6 +140,7 @@ __export(index_exports, {
   agent: () => agent,
   createSpanManager: () => createSpanManager,
   defineConfig: () => defineConfig,
+  resolveThinkingOptions: () => resolveThinkingOptions,
   tool: () => tool,
   workflow: () => workflow,
   zodToJsonSchema: () => zodToJsonSchema
@@ -185,7 +186,7 @@ function tool(config) {
     on: config.retry?.on
   };
   const maxStringLen = config.maxStringLength ?? DEFAULT_MAX_STRING_LENGTH;
-  const execute = async (input) => {
+  const execute = async (input, ctx) => {
     const parsed = config.input.parse(input);
     if (maxStringLen > 0) {
       validateStringLengths(parsed, maxStringLen);
@@ -194,7 +195,7 @@ function tool(config) {
     let lastError;
     for (let attempt = 1; attempt <= maxAttempts; attempt++) {
       try {
-        return await config.handler(parsed);
+        return await config.handler(parsed, ctx);
       } catch (err) {
         lastError = err instanceof Error ? err : new Error(String(err));
         if (attempt === maxAttempts) break;
@@ -224,7 +225,7 @@ function tool(config) {
         if (config.hooks?.before) {
           processedInput = await config.hooks.before(processedInput, ctx);
         }
-        let result = await execute(processedInput);
+        let result = await execute(processedInput, ctx);
         if (config.hooks?.after) {
           result = await config.hooks.after(result, ctx);
         }
@@ -246,6 +247,25 @@ function tool(config) {
   };
 }
 
+// src/providers/types.ts
+function resolveThinkingOptions(options) {
+  if (options.thinkingBudget !== void 0 && options.thinkingBudget < 0) {
+    throw new Error(`thinkingBudget must be non-negative, got ${options.thinkingBudget}`);
+  }
+  const effort = options.effort;
+  const thinkingBudget = options.thinkingBudget;
+  const hasBudgetOverride = thinkingBudget !== void 0 && thinkingBudget > 0;
+  return {
+    effort,
+    thinkingBudget,
+    includeThoughts: options.includeThoughts ?? false,
+    // Budget override wins: effort: 'none' + thinkingBudget: 5000 → thinking enabled
+    thinkingDisabled: (effort === "none" || thinkingBudget === 0) && !hasBudgetOverride,
+    activeEffort: effort && effort !== "none" ? effort : void 0,
+    hasBudgetOverride
+  };
+}
+
 // src/providers/retry.ts
 var RETRYABLE_STATUS_CODES = /* @__PURE__ */ new Set([429, 503, 529]);
 var MAX_RETRIES = 2;
@@ -300,6 +320,9 @@ var OPENAI_PRICING = {
   "gpt-5-nano": [5e-8, 4e-7],
   "gpt-5.1": [125e-8, 1e-5],
   "gpt-5.2": [175e-8, 14e-6],
+  "gpt-5.3": [175e-8, 14e-6],
+  "gpt-5.4": [25e-7, 15e-6],
+  "gpt-5.4-pro": [3e-5, 18e-5],
   o1: [15e-6, 6e-5],
   "o1-mini": [3e-6, 12e-6],
   "o1-pro": [15e-5, 6e-4],
@@ -328,26 +351,31 @@ function estimateOpenAICost(model, promptTokens, completionTokens, cachedTokens)
   const inputCost = (promptTokens - cached) * inputRate + cached * inputRate * 0.5;
   return inputCost + completionTokens * outputRate;
 }
-function isReasoningModel(model) {
+function isOSeriesModel(model) {
   return /^(o1|o3|o4-mini)/.test(model);
 }
-function thinkingToReasoningEffort(thinking) {
-  if (typeof thinking === "object") {
-    const budget = thinking.budgetTokens;
-    if (budget <= 1024) return "low";
-    if (budget <= 8192) return "medium";
-    return "high";
-  }
-  switch (thinking) {
-    case "low":
-      return "low";
-    case "medium":
-      return "medium";
-    case "high":
-      return "high";
-    case "max":
-      return "xhigh";
-  }
+function supportsReasoningEffort(model) {
+  return isOSeriesModel(model) || /^gpt-5/.test(model);
+}
+function supportsReasoningNone(model) {
+  return /^gpt-5\.[1-9]/.test(model);
+}
+function supportsXhigh(model) {
+  return /^gpt-5\.([2-9]|\d{2,})/.test(model);
+}
+function clampReasoningEffort(model, effort) {
+  if (model.startsWith("gpt-5-pro")) return "high";
+  if (effort === "none" && !supportsReasoningNone(model)) return "minimal";
+  if (effort === "xhigh" && !supportsXhigh(model)) return "high";
+  return effort;
+}
+function effortToReasoningEffort(effort) {
+  return effort === "max" ? "xhigh" : effort;
+}
+function budgetToReasoningEffort(budget) {
+  if (budget <= 1024) return "low";
+  if (budget <= 8192) return "medium";
+  return "high";
 }
 var OpenAIProvider = class {
   name = "openai";
@@ -436,13 +464,26 @@ var OpenAIProvider = class {
   // Internal helpers
   // ---------------------------------------------------------------------------
   buildRequestBody(messages, options, stream) {
-    const reasoning = isReasoningModel(options.model);
+    const oSeries = isOSeriesModel(options.model);
+    const reasoningCapable = supportsReasoningEffort(options.model);
+    const { thinkingBudget, thinkingDisabled, activeEffort, hasBudgetOverride } = resolveThinkingOptions(options);
+    let wireEffort;
+    if (reasoningCapable) {
+      if (hasBudgetOverride) {
+        wireEffort = clampReasoningEffort(options.model, budgetToReasoningEffort(thinkingBudget));
+      } else if (!thinkingDisabled && activeEffort) {
+        wireEffort = clampReasoningEffort(options.model, effortToReasoningEffort(activeEffort));
+      } else if (thinkingDisabled) {
+        wireEffort = clampReasoningEffort(options.model, "none");
+      }
+    }
+    const stripTemp = oSeries || reasoningCapable && wireEffort !== void 0;
     const body = {
       model: options.model,
-      messages: messages.map((m) => this.formatMessage(m, reasoning)),
+      messages: messages.map((m) => this.formatMessage(m, oSeries)),
       stream
     };
-    if (options.temperature !== void 0 && !reasoning) {
+    if (options.temperature !== void 0 && !stripTemp) {
       body.temperature = options.temperature;
     }
     if (options.maxTokens !== void 0) {
@@ -451,7 +492,7 @@ var OpenAIProvider = class {
     if (options.stop) body.stop = options.stop;
     if (options.tools && options.tools.length > 0) {
       body.tools = options.tools;
-      if (!reasoning) {
+      if (!oSeries) {
         body.parallel_tool_calls = true;
       }
     }
@@ -461,15 +502,13 @@ var OpenAIProvider = class {
     if (options.responseFormat) {
       body.response_format = options.responseFormat;
     }
-    if (reasoning) {
-      const effort = options.thinking ? thinkingToReasoningEffort(options.thinking) : options.reasoningEffort;
-      if (effort) {
-        body.reasoning_effort = effort;
-      }
-    }
+    if (wireEffort) body.reasoning_effort = wireEffort;
     if (stream) {
       body.stream_options = { include_usage: true };
     }
+    if (options.providerOptions) {
+      Object.assign(body, options.providerOptions);
+    }
     return body;
   }
   /** Extract a human-readable message from an API error response body. */
@@ -483,9 +522,9 @@ var OpenAIProvider = class {
     }
     return `OpenAI API error (${status}): ${body}`;
   }
-  formatMessage(msg, reasoning) {
+  formatMessage(msg, oSeries) {
     const out = {
-      role: msg.role === "system" && reasoning ? "developer" : msg.role,
+      role: msg.role === "system" && oSeries ? "developer" : msg.role,
       content: msg.content
     };
     if (msg.name) out.name = msg.name;
@@ -622,7 +661,20 @@ var OpenAIResponsesProvider = class {
   // Internal: build request body
   // ---------------------------------------------------------------------------
   buildRequestBody(messages, options, stream) {
-    const reasoning = isReasoningModel(options.model);
+    const oSeries = isOSeriesModel(options.model);
+    const reasoningCapable = supportsReasoningEffort(options.model);
+    const { thinkingBudget, includeThoughts, thinkingDisabled, activeEffort, hasBudgetOverride } = resolveThinkingOptions(options);
+    let wireEffort;
+    if (reasoningCapable) {
+      if (hasBudgetOverride) {
+        wireEffort = clampReasoningEffort(options.model, budgetToReasoningEffort(thinkingBudget));
+      } else if (!thinkingDisabled && activeEffort) {
+        wireEffort = clampReasoningEffort(options.model, effortToReasoningEffort(activeEffort));
+      } else if (thinkingDisabled) {
+        wireEffort = clampReasoningEffort(options.model, "none");
+      }
+    }
+    const stripTemp = oSeries || reasoningCapable && wireEffort !== void 0;
     const systemMessages = messages.filter((m) => m.role === "system");
     const nonSystemMessages = messages.filter((m) => m.role !== "system");
     const body = {
@@ -637,7 +689,7 @@ var OpenAIResponsesProvider = class {
     if (options.maxTokens !== void 0) {
       body.max_output_tokens = options.maxTokens;
     }
-    if (options.temperature !== void 0 && !reasoning) {
+    if (options.temperature !== void 0 && !stripTemp) {
       body.temperature = options.temperature;
     }
     if (options.tools && options.tools.length > 0) {
@@ -656,15 +708,21 @@ var OpenAIResponsesProvider = class {
         body.tool_choice = options.toolChoice;
       }
     }
-    if (reasoning) {
-      const effort = options.thinking ? thinkingToReasoningEffort(options.thinking) : options.reasoningEffort;
-      if (effort) {
-        body.reasoning = { effort };
-      }
+    if (reasoningCapable && (wireEffort !== void 0 || includeThoughts)) {
+      const reasoning = {};
+      if (wireEffort !== void 0) reasoning.effort = wireEffort;
+      if (includeThoughts) reasoning.summary = "detailed";
+      if (Object.keys(reasoning).length > 0) body.reasoning = reasoning;
+    }
+    if (reasoningCapable) {
+      body.include = ["reasoning.encrypted_content"];
     }
     if (options.responseFormat) {
       body.text = { format: this.mapResponseFormat(options.responseFormat) };
     }
+    if (options.providerOptions) {
+      Object.assign(body, options.providerOptions);
+    }
     return body;
   }
   // ---------------------------------------------------------------------------
@@ -680,6 +738,12 @@ var OpenAIResponsesProvider = class {
           output: msg.content
         });
       } else if (msg.role === "assistant" && msg.tool_calls && msg.tool_calls.length > 0) {
+        const reasoningItems = msg.providerMetadata?.openaiReasoningItems;
+        if (reasoningItems) {
+          for (const item of reasoningItems) {
+            input.push(item);
+          }
+        }
         if (msg.content) {
           input.push({ type: "message", role: "assistant", content: msg.content });
         }
@@ -692,6 +756,12 @@ var OpenAIResponsesProvider = class {
           });
         }
       } else if (msg.role === "user" || msg.role === "assistant") {
+        if (msg.role === "assistant" && msg.providerMetadata?.openaiReasoningItems) {
+          const reasoningItems = msg.providerMetadata.openaiReasoningItems;
+          for (const item of reasoningItems) {
+            input.push(item);
+          }
+        }
         input.push({
           type: "message",
           role: msg.role,
@@ -724,7 +794,9 @@ var OpenAIResponsesProvider = class {
   // ---------------------------------------------------------------------------
   parseResponse(json, model) {
     let content = "";
+    let thinkingContent = "";
     const toolCalls = [];
+    const reasoningItems = [];
     for (const item of json.output) {
       if (item.type === "message") {
         for (const part of item.content ?? []) {
@@ -741,6 +813,15 @@ var OpenAIResponsesProvider = class {
             arguments: item.arguments
           }
         });
+      } else if (item.type === "reasoning") {
+        reasoningItems.push(item);
+        if (item.summary) {
+          for (const s of item.summary) {
+            if (s.type === "summary_text" && s.text) {
+              thinkingContent += s.text;
+            }
+          }
+        }
       }
     }
     const usage = json.usage ? {
@@ -751,11 +832,14 @@ var OpenAIResponsesProvider = class {
       cached_tokens: json.usage.input_tokens_details?.cached_tokens
     } : void 0;
     const cost = usage ? estimateOpenAICost(model, usage.prompt_tokens, usage.completion_tokens, usage.cached_tokens) : void 0;
+    const providerMetadata = reasoningItems.length > 0 ? { openaiReasoningItems: reasoningItems } : void 0;
     return {
       content,
+      thinking_content: thinkingContent || void 0,
       tool_calls: toolCalls.length > 0 ? toolCalls : void 0,
       usage,
-      cost
+      cost,
+      providerMetadata
     };
   }
   // ---------------------------------------------------------------------------
@@ -807,6 +891,8 @@ var OpenAIResponsesProvider = class {
     switch (eventType) {
       case "response.output_text.delta":
         return { type: "text_delta", content: data.delta ?? "" };
+      case "response.reasoning_summary_text.delta":
+        return { type: "thinking_delta", content: data.delta ?? "" };
       case "response.output_item.added":
         if (data.item?.type === "function_call") {
           const callId = data.item.call_id ?? data.item.id ?? "";
@@ -837,7 +923,9 @@ var OpenAIResponsesProvider = class {
           reasoning_tokens: response.usage.output_tokens_details?.reasoning_tokens,
           cached_tokens: response.usage.input_tokens_details?.cached_tokens
         } : void 0;
-        return { type: "done", usage };
+        const reasoningItems = response?.output?.filter((item) => item.type === "reasoning") ?? [];
+        const providerMetadata = reasoningItems.length > 0 ? { openaiReasoningItems: reasoningItems } : void 0;
+        return { type: "done", usage, providerMetadata };
       }
       case "response.failed": {
         const errorMsg = data.response?.error?.message ?? data.response?.status_details?.error?.message ?? "Unknown error";
@@ -865,9 +953,12 @@ var OpenAIResponsesProvider = class {
 // src/providers/anthropic.ts
 var ANTHROPIC_API_VERSION = "2023-06-01";
 var ANTHROPIC_PRICING = {
-  "claude-opus-4-6": [15e-6, 75e-6],
+  "claude-opus-4-6": [5e-6, 25e-6],
+  "claude-sonnet-4-6": [3e-6, 15e-6],
+  "claude-opus-4-5": [5e-6, 25e-6],
+  "claude-opus-4-1": [15e-6, 75e-6],
   "claude-sonnet-4-5": [3e-6, 15e-6],
-  "claude-haiku-4-5": [8e-7, 4e-6],
+  "claude-haiku-4-5": [1e-6, 5e-6],
   "claude-sonnet-4": [3e-6, 15e-6],
   "claude-opus-4": [15e-6, 75e-6],
   "claude-3-7-sonnet": [3e-6, 15e-6],
@@ -877,12 +968,15 @@ var ANTHROPIC_PRICING = {
   "claude-3-sonnet": [3e-6, 15e-6],
   "claude-3-haiku": [25e-8, 125e-8]
 };
+var ANTHROPIC_PRICING_KEYS_BY_LENGTH = Object.keys(ANTHROPIC_PRICING).sort(
+  (a, b) => b.length - a.length
+);
 function estimateAnthropicCost(model, inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens) {
   let pricing = ANTHROPIC_PRICING[model];
   if (!pricing) {
-    for (const [key, value] of Object.entries(ANTHROPIC_PRICING)) {
+    for (const key of ANTHROPIC_PRICING_KEYS_BY_LENGTH) {
       if (model.startsWith(key)) {
-        pricing = value;
+        pricing = ANTHROPIC_PRICING[key];
         break;
       }
     }
@@ -902,16 +996,15 @@ var THINKING_BUDGETS = {
   // With auto-bump (+1024), max_tokens becomes 31024 which fits all models.
   max: 3e4
 };
-function thinkingToBudgetTokens(thinking) {
-  if (typeof thinking === "string") return THINKING_BUDGETS[thinking] ?? 5e3;
-  return thinking.budgetTokens;
-}
 function supportsAdaptiveThinking(model) {
   return model.startsWith("claude-opus-4-6") || model.startsWith("claude-sonnet-4-6");
 }
 function supportsMaxEffort(model) {
   return model.startsWith("claude-opus-4-6");
 }
+function supportsEffort(model) {
+  return model.startsWith("claude-opus-4-6") || model.startsWith("claude-sonnet-4-6") || model.startsWith("claude-opus-4-5");
+}
 var AnthropicProvider = class {
   name = "anthropic";
   baseUrl;
@@ -1001,9 +1094,6 @@ var AnthropicProvider = class {
     if (systemText) {
       body.system = systemText;
     }
-    if (options.temperature !== void 0 && !options.thinking) {
-      body.temperature = options.temperature;
-    }
     if (options.stop) {
       body.stop_sequences = options.stop;
     }
@@ -1013,19 +1103,39 @@ var AnthropicProvider = class {
     if (options.toolChoice !== void 0) {
       body.tool_choice = this.mapToolChoice(options.toolChoice);
     }
-    if (options.thinking) {
-      if (typeof options.thinking === "string" && supportsAdaptiveThinking(options.model) && // 'max' effort is only supported on Opus 4.6; Sonnet 4.6 falls back to manual mode
-      (options.thinking !== "max" || supportsMaxEffort(options.model))) {
-        body.thinking = { type: "adaptive" };
-        body.output_config = { effort: options.thinking };
-      } else {
-        const budgetTokens = thinkingToBudgetTokens(options.thinking);
-        body.thinking = { type: "enabled", budget_tokens: budgetTokens };
-        const currentMax = body.max_tokens;
-        if (currentMax < budgetTokens + 1024) {
-          body.max_tokens = budgetTokens + 1024;
-        }
+    const { thinkingBudget, thinkingDisabled, activeEffort, hasBudgetOverride } = resolveThinkingOptions(options);
+    let resolvedEffort = activeEffort;
+    if (resolvedEffort === "max" && !supportsMaxEffort(options.model)) {
+      resolvedEffort = "high";
+    }
+    if (hasBudgetOverride) {
+      body.thinking = { type: "enabled", budget_tokens: thinkingBudget };
+      const currentMax = body.max_tokens;
+      if (currentMax < thinkingBudget + 1024) {
+        body.max_tokens = thinkingBudget + 1024;
       }
+      if (resolvedEffort && supportsEffort(options.model)) {
+        body.output_config = { effort: resolvedEffort };
+      }
+    } else if (thinkingDisabled) {
+      if (resolvedEffort && supportsEffort(options.model)) {
+        body.output_config = { effort: resolvedEffort };
+      }
+    } else if (resolvedEffort && supportsAdaptiveThinking(options.model)) {
+      body.thinking = { type: "adaptive" };
+      body.output_config = { effort: resolvedEffort };
+    } else if (resolvedEffort && supportsEffort(options.model)) {
+      body.output_config = { effort: resolvedEffort };
+    } else if (resolvedEffort) {
+      const budget = THINKING_BUDGETS[resolvedEffort] ?? 5e3;
+      body.thinking = { type: "enabled", budget_tokens: budget };
+      const currentMax = body.max_tokens;
+      if (currentMax < budget + 1024) {
+        body.max_tokens = budget + 1024;
+      }
+    }
+    if (options.temperature !== void 0 && !body.thinking) {
+      body.temperature = options.temperature;
     }
     if (options.responseFormat && options.responseFormat.type !== "text") {
       const jsonInstruction = "You must respond with valid JSON only. No markdown fences, no extra text.";
@@ -1033,6 +1143,9 @@ var AnthropicProvider = class {
 
 ${jsonInstruction}` : jsonInstruction;
     }
+    if (options.providerOptions) {
+      Object.assign(body, options.providerOptions);
+    }
     return body;
   }
   /**
@@ -1143,9 +1256,12 @@ ${jsonInstruction}` : jsonInstruction;
   // ---------------------------------------------------------------------------
   parseResponse(json) {
     let content = "";
+    let thinkingContent = "";
     const toolCalls = [];
     for (const block of json.content) {
-      if (block.type === "text") {
+      if (block.type === "thinking") {
+        thinkingContent += block.thinking;
+      } else if (block.type === "text") {
         content += block.text;
       } else if (block.type === "tool_use") {
         toolCalls.push({
@@ -1176,6 +1292,7 @@ ${jsonInstruction}` : jsonInstruction;
     ) : void 0;
     return {
       content,
+      thinking_content: thinkingContent || void 0,
       tool_calls: toolCalls.length > 0 ? toolCalls : void 0,
       usage,
       cost
@@ -1226,7 +1343,9 @@ ${jsonInstruction}` : jsonInstruction;
             }
             case "content_block_delta": {
               const delta = event.delta;
-              if (delta?.type === "text_delta" && delta.text) {
+              if (delta?.type === "thinking_delta" && delta.thinking) {
+                yield { type: "thinking_delta", content: delta.thinking };
+              } else if (delta?.type === "text_delta" && delta.text) {
                 yield { type: "text_delta", content: delta.text };
               } else if (delta?.type === "input_json_delta" && delta.partial_json) {
                 yield {
@@ -1297,14 +1416,19 @@ var GEMINI_PRICING = {
   "gemini-2.0-flash": [1e-7, 4e-7],
   "gemini-2.0-flash-lite": [1e-7, 4e-7],
   "gemini-3-pro-preview": [2e-6, 12e-6],
-  "gemini-3-flash-preview": [5e-7, 3e-6]
+  "gemini-3-flash-preview": [5e-7, 3e-6],
+  "gemini-3.1-pro-preview": [2e-6, 12e-6],
+  "gemini-3.1-flash-lite-preview": [25e-8, 15e-7]
 };
+var GEMINI_PRICING_KEYS_BY_LENGTH = Object.keys(GEMINI_PRICING).sort(
+  (a, b) => b.length - a.length
+);
 function estimateGeminiCost(model, inputTokens, outputTokens, cachedTokens) {
   let pricing = GEMINI_PRICING[model];
   if (!pricing) {
-    for (const [key, value] of Object.entries(GEMINI_PRICING)) {
+    for (const key of GEMINI_PRICING_KEYS_BY_LENGTH) {
       if (model.startsWith(key)) {
-        pricing = value;
+        pricing = GEMINI_PRICING[key];
         break;
       }
     }
@@ -1321,9 +1445,32 @@ var THINKING_BUDGETS2 = {
   high: 1e4,
   max: 24576
 };
-function thinkingToBudgetTokens2(thinking) {
-  if (typeof thinking === "string") return THINKING_BUDGETS2[thinking] ?? 5e3;
-  return thinking.budgetTokens;
+var THINKING_LEVELS = {
+  low: "low",
+  medium: "medium",
+  high: "high",
+  max: "high"
+  // 3.x caps at 'high'
+};
+function isGemini3x(model) {
+  return /^gemini-3[.-]/.test(model);
+}
+function budgetToThinkingLevel(budgetTokens) {
+  if (budgetTokens <= 1024) return "low";
+  if (budgetTokens <= 5e3) return "medium";
+  return "high";
+}
+function minThinkingLevel(model) {
+  if (model.startsWith("gemini-3.1-pro")) return "low";
+  return "minimal";
+}
+var _warned3xEffortNone = /* @__PURE__ */ new Set();
+function warnGemini3xEffortNone(model) {
+  if (_warned3xEffortNone.has(model)) return;
+  _warned3xEffortNone.add(model);
+  console.warn(
+    `[axl] effort: 'none' on Gemini 3.x (${model}) maps to the model's minimum thinking level ('${minThinkingLevel(model)}'), not fully disabled. Gemini 3.x models cannot disable thinking entirely.`
+  );
 }
 var GeminiProvider = class {
   name = "google";
@@ -1438,17 +1585,58 @@ var GeminiProvider = class {
     if (Object.keys(generationConfig).length > 0) {
       body.generationConfig = generationConfig;
     }
-    if (options.thinking) {
-      generationConfig.thinkingConfig = {
-        thinkingBudget: thinkingToBudgetTokens2(options.thinking)
-      };
-      if (!body.generationConfig) {
-        body.generationConfig = generationConfig;
+    const {
+      effort,
+      thinkingBudget,
+      includeThoughts,
+      thinkingDisabled,
+      activeEffort,
+      hasBudgetOverride
+    } = resolveThinkingOptions(options);
+    if (thinkingDisabled) {
+      if (isGemini3x(options.model)) {
+        if (effort === "none") {
+          warnGemini3xEffortNone(options.model);
+        }
+        generationConfig.thinkingConfig = { thinkingLevel: minThinkingLevel(options.model) };
+      } else {
+        generationConfig.thinkingConfig = { thinkingBudget: 0 };
+      }
+      if (!body.generationConfig) body.generationConfig = generationConfig;
+    } else if (hasBudgetOverride) {
+      const config = {};
+      if (isGemini3x(options.model)) {
+        config.thinkingLevel = budgetToThinkingLevel(thinkingBudget);
+      } else {
+        config.thinkingBudget = thinkingBudget;
+      }
+      if (includeThoughts) config.includeThoughts = true;
+      generationConfig.thinkingConfig = config;
+      if (!body.generationConfig) body.generationConfig = generationConfig;
+    } else if (activeEffort) {
+      const config = {};
+      if (isGemini3x(options.model)) {
+        config.thinkingLevel = THINKING_LEVELS[activeEffort] ?? "medium";
+      } else {
+        if (activeEffort === "max" && options.model.startsWith("gemini-2.5-pro")) {
+          config.thinkingBudget = 32768;
+        } else {
+          config.thinkingBudget = THINKING_BUDGETS2[activeEffort] ?? 5e3;
+        }
       }
+      if (includeThoughts) config.includeThoughts = true;
+      generationConfig.thinkingConfig = config;
+      if (!body.generationConfig) body.generationConfig = generationConfig;
+    } else if (includeThoughts) {
+      generationConfig.thinkingConfig = { includeThoughts: true };
+      if (!body.generationConfig) body.generationConfig = generationConfig;
     }
     if (options.toolChoice !== void 0) {
       body.toolConfig = { functionCallingConfig: this.mapToolChoice(options.toolChoice) };
     }
+    if (options.providerOptions) {
+      Object.assign(body, options.providerOptions);
+    }
     return body;
   }
   /**
@@ -1474,28 +1662,33 @@ var GeminiProvider = class {
     const result = [];
     for (const msg of messages) {
       if (msg.role === "assistant") {
-        const parts = [];
-        if (msg.content) {
-          parts.push({ text: msg.content });
-        }
-        if (msg.tool_calls && msg.tool_calls.length > 0) {
-          for (const tc of msg.tool_calls) {
-            let parsedArgs;
-            try {
-              parsedArgs = JSON.parse(tc.function.arguments);
-            } catch {
-              parsedArgs = {};
-            }
-            parts.push({
-              functionCall: {
-                name: tc.function.name,
-                args: parsedArgs
+        const rawParts = msg.providerMetadata?.geminiParts;
+        if (rawParts && rawParts.length > 0) {
+          result.push({ role: "model", parts: rawParts });
+        } else {
+          const parts = [];
+          if (msg.content) {
+            parts.push({ text: msg.content });
+          }
+          if (msg.tool_calls && msg.tool_calls.length > 0) {
+            for (const tc of msg.tool_calls) {
+              let parsedArgs;
+              try {
+                parsedArgs = JSON.parse(tc.function.arguments);
+              } catch {
+                parsedArgs = {};
               }
-            });
+              parts.push({
+                functionCall: {
+                  name: tc.function.name,
+                  args: parsedArgs
+                }
+              });
+            }
+          }
+          if (parts.length > 0) {
+            result.push({ role: "model", parts });
           }
-        }
-        if (parts.length > 0) {
-          result.push({ role: "model", parts });
         }
       } else if (msg.role === "tool") {
         const functionName = toolCallIdToName.get(msg.tool_call_id) ?? "unknown";
@@ -1572,10 +1765,13 @@ var GeminiProvider = class {
   parseResponse(json, model) {
     const candidate = json.candidates?.[0];
     let content = "";
+    let thinkingContent = "";
     const toolCalls = [];
     if (candidate?.content?.parts) {
       for (const part of candidate.content.parts) {
-        if (part.text) {
+        if (part.thought && part.text) {
+          thinkingContent += part.text;
+        } else if (part.text) {
           content += part.text;
         } else if (part.functionCall) {
           toolCalls.push({
@@ -1590,18 +1786,24 @@ var GeminiProvider = class {
       }
     }
     const cachedTokens = json.usageMetadata?.cachedContentTokenCount;
+    const reasoningTokens = json.usageMetadata?.thoughtsTokenCount;
     const usage = json.usageMetadata ? {
       prompt_tokens: json.usageMetadata.promptTokenCount ?? 0,
       completion_tokens: json.usageMetadata.candidatesTokenCount ?? 0,
       total_tokens: json.usageMetadata.totalTokenCount ?? 0,
-      cached_tokens: cachedTokens && cachedTokens > 0 ? cachedTokens : void 0
+      cached_tokens: cachedTokens && cachedTokens > 0 ? cachedTokens : void 0,
+      reasoning_tokens: reasoningTokens && reasoningTokens > 0 ? reasoningTokens : void 0
     } : void 0;
     const cost = usage ? estimateGeminiCost(model, usage.prompt_tokens, usage.completion_tokens, usage.cached_tokens) : void 0;
+    const rawParts = candidate?.content?.parts;
+    const providerMetadata = rawParts ? { geminiParts: rawParts } : void 0;
     return {
       content,
+      thinking_content: thinkingContent || void 0,
       tool_calls: toolCalls.length > 0 ? toolCalls : void 0,
       usage,
-      cost
+      cost,
+      providerMetadata
     };
   }
   // ---------------------------------------------------------------------------
@@ -1612,6 +1814,7 @@ var GeminiProvider = class {
     const decoder = new TextDecoder();
     let buffer = "";
     let usage;
+    const accumulatedParts = [];
     try {
       while (true) {
         const { done, value } = await reader.read();
@@ -1632,17 +1835,22 @@ var GeminiProvider = class {
           }
           if (chunk.usageMetadata) {
             const cached = chunk.usageMetadata.cachedContentTokenCount;
+            const reasoning = chunk.usageMetadata.thoughtsTokenCount;
             usage = {
               prompt_tokens: chunk.usageMetadata.promptTokenCount ?? 0,
               completion_tokens: chunk.usageMetadata.candidatesTokenCount ?? 0,
               total_tokens: chunk.usageMetadata.totalTokenCount ?? 0,
-              cached_tokens: cached && cached > 0 ? cached : void 0
+              cached_tokens: cached && cached > 0 ? cached : void 0,
+              reasoning_tokens: reasoning && reasoning > 0 ? reasoning : void 0
             };
           }
           const candidate = chunk.candidates?.[0];
           if (candidate?.content?.parts) {
             for (const part of candidate.content.parts) {
-              if (part.text) {
+              accumulatedParts.push(part);
+              if (part.thought && part.text) {
+                yield { type: "thinking_delta", content: part.text };
+              } else if (part.text) {
                 yield { type: "text_delta", content: part.text };
               } else if (part.functionCall) {
                 yield {
@@ -1656,7 +1864,8 @@ var GeminiProvider = class {
           }
         }
       }
-      yield { type: "done", usage };
+      const providerMetadata = accumulatedParts.length > 0 ? { geminiParts: accumulatedParts } : void 0;
+      yield { type: "done", usage, providerMetadata };
     } finally {
       reader.releaseLock();
     }
@@ -2029,7 +2238,7 @@ function estimateMessagesTokens(messages) {
   }
   return total;
 }
-var WorkflowContext = class {
+var WorkflowContext = class _WorkflowContext {
   input;
   executionId;
   metadata;
@@ -2082,6 +2291,37 @@ var WorkflowContext = class {
       this.summaryCache = init.metadata.summaryCache;
     }
   }
+  /**
+   * Create a child context for nested agent invocations (e.g., agent-as-tool).
+   * Shares: budget tracking, abort signals, trace emission, provider registry,
+   *         state store, span manager, memory manager, MCP manager, config,
+   *         awaitHuman handler, pending decisions, tool overrides.
+   * Isolates: session history, step counter, streaming callbacks (onToken, onAgentStart, onToolCall).
+   */
+  createChildContext() {
+    return new _WorkflowContext({
+      input: this.input,
+      executionId: this.executionId,
+      config: this.config,
+      providerRegistry: this.providerRegistry,
+      metadata: { ...this.metadata },
+      // Shared infrastructure
+      budgetContext: this.budgetContext,
+      stateStore: this.stateStore,
+      mcpManager: this.mcpManager,
+      spanManager: this.spanManager,
+      memoryManager: this.memoryManager,
+      onTrace: this.onTrace,
+      onAgentCallComplete: this.onAgentCallComplete,
+      awaitHumanHandler: this.awaitHumanHandler,
+      pendingDecisions: this.pendingDecisions,
+      toolOverrides: this.toolOverrides,
+      signal: this.signal,
+      workflowName: this.workflowName
+      // Isolated: sessionHistory (empty), stepCounter (0),
+      // onToken (null), onAgentStart (null), onToolCall (null)
+    });
+  }
   /**
    * Resolve the current abort signal.
    * Branch-scoped signals (from race/spawn/map/budget) in AsyncLocalStorage
@@ -2144,10 +2384,12 @@ var WorkflowContext = class {
         promptVersion: agent2._config.version,
         temperature: options?.temperature ?? agent2._config.temperature,
         maxTokens: options?.maxTokens ?? agent2._config.maxTokens ?? 4096,
-        thinking: options?.thinking ?? agent2._config.thinking,
-        reasoningEffort: options?.reasoningEffort ?? agent2._config.reasoningEffort,
+        effort: options?.effort ?? agent2._config.effort,
+        thinkingBudget: options?.thinkingBudget ?? agent2._config.thinkingBudget,
+        includeThoughts: options?.includeThoughts ?? agent2._config.includeThoughts,
         toolChoice: options?.toolChoice ?? agent2._config.toolChoice,
-        stop: options?.stop ?? agent2._config.stop
+        stop: options?.stop ?? agent2._config.stop,
+        providerOptions: options?.providerOptions ?? agent2._config.providerOptions
       });
       return result;
     });
@@ -2170,7 +2412,21 @@ var WorkflowContext = class {
     const modelUri = agent2.resolveModel(resolveCtx);
     const systemPrompt = agent2.resolveSystem(resolveCtx);
     const { provider, model } = this.providerRegistry.resolve(modelUri, this.config);
-    const toolDefs = this.buildToolDefs(agent2);
+    let resolvedHandoffs;
+    if (typeof agent2._config.handoffs === "function") {
+      try {
+        resolvedHandoffs = agent2._config.handoffs(resolveCtx);
+      } catch (err) {
+        this.log("handoff_resolve_error", {
+          agent: agent2._name,
+          error: err instanceof Error ? err.message : String(err)
+        });
+        resolvedHandoffs = void 0;
+      }
+    } else {
+      resolvedHandoffs = agent2._config.handoffs;
+    }
+    const toolDefs = this.buildToolDefs(agent2, resolvedHandoffs);
     const messages = [];
     if (systemPrompt) {
       messages.push({ role: "system", content: systemPrompt });
@@ -2272,21 +2528,17 @@ Please fix and try again.`;
         throw new TimeoutError("ctx.ask()", timeoutMs);
       }
       turns++;
-      const thinking = options?.thinking ?? agent2._config.thinking;
-      if (thinking && typeof thinking === "object" && thinking.budgetTokens <= 0) {
-        throw new Error(
-          `thinking.budgetTokens must be a positive number, got ${thinking.budgetTokens}`
-        );
-      }
       const chatOptions = {
         model,
         temperature: options?.temperature ?? agent2._config.temperature,
         tools: toolDefs.length > 0 ? toolDefs : void 0,
         maxTokens: options?.maxTokens ?? agent2._config.maxTokens ?? 4096,
-        thinking,
-        reasoningEffort: options?.reasoningEffort ?? agent2._config.reasoningEffort,
+        effort: options?.effort ?? agent2._config.effort,
+        thinkingBudget: options?.thinkingBudget ?? agent2._config.thinkingBudget,
+        includeThoughts: options?.includeThoughts ?? agent2._config.includeThoughts,
         toolChoice: options?.toolChoice ?? agent2._config.toolChoice,
         stop: options?.stop ?? agent2._config.stop,
+        providerOptions: options?.providerOptions ?? agent2._config.providerOptions,
         signal: this.currentSignal
       };
       if (options?.schema && toolDefs.length === 0) {
@@ -2298,10 +2550,14 @@ Please fix and try again.`;
         let content2 = "";
         const toolCalls = [];
         const toolCallBuffers = /* @__PURE__ */ new Map();
+        let streamProviderMetadata;
+        let thinkingContent = "";
         for await (const chunk of provider.stream(currentMessages, chatOptions)) {
           if (chunk.type === "text_delta") {
             content2 += chunk.content;
             this.onToken(chunk.content);
+          } else if (chunk.type === "thinking_delta") {
+            thinkingContent += chunk.content;
           } else if (chunk.type === "tool_call_delta") {
             let buffer = toolCallBuffers.get(chunk.id);
             if (!buffer) {
@@ -2311,6 +2567,7 @@ Please fix and try again.`;
             if (chunk.name) buffer.name = chunk.name;
             if (chunk.arguments) buffer.arguments += chunk.arguments;
           } else if (chunk.type === "done") {
+            streamProviderMetadata = chunk.providerMetadata;
             if (chunk.usage) {
               response = {
                 content: content2,
@@ -2337,6 +2594,12 @@ Please fix and try again.`;
         if (toolCalls.length > 0) {
           response.tool_calls = toolCalls;
         }
+        if (streamProviderMetadata) {
+          response.providerMetadata = streamProviderMetadata;
+        }
+        if (thinkingContent) {
+          response.thinking_content = thinkingContent;
+        }
       } else {
         response = await provider.chat(currentMessages, chatOptions);
       }
@@ -2367,13 +2630,14 @@ Please fix and try again.`;
         currentMessages.push({
           role: "assistant",
           content: response.content || "",
-          tool_calls: response.tool_calls
+          tool_calls: response.tool_calls,
+          ...response.providerMetadata ? { providerMetadata: response.providerMetadata } : {}
         });
         for (const toolCall of response.tool_calls) {
           const toolName = toolCall.function.name;
           if (toolName.startsWith("handoff_to_")) {
             const targetName = toolName.replace("handoff_to_", "");
-            const descriptor = agent2._config.handoffs?.find((h) => h.agent._name === targetName);
+            const descriptor = resolvedHandoffs?.find((h) => h.agent._name === targetName);
             if (descriptor) {
               const mode = descriptor.mode ?? "oneway";
               let handoffPrompt = prompt;
@@ -2626,8 +2890,9 @@ Please fix and try again.`;
                 resultContent2 = JSON.stringify(toolResult2);
               }
             } else if (tool2) {
+              const childCtx = this.createChildContext();
               try {
-                toolResult2 = await tool2._execute(toolArgs);
+                toolResult2 = await tool2._execute(toolArgs, childCtx);
               } catch (err) {
                 toolResult2 = { error: err instanceof Error ? err.message : String(err) };
               }
@@ -2707,7 +2972,8 @@ Please fix and try again.`;
               guardrailOutputRetries++;
               currentMessages.push({
                 role: "assistant",
-                content
+                content,
+                ...response.providerMetadata ? { providerMetadata: response.providerMetadata } : {}
               });
               currentMessages.push({
                 role: "system",
@@ -2728,6 +2994,7 @@ Please fix and try again.`;
         try {
           const parsed = JSON.parse(stripMarkdownFences(content));
           const validated = options.schema.parse(parsed);
+          this.pushAssistantToSessionHistory(content, response.providerMetadata);
           return validated;
         } catch (err) {
           const maxRetries = options.retries ?? 3;
@@ -2754,11 +3021,23 @@ Please fix and try again.`;
           throw new VerifyError(content, zodErr, maxRetries);
         }
       }
+      this.pushAssistantToSessionHistory(content, response.providerMetadata);
       return content;
     }
     throw new MaxTurnsError("ctx.ask()", maxTurns);
   }
-  buildToolDefs(agent2) {
+  /**
+   * Push the final assistant message into session history, preserving providerMetadata
+   * (e.g., Gemini thought signatures needed for multi-turn reasoning context).
+   */
+  pushAssistantToSessionHistory(content, providerMetadata) {
+    this.sessionHistory.push({
+      role: "assistant",
+      content,
+      ...providerMetadata ? { providerMetadata } : {}
+    });
+  }
+  buildToolDefs(agent2, resolvedHandoffs) {
     const defs = [];
     if (agent2._config.tools) {
       for (const tool2 of agent2._config.tools) {
@@ -2772,8 +3051,8 @@ Please fix and try again.`;
         });
       }
     }
-    if (agent2._config.handoffs) {
-      for (const { agent: handoffAgent, description, mode } of agent2._config.handoffs) {
+    if (resolvedHandoffs) {
+      for (const { agent: handoffAgent, description, mode } of resolvedHandoffs) {
         const isRoundtrip = mode === "roundtrip";
         const defaultDesc = isRoundtrip ? `Delegate a task to ${handoffAgent._name} and receive the result back` : `Hand off the conversation to ${handoffAgent._name}`;
         defs.push({
@@ -3462,6 +3741,79 @@ ${summaryResponse.content}`
     const sessionId = this.metadata?.sessionId;
     await this.memoryManager.forget(key, this.stateStore, sessionId, options);
   }
+  // ── ctx.delegate() ──────────────────────────────────────────────────
+  /**
+   * Select the best agent from a list of candidates and invoke it.
+   * Creates a temporary router agent that uses handoffs to pick the right specialist.
+   *
+   * This is convenience sugar over creating a router agent with dynamic handoffs.
+   * For full control over the router's behavior, create the router agent explicitly.
+   *
+   * @param agents - Candidate agents to choose from (at least 1)
+   * @param prompt - The prompt to send to the selected agent
+   * @param options - Optional: schema, routerModel, metadata, retries
+   */
+  async delegate(agents, prompt, options) {
+    if (agents.length === 0) {
+      throw new Error("ctx.delegate() requires at least one candidate agent");
+    }
+    const names = /* @__PURE__ */ new Set();
+    for (const a of agents) {
+      if (names.has(a._name)) {
+        throw new Error(
+          `ctx.delegate() received duplicate agent name '${a._name}'. All candidate agents must have unique names.`
+        );
+      }
+      names.add(a._name);
+    }
+    if (agents.length === 1) {
+      return this.ask(agents[0], prompt, {
+        schema: options?.schema,
+        retries: options?.retries,
+        metadata: options?.metadata
+      });
+    }
+    const resolveCtx = options?.metadata ? { metadata: { ...this.metadata, ...options.metadata } } : { metadata: this.metadata };
+    const routerModelUri = options?.routerModel ?? agents[0].resolveModel(resolveCtx);
+    const handoffs = agents.map((a) => {
+      let description;
+      try {
+        description = a.resolveSystem(resolveCtx).slice(0, 200);
+      } catch {
+        description = `Agent: ${a._name}`;
+      }
+      return { agent: a, description };
+    });
+    const routerSystem = "Route to the best agent for this task. Always hand off; never answer directly.";
+    const routerAgent = {
+      _config: {
+        model: routerModelUri,
+        system: routerSystem,
+        temperature: 0,
+        handoffs,
+        maxTurns: 2
+      },
+      _name: "_delegate_router",
+      ask: async () => {
+        throw new Error("Direct invocation not supported on delegate router");
+      },
+      resolveModel: () => routerModelUri,
+      resolveSystem: () => routerSystem
+    };
+    this.emitTrace({
+      type: "delegate",
+      agent: "_delegate_router",
+      data: {
+        candidates: agents.map((a) => a._name),
+        routerModel: routerModelUri
+      }
+    });
+    return this.ask(routerAgent, prompt, {
+      schema: options?.schema,
+      retries: options?.retries,
+      metadata: options?.metadata
+    });
+  }
   // ── Private ───────────────────────────────────────────────────────────
   emitTrace(partial) {
     let data = partial.data;
@@ -4071,11 +4423,13 @@ var Session = class _Session {
         ...cachedSummary ? { summaryCache: cachedSummary } : {}
       }
     });
-    const assistantMessage = {
-      role: "assistant",
-      content: typeof result === "string" ? result : JSON.stringify(result)
-    };
-    history.push(assistantMessage);
+    const lastMsg = history[history.length - 1];
+    if (!(lastMsg && lastMsg.role === "assistant")) {
+      history.push({
+        role: "assistant",
+        content: typeof result === "string" ? result : JSON.stringify(result)
+      });
+    }
     if (this.options.persist !== false) {
       await this.store.saveSession(this.sessionId, history);
     }
@@ -4118,10 +4472,13 @@ var Session = class _Session {
       }
     });
     const updateHistory = async (result) => {
-      history.push({
-        role: "assistant",
-        content: typeof result === "string" ? result : JSON.stringify(result)
-      });
+      const lastMsg = history[history.length - 1];
+      if (!(lastMsg && lastMsg.role === "assistant")) {
+        history.push({
+          role: "assistant",
+          content: typeof result === "string" ? result : JSON.stringify(result)
+        });
+      }
       if (this.options.persist !== false) {
         await this.store.saveSession(this.sessionId, history);
       }
@@ -5011,6 +5368,24 @@ var AxlRuntime = class extends import_node_events2.EventEmitter {
   getExecutions() {
     return [...this.executions.values()];
   }
+  /**
+   * Create a lightweight WorkflowContext for ad-hoc use (tool testing, prototyping).
+   * The context has access to the runtime's providers, state store, and MCP manager
+   * but no session history, streaming callbacks, or budget tracking.
+   */
+  createContext(options) {
+    return new WorkflowContext({
+      input: void 0,
+      executionId: (0, import_node_crypto2.randomUUID)(),
+      metadata: options?.metadata,
+      config: this.config,
+      providerRegistry: this.providerRegistry,
+      stateStore: this.stateStore,
+      mcpManager: this.mcpManager,
+      spanManager: this.spanManager,
+      memoryManager: this.memoryManager
+    });
+  }
   /** Register a custom provider instance. */
   registerProvider(name, provider) {
     this.providerRegistry.registerInstance(name, provider);
@@ -5709,6 +6084,7 @@ function cosineSimilarity2(a, b) {
   agent,
   createSpanManager,
   defineConfig,
+  resolveThinkingOptions,
   tool,
   workflow,
   zodToJsonSchema