whale-code 6.5.5 → 6.5.6

package/dist/server/lib/server-subagent.js

@@ -4,42 +4,49 @@
  * The parent agent loop injects all dependencies; this module has zero imports
  * from index.ts (avoids circular deps) and is fully testable in isolation.
  */
+
 import { randomUUID } from "node:crypto";
-import { LoopDetector, estimateCostUsd, sanitizeError, isRetryableError, addPromptCaching, } from "../../shared/agent-core.js";
+import { LoopDetector, estimateCostUsd, sanitizeError, isRetryableError, addPromptCaching, AGENT_DEFAULTS } from "../../shared/agent-core.js";
 import { buildAPIRequest } from "../../shared/api-client.js";
 import { MODELS, MODEL_MAP } from "../../shared/constants.js";
 import { dispatchTools } from "../../shared/tool-dispatch.js";
+
+// ============================================================================
+// TYPES
+// ============================================================================
+
 // ============================================================================
 // TOOL DEFINITION — exposed to parent Opus agent
 // ============================================================================
+
 export const DELEGATE_TASK_TOOL_DEF = {
-    name: "delegate_task",
-    description: "Delegate a focused task to a faster sub-agent. Use for: data lookups, " +
-        "multi-step tool chains, analytics queries, inventory checks, customer lookups. " +
-        "Do NOT delegate decisions requiring your judgment or creative writing.",
-    input_schema: {
-        type: "object",
-        properties: {
-            prompt: {
-                type: "string",
-                description: "Clear instruction for the sub-agent. Include all context needed.",
-            },
-            model: {
-                type: "string",
-                enum: ["haiku", "sonnet"],
-                description: "haiku (fast/$1) for simple lookups. sonnet ($3) for multi-step chains. Default: haiku.",
-            },
-            max_turns: {
-                type: "number",
-                description: "Max agentic turns (1-12). Default: 6.",
-            },
-        },
-        required: ["prompt"],
+  name: "delegate_task",
+  description: "Delegate a focused task to a faster sub-agent. Use for: data lookups, " + "multi-step tool chains, analytics queries, inventory checks, customer lookups. " + "Do NOT delegate decisions requiring your judgment or creative writing.",
+  input_schema: {
+    type: "object",
+    properties: {
+      prompt: {
+        type: "string",
+        description: "Clear instruction for the sub-agent. Include all context needed."
+      },
+      model: {
+        type: "string",
+        enum: ["haiku", "sonnet"],
+        description: "haiku (fast/$1) for simple lookups. sonnet ($3) for multi-step chains. Default: haiku."
+      },
+      max_turns: {
+        type: "number",
+        description: "Max agentic turns (1-12). Default: 6."
+      }
     },
+    required: ["prompt"]
+  }
 };
+
 // ============================================================================
 // SUBAGENT SYSTEM PROMPT
 // ============================================================================
+
 const SUBAGENT_SYSTEM_PROMPT = `You are a focused tool executor. Your job is to complete the delegated task using the available tools, then return a clear summary of the results.
 
 Rules:
@@ -48,161 +55,250 @@ Rules:
 - If a tool fails, try an alternative approach or report the failure
 - Be concise in your final response
 - Always include store_id in tool calls when it's available in your context`;
+
 // ============================================================================
 // CORE LOOP
 // ============================================================================
-const SUBAGENT_MAX_OUTPUT_TOKENS = 8192;
+
+const SUBAGENT_MAX_OUTPUT_TOKENS = AGENT_DEFAULTS.subagentMaxTokens;
 export async function runServerSubagent(opts) {
-    const { anthropic, prompt, model: modelAlias, maxTurns, tools, executeTool, onProgress, clientDisconnected, startedAt, maxDurationMs, storeId, } = opts;
-    const subagentId = `sub-${randomUUID().slice(0, 8)}`;
-    const modelId = MODEL_MAP[modelAlias] || MODELS.HAIKU;
-    const loopDetector = new LoopDetector();
-    const toolsUsed = [];
-    let totalIn = 0;
-    let totalOut = 0;
-    let cacheCreation = 0;
-    let cacheRead = 0;
-    // Enable thinking for capable models; adjust max tokens for Opus
-    const shouldThink = modelAlias === "opus" || modelAlias === "sonnet";
-    const baseMaxTokens = opts.maxOutputTokens ?? SUBAGENT_MAX_OUTPUT_TOKENS;
-    const effectiveMaxTokens = modelAlias === "opus" ? Math.max(16384, baseMaxTokens) : baseMaxTokens;
-    const effectiveTemperature = opts.temperature ?? 0.3;
-    // Context management for subagent: clear at 60K, keep 2, no compaction
-    const apiConfig = buildAPIRequest({
-        model: modelId,
-        contextProfile: "subagent",
-        thinkingEnabled: shouldThink,
-        maxOutputTokens: effectiveMaxTokens,
+  const {
+    anthropic,
+    prompt,
+    model: modelAlias,
+    maxTurns,
+    tools,
+    executeTool,
+    onProgress,
+    clientDisconnected,
+    startedAt,
+    maxDurationMs,
+    storeId
+  } = opts;
+  const subagentId = `sub-${randomUUID().slice(0, 8)}`;
+  const modelId = MODEL_MAP[modelAlias] || MODELS.HAIKU;
+  const loopDetector = new LoopDetector();
+  const toolsUsed = [];
+  let totalIn = 0;
+  let totalOut = 0;
+  let cacheCreation = 0;
+  let cacheRead = 0;
+
+  // Enable thinking for capable models; adjust max tokens for Opus
+  const shouldThink = modelAlias === "opus" || modelAlias === "sonnet";
+  const baseMaxTokens = opts.maxOutputTokens ?? SUBAGENT_MAX_OUTPUT_TOKENS;
+  const effectiveMaxTokens = modelAlias === "opus" ? Math.max(16384, baseMaxTokens) : baseMaxTokens;
+  const effectiveTemperature = opts.temperature ?? AGENT_DEFAULTS.subagentTemperature;
+
+  // Context management for subagent: clear at 60K, keep 2, no compaction
+  const apiConfig = buildAPIRequest({
+    model: modelId,
+    contextProfile: "subagent",
+    thinkingEnabled: shouldThink,
+    maxOutputTokens: effectiveMaxTokens
+  });
+  const system = storeId ? `${SUBAGENT_SYSTEM_PROMPT}\n\nstore_id: ${storeId}` : SUBAGENT_SYSTEM_PROMPT;
+  const messages = [{
+    role: "user",
+    content: prompt
+  }];
+  onProgress?.({
+    subagentId,
+    event: "started",
+    model: modelAlias
+  });
+  let turn = 0;
+  let lastText = "";
+  while (turn < maxTurns) {
+    // Abort checks
+    if (clientDisconnected.value) {
+      return makeResult(false, "Client disconnected", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
+    }
+    if (Date.now() - startedAt > maxDurationMs) {
+      return makeResult(false, "Parent request timeout", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
+    }
+    turn++;
+    loopDetector.resetTurn();
+    onProgress?.({
+      subagentId,
+      event: "turn",
+      turn,
+      maxTurns
     });
-    const system = storeId
-        ? `${SUBAGENT_SYSTEM_PROMPT}\n\nstore_id: ${storeId}`
-        : SUBAGENT_SYSTEM_PROMPT;
-    const messages = [{ role: "user", content: prompt }];
-    onProgress?.({ subagentId, event: "started", model: modelAlias });
-    let turn = 0;
-    let lastText = "";
-    while (turn < maxTurns) {
-        // Abort checks
-        if (clientDisconnected.value) {
-            return makeResult(false, "Client disconnected", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
-        }
-        if (Date.now() - startedAt > maxDurationMs) {
-            return makeResult(false, "Parent request timeout", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
-        }
-        turn++;
-        loopDetector.resetTurn();
-        onProgress?.({ subagentId, event: "turn", turn, maxTurns });
-        // Apply prompt caching to tools and messages for cache hits on repeated context
-        const { tools: cachedTools, messages: cachedMessages } = addPromptCaching(tools, messages);
-        // Non-streaming API call with retry
-        let response;
-        try {
-            response = await withSubagentRetry(async () => {
-                return await anthropic.beta.messages.create({
-                    model: modelId,
-                    max_tokens: apiConfig.maxTokens,
-                    temperature: shouldThink ? 1 : effectiveTemperature, // Anthropic requires temp=1 with thinking
-                    system,
-                    tools: cachedTools,
-                    messages: cachedMessages,
-                    betas: apiConfig.betas,
-                    context_management: apiConfig.contextManagement,
-                    ...(apiConfig.thinking ? { thinking: apiConfig.thinking } : {}),
-                });
-            });
-        }
-        catch (err) {
-            const errMsg = sanitizeError(err);
-            onProgress?.({ subagentId, event: "error", output: errMsg });
-            return makeResult(false, `API error: ${errMsg}`, totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
-        }
-        // Track tokens (including cache for accurate cost)
-        totalIn += response.usage?.input_tokens || 0;
-        totalOut += response.usage?.output_tokens || 0;
-        cacheCreation += response.usage?.cache_creation_input_tokens ?? 0;
-        cacheRead += response.usage?.cache_read_input_tokens ?? 0;
-        // Extract text and tool_use blocks
-        const textBlocks = [];
-        const toolUseBlocks = [];
-        for (const block of response.content) {
-            if (block.type === "text") {
-                textBlocks.push(block.text);
-            }
-            else if (block.type === "tool_use") {
-                toolUseBlocks.push({ id: block.id, name: block.name, input: block.input });
-            }
-        }
-        lastText = textBlocks.join("\n");
-        // No tool calls → done
-        if (toolUseBlocks.length === 0) {
-            onProgress?.({ subagentId, event: "done", output: lastText });
-            return makeResult(true, lastText, totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
-        }
-        // Execute tools in parallel via dispatchTools
-        const toolExecutor = async (name, input) => {
-            toolsUsed.push(name);
-            onProgress?.({ subagentId, event: "tool_start", toolName: name });
-            try {
-                const result = await executeTool(name, input);
-                onProgress?.({ subagentId, event: "tool_result", toolName: name, toolSuccess: result.success });
-                return {
-                    success: result.success,
-                    output: result.success ? JSON.stringify(result.data) : (result.error || "Unknown error"),
-                };
-            }
-            catch (err) {
-                const errMsg = sanitizeError(err);
-                onProgress?.({ subagentId, event: "tool_result", toolName: name, toolSuccess: false });
-                return { success: false, output: errMsg };
-            }
-        };
-        const { results: toolResults, bailOut, bailMessage } = await dispatchTools(toolUseBlocks, toolExecutor, { loopDetector, maxConcurrent: 3 });
-        if (bailOut) {
-            onProgress?.({ subagentId, event: "error", output: bailMessage });
-            return makeResult(false, bailMessage || "Too many errors", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
-        }
-        // Append for next turn
-        messages.push({
-            role: "assistant",
-            content: response.content,
+
+    // Apply prompt caching to tools and messages for cache hits on repeated context
+    const {
+      tools: cachedTools,
+      messages: cachedMessages
+    } = addPromptCaching(tools, messages);
+
+    // Non-streaming API call with retry
+    let response;
+    try {
+      response = await withSubagentRetry(async () => {
+        return await anthropic.beta.messages.create({
+          model: modelId,
+          max_tokens: apiConfig.maxTokens,
+          temperature: shouldThink ? 1 : effectiveTemperature,
+          // Anthropic requires temp=1 with thinking
+          system,
+          tools: cachedTools,
+          messages: cachedMessages,
+          betas: apiConfig.betas,
+          context_management: apiConfig.contextManagement,
+          ...(apiConfig.thinking ? {
+            thinking: apiConfig.thinking
+          } : {})
+        });
+      });
+    } catch (err) {
+      const errMsg = sanitizeError(err);
+      onProgress?.({
+        subagentId,
+        event: "error",
+        output: errMsg
+      });
+      return makeResult(false, `API error: ${errMsg}`, totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
+    }
+
+    // Track tokens (including cache for accurate cost)
+    totalIn += response.usage?.input_tokens || 0;
+    totalOut += response.usage?.output_tokens || 0;
+    cacheCreation += response.usage?.cache_creation_input_tokens ?? 0;
+    cacheRead += response.usage?.cache_read_input_tokens ?? 0;
+
+    // Extract text and tool_use blocks
+    const textBlocks = [];
+    const toolUseBlocks = [];
+    for (const block of response.content) {
+      if (block.type === "text") {
+        textBlocks.push(block.text);
+      } else if (block.type === "tool_use") {
+        toolUseBlocks.push({
+          id: block.id,
+          name: block.name,
+          input: block.input
+        });
+      }
+    }
+    lastText = textBlocks.join("\n");
+
+    // No tool calls → done
+    if (toolUseBlocks.length === 0) {
+      onProgress?.({
+        subagentId,
+        event: "done",
+        output: lastText
+      });
+      return makeResult(true, lastText, totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
+    }
+
+    // Execute tools in parallel via dispatchTools
+    const toolExecutor = async (name, input) => {
+      toolsUsed.push(name);
+      onProgress?.({
+        subagentId,
+        event: "tool_start",
+        toolName: name
+      });
+      try {
+        const result = await executeTool(name, input);
+        onProgress?.({
+          subagentId,
+          event: "tool_result",
+          toolName: name,
+          toolSuccess: result.success
         });
-        messages.push({
-            role: "user",
-            content: toolResults,
+        return {
+          success: result.success,
+          output: result.success ? JSON.stringify(result.data) : result.error || "Unknown error"
+        };
+      } catch (err) {
+        const errMsg = sanitizeError(err);
+        onProgress?.({
+          subagentId,
+          event: "tool_result",
+          toolName: name,
+          toolSuccess: false
         });
+        return {
+          success: false,
+          output: errMsg
+        };
+      }
+    };
+    const {
+      results: toolResults,
+      bailOut,
+      bailMessage
+    } = await dispatchTools(toolUseBlocks, toolExecutor, {
+      loopDetector,
+      maxConcurrent: AGENT_DEFAULTS.maxConcurrentTools,
+      batchErrorLimit: AGENT_DEFAULTS.loopBatchErrorLimit
+    });
+    if (bailOut) {
+      onProgress?.({
+        subagentId,
+        event: "error",
+        output: bailMessage
+      });
+      return makeResult(false, bailMessage || "Too many errors", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
     }
-    // Exhausted max turns
-    onProgress?.({ subagentId, event: "done", output: lastText });
-    return makeResult(true, lastText || "(subagent completed without text response)", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
+
+    // Append for next turn
+    messages.push({
+      role: "assistant",
+      content: response.content
+    });
+    messages.push({
+      role: "user",
+      content: toolResults
+    });
+  }
+
+  // Exhausted max turns
+  onProgress?.({
+    subagentId,
+    event: "done",
+    output: lastText
+  });
+  return makeResult(true, lastText || "(subagent completed without text response)", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);
 }
+
 // ============================================================================
 // HELPERS
 // ============================================================================
+
 function makeResult(success, output, inputTokens, outputTokens, modelId, toolsUsed, turnCount, cacheReadTokens = 0, cacheCreationTokens = 0) {
-    return {
-        success,
-        output,
-        tokensUsed: { input: inputTokens, output: outputTokens, cacheRead: cacheReadTokens, cacheCreation: cacheCreationTokens },
-        costUsd: estimateCostUsd(inputTokens, outputTokens, modelId, 0, cacheReadTokens, cacheCreationTokens),
-        toolsUsed: [...new Set(toolsUsed)],
-        turnCount,
-        stopReason: "end_turn", // subagent always runs to completion or error
-    };
+  return {
+    success,
+    output,
+    tokensUsed: {
+      input: inputTokens,
+      output: outputTokens,
+      cacheRead: cacheReadTokens,
+      cacheCreation: cacheCreationTokens
+    },
+    costUsd: estimateCostUsd(inputTokens, outputTokens, modelId, 0, cacheReadTokens, cacheCreationTokens),
+    toolsUsed: [...new Set(toolsUsed)],
+    turnCount,
+    stopReason: "end_turn" // subagent always runs to completion or error
+  };
 }
 async function withSubagentRetry(fn, maxRetries = 2) {
-    let lastError;
-    for (let attempt = 0; attempt <= maxRetries; attempt++) {
-        try {
-            return await fn();
-        }
-        catch (err) {
-            lastError = err;
-            if (attempt < maxRetries && isRetryableError(err)) {
-                await new Promise((resolve) => setTimeout(resolve, 1000 * Math.pow(2, attempt)));
-                continue;
-            }
-            throw err;
-        }
+  let lastError;
+  for (let attempt = 0; attempt <= maxRetries; attempt++) {
+    try {
+      return await fn();
+    } catch (err) {
+      lastError = err;
+      if (attempt < maxRetries && isRetryableError(err)) {
+        await new Promise(resolve => setTimeout(resolve, 1000 * Math.pow(2, attempt)));
+        continue;
+      }
+      throw err;
     }
-    throw lastError;
+  }
+  throw lastError;
 }
+//# sourceMappingURL=server-subagent.js.map

package/dist/server/lib/server-subagent.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"server-subagent.js","names":["randomUUID","LoopDetector","estimateCostUsd","sanitizeError","isRetryableError","addPromptCaching","AGENT_DEFAULTS","buildAPIRequest","MODELS","MODEL_MAP","dispatchTools","DELEGATE_TASK_TOOL_DEF","name","description","input_schema","type","properties","prompt","model","enum","max_turns","required","SUBAGENT_SYSTEM_PROMPT","SUBAGENT_MAX_OUTPUT_TOKENS","subagentMaxTokens","runServerSubagent","opts","anthropic","modelAlias","maxTurns","tools","executeTool","onProgress","clientDisconnected","startedAt","maxDurationMs","storeId","subagentId","slice","modelId","HAIKU","loopDetector","toolsUsed","totalIn","totalOut","cacheCreation","cacheRead","shouldThink","baseMaxTokens","maxOutputTokens","effectiveMaxTokens","Math","max","effectiveTemperature","temperature","subagentTemperature","apiConfig","contextProfile","thinkingEnabled","system","messages","role","content","event","turn","lastText","value","makeResult","Date","now","resetTurn","cachedTools","cachedMessages","response","withSubagentRetry","beta","create","max_tokens","maxTokens","betas","context_management","contextManagement","thinking","err","errMsg","output","usage","input_tokens","output_tokens","cache_creation_input_tokens","cache_read_input_tokens","textBlocks","toolUseBlocks","block","push","text","id","input","join","length","toolExecutor","toolName","result","toolSuccess","success","JSON","stringify","data","error","results","toolResults","bailOut","bailMessage","maxConcurrent","maxConcurrentTools","batchErrorLimit","loopBatchErrorLimit","inputTokens","outputTokens","turnCount","cacheReadTokens","cacheCreationTokens","tokensUsed","costUsd","Set","stopReason","fn","maxRetries","lastError","attempt","Promise","resolve","setTimeout","pow"],"sources":["../../../src/server/lib/server-subagent.ts"],"sourcesContent":["/**\n * Server Subagent — enables Opus to delegate tool execution to Haiku/Sonnet.\n *\n * The parent agent loop injects all dependencies; this module has zero imports\n * from index.ts (avoids circular deps) and is fully testable in isolation.\n */\n\nimport { randomUUID } from \"node:crypto\";\nimport type Anthropic from \"@anthropic-ai/sdk\";\nimport type { SupabaseClient } from \"@supabase/supabase-js\";\n\nimport {\n  LoopDetector,\n  estimateCostUsd,\n  sanitizeError,\n  isRetryableError,\n  addPromptCaching,\n  AGENT_DEFAULTS,\n} from \"../../shared/agent-core.js\";\nimport { buildAPIRequest } from \"../../shared/api-client.js\";\nimport { MODELS, MODEL_MAP } from \"../../shared/constants.js\";\nimport { dispatchTools } from \"../../shared/tool-dispatch.js\";\n\n// ============================================================================\n// TYPES\n// ============================================================================\n\nexport interface SubagentResult {\n  success: boolean;\n  output: string;\n  tokensUsed: { input: number; output: number; cacheRead: number; cacheCreation: number };\n  costUsd: number;\n  toolsUsed: string[];\n  turnCount: number;\n  stopReason: string;\n}\n\nexport interface SubagentProgressEvent {\n  subagentId: string;\n  event: \"started\" | \"tool_start\" | \"tool_result\" | \"turn\" | \"done\" | \"error\";\n  model?: string;\n  toolName?: string;\n  toolSuccess?: boolean;\n  turn?: number;\n  maxTurns?: number;\n  output?: string;\n}\n\nexport type SubagentProgressCallback = (event: SubagentProgressEvent) => void;\n\nexport interface RunServerSubagentOptions {\n  anthropic: Anthropic;\n  supabase: SupabaseClient;\n  storeId: string | undefined;\n  prompt: string;\n  model: \"haiku\" | \"sonnet\" | \"opus\";\n  maxTurns: number;\n  tools: Array<{ name: string; description: string; input_schema: Record<string, unknown> }>;\n  executeTool: (\n    toolName: string,\n    args: Record<string, unknown>,\n  ) => Promise<{ success: boolean; data?: unknown; error?: string }>;\n  onProgress?: SubagentProgressCallback;\n  clientDisconnected: { value: boolean };\n  startedAt: number;\n  maxDurationMs: number;\n  // Parent-controlled overrides from context_config\n  maxOutputTokens?: number;     // Default: SUBAGENT_MAX_OUTPUT_TOKENS (8192)\n  temperature?: number;         // Default: 0.3\n}\n\n// ============================================================================\n// TOOL DEFINITION — exposed to parent Opus agent\n// ============================================================================\n\nexport const DELEGATE_TASK_TOOL_DEF = {\n  name: \"delegate_task\",\n  description:\n    \"Delegate a focused task to a faster sub-agent. Use for: data lookups, \" +\n    \"multi-step tool chains, analytics queries, inventory checks, customer lookups. \" +\n    \"Do NOT delegate decisions requiring your judgment or creative writing.\",\n  input_schema: {\n    type: \"object\" as const,\n    properties: {\n      prompt: {\n        type: \"string\",\n        description: \"Clear instruction for the sub-agent. Include all context needed.\",\n      },\n      model: {\n        type: \"string\",\n        enum: [\"haiku\", \"sonnet\"],\n        description: \"haiku (fast/$1) for simple lookups. sonnet ($3) for multi-step chains. Default: haiku.\",\n      },\n      max_turns: {\n        type: \"number\",\n        description: \"Max agentic turns (1-12). Default: 6.\",\n      },\n    },\n    required: [\"prompt\"],\n  },\n};\n\n// ============================================================================\n// SUBAGENT SYSTEM PROMPT\n// ============================================================================\n\nconst SUBAGENT_SYSTEM_PROMPT = `You are a focused tool executor. Your job is to complete the delegated task using the available tools, then return a clear summary of the results.\n\nRules:\n- Execute tools as needed to fulfill the request\n- Return factual results — do not add opinions or recommendations\n- If a tool fails, try an alternative approach or report the failure\n- Be concise in your final response\n- Always include store_id in tool calls when it's available in your context`;\n\n// ============================================================================\n// CORE LOOP\n// ============================================================================\n\nconst SUBAGENT_MAX_OUTPUT_TOKENS = AGENT_DEFAULTS.subagentMaxTokens;\nexport async function runServerSubagent(opts: RunServerSubagentOptions): Promise<SubagentResult> {\n  const {\n    anthropic,\n    prompt,\n    model: modelAlias,\n    maxTurns,\n    tools,\n    executeTool,\n    onProgress,\n    clientDisconnected,\n    startedAt,\n    maxDurationMs,\n    storeId,\n  } = opts;\n\n  const subagentId = `sub-${randomUUID().slice(0, 8)}`;\n  const modelId = MODEL_MAP[modelAlias] || MODELS.HAIKU;\n  const loopDetector = new LoopDetector();\n  const toolsUsed: string[] = [];\n\n  let totalIn = 0;\n  let totalOut = 0;\n  let cacheCreation = 0;\n  let cacheRead = 0;\n\n  // Enable thinking for capable models; adjust max tokens for Opus\n  const shouldThink = modelAlias === \"opus\" || modelAlias === \"sonnet\";\n  const baseMaxTokens = opts.maxOutputTokens ?? SUBAGENT_MAX_OUTPUT_TOKENS;\n  const effectiveMaxTokens = modelAlias === \"opus\" ? Math.max(16384, baseMaxTokens) : baseMaxTokens;\n  const effectiveTemperature = opts.temperature ?? AGENT_DEFAULTS.subagentTemperature;\n\n  // Context management for subagent: clear at 60K, keep 2, no compaction\n  const apiConfig = buildAPIRequest({\n    model: modelId,\n    contextProfile: \"subagent\",\n    thinkingEnabled: shouldThink,\n    maxOutputTokens: effectiveMaxTokens,\n  });\n\n  const system = storeId\n    ? `${SUBAGENT_SYSTEM_PROMPT}\\n\\nstore_id: ${storeId}`\n    : SUBAGENT_SYSTEM_PROMPT;\n\n  const messages: Anthropic.MessageParam[] = [{ role: \"user\", content: prompt }];\n\n  onProgress?.({ subagentId, event: \"started\", model: modelAlias });\n\n  let turn = 0;\n  let lastText = \"\";\n\n  while (turn < maxTurns) {\n    // Abort checks\n    if (clientDisconnected.value) {\n      return makeResult(false, \"Client disconnected\", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);\n    }\n    if (Date.now() - startedAt > maxDurationMs) {\n      return makeResult(false, \"Parent request timeout\", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);\n    }\n\n    turn++;\n    loopDetector.resetTurn();\n\n    onProgress?.({ subagentId, event: \"turn\", turn, maxTurns });\n\n    // Apply prompt caching to tools and messages for cache hits on repeated context\n    const { tools: cachedTools, messages: cachedMessages } = addPromptCaching(\n      tools as any, messages as any,\n    );\n\n    // Non-streaming API call with retry\n    let response: Anthropic.Message;\n    try {\n      response = await withSubagentRetry(async () => {\n        return await anthropic.beta.messages.create({\n          model: modelId,\n          max_tokens: apiConfig.maxTokens,\n          temperature: shouldThink ? 1 : effectiveTemperature,  // Anthropic requires temp=1 with thinking\n          system,\n          tools: cachedTools as any,\n          messages: cachedMessages as any,\n          betas: apiConfig.betas,\n          context_management: apiConfig.contextManagement,\n          ...(apiConfig.thinking ? { thinking: apiConfig.thinking } : {}),\n        } as any) as unknown as Anthropic.Message;\n      });\n    } catch (err) {\n      const errMsg = sanitizeError(err);\n      onProgress?.({ subagentId, event: \"error\", output: errMsg });\n      return makeResult(false, `API error: ${errMsg}`, totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);\n    }\n\n    // Track tokens (including cache for accurate cost)\n    totalIn += response.usage?.input_tokens || 0;\n    totalOut += response.usage?.output_tokens || 0;\n    cacheCreation += (response.usage as any)?.cache_creation_input_tokens ?? 0;\n    cacheRead += (response.usage as any)?.cache_read_input_tokens ?? 0;\n\n    // Extract text and tool_use blocks\n    const textBlocks: string[] = [];\n    const toolUseBlocks: Array<{ id: string; name: string; input: Record<string, unknown> }> = [];\n\n    for (const block of response.content) {\n      if (block.type === \"text\") {\n        textBlocks.push(block.text);\n      } else if (block.type === \"tool_use\") {\n        toolUseBlocks.push({ id: block.id, name: block.name, input: block.input as Record<string, unknown> });\n      }\n    }\n\n    lastText = textBlocks.join(\"\\n\");\n\n    // No tool calls → done\n    if (toolUseBlocks.length === 0) {\n      onProgress?.({ subagentId, event: \"done\", output: lastText });\n      return makeResult(true, lastText, totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);\n    }\n\n    // Execute tools in parallel via dispatchTools\n    const toolExecutor = async (name: string, input: Record<string, unknown>) => {\n      toolsUsed.push(name);\n      onProgress?.({ subagentId, event: \"tool_start\", toolName: name });\n      try {\n        const result = await executeTool(name, input);\n        onProgress?.({ subagentId, event: \"tool_result\", toolName: name, toolSuccess: result.success });\n        return {\n          success: result.success,\n          output: result.success ? JSON.stringify(result.data) : (result.error || \"Unknown error\"),\n        };\n      } catch (err) {\n        const errMsg = sanitizeError(err);\n        onProgress?.({ subagentId, event: \"tool_result\", toolName: name, toolSuccess: false });\n        return { success: false, output: errMsg };\n      }\n    };\n\n    const { results: toolResults, bailOut, bailMessage } = await dispatchTools(\n      toolUseBlocks as any,\n      toolExecutor,\n      { loopDetector, maxConcurrent: AGENT_DEFAULTS.maxConcurrentTools, batchErrorLimit: AGENT_DEFAULTS.loopBatchErrorLimit },\n    );\n\n    if (bailOut) {\n      onProgress?.({ subagentId, event: \"error\", output: bailMessage });\n      return makeResult(false, bailMessage || \"Too many errors\", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);\n    }\n\n    // Append for next turn\n    messages.push({\n      role: \"assistant\",\n      content: response.content as any,\n    });\n    messages.push({\n      role: \"user\",\n      content: toolResults as any,\n    });\n  }\n\n  // Exhausted max turns\n  onProgress?.({ subagentId, event: \"done\", output: lastText });\n  return makeResult(true, lastText || \"(subagent completed without text response)\", totalIn, totalOut, modelId, toolsUsed, turn, cacheRead, cacheCreation);\n}\n\n// ============================================================================\n// HELPERS\n// ============================================================================\n\nfunction makeResult(\n  success: boolean,\n  output: string,\n  inputTokens: number,\n  outputTokens: number,\n  modelId: string,\n  toolsUsed: string[],\n  turnCount: number,\n  cacheReadTokens = 0,\n  cacheCreationTokens = 0,\n): SubagentResult {\n  return {\n    success,\n    output,\n    tokensUsed: { input: inputTokens, output: outputTokens, cacheRead: cacheReadTokens, cacheCreation: cacheCreationTokens },\n    costUsd: estimateCostUsd(inputTokens, outputTokens, modelId, 0, cacheReadTokens, cacheCreationTokens),\n    toolsUsed: [...new Set(toolsUsed)],\n    turnCount,\n    stopReason: \"end_turn\", // subagent always runs to completion or error\n  };\n}\n\nasync function withSubagentRetry<T>(fn: () => Promise<T>, maxRetries = 2): Promise<T> {\n  let lastError: unknown;\n  for (let attempt = 0; attempt <= maxRetries; attempt++) {\n    try {\n      return await fn();\n    } catch (err) {\n      lastError = err;\n      if (attempt < maxRetries && isRetryableError(err)) {\n        await new Promise((resolve) => setTimeout(resolve, 1000 * Math.pow(2, attempt)));\n        continue;\n      }\n      throw err;\n    }\n  }\n  throw lastError;\n}\n"],"mappings":"AAAA;AACA;AACA;AACA;AACA;AACA;;AAEA,SAASA,UAAU,QAAQ,aAAa;AAIxC,SACEC,YAAY,EACZC,eAAe,EACfC,aAAa,EACbC,gBAAgB,EAChBC,gBAAgB,EAChBC,cAAc,QACT,4BAA4B;AACnC,SAASC,eAAe,QAAQ,4BAA4B;AAC5D,SAASC,MAAM,EAAEC,SAAS,QAAQ,2BAA2B;AAC7D,SAASC,aAAa,QAAQ,+BAA+B;;AAE7D;AACA;AACA;;AA8CA;AACA;AACA;;AAEA,OAAO,MAAMC,sBAAsB,GAAG;EACpCC,IAAI,EAAE,eAAe;EACrBC,WAAW,EACT,wEAAwE,GACxE,iFAAiF,GACjF,wEAAwE;EAC1EC,YAAY,EAAE;IACZC,IAAI,EAAE,QAAiB;IACvBC,UAAU,EAAE;MACVC,MAAM,EAAE;QACNF,IAAI,EAAE,QAAQ;QACdF,WAAW,EAAE;MACf,CAAC;MACDK,KAAK,EAAE;QACLH,IAAI,EAAE,QAAQ;QACdI,IAAI,EAAE,CAAC,OAAO,EAAE,QAAQ,CAAC;QACzBN,WAAW,EAAE;MACf,CAAC;MACDO,SAAS,EAAE;QACTL,IAAI,EAAE,QAAQ;QACdF,WAAW,EAAE;MACf;IACF,CAAC;IACDQ,QAAQ,EAAE,CAAC,QAAQ;EACrB;AACF,CAAC;;AAED;AACA;AACA;;AAEA,MAAMC,sBAAsB,GAAG;AAC/B;AACA;AACA;AACA;AACA;AACA;AACA,4EAA4E;;AAE5E;AACA;AACA;;AAEA,MAAMC,0BAA0B,GAAGjB,cAAc,CAACkB,iBAAiB;AACnE,OAAO,eAAeC,iBAAiBA,CAACC,IAA8B,EAA2B;EAC/F,MAAM;IACJC,SAAS;IACTV,MAAM;IACNC,KAAK,EAAEU,UAAU;IACjBC,QAAQ;IACRC,KAAK;IACLC,WAAW;IACXC,UAAU;IACVC,kBAAkB;IAClBC,SAAS;IACTC,aAAa;IACbC;EACF,CAAC,GAAGV,IAAI;EAER,MAAMW,UAAU,GAAG,OAAOrC,UAAU,CAAC,CAAC,CAACsC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE;EACpD,MAAMC,OAAO,GAAG9B,SAAS,CAACmB,UAAU,CAAC,IAAIpB,MAAM,CAACgC,KAAK;EACrD,MAAMC,YAAY,GAAG,IAAIxC,YAAY,CAAC,CAAC;EACvC,MAAMyC,SAAmB,GAAG,EAAE;EAE9B,IAAIC,OAAO,GAAG,CAAC;EACf,IAAIC,QAAQ,GAAG,CAAC;EAChB,IAAIC,aAAa,GAAG,CAAC;EACrB,IAAIC,SAAS,GAAG,CAAC;;EAEjB;EACA,MAAMC,WAAW,GAAGnB,UAAU,KAAK,MAAM,IAAIA,UAAU,KAAK,QAAQ;EACpE,MAAMoB,aAAa,GAAGtB,IAAI,CAACuB,eAAe,IAAI1B,0BAA0B;EACxE,MAAM2B,kBAAkB,GAAGtB,UAAU,KAAK,MAAM,GAAGuB,IAAI,CAACC,GAAG,CAAC,KAAK,EAAEJ,aAAa,CAAC,GAAGA,aAAa;EACjG,MAAMK,oBAAoB,GAAG3B,IAAI,CAAC4B,WAAW,IAAIhD,cAAc,CAACiD,mBAAmB;;EAEnF;EACA,MAAMC,SAAS,GAAGjD,eAAe,CAAC;IAChCW,KAAK,EAAEqB,OAAO;IACdkB,cAAc,EAAE,UAAU;IAC1BC,eAAe,EAAEX,WAAW;IAC5BE,eAAe,EAAEC;EACnB,CAAC,CAAC;EAEF,MAAMS,MAAM,GAAGvB,OAAO,GAClB,GAAGd,sBAAsB,iBAAiBc,OAAO,EAAE,GACnDd,sBAAsB;EAE1B,MAAMsC,QAAkC,GAAG,CAAC;IAAEC,IAAI,EAAE,MAAM;IAAEC,OAAO,EAAE7C;EAAO,CAAC,CAAC;EAE9Ee,UAAU,GAAG;IAAEK,UAAU;IAAE0B,KAAK,EAAE,SAAS;IAAE7C,KAAK,EAAEU;EAAW,CAAC,CAAC;EAEjE,IAAIoC,IAAI,GAAG,CAAC;EACZ,IAAIC,QAAQ,GAAG,EAAE;EAEjB,OAAOD,IAAI,GAAGnC,QAAQ,EAAE;IACtB;IACA,IAAII,kBAAkB,CAACiC,KAAK,EAAE;MAC5B,OAAOC,UAAU,CAAC,KAAK,EAAE,qBAAqB,EAAExB,OAAO,EAAEC,QAAQ,EAAEL,OAAO,EAAEG,SAAS,EAAEsB,IAAI,EAAElB,SAAS,EAAED,aAAa,CAAC;IACxH;IACA,IAAIuB,IAAI,CAACC,GAAG,CAAC,CAAC,GAAGnC,SAAS,GAAGC,aAAa,EAAE;MAC1C,OAAOgC,UAAU,CAAC,KAAK,EAAE,wBAAwB,EAAExB,OAAO,EAAEC,QAAQ,EAAEL,OAAO,EAAEG,SAAS,EAAEsB,IAAI,EAAElB,SAAS,EAAED,aAAa,CAAC;IAC3H;IAEAmB,IAAI,EAAE;IACNvB,YAAY,CAAC6B,SAAS,CAAC,CAAC;IAExBtC,UAAU,GAAG;MAAEK,UAAU;MAAE0B,KAAK,EAAE,MAAM;MAAEC,IAAI;MAAEnC;IAAS,CAAC,CAAC;;IAE3D;IACA,MAAM;MAAEC,KAAK,EAAEyC,WAAW;MAAEX,QAAQ,EAAEY;IAAe,CAAC,GAAGnE,gBAAgB,CACvEyB,KAAK,EAAS8B,QAChB,CAAC;;IAED;IACA,IAAIa,QAA2B;IAC/B,IAAI;MACFA,QAAQ,GAAG,MAAMC,iBAAiB,CAAC,YAAY;QAC7C,OAAO,MAAM/C,SAAS,CAACgD,IAAI,CAACf,QAAQ,CAACgB,MAAM,CAAC;UAC1C1D,KAAK,EAAEqB,OAAO;UACdsC,UAAU,EAAErB,SAAS,CAACsB,SAAS;UAC/BxB,WAAW,EAAEP,WAAW,GAAG,CAAC,GAAGM,oBAAoB;UAAG;UACtDM,MAAM;UACN7B,KAAK,EAAEyC,WAAkB;UACzBX,QAAQ,EAAEY,cAAqB;UAC/BO,KAAK,EAAEvB,SAAS,CAACuB,KAAK;UACtBC,kBAAkB,EAAExB,SAAS,CAACyB,iBAAiB;UAC/C,IAAIzB,SAAS,CAAC0B,QAAQ,GAAG;YAAEA,QAAQ,EAAE1B,SAAS,CAAC0B;UAAS,CAAC,GAAG,CAAC,CAAC;QAChE,CAAQ,CAAC;MACX,CAAC,CAAC;IACJ,CAAC,CAAC,OAAOC,GAAG,EAAE;MACZ,MAAMC,MAAM,GAAGjF,aAAa,CAACgF,GAAG,CAAC;MACjCnD,UAAU,GAAG;QAAEK,UAAU;QAAE0B,KAAK,EAAE,OAAO;QAAEsB,MAAM,EAAED;MAAO,CAAC,CAAC;MAC5D,OAAOjB,UAAU,CAAC,KAAK,EAAE,cAAciB,MAAM,EAAE,EAAEzC,OAAO,EAAEC,QAAQ,EAAEL,OAAO,EAAEG,SAAS,EAAEsB,IAAI,EAAElB,SAAS,EAAED,aAAa,CAAC;IACzH;;IAEA;IACAF,OAAO,IAAI8B,QAAQ,CAACa,KAAK,EAAEC,YAAY,IAAI,CAAC;IAC5C3C,QAAQ,IAAI6B,QAAQ,CAACa,KAAK,EAAEE,aAAa,IAAI,CAAC;IAC9C3C,aAAa,IAAK4B,QAAQ,CAACa,KAAK,EAAUG,2BAA2B,IAAI,CAAC;IAC1E3C,SAAS,IAAK2B,QAAQ,CAACa,KAAK,EAAUI,uBAAuB,IAAI,CAAC;;IAElE;IACA,MAAMC,UAAoB,GAAG,EAAE;IAC/B,MAAMC,aAAkF,GAAG,EAAE;IAE7F,KAAK,MAAMC,KAAK,IAAIpB,QAAQ,CAACX,OAAO,EAAE;MACpC,IAAI+B,KAAK,CAAC9E,IAAI,KAAK,MAAM,EAAE;QACzB4E,UAAU,CAACG,IAAI,CAACD,KAAK,CAACE,IAAI,CAAC;MAC7B,CAAC,MAAM,IAAIF,KAAK,CAAC9E,IAAI,KAAK,UAAU,EAAE;QACpC6E,aAAa,CAACE,IAAI,CAAC;UAAEE,EAAE,EAAEH,KAAK,CAACG,EAAE;UAAEpF,IAAI,EAAEiF,KAAK,CAACjF,IAAI;UAAEqF,KAAK,EAAEJ,KAAK,CAACI;QAAiC,CAAC,CAAC;MACvG;IACF;IAEAhC,QAAQ,GAAG0B,UAAU,CAACO,IAAI,CAAC,IAAI,CAAC;;IAEhC;IACA,IAAIN,aAAa,CAACO,MAAM,KAAK,CAAC,EAAE;MAC9BnE,UAAU,GAAG;QAAEK,UAAU;QAAE0B,KAAK,EAAE,MAAM;QAAEsB,MAAM,EAAEpB;MAAS,CAAC,CAAC;MAC7D,OAAOE,UAAU,CAAC,IAAI,EAAEF,QAAQ,EAAEtB,OAAO,EAAEC,QAAQ,EAAEL,OAAO,EAAEG,SAAS,EAAEsB,IAAI,EAAElB,SAAS,EAAED,aAAa,CAAC;IAC1G;;IAEA;IACA,MAAMuD,YAAY,GAAG,MAAAA,CAAOxF,IAAY,EAAEqF,KAA8B,KAAK;MAC3EvD,SAAS,CAACoD,IAAI,CAAClF,IAAI,CAAC;MACpBoB,UAAU,GAAG;QAAEK,UAAU;QAAE0B,KAAK,EAAE,YAAY;QAAEsC,QAAQ,EAAEzF;MAAK,CAAC,CAAC;MACjE,IAAI;QACF,MAAM0F,MAAM,GAAG,MAAMvE,WAAW,CAACnB,IAAI,EAAEqF,KAAK,CAAC;QAC7CjE,UAAU,GAAG;UAAEK,UAAU;UAAE0B,KAAK,EAAE,aAAa;UAAEsC,QAAQ,EAAEzF,IAAI;UAAE2F,WAAW,EAAED,MAAM,CAACE;QAAQ,CAAC,CAAC;QAC/F,OAAO;UACLA,OAAO,EAAEF,MAAM,CAACE,OAAO;UACvBnB,MAAM,EAAEiB,MAAM,CAACE,OAAO,GAAGC,IAAI,CAACC,SAAS,CAACJ,MAAM,CAACK,IAAI,CAAC,GAAIL,MAAM,CAACM,KAAK,IAAI;QAC1E,CAAC;MACH,CAAC,CAAC,OAAOzB,GAAG,EAAE;QACZ,MAAMC,MAAM,GAAGjF,aAAa,CAACgF,GAAG,CAAC;QACjCnD,UAAU,GAAG;UAAEK,UAAU;UAAE0B,KAAK,EAAE,aAAa;UAAEsC,QAAQ,EAAEzF,IAAI;UAAE2F,WAAW,EAAE;QAAM,CAAC,CAAC;QACtF,OAAO;UAAEC,OAAO,EAAE,KAAK;UAAEnB,MAAM,EAAED;QAAO,CAAC;MAC3C;IACF,CAAC;IAED,MAAM;MAAEyB,OAAO,EAAEC,WAAW;MAAEC,OAAO;MAAEC;IAAY,CAAC,GAAG,MAAMtG,aAAa,CACxEkF,aAAa,EACbQ,YAAY,EACZ;MAAE3D,YAAY;MAAEwE,aAAa,EAAE3G,cAAc,CAAC4G,kBAAkB;MAAEC,eAAe,EAAE7G,cAAc,CAAC8G;IAAoB,CACxH,CAAC;IAED,IAAIL,OAAO,EAAE;MACX/E,UAAU,GAAG;QAAEK,UAAU;QAAE0B,KAAK,EAAE,OAAO;QAAEsB,MAAM,EAAE2B;MAAY,CAAC,CAAC;MACjE,OAAO7C,UAAU,CAAC,KAAK,EAAE6C,WAAW,IAAI,iBAAiB,EAAErE,OAAO,EAAEC,QAAQ,EAAEL,OAAO,EAAEG,SAAS,EAAEsB,IAAI,EAAElB,SAAS,EAAED,aAAa,CAAC;IACnI;;IAEA;IACAe,QAAQ,CAACkC,IAAI,CAAC;MACZjC,IAAI,EAAE,WAAW;MACjBC,OAAO,EAAEW,QAAQ,CAACX;IACpB,CAAC,CAAC;IACFF,QAAQ,CAACkC,IAAI,CAAC;MACZjC,IAAI,EAAE,MAAM;MACZC,OAAO,EAAEgD;IACX,CAAC,CAAC;EACJ;;EAEA;EACA9E,UAAU,GAAG;IAAEK,UAAU;IAAE0B,KAAK,EAAE,MAAM;IAAEsB,MAAM,EAAEpB;EAAS,CAAC,CAAC;EAC7D,OAAOE,UAAU,CAAC,IAAI,EAAEF,QAAQ,IAAI,4CAA4C,EAAEtB,OAAO,EAAEC,QAAQ,EAAEL,OAAO,EAAEG,SAAS,EAAEsB,IAAI,EAAElB,SAAS,EAAED,aAAa,CAAC;AAC1J;;AAEA;AACA;AACA;;AAEA,SAASsB,UAAUA,CACjBqC,OAAgB,EAChBnB,MAAc,EACdgC,WAAmB,EACnBC,YAAoB,EACpB/E,OAAe,EACfG,SAAmB,EACnB6E,SAAiB,EACjBC,eAAe,GAAG,CAAC,EACnBC,mBAAmB,GAAG,CAAC,EACP;EAChB,OAAO;IACLjB,OAAO;IACPnB,MAAM;IACNqC,UAAU,EAAE;MAAEzB,KAAK,EAAEoB,WAAW;MAAEhC,MAAM,EAAEiC,YAAY;MAAExE,SAAS,EAAE0E,eAAe;MAAE3E,aAAa,EAAE4E;IAAoB,CAAC;IACxHE,OAAO,EAAEzH,eAAe,CAACmH,WAAW,EAAEC,YAAY,EAAE/E,OAAO,EAAE,CAAC,EAAEiF,eAAe,EAAEC,mBAAmB,CAAC;IACrG/E,SAAS,EAAE,CAAC,GAAG,IAAIkF,GAAG,CAAClF,SAAS,CAAC,CAAC;IAClC6E,SAAS;IACTM,UAAU,EAAE,UAAU,CAAE;EAC1B,CAAC;AACH;AAEA,eAAenD,iBAAiBA,CAAIoD,EAAoB,EAAEC,UAAU,GAAG,CAAC,EAAc;EACpF,IAAIC,SAAkB;EACtB,KAAK,IAAIC,OAAO,GAAG,CAAC,EAAEA,OAAO,IAAIF,UAAU,EAAEE,OAAO,EAAE,EAAE;IACtD,IAAI;MACF,OAAO,MAAMH,EAAE,CAAC,CAAC;IACnB,CAAC,CAAC,OAAO3C,GAAG,EAAE;MACZ6C,SAAS,GAAG7C,GAAG;MACf,IAAI8C,OAAO,GAAGF,UAAU,IAAI3H,gBAAgB,CAAC+E,GAAG,CAAC,EAAE;QACjD,MAAM,IAAI+C,OAAO,CAAEC,OAAO,IAAKC,UAAU,CAACD,OAAO,EAAE,IAAI,GAAGhF,IAAI,CAACkF,GAAG,CAAC,CAAC,EAAEJ,OAAO,CAAC,CAAC,CAAC;QAChF;MACF;MACA,MAAM9C,GAAG;IACX;EACF;EACA,MAAM6C,SAAS;AACjB","ignoreList":[]}