fix(maker-core): surface Claude subagent launch failures to the parent task

apps/desktop/src/renderer/components/chat/AgentTaskCard.tsx

@@ -14,6 +14,7 @@ import {
 import { useTranslation } from 'react-i18next';
 import {
   deriveAgentTaskStatus,
+  isSubagentResultError,
   type AgentTaskTerminalStatus,
 } from '@cindy/maker-shared/agent-task';
 
@@ -235,6 +236,7 @@ export function AgentTaskCard({
         resultIsLaunchReceipt:
           subagentSpawnReceiptName(toolCall?.toolName, toolCall?.toolInput, result) !== undefined
           || subagentSpawnResultIndicatesRunning(toolCall?.toolName, result),
+        resultIsError: isSubagentResultError(result),
       });
   const StatusIcon = statusIcon(status);
   const statusIconClassName = cn(

apps/desktop/src/renderer/features/right-sidebar/plugins/background-tasks/listSessionTasks.ts

@@ -19,6 +19,7 @@
 import {
   deriveAgentTaskStatus,
   isAgentTaskToolName,
+  isSubagentResultError,
   subagentSpawnReceiptName,
   subagentSpawnResultIndicatesRunning,
 } from '@cindy/maker-shared/agent-task';
@@ -341,6 +342,7 @@ export function listSessionTasks(input: {
             resultIsLaunchReceipt:
               subagentSpawnReceiptName(toolName, toolInput, resultText) !== undefined
               || subagentSpawnResultIndicatesRunning(toolName, resultText),
+            resultIsError: isSubagentResultError(resultText),
           })
         : (settled ? 'completed' : isSessionStreaming ? 'running' : 'stopped');
     const provider: SessionTaskItem['provider'] =

packages/maker-shared/src/__tests__/agentTask.test.ts

@@ -5,6 +5,7 @@ import {
   deriveAgentTaskStatus,
   findAgentTaskUpdate,
   isAgentTaskToolName,
+  isSubagentResultError,
   isSubagentSpawnToolName,
   mergeAgentTaskUpdate,
   PI_SUBAGENT_TOOL_NAME,
@@ -113,6 +114,31 @@ describe('deriveAgentTaskStatus', () => {
       persistedStatus: 'cancelled' as never,
     })).toBe('completed');
   });
+
+  it('returns failed when resultIsError is true and result is non-empty', () => {
+    expect(deriveAgentTaskStatus(undefined, '<tool_use_error>Auth failed</tool_use_error>', {
+      resultIsError: true,
+    })).toBe('failed');
+  });
+
+  it('returns failed even when updateStatus is running when resultIsError is true', () => {
+    expect(deriveAgentTaskStatus('running', '<tool_use_error>timeout</tool_use_error>', {
+      resultIsError: true,
+    })).toBe('failed');
+  });
+
+  it('persisted status wins over resultIsError', () => {
+    expect(deriveAgentTaskStatus(undefined, '<tool_use_error>fail</tool_use_error>', {
+      resultIsError: true,
+      persistedStatus: 'completed',
+    })).toBe('completed');
+  });
+
+  it('resultIsError is ignored when result is empty', () => {
+    expect(deriveAgentTaskStatus(undefined, '', {
+      resultIsError: true,
+    })).toBe('running');
+  });
 });
 
 describe('mergeAgentTaskUpdate', () => {
@@ -525,3 +551,39 @@ describe('buildAgentTaskCardModel', () => {
     expect(model.summary).toBe('thread-2: done');
   });
 });
+
+describe('isSubagentResultError', () => {
+  it('returns false for empty/undefined/null', () => {
+    expect(isSubagentResultError(undefined)).toBe(false);
+    expect(isSubagentResultError('')).toBe(false);
+    expect(isSubagentResultError('   ')).toBe(false);
+  });
+
+  it('returns true for Claude protocol <tool_use_error>', () => {
+    expect(isSubagentResultError('<tool_use_error>Authentication failed</tool_use_error>')).toBe(true);
+  });
+
+  it('returns false for generic <error> prefix (too generic for work product)', () => {
+    // A subagent returning <error>校验报告</error> as valid output must not be
+    // misclassified as failure. Only <tool_use_error> is a reliable protocol marker.
+    expect(isSubagentResultError('<error>Transport failure</error>')).toBe(false);
+    expect(isSubagentResultError('<error>校验报告</error>')).toBe(false);
+  });
+
+  it('returns false for JSON with error-looking fields (authority boundary)', () => {
+    // Subagent result content is arbitrary user work product.
+    // Fields like "errors", "status", "stderr" are data, not execution signals.
+    expect(isSubagentResultError('{"errors":["not found"]}')).toBe(false);
+    expect(isSubagentResultError('{"status":"failed"}')).toBe(false);
+    expect(isSubagentResultError('{"stderr":"something went wrong"}')).toBe(false);
+    expect(isSubagentResultError('{"error":"custom message"}')).toBe(false);
+    expect(isSubagentResultError('{"success":false}')).toBe(false);
+    expect(isSubagentResultError('{"ok":false}')).toBe(false);
+  });
+
+  it('returns false for natural language error phrases', () => {
+    expect(isSubagentResultError('failed to launch')).toBe(false);
+    expect(isSubagentResultError('unable to start the service')).toBe(false);
+    expect(isSubagentResultError('Error: something happened')).toBe(false);
+  });
+});

packages/maker-shared/src/agentTask.ts

@@ -174,15 +174,50 @@ export function deriveAgentTaskStatus(
   options?: {
     resultIsLaunchReceipt?: boolean;
     persistedStatus?: AgentTaskTerminalStatus;
+    resultIsError?: boolean;
   },
 ): AgentTaskStatus {
   const persistedStatus = normalizeAgentTaskTerminalStatus(options?.persistedStatus);
   if (persistedStatus) return persistedStatus;
   const hasResult = typeof result === 'string' && result.trim().length > 0;
+  if (options?.resultIsError && hasResult) return 'failed';
   if (updateStatus === 'running' && hasResult && !options?.resultIsLaunchReceipt) return 'completed';
   return updateStatus ?? (hasResult ? 'completed' : 'running');
 }
 
+/**
+ * 判断子任务工具结果是否以协议级错误收尾(历史回放恢复 failed 的依据)。
+ *
+ * 仅识别 Claude SDK 协议标记 `<tool_use_error>` — 这是 SDK 在 tool call 失败时
+ * 发出的结构化错误格式。不解析任意 JSON 字段或自然语言错误短语,因为子任务结果
+ * 内容是用户工作产物,其中 "errors"/"status"/"stderr" 等字段是数据而非执行信号。
+ *
+ * 调用方约束:此函数仅应在已确认为子任务上下文的调用点使用
+ * (AgentTaskCard / listSessionTasks)。普通工具结果包含 `<tool_use_error>` 时
+ * 不应传入此函数,否则会将非子任务结果误判为失败。
+ *
+ * Authority: Claude protocol `<tool_use_error>` — SDK 在 tool call 失败时发出。
+ */
+export function isSubagentResultError(result: string | undefined): boolean {
+  const text = typeof result === 'string' ? result.trim() : '';
+  if (text.length === 0) return false;
+  // Only trust protocol-level error markers. Subagent result content is arbitrary
+  // user work product -- fields like "errors", "status", "stderr" in JSON output
+  // are data, not execution failure signals. Parsing arbitrary body for
+  // error-looking fields creates false positives that mark successful tasks as
+  // failed after Desktop reload / Mobile reconnect.
+  //
+  // Authority sources:
+  // 1. Claude protocol <tool_use_error> -- emitted by the SDK when a tool call fails
+  // 2. Persisted structured terminal status (agentTaskStatus) -- written by
+  //    messagePersistBroadcaster on terminal observations
+  //
+  // Note: <error> prefix removed -- too generic. A subagent returning
+  // `<error>校验报告</error>` as work output would be misclassified as failure.
+  // Only <tool_use_error> is a reliable protocol-owned error marker.
+  return text.startsWith('<tool_use_error>');
+}
+
 /**
  * Tool names that spawn a sub-agent task: Claude `Task`/`Agent`, Codex collab agents,
  * PI `subagent`(Cindy 自有扩展注册的工具名,与 pi 社区惯例一致)。