feat: add plan agent & model agent (#5577)

* refactor: agent call (#5572) * feat: Add relevant functions related to agent invocation, including plan parsing, state management and tool invocation * Refactor agent call logic and utilities - Simplified the `runAgentCall` function by removing unnecessary complexity and restructuring the flow for better readability and maintainability. - Introduced helper functions to create tools from tool nodes and to prepare agent messages, enhancing modularity. - Removed the `utils.ts` file as its functions were integrated into the main logic, streamlining the codebase. - Updated the dispatch logic in `index.ts` to utilize the new helper functions and improve clarity. - Adjusted the handling of interactive modes and tool calls to ensure proper response formatting and error handling. * refactor: clean up the processing logic of the interactive mode and remove the unused tool creation functions * feat: add relevant constants for proxy configuration and update the proxy call logic * refactor: remove unused configuration variables from workflow properties * refactor: remove unused configuration variables from dispatchRunAgents props * fix: build error * refactor: update FlowNodeTypeEnum values and consolidate utility functions * refactor: simplify conditional checks in tool call and reasoning handlers * feat: add default agent prompt for improved response handling * refactor: rename directory with agent->tool, agentCall->agnet * refactor: rename dispatchRunAgents to dispatchRunAgent for consistency * refactor: rename toolCall to tools for consistency in FlowNodeTypeEnum * refactor: rename agents to toolCall for consistency in nodeTypes mapping * refactor: remove unused runtimeEdges parameter from dispatchRunAgent * refactor: update runAgentCall and dispatchRunAgent to use structured requestProps and workflowProps * refactor: streamline requestProps and handleToolResponse in runAgentCall and dispatchRunAgent * refactor: restructure RunAgentCallProps and update requestProps to requestParams for clarity * refactor: enhance interactiveEntryToolParams handling in runAgentCall for improved response management * refactor: flatten RunAgentCallProps structure and update dispatchRunAgent to use direct properties * fix: correct initialization of interactiveResponse in runAgentCall * agent call code * fix: agent call stop sign * feat: add plan agent tools and default generated prompts * feat: add model agent tools and related functions * chore: rename enum value * fix: optimize isEndSign assignment and update default plan prompt format * fix: update transferPlanAgent to use histories instead of sharedContext and rename default prompt variable * fix: update transferPlanAgent to use ChatItemType and adapt message structure * feat: add ModelAgentTool and PlanAgentTool with detailed descriptions and parameters fix: update error handling in transferModelAgent and transferPlanAgent to return error messages refactor: simplify isEndSign assignment in runAgentCall * feat: enhance agent call handling and response processing with context support * feat: refactor agent prompts and add utility functions for system prompt parsing * feat: add plan agent tools and default generated prompts * feat: add model agent tools and related functions * chore: rename enum value * fix: optimize isEndSign assignment and update default plan prompt format * fix: update transferPlanAgent to use histories instead of sharedContext and rename default prompt variable * fix: update transferPlanAgent to use ChatItemType and adapt message structure * feat: add ModelAgentTool and PlanAgentTool with detailed descriptions and parameters fix: update error handling in transferModelAgent and transferPlanAgent to return error messages refactor: simplify isEndSign assignment in runAgentCall * feat: enhance agent call handling and response processing with context support * feat: refactor agent prompts and add utility functions for system prompt parsing * feat: add AskAgentTool to support the interactive questioning function * Update request.ts --------- Co-authored-by: archer <545436317@qq.com>
2026-01-14 06:03:34 +08:00 · 2025-09-08 10:55:18 +08:00
parent 868db4ef1a
commit 32ea07cb9c
12 changed files with 700 additions and 67 deletions
--- a/packages/service/core/ai/llm/agentCall.ts
+++ b/packages/service/core/ai/llm/agentCall.ts
@@ -32,7 +32,10 @@ type RunAgentCallProps = {
    name: string;
    avatar: string;
  };
-  handleToolResponse: (e: ChatCompletionMessageToolCall) => Promise<{
+  handleToolResponse: (e: {
+    call: ChatCompletionMessageToolCall;
+    context: ChatCompletionMessageParam[];
+  }) => Promise<{
    response: string;
    usages: ChatNodeUsageType[];
    isEnd: boolean;
@@ -124,7 +127,10 @@ export const runAgentCall = async ({
    let isEndSign = false;
    for await (const tool of toolCalls) {
      // TODO: 加入交互节点处理
-      const { response, usages, isEnd } = await handleToolResponse(tool);
+      const { response, usages, isEnd } = await handleToolResponse({
+        call: tool,
+        context: requestMessages
+      });

      if (isEnd) {
        isEndSign = true;
--- a/packages/service/core/workflow/dispatch/ai/agent/constants.ts
+++ b/packages/service/core/workflow/dispatch/ai/agent/constants.ts
@@ -1,50 +1,30 @@
 import type { ChatCompletionTool } from '@fastgpt/global/core/ai/type';

-export const getTopAgentDefaultPrompt = () => {
-  return `你是一位Supervisor Agent，具备以下核心能力：
+export enum SubAppIds {
+  plan = 'plan_agent',
+  stop = 'stop_agent',
+  model = 'model_agent',
+  fileRead = 'file_read'
+}

-## 核心能力
-1. **计划制定与管理**：根据用户需求制定详细的执行计划，并实时跟踪和调整计划进度
-2. **工具调用编排**：可以调用各种工具来完成特定任务，支持并行和串行工具调用
-3. **上下文理解**：能够理解对话历史、文档内容和当前状态
-4. **自主决策**：根据当前情况和计划进度做出最优决策
-
-## 工作流程
-1. **需求分析**：深入理解用户需求，识别关键目标和约束条件
-2. **计划制定**：使用 plan_agent 工具制定详细的执行计划
-3. **工具编排**：根据计划选择和调用合适的工具
-4. **结果处理**：分析工具返回结果，判断是否满足预期
-5. **计划调整**：根据执行结果动态调整计划
-6. **最终输出**：给出完整、准确的回答
-
-## 特殊指令
+export const getTopAgentConstantPrompt = () => {
+  return `## 特殊指令
 - 对于复杂任务，必须先使用 plan_agent 制定计划
 - 在执行过程中如需调整计划，再次调用 plan_agent
 - 始终保持计划的可见性和可追踪性
- 遇到错误时要有容错和重试机制
-
-请始终保持专业、准确、有条理的回答风格，确保用户能够清楚了解执行进度和结果。`;
+- 每次有新的进度完成时，都要调用 plan_agent 更新计划
+- 遇到错误时要有容错和重试机制`;
 };

-export const PlanAgentTool: ChatCompletionTool = {
-  type: 'function',
-  function: {
-    name: 'plan_agent',
-    description:
-      '如果用户的任务非常复杂，可以先使用 plan_agent 制定计划，然后根据计划使用其他工具来完成任务。'
-  }
-};
-
-export const StopAgentId = 'stop_agent';
 export const StopAgentTool: ChatCompletionTool = {
  type: 'function',
  function: {
-    name: StopAgentId,
-    description: '如果完成了所有的任务，可调用次工具。'
+    name: SubAppIds.stop,
+    description: '如果完成了所有的任务，可调用此工具。'
  }
 };

-/* 
+/*
  结构：
  [url1,url2,url2]
  [
@@ -52,7 +32,7 @@ export const StopAgentTool: ChatCompletionTool = {
    {id:2,url: url2}
  ]
 */
-export const getFileReadTool = (urls: string[]): ChatCompletionTool => {
+export const getFileReadTool = (urls?: string[]): ChatCompletionTool => {
  return {
    type: 'function',
    function: {
@@ -64,7 +44,7 @@ export const getFileReadTool = (urls: string[]): ChatCompletionTool => {
          file_path: {
            type: 'string',
            description: '文件ID',
-            enum: urls.map((url, index) => `${index + 1}`)
+            enum: urls?.map((_, index) => `${index + 1}`)
          }
        },
        required: ['file_path']
--- a/packages/service/core/workflow/dispatch/ai/agent/index.ts
+++ b/packages/service/core/workflow/dispatch/ai/agent/index.ts
@@ -27,18 +27,25 @@ import {
 import { formatModelChars2Points } from '../../../../../support/wallet/usage/utils';
 import { getHistoryPreview } from '@fastgpt/global/core/chat/utils';
 import {
+  applyDiff,
  filterToolResponseToPreview,
  formatToolResponse,
  getToolNodesByIds,
-  initToolNodes
+  initToolNodes,
+  parseToolArgs
 } from '../utils';
-import { getTopAgentDefaultPrompt, StopAgentId, StopAgentTool } from './constants';
+import { getFileReadTool, getTopAgentConstantPrompt, StopAgentTool, SubAppIds } from './constants';
 import { runWorkflow } from '../..';
-import json5 from 'json5';
 import type { ChatCompletionTool } from '@fastgpt/global/core/ai/type';
 import type { ToolNodeItemType } from './type';
 import { textAdaptGptResponse } from '@fastgpt/global/core/workflow/runtime/utils';
 import { sliceStrStartEnd } from '@fastgpt/global/common/string/tools';
+import { transferPlanAgent } from '../sub/plan';
+import { transferModelAgent } from '../sub/model';
+import { PlanAgentTool } from '../sub/plan/constants';
+import { ModelAgentTool } from '../sub/model/constants';
+import { getSubIdsByAgentSystem, parseAgentSystem } from './utils';
+import { getChildAppPreviewNode } from '../../../../../core/app/plugin/controller';

 export type DispatchAgentModuleProps = ModuleDispatchProps<{
  [NodeInputKeyEnum.history]?: ChatItemType[];
@@ -82,7 +89,8 @@ export const dispatchRunAgent = async (props: DispatchAgentModuleProps): Promise
      history = 6,
      fileUrlList: fileLinks,
      temperature,
-      aiChatTopP
+      aiChatTopP,
+      planConfig
    }
  } = props;

@@ -96,7 +104,8 @@ export const dispatchRunAgent = async (props: DispatchAgentModuleProps): Promise
    }

    // Init tool params
-    const toolNodeIds = filterToolNodeIdByEdges({ nodeId, edges: runtimeEdges });
+    // const toolNodeIds = filterToolNodeIdByEdges({ nodeId, edges: runtimeEdges });
+    const toolNodeIds = getSubIdsByAgentSystem(systemPrompt);
    const toolNodes = getToolNodesByIds({ toolNodeIds, runtimeNodes });
    // TODO: 补充系统 agent
    const toolNodesMap = new Map<string, ToolNodeItemType>();
@@ -111,12 +120,14 @@ export const dispatchRunAgent = async (props: DispatchAgentModuleProps): Promise
      };
    };

-    const subApps = getSubApps({ toolNodes });
+    const subApps = getSubApps({ toolNodes, urls: fileLinks });
+
+    const combinedSystemPrompt = `${parseAgentSystem({ systemPrompt, toolNodesMap })}\n\n${getTopAgentConstantPrompt()}`;

    // TODO: 把 files 加入 query 中。
    const messages: ChatItemType[] = (() => {
      const value: ChatItemType[] = [
-        ...getSystemPrompt_ChatItemType(systemPrompt || getTopAgentDefaultPrompt()),
+        ...getSystemPrompt_ChatItemType(combinedSystemPrompt),
        // Add file input prompt to histories
        ...chatHistories,
        {
@@ -160,15 +171,107 @@ export const dispatchRunAgent = async (props: DispatchAgentModuleProps): Promise
        isAborted: res ? () => res.closed : undefined,

        getToolInfo,
-        handleToolResponse: async (call) => {
+        handleToolResponse: async ({ call, context }) => {
          const toolId = call.function.name;

-          if (toolId === StopAgentId) {
+          if (toolId === SubAppIds.stop) {
            return {
              response: '',
              usages: [],
              isEnd: true
            };
+          } else if (toolId === SubAppIds.plan) {
+            const planModel = planConfig?.model ?? model;
+            const { instruction } = parseToolArgs<{ instruction: string }>(call.function.arguments);
+
+            const { content, inputTokens, outputTokens } = await transferPlanAgent({
+              model: planModel,
+              instruction,
+              histories: GPTMessages2Chats({
+                messages: context.slice(1, -1),
+                getToolInfo
+              }),
+              onStreaming({ text, fullText }) {
+                //TODO: 需要一个新的 plan sse event
+                if (!fullText) return;
+                workflowStreamResponse?.({
+                  event: SseResponseEventEnum.toolResponse,
+                  data: {
+                    tool: {
+                      id: call.id,
+                      toolName: '',
+                      toolAvatar: '',
+                      params: '',
+                      response: sliceStrStartEnd(fullText, 5000, 5000)
+                    }
+                  }
+                });
+              }
+            });
+
+            const lastPlanCallIndex = context
+              .slice(0, -1)
+              .findLastIndex(
+                (c) =>
+                  c.role === 'assistant' &&
+                  c.tool_calls?.some((tc) => tc.function?.name === SubAppIds.plan)
+              );
+            const originalContent =
+              lastPlanCallIndex !== -1 ? (context[lastPlanCallIndex + 1].content as string) : '';
+
+            const applyedContent = applyDiff({
+              original: originalContent,
+              patch: content
+            });
+
+            // workflowStreamResponse?.({
+            //   event: SseResponseEventEnum.toolResponse,
+            //   data: {
+            //     tool: {
+            //       id: call.id,
+            //       toolName: '',
+            //       toolAvatar: '',
+            //       params: '',
+            //       response: sliceStrStartEnd(applyedContent, 5000, 5000)
+            //     }
+            //   }
+            // });
+
+            return {
+              response: content,
+              usages: [],
+              isEnd: false
+            };
+          } else if (toolId === SubAppIds.model) {
+            const { systemPrompt, task } = parseToolArgs<{ systemPrompt: string; task: string }>(
+              call.function.arguments
+            );
+
+            const { content, inputTokens, outputTokens } = await transferModelAgent({
+              model,
+              systemPrompt,
+              task,
+              onStreaming({ text, fullText }) {
+                if (!fullText) return;
+                workflowStreamResponse?.({
+                  event: SseResponseEventEnum.toolResponse,
+                  data: {
+                    tool: {
+                      id: call.id,
+                      toolName: '',
+                      toolAvatar: '',
+                      params: '',
+                      response: sliceStrStartEnd(fullText, 5000, 5000)
+                    }
+                  }
+                });
+              }
+            });
+            return {
+              response: content,
+              usages: [],
+              isEnd: false
+            };
          }

          const node = toolNodesMap.get(toolId);
@@ -180,14 +283,7 @@ export const dispatchRunAgent = async (props: DispatchAgentModuleProps): Promise
            };
          }

-          const startParams = (() => {
-            try {
-              return json5.parse(call.function.arguments);
-            } catch {
-              return {};
-            }
-          })();
-
+          const startParams = parseToolArgs(call.function.arguments);
          initToolNodes(runtimeNodes, [node.nodeId], startParams);
          const { toolResponses, flowUsages, flowResponses } = await runWorkflow({
            ...props,
@@ -324,9 +420,20 @@ export const dispatchRunAgent = async (props: DispatchAgentModuleProps): Promise
  }
 };

-const getSubApps = ({ toolNodes }: { toolNodes: ToolNodeItemType[] }): ChatCompletionTool[] => {
+const getSubApps = ({
+  toolNodes,
+  urls
+}: {
+  toolNodes: ToolNodeItemType[];
+  urls?: string[];
+}): ChatCompletionTool[] => {
  // System Tools: Plan Agent, stop sign, model agent.
-  const systemTools: ChatCompletionTool[] = [];
+  const systemTools: ChatCompletionTool[] = [
+    PlanAgentTool,
+    StopAgentTool,
+    ModelAgentTool,
+    getFileReadTool(urls)
+  ];

  // Node Tools
  const nodeTools = toolNodes.map<ChatCompletionTool>((item: ToolNodeItemType) => {
@@ -335,7 +442,7 @@ const getSubApps = ({ toolNodes }: { toolNodes: ToolNodeItemType[] }): ChatCompl
        type: 'function',
        function: {
          name: item.nodeId,
-          description: item.intro || item.name,
+          description: `调用${item.flowNodeType}:${item.name || item.intro}完成任务`,
          parameters: item.jsonSchema
        }
      };
@@ -358,7 +465,7 @@ const getSubApps = ({ toolNodes }: { toolNodes: ToolNodeItemType[] }): ChatCompl
      type: 'function',
      function: {
        name: item.nodeId,
-        description: item.toolDescription || item.intro || item.name,
+        description: `调用${item.flowNodeType}:${item.name || item.toolDescription || item.intro}完成任务`,
        parameters: {
          type: 'object',
          properties,
@@ -367,6 +474,7 @@ const getSubApps = ({ toolNodes }: { toolNodes: ToolNodeItemType[] }): ChatCompl
      }
    };
  });
+  console.dir(nodeTools, { depth: null });

  return [...systemTools, ...nodeTools];
 };
--- a/packages/service/core/workflow/dispatch/ai/agent/utils.ts
+++ b/packages/service/core/workflow/dispatch/ai/agent/utils.ts
@@ -0,0 +1,61 @@
+import type { ToolNodeItemType } from './type';
+
+const namespaceMap = new Map<string, string>([
+  ['a', '子应用'],
+  ['t', '工具'],
+  ['d', '知识库'],
+  ['m', '模型']
+]);
+
+// e.g: {{@a.appId@}} -> a.appId
+const buildPattern = (options?: { prefix?: string }): RegExp => {
+  const config = {
+    prefix: '@',
+    ...options
+  };
+
+  const escapedPrefix = config.prefix.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  return new RegExp(`\\{\\{${escapedPrefix}([^${escapedPrefix}]+)${escapedPrefix}\\}\\}`, 'g');
+};
+
+export const getSubIdsByAgentSystem = (
+  systemPrompt: string,
+  options?: { prefix?: string }
+): string[] => {
+  const pattern = buildPattern(options);
+  const ids: string[] = [];
+  let match;
+
+  while ((match = pattern.exec(systemPrompt)) !== null) {
+    const fullName = match[1];
+    const [, id] = fullName.split('.');
+    if (id) {
+      ids.push(id);
+    }
+  }
+
+  return ids;
+};
+
+export const parseAgentSystem = ({
+  systemPrompt,
+  toolNodesMap,
+  options
+}: {
+  systemPrompt: string;
+  toolNodesMap: Map<string, ToolNodeItemType>;
+  options?: { prefix?: string };
+}): string => {
+  const pattern = buildPattern(options);
+
+  const processedPrompt = systemPrompt.replace(pattern, (_, toolName) => {
+    const [namespace, id] = toolName.split('.');
+    const toolNode = toolNodesMap.get(id);
+    const name = toolNode?.name || toolNode?.toolDescription || toolNode?.intro || 'unknown';
+
+    const prefix = namespaceMap.get(namespace) ?? 'unknown';
+    return `${prefix}:${name}`;
+  });
+
+  return processedPrompt;
+};
--- a/packages/service/core/workflow/dispatch/ai/sub/ask/constants.ts
+++ b/packages/service/core/workflow/dispatch/ai/sub/ask/constants.ts
@@ -0,0 +1,61 @@
+import type { ChatCompletionTool } from '@fastgpt/global/core/ai/type';
+import { SubAppIds } from '../../agent/constants';
+
+export const AskAgentTool: ChatCompletionTool = {
+  type: 'function',
+  function: {
+    name: SubAppIds.ask,
+    description: '调用此工具，向用户发起交互式提问',
+    parameters: {
+      type: 'object',
+      properties: {
+        mode: {
+          type: 'string',
+          enum: ['userSelect', 'formInput', 'userInput'],
+          description: '交互模式'
+        },
+        prompt: {
+          type: 'string',
+          description: '向用户展示的提示信息'
+        },
+        options: {
+          type: 'array',
+          description: '当 mode=userSelect 时可供选择的选项',
+          items: {
+            type: 'string'
+          }
+        },
+        form: {
+          type: 'array',
+          description: '当 mode=formInput 时需要填写的表单字段列表',
+          items: {
+            type: 'object',
+            properties: {
+              field: {
+                type: 'string',
+                description: '字段名，如 name, age, 同时会展示给用户一样的label'
+              },
+              type: {
+                type: 'string',
+                enum: ['textInput', 'numberInput', 'singleSelect', 'multiSelect'],
+                description: '字段输入类型'
+              },
+              required: { type: 'boolean', description: '该字段是否必填', default: false },
+              options: {
+                type: 'array',
+                description: '当 type 为 singleSelect 或 multiSelect 时的可选项',
+                items: { type: 'string' }
+              }
+            },
+            required: ['field', 'type']
+          }
+        },
+        userInput: {
+          type: 'string',
+          description: '当 mode=userInput 时用户自由输入的内容'
+        }
+      },
+      required: ['mode', 'prompt']
+    }
+  }
+};
--- a/packages/service/core/workflow/dispatch/ai/sub/model/constants.ts
+++ b/packages/service/core/workflow/dispatch/ai/sub/model/constants.ts
@@ -0,0 +1,24 @@
+import type { ChatCompletionTool } from '@fastgpt/global/core/ai/type';
+import { SubAppIds } from '../../agent/constants';
+
+export const ModelAgentTool: ChatCompletionTool = {
+  type: 'function',
+  function: {
+    name: SubAppIds.model,
+    description: '完成一些简单通用型任务, 可以调用此工具。',
+    parameters: {
+      type: 'object',
+      properties: {
+        systemPrompt: {
+          type: 'string',
+          description: '注入给此 agent 的系统提示词'
+        },
+        task: {
+          type: 'string',
+          description: '此 agent 本轮需要完成的任务'
+        }
+      },
+      required: ['systemPrompt', 'task']
+    }
+  }
+};
--- a/packages/service/core/workflow/dispatch/ai/sub/model/index.ts
+++ b/packages/service/core/workflow/dispatch/ai/sub/model/index.ts
@@ -0,0 +1,87 @@
+import type { ChatCompletionMessageParam } from '@fastgpt/global/core/ai/type.d';
+import { addLog } from '../../../../../../common/system/log';
+import { createLLMResponse, type ResponseEvents } from '../../../../../ai/llm/request';
+import type { ChatItemType } from '@fastgpt/global/core/chat/type';
+import { chats2GPTMessages, getSystemPrompt_ChatItemType } from '@fastgpt/global/core/chat/adapt';
+import { ChatItemValueTypeEnum, ChatRoleEnum } from '@fastgpt/global/core/chat/constants';
+import { getErrText } from '@fastgpt/global/common/error/utils';
+
+type ModelAgentConfig = {
+  model: string;
+  temperature?: number;
+  top_p?: number;
+  stream?: boolean;
+};
+
+type transferModelAgentProps = {
+  systemPrompt?: string;
+  task?: string;
+} & ModelAgentConfig &
+  Pick<ResponseEvents, 'onStreaming' | 'onReasoning'>;
+
+export async function transferModelAgent({
+  systemPrompt = '',
+  task = '',
+
+  onStreaming,
+  onReasoning,
+
+  model,
+  temperature = 0.7,
+  top_p,
+  stream = true
+}: transferModelAgentProps): Promise<{
+  content: string;
+  inputTokens: number;
+  outputTokens: number;
+}> {
+  try {
+    const messages: ChatItemType[] = [
+      ...getSystemPrompt_ChatItemType(systemPrompt),
+      {
+        obj: ChatRoleEnum.Human,
+        value: [
+          {
+            type: ChatItemValueTypeEnum.text,
+            text: {
+              content: task
+            }
+          }
+        ]
+      }
+    ];
+    const adaptedMessages: ChatCompletionMessageParam[] = chats2GPTMessages({
+      messages,
+      reserveId: false
+    });
+
+    const {
+      answerText,
+      usage: { inputTokens, outputTokens }
+    } = await createLLMResponse({
+      body: {
+        model,
+        temperature,
+        messages: adaptedMessages,
+        top_p,
+        stream
+      },
+      onStreaming,
+      onReasoning
+    });
+
+    return {
+      content: answerText,
+      inputTokens,
+      outputTokens
+    };
+  } catch (error) {
+    const err = getErrText(error);
+    addLog.warn('call model_agent failed');
+    return {
+      content: err,
+      inputTokens: 0,
+      outputTokens: 0
+    };
+  }
+}
--- a/packages/service/core/workflow/dispatch/ai/sub/plan/constants.ts
+++ b/packages/service/core/workflow/dispatch/ai/sub/plan/constants.ts
@@ -0,0 +1,22 @@
+import type { ChatCompletionTool } from '@fastgpt/global/core/ai/type';
+import { SubAppIds } from '../../agent/constants';
+
+export const PlanAgentTool: ChatCompletionTool = {
+  type: 'function',
+  function: {
+    name: SubAppIds.plan,
+    description:
+      '如果用户的任务非常复杂，可以先使用 plan_agent 制定计划，然后根据计划使用其他工具来完成任务。同时，plan_agent 负责维护整个任务的上下文和状态。可以更新或修改计划中的内容. 但是 plan_agent 不能直接执行任务。',
+    parameters: {
+      type: 'object',
+      properties: {
+        instruction: {
+          type: 'string',
+          description:
+            '给 plan_agent 的指令, 例如: "制定一个包含以下步骤的计划:xxx", "将 xxx 待办事项标记为已完成"'
+        }
+      },
+      required: ['instruction']
+    }
+  }
+};
--- a/packages/service/core/workflow/dispatch/ai/sub/plan/index.ts
+++ b/packages/service/core/workflow/dispatch/ai/sub/plan/index.ts
@@ -0,0 +1,96 @@
+import type { ChatCompletionMessageParam } from '@fastgpt/global/core/ai/type.d';
+import { addLog } from '../../../../../../common/system/log';
+import { createLLMResponse, type ResponseEvents } from '../../../../../ai/llm/request';
+import { defaultPlanAgentPrompt } from './prompt';
+import { replaceVariable } from '@fastgpt/global/common/string/tools';
+import { chats2GPTMessages, getSystemPrompt_ChatItemType } from '@fastgpt/global/core/chat/adapt';
+import type { ChatItemType } from '@fastgpt/global/core/chat/type';
+import { ChatItemValueTypeEnum, ChatRoleEnum } from '@fastgpt/global/core/chat/constants';
+import { getErrText } from '@fastgpt/global/common/error/utils';
+
+type PlanAgentConfig = {
+  model: string;
+  customSystemPrompt?: string;
+  temperature?: number;
+  top_p?: number;
+  stream?: boolean;
+};
+
+type transferPlanAgentProps = {
+  histories: ChatItemType[];
+  instruction?: string;
+} & PlanAgentConfig &
+  Pick<ResponseEvents, 'onStreaming' | 'onReasoning'>;
+
+export async function transferPlanAgent({
+  instruction = '',
+  histories,
+
+  onStreaming,
+  onReasoning,
+
+  model,
+  customSystemPrompt,
+  temperature = 0,
+  top_p,
+  stream = true
+}: transferPlanAgentProps): Promise<{
+  content: string;
+  inputTokens: number;
+  outputTokens: number;
+}> {
+  try {
+    const messages: ChatItemType[] = [
+      ...getSystemPrompt_ChatItemType(
+        replaceVariable(defaultPlanAgentPrompt, {
+          userRole: customSystemPrompt
+        })
+      ),
+      ...histories,
+      {
+        obj: ChatRoleEnum.Human,
+        value: [
+          {
+            type: ChatItemValueTypeEnum.text,
+            text: {
+              content: instruction
+            }
+          }
+        ]
+      }
+    ];
+    const adaptedMessages: ChatCompletionMessageParam[] = chats2GPTMessages({
+      messages,
+      reserveId: false
+    });
+
+    const {
+      answerText,
+      usage: { inputTokens, outputTokens }
+    } = await createLLMResponse({
+      body: {
+        model,
+        temperature,
+        messages: adaptedMessages,
+        top_p,
+        stream
+      },
+      onStreaming,
+      onReasoning
+    });
+
+    return {
+      content: answerText,
+      inputTokens,
+      outputTokens
+    };
+  } catch (error) {
+    const err = getErrText(error);
+    addLog.warn('call plan_agent failed');
+    return {
+      content: err,
+      inputTokens: 0,
+      outputTokens: 0
+    };
+  }
+}
--- a/packages/service/core/workflow/dispatch/ai/sub/plan/prompt.ts
+++ b/packages/service/core/workflow/dispatch/ai/sub/plan/prompt.ts
@@ -0,0 +1,145 @@
+export const defaultPlanAgentPrompt = `<role>
+你是一个专业的项目规划助手，擅长将复杂任务分解为结构化的执行计划；同时支持对既有计划进行“最小差异（+- Diff）”式修改。你会严格遵循指定的注释标记格式，并在修改模式下输出可直接应用的补丁（patch）。
+</role>
+
+<user_role>
+{{userRole}}
+</user_role>
+
+<modes>
+- 自动识别两种模式：
+  1) 创建模式（create）：用户未提供现有计划时，生成全新的计划文档。
+  2) 修改模式（patch）：用户提供现有计划或明确提出“增删改”需求时，输出 Diff Patch。
+- 若用户提供了现有计划文本（含注释标记），一律进入修改模式；否则进入创建模式。
+</modes>
+
+<task>
+根据用户提供的主题或目标，生成或修改一份详细、可执行的项目计划文档，包含合理的阶段划分与具体待办事项；修改时以“补丁”为最小输出单元，确保变更可定位、可回放、可审计。
+</task>
+
+<inputs>
+- 用户输入：一个需要制定或更新的主题、目标或任务描述；可选的现有计划文档；可选的变更请求（自然语言或指令式）。
+- 输入格式：自然语言描述，可能包含背景、目标、约束、优先级、本地化偏好、以及现有计划全文。
+</inputs>
+
+<process>
+通用步骤
+1. 解析用户输入，提取核心目标、关键要素、约束与本地化偏好。
+2. 评估任务复杂度（简单：2-3 步；复杂：4-7 步），据此确定阶段数量。
+3. 各阶段生成 3-5 条可执行 Todo，动词开头，MECE 且无重叠。
+4. 语言风格本地化（根据用户输入语言进行术语与语序调整）。
+
+创建模式（create）
+5. 产出完整计划，严格使用占位符 [主题] 与标记体系；确保编号连续、标签闭合、结构清晰。
+
+修改模式（patch）
+5. 解析“现有计划”（锚点优先级：\`<!--@step:N:start-->\`、\`<!--@todo:N.X-->\`、\`<!--@note:N-->\` 等）。
+6. 将用户变更需求映射为原子操作（见 <diff_ops>），生成最小必要的行级 Diff：
+   - 仅对变更涉及的行输出 \`+\`（新增）或 \`-\`（删除）；未变更行不重复输出。
+   - 修改视为“-旧行”与“+新行”的并列呈现。
+   - 插入请贴靠最稳固的锚点（如 \`<!--@step:N:start-->\` 下的标题或 \`<!--@todos:N:start-->\` 前后）。
+7. 自动重排：对步骤编号 N、待办编号 X 做连续性校正；若移动/插入造成编号漂移，补丁中体现校正后的行。
+8. 校验：所有必须标签完整闭合；编号连续；每步 3-5 条待办；MECE；无空段落；无悬挂标记。
+9. 产出补丁；如 render=full，则在补丁后附上“更新后的完整文档”。
+
+<diff_ops>
+支持的原子操作（内部推理用，输出仍为行级 Diff）：
+- ADD_STEP(after N | before N | at end)    新增步骤（含标题、描述、Todos、可选备注）
+- REMOVE_STEP(N)                           删除步骤
+- UPDATE_STEP(N, title/desc/note=…)        更新步骤标题/描述/备注
+- MOVE_STEP(N -> M)                        移动步骤至序号 M（重排后编号连续）
+- ADD_TODO(N, at k)                        在步骤 N 的第 k 个位置插入 Todo
+- REMOVE_TODO(N.k)                         删除 Todo
+- UPDATE_TODO(N.k, text=…)                 更新 Todo 文本
+- MOVE_TODO(N.k -> M.t)                    移动 Todo 到其他步骤/位置
+- RENAME_THEME(text=…)                     更新主标题或整体描述中的 [主题] 文本描述（占位仍用 [主题]）
+说明：若用户未显式给出操作，你需从自然语言中归纳为上述操作的序列并生成对应行级 Diff。
+</diff_ops>
+
+<diff_rules>
+- 补丁仅包含变更行，以“前缀字符 +|-”表示新增/删除。
+- 不在上述集合内的行（例如空行）若因结构需要调整，可一并纳入补丁。
+- 修改必须保持“标记尾注不变更其语义角色”，即：当你替换内容时，保留原注释标签并只更改标签左侧的可读文本。
+- 重排后编号以补丁中的最新数字为准；不要在同一补丁里出现对同一元素的多次相互抵消的改动。
+- 若原文缺少稳固锚点，优先在最近的 \`<!--@step:*:start/end-->\` 或 \`<!--@todos:*:start/end-->\` 相邻位置插入。
+</diff_rules>
+
+<requirements>
+- 必须严格遵循以下注释标记格式：
+  * <!--@title--> 标记主标题
+  * <!--@desc--> 标记整体描述
+  * <!--@step:N:start--> 和 <!--@step:N:end--> 包裹步骤块
+  * <!--@step:N:title--> 标记步骤标题
+  * <!--@step:N:desc--> 标记步骤描述
+  * <!--@todos:N:start--> 和 <!--@todos:N:end--> 包裹待办列表
+  * <!--@todo:N.X--> 标记单个待办事项
+  * <!--@note:N--> 添加重要注释或备注
+- 步骤数量随复杂度自动调整；每步 3-5 条 Todo。
+- 编号（N、X）必须连续、准确，修改模式下需自动校正。
+- 描述语言简洁、专业、可操作；各阶段逻辑递进、MECE。
+- 进行本地化调整（术语、量词、表达习惯）。
+</requirements>
+
+<guardrails>
+- 不生成违法、不道德或有害内容；敏感主题输出合规替代方案。
+- 避免过于具体的时间/预算承诺与无法验证的保证。
+- 保持中立、客观；必要时指出风险与依赖。
+- 你拥有的记忆是通过别的 Agent 共享给你的, 你只需要专注于输出内容, 不必担心上下文的完整性。
+</guardrails>
+
+<output>
+  <format_create>
+  # [主题] 深度调研计划 <!--@title-->
+
+  全面了解 [主题] 的 [核心维度描述] <!--@desc-->
+
+  <!--@step:1:start-->
+  ## Step 1: [阶段名称] <!--@step:1:title-->
+  [阶段目标描述] <!--@step:1:desc-->
+  ### Todo List
+  <!--@todos:1:start-->
+  - [ ] [具体任务描述] <!--@todo:1.1-->
+  - [ ] [具体任务描述] <!--@todo:1.2-->
+  - [ ] [具体任务描述] <!--@todo:1.3-->
+  <!--@todos:1:end-->
+  <!--@note:1--> [可选备注]
+  <!--@step:1:end-->
+
+  <!--@step:2:start-->
+  ## Step 2: [阶段名称] <!--@step:2:title-->
+  [阶段目标描述] <!--@step:2:desc-->
+  ### Todo List
+  <!--@todos:2:start-->
+  - [ ] [具体任务描述] <!--@todo:2.1-->
+  - [ ] [具体任务描述] <!--@todo:2.2-->
+  - [ ] [具体任务描述] <!--@todo:2.3-->
+  <!--@todos:2:end-->
+  <!--@note:2--> [可选备注]
+  <!--@step:2:end-->
+  </format_create>
+
+  <format_patch>
+  # 仅输出变更行，以 +- 表示；无代码块围栏；保持原有缩进与空行风格。确保旧行的准确性和完整性
+  # 如果是 todo list 的变更，请确保 todo 前的 - [ ] 符号正确。删除或替换的行前用 - - [ ] 来表示 todo 的变更
+  # 禁止输出代码块标记\`\`\`
+  + 新增或替换后的行
+  - 被删除或被替换的旧行
+  </format_patch>
+
+  <style>
+  - 标题简洁有力，突出核心主题
+  - 描述准确概括该阶段的核心目标
+  - 待办事项以动词开头，明确可执行
+  - 保持专业术语的准确性
+  - 语言流畅、逻辑清晰
+  </style>
+</output>
+
+<examples>
+- 例：标记 todo 2.1 完成状态
+<patch>
+- - [ ] 完成数据采集与清洗 <!--@todo:2.1-->
+ - [x] 完成数据采集与清洗 <!--@todo:2.1-->
+</patch>
+</examples>
+`;
--- a/packages/service/core/workflow/dispatch/ai/tool/toolCall.ts
+++ b/packages/service/core/workflow/dispatch/ai/tool/toolCall.ts
@@ -13,7 +13,7 @@ import json5 from 'json5';
 import type { DispatchFlowResponse } from '../../type';
 import { GPTMessages2Chats } from '@fastgpt/global/core/chat/adapt';
 import type { AIChatItemType } from '@fastgpt/global/core/chat/type';
-import { formatToolResponse, initToolCallEdges, initToolNodes } from '../utils';
+import { formatToolResponse, initToolCallEdges, initToolNodes, parseToolArgs } from '../utils';
 import { computedMaxToken } from '../../../../ai/utils';
 import { sliceStrStartEnd } from '@fastgpt/global/common/string/tools';
 import type { WorkflowInteractiveResponseType } from '@fastgpt/global/core/workflow/template/system/interactive/type';
@@ -368,13 +368,7 @@ export const runToolCall = async (

      if (!toolNode) continue;

-      const startParams = (() => {
-        try {
-          return json5.parse(tool.function.arguments);
-        } catch (error) {
-          return {};
-        }
-      })();
+      const startParams = parseToolArgs(tool.function.arguments);

      initToolNodes(runtimeNodes, [toolNode.nodeId], startParams);
      const toolRunResponse = await runWorkflow({
@@ -384,7 +378,6 @@ export const runToolCall = async (
      });

      const stringToolResponse = formatToolResponse(toolRunResponse.toolResponses);
-
      const toolMsgParams: ChatCompletionToolMessageParam = {
        tool_call_id: tool.id,
        role: ChatCompletionRequestMessageRoleEnum.Tool,
--- a/packages/service/core/workflow/dispatch/ai/utils.ts
+++ b/packages/service/core/workflow/dispatch/ai/utils.ts
@@ -11,6 +11,7 @@ import { NodeInputKeyEnum } from '@fastgpt/global/core/workflow/constants';
 import type { McpToolDataType } from '@fastgpt/global/core/app/mcpTools/type';
 import type { JSONSchemaInputType } from '@fastgpt/global/core/app/jsonschema';
 import type { ToolNodeItemType } from './tool/type';
+import json5 from 'json5';

 export const filterToolResponseToPreview = (response: AIChatItemValueItemType[]) => {
  return response.map((item) => {
@@ -167,3 +168,52 @@ export const getToolNodesByIds = ({
      };
    });
 };
+
+export const parseToolArgs = <T = Record<string, any>>(toolArgs: string): T => {
+  try {
+    return json5.parse(toolArgs) as T;
+  } catch {
+    return {} as T;
+  }
+};
+
+/**
+ * 简单版 diff apply
+ * @param original 原始文本
+ * @param patch diff patch 文本（带 + 和 -）
+ */
+export const applyDiff = ({ original, patch }: { original: string; patch: string }): string => {
+  if (!original) return patch;
+
+  let result = original.split('\n');
+
+  const patchLines = patch.split('\n');
+  for (let i = 0; i < patchLines.length; i++) {
+    const line = patchLines[i];
+
+    if (line.startsWith('-')) {
+      const oldContent = line.slice(1).trim();
+      const next = patchLines[i + 1];
+
+      // 下一个是对应的 + 行 → 替换
+      if (next && next.startsWith('+')) {
+        const newContent = next.slice(1).trim(); // 也要 trim
+        const idx = result.findIndex((l) => l.trim() === oldContent);
+        if (idx !== -1) {
+          // 保留原有的缩进
+          const indent = result[idx].match(/^(\s*)/)?.[1] || '';
+          result[idx] = indent + newContent; // 保留缩进
+        }
+        i++; // 跳过下一个 + 行
+      } else {
+        // 单独的删除行
+        const idx = result.findIndex((l) => l.trim() === oldContent);
+        if (idx !== -1) {
+          result.splice(idx, 1);
+        }
+      }
+    }
+  }
+
+  return result.join('\n');
+};