Add /background commands (when background agent is configured) (#4407)

Co-authored-by: Bryan Morgan <bryanmorgan@google.com>

packages/core/src/background/backgroundAgent.ts

@@ -0,0 +1,126 @@
+/**
+ * @license
+ * Copyright 2025 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { MCPServerConfig } from '../config/config.js';
+import { connectToMcpServer, discoverTools } from '../tools/mcp-client.js';
+import { DiscoveredMCPTool } from '../tools/mcp-tool.js';
+import {
+  BackgroundAgentTasksResponseSchema,
+  BackgroundAgentTaskResponseSchema,
+  BackgroundAgentTask,
+} from './types.js';
+
+export async function loadBackgroundAgent(
+  name: string,
+  config: MCPServerConfig,
+  debugMode: boolean,
+): Promise<BackgroundAgent> {
+  const server = await connectToMcpServer(name, config, debugMode);
+  try {
+    const tools = await discoverTools(name, config, server);
+    return new BackgroundAgent(name, tools);
+  } catch (error) {
+    await server.close();
+    throw error;
+  }
+}
+
+export class BackgroundAgent {
+  readonly startTaskTool: DiscoveredMCPTool;
+  readonly getTaskTool: DiscoveredMCPTool;
+  readonly listTasksTool: DiscoveredMCPTool;
+  readonly messageTaskTool: DiscoveredMCPTool;
+  readonly deleteTaskTool: DiscoveredMCPTool;
+  readonly cancelTaskTool: DiscoveredMCPTool;
+
+  constructor(
+    readonly serverName: string,
+    tools: DiscoveredMCPTool[],
+  ) {
+    const getToolOrFail = (name: string): DiscoveredMCPTool => {
+      for (const tool of tools) {
+        if (tool.serverToolName === name) {
+          return tool;
+        }
+      }
+      throw new Error(`missing expected tool: ${name}`);
+    };
+
+    this.startTaskTool = getToolOrFail('startTask');
+    this.getTaskTool = getToolOrFail('getTask');
+    this.listTasksTool = getToolOrFail('listTasks');
+    this.messageTaskTool = getToolOrFail('messageTask');
+    this.deleteTaskTool = getToolOrFail('deleteTask');
+    this.cancelTaskTool = getToolOrFail('cancelTask');
+  }
+
+  async startTask(prompt: string): Promise<BackgroundAgentTask> {
+    const resp = await this.callTool(this.startTaskTool, {
+      prompt: {
+        role: 'user',
+        parts: [{ text: prompt }],
+      },
+    });
+    const taskResp = await BackgroundAgentTaskResponseSchema.parseAsync(resp);
+    return taskResp.structuredContent;
+  }
+
+  async getTask(
+    id: string,
+    historyLength?: number,
+  ): Promise<BackgroundAgentTask> {
+    const resp = await this.callTool(this.getTaskTool, {
+      id,
+      historyLength,
+    });
+    const taskResp = await BackgroundAgentTaskResponseSchema.parseAsync(resp);
+    return taskResp.structuredContent;
+  }
+
+  async listTasks(): Promise<BackgroundAgentTask[]> {
+    const resp = await this.callTool(this.listTasksTool, {});
+    const tasksResp = await BackgroundAgentTasksResponseSchema.parseAsync(resp);
+    return tasksResp.structuredContent;
+  }
+
+  async messageTask(id: string, message: string) {
+    await this.callTool(this.messageTaskTool, {
+      id,
+      message: {
+        role: 'user',
+        parts: [{ text: message }],
+      },
+    });
+  }
+
+  async deleteTask(id: string) {
+    await this.callTool(this.deleteTaskTool, { id });
+  }
+
+  async cancelTask(id: string) {
+    await this.callTool(this.cancelTaskTool, { id });
+  }
+
+  private async callTool(
+    tool: DiscoveredMCPTool,
+    params: Record<string, unknown>,
+  ): Promise<Record<string, unknown>> {
+    const { llmContent: parts } = await tool.execute(params);
+    if (
+      !Array.isArray(parts) ||
+      parts.length !== 1 ||
+      typeof parts[0] !== 'object' ||
+      parts[0]?.functionResponse?.response === undefined
+    ) {
+      throw new Error('Expected exactly one part with a functionResponse');
+    }
+    const resp = parts[0].functionResponse.response;
+    if ('isError' in resp && resp.isError) {
+      throw new Error(`Error calling ${tool.displayName}: ${resp}`);
+    }
+    return resp;
+  }
+}

packages/core/src/background/backgroundManager.ts

@@ -0,0 +1,40 @@
+/**
+ * @license
+ * Copyright 2025 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { MCPServerConfig } from '../config/config.js';
+import { BackgroundAgent, loadBackgroundAgent } from './backgroundAgent.js';
+
+export async function loadBackgroundAgentManager(
+  backgroundAgentConfigs: Record<string, MCPServerConfig> | undefined,
+  debugMode: boolean,
+): Promise<BackgroundAgentManager> {
+  const agents = await Promise.all(
+    Object.entries(backgroundAgentConfigs ?? {}).map(([name, config]) =>
+      loadBackgroundAgent(name, config, debugMode).catch((error) => {
+        console.error(`Error loading background agent '${name}': ${error}`);
+        return null;
+      }),
+    ),
+  ).then((agents) => agents.filter((agent) => agent !== null));
+  return new BackgroundAgentManager(agents);
+}
+
+export class BackgroundAgentManager {
+  // The active agent. May be empty if none are confgured.
+  activeAgent?: BackgroundAgent;
+
+  constructor(readonly backgroundAgents: BackgroundAgent[]) {
+    if (backgroundAgents.length !== 0) {
+      this.activeAgent = backgroundAgents[0];
+    }
+  }
+
+  setActiveAgentByName(name: string) {
+    this.activeAgent = this.backgroundAgents.find(
+      (agent) => agent.serverName === name,
+    );
+  }
+}

packages/core/src/background/types.ts

@@ -0,0 +1,107 @@
+/**
+ * @license
+ * Copyright 2025 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { z } from 'zod';
+import { Outcome, Language, FunctionResponseScheduling } from '@google/genai';
+
+// Should conform to Part in @google/genai
+export const PartSchema = z.object({
+  videoMetadata: z
+    .object({
+      fps: z.number().optional(),
+      endOffset: z.string().optional(),
+      startOffset: z.string().optional(),
+    })
+    .optional(),
+  thought: z.boolean().optional(),
+  inlineData: z
+    .object({
+      displayName: z.string().optional(),
+      data: z.string(),
+      mimeType: z.string(),
+    })
+    .optional(),
+  fileData: z
+    .object({
+      displayName: z.string().optional(),
+      fileUri: z.string(),
+      mimeType: z.string(),
+    })
+    .optional(),
+  thoughtSignature: z.string().optional(),
+  codeExecutionResult: z
+    .object({
+      outcome: z.nativeEnum(Outcome).optional(),
+      output: z.string().optional(),
+    })
+    .optional(),
+  executableCode: z
+    .object({
+      code: z.string().optional(),
+      language: z.nativeEnum(Language).optional(),
+    })
+    .optional(),
+  functionCall: z
+    .object({
+      id: z.string().optional(),
+      args: z.record(z.unknown()).optional(),
+      name: z.string(),
+    })
+    .optional(),
+  functionResponse: z
+    .object({
+      willContinue: z.boolean().optional(),
+      scheduling: z.nativeEnum(FunctionResponseScheduling).optional(),
+      id: z.string().optional(),
+      name: z.string(),
+      response: z.record(z.unknown()).optional(),
+    })
+    .optional(),
+  text: z.string().optional(),
+});
+
+export const BackgroundAgentMessageSchema = z.object({
+  role: z.enum(['user', 'agent']).describe('The role of the sender.'),
+  parts: z.array(PartSchema).describe('The parts of the message.'),
+});
+
+export const BackgroundAgentTaskStatusSchema = z.object({
+  state: z.enum([
+    'submitted',
+    'working',
+    'input-required',
+    'completed',
+    'failed',
+  ]),
+  message: BackgroundAgentMessageSchema.describe(
+    'Message describing the state of the task.',
+  ).optional(),
+});
+
+export const BackgroundAgentTaskSchema = z.object({
+  id: z.string().describe('The id of the task. Must match `[a-zA-Z0-9.-_]+`'),
+  status: BackgroundAgentTaskStatusSchema.describe(
+    'The current status of the task.',
+  ),
+  history: z
+    .array(BackgroundAgentMessageSchema)
+    .describe('Recent history of messages associated with this task')
+    .optional(),
+});
+
+export type BackgroundAgentMessage = z.infer<
+  typeof BackgroundAgentMessageSchema
+>;
+
+export type BackgroundAgentTask = z.infer<typeof BackgroundAgentTaskSchema>;
+
+export const BackgroundAgentTaskResponseSchema = z.object({
+  structuredContent: BackgroundAgentTaskSchema,
+});
+
+export const BackgroundAgentTasksResponseSchema = z.object({
+  structuredContent: z.array(BackgroundAgentTaskSchema),
+});

packages/core/src/config/config.ts

@@ -45,6 +45,10 @@ import {
   DEFAULT_GEMINI_FLASH_MODEL,
 } from './models.js';
 import { ClearcutLogger } from '../telemetry/clearcut-logger/clearcut-logger.js';
+import {
+  BackgroundAgentManager,
+  loadBackgroundAgentManager,
+} from '../background/backgroundManager.js';
 
 export enum ApprovalMode {
   DEFAULT = 'default',
@@ -127,6 +131,7 @@ export interface ConfigParameters {
   toolCallCommand?: string;
   mcpServerCommand?: string;
   mcpServers?: Record<string, MCPServerConfig>;
+  backgroundAgents?: Record<string, MCPServerConfig>;
   userMemory?: string;
   geminiMdFileCount?: number;
   approvalMode?: ApprovalMode;
@@ -158,6 +163,7 @@ export interface ConfigParameters {
 
 export class Config {
   private toolRegistry!: ToolRegistry;
+  private backgroundAgentManager?: BackgroundAgentManager;
   private readonly sessionId: string;
   private contentGeneratorConfig!: ContentGeneratorConfig;
   private readonly embeddingModel: string;
@@ -172,6 +178,7 @@ export class Config {
   private readonly toolCallCommand: string | undefined;
   private readonly mcpServerCommand: string | undefined;
   private readonly mcpServers: Record<string, MCPServerConfig> | undefined;
+  private readonly backgroundAgents?: Record<string, MCPServerConfig>;
   private userMemory: string;
   private geminiMdFileCount: number;
   private approvalMode: ApprovalMode;
@@ -224,6 +231,7 @@ export class Config {
     this.toolCallCommand = params.toolCallCommand;
     this.mcpServerCommand = params.mcpServerCommand;
     this.mcpServers = params.mcpServers;
+    this.backgroundAgents = params.backgroundAgents;
     this.userMemory = params.userMemory ?? '';
     this.geminiMdFileCount = params.geminiMdFileCount ?? 0;
     this.approvalMode = params.approvalMode ?? ApprovalMode.DEFAULT;
@@ -281,6 +289,10 @@ export class Config {
     if (this.getCheckpointingEnabled()) {
       await this.getGitService();
     }
+    this.backgroundAgentManager = await loadBackgroundAgentManager(
+      this.backgroundAgents,
+      this.debugMode,
+    );
     this.toolRegistry = await this.createToolRegistry();
   }
 
@@ -406,6 +418,10 @@ export class Config {
     return this.mcpServers;
   }
 
+  getBackgroundAgentManager(): BackgroundAgentManager | undefined {
+    return this.backgroundAgentManager;
+  }
+
   getUserMemory(): string {
     return this.userMemory;
   }

packages/core/src/index.ts

@@ -24,6 +24,9 @@ export * from './code_assist/oauth2.js';
 export * from './code_assist/server.js';
 export * from './code_assist/types.js';
 
+export * from './background/types.js';
+export * from './background/backgroundManager.js';
+
 // Export utilities
 export * from './utils/paths.js';
 export * from './utils/schemaValidator.js';

packages/core/src/tools/mcp-tool.ts

@@ -113,9 +113,7 @@ export class DiscoveredMCPTool extends BaseTool<ToolParams, ToolResult> {
         args: params,
       },
     ];
-
-    const responseParts: Part[] = await this.mcpTool.callTool(functionCalls);
-
+    const responseParts = await this.mcpTool.callTool(functionCalls);
     return {
       llmContent: responseParts,
       returnDisplay: getStringifiedResultForDisplay(responseParts),