@ backendkit-labs/ 代理编码

生产准备编码代理引擎 @backendkit-labs/agent-core. Bundles 10个专业代理、文件系统工具、多供应商支持(DeepSeek、OpenAI、Anthropic、Ollama、Kimi、Grok)、自动堆栈检测、持续内存、QA审查、反射和管弦管——全部为单一 createCodingEngine() 电话.

状态 : stable ——保持并做好生产准备. 在试验阶段没有新的特点.

目录

• 安装
• 快开始
• 创建编码引擎
• 从配置创建编码引擎
• 供应商
• 内设特工
• 内建工具
• 技术包
• 管弦管
• 反射引擎
• MCP 服务器集成
• 多代理实例
• 配置文件

安装

npm install @backendkit-labs/agent-coding

所有提供者适配器被捆绑——不需要单独的包.

快开始

import { createCodingEngine } from '@backendkit-labs/agent-coding';

const engine = createCodingEngine({
  providers:       { deepseek: { apiKey: process.env.DEEPSEEK_API_KEY! } },
  defaultProvider: 'deepseek',
});

await engine.run('List all TypeScript files in src/ and tell me which ones have no tests');

这个单调给你:10个专门的代理,文件系统工具(读/写/编辑/搜索),持续的项目内存,技能加载,和堆栈自动检测.

创建编码引擎

初等工厂. 返回配置齐全 AgentEngine。 。 。 .

选项参考

interface CodingEngineOptions {
  // Required
  providers:       ProvidersConfig;
  defaultProvider: string;

  // Agents and engine
  defaultAgent?:          string;       // default: 'general'
  workingDir?:            string;       // default: process.cwd()
  maxIterations?:         number;       // default: 100
  maxParallelAgents?:     number;       // default: 6

  // I/O
  transport?:             Transport;    // default: StdoutTransport

  // App identity
  appName?:               string;       // default: 'bk-agent' — used for ~/.{appName}/

  // Persistence
  disablePersistence?:    boolean;      // skip ProjectStore (useful for testing)

  // MCP servers
  mcpServers?:            MCPServerConfig[];

  // Context sharing
  subAgentContextMessages?: number;    // messages visible to sub-agents (recommend 10-20)

  // Iteration control
  onToolApproval?:   (toolName, agentId, argsPreview) => Promise<ToolApprovalDecision>;
  onIterationLimit?: (stats: IterationStats) => Promise<boolean>;
  onStep?:           (stats: IterationStats) => Promise<boolean>;

  // Orchestration
  orchestration?:    boolean | OrchestrationOptions;
}

最小设置

const engine = createCodingEngine({
  providers:       { deepseek: { apiKey: 'sk-...' } },
  defaultProvider: 'deepseek',
});

多个供应商

const engine = createCodingEngine({
  providers: {
    deepseek:  { apiKey: process.env.DEEPSEEK_API_KEY! },
    anthropic: { apiKey: process.env.ANTHROPIC_API_KEY!, model: 'claude-opus-4-8' },
    ollama:    { host: 'http://localhost:11434', model: 'llama3.2' },
  },
  defaultProvider: 'deepseek',
  // Each agent profile can specify a provider override
});

满腔乐曲

const engine = createCodingEngine({
  providers:       { deepseek: { apiKey: '...' } },
  defaultProvider: 'deepseek',
  orchestration:   true,
  // Enables:
  //   intent detection → domain detection → risk scoring
  //   policy engine → agent routing → QA auto-review → reflection
});

使用 MCP 服务器

const engine = createCodingEngine({
  providers:       { openai: { apiKey: '...' } },
  defaultProvider: 'openai',
  mcpServers: [
    { name: 'github',   command: 'npx', args: ['-y', '@modelcontextprotocol/server-github'],  env: { GITHUB_TOKEN: '...' } },
    { name: 'postgres', command: 'npx', args: ['-y', '@modelcontextprotocol/server-postgres'], env: { DATABASE_URL: '...' } },
  ],
  subAgentContextMessages: 15,
});

手动批准工具

const engine = createCodingEngine({
  providers:       { deepseek: { apiKey: '...' } },
  defaultProvider: 'deepseek',
  onToolApproval:  async (toolName, agentId, argsPreview) => {
    const write = ['write_file', 'edit_file', 'run_command'];
    if (!write.includes(toolName)) return 'approve';
    console.log(`\n⚠  [${agentId}] ${toolName}:\n${argsPreview}\n`);
    const a = await prompt('Approve? (y/n/all): ');
    if (a === 'y')   return 'approve';
    if (a === 'all') return 'approve_all';
    return 'reject';
  },
});

从配置创建编码引擎

从 ~/.{appName}/config.json。不需要更改代码以切换提供者 :

import { createCodingEngineFromConfig } from '@backendkit-labs/agent-coding';

const engine = createCodingEngineFromConfig();

// With overrides
const engine = createCodingEngineFromConfig({
  workingDir:    '/my-project',
  orchestration: true,
});

供应商

深层搜索

providers: {
  deepseek: {
    apiKey:       process.env.DEEPSEEK_API_KEY!,
    model?:       'deepseek-chat',    // default
    maxTokens?:   8192,
    temperature?: 0.7,
  }
}

开放AI

providers: {
  openai: {
    apiKey:   process.env.OPENAI_API_KEY!,
    model?:   'gpt-4o',
    baseUrl?: 'https://api.openai.com/v1',
  }
}

Anthropic (克罗地亚)

providers: {
  anthropic: {
    apiKey: process.env.ANTHROPIC_API_KEY!,
    model?: 'claude-sonnet-4-6',    // default
  }
}

Ollama( 本地- 无 API 密钥)

# Prerequisites
ollama serve
ollama pull llama3.2

providers: {
  ollama: {
    host?:  'http://localhost:11434',
    model?: 'llama3.2',
  }
}

OpenAI 兼容( Kimi, Grok, 任意端点)

providers: {
  kimi: { apiKey: process.env.KIMI_API_KEY!, model: 'moonshot-v1-8k' },
  grok: { apiKey: process.env.GROK_API_KEY!, model: 'grok-3' },
  // Any OpenAI-compatible endpoint
  vllm: { apiKey: 'local', baseUrl: 'http://localhost:8080/v1', model: 'my-model' },
}

内设特工

十个专家特工被捆绑起来。 Orchestration 路由自动任务;也可以明确选择.

选择代理

await engine.run('Audit src/ for SQL injection risks', { agentId: 'security-reviewer' });

添加自定义代理

创建 ~/.bk-agent/agents/my-agent.json数字 :

{
  "id":           "my-agent",
  "name":         "My Custom Agent",
  "systemPrompt": "You are specialized in ...",
  "tools":        ["read_file", "run_command"],
  "provider":     "anthropic"
}

已自动装入——不需要更改代码 .

项目代理人和技能需要可信赖的项目

探员在 {project}/.bk-agent/agents/ 和技能 {project}/.bk-agent/skills/ 这是 存储器 内容,而不是你的:一个被克隆的回波可以运送一个代理,用它自己的即时和工具来取代内置的管弦乐器. 所以他们只为你信任的一个项目装上子弹 你自己的经纪人和技能 ~/.bk-agent/ 总是装入.

/trust              # is this project trusted, and what would its project layer load
/trust add          # trust it (records the real path in ~/.bk-agent/trusted-projects.json)
/trust remove       # stop trusting it
/trust list

• 信任来自 准确路径: 信任 ~/code 不相信里面克隆的仓库.
• 对于一个不值得信赖的项目,引擎发出警告,命名了被遗漏的东西. 技能负载紧接其后 /trust add; 代理下次开始时代理.
• BK_TRUST_PROJECT=1 信任每个路径( CI, 容器) 。 自己决定的宿主可以通过 trustedProject: true | false 改为 createCodingEngine。 。 。 .
• 一个缺失或被损坏的信任文件意味着没有人被信任.

内建工具

所有特工都可以通过 CODING_TOOLS数字 :

技术包

根据检测到的项目堆栈激活内置包 :

管弦管

何时 orchestration: true在每次LLM呼叫之前,一个智能的路由层运行:

Input
  → Intent Detection (feature? bug? refactor? question?)
  → Domain Detection (backend? frontend? security? ops?)
  → Risk Scoring (reads only? writes files? runs commands?)
  → Policy Engine (enforce domain-level rules)
  → Agent Selection (best agent for detected domain + risk)
  → LLM Call
  → QA Review (auto-review code-heavy responses)

自定义管弦

orchestration: {
  capabilityMatrix: {
    'my-specialist': {
      owns:       ['payments', 'billing'],
      skills:     ['typescript', 'stripe'],
      baseWeight: 0.95,
    },
  },
  customRules: [
    {
      if:   { domain: 'security' },
      then: { mustInclude: ['security-reviewer'] },
    },
  ],
  enableQA:   true,
  noQA:       false,
  onQAReview: (review) => logger.info('[QA]', review),
}

反射引擎

纵观失败

orchestration: {
  reflection: true,
  // After each sprint:
  // 1. Failures logged to FailureCatalog
  // 2. Patterns with ≥3 occurrences promoted to PolicyRules
  // 3. lecciones-aprendidas.md updated
}

MCP 服务器集成

mcpServers: [
  {
    name:    'code-intelligence',
    command: 'bk-mcp',               // @backendkit-labs/mcp-server
    args:    [],
    env:     { BK_APP_NAME: 'bk-agent', BK_CWD: process.cwd() },
  },
  {
    name:    'github',
    command: 'npx',
    args:    ['-y', '@modelcontextprotocol/server-github'],
    env:     { GITHUB_TOKEN: process.env.GITHUB_TOKEN! },
  },
],
subAgentContextMessages: 20,

多代理实例

特性实施

await engine.run(`
  Implement JWT authentication middleware for Express:
  - Validate Bearer token in Authorization header
  - Attach decoded user to req.user
  - Return 401 with clear error messages
  - Include unit tests with valid/expired/malformed token cases
`);
// architect plans → backend-agent implements → qa-engineer writes tests

安保审计

await engine.run(`
  Audit all routes in src/routes/ for:
  1. Missing auth middleware
  2. SQL injection vulnerabilities
  3. Exposed sensitive data
  Output a prioritized fix list.
`, { agentId: 'security-reviewer' });

数据库迁移

await engine.run(`
  Write a zero-downtime migration: add NOT NULL column 
  'last_login_at TIMESTAMP DEFAULT NOW()' to the users table (50M rows).
  Include rollback script and backfill strategy.
`, { agentId: 'database-agent' });

完整堆栈重构

await engine.run(`
  Refactor auth from session cookies to JWT.
  Update both Express middleware and React login flow.
  Keep backward compatibility for 2 more sprints.
`);
// backend-agent + frontend-agent + security-reviewer

配置文件

~/.bk-agent/config.json数字 :

{
  "defaultProvider": "deepseek",
  "providers": {
    "deepseek": { "apiKey": "sk-xxxx", "model": "deepseek-chat" },
    "anthropic": { "apiKey": "sk-ant-xxxx", "model": "claude-sonnet-4-6" },
    "ollama":    { "host": "http://localhost:11434", "model": "llama3.2" }
  },
  "mcpServers": [
    {
      "name": "github",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_TOKEN": "ghp_xxxx" }
    }
  ]
}