letta-fleet-management

lettactl

用于以声明式方式管理 Letta AI Agent 群组的 kubectl 风格 CLI。

适用场景

• 使用共享配置部署多个 Agent
• 管理 Agent 的记忆块、工具和文件夹
• 将模板应用到现有 Agent
• 在推广到生产环境前运行金丝雀部署
• 多租户 Agent 管理（B2B / B2B2C）
• 在 Agent 群组中批量发送消息
• 在不同环境之间导入/导出 Agent
• 分析 Agent 记忆健康状况（自我诊断）
• 使用首条消息启动序列校准 Agent
• 通过 SDK 以编程方式管理 Agent 群组

核心工作流

1. 在 fleet.yaml 中定义 Agent
2. 使用 lettactl apply -f fleet.yaml 应用配置
3. 使用 lettactl get agents 和 lettactl describe agent <name> 验证

群组 YAML 结构

shared_blocks:
  - name: company-context
    description: Shared company knowledge
    limit: 5000
    from_file: ./context/company.md

shared_folders:
  - name: brand_docs
    files:
      - "docs/*.md"

mcp_servers:
  - name: firecrawl
    type: sse
    server_url: "https://sse.firecrawl.dev"
    auth_header: "Authorization"
    auth_token: "Bearer ${FIRECRAWL_API_KEY}"

agents:
  - name: support-agent
    description: Customer support assistant
    tags:
      - "tenant:acme-corp"
      - "role:support"
    system_prompt:
      from_file: ./prompts/support.md
    llm_config:
      model: google_ai/gemini-2.5-pro
      context_window: 128000
    reasoning: true
    first_message: "Initialize and confirm readiness."
    memory_blocks:
      - name: persona
        description: Agent personality
        limit: 2000
        value: "You are a helpful support agent."
        agent_owned: true
    archives:
      - name: knowledge_base
        description: Long-term knowledge storage
    shared_blocks:
      - company-context
    shared_folders:
      - brand_docs
    tools:
      - send_email
      - search_docs
      - "tools/*"
    mcp_tools:
      - server: firecrawl
        tools: ["scrape", "crawl"]

完整架构请参阅 reference/fleet-config.md。

CLI 命令

应用配置

lettactl apply -f fleet.yaml                    # Create/update agents
lettactl apply -f fleet.yaml --dry-run          # Preview changes
lettactl apply -f fleet.yaml --match "*-prod"   # Template mode
lettactl apply -f fleet.yaml --canary           # Deploy canary copies
lettactl apply -f fleet.yaml --promote          # Promote canary to production
lettactl apply -f fleet.yaml --recalibrate      # Re-send calibration messages

检查资源

lettactl get agents                        # List all agents
lettactl get agents -o wide                # With details
lettactl get agents --tags "tenant:acme"   # Filter by tags
lettactl get blocks --shared               # Shared blocks only
lettactl get tools --orphaned              # Unused tools
lettactl describe agent <name>             # Full agent details

消息

lettactl send <agent> "Hello"              # Send message
lettactl send <agent> "Hi" --stream        # Stream response
lettactl send --all "support-*" "Update"   # Bulk send by pattern
lettactl send --tags "role:support" "Hi"   # Bulk send by tags
lettactl messages list <agent>             # View history
lettactl messages reset <agent>            # Clear history
lettactl messages compact <agent>          # Summarize history

导入 / 导出

lettactl export agent <name> -f yaml       # Export single agent
lettactl export agents --all               # Export entire fleet
lettactl import agent-export.yaml          # Import agent

群组报告

lettactl report memory                     # Memory usage report
lettactl report memory --analyze           # LLM-powered deep analysis

所有选项请参阅 reference/cli-commands.md。

金丝雀部署

先在隔离副本上测试更改，再推广到生产环境：

lettactl apply -f fleet.yaml --canary           # Create CANARY-* copies
lettactl send CANARY-support-agent "test msg"   # Test the canary
lettactl apply -f fleet.yaml --promote          # Promote to production
lettactl apply -f fleet.yaml --cleanup          # Remove canary agents

请参阅 reference/canary-deployments.md。

多租户

为 Agent 添加标签，以便按 B2B 和 B2B2C 筛选：

agents:
  - name: acme-support
    tags:
      - "tenant:acme-corp"
      - "role:support"
      - "env:production"

lettactl get agents --tags "tenant:acme-corp"
lettactl send --tags "tenant:acme-corp,role:support" "Policy update"

请参阅 reference/multi-tenancy.md。

自我诊断

分析整个 Agent 群组的记忆健康状况：

lettactl report memory                  # Usage stats for all agents
lettactl report memory --analyze        # LLM-powered analysis per agent

报告会列出填充百分比、过时数据、冗余、缺失知识和拆分建议。请参阅 reference/self-diagnosis.md。

Agent 校准

创建 Agent 时使用启动消息进行预热：

agents:
  - name: support-agent
    first_message: "Review your persona and confirm you understand your role."

更新后重新校准现有 Agent：

lettactl apply -f fleet.yaml --recalibrate
lettactl apply -f fleet.yaml --recalibrate --recalibrate-tags "role:support"

请参阅 reference/agent-calibration.md。

模板模式

将配置应用于符合某种模式的现有 Agent：

lettactl apply -f template.yaml --match "*-draper"

采用三方合并：保留用户添加的资源，同时更新受管理的资源。请参阅 reference/template-mode.md。

SDK 用法

import { LettaCtl } from 'lettactl';

const ctl = new LettaCtl({ lettaBaseUrl: 'http://localhost:8283' });

// Deploy from YAML
await ctl.deployFromYaml('./fleet.yaml');

// Programmatic fleet config
const config = ctl.createFleetConfig()
  .addSharedBlock({ name: 'kb', description: 'Knowledge', limit: 5000, from_file: 'kb.md' })
  .addAgent({
    name: 'support-agent',
    description: 'Support AI',
    system_prompt: { from_file: 'prompts/support.md' },
    llm_config: { model: 'google_ai/gemini-2.5-pro', context_window: 32000 },
    shared_blocks: ['kb'],
    tags: ['team:support'],
  })
  .build();

await ctl.deployFleet(config);

// Send message with callbacks
await ctl.sendMessage('agent-id', 'Hello', {
  onComplete: (run) => console.log('Done:', run.id),
});

// Template mode
await ctl.deployFromYaml('./template.yaml', { match: '*-prod' });

完整 API 请参阅 reference/sdk-usage.md。