LBS Agent

LBS Agent 是一个自主 Agent 运行时系统。当前仓库的正式架构已经收口为五层：

• application/：composition root + application services
• runtime/：runtime kernel（LLM / context / memory / orchestration / perception）
• interfaces/：CLI / API / gateway / web 入口
• extensions/：skills 与 optional MCP catalogs
• tools/：tool runtime、skill/mcp provider、执行基础设施

legacy 顶层包（如 app/、core/、cli/）目前仅作为 compatibility shim 保留，不再是正式实现目录。

5 分钟上手

安装 / 启动

git clone <repo>
cd lbs-agent
pip install -r requirements.txt   # 或 poetry install
python lbs.py --help

也可以继续使用顶层入口：

python lbs.py --help

配置 LLM Provider

python lbs.py llm inspect
python lbs.py models
python lbs.py config

模型与 agent 内置配置位于：

• runtime/core/llm/registry/models.yaml
• runtime/core/sub_agents/agents.yaml

第一次跑 Agent

python lbs.py run "say hello"
python lbs.py run-agent code "write fizzbuzz in python"

多 Agent 编排

python lbs.py run-agent research "调研 Hermes 和 LBS 的区别"
python lbs.py run-agent summary "总结上一条结果"

当前架构

1. application/

• application/container.py：系统装配根
• application/runtime_factory.py：构建单 agent runtime config
• application/services/：use-case services（model / agent / orchestration / session / cron / tool / interactive / gateway / skill / mcp）

2. runtime/

• runtime/core/：ModelRouter、provider、sub-agent registry
• runtime/context/：context builder / manager / trimming / compression
• runtime/memory/：session store、working memory、long-term memory、task contracts
• runtime/orchestration/：Orchestrator、AgentLoop、scheduler、verification
• runtime/perception/：intent / entity / input preprocess

3. interfaces/

• interfaces/cli/_impl/：CLI 命令实现
• interfaces/api/_impl/server.py：FastAPI 服务
• interfaces/gateway/_impl/：gateway adapters
• interfaces/web/frontend/：前端
• interfaces/web/legacy_web/：历史 web 资源承接区

4. extensions/

• extensions/skills/builtin/
• extensions/skills/optional/
• extensions/mcps/optional/

5. tools/

• tools/tool/：ToolRuntime、tool policy、tool registry
• tools/skill/：skill registry / provider
• tools/mcp/：MCP registry / provider infra

依赖方向

正式依赖方向应保持：

interfaces -> application -> runtime
tools -----^            -> tools infra
extensions (asset catalogs, loaded by runtime/tools when needed)

约束：

• interfaces/ 不直接依赖 runtime internals
• runtime/ 不反向 import application/
• legacy shim 只做兼容转发，不承载真实逻辑

常用命令

lbs models
lbs model list [--config]
lbs model switch <provider> <model>
lbs model health [provider]

lbs llm inspect
lbs llm validate [--json]
lbs llm diagnostics [--json]
lbs llm route <agent>

lbs agents
lbs tools inspect
lbs tools list [--agent name]
lbs tools profiles

lbs skills list
lbs skills view <name>
lbs skill list
lbs skill sources
lbs skill sync <repo> [alias]

lbs mcp list

lbs sessions list
lbs sessions show <id>
lbs sessions export <id> <out>
lbs sessions report [id]

lbs workers list
lbs workers show <id>

lbs cron list
lbs run "<task>"
lbs run-agent <name> "<task>"

编程 API

Container

from application.container import AgentSystemContainer
from config.config import AppConfig

container = AgentSystemContainer(config=AppConfig())
container.session_service().list_sessions(limit=10)
container.model_service().inspect_llm()
container.build_orchestrator()

Orchestrator

result = container.build_orchestrator().run_agent_sync(
    agent_name="code",
    task="write hello world",
    context="",
    session_id="",
)
print(result.success, result.content)

开发

focused tests

python -m pytest tests/architecture/ -v
python -m pytest tests/test_app_services_boundary.py tests/test_server_api.py -q

目录收口说明

目录大迁移已完成；当前剩余 legacy 顶层目录主要用于 import compatibility。后续收尾以：

1. 缩小 shim 表面积
2. 让测试与文档全面直连新路径
3. 最终删除不再需要的 legacy 目录 为主。