04 Context compression：自动压缩与手动 compact

它是什么

Context compression 是 Deep Agents 控制长任务上下文膨胀的内置机制。大工具输入或输出会被 offload 到虚拟文件系统，消息历史逼近上下文限制时会触发 summarization。需要主动清理上下文时，可以额外给 Agent 加 compact_conversation 工具。

最小代码

文件：deepagent_src/context_engineering_teach/04_context_compression.py

model = get_real_model()
compaction_middleware = create_summarization_tool_middleware(model, StateBackend)
agent = create_deep_agent(model=model, middleware=[compaction_middleware])

运行

uv run python deepagent_src/context_engineering_teach/04_context_compression.py

预期输出：

context compression real agent ok

验证方式

脚本断言 middleware 暴露了 compact_conversation 工具，并真实调用一次 Agent。自动 offloading 和阈值 summarization 需要真实长对话才会触发，所以这里不伪造大上下文。

常见误区

别以为加了手动 compact 就关闭了自动压缩。Deep Agents 默认仍会在上下文接近限制时自动总结。

相关资源

查看示例代码：deepagent_src/context_engineering_teach/04_context_compression.py

from __future__ import annotations

from deepagents import create_deep_agent
from deepagents.backends import StateBackend
from deepagents.middleware.summarization import create_summarization_tool_middleware

from _model import get_real_model
from deepagent_src.agent_output import invoke_and_pretty_print


def main() -> None:
    model = get_real_model()
    compaction_middleware = create_summarization_tool_middleware(model, StateBackend)
    agent = create_deep_agent(model=model, middleware=[compaction_middleware])

    tool_names = [tool.name for tool in compaction_middleware.tools]
    result = invoke_and_pretty_print(
        agent,
        {
            "messages": [
                {
                    "role": "user",
                    "content": "用一句中文回答：compact_conversation 工具什么时候适合使用？",
                }
            ]
        }
    )

    assert agent is not None
    assert "compact_conversation" in tool_names
    assert result["messages"][-1].content
    print("context compression real agent ok")


if __name__ == "__main__":
    main()