83 lines
2.3 KiB
Markdown
83 lines
2.3 KiB
Markdown
---
|
||
title: "交接协议（Handoff Protocol）"
|
||
type: concept
|
||
tags: [openclaw, memory, model-switch, agentic-ai]
|
||
sources: [养龙虾5天血泪史]
|
||
last_updated: 2026-04-23
|
||
---
|
||
|
||
## Definition
|
||
|
||
交接协议是 AI Agent 在模型切换或会话结束时，将当前上下文状态写入每日日志的规程。解决 OpenClaw Agent 切换模型时丢失所有上下文的核心问题。
|
||
|
||
## The Problem
|
||
|
||
OpenClaw Agent 在切换模型时丢失所有上下文：
|
||
- 新模型以新鲜上下文窗口开始
|
||
- 只看到自动加载的文件
|
||
- 当前会话状态、进行中的任务、待处理决定全部丢失
|
||
|
||
> "切换模型后，Agent 表现得像我们从未交谈过。我提到两天前的讨论决定，它一脸茫然。"
|
||
|
||
## Solution
|
||
|
||
在任何模型切换或会话结束前执行交接：
|
||
|
||
```markdown
|
||
# Handoff Protocol
|
||
|
||
## Current Session State
|
||
- Current task: [task description]
|
||
- Progress: [X% complete]
|
||
- Pending decisions: [list]
|
||
- Next steps: [action items]
|
||
|
||
## What Worked
|
||
- [insight 1]
|
||
- [insight 2]
|
||
|
||
## What Didn't Work
|
||
- [failed approach 1]
|
||
- [failed approach 2]
|
||
```
|
||
|
||
## Implementation
|
||
|
||
### 在 AGENTS.md 顶部添加交接指令
|
||
|
||
```markdown
|
||
# Handoff Protocol (必须执行)
|
||
Before any model switch or session end:
|
||
1. Write current task state to memory/YYYY-MM-DD.md
|
||
2. Include: progress, pending decisions, next steps
|
||
3. Note what worked and what didn't
|
||
4. This is non-negotiable — DO NOT skip
|
||
```
|
||
|
||
### 触发时机
|
||
|
||
- `/model` 命令切换模型
|
||
- `/exit` 或 `/quit` 结束会话
|
||
- 长时间无响应后的新会话
|
||
- 主动要求交接时
|
||
|
||
## Key Insight
|
||
|
||
> "交接协议是模型切换的修复"
|
||
|
||
没有交接协议，新模型不知道发生了什么。有了交接协议，新模型从每日日志读取当前状态，无缝继续工作。
|
||
|
||
## 与上下文刷新的关系
|
||
|
||
- **上下文刷新**（Memory Flush）：防止单次压缩周期内的信息丢失
|
||
- **交接协议**：防止模型切换时的信息丢失
|
||
|
||
两者互补，共同确保长期会话的信息完整性。
|
||
|
||
## Connections
|
||
- [[上下文压缩]] ← 交接协议解决压缩无法覆盖的多次压缩问题
|
||
- [[上下文刷新]] ← 互补机制
|
||
- [[写入纪律]] ← 交接协议是写入纪律的具体场景
|
||
- [[自动加载文件]] ← 新模型只看到自动加载的文件，交接日志弥补缺失
|
||
- [[养龙虾5天血泪史]] ← 主要来源
|