48 lines
1.7 KiB
Markdown
48 lines
1.7 KiB
Markdown
---
|
||
title: "Context Anxiety"
|
||
type: concept
|
||
tags:
|
||
- "agentic-ai"
|
||
- "context-window"
|
||
- "failure-mode"
|
||
sources:
|
||
- "Your-AI-Isn-t-Stupid---It-Just-Needs-a-Better-Harness--Lychee-Technology-Engineering-Blog"
|
||
last_updated: 2026-04-20
|
||
---
|
||
|
||
## Overview
|
||
Context Anxiety——当 LLM 的 context window 使用率超过约 70% 容量,或延迟升高时,模型表现出"仓促"行为的现象:跳过步骤、过早完成任务或过早宣告成功。
|
||
|
||
## Mechanism
|
||
- Context window 是模型的唯一记忆空间
|
||
- 当感知到"墙壁在逼近"(token 限制),模型开始优先"快速完成"而非"正确完成"
|
||
- 这不是模型能力问题,而是 context 容量压力的系统性反应
|
||
|
||
## Detection
|
||
- 监控 `tokens_used / max_context > 0.7` 阈值(需按模型和工作负载调优)
|
||
- 延迟 spikes 也是触发信号
|
||
|
||
## Solution: Context Reset
|
||
当 Context Anxiety 触发时,Harness 执行程序化 Context Reset:
|
||
1. `save_state_to_disk(state)` — 完整项目状态写入持久存储
|
||
2. `terminate_current_instance()` — 终止当前 LLM 实例
|
||
3. `launch_fresh_agent(state)` — 启动全新 Agent,从保存状态恢复
|
||
|
||
关键代码:
|
||
```python
|
||
if (tokens_used / max_context) > 0.7:
|
||
save_state_to_disk(state)
|
||
terminate_current_instance()
|
||
launch_fresh_agent(state)
|
||
```
|
||
|
||
## Note on In-Place Summarization
|
||
原地摘要(in-place summarization)不够——它仍然让模型在杂乱、退化的 context 上操作。Context Reset 给予模型干净的处理空间。
|
||
|
||
## Source
|
||
- [[Your-AI-Isn-t-Stupid---It-Just-Needs-a-Better-Harness--Lychee-Technology-Engineering-Blog]]
|
||
|
||
## See Also
|
||
- [[Context-Reset]] — 具体实现机制
|
||
- [[7-Layer-Harness-Stack]] — 第 5 层 Memory & State 和第 7 层 Constraints & Recovery 中处理此问题
|