Direction-Preserving Context Engine

基于 CDA Protocol 的上下文引擎

让 Agent 在只保留 10% 历史信息的前提下，方向依然不跑偏。

For Agents who refuse to walk in circles.

English | 中文

---

中文

痛点：长对话里的“失忆”与“鬼打墙”

当对话超过 500 轮，Agent 开始出现三种典型崩溃：

1. 忘记早期约束 —— 最初定好的目标、格式或禁忌被抛到脑后；
2. 重复走死路 —— 在代码修复、工具调用、多步推理中，反复尝试同一个已被证伪的策略；
3. 主题飘走 —— 讨论逐渐偏离全局任务，陷入无关细节。

这不是上下文窗口不够大，而是方向管理出了问题。

承诺：CDA 的三个可验证赢面

赢面	普通 Agent	启用 CDA
超长对话方向稳定	1500 轮后主题偏离度 > 0.4	偏离度 < 0.25
死路不再走第二次	工具密集型任务死路重复率 20–40%	< 10%，目标 < 5%
10% 历史撑起 80% 一致性	压缩后有效上下文密度低	有效密度提升 2–3×

核心差异：CDA 不是帮你「记住更多」，而是帮你「记住哪些路已经证明是错的」，并在每次组装上下文时过滤掉已知错误方向。

核心机制

• 方向追踪（Direction Tracking）：记录当前语义方向 D 与全局意图的偏差，而非原始 token；
• 死路注册表（Dead-End Registry）：将被证伪的推理链 / 工具调用组合登记为负向经验，assemble 阶段以系统提示注入，作为显式排除约束；
• 语义压缩图（SCG）：以拓扑结构保留推理骨干，即使压缩 90%，逻辑链条依然完整。

5 行接入示例

from cda import SPCContext

ctx = SPCContext()
for turn in conversation:
    ctx.append_turn(turn)
    assembled = ctx.assemble_context(query=turn.user_input, max_tokens=8_000)
    response = llm(assembled)

最小接口

// 写入一轮对话与工具轨迹
appendTurn({ messages, tools, phase })

// 组装本轮需要喂给模型的最小上下文
assembleContext({ query: CurrentIntent, max_tokens: number }) => AssembledContext

// 触发压缩（可由外部按需调用）
compact({ strategy?: 'auto' | 'aggressive' })

// 查询语义方向状态（调试、监控）
getDirectionState() => {
  global_intent: IntentVector,
  drift_score: number,
  dead_end_matches: DeadEndMatch[]
}

// 显式注册一条死路（负向经验）
registerDeadEnd({ session_id, trace, reason })

内容结构

cda_protocol/
├── README.md              # 本页：快速入门与产品定位
├── METRICS.md             # 指标体系与可复现 benchmark 说明
├── cda-book-zh.md         # 中文完整书稿（理论深度）
├── cda-book-en.md         # English complete manuscript
├── skill/                 # 可插拔 skill 实现
│   ├── skill.json
│   └── hooks/
├── images/                # 双语配图
└── chapters/              # 分章节版本

许可

本作品采用 CC BY-ND 4.0（署名-禁止演绎 4.0 国际）许可协议。

---

English

The Problem: Amnesia and Déjà Vu in Long Conversations

Beyond 500 turns, Agents exhibit three classic failure modes:

1. Forgotten constraints — early goals, formats, or prohibitions slip away;
2. Dead-end loops — in code repair, tool use, or multi-step reasoning, the same disproven strategy is tried again and again;
3. Topic drift — the conversation slowly slides into irrelevant details, losing sight of the global objective.

This is not a context-window size problem. It is a direction-management problem.

The Promise: Three Verified Wins with CDA

Win	Baseline Agent	CDA Enabled
Ultra-long alignment stability	Topic drift > 0.4 after 1,500 turns	Drift < 0.25
Dead ends are not repeated	Repetition rate 20–40% in tool-heavy tasks	< 10%, target < 5%
10% history → 80% consistency	Low effective context density after compression	Effective density 2–3× higher

Key difference: CDA does not help you "remember more." It helps you remember which directions are already proven wrong, and filters them out at every assemble stage.

Core Mechanisms

• Direction Tracking: Tracks the current semantic direction D and its deviation from global intent, rather than raw tokens.
• Dead-End Registry: Records disproven reasoning chains / tool-call patterns as negative experience, injecting them as explicit exclusion constraints during assemble.
• Semantic Compression Graph (SCG): Preserves the topological shape of reasoning; even after 90% compression, the logical skeleton survives.

5-Line Integration Example

from cda import SPCContext

ctx = SPCContext()
for turn in conversation:
    ctx.append_turn(turn)
    assembled = ctx.assemble_context(query=turn.user_input, max_tokens=8_000)
    response = llm(assembled)

Minimal API

// Append one turn plus tool trace
appendTurn({ messages, tools, phase })

// Assemble the minimal context needed for this turn
assembleContext({ query: CurrentIntent, max_tokens: number }) => AssembledContext

// Trigger compression on demand
compact({ strategy?: 'auto' | 'aggressive' })

// Inspect semantic direction state (for debugging / monitoring)
getDirectionState() => {
  global_intent: IntentVector,
  drift_score: number,
  dead_end_matches: DeadEndMatch[]
}

// Explicitly register a dead-end (negative experience)
registerDeadEnd({ session_id, trace, reason })

Repository Structure

cda_protocol/
├── README.md              # This page: quick start & product positioning
├── METRICS.md             # Metrics & reproducible benchmark guide
├── cda-book-zh.md         # Chinese manuscript (deep theory)
├── cda-book-en.md         # English manuscript
├── skill/                 # Pluggable skill implementation
│   ├── skill.json
│   └── hooks/
├── images/                # Bilingual figures
└── chapters/              # Chapter-by-chapter versions

License

This work is licensed under CC BY-ND 4.0 (Attribution-NoDerivatives 4.0 International).