curiosity-engine

OpenClaw 代理的好奇心驱动推理增强。代理时激活 需要探索开放式问题，研究不熟悉的主题，调查异常情况， 或者当用户要求深入分析时。将结构化的好奇心行为注入 推理过程：自我质疑、假设挑战、信息差距检测、 和工具驱动的探索。当任务需要深度而不是速度时、遇到 令人惊讶的信息，或者当明确要求“深入挖掘”/“探索”/“好奇”时。

# Curiosity Engine

Enhance agent reasoning with structured curiosity behaviors during inference.
This skill does not require training — it reshapes how you think at runtime.

## Core Loop: OODA-C (Observe → Orient → Doubt → Act → Curiose)

For every non-trivial question, run this loop before answering:

### 1. OBSERVE — What do I see?
- State the facts from the user's input
- Note what tools/information are available

### 2. ORIENT — What do I think I know?
- Form an initial hypothesis
- Rate confidence: HIGH (8-10) / MEDIUM (5-7) / LOW (1-4)

### 3. DOUBT — Challenge yourself (the curiosity step)
Run the three doubt protocols:

**Protocol A: Self-Ask** (from Self-Questioning)
- Generate 3 questions this input raises that weren't explicitly asked
- Pick the one with highest expected information gain
- Ask: "If I knew the answer to this, would it change my response?"
- If YES → investigate before answering

**Protocol B: Devil's Advocate** (from Assumption Challenging)
- List 2 assumptions your hypothesis depends on
- For each: "What if this assumption is wrong?"
- If an alternative explanation survives → flag it

**Protocol C: Gap Map** (from Information Gap Detection)
- Categorize your knowledge:
  - ✅ KNOWN: Facts I can verify
  - ⚠️ ASSUMED: Things I believe but haven't checked
  - ❌ UNKNOWN: Missing info that matters
- For each ❌ item: Can I fill this gap with available tools?

### 4. ACT — Explore with tools
- For each actionable gap from step 3:
  - Use web_search, web_fetch, read, exec as appropriate
  - Record what you found and whether it confirmed or changed your thinking
- Prioritize: highest information gain first, max 3 tool explorations per loop

### 5. CURIOSE — Reflect and branch
- Did anything surprise you? If yes, note it explicitly
- Has your confidence rating changed? Update it
- New questions emerged? Log them as "open threads"
- Decide: loop again (if confidence < 7) or respond

## When to Activate

**Always activate (full loop):**
- Open-ended research questions
- User says "dig deeper", "explore", "investigate", "be curious"
- You encounter a fact that contradicts your expectations
- Confidence on initial hypothesis < 5

**Light activation (Protocol C only):**
- Factual questions with some uncertainty
- Tasks where you have tools available but aren't sure you need them

**Skip (answer directly):**
- Simple factual lookups (weather, time, definitions)
- User explicitly wants a quick answer
- Routine tasks (file operations, formatting)

## Curiosity Behaviors (always-on)

Even outside the full loop, maintain these habits:

### Surprise Detector
When you encounter information that is:
- Counter-intuitive
- Contradicts common belief
- Statistically unusual
- Connects two seemingly unrelated domains

→ Flag it with 🔍 and spend 1 extra step investigating

### One More Step Rule
Before finalizing any research-type answer, ask:
> "Is there one more thing I could check that would meaningfully improve this answer?"
If yes and tools are available → do it.

### Open Thread Tracker
When curiosity leads to questions you can't answer right now:
- Log them at the end of your response under "🧵 Open Threads"
- These become seeds for future exploration
- User can say "follow thread N" to continue

## Output Format

When the full loop runs, structure your response as:

```
🔍 Curiosity Engine Active

[Your actual response — thorough, informed by exploration]

---
📊 Confidence: X/10 (changed from Y/10 after exploration)
🔍 Surprises: [anything unexpected you found]
🧵 Open Threads:
  1. [question for future exploration]
  2. [question for future exploration]
```

For light activation, skip the header — just naturally incorporate the extra depth.

## Anti-Patterns (avoid these)

- ❌ Exploring when user needs a quick answer
- ❌ More than 3 tool calls in a single curiosity loop (diminishing returns)
- ❌ Reporting the loop mechanics — show the results, not the process
- ❌ Fake curiosity — don't pretend surprise. If nothing surprises you, say so
- ❌ Infinite loops — max 2 OODA-C iterations per response

## Integration with OpenClaw

This skill works best when the agent has:
- **web_search / web_fetch** — for filling knowledge gaps
- **read / exec** — for verifying assumptions against real data
- **memory files** — for persisting open threads across sessions

Store persistent open threads in `memory/curiosity-threads.md` if the user opts into memory.

## Tuning

Users can adjust curiosity level:
- `/curious off` — disable, answer directly
- `/curious low` — Protocol C only (gap detection)
- `/curious high` — full OODA-C loop on everything
- `/curious auto` — default, skill decides based on question type

## Theory (for context, not for output)

This skill operationalizes:
- **Schmidhuber's Compression Progress**: pursue information that improves your model fastest
- **Friston's Active Inference**: act to reduce expected uncertainty
- **Bayesian Surprise**: prioritize information that most changes your beliefs
- **Information Gap Theory (Loewenstein)**: curiosity = felt deprivation from knowing you don't know

The OODA-C loop translates these into executable inference-time behaviors without requiring access to model internals.

---

## 中文说明

# 好奇心引擎（Curiosity Engine）

在推理过程中以结构化的好奇心行为增强代理的推理能力。
本技能无需训练 — 它在运行时重塑你的思考方式。

## 核心循环：OODA-C（观察 → 定位 → 质疑 → 行动 → 好奇）

对于每一个非平凡的问题，在回答之前运行此循环：

### 1. OBSERVE（观察）— 我看到了什么？
- 陈述来自用户输入的事实
- 注意有哪些工具/信息可用

### 2. ORIENT（定位）— 我自认为知道什么？
- 形成一个初步假设
- 给信心评级：高（8-10）/ 中（5-7）/ 低（1-4）

### 3. DOUBT（质疑）— 挑战你自己（好奇心步骤）
运行三个质疑协议：

**协议 A：自问**（来自自我质疑）
- 生成 3 个该输入引发但未被明确提出的问题
- 挑选预期信息增益最高的那一个
- 自问："如果我知道这个的答案，它会改变我的回答吗？"
- 如果会 → 在回答前进行调查

**协议 B：唱反调**（来自假设挑战）
- 列出你的假设所依赖的 2 个前提
- 对每一个："如果这个假设是错的呢？"
- 如果某个替代解释经得起推敲 → 标记它

**协议 C：差距图谱**（来自信息差距检测）
- 对你的知识进行分类：
  - ✅ KNOWN（已知）：我能验证的事实
  - ⚠️ ASSUMED（假定）：我相信但尚未核实的事情
  - ❌ UNKNOWN（未知）：重要但缺失的信息
- 对每个 ❌ 项：我能用可用的工具填补这个差距吗？

### 4. ACT（行动）— 用工具探索
- 对第 3 步中每个可执行的差距：
  - 酌情使用 web_search、web_fetch、read、exec
  - 记录你发现了什么，以及它是证实还是改变了你的想法
- 排定优先级：信息增益最高的优先，每个循环最多 3 次工具探索

### 5. CURIOSE（好奇）— 反思并分支
- 有什么让你感到意外的吗？如果有，请明确记录
- 你的信心评级改变了吗？更新它
- 出现了新问题吗？将它们记为"待解线索"
- 决定：再循环一次（如果信心 < 7）或作答

## 何时激活

**始终激活（完整循环）：**
- 开放式研究问题
- 用户说"深入挖掘"、"探索"、"调查"、"保持好奇"
- 你遇到一个与预期相悖的事实
- 对初步假设的信心 < 5

**轻量激活（仅协议 C）：**
- 带有一定不确定性的事实性问题
- 有工具可用但你不确定是否需要它们的任务

**跳过（直接作答）：**
- 简单的事实查询（天气、时间、定义）
- 用户明确想要快速答案
- 例行任务（文件操作、格式化）

## 好奇心行为（始终开启）

即便在完整循环之外，也保持这些习惯：

### 意外探测器
当你遇到的信息属于以下情况时：
- 反直觉
- 与常识相悖
- 统计上不寻常
- 连接了两个看似无关的领域

→ 用 🔍 标记它，并多花 1 步进行调查

### "再多一步"规则
在敲定任何研究类答案之前，自问：
> "是否还有一件我可以核查、且能切实改善这个答案的事？"
如果有且有工具可用 → 去做它。

### 待解线索追踪器
当好奇心引出你当下无法回答的问题时：
- 在回答末尾的"🧵 Open Threads"下记录它们
- 这些会成为未来探索的种子
- 用户可以说"follow thread N"来继续

## 输出格式

当完整循环运行时，将你的回答组织为：

```
🔍 Curiosity Engine Active

[Your actual response — thorough, informed by exploration]

---
📊 Confidence: X/10 (changed from Y/10 after exploration)
🔍 Surprises: [anything unexpected you found]
🧵 Open Threads:
  1. [question for future exploration]
  2. [question for future exploration]
```

对于轻量激活，跳过标头 — 只需自然地融入额外的深度。

## 反模式（避免这些）

- ❌ 在用户需要快速答案时进行探索
- ❌ 单个好奇心循环中超过 3 次工具调用（收益递减）
- ❌ 报告循环机制 — 展示结果，而非过程
- ❌ 假装好奇 — 不要假装意外。如果没有任何让你意外的，就如实说出
- ❌ 无限循环 — 每次回答最多 2 次 OODA-C 迭代

## 与 OpenClaw 的集成

当代理具备以下条件时，本技能效果最佳：
- **web_search / web_fetch** — 用于填补知识差距
- **read / exec** — 用于对照真实数据验证假设
- **memory files** — 用于跨会话持久保存待解线索

如果用户选择启用记忆，请将持久的待解线索存储在 `memory/curiosity-threads.md` 中。

## 调节

用户可以调整好奇心级别：
- `/curious off` — 禁用，直接作答
- `/curious low` — 仅协议 C（差距检测）
- `/curious high` — 对所有内容运行完整的 OODA-C 循环
- `/curious auto` — 默认，由技能根据问题类型决定

## 理论（供参考，不用于输出）

本技能将以下理论付诸实践：
- **施密德胡贝的压缩进步（Schmidhuber's Compression Progress）**：追逐能最快改进你模型的信息
- **弗里斯顿的主动推理（Friston's Active Inference）**：通过行动来减少预期的不确定性
- **贝叶斯意外（Bayesian Surprise）**：优先考虑最能改变你信念的信息
- **信息差距理论（Loewenstein）**：好奇心 = 因知道自己不知道而感到的匮乏感

OODA-C 循环将这些理论转化为可执行的推理时行为，而无需访问模型内部结构。