skill-upgrader

技能升级器

一种元技能，使用 4 个由 Ragie 支持的并行 Agent，将任意 SKILL.md 升级为 Decision Theory v5 Hybrid 格式。

使用时机

• “将此技能升级到 v5”
• “使用决策理论形式化此技能”
• “为此技能添加 MDP 结构”
• “对 X 应用 skill-upgrader”

前置条件

Ragie RAG 已索引以下书籍：

• decision-theory 分区：LaValle Planning Algorithms、Sutton & Barto RL
• modal-logic 分区：Blackburn Modal Logic、Huth & Ryan Logic in CS

工作流

步骤 1：设置会话

SESSION=$(date +%Y%m%d-%H%M%S)-upgrade-{skill_name}
mkdir -p thoughts/skill-builds/${SESSION}

步骤 2：初始化黑板

创建 thoughts/skill-builds/{session}/00-blackboard.md：

# Skill Upgrade: {skill_name}
Started: {timestamp}

## Input Skill
{path_to_skill}

## Target Format
Decision Theory v5 Hybrid

## Agent Findings
(Agents append below)

---

步骤 3：并行启动 4 个 Agent

使用 Task 工具同时启动全部 4 个 Agent。每个 Agent：

1. 读取输入技能
2. 在 Ragie 中查询指定书籍
3. 将发现结果追加到黑板

---

Agent 1：LaValle 规划器

书籍： LaValle 的《Planning Algorithms》（decision-theory 分区） 重点： 状态、动作、转移

Task(
  subagent_type="general-purpose",
  prompt="""
INPUT SKILL: {path}
BLACKBOARD: thoughts/skill-builds/{session}/00-blackboard.md

YOUR BOOK: LaValle's "Planning Algorithms" in Ragie partition 'decision-theory'

TASK: Identify MDP structure in the skill.

Query Ragie:
```bash
uv run python scripts/ragie_query.py -q "MDP state space definition" -p decision-theory
uv run python scripts/ragie_query.py -q "action space sequential decisions" -p decision-theory
uv run python scripts/ragie_query.py -q "POMDP partial observability" -p decision-theory

读取输入技能并回答：

1. 什么是状态？（阶段、模式、跟踪的信息）
2. 什么是动作？（Agent 在每种状态下可以做什么）
3. 转移如何运作？（确定性还是随机性）
4. 这是 POMDP 还是完全可观测？

写入黑板章节：## Agent 1: States, Actions & Transitions

用通俗英语撰写，并引用 LaValle 的章节。 """ )

---

## Agent 2: Sutton & Barto Optimizer

**Book:** Sutton & Barto's "Reinforcement Learning" (decision-theory partition)
**Focus:** Policy, Termination, Value
**Depends on:** Agent 1

Task( subagent_type="general-purpose", prompt=""" INPUT SKILL: {path} BLACKBOARD: thoughts/skill-builds/{session}/00-blackboard.md

YOUR BOOK: Sutton & Barto's "Reinforcement Learning" in Ragie partition 'decision-theory'

WAIT: Read Agent 1's findings from blackboard first.

TASK: Design policy and termination conditions.

Query Ragie:

uv run python scripts/ragie_query.py -q "policy deterministic stochastic" -p decision-theory
uv run python scripts/ragie_query.py -q "episodic termination conditions" -p decision-theory
uv run python scripts/ragie_query.py -q "reward function design" -p decision-theory

Using Agent 1's states and actions, answer:

1. What's the POLICY? (state → action rules)
2. When does it END? (terminal states, success/failure)
3. What are REWARDS? (goals +, costs -)
4. Which states are HIGH/LOW value?

WRITE to blackboard section: ## Agent 2: Policy & Values

Format as plain English with Sutton & Barto section citations. """ )

---

## Agent 3: Blackburn Modal Logician

**Book:** Blackburn's "Modal Logic" (modal-logic partition)
**Focus:** Constraints (temporal, epistemic, deontic)

Task( subagent_type="general-purpose", prompt=""" INPUT SKILL: {path} BLACKBOARD: thoughts/skill-builds/{session}/00-blackboard.md

YOUR BOOK: Blackburn's "Modal Logic" in Ragie partition 'modal-logic'

TASK: Extract constraints from the skill.

Query Ragie:

uv run python scripts/ragie_query.py -q "temporal logic LTL operators" -p modal-logic
uv run python scripts/ragie_query.py -q "epistemic logic knowledge" -p modal-logic
uv run python scripts/ragie_query.py -q "deontic logic obligations" -p modal-logic

Read the input skill and identify:

1. TEMPORAL: "must do X before Y" → □, ◇, U
2. EPISTEMIC: "must know X" → K operator
3. DEONTIC: "must/forbidden/may" → O, F, P
4. DYNAMIC: "action causes effect" → [action]

WRITE to blackboard section: ## Agent 3: Constraints

For each constraint:

• Plain English description
• Modal logic notation
• Why it matters
• Blackburn chapter citation """ )

---

## Agent 4: Huth & Ryan Verifier

**Book:** Huth & Ryan's "Logic in Computer Science" (modal-logic partition)
**Focus:** Validation, Safety, Liveness
**Depends on:** Agents 1-3

Task( subagent_type="general-purpose", prompt=""" INPUT SKILL: {path} BLACKBOARD: thoughts/skill-builds/{session}/00-blackboard.md

YOUR BOOK: Huth & Ryan's "Logic in Computer Science" in Ragie partition 'modal-logic'

WAIT: Read Agents 1-3 findings from blackboard first.

TASK: Verify consistency and completeness.

Query Ragie:

uv run python scripts/ragie_query.py -q "safety properties verification" -p modal-logic
uv run python scripts/ragie_query.py -q "liveness properties eventually" -p modal-logic
uv run python scripts/ragie_query.py -q "model checking CTL" -p modal-logic

Check:

1. SAFETY: What bad things never happen? □¬(bad)
2. LIVENESS: What good things eventually happen? ◇(good)
3. CONSISTENCY: Any contradictions between agents?
4. COMPLETENESS: Any gaps in coverage?

WRITE to blackboard section: ## Agent 4: Verification

Report with ✓/✗ for each property. Overall verdict: PASS or NEEDS_WORK Huth & Ryan section citations. """ )

---

## Step 4: Synthesize Final Skill

After all agents complete, read the blackboard and create:

**Output:** `thoughts/skill-builds/{session}/SKILL-upgraded.md`

Use v5 Hybrid template:

```yaml
---
name: {original_name}
description: {original_description}
version: 5.1-hybrid
---

# Option: {name}

## Initiation (I)
[From original + Agent 1 state analysis]

## Observation Space (Y)
[From Agent 1 POMDP analysis]

## Action Space (U)
[From Agent 1 actions]

## Policy (pi)
[From Agent 2 state→action rules]

## Termination (beta)
[From Agent 2 episode structure]

## Q-Heuristics
[From Agent 2 value guidance]

## Constraints
[From Agent 3 modal logic]

## Verification
[From Agent 4 safety/liveness]

---

使用示例

User: "Upgrade .claude/skills/implement_plan/SKILL.md to v5 Hybrid"

Claude:
1. Creates session directory
2. Initializes blackboard
3. Launches 4 agents in parallel (Task tool)
4. Waits for completion
5. Reads blackboard
6. Synthesizes upgraded skill
7. Reports: "Upgraded skill at thoughts/skill-builds/.../SKILL-upgraded.md"

Ragie 查询参考

# Decision theory partition
uv run python scripts/ragie_query.py -q "your question" -p decision-theory

# Modal logic partition
uv run python scripts/ragie_query.py -q "your question" -p modal-logic

# With reranking for better results
uv run python scripts/ragie_query.py -q "your question" -p decision-theory --rerank

创建的文件

升级后：

thoughts/skill-builds/{session}/
├── 00-blackboard.md      # Agent collaboration
├── SKILL-upgraded.md     # Final v5 Hybrid skill
└── validation-report.md  # Agent 4 verification