Requirement Comprehension Engine

Advanced requirement comprehension and skill orchestration engine. Analyzes complex/multi-step user requests, determines which skills to invoke and in what order, builds task decomposition trees. Delegates memory management to complex-memory-manager and self-iteration to self-iteration-engine for cross-skill consistency. Triggers: multi-intent requests, ambiguous instructions, requests spanning multiple skill domains, structured requirement analysis.

A metacognitive skill that equips agents with structured requirement analysis, skill selection & orchestration. Delegates memory management to complex-memory-manager and self-iteration to self-iteration-engine for cross-skill consistency.

Core Capabilities

1. Requirement Analysis & Decomposition

When the agent receives a complex or ambiguous request:

Step 1: Intent Classification

• Parse incoming message for explicit intent signals (verbs, domain keywords, output expectations)
• Classify into: Direct Action / Information Request / Creative Generation / Analysis & Insight / Meta-Request / Multi-Step Workflow
• If ambiguous, flag as Under-Specified and use elicitation templates

Step 2: Structured Decomposition Break the request into a tree:

text

Original Request
├── Intent A (primary)
│   ├── Sub-goal A1 → skill/action candidate
│   ├── Sub-goal A2 → skill/action candidate
│   └── Dependency check
├── Intent B (secondary)
│   └── ...
└── Cross-cutting concerns (privacy, cost, performance)

Step 3: Skill Mapping For each sub-goal, select the best matching skill:

• Check name + description of all available skills (always in context as metadata)
• Rank matching skills by semantic relevance
• If no match → fallback to general knowledge
• If multiple → pick the most specific or orchestrate

2. Skill Orchestration Logic

Situation	Action
Single clear intent, single skill	Direct invocation
Multiple independent intents	Parallel execution
Multi-step pipeline (output A → input B)	Sequential orchestration with handoff
Full autonomy requested	Analyze, decide, execute, explain
Conflicting skill capabilities	Prefer skill with more specific description

Handoff Protocol:

text

[ComprehensionEngine] Handoff: Skill-A → Skill-B
  Context: {user's original intent excerpt}
  Skill-A result: {summary}
  Skill-B requirement: {what Skill-B needs to continue}

3. Requirement Elicitation Templates

When a request is under-specified:

Pattern A: Missing Output Format

"I need to confirm the output format — do you want a daily report, summary, table, chart, code, or something else?"

Pattern B: Missing Scope

"This request seems broad. Could you narrow down the focus? For example: Option A, Option B, or Option C?"

Pattern C: Ambiguous Intent

"I see two possible directions: do you want me to analyze existing data, or go out and collect new data?"

Pattern D: Multi-Step Confirmation

"I plan to do this in three steps: ① collect latest market data ② analyze trends ③ generate a visualization. Does that work?"

4. Delegation to Shared Components

Function	Delegated To	How
Persistent memory (T1/T2/T3)	complex-memory-manager	Use its encrypt/decrypt, storage conventions, and cleanup protocols
Usage logging & performance tracking	self-iteration-engine	Use its log format, review cycles, and update matrix
Self-iteration decisions	self-iteration-engine	Check triggers, run update decision matrix
New skill creation proposals	self-iteration-engine	Use its proposal template

Quick Reference: Skill Selection Flow

text

User speaks
  ↓
Parse & Classify Intent
  ↓
Is it clear? ──No──→ Use Elicitation Templates
  │                     ↓
  Yes←── User clarifies
  ↓
Decompose into sub-goals
  ↓
For each sub-goal:
  ├── Match skill by name+description
  ├── If match: check last-review-date; if >30d → quick scan SKILL.md
  ├── If no match: fallback to general knowledge
  └── If multi-skill: orchestrate (parallel or sequential)
  ↓
Execute & Deliver
  ↓
Log outcome → Delegate to self-iteration-engine
  ↓
[Periodic] Delegate review to self-iteration-engine + memory cleanup to complex-memory-manager

Dependency Check

Before each orchestration run:

1. Verify complex-memory-manager metadata description is in context
2. Verify self-iteration-engine metadata description is in context
3. If either is missing, they may not be installed — fallback to built-in behavior

---

🧠 需求理解引擎

一个元认知技能，赋予Agent结构化的需求分析、技能选择与编排能力。将持久化记忆委托给 complex-memory-manager，将自迭代委托给 self-iteration-engine，实现跨技能一致性。

核心能力

1. 需求分析与分解

当Agent收到复杂或模糊请求时：

第一步：意图分类

• 解析输入信息的意图信号
• 分类：直接操作 / 信息查询 / 创意生成 / 分析洞察 / 元请求 / 多步工作流
• 若模棱两可，标记为"信息不足"并使用澄清模板

第二步：结构化分解 将请求拆解为树状结构

第三步：Skill映射 为每个子目标选择最佳匹配skill

2. 技能编排逻辑

场景	行动
单一明确意图	直接调用
多个独立意图	并行执行
多步流水线	顺序编排+交接记录
用户要求自主	分析→决策→执行→解释
技能冲突	选择描述更具体的

3. 需求澄清模板

在请求信息不足时使用四种追问模式（详见英文版）。

4. 委托给共享组件

功能	委托给	方式
持久化记忆(T1/T2/T3)	complex-memory-manager	使用其加密/解密、存储规范、清理协议
使用日志与性能追踪	self-iteration-engine	使用其日志格式、审查周期、更新矩阵
自迭代决策	self-iteration-engine	检查触发条件、运行更新决策矩阵
新技能创建建议	self-iteration-engine	使用其提案模板

依赖检查

每次编排运行前：

1. 检查 complex-memory-manager 的 metadata 是否在上下文中
2. 检查 self-iteration-engine 的 metadata 是否在上下文中
3. 若缺失，可能未安装——回退到内置基础行为

参考文件

• references/orchestration-patterns.md — 高级多skill编排示例