/deep-search — 逻辑链深度搜索

Research Board (Working Memory)

Before ANY search begins, you MUST initialize and maintain a Research Board — a structured working memory that prevents multi-round divergence and hallucination.

Three State Containers

Initialization

At the start of every investigation:

1. Parse user input → extract the Ultimate_Goal as a clear, actionable statement
2. Decompose the goal into 3-7 initial Knowledge_Gaps — the questions that must be answered
3. Set Accumulated_Facts to empty

Board Discipline

• Every fact added to Accumulated_Facts must cite its source (URL, file path, line number, or command output)
• Facts must be objective and verifiable — no speculation, no "probably", no "might be"
• Knowledge_Gaps must be re-evaluated every round — gaps may split, merge, or be marked irrelevant
• The Board is your single source of truth. Do not rely on chat history or stale context.

See references/research-board.md for formal schema and quality criteria.

The Cognitive Loop

Each round of the investigation follows a strict four-step engine cycle. You MUST execute all four steps in order. Maximum 7 rounds before forced convergence.

Step 1: Synthesize (提纯)

From the previous round's raw output (web pages, terminal logs, code files — potentially thousands of lines), extract ONLY the facts that fill a Knowledge Gap.

Rules:

• Strip marketing language, opinions, tangential information
• Each extracted fact must directly address a specific Knowledge Gap
• Update Accumulated_Facts with sourced facts
• Update Knowledge_Gaps — mark filled gaps, refine remaining ones

Step 2: Route (逻辑路由)

You are required to select exactly ONE of the five Logical Operators to bridge from this round to the next. You MUST state your choice explicitly with reasoning.

The five operators are:

You MUST output a Reasoning Block before executing the next search:

┌─ Reasoning Block ─────────────────────────────────┐
│ 🏷️ Round: [N]                                      │
│ 📋 Previous Findings: [1-2 sentence distillation]  │
│ 🔗 Logical Operator: [Operator Name + Symbol]      │
│ 💭 Reasoning: [Why this operator? What gap does     │
│    the next search target?]                        │
│ 🎯 Next Query: [Precise search query or command]   │
│ 🛠️ Tool: [Which tool to dispatch to]               │
└────────────────────────────────────────────────────┘

See references/logical-operators.md for detailed operator definitions, decision tree, and examples.

Step 3: Dispatch (跨域寻址)

Based on the Routing decision, dispatch to the appropriate tool. Choose dynamically:

Cross-surface principle: A single investigation typically spans 2-3 tool categories. The power of this Skill lies in seamlessly weaving across surfaces.

Step 4: Evaluate (终局评估)

After executing the dispatched search:

1. Check Knowledge_Gaps — are ALL gaps filled?
2. If YES → Exit loop, proceed to Artifact generation
3. If NO → Return to Step 1 with fresh results

Forced convergence rules:

• If 7 rounds reached → force convergence, report remaining gaps honestly
• If 2 consecutive rounds yield no new facts (information gain ≈ 0) → convergence (diminishing returns)
• If a gap is provably unanswerable with available tools → mark as "out of scope" and continue

Convergence & Artifact Output

When the loop exits, you MUST produce an Artifact (never dump in chat). The artifact format depends on the investigation type:

Type A: Root Cause Analysis (for debugging / troubleshooting)

# Root Cause Analysis: [Problem Title]

## Executive Summary
[2-3 sentences: what was wrong, why, and the fix]

## Root Cause
[Precise technical explanation]

## Evidence Chain
| Round | Operator | Query | Key Finding |
|-------|----------|-------|-------------|
| 1     | 🔍 Drill-down | ... | ... |
| 2     | ⚖️ Verify | ... | ... |
| ...   | ...      | ...   | ... |

## Recommended Fix
[Actionable steps or code diff]

## Remaining Risks
[Any gaps that could not be filled]

Type B: Research Report (for tech evaluation / architecture research)

# Research Report: [Topic]

## Executive Summary
[Main conclusions]

## Findings by Dimension
### [Dimension 1]
...
### [Dimension 2]
...

## Evidence Chain
[Same table format as Type A]

## Recommendation
[Actionable recommendation with rationale]

## Sources
[All references with URLs]

Type C: Comparison Matrix (for technology selection / alternatives)

# Comparison: [Options]

## Decision Matrix
| Criterion | Option A | Option B | Option C |
|-----------|----------|----------|----------|
| ...       | ...      | ...      | ...      |

## Evidence Chain
[Same table format]

## Recommendation
[Which option and why]

Anti-Patterns (What NOT to Do)

Interaction with User

Mid-Investigation Steering

If the user provides feedback during the investigation (e.g., "方向不对，去查数据库" or "Skip networking, focus on the ORM layer"):

1. Accept the steering as a logical intervention
2. Treat it as a forced Pivot or Drill-down operator
3. Incorporate into the Reasoning Block for the next round
4. Do NOT restart — continue from current Accumulated_Facts

Asking for Clarification

If Knowledge_Gaps cannot be refined without user input (ambiguous scope, missing access, etc.):

1. Use notify_user to ask specific, bounded questions (max 3)
2. Resume the loop from where you paused after receiving answers

Keywords for Automatic Detection

Entities: search, investigate, research, debug, trace, analyze, compare, evaluate, diagnose, 搜索, 调研, 排查, 追查, 分析, 对比, 诊断 Actions: dig into, trace down, find out, figure out, look into, root cause, 查一下, 搞清楚, 追溯, 定位 Qualifiers: deep, thorough, systematic, logical, chain, multi-round, 深度, 系统性, 逻辑链, 多轮 Domains: code, architecture, technology, performance, error, bug, 代码, 架构, 技术, 性能, 报错, 缺陷

Does NOT activate for:

• Simple factual questions ("What is X?") → direct answer
• Single-shot lookups ("Find the file for class Foo") → grep/search
• Writing or formatting tasks → other skills