debug-council（调试委员会）

Debug Council: Research-Aligned Self-Consistency

Pure implementation of self-consistency (Wang et al., 2022). Each agent receives the raw user prompt and explores/debugs independently. No pre-processing, no shared context. Majority voting selects the answer.

Use this for bugs and problems with ONE correct answer.

Step 0: Ask User How Many Agents

Before doing anything else, ask the user how many solver agents to use:

How many debug agents would you like me to use? (3-10)

Recommendations:
- 3 agents: Faster, still reliable
- 5 agents: Good balance
- 7 agents: High confidence
- 10 agents: Maximum confidence (critical bugs)

Note: Each agent will independently explore the codebase and find the bug.
This takes longer but provides true independence per the research.

Wait for the user's response. If they specified a number (e.g., "debug council of 5"), use that.

Minimum: 3 agents | Maximum: 10 agents

CRITICAL: Pure Research Alignment

What This Means

1. NO orchestrator exploration - Do NOT read files or gather context before spawning agents
2. Raw user prompt to all agents - Each agent gets the user's original request, unchanged
3. Each agent explores independently - Agents discover the codebase themselves
4. True independence - No shared context, no cross-contamination

Why This Matters

The research shows that independent samples converge on correct answers. If we pre-process or share context, we:

• Introduce orchestrator bias
• Reduce independence
• May miss what individual agents would discover

Workflow

Step 1: Capture the Raw User Prompt

Take the user's request exactly as stated. Do NOT:

• ❌ Read files first
• ❌ Explore the codebase
• ❌ Add context
• ❌ Rephrase or enhance the prompt

Just capture what the user said.

Step 2: Spawn Agents IN PARALLEL with RAW PROMPT

Spawn ALL agents simultaneously. Each gets the exact same raw prompt:

Task(agent: "debug-solver-1", prompt: "[USER'S EXACT WORDS]")
Task(agent: "debug-solver-2", prompt: "[USER'S EXACT WORDS]")
Task(agent: "debug-solver-3", prompt: "[USER'S EXACT WORDS]")
... (all in the SAME batch - parallel execution)

DO NOT modify the prompt. DO NOT add context. Raw user words only.

Step 3: Agents Work Independently

Each agent will:

1. Read and understand the user's request
2. Explore the codebase using their tools (Read, Grep, Glob, LS)
3. Identify the root cause
4. Reason through solutions (chain-of-thought)
5. Generate a complete fix

Each agent works in complete isolation - they cannot see what other agents are doing or have found.

Step 4: Track Progress & Collect Solutions

As agents complete, show progress to the user:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
                     AGENT PROGRESS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
☑ Agent 1 - Complete
☑ Agent 2 - Complete
☑ Agent 3 - Complete
☐ Agent 4 - Working...
☐ Agent 5 - Working...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Update this display as each agent finishes. When all complete:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
                     AGENT PROGRESS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
☑ Agent 1 - Complete ✓
☑ Agent 2 - Complete ✓
☑ Agent 3 - Complete ✓
☑ Agent 4 - Complete ✓
☑ Agent 5 - Complete ✓

All agents finished! Analyzing solutions...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Collect all outputs for voting.

Step 5: Majority Voting

Group solutions by their core approach/answer:

1. Identify the key decision in each solution
2. Group solutions that make the same key decision
3. Count how many agents chose each approach

Voting rules:

• Clear majority (≥50%): Select that solution, HIGH confidence
• Plurality (highest < 50%): Select that solution, MEDIUM confidence
• No clear winner: Analyze disagreement, LOW confidence

Step 6: Implement the Winner

Implement the majority solution. Do NOT synthesize or merge - use the winning answer as-is.

Step 7: Report Results

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
                    DEBUG COUNCIL RESULTS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## 📊 Voting Summary

| Approach | Description | Agents | Votes |
|----------|-------------|--------|-------|
| ✅ A | [description] | 1, 2, 4, 5, 7 | **5/7** |
| B | [description] | 3, 6 | 2/7 |

**Winner: Approach A** (71% consensus)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## 🔍 What Each Agent Found

### Agent 1
- Files explored: [list]
- Root cause identified: [summary]
- Solution: [brief]

### Agent 2
- Files explored: [list]
- Root cause identified: [summary]
- Solution: [brief]

... (for each agent)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## 🧠 Reasoning Highlights

### Why majority chose Approach A:
- Agent 1: "[key insight]"
- Agent 2: "[key insight]"
- Agent 4: "[key insight]"

### Why minority chose differently:
- Agent 3: "[different perspective]"

### Valuable minority insight:
[Any good ideas from minority that might be worth noting]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## 📈 Confidence: HIGH/MEDIUM/LOW

[Explanation based on voting distribution and reasoning quality]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## ✅ Selected Solution

[The complete winning solution]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## 🔧 Implementation

[The actual code change being made]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Configuration

If user just says debug council, ask them to choose.

Research Basis

Based on "Self-Consistency Improves Chain of Thought Reasoning in Language Models" (Wang et al., 2022):

Why This is Slower (And Why That's OK)

Each agent independently:

• Explores the codebase
• Reads relevant files
• Reasons through the problem
• Generates a solution

This takes 3-10x longer than shared-context approaches, but provides:

• True independence - no orchestrator bias
• Diverse exploration - agents may find different things
• Research alignment - matches the paper exactly
• Maximum reliability - for when accuracy matters most

Use this for critical problems where getting it right matters more than getting it fast.

Agents

10 identical debug solver agents in agents/ directory:

• debug-solver-1 through debug-solver-10

All agents:

• Same instructions
• Same temperature (0.7)
• Same tools (Read, Grep, Glob, LS)
• Use ultrathink (extended thinking)
• Focus on finding the ONE correct answer

Diversity comes from sampling randomness and independent exploration, not different prompts.