Debate Skill

Two fully independent agents debate a proposal through iterative rounds until convergence,

then a neutral Synthesizer produces the final output.

Architecture

User Input → Auto-detect domain → Agent A (propose) → Agent B (critique)
→ Orchestrator (score + converge?) → [loop] → Synthesizer → Final Output

• Agent A (Proposer): Knows only the original input + previous version + B's last criticism
• Agent B (Critic): Knows only the original input + current version + its own last criticism
• Agents are fully context-isolated — neither knows the other exists
• Synthesizer: Sees the full debate history, produces neutral final output

Step 1 — Parse Arguments

Extract from user input:

• --verbose → show all round details (default: false)
• --rounds N → max rounds (default: 5)
• remaining text → {subject}

Step 2 — Auto-Detect Domain

Analyze {subject} and select one domain. See references/domains.md for role definitions and focus areas per domain.

Step 3 — Iterative Debate Loop

Run rounds 1 → max_rounds. Each round:

1. Agent A (independent context) — see references/prompts.md § Agent A
• Round 1: Generate initial proposal from {subject}
• Round N: Respond to B's criticism + output improved full version

1. Agent B (independent context) — see references/prompts.md § Agent B
• Critique current version; reference own last criticism to avoid repetition

1. Orchestrator scoring (inline logic, not a separate agent):
• Parse B's output for 🔴/🟡/🟢 counts
• Converge early if: 🔴 == 0 AND 🟡 ≤ 1
• Force converge if: round == max_rounds
• Otherwise: pass context to next round

1. Print round status panel (always visible, even without --verbose):

```

━━━ Round N/max ━━━ 🔴×N 🟡×N 🟢×N → [继续迭代 | 收敛 ✓]

```

Step 4 — Synthesizer

After convergence, run Synthesizer with full debate history.

See references/prompts.md § Synthesizer.

Step 5 — Output

Default (精简) mode:

━━━ 对抗分析完成：共 N 轮 ━━━

## 最终方案
[complete, ready-to-use content]

## 关键决策
| 议题 | 决策 | 依据 |
|------|------|------|

## 遗留风险
- ...

--verbose mode (prepend before final output):

## 对抗过程

### Round N
**Agent A 方案：**
[full proposal]

**Agent B 批评：**
[full criticism with 🔴🟡🟢 breakdown]