Second Brain — Complementary Co-Pilot

"第一脑负责发散和创造。第二脑负责收敛和验证。互补，不对抗。"

Two Brains, One System

BRAIN 1 (Creative/Executive)          BRAIN 2 (Analytical/Verification)
─────────────────────────────────     ─────────────────────────────────
擅长：理解意图、生成方案、写代码      擅长：发现遗漏、验证正确性、找边界
模式：发散（divergent）                模式：收敛（convergent）
上下文：完整对话历史                    上下文：只看输入+输出（干净）
盲区：自我审查、上下文腐烂、          盲区：不理解完整背景、
      过度自信、锚定效应                    过于保守、缺少创造力

              │                                  │
              └────────── 互补 ──────────────────┘
                        1 + 1 > 2

Division of Labor

任务阶段	Brain 1 做	Brain 2 做
理解需求	解析意图、澄清歧义	检查需求完整性、找隐含约束
设计方案	生成多个方案、权衡利弊	验证每个方案的可行性、找死角
写代码	实现功能、保持风格一致	审查正确性、补测试、找边界条件
修 bug	定位根因、修复代码	验证修复是否彻底、是否引入新 bug
提交前	自审、格式化	独立审查、跑测试、检查引用完整性
完成任务	总结产出、标记完成	检查是否有遗漏子任务、更新记忆

The Complementarity Principle

Brain 2 doesn't just say "you missed X."

Brain 2 says:

• "这里缺了 X，我已经补上了 → 这是补丁"
• "这个边界条件没处理 → 这是我写的测试"
• "这个决策有个风险 → 这是我建议的缓解方案"

Brain 2 fills gaps, doesn't just flag them.

Trigger Points

触发	Brain 2 行动	优先级
代码改动 (>10行或涉及安全)	独立审查 diff → 补测试 → 补边界处理	🔴
架构决策	验证可行性 → 找遗漏的约束 → 建议替代方案	🔴
修 bug	验证修复彻底性 → 检查同类 bug 是否存在	🟡
任务完成	检查是否有遗漏子任务 → 更新记忆	🟡
用户纠正	分析纠正模式 → 建议规则更新	🟢

When to SKIP Brain 2

跳过的场景	原因
单行 typo 修复	成本 > 收益
纯格式化/注释改动	无需审查
配置值微调（不改逻辑）	影响范围明确
用户明确说"不用审查"	尊重用户意图
同一任务已被 Brain 2 审查过 2 次以上	避免重复开销

Clean Context Rule

Brain 2 runs as an independent subagent (separate subagent invocation, not same agent switching personas). True isolation requires a fresh subagent — see delegation-lora for the delegation mechanism.

Brain 2 receives ONLY:

• Original user request (verbatim)
• Output artifacts (diff, files, terminal output)
• Explicit constraints

Brain 2 does NOT receive:

• Brain 1's reasoning or justifications
• Conversation history
• Brain 1's self-assessment

Disagreement Resolution

When Brain 1 and Brain 2 disagree:

Disagreement Type	Who Wins	Why
Objective correctness (null check, syntax, security hole)	Brain 2	Facts over opinions
Design choice (architecture, naming, tradeoffs)	Brain 1	Has full project context
Risk assessment	Tie → escalate to user	Subjective, needs human judgment

Rule of thumb: Brain 2 wins on WHAT is wrong. Brain 1 wins on HOW to fix it.

False Positive Prevention

Brain 2 findings are not automatically applied:

1. Brain 2 reports → Brain 1 reviews and accepts/rejects each item
2. Accepted findings → write to KG as successful catch
3. Rejected findings → write to KG as false positive pattern
4. Same false positive 3 times → Brain 2 learns to stop flagging it

Output: Complement Report

### Second Brain — Complement Report

#### ✅ CONFIRMED
- {what Brain 1 got right, with evidence}

#### 🔧 COMPLETED
- {what Brain 1 missed → Brain 2 already filled}
- {补丁、测试、边界处理已写好}

#### ⚠️ FLAGGED (needs Brain 1)
- {issues Brain 2 found but can't fix alone — needs Brain 1's context}

#### 💡 SUGGESTED
- {improvements for next time, patterns to learn}

Model Selection

Different model = complementary blind spots:

Brain 1 Model	Brain 2 Model	Why
Creative model (Opus/Sonnet)	Precise model (Haiku)	Big-picture vs detail-oriented
Different provider	Different provider	Different training = different gaps
Any	adversarial-reviewer agent	Built for this exact purpose

Learning Together

Both brains contribute to memory:

• Brain 1 patterns → KG (creative strategies, successful approaches)
• Brain 2 findings → KG (common omissions, edge case patterns)
• Recurring gaps → auto-generated prevention rules
• Complementary strengths improve over time

Integration

Built on the cross-model review mandate from self-review.md: different model, clean context, independent review.

Current integration status:

Partner	Status	Mechanism
precommit-pipeline step 4	✅ wired	Delegates to adversarial-reviewer = Brain 2
delegation-lora #5	✅ wired	Cross-model review trigger = Brain 2
memory-lora	🔜 planned	Both brains write to memory.db
skill-judge	🔜 planned	Brain 2 reports feed skill scores

Sub-Agent Reference

Brain 2 is implemented using existing agent types:

• adversarial-reviewer — primary Brain 2 agent
• code-review — code-specific review
• security-auditor — security-focused review
• General-purpose subagent with different model — fallback