Before You Fix: Stop and Think

Read this BEFORE making any code changes.

The Rule

落ち着け。
じっくり考えて、one by one で確かめろ。
深掘りしろ。

Why This Exists

Symptom-driven fixes are wrong fixes. Seeing an error and immediately "fixing" it without understanding WHY the error occurs leads to:

• •Breaking the design
• •Introducing new bugs
• •Wasting time on wrong solutions

Checklist (Do Not Skip)

1. Stop

Do NOT write code yet. Do NOT edit files yet.

2. Ask "Why?"

The error is a symptom. Ask:

• •Why is this error occurring?
• •What is the system trying to do?
• •What did the system expect vs what happened?

3. Read the Design

Before changing ANY code:

• •Find the relevant design document
• •Read it completely
• •Understand the intended behavior

Common locations:

• •agents/docs/design/ - Agent architecture
• •docs/ - Project documentation
• •Schema files - Structural contracts

4. Trace the Flow

Follow the execution path:

1. •What triggered this code?
2. •What state was expected?
3. •Where did it diverge?

5. Identify Root Cause

The fix location is often NOT where the error appears.

6. Verify Your Understanding

Before fixing, state:

• •What is the root cause?
• •Why does the design work this way?
• •What is the minimal correct fix?

7. Then Fix

Only now: make the smallest change that addresses the root cause.

Anti-Patterns

Bad: "Error says X not in Y, so add X to Y"

Good: "Error says X not in Y. Why is code producing X? Is X correct? What does design say about X?"

Example

Error: Intent 'handoff' not in allowedIntents

Bad fix: Add 'handoff' to allowedIntents
  → Doesn't ask WHY handoff is being returned

Good process:
  1. Why is agent returning 'handoff'?
  2. Check design: initial.* steps should NOT use handoff
  3. Root cause: prompt tells agent to use handoff
  4. Correct fix: update prompt to use 'next' instead

Investigation Output: tmp/ Directory

調査・思考結果は tmp/ 配下に階層を作り整理する。

Structure

tmp/
└── investigation/
    └── <issue-name>/
        ├── overview.md      # 問題の全容（mermaid図必須）
        ├── trace.md         # 実行フロー追跡
        └── root-cause.md    # 根本原因と修正方針

Rules

1. •端的なmermaid図で整理する - 文章より図で構造を示す
2. •1ファイル500行以内 - 超えたら要約して分割
3. •問題の全容と本質的な構造を見極める

Mermaid Example

## Problem Structure

\`\`\`mermaid flowchart TD A[Error: X not found] --> B{Where?} B -->
C[Validation layer] B --> D[Schema definition] C --> E[Symptom] D --> F[Root
Cause] F --> G[Fix: Update schema] \`\`\`

When to Use

• •複雑な問題で即座に原因が分からない時
• •複数のコンポーネントが絡む時
• •後で振り返りが必要になりそうな時

Output Quality