debug

Files (4)

• SKILL.md
• failure-taxonomy.md · references/
• probe-protocol.md · references/
• report-template.md · references/

SKILL.md (first lines)

# Debug Skill

## Trigger

- Keywords: debug, 除錯, troubleshoot, diagnose, 查問題, 找 bug, 為什麼不動, 為什麼不 work, something wrong, not working

## When NOT to Use

| Scenario | Alternative |
|----------|------------|
| GitHub Issue / PR thread 分析 | `/issue-analyze` |
| 已知根因、直接修復 | `/bug-fix` |
| 理解程式碼如何運作 | `/code-explore` |
| 需要雙視角確認邏輯 | `/code-investigate` |
| 部署後唯讀功能驗證 | `/feature-verify` |

## Prohibited Actions

```
❌ git add | git commit | git push — per @rules/git-workflow.md
```

This skill debugs and may fix code but does **not** commit. `/precommit` is a quality gate only. To commit, offer the menu per `rules/git-workflow.md` § Proactive Offer — a commit option on any real branch, a push option only where `review-state.js offer` allows one — and invoke `/smart-commit --execute` on selection; never print the command for the user to copy.

## Workflow

```
Phase 0    Phase 1     Phase 2          Phase 3           Phase 4        Phase 5
Intake  →  Classify →  Probe Loop   →   Root Cause    →   Fix +       →  Report
+ Repro    (Taxonomy)  (max 6)          Confirmation      Verify
  │          │           │                 │                │              │
  ▼          ▼           ▼                 ▼                ▼              ▼
Execute    refs/       Bash            /seek-verdict    Edit+test     --export
reproduce  failure-    /codex-         /code-investigate /bug-fix      (optional)
           taxonomy    brainstorm      (optional)
```

## Phase 0: Intake + Reproduce

| Step | Action | Output |
|------|--------|--------|
| 0a | 解析問題描述（症狀、範圍、環境） | 結構化問題描述 |
| 0b | 嘗試重現（執行命令/腳本） | 預期 vs 實際結果 |
| 0c | 重現成功？ | Yes → Phase 1 / No → `⚠️ Need Human` |

**Repro Contract**: 根因聲稱必須基於可重現的觀察，不可基於靜態推測。

## Phase 1: Classify

根據觀察到的失敗模式，使用 Failure Taxonomy 分類並選擇 first-probe 策略。

See `references/failure-taxonomy.md` for:
- 6 problem types + detection signals + first-probe commands
- Classification decision tree
- Escalation paths

## Phase 2: Probe Loop

假設驅動的互動式探測迴圈。See `references/probe-protocol.md` for full rules.

每個 probe 是一個假設測試：

```
1. 定義假設 H
2. 設計探測命令 C
3. 預測「若 H 成立 → 期望 O」
4. 執行 C → 觀察實際結果
5. 更新假設集合
6. 選下一個最具鑑別力的探測
```

### Termination

| Condition | Action |
|-----------|--------|
| 根因已定位 + ≥1 執行結果佐證 | **Stop** → Phase 3 |
| ≥2 競爭假設無法區分 | **Brainstorm** → `/codex-brainstorm` 對抗辯論 |
| 連續 2 輪無新資訊 | **Escalate** → `⚠️ Need Human` |
| Max rounds (6) reached | **Escalate** → `⚠️ Need Human` |

### `/codex-brainstorm` Integration

When ≥2 equally credible hypotheses exist:
1. Pause probe loop
2. Invoke `/codex-brainstorm` via Skill tool — topic = hypothesis comparison
3. Use Nash equilibrium result as new primary hypothesis
4. Resume probe loop to verify

## Phase 3: Root Cause Confirmation

| Step | Action | Required |
|------|--------|----------|
| 3a | 總結根因（What + Why + Impact + Evidence） | ✅ |
| 3b | `/seek-verdict --intent confirm` — Codex 獨立驗證 | ✅ Mandatory |
| 3c | `/code-investigate` 雙視角深入驗證 | Optional |
| 3d | `/git-investigate` 追蹤引入點（若 regression） | Conditional |

### `/seek-verdict` Integration (Step 3b — Mandatory)

Invoke `/seek-verdict --intent confirm` for independent root cause verification.

**Anti-anchoring contract** (per `@skills/seek-verdict/SKILL.md`):
- **Fresh thread**: a new § Start dispatch, never a § Resume onto an existing thread
- **No Claude conclusions**: prompt must not contain Claude's probe findings or root cause judgment
- **Finding packet only**: provide symptoms, files, observed behavior — let Codex judge independently

**Result routing**:

| Result | Action |
|--------|--------|
| ACTIONABLE (confirm) | → Phase 4 |
| NON_ACTIONABLE (high confidence) | Re-enter Phase 2 or `⚠️ Need Human` |
| UNCERTAIN | → Phase 4 (conservative), note low confidence |

## Phase 4: Fix + Verify

| Condition | Path |
|-----------|------|
| Simple fix (≤3 lines) | Edit + regression test + verify |

From sd0xdev/sd0x-harness, MIT.

More from sd0xdev/sd0x-harness