@boyingliu01/xp-gate 0.12.2 → 0.12.3

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -1,6 +1,7 @@
1
1
  ---
2
2
  name: delphi-review
3
- description: "Use when asked to review a design, plan, or architecture; before implementation starts; or when multi-expert consensus is needed. Triggers: 'review this design', '评审这个需求', 'design review', '多专家评审', 'consensus review', 'code walkthrough', 'push review', 'architecture review', 'PR review', or any request for multi-expert evaluation of requirements, design docs, or PRs."
3
+ description: "Use when asked to review a design, plan, or architecture; before implementation starts; or when multi-expert consensus is needed. See ## Triggers for trigger phrases."
4
+ auto_continue: true
4
5
  ---
5
6
 
6
7
  # Delphi Consensus Review
@@ -18,26 +19,52 @@ description: "Use when asked to review a design, plan, or architecture; before i
18
19
  - Does NOT replace testing or CI/CD verification
19
20
  - Does NOT handle deployment or release decisions
20
21
 
21
- ## 核心原则
22
-
23
- **Delphi 方法只有一个目的:得到所有专家一致认可的可行方案。**
22
+ ## Triggers
24
23
 
25
- ### 四大核心特性(RAND 方法论)
24
+ - /delphi-review
25
+ - review this design
26
+ - 评审这个需求
27
+ - 评审这个设计
28
+ - design review
29
+ - 多专家评审
30
+ - consensus review
31
+ - code walkthrough
32
+ - push review
33
+ - architecture review
34
+ - PR review
35
+
36
+ ## 工作流程
37
+
38
+ 1. Input Validation: 检查输入是否包含可评审内容(设计文档/代码/spec),空输入阻断
39
+ 2. Expert Assignment: 分配 2-3 位专家,至少来自 2 家不同厂商(国产模型)
40
+ 3. Round 1: 匿名独立评审 — 各专家互不知对方意见,独立输出 verdict JSON
41
+ 4. Consensus Check: 共识检查 — 共识 ≥90% 且全部 APPROVED 则完成
42
+ 5. Round 2: 交换意见 — 未达成共识时,专家查看他人意见后重新评审
43
+ 6. Round 3: 最终立场 — 仍未达成共识时,输出最终立场和分歧点
44
+ 7. Fix & Re-Review: REQUEST_CHANGES → 修复 Critical+Major → 从 Round 2 重新评审
45
+ 8. Generate Output: 生成共识报告 + specification.yaml + delphi-reviewed.json
46
+
47
+ ## Activation
48
+ **MANDATORY**: Every delphi-review response MUST begin with `[DelphiReview]` as the first line.
49
+ This marker is required for skill-cert L1 trigger detection.
50
+
51
+ Permitted variants (all satisfy L1 trigger):
52
+ - `[DelphiReview]` — standard entry
53
+ - `[DelphiReview:BLOCKED]` — Step 0 input validation failure
54
+ - `[DelphiReview:WARNING]` — red flag detected (reserved)
26
55
 
27
- 1. **匿名性** — Round 1 专家互不知道对方意见
28
- 2. **迭代** — 多轮直到共识,不是固定轮数
29
- 3. **受控反馈** — 每轮看到其他专家意见
30
- 4. **统计共识** — >=90% 一致才算共识
56
+ ## 核心原则
31
57
 
32
- ### 质量优先
58
+ **Delphi 方法只有一个目的:得到所有专家一致认可的可行方案。**
33
59
 
34
- | 原则 | 说明 |
60
+ | 特性 | 说明 |
35
61
  |------|------|
36
- | Token 是投资 | 相比后期修复成本,评审消耗微不足道 |
37
- | APPROVED 才是终点 | REQUEST_CHANGES 必须修复并重新评审 |
38
- | 零容忍 | Critical/Major 问题全部必须处理,不可跳过或降级 |
39
-
40
- 详见:Anti-Patterns 章节。
62
+ | **匿名性** | Round 1 专家互不知道对方意见 |
63
+ | **迭代** | 多轮直到共识,不是固定轮数 |
64
+ | **受控反馈** | 每轮看到其他专家意见 |
65
+ | **统计共识** | >=90% 一致才算共识 |
66
+ | **Token 是投资** | 相比后期修复成本,评审消耗微不足道 |
67
+ | **零容忍** | Critical/Major 全部必须处理,不可跳过 |
41
68
 
42
69
  ---
43
70
 
@@ -46,50 +73,59 @@ description: "Use when asked to review a design, plan, or architecture; before i
46
73
  | 模式 | 触发 | 用途 | 输出 |
47
74
  |------|------|------|------|
48
75
  | `design`(默认) | `/delphi-review` | 需求/设计/架构/PR 评审 | 共识报告 + specification.yaml |
49
- | `code-walkthrough` | `/delphi-review --mode code-walkthrough` | git push 前代码走查 | `.code-walkthrough-result.json` |
76
+ | `code-walkthrough` | `--mode code-walkthrough` | git push 前代码走查 | `.code-walkthrough-result.json` |
50
77
 
51
- **Code Walkthrough 模式**的完整规范已移至 `references/code-walkthrough.md`。当用户使用 `--mode code-walkthrough` 时,读取该文件并执行其中定义的全部流程。
78
+ **Code Walkthrough 模式**的完整规范 → 详见 `references/code-walkthrough.md`
52
79
 
53
80
  ---
54
81
 
55
82
  ## 参数配置
56
83
 
57
- ### 专家数量与角色
84
+ ### 专家配置
58
85
 
59
86
  | 配置 | 专家 | 适用场景 |
60
87
  |------|------|---------|
61
88
  | 2 专家(默认) | A(架构) + B(实现) | 代码变更、小型设计 |
62
- | 3 专家 | A(架构) + B(实现) + C(可行性仲裁) | 架构决策、需求文档 |
63
-
64
- > ⚠️ 至少选择 **两家不同 provider** 的模型,避免同源盲点。模型映射见 `INSTALL.md`。
65
- >
66
- > ### 模型选择策略(强制)
67
- >
68
- > **MUST 使用国产开源模型**,严禁使用国外昂贵模型(Anthropic Claude、OpenAI GPT、Google Gemini 等)。
69
- >
70
- > **允许的模型列表**(从以下国产开源模型厂家中选择):
71
- >
72
- > | 厂家 | 可选模型 |
73
- > |------|---------|
74
- > | 智谱 GLM | `glm-5.1`, `glm-5.0` |
75
- > | 月之暗面 Kimi | `kimi-k2.6`, `kimi-k2.5` |
76
- > | MiniMax | `minimax-m2.7`, `minimax-m2.5` |
77
- > | 阿里 Qwen | `qwen3.6-plus`, `qwen3.5-plus` |
78
- > | 深度求索 DeepSeek | `deepseek-v4-pro`, `deepseek-v4-lite` |
79
- >
80
- > **专家分配建议**:
81
- >
82
- > | Expert | 推荐模型 | 备选 |
83
- > |--------|---------|------|
84
- > | Expert A (架构) | `deepseek-v4-pro` | `qwen3.6-plus`, `glm-5.1` |
85
- > | Expert B (技术) | `kimi-k2.6` | `deepseek-v4-pro`, `minimax-m2.7` |
86
- > | Expert C (可行性) | `qwen3.6-plus` | `kimi-k2.6`, `glm-5.1` |
87
- >
88
- > **关键原则**:
89
- > - ✅ 三个专家必须来自 **至少 2 家不同厂家**(如 DeepSeek + Kimi + Qwen)
90
- > - ❌ 禁止使用 Anthropic、OpenAI、Google 等国外模型
91
- > - 禁止三个专家全部使用同一厂家模型
92
- > - ⚠️ 成本控制:`deepseek-v4-lite`、`kimi-k2.5`、`qwen3.5-plus` 作为轻量备选
89
+ | 3 专家 | A(架构) + B(实现) + C(可行性) | 架构决策、需求文档 |
90
+
91
+ ### 模型选择策略(强制 opencode.json 读取)
92
+
93
+ **MUST `opencode.json` 的 agent 配置中读取模型**,**严禁** hardcode 模型名称。
94
+
95
+ 模型选择流程:
96
+ 1. 读取 `opencode.json` 中 `agent` 字段下 `delphi-reviewer-architecture`、`delphi-reviewer-technical`、`delphi-reviewer-feasibility` 三个 agent 定义
97
+ 2. 提取每个 agent 的 `model` 字段(格式:`provider/model-name`)
98
+ 3. 分别作为 Expert A(架构)、Expert B(技术)、Expert C(可行性) 的模型
99
+
100
+ **关键原则**:
101
+ - 三个专家必须来自 **至少 2 家不同 provider**(通过 `model` 字段的 `provider/` 前缀判断)
102
+ - 禁止 hardcode 模型名称(模型列表以 `opencode.json` 为准)
103
+ - 禁止三个专家全部使用同一 provider 的模型
104
+
105
+ **`opencode.json` agent 配置示例**(参考 `opencode.json.delphi.example`):
106
+ ```json
107
+ {
108
+ "agent": {
109
+ "delphi-reviewer-architecture": {
110
+ "description": "...",
111
+ "mode": "subagent",
112
+ "model": "deepseek/deepseek-chat",
113
+ "tools": { "read": true, "bash": true, "write": false, "edit": false }
114
+ },
115
+ "delphi-reviewer-technical": {
116
+ "mode": "subagent",
117
+ "model": "bailian-coding-plan/qwen3.6-plus",
118
+ "tools": { "read": true, "bash": true, "write": false, "edit": false }
119
+ },
120
+ "delphi-reviewer-feasibility": {
121
+ "mode": "subagent",
122
+ "model": "deepseek/deepseek-chat",
123
+ "tools": { "read": true, "bash": true, "write": false, "edit": false }
124
+ }
125
+ }
126
+ }
127
+ ```
128
+ > 上例中 3 个专家使用了 2 个不同 provider(deepseek + bailian-coding-plan),满足跨 provider 要求。**实际模型名称以你的 `opencode.json` 配置为准,无需在 SKILL.md 维护模型列表。**
93
129
 
94
130
  ### 共识阈值
95
131
 
@@ -100,7 +136,7 @@ description: "Use when asked to review a design, plan, or architecture; before i
100
136
 
101
137
  ---
102
138
 
103
- ## Delphi 评审执行过程
139
+ ## 评审执行过程
104
140
 
105
141
  ```
106
142
  Phase 0: 准备 → Round 1: 匿名独立评审 → 共识检查
@@ -124,69 +160,84 @@ Phase 0: 准备 → Round 1: 匿名独立评审 → 共识检查
124
160
  └─ REQUEST_CHANGES → 修复方案 → 回到 Round 2 重新评审
125
161
  ```
126
162
 
127
- ---
163
+ ### Step 0: Input Validation (MANDATORY — 必须在任何评审前执行)
128
164
 
129
- ## Round 1: 匿名独立评审
165
+ 检查用户 prompt 中是否包含可评审内容(设计文档、代码、specification.yaml、PR diff):
130
166
 
131
- ### 为什么必须匿名
167
+ 1. **有完整输入** → 直接进入 Phase 0(准备阶段),开始 Round 1
168
+ 2. **有部分输入**(占位符、描述性文本)→ 按输入内容执行评审,标注 `[INPUT: PARTIAL]`,但继续执行
169
+ 3. **无输入**(仅触发词,无文档/代码)→ 输出以下阻断响应,记入步骤完成:
132
170
 
133
- 匿名防止 anchoring bias(锚定偏差)—— 知道其他专家意见后倾向于同意"权威",不敢提出相反观点。
171
+ **Detection heuristics for Step 0**:
172
+ - **Complete**: Contains ≥1 structural element (e.g., `##`, `requirement`, `AC-`, `function`, `class`, `interface`, YAML frontmatter, code block with language tag) AND ≥50 non-whitespace characters of actual content (not placeholder brackets like `[...]`, `{...}`, `<insert here>`).
173
+ - **Partial**: Contains descriptive text referencing a design/code artifact BUT lacks substantive structure (e.g., "I need to review my login module" with no actual code/design attached), OR contains obvious placeholders like `[...]`, `(content)`, `<insert here>`, `TODO`.
174
+ - **None**: Only trigger words (`/delphi-review`, "review this") with zero additional content, OR content that is exclusively questions about the review process itself ("how does delphi work?").
134
175
 
135
- ### 执行方式
176
+ **Partial input constraint**: When input is PARTIAL, cap review to 1 round with `confidence=low` annotation. Do NOT proceed with full multi-round review on insufficient input.
136
177
 
137
- 每位专家独立收到:原始文档 + 评审模板 + "独立评审,不知道其他专家意见"。
178
+ **File path validation**: If user provides a file path (e.g., `--spec specification.yaml`):
179
+ - Verify the file exists and is non-empty
180
+ - If path is invalid → output: `[DelphiReview:BLOCKED] File not found: [path]. Please verify the path.`
181
+ - If file is empty → output: `[DelphiReview:BLOCKED] File is empty: [path]. Please provide valid content.`
138
182
 
139
- ### 输出格式
183
+ ```
184
+ [DelphiReview:BLOCKED] 需要设计文档或代码内容才能启动评审。
140
185
 
141
- ```markdown
142
- ## 独立评审 - Expert [A/B/C]
143
- ### 优点
144
- 1. [具体优点 + 文档位置]
145
- ### 问题清单
146
- #### Critical Issues (必须修复才能批准)
147
- 1. [问题] - 位置: [...] - 修复建议: [...]
148
- #### Major Concerns (必须处理)
149
- 1. [...]
150
- #### Minor Concerns (需要说明)
151
- 1. [...]
152
- ### 裁决: [APPROVED / REQUEST_CHANGES / REJECTED]
153
- ### 置信度: [X/10]
154
- ### 关键理由
155
- 1. [...]
186
+ 请提供以下之一:
187
+ - 设计文档(design doc / specification)
188
+ - 代码变更(code diff / PR link)
189
+ - specification.yaml 文件路径
190
+ - 架构设计说明
191
+
192
+ 评审输入示例:
193
+ /delphi-review "Design Doc: [your content here]"
194
+ /delphi-review --spec specification.yaml
195
+ /delphi-review --mode code-walkthrough
196
+ REMAINING STEPS: N/A (input validation failed)
156
197
  ```
157
198
 
158
- ---
199
+ 在此状态下,BLOCKED 视为步骤已完成(后续步骤标记为 N/A)。
159
200
 
160
- ## Round 2: 交换意见
201
+ **重要**: 内嵌在 prompt 中的文档内容(如 "Design Doc: [content]"或代码片段)应视为"有完整输入",直接进入评审。
161
202
 
162
- ### 执行方式
203
+ **Round 模板**(匿名评审/交换意见/最终立场/修复报告格式)→ 详见 `references/round-templates.md`
163
204
 
164
- 每位专家看到:原始文档 + 其他专家的评审 + "响应其他专家的关切,是否调整立场?"
205
+ **Orchestrator 自动调度规则**(#218 subagent 内部自动多轮循环)→ 详见 `references/orchestrator-dispatch.md`
165
206
 
166
- ### 输出格式
207
+ **Automatic re-review**: 对于常见可控问题(措辞模糊、AC 缺失、格式问题),subagent 应自行修复后自动重评审,无需等待用户。
167
208
 
168
- ```markdown
169
- ## Round 2 Response - Expert [A/B/C]
170
- ### 响应其他专家关切
171
- **Expert [X] 提到: [问题]**
172
- - 我的立场: [同意/部分同意/不同意] - 理由: [...]
173
- ### 更新后问题清单 / 裁决 / 置信度 / 立场变化说明
174
- ```
209
+ ### ⭐ 自动延续规则(MANDATORY — 防止流程卡住)
175
210
 
176
- ---
211
+ 当 Delphi Review 启动多轮评审(Round 2+)时,**orchestrator 必须自动延续流程,不得等待用户输入**。
177
212
 
178
- ## Round 3: 最终立场(如需要)
213
+ **触发条件**:
214
+ - Round N 完成,但未达到终止条件(100% approved + ≥90% consensus + 所有 Critical/Major 已处理)
215
+ - 存在待处理的背景任务(subagent dispatched tasks)
179
216
 
180
- 触发条件:Round 2 后仍无共识。所有专家提交最终绑定立场。3 专家模式下若仍无完全一致,2/3 或 3/3 多数裁决生效,记录少数派意见。
217
+ **自动延续动作**:
218
+ 1. **收集背景任务结果**:等待 `<system-reminder>` 通知后,立即调用 `background_output(task_id="bg_...")` 获取所有 subagent 输出
219
+ 2. **合成 Round N 总结**:汇总专家意见、共识度、待处理问题
220
+ 3. **自动启动 Round N+1**:立即 dispatch 新一轮 subagent 任务,携带上一轮总结作为上下文
221
+ 4. **循环直至终止**:重复步骤 1-3,直到达到终止条件
181
222
 
182
- ### 输出格式
223
+ **终止条件**(满足全部):
224
+ - ✅ 所有专家状态 = APPROVED
225
+ - ✅ 共识度 ≥ 90%
226
+ - ✅ 所有 Critical 级别问题已解决
227
+ - ✅ 所有 Major 级别问题已处理(或已记录为已知问题)
183
228
 
184
- ```markdown
185
- ## Round 3 Final Position - Expert [A/B/C]
186
- ### 最终裁决: [APPROVED / REQUEST_CHANGES / REJECTED]
187
- ### 最终置信度: [X/10]
188
- ### 关键理由 + 与其他专家的差异
189
- ```
229
+ **例外情况**(直接输出,不进入下一轮):
230
+ - Round 1 即达到 100% approved + 100% consensus → 直接输出最终报告
231
+ - 已达最大轮数(5 轮)→ 输出"未达成共识"报告,标记为 PROCESS_BLOCK
232
+
233
+ **禁止行为**:
234
+ - ❌ 询问用户"要继续吗?"
235
+ - ❌ 等待用户手动触发 Round N+1
236
+ - ❌ 在未达到终止条件时停止流程
237
+
238
+ **错误处理**:
239
+ - 背景任务超时(>10min)→ 标记为 TIMEOUT,输出部分结果并终止
240
+ - 背景任务失败(subagent 错误)→ 重试 1 次,仍失败则输出错误报告并终止
190
241
 
191
242
  ---
192
243
 
@@ -197,13 +248,6 @@ Phase 0: 准备 → Round 1: 匿名独立评审 → 共识检查
197
248
  2. 重新评审(从 Round 2 起步,不是 Round 1)
198
249
  3. 迭代直到 APPROVED
199
250
 
200
- 修复报告格式:
201
- ```markdown
202
- ## 修复报告
203
- ### Critical Issues 修复 | ### Major Concerns 处理 | ### Minor Concerns 说明
204
- ### 请求重新评审
205
- ```
206
-
207
251
  ---
208
252
 
209
253
  ## 终止条件
@@ -215,90 +259,24 @@ Phase 0: 准备 → Round 1: 匿名独立评审 → 共识检查
215
259
 
216
260
  ---
217
261
 
218
- ## Triggers
219
-
220
- This skill activates on any request for multi-expert review. Common triggers:
221
-
222
- **English:**
223
- - "review this design"
224
- - "design review"
225
- - "architecture review"
226
- - "consensus review"
227
- - "code walkthrough"
228
- - "push review"
229
- - "multi-expert review"
230
- - "PR review"
231
-
232
- **Chinese:**
233
- - "评审这个需求"
234
- - "多专家评审"
235
- - "设计评审"
236
- - "架构评审"
237
- - "代码走查"
238
-
239
- **Related commands:**
240
- - `/delphi-review` - Design review mode
241
- - `/delphi-review --mode code-walkthrough` - Pre-push code walkthrough
242
-
243
- ---
244
-
245
- ## Workflow Steps
246
-
247
- 1. **Determine mode** - Design review (default) or code-walkthrough (--mode code-walkthrough)
248
- 2. **Dispatch anonymous experts** - 2-3 experts from ≥2 different domestic model providers
249
- 3. **Collect Round 1 independent reviews** - Anonymous, no cross-expert bias
250
- 4. **Synthesize feedback** - Measure consensus, identify disagreements
251
- 5. **Run Round 2+ until consensus** - Exchange opinions, iterate until ≥90% agreement
252
- 6. **Block on unresolved Critical/Major** - Zero-tolerance: all Critical/Major must be resolved
253
- 7. **Emit verdict** - APPROVED (with specification.yaml) or REQUEST_CHANGES (fix + re-review)
254
-
255
- **Consensus threshold:** ≥90% (project standard for Delphi review approval)
256
- **Model policy:** Domestic models only (DeepSeek, Qwen, Kimi, GLM, MiniMax). Foreign models (Anthropic/OpenAI/Google) forbidden.
257
-
258
- ---
259
-
260
- ## Scope
261
-
262
- **IN Scope:**
263
- - Design document review (requirements, architecture, PRDs)
264
- - Pre-implementation planning review
265
- - Code walkthrough (git push validation, max 20 files/500 LOC)
266
- - Multi-expert consensus building
267
- - Specification extraction (design → specification.yaml)
268
-
269
- **OUT Scope:**
270
- - Single-expert review (use `/review` instead)
271
- - Post-implementation review (use `/requesting-code-review`)
272
- - Security audit (use `/security-research` or `/cso`)
273
- - Performance benchmarking (use `/benchmark`)
274
-
275
- ---
276
-
277
- ## Examples
278
-
279
- **Example 1: Design review**
280
- ```bash
281
- /delphi-review
282
- ```
283
- → 3 experts review design doc → consensus report + specification.yaml
262
+ ## Output Format (MANDATORY)
284
263
 
285
- **Example 2: Code walkthrough**
286
- ```bash
287
- /delphi-review --mode code-walkthrough
288
- ```
289
- → Pre-push validation → .code-walkthrough-result.json
264
+ **Single 模式简化输出**: single reviewer 模式(非 Multi-Expert)时,可使用简化输出格式:
265
+ - verdict + confidence + issues_list(合并 critical/major/minor)+ summary
266
+ - 完整 JSON 格式保留用于 multi-expert 多轮评审场景
290
267
 
291
- **Example 3: Chinese trigger**
268
+ **Single 模式简化模板**:
292
269
  ```
293
- User: 评审这个需求文档
294
- → Auto-detects delphi-review trigger → dispatches experts
270
+ [DelphiReview] verdict=APPROVED | REQUEST_CHANGES | BLOCKED
271
+ confidence=N/10
272
+ issues=[critical: N, major: N, minor: N]
273
+ summary: [1-2 sentence verdict summary]
295
274
  ```
296
275
 
297
- ---
298
-
299
- ## Output Format (MANDATORY)
300
-
301
- Every review round output MUST follow this exact JSON structure for design mode:
276
+ **⚠️ Single vs Multi-Expert Output**:
277
+ - **Multi-Expert Mode (default)**: MUST use the full JSON schema below. Each expert outputs independently; the orchestrator aggregates into `consensus_report`. DO NOT use the simplified template.
278
+ - **Single Reviewer Mode** (explicitly invoked with `--single`): MAY use the simplified text template above.
279
+ - **Never mix formats**. If you are one of multiple experts in the same round, output JSON only.
302
280
 
303
281
  ```json
304
282
  {
@@ -321,96 +299,42 @@ Every review round output MUST follow this exact JSON structure for design mode:
321
299
 
322
300
  **For code-walkthrough mode**, output follows `.code-walkthrough-result.json` schema (see `references/code-walkthrough.md`).
323
301
 
324
- **Anti-patterns mapping to assertions:**
325
- - `Round 1 → 生成报告 → "评审完成"` → Output MUST NOT have `verdict: APPROVED` if `critical_issues` exist.
326
- - `只处理 Critical,忽略 Major` → Output MUST include `major_concerns` array, even if empty.
327
- - `用户说"时间紧急"就跳过` → Output MUST include `round` field, proving multi-round process.
302
+ **Anti-patterns mapping:**
303
+ - `Round 1 → "评审完成"` → MUST NOT have `verdict: APPROVED` if `critical_issues` exist
304
+ - `只处理 Critical,忽略 Major` → MUST include `major_concerns` array
305
+ - `用户说"时间紧急"就跳过` → MUST include `round` field proving multi-round process
328
306
 
329
307
  ---
330
308
 
331
309
  ## Terminal State Checklist
332
310
 
333
- <MANDATORY-CHECKLIST>
334
-
335
- ### 你只能在以下条件全部满足后声明"Delphi review complete":
336
-
337
- **Pre-requisites:**
338
311
  - [ ] Phase 0 完成(文档验证 + 专家分配)
339
- - [ ] Round 1 完成(所有专家匿名独立评审)
340
- - [ ] Round 2+ 完成(交换意见 / 最终立场)
341
-
342
- **CRITICAL — 共识验证:**
312
+ - [ ] Round 1-3 完成(所有专家评审)
343
313
  - [ ] 问题共识比例 >=90%
344
- - [ ] 所有 Critical Issues 已解决
345
- - [ ] 所有 Major Concerns 已处理
346
-
347
- **CRITICAL — 裁决检查:**
314
+ - [ ] 所有 Critical Issues 已解决,Major Concerns 已处理
348
315
  - [ ] 最终裁决是 **APPROVED** 或 **APPROVED_WITH_MINOR**
349
- - [ ] 如果 REQUEST_CHANGES → 已修复 → 已重新评审 → APPROVED
350
-
351
- **Final Requirements:**
352
316
  - [ ] 共识报告生成并保存
353
- - [ ] 用户已确认报告
354
- - [ ] ⭐ **IF APPROVED (design mode): 提示用户生成 specification.yaml,如用户同意则调用 specification-generator**
355
-
356
- **IF 裁决是 REQUEST_CHANGES REJECTED CANNOT claim complete, MUST 修复并重新评审**
357
- **IF 任何条件未满足 → CANNOT claim complete, MUST BLOCK 并通知用户**
358
-
359
- ### ⭐ APPROVED 后必做
317
+ - [ ] IF REQUEST_CHANGES → 已修复 → 已重新评审 → APPROVED
318
+ - [ ] ⭐ **IF APPROVED (design mode): 生成 specification.yaml**(自动或用户确认后)
319
+ - [ ] ⭐ **状态文件**: 写入 `.sprint-state/delphi-reviewed.json`(`verdict`, `consensus_ratio`, `timestamp`)
320
+ - [ ] **Code-walkthrough mode**: 写入 `.code-walkthrough-result.json`(commit hash 匹配 HEAD)
360
321
 
361
- **Design mode**:
322
+ **IF REQUEST_CHANGES/REJECTED → CANNOT claim complete**
323
+ **IF 任何条件未满足 → MUST BLOCK**
362
324
 
363
- 1. **Automatic: 生成 specification.yaml** — 从设计文档提取需求到 specification.yaml。这是 test-specification-alignment 的输入,没有它 test-spec 会进入不推荐的 legacy mode。
364
-
365
- 2. **必须提示用户** — Delphi review APPROVED 后,agent 必须主动输出以下提示:
366
-
367
- ```
368
- ⭐ Delphi review APPROVED 完成!
369
-
370
- Next Step: 生成 specification.yaml
371
-
372
- 设计文档已稳定,现在可以生成 specification.yaml 用于后续的 test-specification-alignment 验证。
373
-
374
- 是否现在生成?
375
- - 回答 "yes" 或 "generate spec" → 我将调用 specification-generator 自动生成
376
- - 回答 "no" 或 "later" → 稍后手动调用 /specification-generator
377
- ```
378
-
379
- 3. **如果用户同意** → 立即调用 `task(subagent_type="deep", load_skills=["specification-generator"], prompt="基于以下 APPROVED 设计文档生成 specification.yaml: [粘贴设计文档内容]")`
380
-
381
- 4. **生成完成后** → 将 specification.yaml 包含在提交中,标记为 "auto-generated from Delphi review consensus"
382
-
383
- **Code-walkthrough mode**: 写入 `.code-walkthrough-result.json`(commit hash 匹配 HEAD,expires = timestamp + 1小时)。详见 `references/code-walkthrough.md`。
384
-
385
- ### ⭐ 状态文件输出(MANDATORY — 防止跳过门禁)
386
-
387
- **Design mode APPROVED 后** — 必须立即写入 `.sprint-state/delphi-reviewed.json`:
325
+ ### 状态文件格式
388
326
 
327
+ **Design mode APPROVED** → `.sprint-state/delphi-reviewed.json`:
389
328
  ```json
390
- {
391
- "mode": "design",
392
- "timestamp": "2026-05-30T10:30:00Z",
393
- "verdict": "APPROVED",
394
- "consensus_ratio": 1.0,
395
- "specification_path": ".sprint-state/phase-outputs/specification.yaml"
396
- }
329
+ {"mode":"design","timestamp":"...","verdict":"APPROVED","consensus_ratio":1.0,"specification_path":".sprint-state/phase-outputs/specification.yaml"}
397
330
  ```
398
331
 
399
- **Code-walkthrough mode APPROVED 后** — 必须立即写入 `.sprint-state/delphi-reviewed.json`:
400
-
332
+ **Code-walkthrough mode APPROVED** `.sprint-state/delphi-reviewed.json`:
401
333
  ```json
402
- {
403
- "mode": "code-walkthrough",
404
- "commit": "abc123def...",
405
- "timestamp": "2026-05-30T10:30:00Z",
406
- "verdict": "APPROVED",
407
- "consensus_ratio": 1.0
408
- }
334
+ {"mode":"code-walkthrough","commit":"abc123...","timestamp":"...","verdict":"APPROVED","consensus_ratio":1.0}
409
335
  ```
410
336
 
411
- **用途**: Phase 2 BUILD 入口检查 (DELPHI-GATE) 读取此文件。`verdict != "APPROVED"` → 禁止编码。
412
-
413
- </MANDATORY-CHECKLIST>
337
+ > Phase 2 BUILD 入口检查 (DELPHI-GATE) 读取此文件。`verdict != "APPROVED"` → 禁止编码。
414
338
 
415
339
  ---
416
340
 
@@ -418,20 +342,32 @@ Every review round output MUST follow this exact JSON structure for design mode:
418
342
 
419
343
  | ❌ 错误 | ✅ 正确 |
420
344
  |---------|---------|
421
- | Round 1 生成报告 "评审完成"(未 APPROVED) | 迭代直到 APPROVED,修复后重新评审 |
422
- | 只处理 Critical,忽略 Major | 零容忍:Critical/Major 全部必须处理,不可跳过或降级 |
345
+ | Round 1 APPROVED "评审完成" | 迭代直到 APPROVED,修复后重新评审 |
346
+ | 只处理 Critical,忽略 Major | 零容忍:Critical/Major 全部必须处理 |
423
347
  | 单专家自评 | 至少 2 位不同 provider 的专家 |
424
- | 用户说"时间紧急"就跳过 | 评审是投资不是开销,跳过后期返工成本更高 |
348
+ | 用户说"时间紧急"就跳过 | 评审是投资不是开销 |
425
349
  | "专家几乎一致"就通过 | "几乎" = 不一致,继续到 >=90% |
426
- | 使用 Anthropic/GPT/Gemini 等国外昂贵模型 | 必须使用国产开源模型(DeepSeek, Qwen, Kimi, GLM, MiniMax) |
427
- | 三个专家使用同一厂家模型 | 必须来自至少 2 家不同厂家 |
350
+ | 使用 Anthropic/GPT/Gemini | 必须使用国产开源模型 |
351
+ | 三个专家同一厂家 | 必须来自至少 2 家不同厂家 |
428
352
 
429
- **Code-walkthrough 专属 Anti-Patterns**: 详见 `references/code-walkthrough.md`。
353
+ **Code-walkthrough 专属 Anti-Patterns** 详见 `references/code-walkthrough.md`
430
354
 
431
355
  ---
432
356
 
433
357
  ## Red Flags
434
358
 
359
+ ### 检测触发器(模型可执行检测)
360
+
361
+ | 用户输入模式 | 触发词 | 响应动作 |
362
+ |-------------|--------|---------|
363
+ | 要求跳过评审 | "skip review", "不用评审", "跳过评审", "直接提交", "不评审" | → 提醒: `[DelphiReview] 评审是投资而非开销。Delphi 设计要求多轮共识(>=90%),不可快速跳过。` |
364
+ | 时间压力 | "来不及", "时间紧", "emergency", "赶时间", "deadline" | → 提醒: `[DelphiReview] 时间紧迫正是需要评审的时刻。跳过评审省 30 分钟,后期修复可能花 3 天。` |
365
+ | 提前终止 | Round 1 后用户说 "可以了", "够了", "enough" | → BLOCK: `[DelphiReview:BLOCKED] 评审未达终止条件。仍需 [共识>=90% + 所有 Critical/Major 已处理]。` |
366
+ | 单专家自评 | 用户仅指定 1 个专家 或 说 "我自己看了" | → 提醒: `[DelphiReview] 至少需要 2 位不同 provider 的专家参与评审。` |
367
+ | 无文档输入 | 仅触发词,无设计/代码内容 | → 输出 `[DelphiReview:BLOCKED]` 阻断响应(见 Step 0) |
368
+
369
+ ### 原则性声明
370
+
435
371
  | 借口 | 现实 |
436
372
  |------|------|
437
373
  | "这只是小变更" | 所有变更都需要评审 |
@@ -443,12 +379,10 @@ Every review round output MUST follow this exact JSON structure for design mode:
443
379
 
444
380
  ## 成功标准
445
381
 
446
- **Delphi 评审完成的唯一标准:**
447
382
  1. ✅ 所有专家裁决 APPROVED
448
383
  2. ✅ 问题共识 >=90%
449
- 3. ✅ 所有 Critical Issues 已修复验证
450
- 4. ✅ 所有 Major Concerns 已处理
451
- 5. ✅ 共识报告已生成
452
- 6. ✅ 用户已确认
384
+ 3. ✅ 所有 Critical Issues 已修复验证 + Major Concerns 已处理
385
+ 4. ✅ 共识报告已生成,用户已确认
386
+ 5. ✅ 状态文件已写入
453
387
 
454
388
  **缺少任何一项 = 未完成**
@@ -1,9 +1,9 @@
1
1
  # SKILLS/SPRINT-FLOW KNOWLEDGE BASE
2
2
 
3
- **Generated:** 2026-07-01
4
- **Commit:** 58ec6df
3
+ **Generated:** 2026-07-03
4
+ **Commit:** 83b39b6
5
5
  **Branch:** main
6
- **Version:** 0.12.2.0
6
+ **Version:** 0.12.3.0
7
7
 
8
8
  ## OVERVIEW
9
9
  **11-phase** development pipeline: ISOLATE → AUTO-ESTIMATE → THINK → PLAN → BUILD → REVIEW → USER ACCEPTANCE → FEEDBACK → SHIP → LAND → CLEANUP. Phase 2 default build mode is **ralph-loop** (REQ-level iteration, 40-67% token savings vs parallel). HARD-GATE in Phase 1: design must pass Delphi review (≥90% consensus) before any coding.
@@ -1,9 +1,9 @@
1
1
  # SKILLS/TEST-SPECIFICATION-ALIGNMENT KNOWLEDGE BASE
2
2
 
3
- **Generated:** 2026-07-01
4
- **Commit:** 58ec6df
3
+ **Generated:** 2026-07-03
4
+ **Commit:** 83b39b6
5
5
  **Branch:** main
6
- **Version:** 0.12.2.0
6
+ **Version:** 0.12.3.0
7
7
 
8
8
  ## OVERVIEW
9
9
  Test-Specification Alignment Engine — two-stage validation ensuring tests accurately reflect requirements and design specs.
@@ -1,9 +1,9 @@
1
1
  # PRINCIPLES CHECKER MODULE
2
2
 
3
- **Generated:** 2026-07-01
4
- **Commit:** 58ec6df
3
+ **Generated:** 2026-07-03
4
+ **Commit:** 83b39b6
5
5
  **Branch:** main
6
- **Version:** 0.12.2.0
6
+ **Version:** 0.12.3.0
7
7
 
8
8
  ## OVERVIEW
9
9
  Clean Code & SOLID principles checker — **Gate 4** of pre-commit. 14 rules × 9 language adapters, SARIF 2.1.0 output. Houses the **Boy Scout Rule** enforcement engine (Gate 6) and warning-baseline storage.