@boyingliu01/xp-gate 0.5.3 → 0.6.1

package/skills/delphi-review/SKILL.md

@@ -1,10 +1,23 @@
 ---
 name: delphi-review
-description: "Delphi consensus review: multi-round anonymous expert review until unanimous APPROVAL. Supports design/code-walkthrough modes. 2-3 experts from different providers. MANDATORY before implementation, design, or architecture decisions. Trigger: 'review this design', '评审这个需求', 'design review', '多专家评审', 'consensus review', 'code walkthrough', 'push review', or any request for multi-expert review of requirements, design docs, architecture, or PRs."
+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."
 ---
 
 # Delphi Consensus Review
 
+## Scope
+
+**In Scope:**
+- Multi-round anonymous expert consensus review (design + code-walkthrough modes)
+- 2-3 experts from different providers with statistical consensus (>= 91%)
+- Structured verdict: APPROVED / PASS_WITH_CAVEATS / REQUEST_CHANGES
+- Domestic models only (no Anthropic/OpenAI/Google)
+
+**Out of Scope:**
+- Does NOT implement code changes (review only, implementation is separate)
+- Does NOT replace testing or CI/CD verification
+- Does NOT handle deployment or release decisions
+
 ## 核心原则
 
 **Delphi 方法只有一个目的：得到所有专家一致认可的可行方案。**
@@ -14,7 +27,7 @@ description: "Delphi consensus review: multi-round anonymous expert review until
 1. **匿名性** — Round 1 专家互不知道对方意见
 2. **迭代** — 多轮直到共识，不是固定轮数
 3. **受控反馈** — 每轮看到其他专家意见
-4. **统计共识** — >=95% 一致才算共识
+4. **统计共识** — >=91% 一致才算共识
 
 ### 质量优先
 
@@ -82,24 +95,24 @@ description: "Delphi consensus review: multi-round anonymous expert review until
 
 | 阈值 | 说明 |
 |------|------|
-| **>=95%** | 推荐默认 |
+| **>=91%** | 推荐默认 |
 | 100% | 完全一致（更严格） |
 
 ---
 
-## 完整流程
+## Delphi 评审执行过程
 
 ```
 Phase 0: 准备 → Round 1: 匿名独立评审 → 共识检查
     │
-    ├─ 一致 + >=95% + APPROVED → ✅ 完成
+    ├─ 一致 + >=91% + APPROVED → ✅ 完成
     │
-    └─ 不一致 或 <95% 或 REQUEST_CHANGES
+    └─ 不一致 或 <91% 或 REQUEST_CHANGES
           │
           ▼
        Round 2: 交换意见 → 共识检查
           │
-          ├─ 一致 + >=95% + APPROVED → ✅ 完成
+          ├─ 一致 + >=91% + APPROVED → ✅ 完成
           │
           └─ 仍分歧 或 REQUEST_CHANGES
                 │
@@ -202,14 +215,96 @@ Phase 0: 准备 → Round 1: 匿名独立评审 → 共识检查
 
 ---
 
+## Triggers
+
+This skill activates on any request for multi-expert review. Common triggers:
+
+**English:**
+- "review this design"
+- "design review"
+- "architecture review"
+- "consensus review"
+- "code walkthrough"
+- "push review"
+- "multi-expert review"
+- "PR review"
+
+**Chinese:**
+- "评审这个需求"
+- "多专家评审"
+- "设计评审"
+- "架构评审"
+- "代码走查"
+
+**Related commands:**
+- `/delphi-review` - Design review mode
+- `/delphi-review --mode code-walkthrough` - Pre-push code walkthrough
+
+---
+
+## Workflow Steps
+
+1. **Determine mode** - Design review (default) or code-walkthrough (--mode code-walkthrough)
+2. **Dispatch anonymous experts** - 2-3 experts from ≥2 different domestic model providers
+3. **Collect Round 1 independent reviews** - Anonymous, no cross-expert bias
+4. **Synthesize feedback** - Measure consensus, identify disagreements
+5. **Run Round 2+ until consensus** - Exchange opinions, iterate until ≥91% agreement
+6. **Block on unresolved Critical/Major** - Zero-tolerance: all Critical/Major must be resolved
+7. **Emit verdict** - APPROVED (with specification.yaml) or REQUEST_CHANGES (fix + re-review)
+
+**Consensus threshold:** ≥91% (project standard for Delphi review approval)
+**Model policy:** Domestic models only (DeepSeek, Qwen, Kimi, GLM, MiniMax). Foreign models (Anthropic/OpenAI/Google) forbidden.
+
+---
+
+## Scope
+
+**IN Scope:**
+- Design document review (requirements, architecture, PRDs)
+- Pre-implementation planning review
+- Code walkthrough (git push validation, max 20 files/500 LOC)
+- Multi-expert consensus building
+- Specification extraction (design → specification.yaml)
+
+**OUT Scope:**
+- Single-expert review (use `/review` instead)
+- Post-implementation review (use `/requesting-code-review`)
+- Security audit (use `/security-research` or `/cso`)
+- Performance benchmarking (use `/benchmark`)
+
+---
+
+## Examples
+
+**Example 1: Design review**
+```bash
+/delphi-review
+```
+→ 3 experts review design doc → consensus report + specification.yaml
+
+**Example 2: Code walkthrough**
+```bash
+/delphi-review --mode code-walkthrough
+```
+→ Pre-push validation → .code-walkthrough-result.json
+
+**Example 3: Chinese trigger**
+```
+User: 评审这个需求文档
+→ Auto-detects delphi-review trigger → dispatches experts
+```
+
+---
+
 ## Output Format (MANDATORY)
-Every review round output MUST follow this exact JSON structure:
+
+Every review round output MUST follow this exact JSON structure for design mode:
 
 ```json
 {
   "expert_id": "A|B|C",
   "round": 1,
-  "mode": "design|code-walkthrough",
+  "mode": "design",
   "verdict": "APPROVED|REQUEST_CHANGES|REJECTED",
   "confidence": 9,
   "critical_issues": ["..."],
@@ -218,11 +313,14 @@ Every review round output MUST follow this exact JSON structure:
   "consensus_report": {
     "agreed_items": ["..."],
     "disagreed_items": ["..."],
-    "final_verdict": "APPROVED|REQUEST_CHANGES"
+    "final_verdict": "APPROVED|REQUEST_CHANGES",
+    "consensus_ratio": 0.95
   }
 }
 ```
 
+**For code-walkthrough mode**, output follows `.code-walkthrough-result.json` schema (see `references/code-walkthrough.md`).
+
 **Anti-patterns mapping to assertions:**
 - `Round 1 → 生成报告 → "评审完成"` → Output MUST NOT have `verdict: APPROVED` if `critical_issues` exist.
 - `只处理 Critical，忽略 Major` → Output MUST include `major_concerns` array, even if empty.
@@ -242,7 +340,7 @@ Every review round output MUST follow this exact JSON structure:
 - [ ] Round 2+ 完成（交换意见 / 最终立场）
 
 **CRITICAL — 共识验证:**
-- [ ] 问题共识比例 >=95%
+- [ ] 问题共识比例 >=91%
 - [ ] 所有 Critical Issues 已解决
 - [ ] 所有 Major Concerns 已处理
 
@@ -324,7 +422,7 @@ Every review round output MUST follow this exact JSON structure:
 | 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 |
 | 单专家自评 | 至少 2 位不同 provider 的专家 |
 | 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 |
-| "专家几乎一致"就通过 | "几乎" = 不一致，继续到 >=95% |
+| "专家几乎一致"就通过 | "几乎" = 不一致，继续到 >=91% |
 | 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） |
 | 三个专家使用同一厂家模型 | 必须来自至少 2 家不同厂家 |
 
@@ -339,7 +437,7 @@ Every review round output MUST follow this exact JSON structure:
 | "这只是小变更" | 所有变更都需要评审 |
 | "Round 1 就够了" | 不够，必须多轮直到共识 |
 | "生成报告就完成了" | APPROVED 才算完成 |
-| "2/3 同意就是共识" | 还要检查问题共识比例 >=95% |
+| "2/3 同意就是共识" | 还要检查问题共识比例 >=91% |
 
 ---
 
@@ -347,25 +445,10 @@ Every review round output MUST follow this exact JSON structure:
 
 **Delphi 评审完成的唯一标准：**
 1. ✅ 所有专家裁决 APPROVED
-2. ✅ 问题共识 >=95%
+2. ✅ 问题共识 >=91%
 3. ✅ 所有 Critical Issues 已修复验证
 4. ✅ 所有 Major Concerns 已处理
 5. ✅ 共识报告已生成
 6. ✅ 用户已确认
 
 **缺少任何一项 = 未完成**
-## Output Format (MANDATORY)
-Every delphi review round MUST output valid JSON:
-```json
-{
-  "skill_name": "delphi-review",
-  "mode": "design|code-walkthrough",
-  "phase": "Round 1|Round 2|Round 3|Consensus",
-  "expert_id": "A|B|C",
-  "verdict": "APPROVED|REQUEST_CHANGES|REJECTED",
-  "confidence": 8,
-  "issues": [{"id": "string", "severity": "critical|major|minor", "description": "string"}],
-  "consensus_report": {"status": "pending|consensus|disagreement"}
-}
-```
-**Eval assertions check for:** `verdict` enum values, `confidence` range, `issues` structure, `consensus_report.status`.

package/skills/ralph-loop/SKILL.md

@@ -1,13 +1,6 @@
 ---
 name: ralph-loop
-description: >
-  Default Phase 2 BUILD mode for Sprint-Flow. Processes ONE REQ at a time from
-  specification.yaml with clean isolated context per subagent dispatch, persists
-  memory via git history + classified learnings, runs full regression tests on each
-  REQ, and only commits on verification pass. Designed to reduce token consumption
-  from linear context accumulation to per-REQ fixed budgets (40-67% savings).
-  Token-constraint is the default — every Phase 2 uses this mode automatically.
-  Use `--mode parallel` to opt into the legacy all-at-once mode if needed.
+description: Use when executing Sprint-Flow Phase 2 BUILD, processing one REQ, building the next requirement, iterating with clean context, or requiring full regression per requirement.
 maturity: stable
 ---
 
@@ -25,6 +18,25 @@ maturity: stable
 | **浓缩型分类学习** | permanent（架构级）+ contextual（最近 3 条）双层 learnings |
 | **AGENTS.md 统一更新** | orchestrator 统一写，subagent 不直接修改 — 无竞态 |
 
+---
+
+## Triggers
+
+**关键词触发**：ralph-loop, process one REQ, build next requirement, Phase 2 BUILD, iterate REQ, isolated REQ build, clean context, full regression
+
+**使用场景**：
+- 执行 Sprint-Flow Phase 2 BUILD 阶段
+- 从 specification.yaml 读取需求逐个实现
+- 需要控制每个 REQ 的 token 预算
+- 防止上下文污染（前一个 REQ 的实现细节影响后一个 REQ）
+- 需要全量回归测试保证不引入回归
+- 需要持久化 learnings 供后续 REQ 参考
+
+**何时不使用**：
+- 需求数量极少（1-2 个）且相互独立，可以并行
+- 紧急热修复，不需要完整 Sprint Flow
+- 明确使用 `--mode parallel` 切换到并行模式
+
 ## 作为 Phase 2 默认行为
 
 Ralph Loop 是 Sprint-Flow **Phase 2 BUILD 的默认模式**：
@@ -297,6 +309,165 @@ Phase 0, 1, 3-6 行为完全不变。
 
 ---
 
+## Workflow Steps (Mandatory Execution Order)
+
+**每个 REQ 必须按以下顺序执行，不得跳过或重排**：
+
+1. **Load next READY REQ**
+   - 从 specification.yaml 读取下一个依赖已满足的 REQ
+   - 检查 priority 和 status (pending 状态)
+   - 构建依赖图，确认无循环依赖
+
+2. **Create isolated context**
+   - 清空前一个 REQ 的对话历史
+   - 注入：当前 REQ + AC + permanent learnings + contextual learnings (最近 3 条)
+   - 注入：AGENTS.md + git log --oneline -5 + 测试基础设施摘要
+
+3. **Dispatch build subagent**
+   - 使用 `task(category="unspecified-high", load_skills=["test-driven-development"], timeout=300)`
+   - 强制 TDD 流程：RED → GREEN → REFACTOR
+   - 强制执行 Mock 边界策略（详见 Mock 边界表）
+
+4. **Run full regression tests**
+   - L1: typecheck + lint on changed files
+   - L1b: 检查测试先行比率 (新增测试行数 / 总新增行数 ≥ 40%)
+   - **L2: 全量测试运行**（不只是 @test 当前 REQ 的测试）
+   - L3: 检查整体覆盖率 ≥ 80%
+
+5. **Fix until green or block**
+   - 失败 → retry (max 3 次，每次注入失败摘要)
+   - 第 3 次仍失败 → BLOCK → 等待用户决策 (skip/manual/stop/rollback)
+   - 通过 → git commit + 标记 done
+
+6. **Persist learnings**
+   - 分类为 permanent（架构级）或 contextual（临时性）
+   - 自动升级条件：被≥2 个 REQ 引用 或 涉及接口/数据结构
+   - 调用 `gstack/learn` 总结经验教训
+
+7. **Update sprint state**
+   - orchestrator 统一更新 AGENTS.md（append `## ralph-loop: [REQ-XXX title]`）
+   - 原子写 checkpoint (progress.log)
+   - 继续下一个 READY REQ
+
+**禁止行为**：
+- ❌ 跳过测试基础设施检查
+- ❌ 只运行部分测试（必须全量回归）
+- ❌ 验证失败仍 commit
+- ❌ 多个 subagent 同时写 AGENTS.md
+- ❌ 修改前一个 REQ 的代码（除非修复回归）
+
+---
+
+## Scope
+
+**In Scope**：
+- Phase 2 BUILD 阶段的 REQ 级迭代构建
+- 从 specification.yaml 读取需求并逐个实现
+- 测试先行（TDD）+ 全量回归测试
+- 持久化 learnings 供后续 REQ 参考
+- 依赖排序和循环依赖检测
+- 崩溃恢复（从 checkpoint 继续）
+
+**Out of Scope**：
+- Phase 0-1 (THINK/PLAN) 的需求分析和设计
+- Phase 3-6 (REVIEW/USER ACCEPT/FEEDBACK/SHIP) 的后续阶段
+- 并行模式（需显式使用 `--mode parallel` 切换到 `dispatching-parallel-agents`）
+- 紧急热修复（应使用标准 sprint-flow 或手动处理）
+- 非 specification.yaml 定义的需求（临时需求应先纳入 spec 再执行）
+
+---
+
+## Examples
+
+**Example 1: 标准 Sprint Flow 启动**
+```bash
+/sprint-flow "开发用户登录功能，支持 JWT 和 OAuth2" --type web-nextjs
+# → Phase 0-1 完成设计评审后，Phase 2 自动进入 ralph-loop 模式
+# → 按拓扑顺序处理 REQ-001 (User Model) → REQ-002 (Auth Middleware) → REQ-003 (Login API)
+```
+
+**Example 2: 单个 REQ 构建**
+```bash
+/ralph-loop
+# → 从 specification.yaml 读取下一个 READY REQ
+# → 创建孤立上下文，dispatch TDD subagent
+# → 全量回归测试通过后 commit
+```
+
+**Example 3: 处理阻塞 REQ**
+```
+REQ-005 失败 3 次 → BLOCKED
+→ 用户决策：
+  - skip: 跳过此 REQ，继续 REQ-006
+  - manual: 手动修复后重新 dispatch
+  - stop: 停止整个 ralph-loop，保留已 done commits
+  - rollback: git reset 回上一个稳定版本
+```
+
+**Example 4: 崩溃恢复**
+```bash
+# 系统崩溃后重新启动
+/ralph-loop --resume
+# → 检测 progress.log → 跳过已 done REQs
+# → 从最后一个 pending REQ 继续，无需重做已完成工作
+```
+
+**Example 5: 测试基础设施缺失**
+```
+REQ-001 开始 → 检查 test-utils.ts → 不存在
+→ dispatch "生成测试基础设施" subagent
+→ retry max 2 次 → 仍失败 → BLOCK 或 fallback inline 生成
+→ 基础设施就绪后继续业务代码开发
+```
+
+---
+
+## Output Contract (Mandatory Checklist)
+
+**每个 REQ 完成时必须输出以下结构化检查清单**：
+
+```markdown
+## REQ-XXX: [Title] - Execution Summary
+
+**Status**: ✅ PASS / ❌ FAIL (retry n/3) / 🚫 BLOCKED
+
+**Context Isolation**:
+- [ ] Previous REQ context cleared
+- [ ] Permanent learnings injected (N items)
+- [ ] Contextual learnings injected (≤3 items)
+- [ ] Test infrastructure confirmed ready
+
+**TDD Compliance**:
+- [ ] Tests written BEFORE implementation
+- [ ] Test/implementation ratio ≥ 40%
+- [ ] Mock usage justified (if >30% mock density)
+
+**Verification Layers**:
+- [ ] L1: typecheck + lint pass
+- [ ] L1b: Test-first ratio check pass
+- [ ] L2: **FULL regression tests** (all tests, not just @test REQ-XXX)
+- [ ] L3: coverage ≥ 80%
+
+**Learnings Persisted**:
+- [ ] Permanent: [N items, list key ones]
+- [ ] Contextual: [N items, sliding window]
+- [ ] `gstack/learn` called
+
+**State Updated**:
+- [ ] AGENTS.md updated (orchestrator)
+- [ ] progress.log atomically written
+- [ ] Git commit created (if PASS)
+
+**Next Step**: [REQ-YYY title / COMPLETE / BLOCKED]
+```
+
+**Eval Criteria** (Issue #120, #121):
+- ✅ L2 pass: Structured execution plan visible (checklist format)
+- ✅ L3 pass: All 7 workflow steps followed, full regression tests run
+- ❌ Fail: Missing checklist, partial tests, context pollution
+
+---
+
 ## Anti-Patterns
 
 | ❌ 错误 | ✅ 正确 |