ai-spec-dev 0.30.1 → 0.33.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ai-spec-dev",
-  "version": "0.30.1",
+  "version": "0.33.0",
   "description": "AI-driven Development Orchestrator SDK & CLI",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -9,7 +9,9 @@
   },
   "scripts": {
     "build": "tsup",
-    "dev": "tsup --watch"
+    "dev": "tsup --watch",
+    "test": "vitest run",
+    "test:watch": "vitest"
   },
   "keywords": [
     "ai",
@@ -41,6 +43,7 @@
     "glob": "^13.0.6",
     "ts-node": "^10.9.2",
     "tsup": "^8.4.0",
-    "typescript": "^5.7.3"
+    "typescript": "^5.7.3",
+    "vitest": "^2.1.0"
   }
 }

package/purpose.md

@@ -2,7 +2,7 @@
 
 > 痛点 · 架构创新 · 边界处理 · DSL 的意义 · 当前局限 · 未来方向
 >
-> 当前版本：v0.30.0 · 最后更新：2026-03-29
+> 当前版本：v0.31.0 · 最后更新：2026-03-29
 
 ***
 
@@ -19,7 +19,7 @@
 7. [当前局限](#7-当前局限)
 8. [未来优化方向](#8-未来优化方向)
 
-> **版本记录速览**：v0.17.0 宪法截断修复 · v0.18.0 `learn` + `minSpecScore` + 行为契约提取 · v0.19.0 错误解析重写 + Auto Gate 修复 · v0.20.0 `mock --serve` 一键联调 · v0.21.0 store 公开 API 提取修复 · v0.22.0 service/api 层分离 · v0.23.0 view/route 双层 + 文件名幻觉修复 · v0.24.0 四项质量修复（export default、impliesRegistration、依赖拓扑排序、lesson 计数）· v0.25.0 HTTP import 正则、分页提取、isToolCrash 三项修复 · v0.26.0 多仓库 review 目录、batch 容错、tasks JSON 健壮性 · **v0.27.0 可靠性三件套**（Provider retry/timeout/分类、文件快照 + `restore`、RunId 结构化日志）· **v0.28.0 3-pass review**（Pass 3 影响面评估 + 代码复杂度评估）· **v0.29.0 全量审查修复**（RunLogger 完整插桩、update 快照/日志/knowledge、Score Trend 显示影响/复杂度等级、死代码清理）· **v0.30.0 错误修复依赖图排序 + 前端 Import 多行感知解析**
+> **版本记录速览**：v0.17.0 宪法截断修复 · v0.18.0 `learn` + `minSpecScore` + 行为契约提取 · v0.19.0 错误解析重写 + Auto Gate 修复 · v0.20.0 `mock --serve` 一键联调 · v0.21.0 store 公开 API 提取修复 · v0.22.0 service/api 层分离 · v0.23.0 view/route 双层 + 文件名幻觉修复 · v0.24.0 四项质量修复（export default、impliesRegistration、依赖拓扑排序、lesson 计数）· v0.25.0 HTTP import 正则、分页提取、isToolCrash 三项修复 · v0.26.0 多仓库 review 目录、batch 容错、tasks JSON 健壮性 · **v0.27.0 可靠性三件套**（Provider retry/timeout/分类、文件快照 + `restore`、RunId 结构化日志）· **v0.28.0 3-pass review**（Pass 3 影响面评估 + 代码复杂度评估）· **v0.29.0 全量审查修复**（RunLogger 完整插桩、update 快照/日志/knowledge、Score Trend 显示影响/复杂度等级、死代码清理）· **v0.30.0 错误修复依赖图排序 + 前端 Import 多行感知解析** · **v0.31.0 Harness Engineer：Prompt Hash + Create 内联 Self-Eval**
 
 ***
 
@@ -55,10 +55,107 @@ ai-spec 对每个痛点都有对应的架构设计，不是功能堆砌，而是
 
 **核心定位**：ai-spec 不是代码补全工具，而是一个「AI 辅助工程流程编排器」。它的目标是让工程师用最少的时间获得一个符合项目规范、通过基本质检、可直接进入 Review 的代码分支。
 
+### 1.3 整体架构鸟瞰
+
+> 在进入各模块细节之前，先建立一个全局视图。
+
+```mermaid
+flowchart TD
+    subgraph INPUT["输入层"]
+        IDEA["💬 需求描述\n自然语言"]
+        CONST["📜 项目宪法\n§1-§8 规则 + §9 教训"]
+        CTX["🗂️ 项目上下文\n代码结构 / 依赖 / 路由"]
+    end
+
+    subgraph CONTRACT["双层契约"]
+        SPEC["📄 Spec\nMarkdown 人类可读"]
+        DSL["📊 DSL\nJSON 机器可读"]
+    end
+
+    subgraph GATE["质量门控"]
+        SCORE_GATE["🎯 Spec 质量评分\nminSpecScore 阈值"]
+        APPROVAL["🧑‍💻 Approval Gate\n人工确认后才开始生成"]
+    end
+
+    subgraph GENERATE["生成层"]
+        CODEGEN["⚙️ Task 分层代码生成\ndata→service→api→view→route\n层内拓扑排序 + batch 并行"]
+        CACHE[("🗄️ File Cache\n行为契约 / 函数签名")]
+    end
+
+    subgraph VERIFY["验证层"]
+        ERRLOOP["🔄 错误反馈闭环\n≤2 cycle · 依赖图排序修复"]
+        REVIEW["🔬 3-pass 代码审查\n架构 + 实现 + 影响面"]
+    end
+
+    subgraph LEARN["学习层（闭环）"]
+        KNOW["📚 §9 知识积累\n审查 issue 自动写入"]
+        EVAL["📈 Harness Self-Eval\nharnessScore + promptHash"]
+    end
+
+    IDEA --> SPEC
+    CONST -->|"全文注入所有 prompt"| SPEC
+    CTX --> SPEC
+
+    SPEC --> SCORE_GATE
+    SCORE_GATE -->|"通过"| APPROVAL
+    SCORE_GATE -->|"不足"| STOP1(["🚫 中止"])
+    APPROVAL -->|"Proceed"| DSL
+    APPROVAL -->|"Abort"| STOP2(["🚫 退出，无残留"])
+
+    DSL --> CODEGEN
+    CODEGEN <-->|"读写"| CACHE
+    CODEGEN --> ERRLOOP
+    ERRLOOP --> REVIEW
+    REVIEW --> KNOW
+    KNOW -->|"更新宪法 §9"| CONST
+    REVIEW --> EVAL
+```
+
 ***
 
 ## 2. 核心架构设计
 
+### 2.0 `ai-spec create` 完整流水线
+
+> 以下流程图展示了运行一次 `ai-spec create` 时所有步骤的完整执行路径，包括每个决策门和反馈循环。后续 §2.1—§2.13 各节是对图中各模块的深度解析。
+
+```mermaid
+flowchart TD
+    START(["▶ ai-spec create &lt;idea&gt;"])
+
+    START --> S1["Step 1 · 加载项目上下文\nContextLoader 扫描代码结构 / 依赖 / 路由 / schema"]
+    S1 --> S2["Step 2 · Spec + Tasks 生成\n宪法全文注入 prompt 最高优先级"]
+    S2 --> S3["Step 3 · 交互式润色\nDiff 预览，可多轮修改"]
+    S3 --> S34["Step 3.4 · Spec 质量评估\n覆盖度 / 清晰度 / 宪法符合度打分"]
+
+    S34 --> G1{Score ≥ minSpecScore?}
+    G1 -->|"否"| ABORT1(["🚫 exit(1)\n--force 可强制继续"])
+    G1 -->|"是"| G2{"Approval Gate\n展示 Spec + DSL 摘要\n等待人工决策"}
+    G2 -->|"Abort"| ABORT2(["🚫 退出\nSpec 不写入磁盘"])
+    G2 -->|"Proceed"| DSL["DSL 提取 + 9 条规则校验\n失败最多重试 2 次"]
+
+    DSL --> TRACK["RunId 生成 + 文件快照初始化\nPrompt Hash 写入 RunLog"]
+    TRACK --> CG["Step 5–6 · Task 分层代码生成\ndata → infra → service → api → view → route → test\n层内拓扑排序 → batch 并行 → 缓存更新"]
+    CG --> TG["Step 7 · 测试骨架生成"]
+
+    TG --> EF["Step 8 · 错误反馈闭环"]
+    EF --> EF_RUN["运行 test / lint / tsc"]
+    EF_RUN --> EF_CHECK{全部通过?}
+    EF_CHECK -->|"通过"| RV
+    EF_CHECK -->|"有错误 · cycle ≤ 2"| EF_FIX["依赖图排序\nAI 逐文件修复\n携带 DSL 上下文"]
+    EF_FIX --> EF_RUN
+    EF_CHECK -->|"cycle 2 仍失败"| EF_WARN["⚠️ 黄色警告，继续"]
+    EF_WARN --> RV
+
+    RV["Step 9 · 3-pass 代码审查\nPass1 架构 · Pass2 实现 · Pass3 影响面/复杂度"]
+    RV --> KNOW["§9 知识积累\n审查 issue 自动追加宪法"]
+
+    KNOW --> SE["Step 10 · Harness Self-Eval\nDSL 覆盖 + Compile + Review → harnessScore\nPromptHash 关联，零 AI 调用"]
+    SE --> DONE(["✔ Done\nSpec / DSL / 代码 / RunLog 全部落盘"])
+```
+
+***
+
 ### 2.1 项目宪法：可进化的项目记忆
 
 绝大多数 AI 工具的「上下文注入」是静态的——你手动写一段 prompt，每次带进去。ai-spec 的宪法系统不同，它是一个会随项目迭代自动更新的活文档。
@@ -148,6 +245,43 @@ data → infra → service → api → view → route → test
 | route | — | 路由模块文件（`src/router/routes/`） |
 | test | 单测、集成测试 | 同左 |
 
+**执行模型图解：**
+
+```mermaid
+flowchart TB
+    subgraph LAYERS["七层顺序（跨层串行）"]
+        direction LR
+        LA["data"] --> LB["infra"] --> LC["service"] --> LD["api"] --> LE["view"] --> LF["route"] --> LG["test"]
+    end
+
+    subgraph WITHIN["单层内部执行（以 api 层为例）"]
+        direction TB
+        TOPO["拓扑排序\n按 dependencies 字段分 batch"]
+
+        subgraph B1["Batch 1 · 无依赖 → 并行"]
+            T1["userController.ts"]
+            T2["authController.ts"]
+        end
+
+        subgraph B2["Batch 2 · 依赖 Batch 1 → 并行"]
+            T3["adminController.ts\n需要 import userController"]
+        end
+
+        TOPO --> B1
+        B1 -->|"batch 完成\n更新 FileCache"| B2
+    end
+
+    CACHE[("generatedFileCache\n函数签名 / 文件路径\nbehavioral contracts")]
+
+    B1 <-->|"读写"| CACHE
+    B2 <-->|"读写"| CACHE
+
+    LAYERS -->|"进入当层"| WITHIN
+    WITHIN -->|"层完成\n更新共享 config 文件\n(routes/index.ts 等)"| NEXT(["下一层"])
+```
+
+> **为什么 route 必须在 view 之后？** view 完成后，`TaskManagement.vue` 的真实路径进入 FileCache；route task 生成时 prompt 里已经有 `// exists: src/views/task-management/TaskManagement.vue`，AI 不会再猜测 `index.vue`（v0.23.0 文件名幻觉修复）。
+
 **前端四层链路设计（v0.22.0+）**：`service`（api 调用函数）→ `api`（store，导入 service 函数）→ `view`（页面，使用 store action）→ `route`（路由模块，导入 view 组件）。`route` 必须在 `view` 之后生成，因为路由文件需要知道 view 组件的确切文件名（如 `TaskManagement.vue` 而非猜测的 `index.vue`）——这是 v0.23.0 修复文件名幻觉的核心机制。
 
 每个 task 完成后立即写入 `status: done` 到 `tasks.json`，`--resume` 标志让流水线跳过已完成 task，中断恢复精确到 task 粒度。`tasks.json` 文件损坏时（意外截断等）能检测并优雅降级，提示重新生成（v0.26.0）。
@@ -415,6 +549,59 @@ Pass 3 不重复 Pass 1 / Pass 2 的发现，而是站在更高的系统视角
 
 ***
 
+### 2.13 Harness Engineer：Prompt Hash + Self-Eval（v0.31.0+）
+
+#### 背景：缺失的自我评估能力
+
+v0.30.0 之前，ai-spec 是一个能自动生成代码的 Harness，但它是一个**没有自我量化能力的 Harness**。你可以感知到某次生成质量好坏，但无法回答：
+
+- 修改了 codegen prompt 之后，整体质量是提升还是下降了？
+- 哪个 provider / model 在这个项目上生成质量最稳定？
+- 加严宪法 §9 之后，compile 通过率有没有提高？
+
+这正是 **Harness Engineer** 理念的核心缺口：工程师不只是构建 AI 生成系统，还必须能**量化证明这个系统在变好**。
+
+#### Prompt Hash（`core/prompt-hasher.ts`）
+
+每次 `ai-spec create` 启动时，对 6 个核心 prompt 字符串（codegen、DSL extractor、spec generator、review 三 pass）计算 SHA-256 并取前 8 位（如 `a3f2c1d8`），写入 RunLog 根级字段 `promptHash`。
+
+```json
+{
+  "runId": "20260329-143022-a7f2",
+  "promptHash": "a3f2c1d8",
+  "harnessScore": 7.8,
+  ...
+}
+```
+
+任何 prompt 文件的改动都会产生不同的 hash。跨多个 RunLog 对比 `harnessScore` 时，先按 `promptHash` 分组，就能把「prompt 版本差异」从「模型随机性」中解耦，知道分数变化是因为 prompt 改了还是 LLM 本身的波动。
+
+#### Create 内联 Self-Eval（`core/self-evaluator.ts`）
+
+`ai-spec create` 在 Step 9（code review）之后新增 **Step 10: Harness Self-Eval**，零 AI 调用，纯确定性评分：
+
+| 维度 | 评分逻辑 | 权重 |
+|------|---------|------|
+| **DSL Coverage** (0-10) | 检查生成文件是否覆盖了 DSL 声明的 endpoint 层和 model 层 | 40% |
+| **Compile Score** (0-10) | error feedback 全部通过 → 10；未通过 / 跳过 → 5 | 30% |
+| **Review Score** (0-10) | 从 3-pass review 文本提取 `Score: X/10` | 30% |
+
+当 review 被跳过（`--skip-review`）时，权重自动调整为 DSL 55% + Compile 45%。
+
+**输出示例：**
+
+```
+─── Harness Self-Eval ───────────────────────────
+  Score  : [████████░░] 7.8/10
+  DSL    : 8/10  Compile: pass  Review: 7.2/10
+  Prompt : a3f2c1d8
+─────────────────────────────────────────────────
+```
+
+`harnessScore` 和 `promptHash` 同时写入 RunLog，为日后实现跨运行趋势分析奠定数据基础。
+
+***
+
 ## 3. DSL 层的意义
 
 DSL 是整个系统中设计投入最大的模块，也是最容易被误解为「多此一举」的部分。

package/tests/dsl-extractor.test.ts

@@ -0,0 +1,264 @@
+import { describe, it, expect, vi } from "vitest";
+import {
+  dslFilePath,
+  buildDslContextSection,
+  DslExtractor,
+} from "../core/dsl-extractor";
+import type { SpecDSL } from "../core/dsl-types";
+import type { AIProvider } from "../core/spec-generator";
+
+// ─── Fixtures ─────────────────────────────────────────────────────────────────
+
+const VALID_DSL: SpecDSL = {
+  version: "1.0",
+  feature: {
+    id: "user-login",
+    title: "User Login",
+    description: "Authenticate users with email and password",
+  },
+  models: [
+    {
+      name: "User",
+      fields: [
+        { name: "id", type: "String", required: true },
+        { name: "email", type: "String", required: true, unique: true },
+      ],
+      relations: ["has many Session"],
+    },
+  ],
+  endpoints: [
+    {
+      id: "EP-001",
+      method: "POST",
+      path: "/api/auth/login",
+      description: "Authenticate and return JWT",
+      auth: false,
+      successStatus: 200,
+      successDescription: "JWT token",
+      request: { body: { email: "string", password: "string" } },
+      errors: [
+        { status: 401, code: "INVALID_CREDENTIALS", description: "Bad password" },
+      ],
+    },
+  ],
+  behaviors: [
+    {
+      id: "BHV-001",
+      description: "Rate-limit login to 5 attempts per minute",
+      trigger: "POST /api/auth/login",
+      constraints: ["block after 5 failures"],
+    },
+  ],
+};
+
+function makeProvider(response: string): AIProvider {
+  return { generate: vi.fn().mockResolvedValue(response) };
+}
+
+// ─── dslFilePath ──────────────────────────────────────────────────────────────
+
+describe("dslFilePath", () => {
+  it("replaces .md extension with .dsl.json", () => {
+    expect(dslFilePath("/specs/feature-login-v1.md")).toBe("/specs/feature-login-v1.dsl.json");
+  });
+
+  it("works with relative paths", () => {
+    expect(dslFilePath("specs/my-feature-v2.md")).toBe("specs/my-feature-v2.dsl.json");
+  });
+
+  it("handles files in the current directory", () => {
+    expect(dslFilePath("feature.md")).toBe("feature.dsl.json");
+  });
+
+  it("preserves directory structure", () => {
+    const result = dslFilePath("/a/b/c/feature-v3.md");
+    expect(result).toContain("/a/b/c/");
+    expect(result.endsWith(".dsl.json")).toBe(true);
+  });
+});
+
+// ─── buildDslContextSection ───────────────────────────────────────────────────
+
+describe("buildDslContextSection", () => {
+  it("includes the section header and footer", () => {
+    const result = buildDslContextSection(VALID_DSL);
+    expect(result).toContain("=== Feature DSL");
+    expect(result).toContain("=== End of DSL ===");
+  });
+
+  it("lists model names and fields", () => {
+    const result = buildDslContextSection(VALID_DSL);
+    expect(result).toContain("User:");
+    expect(result).toContain("email: String");
+  });
+
+  it("marks required fields", () => {
+    const result = buildDslContextSection(VALID_DSL);
+    expect(result).toContain("required");
+  });
+
+  it("marks unique fields", () => {
+    const result = buildDslContextSection(VALID_DSL);
+    expect(result).toContain("unique");
+  });
+
+  it("includes model relations", () => {
+    const result = buildDslContextSection(VALID_DSL);
+    expect(result).toContain("has many Session");
+  });
+
+  it("includes endpoint method, path, and auth", () => {
+    const result = buildDslContextSection(VALID_DSL);
+    expect(result).toContain("POST");
+    expect(result).toContain("/api/auth/login");
+    expect(result).toContain("auth: false");
+  });
+
+  it("includes endpoint error codes", () => {
+    const result = buildDslContextSection(VALID_DSL);
+    expect(result).toContain("INVALID_CREDENTIALS");
+  });
+
+  it("includes request body fields", () => {
+    const result = buildDslContextSection(VALID_DSL);
+    expect(result).toContain("email");
+    expect(result).toContain("password");
+  });
+
+  it("includes behaviors with trigger and constraints", () => {
+    const result = buildDslContextSection(VALID_DSL);
+    expect(result).toContain("Rate-limit login");
+    expect(result).toContain("POST /api/auth/login");
+    expect(result).toContain("block after 5 failures");
+  });
+
+  it("handles empty models array gracefully", () => {
+    const dsl: SpecDSL = { ...VALID_DSL, models: [] };
+    const result = buildDslContextSection(dsl);
+    expect(result).not.toContain("-- Data Models --");
+  });
+
+  it("handles empty endpoints array gracefully", () => {
+    const dsl: SpecDSL = { ...VALID_DSL, endpoints: [] };
+    const result = buildDslContextSection(dsl);
+    expect(result).not.toContain("-- API Endpoints --");
+  });
+
+  it("handles empty behaviors array gracefully", () => {
+    const dsl: SpecDSL = { ...VALID_DSL, behaviors: [] };
+    const result = buildDslContextSection(dsl);
+    expect(result).not.toContain("-- Business Behaviors --");
+  });
+
+  it("includes UI components section when components are present", () => {
+    const dsl: SpecDSL = {
+      ...VALID_DSL,
+      components: [
+        {
+          id: "CMP-001",
+          name: "LoginForm",
+          description: "Login form component",
+          props: [{ name: "onSuccess", type: "() => void", required: true }],
+          events: [{ name: "onSubmit", payload: "FormData" }],
+          state: { isLoading: "boolean" },
+          apiCalls: ["EP-001"],
+        },
+      ],
+    };
+    const result = buildDslContextSection(dsl);
+    expect(result).toContain("-- UI Components --");
+    expect(result).toContain("LoginForm");
+    expect(result).toContain("onSuccess");
+    expect(result).toContain("onSubmit");
+    expect(result).toContain("isLoading:boolean");
+    expect(result).toContain("EP-001");
+  });
+});
+
+// ─── DslExtractor.extract — success path ─────────────────────────────────────
+
+describe("DslExtractor.extract — success", () => {
+  it("returns a valid SpecDSL when AI output is bare JSON", async () => {
+    const provider = makeProvider(JSON.stringify(VALID_DSL));
+    const extractor = new DslExtractor(provider);
+    const consoleSpy = vi.spyOn(console, "log").mockImplementation(() => {});
+    const result = await extractor.extract("spec content", { auto: true });
+    consoleSpy.mockRestore();
+    expect(result).not.toBeNull();
+    expect(result?.feature.id).toBe("user-login");
+  });
+
+  it("returns a valid SpecDSL when AI wraps output in a JSON fence", async () => {
+    const fenced = "```json\n" + JSON.stringify(VALID_DSL) + "\n```";
+    const provider = makeProvider(fenced);
+    const extractor = new DslExtractor(provider);
+    const consoleSpy = vi.spyOn(console, "log").mockImplementation(() => {});
+    const result = await extractor.extract("spec content", { auto: true });
+    consoleSpy.mockRestore();
+    expect(result?.feature.title).toBe("User Login");
+  });
+
+  it("sanitizes empty error entries before validation", async () => {
+    const dslWithEmptyErrors = {
+      ...VALID_DSL,
+      endpoints: [
+        {
+          ...VALID_DSL.endpoints[0],
+          errors: [
+            { status: 400, code: "", description: "" }, // invalid — should be stripped
+            { status: 401, code: "INVALID_CREDENTIALS", description: "Bad password" },
+          ],
+        },
+      ],
+    };
+    const provider = makeProvider(JSON.stringify(dslWithEmptyErrors));
+    const extractor = new DslExtractor(provider);
+    const consoleSpy = vi.spyOn(console, "log").mockImplementation(() => {});
+    const result = await extractor.extract("spec content", { auto: true });
+    consoleSpy.mockRestore();
+    expect(result).not.toBeNull();
+    // The empty error entry should have been stripped
+    expect(result?.endpoints[0].errors).toHaveLength(1);
+    expect(result?.endpoints[0].errors?.[0].code).toBe("INVALID_CREDENTIALS");
+  });
+});
+
+// ─── DslExtractor.extract — failure paths ────────────────────────────────────
+
+describe("DslExtractor.extract — failure / auto mode", () => {
+  it("returns null in auto mode when AI returns invalid JSON", async () => {
+    const provider = makeProvider("Not JSON at all");
+    const extractor = new DslExtractor(provider);
+    const consoleSpy = vi.spyOn(console, "log").mockImplementation(() => {});
+    const result = await extractor.extract("spec", { auto: true });
+    consoleSpy.mockRestore();
+    expect(result).toBeNull();
+  });
+
+  it("returns null in auto mode when provider throws", async () => {
+    const provider: AIProvider = { generate: vi.fn().mockRejectedValue(new Error("network")) };
+    const extractor = new DslExtractor(provider);
+    const consoleSpy = vi.spyOn(console, "log").mockImplementation(() => {});
+    const result = await extractor.extract("spec", { auto: true });
+    consoleSpy.mockRestore();
+    expect(result).toBeNull();
+  });
+
+  it("retries when first attempt produces invalid DSL (missing required field)", async () => {
+    // First response: invalid DSL (missing feature.description)
+    const badDsl = { ...VALID_DSL, feature: { id: "x", title: "X", description: "" } };
+    // Second response: valid DSL
+    const provider: AIProvider = {
+      generate: vi.fn()
+        .mockResolvedValueOnce(JSON.stringify(badDsl))
+        .mockResolvedValueOnce(JSON.stringify(VALID_DSL)),
+    };
+    const extractor = new DslExtractor(provider);
+    const consoleSpy = vi.spyOn(console, "log").mockImplementation(() => {});
+    const result = await extractor.extract("spec", { auto: true });
+    consoleSpy.mockRestore();
+    // Should have retried — provider called at least twice
+    expect((provider.generate as ReturnType<typeof vi.fn>).mock.calls.length).toBeGreaterThanOrEqual(2);
+    expect(result?.feature.id).toBe("user-login");
+  });
+});