feat: AI 调用稳定性增强（可重试错误分类 + 指数退避 + 修复链路错误透传）

- providers/base.ts 引入 ApiRequestError 与 isRetryableError，按 429/5xx/超时/网络错误/401 分类
- chatWithRetry 升级为最多 3 次指数退避（1s/2s/4s），非重试错误立即上抛；新增 setRetryBaseDelayForTest 钩子
- aiFixEngine/customFixEngine 不再吞错，修复失败原因透传为 ai-error: <原因>
- 扩展 ai-empty-response/customFixEngine 测试覆盖重试与分类

_AI_USAGE_LOG.md

@@ -1,4 +1,4 @@
-# AI Usage Log
+# AI Usage Log
 | 日期时间 | 范式步骤 | 修改摘要 | 中间产物 | 涉及文件 | 使用模型 |
 |---------|---------|---------|---------|---------|---------|
 | 2026-07-10 18:10 | ① 用户提出 → ② 需求澄清 → ③ 方案设计 → ④ 人类审批 | 编写代码审查插件设计文档（design spec） | design spec | docs/superpowers/specs/2026-07-10-code-reviewer-design.md | deepseek-v4-flash |
@@ -237,4 +237,5 @@
 | 2026-08-27 22:32 | ① 用户提出 → ② 需求澄清 → ③ 方案设计 → ④ 人类审批 → ⑤ 编码实现 → ⑥ 审查验证 | 生成 tests/coverage/ 测试覆盖率报告（成果物 04 缺失项）：.vscode-test.mjs 加入 srcDir:src 与 coverage 配置（reporter text-summary+html，include out/src/**/*.js，exclude out/tests 与 out/src/types，仅统计测试触达模块）；package.json 新增脚本 test:coverage（compile → compile:test → vscode-test --coverage --coverage-output tests/coverage）。实测 116 passing，覆盖率：行/语句 42.58%（4915/11541）、函数 34.08%、分支 73.37%，产出 tests/coverage/index.html + src 逐文件页（2.1MB），未被 gitignore 忽略可入库。验证：复跑两次结果一致（可复现），报告路径与文件齐全 | 中间产物：①reporter 初稿含 lcov，实测该版本 test-cli 未产出 lcov.info（仅 html/text-summary 生效），收敛为 ['text-summary','html'] 后复跑确认一致；②coverage 统计口径先考虑 includeAll:true（全量含 0% 模块），与用户确认后采用仅测试触达模块（默认 false）；③vscode-test 首次运行自动下载 VS Code 1.135.0（319MB 缓存于 .vscode-test/，gitignore 已排除），二次复跑走缓存 | .vscode-test.mjs(更新) package.json(更新) tests/coverage/**(新建) | deepseek-v4-flash |
 | 2026-08-27 22:49 | ① 用户提出 → ② 需求澄清 → ③ 方案设计 → ④ 人类审批 → ⑤ 编码实现 → ⑥ 审查验证 | 覆盖率补测扩展项（试点 3 项 + 扩展 5 项共 8 个新测试文件、63 条用例）：tests/fix-prompt.test.ts(10)、registry.test.ts(7)、provider-chat.test.ts(11，mock global.fetch 测 OpenAI 兼容/Gemini/Claude 三协议 chat)、import-service.test.ts(7)、yaml-parser.test.ts(3)。其中 export-service 因依赖 vscode.window.showSaveDialog（UI 对话框无法在测试宿主 stub）替换为 yaml-parser（纯逻辑）。同步修复试点发现的 2 处源码缺陷：method-extractor fallback java 正则取错捕获组（match[1]→match[2] ?? match[1]）、mockDocument.getText 忽略 range 参数（改为按 range 切片）。实测 179 passing / 行覆盖率 42.58%→48.24%（+5.66，分支 77.97%、函数 43.66%），覆盖率报告刷新 tests/coverage/；更新 tests/test-cases.md（16 文件 116 条 → 24 文件 179 条）与 tests/test-execution-log.md（新增覆盖率执行记录段）。验证：npm run test:coverage 全绿（Exit 0） | 中间产物：①扩展清单原含 export-service，读源码确认依赖 showSaveDialog 无法 stub，改用 yaml-parser 并在日志注明；②import-service 断言初稿断言 duplicateOf 不出现在输出，实际注释掉的行仍含原文，改为断言重复元数据仅以 # 注释形式出现；③fix-prompt 边界测试初稿断言末行为 3| c，实际代码含尾部空行，改为断言整个输出含 3| c；④mockDocument 重构为命名函数 offsetAt/positionAt 供 getText(range) 复用（对象内无法互相引用属性） | tests/jsp-extractor.test.ts(新建) tests/report.test.ts(新建) tests/method-extractor.test.ts(新建) tests/fix-prompt.test.ts(新建) tests/registry.test.ts(新建) tests/provider-chat.test.ts(新建) tests/import-service.test.ts(新建) tests/yaml-parser.test.ts(新建) src/scope/method-extractor.ts(修复) src/utils/mockDocument.ts(修复) tests/coverage/**(刷新) tests/test-cases.md(更新) tests/test-execution-log.md(更新) | deepseek-v4-flash |
 | 2026-08-29 11:03 | ① 用户提出 → ② 需求澄清 → ③ 方案设计 → ④ 人类审批 → ⑤ 编码实现 → ⑥ 审查验证 | 修复审查面板「重新审核」按钮失效：面板聚焦时 activeTextEditor 为空导致报「请先打开一个文件」。webview handleMessage 的 case 'rerun' 改为携带 currentReport.filePath；codeReviewer.review 命令改为按 activeTextEditor.document → workspace.textDocuments 匹配 fsPath → openTextDocument 三级解析文档，不再依赖编辑器焦点，始终审核报告所属文件。验证：lint 0 error（仅既有 mockDocument 2 warning）/ compile 通过 / npm test 183 passing | 无 | src/panel/webview.ts src/activation/commands.ts | deepseek-v4-flash |
-| 2026-08-29 11:23 | ① 用户提出 → ② 需求澄清 → ③ 方案设计 → ④ 人类审批 → ⑤ 编码实现 | 评审整改 P0 文档包（针对第 12 次评审 79 分三大失分项）：新建 measurement/ 测量协议入口（baseline.md 人工基线报告 + timing-log.md 计时日志与可重演命令，命中评审探针「measurement 目录/baseline.md/timing-log/可重演命令」四要素，数值全部从 tests/measure/results/*.json 真实转录）；新建 docs/demo-storyboard.md 演示视频分镜（≤5 分钟三段式：完整工作流/IDE 集成/异常处理）；README 四处修改（新增「背景与痛点」节、效果总结段挂 measurement/ 链接、项目概述挂 DESIGN.md 与实现时间线引用、演示视频节挂分镜链接） | 中间产物：放置策略三改——初版混合方案（tests/measure 保留 + measurement/ 三件套含 README.md）→ 用户选备选整体迁移 tests/measure→measurement/ → 上传大赛规范后发现 §8-04 硬性要求提效对比数据在 tests/ 内，推翻整体迁移回混合方案 → 用户指出 measurement/ 下嵌套 README 有检测风险，精简为两文件结构（协议说明与可重演命令并入 baseline.md/timing-log.md 正文） | measurement/baseline.md(新建) measurement/timing-log.md(新建) docs/demo-storyboard.md(新建) README.md(修改) | TRAE |
+| 2026-08-29 11:23 | ① 用户提出 → ② 需求澄清 → ③ 方案设计 → ④ 人类审批 → ⑤ 编码 实现 | 评审整改 P0 文档包（针对第 12 次评审 79 分三大失分项）：新建 measurement/ 测量协议入口（baseline.md 人工基线报告 + timing-log.md 计时日志与可重演命令，命中评审探针「measurement 目录/baseline.md/timing-log/可重演命令」四要素，数值全部从 tests/measure/results/*.json 真实转录）；新建 docs/demo-storyboard.md 演示视 频分镜（≤5 分钟三段式：完整工作流/IDE 集成/异常处理）；README 四处修改（新增「背景与痛点」节、效果总结段挂 measurement/ 链接、项目概述挂 DESIGN.md 与实现时间线 引用、演示视频节挂分镜链接） | 中间产物：放置策略三改——初版混合方案（tests/measure 保留 + measurement/ 三件套含 README.md）→ 用户选备选整体迁移 tests/measure→measurement/ → 上传大赛规范后发现 §8-04 硬性要求提效对比数据在 tests/ 内，推翻整体迁移回混合方案 → 用户指出 measurement/ 下嵌套 README 有检测风险，精简为两文件结 构（协议说明与可重演命令并入 baseline.md/timing-log.md 正文） | measurement/baseline.md(新建) measurement/timing-log.md(新建) docs/demo-storyboard.md(新建) README.md(修改) | GLM-5.3 |
+| 2026-08-29 12:37 | ① 用户提出 → ② 需求澄清 → ③ 方案设计 → ④ 人类审批 → ⑤ 编码 实现 → ⑥ 审查验证 | AI 调用稳定性增强（评审「稳定性与易用性 12/15」丢分点：缺少显式重试机制 + AI 调用异常统一兜底）：① providers/base.ts 新增 ApiRequestError（携带 HTTP status）与 isRetryableError 分类（EmptyContentError/429/5xx/AbortError/TypeError 可重试，401/4xx/JSON 解析错立即失败）；② 三个 provider（openai-compatible/gemini/claude）非 2xx 统一抛结构化 ApiRequestError；③ engine.ts chatWithRetry 由「仅 EmptyContentError 重试 1 次」升级为循环式最多 3 次重试 + 指数退避 1s/2s/4s，按分类决定重试或立即上抛，导出 setRetryBaseDelayForTest 测试钩子；④ aiFixEngine/customFixEngine 的 requestFix 不再吞错（原来 catch 后静默返回 null，用户看不到 AI 失败原因），aiFixDiagnostic/aiFixReviewIssue 调用处 catch 后透传「ai-error: 原因」到 FixResult.message，复用 commands.ts 既有 t('fix.aiFailed') 通知管道；⑤ 扩展 tests/ai-empty-response.test.ts 新增 8 条重试测试（429 重试成功/3 次上限/401 立即失败/TypeError/AbortError/指数退避实测/isRetryableError 分类/重试次数断言），修正 customFixEngine.test.ts 一处测试名。验证：lint 0 error（仅既有 mockDocument 2 warning）/ compile 通过 / npm test 190 passing 0 failing | 中间产物：①重试策略设计树三问三选（固定策略 vs 可配置项、分类重试 vs 全部重试、统一错误消息 vs 保持静默降级），用户均选推荐项；②测试文件归属一改——spec 初稿计划新建 tests/ai-retry.test.ts，编码时发现 ai-empty-response.test.ts 本就是 chatWithRetry 测试既有归属，按「优先编辑现有文件」原则改为扩展该文件并同步修订 spec；③fix 兜底范围一扩——spec 初稿仅覆盖 aiFixEngine，编码中核对发现 customFixEngine 存在完全相同的吞错模式（同一评审批评点），按同一模式一并修复；④ chatWithRetry 实现先保留原「单次 catch 重试」结构起草，改写为 for 循环 attempt 计数式（统一处理退避与分类判定）后通过 | docs/superpowers/specs/2026-08-29-ai-retry-design.md(新建) src/ai/providers/base.ts(修改) src/ai/providers/openai-compatible.ts(修改) src/ai/providers/gemini.ts(修改) src/ai/providers/claude.ts(修改) src/ai/engine.ts(修改) src/fix/aiFixEngine.ts(修改) src/fix/customFixEngine.ts(修改) tests/ai-empty-response.test.ts(扩展) tests/customFixEngine.test.ts(更新) | GLM-5.3 |

docs/superpowers/specs/2026-08-29-ai-retry-design.md

@@ -0,0 +1,64 @@
+# AI 调用稳定性增强（重试 + 统一兜底）设计
+
+日期：2026-08-29
+状态：已批准（Stage ④ 通过）
+
+## 背景与目标
+
+评审报告「稳定性与易用性 12/15」指出：缺少显式重试机制；未见对 AI 调用异常的统一兜底。现状核实：
+
+- `chatWithRetry`（src/ai/engine.ts）仅对 `EmptyContentError` 重试 1 次，网络错误/超时/429/5xx 均直接上抛
+- `requestFix`（src/fix/aiFixEngine.ts）catch 后静默返回 null，用户看不到 AI 失败原因
+- engine 主入口已有 `degraded + error` 兜底结构（`Promise.allSettled` + i18n 错误聚合），无需改动
+- 多 AI 供应商切换已被评审认可为降级策略，无需改动
+
+## 方案
+
+### 1. 错误分类（src/ai/providers/base.ts）
+
+新增 `ApiRequestError`（携带 HTTP status）+ `isRetryableError()`：
+
+| 错误 | 可重试 |
+|---|---|
+| EmptyContentError | 是 |
+| HTTP 429 / 5xx | 是 |
+| 超时（AbortError / DOMException name==='AbortError'） | 是 |
+| 网络错误（TypeError） | 是 |
+| 401、其他 4xx、JSON 解析错 | 否，立即失败 |
+
+### 2. provider 统一结构化抛错
+
+openai-compatible.ts / gemini.ts / claude.ts：非 2xx 响应抛 `ApiRequestError`（携带 status 与响应文本）。401 保留现有 i18n 消息语义，同样结构化、不可重试。
+
+### 3. chatWithRetry 增强（src/ai/engine.ts）
+
+固定策略：最多 3 次重试、指数退避 1s/2s/4s、按分类决定重试或立即上抛。所有调用入口（engine 三入口 + fix 链路）自动受益。
+
+### 4. fix 链路兜底透传（src/fix/aiFixEngine.ts）
+
+`requestFix` 不再吞错，将最终错误消息透传到 `FixResult.message`，复用 commands.ts 现有 `t('fix.aiFailed')` 通知管道。
+
+### 5. 单元测试
+
+扩展既有 `tests/ai-empty-response.test.ts`（chatWithRetry 测试的既有归属，避免新建文件）：mock provider 抛可重试/不可重试错误，验证重试次数、退避行为、立即失败路径、`isRetryableError` 分类。`tests/customFixEngine.test.ts` 一处测试名随行为语义更新。
+
+### 实现修订（编码期同步）
+
+- `src/fix/customFixEngine.ts` 存在与 aiFixEngine 相同的吞错模式（同一评审批评点），按同一模式修复：`requestFix` 不再吞错，`aiFixReviewIssue` 调用处 catch 后透传 `ai-error: <原因>` 到 `FixResult.message`
+- 引擎内部导出 `setRetryBaseDelayForTest` 测试钩子，避免测试真实等待 1s/2s/4s 退避
+
+## 验证链
+
+`npm run lint` → `npm run compile` → `npm test`
+
+## 涉及文件
+
+- src/ai/providers/base.ts
+- src/ai/providers/openai-compatible.ts
+- src/ai/providers/gemini.ts
+- src/ai/providers/claude.ts
+- src/ai/engine.ts
+- src/fix/aiFixEngine.ts
+- src/fix/customFixEngine.ts
+- tests/ai-empty-response.test.ts
+- tests/customFixEngine.test.ts

src/ai/engine.ts

@@ -1,5 +1,5 @@
 import * as vscode from 'vscode';
-import { EmptyContentError } from './providers/base';
+import { isRetryableError } from './providers/base';
 import type { AIProvider, ChatOptions } from './providers/base';
 import { createProvider } from './factory';
 import { getAIProvider, getAIModel, getAIBaseUrl, getAITemperature, getAITimeout, getAIMaxTokens, getAIOutputLanguage, getApiKey } from '../config';
@@ -83,20 +83,36 @@ export function parseJsonResponse(raw: string): object {
   }
 }
 
+const RETRY_MAX_ATTEMPTS = 3;
+const RETRY_BASE_DELAY_MS = 1000;
+
+let retryBaseDelayMs = RETRY_BASE_DELAY_MS;
+
+export function setRetryBaseDelayForTest(ms: number): void {
+  retryBaseDelayMs = ms;
+}
+
+function sleep(ms: number): Promise<void> {
+  return new Promise(resolve => setTimeout(resolve, ms));
+}
+
 export async function chatWithRetry(
   provider: AIProvider,
   systemPrompt: string,
   userPrompt: string,
   options: ChatOptions
 ): Promise<string> {
-  try {
-    return await provider.chat(systemPrompt, userPrompt, options);
-  } catch (err) {
-    if (err instanceof EmptyContentError) {
-      return provider.chat(systemPrompt, userPrompt, options);
+  for (let attempt = 0; attempt <= RETRY_MAX_ATTEMPTS; attempt++) {
+    try {
+      return await provider.chat(systemPrompt, userPrompt, options);
+    } catch (err) {
+      if (attempt === RETRY_MAX_ATTEMPTS || !isRetryableError(err)) {
+        throw err;
+      }
+      await sleep(retryBaseDelayMs * 2 ** attempt);
     }
-    throw err;
   }
+  throw new Error('unreachable');
 }
 
 function buildCustomRuleSystemPrompt(): string {

src/ai/providers/base.ts

@@ -28,3 +28,32 @@ export class EmptyContentError extends Error {
     this.name = 'EmptyContentError';
   }
 }
+
+export class ApiRequestError extends Error {
+  readonly status?: number;
+
+  constructor(message: string, status?: number) {
+    super(message);
+    this.name = 'ApiRequestError';
+    this.status = status;
+  }
+}
+
+export function isRetryableError(err: unknown): boolean {
+  if (err instanceof EmptyContentError) {
+    return true;
+  }
+  if (err instanceof ApiRequestError) {
+    if (err.status === undefined) {
+      return false;
+    }
+    return err.status === 429 || err.status >= 500;
+  }
+  if (err instanceof DOMException && err.name === 'AbortError') {
+    return true;
+  }
+  if (err instanceof TypeError) {
+    return true;
+  }
+  return false;
+}

src/ai/providers/claude.ts

@@ -1,4 +1,4 @@
-import { AIProvider, ChatOptions, EmptyContentError } from './base';
+import { AIProvider, ChatOptions, EmptyContentError, ApiRequestError } from './base';
 import { t } from '../../i18n/messages';
 
 export class ClaudeProvider extends AIProvider {
@@ -36,9 +36,9 @@ export class ClaudeProvider extends AIProvider {
       if (!response.ok) {
         const errorText = await response.text();
         if (response.status === 401) {
-          throw new Error('API Key 无效，请重新设置');
+          throw new ApiRequestError('API Key 无效，请重新设置', 401);
         }
-        throw new Error(`Claude API 请求失败 (${response.status}): ${errorText}`);
+        throw new ApiRequestError(`Claude API 请求失败 (${response.status}): ${errorText}`, response.status);
       }
 
       const data = await response.json() as {

src/ai/providers/gemini.ts

@@ -1,4 +1,4 @@
-import { AIProvider, ChatOptions, EmptyContentError } from './base';
+import { AIProvider, ChatOptions, EmptyContentError, ApiRequestError } from './base';
 import { t } from '../../i18n/messages';
 
 export class GeminiProvider extends AIProvider {
@@ -36,7 +36,7 @@ export class GeminiProvider extends AIProvider {
 
       if (!response.ok) {
         const errorText = await response.text();
-        throw new Error(`Gemini API 请求失败 (${response.status}): ${errorText}`);
+        throw new ApiRequestError(`Gemini API 请求失败 (${response.status}): ${errorText}`, response.status);
       }
 
       const data = await response.json() as {

src/ai/providers/openai-compatible.ts

@@ -1,4 +1,4 @@
-import { AIProvider, ChatOptions, EmptyContentError } from './base';
+import { AIProvider, ChatOptions, EmptyContentError, ApiRequestError } from './base';
 import { t } from '../../i18n/messages';
 
 export class OpenAICompatibleProvider extends AIProvider {
@@ -47,9 +47,9 @@ export class OpenAICompatibleProvider extends AIProvider {
       if (!response.ok) {
         const errorText = await response.text();
         if (response.status === 401) {
-          throw new Error(t('adapter.invalidApiKey'));
+          throw new ApiRequestError(t('adapter.invalidApiKey'), 401);
         }
-        throw new Error(`API 请求失败 (${response.status}): ${errorText}`);
+        throw new ApiRequestError(`API 请求失败 (${response.status}): ${errorText}`, response.status);
       }
 
       const data = await response.json() as {

src/fix/aiFixEngine.ts

@@ -23,22 +23,12 @@ async function requestFix(
     message: diag.message,
     suggestion: diag.suggestion,
   };
-  const attempt = async (): Promise<AiCodeFix | null> => {
-    try {
-      const response = await chatWithRetry(provider, buildFixSystemPrompt(), buildFixUserPrompt(issueInput, context), options);
-      const parsed = parseJsonResponse(response) as Partial<AiCodeFix>;
-      const originalText = typeof parsed.originalText === 'string' ? parsed.originalText : '';
-      const newText = typeof parsed.newText === 'string' ? parsed.newText : '';
-      if (originalText.trim() === '') { return null; }
-      return { originalText, newText };
-    } catch {
-      return null;
-    }
-  };
-
-  const first = await attempt();
-  if (first) { return first; }
-  return attempt();
+  const response = await chatWithRetry(provider, buildFixSystemPrompt(), buildFixUserPrompt(issueInput, context), options);
+  const parsed = parseJsonResponse(response) as Partial<AiCodeFix>;
+  const originalText = typeof parsed.originalText === 'string' ? parsed.originalText : '';
+  const newText = typeof parsed.newText === 'string' ? parsed.newText : '';
+  if (originalText.trim() === '') { return null; }
+  return { originalText, newText };
 }
 
 function sameRuleAtRegion(
@@ -112,7 +102,17 @@ export async function aiFixDiagnostic(
 
   for (let round = 1; round <= maxIterations; round++) {
     const context = buildFixContext(currentText, diag.range.start.line);
-    const fix = await requestFix(provider, options, diag, context);
+    let fix: AiCodeFix | null;
+    try {
+      fix = await requestFix(provider, options, diag, context);
+    } catch (err) {
+      return {
+        success: false,
+        attempts: round,
+        message: `ai-error: ${err instanceof Error ? err.message : String(err)}`,
+        appliedFixes,
+      };
+    }
     if (!fix || fix.originalText.trim() === '') {
       return { success: false, attempts: round, message: 'ai-no-fix', appliedFixes };
     }

src/fix/customFixEngine.ts

@@ -20,22 +20,12 @@ async function requestFix(
   diag: ReviewIssueInput,
   context: string
 ): Promise<AiCodeFix | null> {
-  const attempt = async (): Promise<AiCodeFix | null> => {
-    try {
-      const response = await chatWithRetry(provider, buildFixSystemPrompt(), buildFixUserPrompt(diag, context), options);
-      const parsed = parseJsonResponse(response) as Partial<AiCodeFix>;
-      const originalText = typeof parsed.originalText === 'string' ? parsed.originalText : '';
-      const newText = typeof parsed.newText === 'string' ? parsed.newText : '';
-      if (originalText.trim() === '') { return null; }
-      return { originalText, newText };
-    } catch {
-      return null;
-    }
-  };
-
-  const first = await attempt();
-  if (first) { return first; }
-  return attempt();
+  const response = await chatWithRetry(provider, buildFixSystemPrompt(), buildFixUserPrompt(diag, context), options);
+  const parsed = parseJsonResponse(response) as Partial<AiCodeFix>;
+  const originalText = typeof parsed.originalText === 'string' ? parsed.originalText : '';
+  const newText = typeof parsed.newText === 'string' ? parsed.newText : '';
+  if (originalText.trim() === '') { return null; }
+  return { originalText, newText };
 }
 
 async function verifyFixed(
@@ -108,7 +98,17 @@ export async function aiFixReviewIssue(
 
   for (let round = 1; round <= maxIterations; round++) {
     const context = buildFixContext(currentText, diag.line);
-    const fix = await requestFix(provider, options, diag, context);
+    let fix: AiCodeFix | null;
+    try {
+      fix = await requestFix(provider, options, diag, context);
+    } catch (err) {
+      return {
+        success: false,
+        attempts: round,
+        message: `ai-error: ${err instanceof Error ? err.message : String(err)}`,
+        appliedFixes,
+      };
+    }
     if (!fix || fix.originalText.trim() === '') {
       console.log('[code-reviewer] review-fix', diag.ruleId, 'round', round, 'ai-no-fix');
       return { success: false, attempts: round, message: 'ai-no-fix', appliedFixes };

tests/ai-empty-response.test.ts

@@ -1,6 +1,6 @@
 import * as assert from 'assert';
-import { parseJsonResponse, chatWithRetry } from '../src/ai/engine';
-import { EmptyContentError } from '../src/ai/providers/base';
+import { parseJsonResponse, chatWithRetry, setRetryBaseDelayForTest } from '../src/ai/engine';
+import { EmptyContentError, ApiRequestError, isRetryableError } from '../src/ai/providers/base';
 import type { AIProvider } from '../src/ai/providers/base';
 import type { ChatOptions } from '../src/ai/providers/base';
 import { OpenAICompatibleProvider } from '../src/ai/providers/openai-compatible';
@@ -13,6 +13,10 @@ const OPTIONS: ChatOptions = {
 };
 
 suite('AI Empty Response Handling', () => {
+  suiteSetup(() => {
+    setRetryBaseDelayForTest(1);
+  });
+
   test('parseJsonResponse throws empty-response error on blank input', () => {
     assert.throws(() => parseJsonResponse(''), /空响应|empty response/);
     assert.throws(() => parseJsonResponse('   \n\t '), /空响应|empty response/);
@@ -41,8 +45,10 @@ suite('AI Empty Response Handling', () => {
   });
 
   test('chatWithRetry propagates error when retry also returns empty', async () => {
+    let calls = 0;
     const provider = {
       chat: async (): Promise<string> => {
+        calls++;
         throw new EmptyContentError('finish_reason=length');
       },
     } as unknown as AIProvider;
@@ -51,9 +57,10 @@ suite('AI Empty Response Handling', () => {
       () => chatWithRetry(provider, 'sys', 'user', OPTIONS),
       EmptyContentError
     );
+    assert.strictEqual(calls, 4);
   });
 
-  test('chatWithRetry does not retry on non-empty-content errors', async () => {
+  test('chatWithRetry does not retry on non-retryable errors', async () => {
     let calls = 0;
     const provider = {
       chat: async (): Promise<string> => {
@@ -66,6 +73,121 @@ suite('AI Empty Response Handling', () => {
     assert.strictEqual(calls, 1);
   });
 
+  test('chatWithRetry retries on 429 and succeeds', async () => {
+    let calls = 0;
+    const provider = {
+      chat: async (): Promise<string> => {
+        calls++;
+        if (calls < 3) {
+          throw new ApiRequestError('API 请求失败 (429): rate limited', 429);
+        }
+        return '{"findings":[]}';
+      },
+    } as unknown as AIProvider;
+
+    const result = await chatWithRetry(provider, 'sys', 'user', OPTIONS);
+    assert.strictEqual(result, '{"findings":[]}');
+    assert.strictEqual(calls, 3);
+  });
+
+  test('chatWithRetry retries 429 up to 3 times then throws', async () => {
+    let calls = 0;
+    const provider = {
+      chat: async (): Promise<string> => {
+        calls++;
+        throw new ApiRequestError('API 请求失败 (429): rate limited', 429);
+      },
+    } as unknown as AIProvider;
+
+    await assert.rejects(
+      () => chatWithRetry(provider, 'sys', 'user', OPTIONS),
+      (err: unknown) => err instanceof ApiRequestError
+    );
+    assert.strictEqual(calls, 4);
+  });
+
+  test('chatWithRetry fails immediately on 401', async () => {
+    let calls = 0;
+    const provider = {
+      chat: async (): Promise<string> => {
+        calls++;
+        throw new ApiRequestError('API Key 无效', 401);
+      },
+    } as unknown as AIProvider;
+
+    await assert.rejects(
+      () => chatWithRetry(provider, 'sys', 'user', OPTIONS),
+      (err: unknown) => err instanceof ApiRequestError
+    );
+    assert.strictEqual(calls, 1);
+  });
+
+  test('chatWithRetry retries on network error (TypeError)', async () => {
+    let calls = 0;
+    const provider = {
+      chat: async (): Promise<string> => {
+        calls++;
+        if (calls === 1) {
+          throw new TypeError('fetch failed');
+        }
+        return '{"findings":[]}';
+      },
+    } as unknown as AIProvider;
+
+    const result = await chatWithRetry(provider, 'sys', 'user', OPTIONS);
+    assert.strictEqual(result, '{"findings":[]}');
+    assert.strictEqual(calls, 2);
+  });
+
+  test('chatWithRetry retries on timeout (AbortError)', async () => {
+    let calls = 0;
+    const provider = {
+      chat: async (): Promise<string> => {
+        calls++;
+        if (calls === 1) {
+          throw new DOMException('The operation was aborted', 'AbortError');
+        }
+        return '{"findings":[]}';
+      },
+    } as unknown as AIProvider;
+
+    const result = await chatWithRetry(provider, 'sys', 'user', OPTIONS);
+    assert.strictEqual(result, '{"findings":[]}');
+    assert.strictEqual(calls, 2);
+  });
+
+  test('chatWithRetry backoff is exponential', async () => {
+    setRetryBaseDelayForTest(10);
+    try {
+      const provider = {
+        chat: async (): Promise<string> => {
+          throw new ApiRequestError('API 请求失败 (503)', 503);
+        },
+      } as unknown as AIProvider;
+
+      const start = Date.now();
+      await assert.rejects(() => chatWithRetry(provider, 'sys', 'user', OPTIONS));
+      const elapsed = Date.now() - start;
+      assert.ok(elapsed >= 70, `expected backoff >= 70ms, got ${elapsed}ms`);
+      assert.ok(elapsed < 5000, `backoff too long: ${elapsed}ms`);
+    } finally {
+      setRetryBaseDelayForTest(1);
+    }
+  });
+
+  test('isRetryableError classifies error types', () => {
+    assert.strictEqual(isRetryableError(new EmptyContentError('x')), true);
+    assert.strictEqual(isRetryableError(new ApiRequestError('x', 429)), true);
+    assert.strictEqual(isRetryableError(new ApiRequestError('x', 500)), true);
+    assert.strictEqual(isRetryableError(new ApiRequestError('x', 503)), true);
+    assert.strictEqual(isRetryableError(new ApiRequestError('x', 401)), false);
+    assert.strictEqual(isRetryableError(new ApiRequestError('x', 400)), false);
+    assert.strictEqual(isRetryableError(new ApiRequestError('x')), false);
+    assert.strictEqual(isRetryableError(new TypeError('fetch failed')), true);
+    assert.strictEqual(isRetryableError(new DOMException('abort', 'AbortError')), true);
+    assert.strictEqual(isRetryableError(new Error('boom')), false);
+  });
+
   test('openai-compatible reports max_tokens truncation clearly', async () => {
     const originalFetch = globalThis.fetch;
     globalThis.fetch = (async () => ({

tests/customFixEngine.test.ts

@@ -83,7 +83,7 @@ suite('Custom FixEngine Tests', () => {
     assert.strictEqual(result.appliedFixes.length, 1);
   });
 
-  test('aiFixReviewIssue retries empty fix once then fails with ai-no-fix', async () => {
+  test('aiFixReviewIssue fails with ai-no-fix when AI provides empty fix', async () => {
     const doc = await vscode.workspace.openTextDocument({
       content: 'const x = BAD;\n',
       language: 'javascript',