wangdl/api-server
2
0
Fork 0
Code Issues 28 Pull Requests Actions Packages Projects Releases Wiki Activity
408 Commits 1 Branch 0 Tags
Commit Graph

12 Commits

Author SHA1 Message Date
wangdl
cdcf056ba5 fix: B1 — lockJob 原子递增 attemptCount + Engine 重试解锁逻辑
Some checks failed
Deploy API Server / build-and-unit (push) Failing after 28s
Details
Deploy API Server / current-integration (push) Has been skipped
Details
Deploy API Server / backward-compat (push) Has been skipped
Details
Deploy API Server / deploy (push) Has been skipped
Details 
- lockJob updateMany 新增 attemptCount: { increment: 1 }
- Engine retryable 错误: unlockForRetry() 重置为 queued 后再抛给 BullMQ
- Engine permanent 错误: 直接 markFailed + return（不抛错，BullMQ 不重试）
- unlockForRetry: 重置 lifecycleStatus→queued + 清除 lockedBy/lockedAt/lockUntil
- 更新测试: lockJob 验证 attemptCount 递增; schema validation 验证不抛错

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 18:42:19 +08:00
wangdl
d82ae8ba71 feat: M-AI-03 统一 AiJob Worker Execution Engine
Some checks failed
Deploy API Server / build-and-unit (push) Failing after 29s
Details
Deploy API Server / current-integration (push) Has been skipped
Details
Deploy API Server / backward-compat (push) Has been skipped
Details
Deploy API Server / deploy (push) Has been skipped
Details 
- AiJobExecutionEngineImpl: 完整 5 阶段管线
  PREPARE: read Job → terminal check → cancel check → Registry.get → CAS lockJob
  RESOLVE: load Snapshot → schemaVersion check → AbortController timeout
  EXECUTE: AiGatewayService.generate() (不直接导入 Provider SDK)
  PROJECT: write validatedOutput + outputHash (#292 替换为 Projector)
  COMPLETE: usageLog → markSucceeded
- 错误分类: classifyError() — 7 种错误码, retryable/permanent 分流
  provider_rate_limited/timeout/unavailable → retryable → 抛给 BullMQ
  schema/business/reference validation → permanent → markFailed
  cancelled → markCancelled (3 个检查点: 执行前/Provider前/Projector前)
- Usage logging: 每次 Provider 调用写 AiUsageLog (jobId/attemptNo/credentialMode/token)
- AiJobModule: 使用 AI_JOB_EXECUTION_ENGINE token 注册，替换 #288 的 no-op
- 10 个单元测试: 全链路成功/终态跳过/取消(2)/错误分类(3)/锁冲突/Schema不兼容/UsageLog

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 18:30:02 +08:00
wangdl
471aa2b872 fix: #289 检修 — TOCTOU 幂等 + promptKey 传递 + require 清理
Some checks failed
Deploy API Server / build-and-unit (push) Failing after 29s
Details
Deploy API Server / current-integration (push) Has been skipped
Details
Deploy API Server / backward-compat (push) Has been skipped
Details
Deploy API Server / deploy (push) Has been skipped
Details 
- 幂等保护从 findUnique 预检改为 P2002 catch（消除 TOCTOU 竞态窗口）
- AiJobLifecycleRepository: CreateJobInput 新增 promptKey，createJob 写入
- AiJobCreationService: 从 JobDefinition 传递 promptKey 到 Job
- computeHash: 内联 require('crypto') → 顶层 import * as crypto
- 测试更新: 幂等场景改为 P2002 冲突 → findUniqueOrThrow 返回已有 Job

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 18:22:39 +08:00
wangdl
6668bba9b4 feat: M-AI-03 Outbox Dispatcher — 可靠投递 ai.job.enqueue 到 BullMQ
Some checks failed
Deploy API Server / build-and-unit (push) Failing after 29s
Details
Deploy API Server / current-integration (push) Has been skipped
Details
Deploy API Server / backward-compat (push) Has been skipped
Details
Deploy API Server / deploy (push) Has been skipped
Details 
- OutboxDispatcher: 每 5s 轮询 OutboxEvent 表，CAS 领取 → 验证 → BullMQ 投递 → 标记 published
- 采用 ADR-003 §4.3 先 CAS 后投递算法：markProcessing(CAS) → queue.add() → markPublished
- BullMQ jobId = AiJob.id（幂等重复 add）
- 崩溃窗口保护：bullmq jobId 幂等 + releaseExpiredLocks(60s)
- 错误处理：永久失败（未知eventType/Job不存在/非法queueName/超过最大尝试）
  vs 暂时失败（exponential backoff 回退 pending）
- 仅在 WorkerModule 注册（API 进程不运行 Dispatcher）

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 18:20:22 +08:00
wangdl
f0183c03c4 feat: M-AI-03 事务化 AiJob 创建、Snapshot 与 Enqueue Outbox
Some checks failed
Deploy API Server / build-and-unit (push) Failing after 30s
Details
Deploy API Server / current-integration (push) Has been skipped
Details
Deploy API Server / backward-compat (push) Has been skipped
Details
Deploy API Server / deploy (push) Has been skipped
Details 
- AiJobCreationService: 在一个 Prisma Transaction 内原子创建 Job+Snapshot+Outbox
- 幂等性: (userId, jobType, idempotencyKey) UNIQUE 约束防重复
- Job 元数据从 JobDefinition 写入 (queueName/priority/maxAttempts/timeoutMs 等)
- overrides 支持覆盖 priority/credentialMode/credentialId
- Outbox payload 仅含 {jobId}，不含 userId/credentialId/snapshot
- 依赖 SnapshotBuilderService + OutboxRepository，AiJobModule 导入 AiRuntimeModule

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 18:13:15 +08:00
wangdl
fcce35742f feat: M-AI-03 注册新队列 ai-interactive/ai-background 与薄 Worker Processor
Some checks failed
Deploy API Server / build-and-unit (push) Failing after 29s
Details
Deploy API Server / current-integration (push) Has been skipped
Details
Deploy API Server / backward-compat (push) Has been skipped
Details
Deploy API Server / deploy (push) Has been skipped
Details 
- queue.constants.ts: 新增 QUEUE_AI_INTERACTIVE / QUEUE_AI_BACKGROUND
- queue-definitions.ts: 队列配置对齐 ADR-003 §3.1
  ai-interactive: concurrency=2, attempts=2, backoff=2000ms
  ai-background: concurrency=1, attempts=3, backoff=5000ms
- queue.service.ts: QueueService 注入并路由新队列
- queue.module.ts: BullModule.registerQueue 注册新队列
- AiJobExecutionEngine 接口 + AI_JOB_EXECUTION_ENGINE 注入 Token
- AiInteractiveJobWorker + AiBackgroundJobWorker: 薄 Processor，仅接收 {jobId} 调用 Engine
- worker.module.ts: 注册两个新 Processor（#291 提供 Engine 真实实现前为 no-op）

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 18:02:26 +08:00
wangdl
373e5e09fa refactor: ALLOWED_CREDENTIAL_MODES 集中到 job-definition.types.ts
All checks were successful
Deploy API Server / build-and-unit (push) Successful in 32s
Details
Deploy API Server / current-integration (push) Successful in 29s
Details
Deploy API Server / backward-compat (push) Successful in 0s
Details
Deploy API Server / deploy (push) Successful in 1m0s
Details 
消除 Registry 中的重复运行时数组，改为从 types 文件导入。
类型定义与运行时数组同源，避免异步风险。

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 17:28:58 +08:00
wangdl
e4b922e7e4 feat: M-AI-03 Job Definition Registry 与注册校验
All checks were successful
Deploy API Server / build-and-unit (push) Successful in 32s
Details
Deploy API Server / current-integration (push) Successful in 28s
Details
Deploy API Server / backward-compat (push) Successful in 0s
Details
Deploy API Server / deploy (push) Successful in 58s
Details 
- JobDefinition 接口冻结（job-definition.types.ts）：对齐 ADR-003 §2.1，
  涵盖 queue/execution/input/output/prompt/model/credential/security
- JobDefinitionRegistry：启动时注册、运行时只读、Object.freeze 防篡改
- 10 类校验错误：DuplicateJobType/UnknownJobType/InvalidJobType/
  InvalidQueueName/MissingSchema/InvalidTimeout/InvalidRetries/
  InvalidCredentialMode/InvalidPromptKey/MissingBackoff
- 43 个单元测试覆盖所有校验路径 + 顺序稳定性 + 禁止覆盖
- 更新 AiJobModule 注册新 Provider

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 17:24:57 +08:00
wangdl
4c736f3658 fix: Repository 层移除 modelProvider/modelName 硬编码默认值
All checks were successful
Deploy API Server / build-and-unit (push) Successful in 33s
Details
Deploy API Server / current-integration (push) Successful in 29s
Details
Deploy API Server / backward-compat (push) Successful in 0s
Details
Deploy API Server / deploy (push) Successful in 1m3s
Details 
modelProvider/modelName 默认值从 Repository 移至 Service/JobDefinition 层。
若调用方未传入，依赖 Prisma schema @default 保证列非空。
避免 AiJobService 忘记传值时静默使用错误 provider。

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 17:22:11 +08:00
wangdl
333632dc5b test: 补充 7 个缺失测试用例，覆盖所有 CAS 失败和只读路径
Some checks failed
Deploy API Server / build-and-unit (push) Successful in 33s
Details
Deploy API Server / backward-compat (push) Has been cancelled
Details
Deploy API Server / deploy (push) Has been cancelled
Details
Deploy API Server / current-integration (push) Has been cancelled
Details 
新增用例：
- markSucceeded CAS 失败（非终态 queued）→ JobNotActiveError
- markFailed CAS 失败（已终态）→ JobAlreadyTerminalError
- markFailed 接受可选 tx 参数
- extendLock CAS 失败（Job 不在 running）→ JobNotActiveError
- requestCancellation 终态 Job → JobNotCancellableError
- findExpiredLocks 过期锁查询
- findById 按 ID 查询
总测试数：43 → 50

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 17:21:22 +08:00
wangdl
38d1c4f414 fix: requestCancellation TOCTOU 竞态 — update 改为 updateMany CAS
Some checks failed
Deploy API Server / build-and-unit (push) Successful in 33s
Details
Deploy API Server / current-integration (push) Successful in 29s
Details
Deploy API Server / deploy (push) Has been cancelled
Details
Deploy API Server / backward-compat (push) Has been cancelled
Details 
- running 路径：prisma.aiJob.update() → updateMany() CAS
- CAS 失败时重读当前状态：已取消则幂等返回，已终态则抛出 JobNotCancellableError
- 新增 2 个 TOCTOU 场景测试（已变 cancelled → 幂等，已变 succeeded → 报错）

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 17:20:15 +08:00
wangdl
b08d27a449 feat: M-AI-03 AiJob 状态机与原子状态仓储
All checks were successful
Deploy API Server / build-and-unit (push) Successful in 33s
Details
Deploy API Server / current-integration (push) Successful in 29s
Details
Deploy API Server / backward-compat (push) Successful in 0s
Details
Deploy API Server / deploy (push) Successful in 1m4s
Details 
- AiJobStateMachine: 5 状态（queued/running/succeeded/failed/cancelled）、
  合法/非法迁移验证、终态判断、双向 Shadow Write 映射
- AiJobLifecycleRepository: 唯一 lifecycleStatus 写入入口、
  CAS（条件 UPDATE WHERE lifecycleStatus IN (...)）抢锁/续锁/标记成功/失败/取消、
  可选 tx 参数支持 Projector 事务内调用
- requestCancellation: queued 直接取消 → cancelled、
  running 设置 cancelRequestedAt 信号
- 6 个类型化错误（JobLockConflict/JobNotActive/JobAlreadyTerminal/
  JobNotCancellable/JobExtendLock/IllegalStateTransition）
- 41 个单元测试全部通过

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-06-20 17:17:12 +08:00