DeepSeek Harness 与 Cordis：热插拔、服务共享和并发问答
工具可以随时换，正在使用工具的插件怎么办？从一个视觉服务的例子，讲清 Cordis 的依赖、清理、共享和并发，以及 Python、Go 能否实现同样的机制。
Test platforms & engineering
Layered test SDKs, testability API design, flaky-test root-cause analysis, and the engineering habits, learned on money-movement systems, that keep software maintainable.
Articles
从零到一搭建 AI 测试工作流：从需求到可执行测试
用一个可运行的 Python 工程，逐步建立需求识别、规则审阅、测试剧本生成、骨架填充和结果验证流程，附完整 Prompt、Skill 和公司接入方法。
AI testing workflow: from API material to acceptance
Explain how requirements, domain rules, scenarios, and existing SDKs connect in testing delivery, including where failures should return for correction.
Payments testing SDK: sharing requests without erasing business differences
Use multi-business reuse and existing subclass compatibility to explain request layering, calling contracts, and the limits of return-code retries.
XAT concurrency: method isolation, fixture scope, and deadline boundaries
Compare method and file subprocesses through shared fixtures, then follow discovery isolation, checkpoints, and parent-driven termination.
Who cleans up after test preparation fails? unittest and XAT
Follow partial setup, cleanup exceptions, and worker termination to understand cleanup registration, retained errors, and resource lifetime.
XAT reports: separating failure, missing results, and exit codes
Follow assertion failure, cleanup errors, and a worker crash through multiple issues, completeness, task identity checks, and command exit codes.
Agent 怎么做回归：从回答相似度到任务证据
用公开库存工程重放正常与缺陷版本，建立夹具、报告判定和版本比较，并区分验收链路检查与模型生成质量。
推理压测怎么测：从流式时间戳到满足 SLO 的吞吐
以测试工程的测量边界为起点，区分 TTFT、TPOT、ITL、开闭环负载与 goodput，避免把压力机限制和失败请求隐藏在漂亮报表里。
JSON 合法以后：结构化输出与工具执行契约
以日志查询和创建外部对象为例，区分 Schema、业务校验、权限和副作用状态，说明函数调用为什么不能替运行时完成验收。
XAT: making concurrent test results trustworthy
Follow worker crashes and missing results through discovery, process scheduling, fixtures, method deadlines, and report validation.
Related projects
Pythonunittestrequests
Follow discovery and report validation through concurrent failures, deadlines, and missing results.
Defined scope and execution behavior; collaborated with AI on implementation and code review.
PythonLLMAI Agent
Connect staged LLM generation, human review, and execution checks to turn generated code into useful test cases.
Build a staged LLM test-generation pipeline at work, combining AI-agent delivery with human review gates.
PythonAPISDK
Share request behavior across business APIs while handling protocol changes, existing callers, and retry conditions.
Contributed to APIs and led the merchant-funds SDK refactor and compatibility work.