同样调 Claude API,为什么 Claude Code、Cursor 各有各的手感?差异在 LLM 外面那层"壳"。我把这层壳从零造了出来(HarWork),并写成了 18 篇双语工程拆解 —— 每一处架构决策都有真实代码与数字支撑。现承接 Agent 基础设施 / LLM 应用工程 的咨询与开发。 Same Claude API — yet Claude Code, Cursor, Aider all feel different. The difference is the engineering shell around the LLM. I built that shell from scratch (HarWork) and documented it in 18 bilingual deep-dives, every decision backed by real code and numbers. Available for agent-infrastructure / LLM-application consulting & build work.
不是"会调 API",是把 LLM 变成断网能续、超窗不崩、删库被拦的生产系统。Not "I can call an API" — turning an LLM into a system that survives disconnects, never overflows, and refuses to rm -rf.
Agent Loop、工具协议、权限沙箱、Hooks、中断与恢复 —— 从零搭可维护的 harness,或审查你现有的。Agent loops, tool protocols, permission sandboxes, hooks, interrupt/resume — build a maintainable harness from scratch, or audit your existing one.
上下文压缩、多模型路由、流式 + WebSocket 韧性、token 成本治理 —— 让长对话不爆窗口、不烧钱。Context compaction, multi-model routing, streaming + WebSocket resilience, token-cost control — long conversations that don't overflow or burn cash.
会话存储、每用户持久 Docker、金丝雀发布 + 多探针自动回滚 —— 把 demo 变成扛得住真实流量的服务。Session storage, per-user persistent Docker, canary releases + multi-probe auto-rollback — turn a demo into a service that holds under real traffic.
每篇都从"朴素方案为什么不行"讲到真实代码。完整阅读地图 →Each part goes from "why the naive approach fails" to real code. Start reading →