Kane 的个人网站，记录 AI Coding Agent、Agent Harness、开发者工具与 AI 原生软件工程实践。https://kanelogger.comzh-CNhttps://kanelogger.com/writing/jev-agent-decision-modelhttps://kanelogger.com/writing/jev-agent-decision-modelTypeSafe 的 Jev 是面向 Agent 的决策模型，通过 Choice、Score、Noul 三类接口直接返回固定类型与概率，适合高频、选项可枚举的单步判断。Mon, 21 Sep 2026 00:00:00 GMTAI 模型JevAgent决策模型TypeSafehttps://kanelogger.com/writing/learning-pi-in-practicehttps://kanelogger.com/writing/learning-pi-in-practice按实践优先的顺序学习 Pi，并以三层 Harness 和分页搜索控制上下文成本。Mon, 31 Aug 2026 00:00:00 GMTPiCoding AgentHarness上下文工程https://kanelogger.com/writing/agent-project-recoveryhttps://kanelogger.com/writing/agent-project-recovery把代码仓库升级为 Agent 工作环境，使新会话仅靠仓库即可恢复项目事实、状态和正确流程。Sun, 30 Aug 2026 00:00:00 GMTCoding AgentAGENTS.mdAgent Harness上下文工程https://kanelogger.com/writing/agent-hard-gateshttps://kanelogger.com/writing/agent-hard-gates用可观察验收条件、风险测试、硬门禁和独立 Review 控制 Coding Agent 的高速推进。Sun, 30 Aug 2026 00:00:00 GMTCoding Agent阶段门禁验收测试策略https://kanelogger.com/writing/agent-token-economicshttps://kanelogger.com/writing/agent-token-economics拆解 Agent 全链路成本，区分并行的时间收益与 Token 账单，并给出工具和多 Agent 协作的选择方法。Sun, 30 Aug 2026 00:00:00 GMTCoding AgentToken 成本多 Agent模型路由https://kanelogger.com/writing/coding-agent-token-costhttps://kanelogger.com/writing/coding-agent-token-cost从成本结构、使用习惯和模型路由，到压缩工具、代码图谱与多 Agent 协作的完整方法。Fri, 07 Aug 2026 00:00:00 GMTCoding AgentToken 成本上下文压缩多 Agenthttps://kanelogger.com/writing/ai-native-development-systemhttps://kanelogger.com/writing/ai-native-development-system以仓库承载事实、过程、状态和经验，用可运行小版本、硬门禁与反馈回路控制 Agent 风险。Sun, 02 Aug 2026 00:00:00 GMTAI 原生开发Coding Agent软件工程反馈回路https://kanelogger.com/writing/ai-native-from-design-to-runninghttps://kanelogger.com/writing/ai-native-from-design-to-running以可执行原型、可运行小版本、阶段门禁和可观察反馈约束高速 Coding Agent。Fri, 31 Jul 2026 00:00:00 GMTAI 原生开发Coding Agent软件工程质量控制https://kanelogger.com/writing/coding-agent-landscapehttps://kanelogger.com/writing/coding-agent-landscape对 8 款 AI 编程 Agent 的架构、工程能力、企业门槛和 FDE 场景进行分层比较与决策分析。Thu, 04 Jun 2026 00:00:00 GMTAI 编程Agent工具选型Harness企业部署https://kanelogger.com/writing/role-agent-pipelinehttps://kanelogger.com/writing/role-agent-pipeline从角色边界、文档交接与缓存连续性分析多 Agent 流水线的成本，整理主 Agent、独立子任务和外部状态文件的维护型项目工作流。Wed, 29 Apr 2026 00:00:00 GMTAIAgent架构设计Multi-Agent上下文工程模型路由缓存https://kanelogger.com/writing/llm-structured-wikihttps://kanelogger.com/writing/llm-structured-wiki把传统 Obsidian 笔记迁入由 LLM 维护的结构化 Wiki，通过原始资料、知识网络和操作规范，形成录入、查询与持续维护的知识工作流。Tue, 07 Apr 2026 00:00:00 GMTAI知识管理ObsidianLLMWiki