[{"data":1,"prerenderedAt":268},["ShallowReactive",2],{"\u002Fblog\u002Fmodel-harness-decoupling\u002F:zh":3,"\u002Fblog\u002Fmodel-harness-decoupling\u002F-surround:zh":163,"blog-all-posts:zh":164},{"id":4,"title":5,"author":6,"body":10,"date":145,"description":146,"extension":147,"image":148,"locale":149,"meta":150,"minRead":151,"navigation":152,"original":153,"path":154,"pinned":153,"seo":155,"source":156,"sourceUrl":23,"stem":157,"tags":158,"updated":148,"__hash__":162},"blog\u002Fblog\u002Fmodel-harness-decoupling.md","原厂 Harness 未必是版本答案：开放模型时代，模型和 Harness 要分开选",{"name":7,"avatar":8},"陈大黄",{"src":9,"alt":7},"\u002Favatar.jpg",{"type":11,"value":12,"toc":136},"minimark",[13,28,31,34,38,41,44,48,51,54,70,73,77,80,83,86,90,93,100,106,112,118,124,127,130],[14,15,16],"blockquote",{},[17,18,19,20,27],"p",{},"核心观点首发于 ",[21,22,26],"a",{"href":23,"rel":24},"https:\u002F\u002Fx.com\u002Frealchendahuang\u002Fstatus\u002F2093890559874388141",[25],"nofollow","X","，这篇文章是完整版论述，附五家 Harness 的定位地图。",[17,29,30],{},"很多人选 Coding Agent 的第一反应都很自然：用 Claude，就上 Claude Code；用 GPT，就上 Codex；用 GLM，就上 ZCode；用 DeepSeek，当然优先考虑 DeepSeek 自己的 Harness。",[17,32,33],{},"这个思路其实完全合理。原厂最大的优势，就是最了解自己的模型。",[35,36,37],"h2",{"id":37},"原厂的优势是真的",[17,39,40],{},"模型喜欢什么样的 Prompt，Tool Schema 怎么设计最稳定，长上下文怎么组织，新版本增加了什么能力，哪些地方最容易翻车——原厂通常都比第三方更早知道。",[17,42,43],{},"所以 Claude Code 和 Codex 这种模型与 Harness 一起长期迭代的产品，原厂组合往往就是很强的版本答案。这一点我不反驳。",[35,45,47],{"id":46},"但训练模型和做-harness是两门完全不同的工程","但训练模型和做 Harness，是两门完全不同的工程",[17,49,50],{},"到了 DeepSeek、GLM 这些开放模型，事情就开始变得有意思了。因为训练一个好模型，和做一个好 Harness，其实是两门完全不同的工程。",[17,52,53],{},"Coding Agent 真正跑起来以后，还有大量模型之外的问题：",[55,56,57,61,64,67],"ul",{},[58,59,60],"li",{},"文件怎么读、代码怎么改",[58,62,63],{},"Agent Loop 怎么控制、Context 怎么压缩",[58,65,66],{},"缓存怎么利用、Tool Call 出错以后怎么恢复",[58,68,69],{},"Subagent 怎么调度、权限怎么管理",[17,71,72],{},"这些地方做得好不好，会直接影响同一个模型最后到底好不好用。同一个模型，换个 Harness，体验可以是天壤之别。",[35,74,76],{"id":75},"更大的问题模型更新速度太快了","更大的问题：模型更新速度太快了",[17,78,79],{},"今天 GLM 强，下个月 DeepSeek 可能又出一个更能打的 Flash，再过一阵子又有新的模型追上来。",[17,81,82],{},"如果整个 Coding 工作流都绑定在某一家原厂产品上，换模型往往连工具和习惯也要跟着换。你花了几个月调教出来的配置、记忆、工作流，全部推倒重来。",[17,84,85],{},"第三方 Harness 的价值就在这里：**你可以把自己熟悉的工具、Skills、MCP、权限和工作流固定下来，只替换底下的模型。**今天跑 DeepSeek，明天换 GLM，后天再换别的，工作环境不用推倒重来。",[35,87,89],{"id":88},"五家-harness-定位地图","五家 Harness 定位地图",[17,91,92],{},"把主流五家的定位说清楚（截至 2026 年 8 月底我的使用体验）：",[17,94,95,99],{},[96,97,98],"strong",{},"Pi","：思路极简，Harness 尽量少干预模型，轻、快、Token 开销低，可塑性极强。适合做自己的长期 Agent 基座，简单、干净、随便魔改。",[17,101,102,105],{},[96,103,104],{},"OMP","：在 Pi 上继续堆 LSP、Debugger、Browser、AST 这些重型 Coding 能力，像给 Agent 装了一套完整 IDE。适合真正重度 Coding、需要复杂 Repo 导航的场景。",[17,107,108,111],{},[96,109,110],{},"DeepSeek Harness","：走得最远，Everything is Plugin，Agent Loop、工具、权限、Preset、UI 都可以拆开重组。适合折腾 Agent 架构、Preset、多 Agent 和下一代 Runtime 的人。多试 PTC 模式，速度更快、更省 token。",[17,113,114,117],{},[96,115,116],{},"OpenCode","：目前最均衡的一类，开源、Provider 多、生态大、Client\u002FServer、桌面端、Subagent 都比较成熟。适合想要一套成熟通用型多模型 Coding Agent 的人。",[17,119,120,123],{},[96,121,122],{},"Command Code","：路子完全不同——它特别喜欢替模型补位。Tool Call 参数写错了本地修；文件重复读帮你去重；长 Session 维持 Stable Prefix 提高 Cache Hit；Context 快爆了就做 Compaction。这套思路放到 DeepSeek V4 Flash、GLM-5.3 Flash 这种苦力模型上，价值最大：模型差一点，Harness 给你补。",[35,125,126],{"id":126},"我的选法",[17,128,129],{},"如果只看长期可塑性，我依然更喜欢 Pi。但如果今天就让我拿 DeepSeek V4 Flash、GLM-5.3 狠狠干活，我会认真尝尝 Command Code 的咸淡——模型和 Harness 的组合是按任务配的，不是按阵营站的。",[17,131,132,133],{},"到了开放模型时代，模型和 Harness 已经完全可以分开选了。",[96,134,135],{},"别再问「用谁家的模型该用谁家的工具」，改问「这个模型放在哪个 Harness 里，能把活干得最好」。",{"title":137,"searchDepth":138,"depth":138,"links":139},"",2,[140,141,142,143,144],{"id":37,"depth":138,"text":37},{"id":46,"depth":138,"text":47},{"id":75,"depth":138,"text":76},{"id":88,"depth":138,"text":89},{"id":126,"depth":138,"text":126},"2026-09-09","用 Claude 就上 Claude Code？开放模型时代这个直觉该升级了。原厂的优势是真的，但训练好模型和做好 Harness 是两门工程——这篇文章讲清楚为什么模型与 Harness 已经可以分开选，以及五家 Harness 的定位地图。","md",null,"zh",{},8,true,false,"\u002Fblog\u002Fmodel-harness-decoupling",{"title":5,"description":146},"x","blog\u002Fmodel-harness-decoupling",[159,160,161],"AI Agent","Harness","选型","Hm3kYuIlRwgNwuhCmKhjosJJQGiOTYP8p72Ew7mHE0c",[148,148],[165,173,182,191,198,205,207,214,222,228,235,244,251,259],{"path":166,"title":167,"description":168,"date":169,"minRead":170,"tags":171},"\u002Fblog\u002Fagent-harness-selection","折腾了 Pi Agent、OMP、Codex、ZCode，我为什么最终选了 OpenCode + OpenChamber","折腾 Agent Harness 的选型复盘：GUI 体验、供应商锁定、二次开发自由度三个标准，排除了 Pi Agent、OMP、Codex、ZCode，最后选了 OpenCode 内核 + OpenChamber 界面。","2026-08-06",6,[159,172],"工具选型",{"path":174,"title":175,"description":176,"date":177,"minRead":170,"tags":178},"\u002Fblog\u002Fcloudflare-broke-stack","2026 独立开发最佳实践：Cloudflare 穷鬼全家桶","独立开发者的零成本技术栈：Codex 写代码、GitHub 管版本、Stripe 收钱，前端 TanStack Start、后端 Hono + Workers、数据库 D1、存储 R2、缓存 KV，全部跑在 Cloudflare 上。","2026-06-15",[179,180,181],"Cloudflare","独立开发","技术栈",{"path":183,"title":184,"description":185,"date":186,"minRead":170,"tags":187},"\u002Fblog\u002Fdeepseek-api-web-search","DeepSeek API 内置联网搜索，Responses API 白嫖官方搜索能力","DeepSeek 官方在 API 里内置了联网搜索：用 Responses 接口调用 deepseek-v4-flash，声明 web_search 工具即可，不用对接第三方搜索引擎，不用申请搜索密钥。","2026-08-05",[188,189,190],"DeepSeek","API","AI 工具",{"path":192,"title":193,"description":194,"date":186,"minRead":195,"tags":196},"\u002Fblog\u002Fdeepseek-v4-flash-review","DeepSeek V4 Flash 正式版深度体验：便宜、快、1M 上下文、内置搜索","深度体验几天 DeepSeek V4 Flash 正式版：极致的便宜、快如闪电、1M 上下文、官方内置联网搜索、完全开源。唯一的短板是多模态，但可以组合其他模型补上。",7,[188,197,190],"模型评测",{"path":199,"title":200,"description":201,"date":177,"minRead":151,"tags":202},"\u002Fblog\u002Ffree-cloudflare","免费用户如何榨干 Cloudflare，免费版到底能白嫖到什么程度？","Cloudflare 免费版能撑起一整套个人互联网基础设施：DNS、CDN、Pages、Workers、KV、D1、R2、邮箱、Tunnel、AI Gateway 等。",[179,203,204],"免费额度","部署",{"path":154,"title":5,"description":146,"date":145,"minRead":151,"tags":206},[159,160,161],{"path":208,"title":209,"description":210,"date":145,"minRead":151,"tags":211},"\u002Fblog\u002Fopencode-productive-stack","稳定、快速、高产、便宜：我的 AI Coding 全家桶实录","OpenCode + OpenChamber + 两份 DeepSeek V4 Flash 订阅，15 个项目同时跑代码，额度只掉一点点。这篇文章摊开我的完整配置：上下文剪枝、分层记忆、桌面自动化，以及「原汤化原食」的选型教训。",[116,212,213],"AI 编程","配置",{"path":215,"title":216,"description":217,"date":218,"minRead":219,"tags":220},"\u002Fblog\u002Fproduct-faxin-principle","发心原理：独立开发者为什么要学会舍弃","你为什么出发？你为谁出发？你到底想解决哪一个问题？产品锋利感来自舍得和放弃。","2026-06-20",10,[221,180],"产品思维",{"path":223,"title":224,"description":225,"date":218,"minRead":151,"tags":226},"\u002Fblog\u002Fproduct-pain-vs-itch","用户说「牛逼」别上头：痛点与痒点","口头喜欢一文不值。判断需求只有一个硬标准：看用户愿意付出什么。",[221,227],"需求判断",{"path":229,"title":230,"description":231,"date":169,"minRead":232,"tags":233},"\u002Fblog\u002Fqwen-vision-for-deepseek","给 DeepSeek 补上多模态：Qwen-3.7-Flash 识图方案","DeepSeek V4 Flash 没有多模态，识图怎么办？调研后目前性价比最高的方案是 Qwen-3.7-Flash：识别一张图片成本极低，和 V4 Flash 组合使用补齐短板。",5,[188,234,197],"多模态",{"path":236,"title":237,"description":238,"date":239,"minRead":240,"tags":241},"\u002Fblog\u002Fsub-store-cloudflare","我把机场订阅聚合搬到了 Cloudflare 上","多个机场加自建节点揉成一条订阅，分流规则在服务端配好，客户端只管订阅。","2026-06-28",4,[179,242,243],"科学上网","开源项目",{"path":245,"title":246,"description":247,"date":145,"minRead":195,"tags":248},"\u002Fblog\u002Ftui-cognitive-bandwidth","TUI 正在杀死你的认知带宽：AI Coding 的「极客滤镜」该碎了","一堆 AI Coding Agent 争相上 TUI，把交互塞回 80 年代的终端范式，还美其名曰「沉浸式」「尊重程序员」。这篇文章拆解 TUI 系统性降低认知带宽的三个机制，以及 Web UI 为什么才是正解。",[212,249,250],"交互设计","随笔",{"path":252,"title":253,"description":254,"date":255,"minRead":170,"tags":256},"\u002Fblog\u002Fvalue-not-external","讨论有没有价值，这件事本身就没有价值","我们的价值不由任何人定义。人在 AI 时代，到底应该如何理解自己。","2026-06-18",[257,258],"AI 思考","人文随笔",{"path":260,"title":261,"description":262,"date":263,"minRead":264,"tags":265},"\u002Fblog\u002Fx-growth-1000-followers","一天多涨粉 1000+，我做了什么","6 月认真运营 X，一天多涨粉 1000+。这篇记录我当时做过的事和得到的反馈。","2026-06-16",12,[266,267],"X 运营","内容增长",1788977132530]