[{"data":1,"prerenderedAt":268},["ShallowReactive",2],{"\u002Fblog\u002Fmodel-harness-decoupling\u002F:zh-Hant":3,"\u002Fblog\u002Fmodel-harness-decoupling\u002F-surround:zh-Hant":163,"blog-all-posts:zh-Hant":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.zh-Hant.md","原廠 Harness 未必是版本答案：開放模型時代，模型和 Harness 要分開選",{"name":7,"avatar":8},"陳大黃",{"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",{},"核心觀點首發於 ",[21,22,26],"a",{"href":23,"rel":24},"https:\u002F\u002Fx.com\u002Frealchendahuang\u002Fstatus\u002F2093890559874388141",[25],"nofollow","X","，這篇文章是完整版論述，附五家 Harness 的定位地圖。",[17,29,30],{},"很多人選 Coding Agent 的第一反應都很自然：用 Claude，就上 Claude Code；用 GPT，就上 Codex；用 GLM，就上 ZCode；用 DeepSeek，當然優先考慮 DeepSeek 自己的 Harness。",[17,32,33],{},"這個思路其實完全合理。原廠最大的優勢，就是最了解自己的模型。",[35,36,37],"h2",{"id":37},"原廠的優勢是真的",[17,39,40],{},"模型喜歡什麼樣的 Prompt，Tool Schema 怎麼設計最穩定，長上下文怎麼組織，新版本增加了什麼能力，哪些地方最容易翻車——原廠通常都比第三方更早知道。",[17,42,43],{},"所以 Claude Code 和 Codex 這種模型與 Harness 一起長期迭代的產品，原廠組合往往就是很強的版本答案。這一點我不反駁。",[35,45,47],{"id":46},"但訓練模型和做-harness是兩門完全不同的工程","但訓練模型和做 Harness，是兩門完全不同的工程",[17,49,50],{},"到了 DeepSeek、GLM 這些開放模型，事情就開始變得有意思了。因為訓練一個好模型，和做一個好 Harness，其實是兩門完全不同的工程。",[17,52,53],{},"Coding Agent 真正跑起來以後，還有大量模型之外的問題：",[55,56,57,61,64,67],"ul",{},[58,59,60],"li",{},"文件怎麼讀、代碼怎麼改",[58,62,63],{},"Agent Loop 怎麼控制、Context 怎麼壓縮",[58,65,66],{},"緩存怎麼利用、Tool Call 出錯以後怎麼恢復",[58,68,69],{},"Subagent 怎麼調度、權限怎麼管理",[17,71,72],{},"這些地方做得好不好，會直接影響同一個模型最後到底好不好用。同一個模型，換個 Harness，體驗可以是天壤之別。",[35,74,76],{"id":75},"更大的問題模型更新速度太快了","更大的問題：模型更新速度太快了",[17,78,79],{},"今天 GLM 強，下個月 DeepSeek 可能又出一個更能打的 Flash，再過一陣子又有新的模型追上來。",[17,81,82],{},"如果整個 Coding 工作流都綁定在某一家原廠產品上，換模型往往連工具和習慣也要跟著換。你花了几個月調教出來的配置、記憶、工作流，全部推倒重來。",[17,84,85],{},"第三方 Harness 的價值就在這裡：**你可以把自己熟悉的工具、Skills、MCP、權限和工作流固定下來，只替換底下的模型。**今天跑 DeepSeek，明天換 GLM，後天再換別的，工作環境不用推倒重來。",[35,87,89],{"id":88},"五家-harness-定位地圖","五家 Harness 定位地圖",[17,91,92],{},"把主流五家的定位說清楚（截至 2026 年 8 月底我的使用體驗）：",[17,94,95,99],{},[96,97,98],"strong",{},"Pi","：思路極簡，Harness 盡量少干預模型，輕、快、Token 開銷低，可塑性極強。適合做自己的長期 Agent 基座，簡單、乾淨、隨便魔改。",[17,101,102,105],{},[96,103,104],{},"OMP","：在 Pi 上繼續堆 LSP、Debugger、Browser、AST 這些重型 Coding 能力，像給 Agent 裝了一套完整 IDE。適合真正重度 Coding、需要複雜 Repo 導航的場景。",[17,107,108,111],{},[96,109,110],{},"DeepSeek Harness","：走得最遠，Everything is Plugin，Agent Loop、工具、權限、Preset、UI 都可以拆開重組。適合折騰 Agent 架構、Preset、多 Agent 和下一代 Runtime 的人。多試 PTC 模式，速度更快、更省 token。",[17,113,114,117],{},[96,115,116],{},"OpenCode","：目前最均衡的一類，開源、Provider 多、生態大、Client\u002FServer、桌面端、Subagent 都比較成熟。適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——它特別喜歡替模型補位。Tool Call 參數寫錯了本地修；文件重複讀幫你去重；長 Session 維持 Stable Prefix 提高 Cache Hit；Context 快爆了就做 Compaction。這套思路放到 DeepSeek V4 Flash、GLM-5.3 Flash 這種苦力模型上，價值最大：模型差一點，Harness 給你補。",[35,125,126],{"id":126},"我的選法",[17,128,129],{},"如果只看長期可塑性，我依然更喜歡 Pi。但如果今天就讓我拿 DeepSeek V4 Flash、GLM-5.3 狠狠幹活，我會認真嚐嚐 Command Code 的鹹淡——模型和 Harness 的組合是按任務配的，不是按陣營站的。",[17,131,132,133],{},"到了開放模型時代，模型和 Harness 已經完全可以分開選了。",[96,134,135],{},"別再問「用誰家的模型該用誰家的工具」，改問「這個模型放在哪個 Harness 裡，能把活幹得最好」。",{"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？開放模型時代這個直覺該升級了。原廠的優勢是真的，但訓練好模型和做好 Harness 是兩門工程——這篇文章講清楚為什麼模型與 Harness 已經可以分開選，以及五家 Harness 的定位地圖。","md",null,"zh-Hant",{},8,true,false,"\u002Fblog\u002Fmodel-harness-decoupling",{"title":5,"description":146},"x","blog\u002Fmodel-harness-decoupling.zh-Hant",[159,160,161],"AI Agent","Harness","選型","mAUa-uvfTnZesbwZvJpPumSZEnqv18Hzf9t1eX0a4ZE",[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","試遍了 Pi Agent、OMP、Codex、ZCode，我最後為什麼選了 OpenCode + OpenChamber","試過一輪 Agent Harness 後的選型復盤：以 GUI 體驗、供應商綁定、二次開發自由度三個標準，排除了 Pi Agent、OMP、Codex、ZCode，最後選了 OpenCode 核心 + OpenChamber 介面。","2026-08-06",6,[159,172],"工具選型",{"path":174,"title":175,"description":176,"date":177,"minRead":170,"tags":178},"\u002Fblog\u002Fcloudflare-broke-stack","2026 獨立開發最佳實踐：Cloudflare 窮鬼全家餐","獨立開發者的零成本技術棧：Codex 寫程式、GitHub 管版本、Stripe 收款，前端 TanStack Start、後端 Hono + Workers、資料庫 D1、儲存 R2、快取 KV，全部跑在 Cloudflare 上。","2026-06-15",[179,180,181],"Cloudflare","獨立開發","技術棧",{"path":183,"title":184,"description":185,"date":186,"minRead":170,"tags":187},"\u002Fblog\u002Fdeepseek-api-web-search","DeepSeek API 內建聯網搜尋，Responses API 白拿官方搜尋能力","DeepSeek 官方在 API 裡內建了聯網搜尋：用 Responses 介面呼叫 deepseek-v4-flash，宣告 web_search 工具即可，不用自己接第三方搜尋引擎，也不用申請搜尋金鑰。","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 正式版深度體驗：便宜、快、1M 上下文、內建搜尋","深度體驗了幾天 DeepSeek V4 Flash 正式版：極致的便宜、快如閃電、1M 上下文、官方內建聯網搜尋、完全開源。唯一的短板是多模態，但可以組合其他模型補上。",7,[188,197,190],"模型評測",{"path":199,"title":200,"description":201,"date":177,"minRead":151,"tags":202},"\u002Fblog\u002Ffree-cloudflare","免費使用者如何榨乾 Cloudflare，免費版到底能白嫖到什麼程度？","Cloudflare 免費版能撐起一整套個人網路基礎設施：DNS、CDN、Pages、Workers、KV、D1、R2、信箱、Tunnel、AI 閘道等。",[179,203,204],"免費額度","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","穩定、快速、高產、便宜：我的 AI Coding 全家桶實錄","OpenCode + OpenChamber + 兩份 DeepSeek V4 Flash 訂閱，15 個項目同時跑代碼，額度只掉一點點。這篇文章攤開我的完整配置：上下文剪枝、分層記憶、桌面自動化，以及「原湯化原食」的選型教訓。",[116,212,213],"AI 程式設計","配置",{"path":215,"title":216,"description":217,"date":218,"minRead":219,"tags":220},"\u002Fblog\u002Fproduct-faxin-principle","發心原理：獨立開發者為什麼要學會捨棄","你為什麼出發？你為誰出發？你到底想解決哪一個問題？產品的鋒利感來自捨得和放棄。","2026-06-20",10,[221,180],"產品思維",{"path":223,"title":224,"description":225,"date":218,"minRead":151,"tags":226},"\u002Fblog\u002Fproduct-pain-vs-itch","使用者說「很屌」別上頭：痛點與癢點","口頭喜歡一文不值。判斷需求只有一個硬標準：看使用者願意付出什麼。",[221,227],"需求判斷",{"path":229,"title":230,"description":231,"date":169,"minRead":232,"tags":233},"\u002Fblog\u002Fqwen-vision-for-deepseek","給 DeepSeek 補上多模態：Qwen-3.7-Flash 識圖方案","DeepSeek V4 Flash 沒有多模態，識圖怎麼辦？研究後目前 CP 值最高的方案是 Qwen-3.7-Flash：辨識一張圖片成本極低，和 V4 Flash 組合使用補齊短板。",5,[188,234,197],"多模態",{"path":236,"title":237,"description":238,"date":239,"minRead":240,"tags":241},"\u002Fblog\u002Fsub-store-cloudflare","我把機場訂閱聚合搬到了 Cloudflare 上","多個機場加自建節點合成一條訂閱，分流規則在伺服器端配好，用戶端只管訂閱。","2026-06-28",4,[179,242,243],"翻牆","開源專案",{"path":245,"title":246,"description":247,"date":145,"minRead":195,"tags":248},"\u002Fblog\u002Ftui-cognitive-bandwidth","TUI 正在殺死你的認知頻寬：AI Coding 的「極客濾鏡」該碎了","一堆 AI Coding Agent 搶著上 TUI，把互動塞回 80 年代的終端範式，還美其名曰「沉浸式」「尊重程式設計師」。這篇文章拆解 TUI 系統性降低認知頻寬的三個機制，以及 Web UI 為什麼才是正解。",[212,249,250],"互動設計","隨筆",{"path":252,"title":253,"description":254,"date":255,"minRead":170,"tags":256},"\u002Fblog\u002Fvalue-not-external","討論有沒有價值，這件事本身就沒有價值","我們的價值不由任何人定義。人在 AI 時代，到底應該如何理解自己。","2026-06-18",[257,258],"AI 思考","人文隨筆",{"path":260,"title":261,"description":262,"date":263,"minRead":264,"tags":265},"\u002Fblog\u002Fx-growth-1000-followers","一天多漲粉 1000+，我做了什麼","6 月認真經營 X，一天多漲粉 1000+。這篇記錄我當時做過的事和得到的回饋。","2026-06-16",12,[266,267],"X 經營","內容增長",1789752317936]