Apache-2.0 通用型 Skill，自動路由本地 Skill 並編排 harness 工作流；作為本地 Skill 透過 $vibe / /vibe 在 DeepSeek Harness 中呼叫。
DSH 整合
相容
作者聲明
安全稽核
未稽核
最後核實
2026-08-21
授權條款
Apache-2.0
01它能幫你完成什麼？
Break a complex multi-step task into planned work units, route local Skills to each part, and check the combined result
Approved requirement, L/XL level plan, per-Skill assignments with expected deliveries, and a final acceptance check across all planned items
Users of harness-style AI agents who have accumulated many local Skills and want complex tasks planned and delivered end to end
Auto-discover and select the best-fitting local Skills for each part of a task
A shortlist of candidate Skills (from 100+ installed) with each selected Skill tied to concrete work, expected delivery, and a check
Teams with a growing personal / team / third-party Skill library who no longer want to remember which Skill to invoke when
Resume and review long-running work from durable task records
Task records storing requirement, plan, execution progress, and final check in one place, so a later session can continue and a review can compare plan vs result
Users running multi-session tasks who need continuation and an auditable delivery review
02如何將外掛接入 DeepSeek Harness？
作者未說明
03DSH 整合程度與能力邊界
Installed as a general-purpose local Skill and invoked through any harness-style agent's Skills entry ($vibe / /vibe); the README documents no DeepSeek Harness-specific setup
Skill discovery & routing
configured local Skill folders (Skills install directory plus folders in skill-roots.json)→local index and shortlisted candidates; only Skills written into the plan are read and executed
Discovery and index generation happen locally and write index/record filesExtra Skill folders are configured via `~/.vibeskills/skill-roots.json` or `<workspace>/.vibeskills/skill-roots.json`Planning state machine (L / XL)
task scope, steps, dependencies, and parallel-work opportunities→a recommended L or XL level and a modular work breakdown; XL can run up to two non-conflicting parts at the same time
Delivery acceptance checks
actual results vs every planned item→a final acceptance decision; incomplete, failed, or blocked required work prevents acceptance
Durable task records
one task's requirement, plan, execution, and final check→session_root, module-work-plan.json, module-execution.json, and delivery-acceptance-report files enabling resume and review
Writes record files to disk (session_root, module-work-plan.json, module-execution.json, delivery-acceptance-report.json/.md)install-receipt.json records the files written by the installer so `check` can find missing or changed files
04適合誰？何時不該用？
適合
- Users of harness-style AI agents who have accumulated many local Skills and want complex tasks planned and delivered end to end
- Teams with a growing personal / team / third-party Skill library who no longer want to remember which Skill to invoke when
- Users running multi-session tasks who need continuation and an auditable delivery review
不適合
- Orchestration adds model-context overhead: more retained candidates, longer Skill documents, or a more finely divided task use more tokens; the cost is bounded by a local index, candidate shortlisting, and on-demand reading.
05相容性、維護與安全提醒
- The README describes VibeSkills as a Skill for any AI application that supports local Skills and mentions harness frameworks similar to Superpower and GSD, but documents no DeepSeek Harness (dsh)-specific installation or integration path.
- A local Skill needs a readable SKILL.md, a non-conflicting name, and a clear fit for the current work before it can be selected; Skills without these stay out of the plan.
- Orchestration adds model-context overhead: more retained candidates, longer Skill documents, or a more finely divided task use more tokens; the cost is bounded by a local index, candidate shortlisting, and on-demand reading.
Apache-2.0 · actively maintained (latest release v4.0.0, 2026-07-17)
06常見問題
VibeSkills 如何接入 DeepSeek Harness？
VibeSkills 是面向任何支援本地 Skill 的 AI 應用程式的通用 Skill：把它安裝為本地 Skill 後，透過應用程式的 Skills 入口用 `$vibe`、`/vibe` 或應用程式提供的語法呼叫即可。README 未提供針對 DeepSeek Harness 的專門設定說明。
它是原生整合還是 MCP？
都不是——它是作為本地 Skill 透過宿主應用程式的 Skills 入口呼叫的，不走 MCP 協定。它遵循類似 Superpower 和 GSD 的 harness 框架模式。
使用前需要什麼前置條件？
需要一個支援本地 Skill 的 AI 應用程式，以及包含可讀 SKILL.md 檔案的本地 Skill 資料夾。額外的 Skill 資料夾可以寫在 `~/.vibeskills/skill-roots.json` 或 `<workspace>/.vibeskills/skill-roots.json` 中。
Skill 庫很大時會消耗很多 token 嗎？
發現與索引生成都在本地完成；只有被保留的候選 Skill 才會以完整 SKILL.md 形式讀入，執行階段只使用寫入計畫的 Skill。token 用量取決於候選數量、文件長度與任務複雜度——開銷有上限，但不為零。
任務紀錄和資料儲存在哪裡？
已核准的需求、計畫、執行進度與最終檢查都保存在同一份任務紀錄中（session_root、module-work-plan.json、module-execution.json、delivery-acceptance-report），安裝狀態單獨記錄在 install-receipt.json 中。
07相關的 DSH 工作流程
archify
作者 tt-a1i
為編碼智慧體生成美觀可驗證的架構圖、時序圖與資料流圖，輸出自包含 HTML，支援動效與清晰匯出。
openviking
作者 volcengine
為 AI 智慧體打造的自進化上下文資料庫，統一智慧體記憶、知識 RAG 與技能。
nocobase
作者 nocobase
開源的 AI + 無程式碼應用搭建平臺：AI 在久經生產驗證的底層基礎設施與所見即所得的無程式碼介面之上工作，幫你快速構建 CRM、ERP 等業務系統，兼顧速度與可靠性。
learn-harness-engineering
作者 walkinglabs
Harness 工程新手教程，從 0 到 1 系統學習智慧體工作流框架。
08資料與來源
VibeSkills is a general-purpose Skill that automatically routes local Skills and intelligently orchestrates harness work…
In any AI application that supports local Skills, invoke VibeSkills through the application's Skills entry, using <code>…
此頁面根據專案公開文件、儲存庫中繼資料與 DSH Plugins 的結構化解析所產生；最後核實於 2026-08-21。發現錯誤？提交更正。
最佳 DeepSeek Harness 外掛
從全目錄挑出的 12 個值得優先安裝的外掛，涵蓋各個分類。