Context window 會碎片化
模型可以讀取少量檔案,但在長期工程任務中,很難穩定維持完整依賴與架構脈絡。
AI 原生圖譜平台
把 Repository 轉換為可持久保存的 Symbol Graph、Repo Map 與唯讀 MCP Context,讓 AI Coding Agent 能在長期工程工作階段中穩定接續脈絡。
為什麼需要 CodeMind Graph
大型 Repository 會快速耗盡 context window,架構關係也常被檔案樹掩蓋。CodeMind Graph 把程式碼結構轉成可持久保存的知識層,讓人類與 Agent 都能穩定查詢。
模型可以讀取少量檔案,但在長期工程任務中,很難穩定維持完整依賴與架構脈絡。
相似度搜尋能協助探索,卻無法保證符號歸屬、import path 與 module boundary 的 deterministic 關係。
如果決策、Repo Map 與交接狀態沒有與目前 graph 對齊,Agent resume 後很容易依賴過期資訊。
把檔案、模組、符號、imports、exports 與 trace 轉為可定位的 graph entities,而不是鬆散文字片段。
Repository intelligence 在本機產生並保存於 `.codemind/`,兼顧隱私、資料主權與低延遲查詢。
Repo Map 與唯讀 MCP tools 讓 Agent 跨 session 保持穩定上下文,同時不暴露私有工程筆記。
Workspace graph
CodeMind Graph 將檔案、模組、符號、imports、exports 與 Agent Context 整合為 deterministic workspace layer。
工程實踐
CodeMind Graph 把結構化 context 視為一級工程產物。產品核心建立在 deterministic graph、隱私優先的本機處理,以及唯讀 Agent integration 之上。
Repository 同時為 Agent 與人類設計:指令輸出必須 deterministic、可 diff,且 context 必須結構化。
Source analysis 在開發者本機執行。除非專案擁有者明確發布 public demo,否則 graph data 不離開本機。
MCP tools 預設唯讀,不暴露 session notes、私有 docs、環境變數,也不提供寫入操作。
透過 persisted graph、Repo Map 與 docs-as-state,讓長期工程 session 能被 resume、review 與 audit。
架構標準
v0.1 架構以 TypeScript Compiler API adapter、deterministic core graph model、CLI-first query surface 與 read-only MCP server 為核心。
品牌定位
CodeMind Graph 為大型 Repository、AI Agent 與長期工程開發而設計,結合 graph-driven code intelligence 與 local-first safety。
Core capabilities
以節點與關係呈現專案結構,而不是把原始檔案直接塞給 Agent。
在修改大型程式碼前,先給 Agent 穩定的符號與依賴上下文。
把決策、狀態與 Repo Map 與私有原始碼曝光風險分離。
為 Implementer、Evaluator 與 Librarian 建立共享圖譜上下文。
讓 Repository 本身成為 Agent 可讀、可查詢、可接續的工程工作區。
從模糊搜尋轉向符號、模組與依賴關係的結構化導航。
Agent workflow
掃描 TypeScript 原始碼並寫入 .codemind/graph.json。
產生 CODEMIND.md,供人類與 Agent 共同閱讀。
透過唯讀 MCP tools 精準取回圖譜上下文。
用圖譜結構支援長期持久化工程工作階段。
Developer-first
CodeMind Graph 以 local-first 工程流程為核心。先產生圖譜、檢查符號,再把安全的唯讀上下文交給 Agent workflow。
pnpm check && node packages/cli/dist/index.js map --root examples/ts-basic --format markdownGitHub