Codesight Ai Context

作者 reason-machines2384a003145a無授權條款83 個星標收錄於 2026年10月8日更新於 2026年10月8日儲存庫3 個月前更新

Universal AI context generator that compiles codebase maps, wiki knowledge bases, and MCP tools to save thousands of tokens per AI conversation.

AI 產生的概覽

產生精簡的程式碼庫脈絡地圖、wiki 知識庫與 AI 工具設定檔,以減少程式助理的 token 用量。

功能
CodeSight 會將程式碼庫編譯成結構化脈絡地圖,涵蓋路由、資料庫模型、UI 元件、相依關係與高影響檔案,也能產生持久化的 wiki 知識庫,或從 markdown 筆記建立知識地圖。它會產生 CLAUDE.md、.cursorrules、codex.md 與 AGENTS.md 等 AI 工具設定檔,並可作為 MCP 伺服器執行,提供脈絡、路由、模型、影響範圍與 wiki 存取等工具。它也會回報 token 節省情形,並可在檔案變更或提交時重新產生輸出。
適用情境
當 AI 程式助理需要專案脈絡、又不想把 token 花在手動瀏覽檔案時使用,也適合在對話開始時載入精簡概觀。重構共用檔案前查看影響範圍,或在儲存庫中保持脈絡檔案最新時同樣適用。
執行需求
需要 Node.js 18 或更新版本,透過 npx 或全域 npm 安裝執行。不需要 API 金鑰或設定檔。此技能資料夾不含指令碼;相關說明會呼叫外部 codesight 套件,MCP 模式需要設定 Claude Code 或 Cursor 等 MCP 用戶端。

CodeSight — AI Context Generator

Skill by ara.so — Daily 2026 Skills collection.

CodeSight compiles your codebase into a compact, structured context map (routes, models, components, dependencies) that AI coding assistants can read in one shot — eliminating thousands of tokens spent on manual file exploration. Supports 14 languages, 30+ frameworks, 13 ORM parsers, and an MCP server with 13 tools.

Installation

No installation required. Run directly with npx:

bash
npx codesight

Or install globally:

bash
npm install -g codesightcodesight

Requirements: Node.js >= 18, no API keys, no config files needed.

Core Commands

bash
# Generate context map (default — outputs .codesight/CODESIGHT.md)npx codesight
# Generate persistent wiki knowledge base (.codesight/wiki/)npx codesight --wiki
# Generate AI tool config files (CLAUDE.md, .cursorrules, codex.md, AGENTS.md)npx codesight --init
# Open interactive HTML report in browsernpx codesight --open
# Start as MCP server (13 tools) for Claude Code / Cursornpx codesight --mcp
# Show blast radius for a specific filenpx codesight --blast src/lib/db.ts
# Generate optimized config for a specific AI toolnpx codesight --profile claude-codenpx codesight --profile cursornpx codesight --profile copilotnpx codesight --profile codex
# Show token savings breakdownnpx codesight --benchmark
# Map markdown knowledge base (ADRs, meeting notes, Obsidian vault)npx codesight --mode knowledgenpx codesight --mode knowledge ~/vaultnpx codesight --mode knowledge ./docs
# Watch mode — regenerate on file changesnpx codesight --watch
# Git hook — regenerate on every commitnpx codesight --hook

What Gets Generated

Default Scan (npx codesight)

Outputs .codesight/CODESIGHT.md — a structured map including:

  • Project metadata (stack, language, framework, package manager)
  • All routes with HTTP methods and handler locations
  • Database models with fields and relations
  • UI components with props
  • High-impact files ranked by dependency count
  • Framework and ORM detection results

Wiki Knowledge Base (--wiki)

Outputs .codesight/wiki/ directory:

.codesight/wiki/  index.md      — catalog of all articles (~200 tokens)  overview.md   — architecture, subsystems, high-impact files  auth.md       — auth routes, middleware, session flow  payments.md   — payment routes, webhook handling, billing flow  database.md   — all models, fields, relations  users.md      — user management routes and models  ui.md         — UI components with props  log.md        — append-only operation log

AI Tool Config Files (--init)

Generates project-root files for each AI tool:

  • CLAUDE.md — Claude Code project instructions
  • .cursorrules — Cursor rules file
  • codex.md — OpenAI Codex context
  • AGENTS.md — general agent instructions

MCP Server Mode

Start CodeSight as an MCP server to give Claude Code or Cursor direct tool access:

bash
npx codesight --mcp

MCP Tool Reference

ToolDescription
codesight_get_contextFull codebase context map
codesight_get_routesAll API routes with methods and handlers
codesight_get_modelsAll database models and schema
codesight_get_componentsUI components with props
codesight_get_blast_radiusImpact analysis for a specific file
codesight_get_high_impact_filesFiles ranked by dependency count
codesight_get_framework_infoDetected frameworks and ORMs
codesight_get_wiki_indexWiki catalog (~200 tokens) for session start
codesight_get_wiki_articleRead one wiki article by name
codesight_lint_wikiWiki health check (orphans, stale, missing links)
codesight_get_knowledgeKnowledge map from markdown notes
codesight_benchmarkToken savings analysis
codesight_get_overviewProject overview summary

Configuring MCP in Claude Code

Add to your Claude Code MCP config (~/.claude/mcp_settings.json or project .mcp.json):

json
{  "mcpServers": {    "codesight": {      "command": "npx",      "args": ["codesight", "--mcp"],      "cwd": "/path/to/your/project"    }  }}

Configuring MCP in Cursor

Add to .cursor/mcp.json in your project root:

json
{  "mcpServers": {    "codesight": {      "command": "npx",      "args": ["codesight", "--mcp"]    }  }}

Language & Framework Support

LanguageAST PrecisionFrameworks Detected
TypeScript✅ Full ASTNext.js, NestJS, Hono, Remix, SvelteKit, Nuxt, Express, Fastify
JavaScriptRegexExpress, Fastify, Koa, Hapi
PythonRegexDjango, FastAPI, Flask, SQLAlchemy
GoRegexGin, Echo, Fiber, Chi
RubyRegexRails, Sinatra
PHPRegexLaravel, Symfony
ElixirRegexPhoenix
JavaRegexSpring Boot
KotlinRegexKtor, Spring Boot
RustRegexAxum, Actix
DartRegexFlutter
SwiftRegexVapor
C#RegexASP.NET Core

ORM / Database Support

Drizzle, Prisma, TypeORM, Sequelize, Mongoose, MikroORM, SQLAlchemy, Django ORM, ActiveRecord, Eloquent, Ecto, GORM, and more (13 total).

Real Usage Patterns

Pattern 1: Session Start with Wiki

At the beginning of every AI session, load the wiki index instead of the full context map:

# In Claude Code or Cursor, at session start:Use codesight_get_wiki_index to get project overview, then codesight_get_wiki_article for "auth" to understand authentication.

Pattern 2: Blast Radius Before Refactoring

Before modifying a shared file, check what breaks:

bash
npx codesight --blast src/lib/database.ts

Output shows every file that imports the target, ranked by impact — critical before refactoring database connections, shared utilities, or types.

Pattern 3: CI/CD Integration

Keep context fresh on every push:

yaml
# .github/workflows/codesight.ymlname: Update AI Contexton:  push:    branches: [main]
jobs:  update-context:    runs-on: ubuntu-latest    steps:      - uses: actions/checkout@v4      - uses: actions/setup-node@v4        with:          node-version: '20'      - name: Generate context map        run: npx codesight      - name: Generate wiki        run: npx codesight --wiki      - name: Generate knowledge map        run: npx codesight --mode knowledge ./docs      - name: Commit updated context        run: |          git config --local user.email "[email protected]"          git config --local user.name "GitHub Action"          git add .codesight/          git diff --staged --quiet || git commit -m "chore: update AI context [skip ci]"          git push

Pattern 4: Git Hook for Local Freshness

bash
npx codesight --hook

This installs a post-commit hook that regenerates .codesight/CODESIGHT.md automatically after every commit.

Pattern 5: Knowledge Mapping for Decision Records

bash
# Map ADRs and architecture docsnpx codesight --mode knowledge ./docs/decisions
# Map full Obsidian vaultnpx codesight --mode knowledge ~/vault
# Outputs .codesight/KNOWLEDGE.md with:# - Key decisions extracted from ADR files# - Open questions surfaced from notes# - Meeting notes indexed by date# - Specs and PRDs cataloged

Consuming the Output in AI Sessions

Loading Context in Claude Code (manual)

Read .codesight/CODESIGHT.md for project context before answering questions.

Loading Wiki at Session Start

Read .codesight/wiki/index.md first. Then read the relevant article (e.g. .codesight/wiki/auth.md) only when questions about that domain arise.

Loading Combined Context

Read .codesight/CODESIGHT.md for code structure and .codesight/KNOWLEDGE.md for architectural decisions.

Token Savings Reference

ApproachTokens per sessionSavings vs baseline
Manual file exploration26K–47Kbaseline
npx codesight (context map)3K–5K7x–12x
--wiki targeted article~200 + ~300/question60x–130x

Real benchmark results from production codebases:

ProjectStackFilesOutput TokensExploration TokensSavings
SaaS AHono + Drizzle1383,93646,02011.7x
SaaS BHono + Drizzle (monorepo)533,62926,1307.2x
SaaS CFastAPI + MongoDB404,16247,45011.4x
Rails appRails + ActiveRecord4,17221,711386,10017.8x
Laravel appLaravel + Eloquent3,89630,739493,28516x

.codesight Directory Structure

.codesight/  CODESIGHT.md       — main context map (commit this)  KNOWLEDGE.md       — knowledge/decision map (commit this)  wiki/    index.md         — wiki catalog (commit this)    overview.md    auth.md    database.md    payments.md    users.md    ui.md    log.md

Add .codesight/ to git to persist context across sessions and team members. The wiki survives session restarts — every new conversation starts with full codebase knowledge from the first message.

Troubleshooting

Routes not detected:

  • Ensure you're running from the project root (where package.json or equivalent lives)
  • TypeScript projects get full AST parsing; other languages use regex — check that route files follow standard framework conventions
  • For Python FastAPI, ensure router files use standard @app.get() / @router.post() decorators

Models showing 0 for MongoDB:

  • This is correct behavior — codesight parses SQL ORM declarations. MongoDB collections without Mongoose schemas won't appear. Define Mongoose schemas explicitly to get model detection.

Wiki articles missing:

  • Run npx codesight --wiki first to generate the wiki directory
  • Use codesight_lint_wiki MCP tool or check .codesight/wiki/log.md for generation errors

Large monorepos (10K+ files) are slow:

  • Add a .codesightignore file (same syntax as .gitignore) to exclude build artifacts, generated files, and node_modules subdirectories
  • Typical scan time: 186ms–890ms for projects under 5K files

MCP server not connecting:

  • Verify Node.js >= 18: node --version
  • Check cwd in MCP config points to the actual project root
  • Run npx codesight --mcp manually first to confirm it starts without errors

--init overwrites existing CLAUDE.md:

  • Back up existing files before running --init on a project with custom AI instructions
  • Generated files are additive starting points — merge with your existing content manually

來源與署名

來源:reason-machines/trending-skills位於skills/codesight-ai-context提交2384a00

授權條款: 無授權條款

內容歸原作者所有。SourceWeft 從公開儲存庫中收錄這些內容。

檢舉或申請下架