cull

cull uses Project Think and virtual file system cloudflare/shell to triage your docs/codebase and answer questions... you know RAG IS DEAD (they said.)

You point it at one repo with env vars, deploy it, and call:

• GET /health
• GET /status
• POST /query
• POST /query/stream

It keeps the repo warm in the background and answers questions from the repo contents.

Deploy

1. Install dependencies.

npm install

2. Create the Cloudflare resources once:

npx wrangler r2 bucket create docs-agent-cf-repo-files
npx wrangler d1 create docs-agent-cf-repo-db

Then copy the returned D1 database_id into wrangler.jsonc.

3. Set your repo in .env for local dev, or as Worker vars for deploy:

REPO_URL=https://github.com/your-org/your-repo
REPO_BRANCH=main
QUERY_TIMEOUT_MS=120000
MODEL_ID=@cf/moonshotai/kimi-k2.5
REPO_SYNC_INTERVAL_SECONDS=3600

4. If the repo is private, add a token:

npx wrangler secret put REPO_TOKEN

5. Deploy:

npx wrangler deploy

6. Ask a question:

curl -X POST https://your-worker.your-subdomain.workers.dev/query \
  -H "content-type: application/json" \
  -d '{"question":"How does authentication work?"}'

Response Shape

POST /query returns:

{
  "answer": "...",
  "sources": [
    { "title": "Authentication", "path": "docs/auth.md" },
    { "title": "auth.ts", "path": "src/auth.ts" }
  ]
}

POST /query/stream streams SSE events:

• status
• delta
• sources
• done
• error