cursorrules

GraphRAG — Code Discovery Rules

Always pass active_project_root (absolute repo root) on every tool call. Reuse the same path throughout a session.

Primary goal: answer or edit with the least token usage possible. GraphRAG is a swiss-army-knife for repo discovery — far better than grep — but only when you use it surgically. Every extra search or read wastes tokens.

---

The only discovery tool

search_codebase_intent is the only GraphRAG tool. Use it for all repo discovery.

Do not use native Grep, SemanticSearch, or iterative file hunting to explore the codebase.

---

Efficiency first

• Answer with what you have. If search + reads already cover the question, stop — do not search or read again.
• One search, then read. Start with a single search_codebase_intent call. A second search is allowed only when something specific is still missing (wrong anchor, no cite for a symbol you need, gap in the call path).
• Read each cite once. Open only the cite spans you actually need. Skip spans that do not help answer the question.
• No overlapping reads. If two cites cover the same file and ranges overlap, read the union once — never re-read lines you already have in context.
• No speculative exploration. Do not read "just in case", prefetch adjacent files, or widen ranges beyond what the question requires.
• Fewest tool calls wins. Prefer one good search + targeted reads over multiple searches, broad reads, or re-fetching content already in the conversation.

---

Workflow

1. Call search_codebase_intent once with intent queries and any known symbol names.
2. Read only the cite spans required to answer (format: startLine:endLine:filepath). Merge overlapping spans in the same file into one read.
3. Answer or edit using what you already have — stop here if sufficient.

Second search — only when a specific gap remains after step 2 (missing symbol, wrong file, incomplete path). Refine queries or symbols; still no grep.