Claude-Context — Semantic Code Search

Semantic code search via AST-based indexing, Milvus vector database, and OpenAI embeddings.

Index path: /Volumes/DATA/GitHub/StakTrakr (40 files, ~885 chunks as of Feb 2026).

Rules

• •Always use search_code as the first step for any codebase lookup — it's the fastest way to orient yourself (~2s, zero subprocess tokens). For single-function or architectural questions, it's often all you need.
• •For cross-cutting or scattered concerns, Claude-Context may return incomplete results (e.g., finding 3 of 7 escape functions). When results seem partial, pass the initial findings as seed context to an Explore agent for comprehensive coverage.
• •For literal string matches, use Grep directly — Claude-Context uses semantic embeddings and can confuse similar concepts (e.g., ranking importCsv above exportCsv for an "export" query).
• •Always pass the absolute path /Volumes/DATA/GitHub/StakTrakr — relative paths will fail.
• •Use natural language queries, not code patterns. Good: "how are spot prices fetched from the API". Bad: "fetchSpotPrice function" (use Grep for literal matches).
• •Use extensionFilter to narrow results when you know the file type (e.g., [".js"] for JS-only results).
• •Raise limit (default 10, max 50) for broad exploratory searches; keep it low for targeted lookups.
• •Check get_indexing_status before searching if results seem stale or a search returns an error — the index may need rebuilding.
• •Re-index with index_codebase (with force: true) after significant structural changes (new files, renamed modules, major refactors). Confirm with the user before force-indexing.
• •Never use clear_index without explicit user request — it destroys the index and requires a full rebuild.