benmarte

Use Language Server Protocol operations directly for precise cross-file navigation. Find all references to a symbol, go to definition, get hover documentation, and rename symbols across the codebase. Only available when a language server is configured for the project language.

Fetch the exact source code of a named symbol (function, class, method, type) from the codebase using the codemunch index. Reads only the lines for that symbol — not the whole file. Uses LSP for hover docs and type info when available. This is the core token-saving operation.

Index the codebase using Universal ctags when LSP is unavailable for a language. Generates a JSON symbol table with name, kind, file, line, and end line for all detected symbols. Supports 40+ languages natively.

Search the codemunch index for symbols matching a name, pattern, kind (function/class/method), or file. Returns a compact list of matches with location and signature — no source code read until you fetch. Use this to explore unfamiliar codebases token-efficiently.

Automatic index freshness gate. Called before any query command (search, fetch, refs, explore) to ensure the index exists and is up-to-date. Handles first-run setup, incremental re-indexing, and staleness detection — so users never need to manually index.

Index symbols using ripgrep regex patterns when neither LSP nor ctags is available. Covers TypeScript, JavaScript, Python, Go, Rust, Ruby, Java, Kotlin, C, C++, PHP, and can be extended with custom patterns for any language via .claude/codemunch/config.json.

Detect which Language Server Protocol (LSP) servers are available for the current project and configure codemunch to use them. Checks for installed language servers across all major languages and writes the LSP config to .claude/codemunch/config.json.

Build or refresh the codemunch symbol index for the current project. Walks every source file, extracts all symbols (functions, classes, methods, constants, types) with their exact file location and line range, and writes a fast-lookup index to .claude/codemunch/index.json. Uses LSP when available, falls back to ctags, then ripgrep patterns.

Scan the codebase for deficiencies, generate a prioritized report ranked by efficiency (points per iteration), and optionally transition into focused improve loops area-by-area until the user stops or the score hits 100.

Autonomous improvement loop for any codebase. Uses git worktrees to run every experiment in isolation — the main branch is never touched until a winning change is explicitly merged. Reads .claude/autoimprove/config.md for the measurement suite, then iterates: create worktree → propose → implement in worktree → measure → merge if improved, delete if not → log → repeat.

Manage git worktrees for autoimprove experiments. Creates isolated worktrees for each experiment so the main codebase is never touched. Handles setup, teardown, and merging winning experiments back to main.

Emergency cleanup of all autoimprove experiment worktrees. Use this if a session was interrupted, Claude crashed mid-experiment, or you want to abort the loop and clean up all branches and worktrees.