mgrep-code-search_skill

Overview

This skill provides semantic code search using mgrep to explore medium-to-large codebases quickly. It complements grep/ripgrep by answering natural language queries about features, intent, and implementation. Use it when repository size or nesting makes simple pattern search inefficient.

How this skill works

The skill indexes the repository and performs semantic matching of natural language queries against code, text, PDFs, and images. A watcher process keeps the index in sync with file changes and respects .gitignore and .mgrepignore. Search commands accept path constraints and options to control result count, content display, and AI-driven synthesis.

When to use it

• Codebases with more than ~30 non-gitignored files or deep directory nesting
• Looking for concepts, features, or intent rather than exact identifiers
• Exploring an unfamiliar project to find where behavior is implemented
• Tracing architectural responsibilities (e.g., auth, caching, error handling)
• When grep/ripgrep returns too many false positives or misses semantic matches

Best practices

• Start the watcher (bunx @mixedbread/mgrep watch) before heavy searching to keep the index current
• Constrain searches to a path when possible to improve relevance and speed
• Use -m to increase result count for broad queries and -c to inspect full matches
• Enable AI synthesis (-a) when you need a concise summary of scattered findings
• Add a .mgrepignore to exclude large generated folders that bloat the index

Example use cases

• Find where authentication is configured across backend and frontend code
• Locate all error handling patterns for API calls to standardize responses
• Discover database connection and migration setup spread across modules
• Understand how caching is implemented and where cache keys are defined
• Quickly map feature ownership when joining an unfamiliar repository