RepoDigest.

- Approach

The design keeps the digest pipeline and the AI layer decoupled on purpose. Core processing (`query-parser`, `repository-cloner`, `file-processor`, `output-generator`) produces a fully-formed digest with no dependency on any model; the AI service only activates when one of the `--ai-*` flags is set and bails gracefully if the key is missing or the network fails. Filters compose instead of nest: language and glob and size and depth all combine through fast-glob and ignore, so real-world repositories with noisy `node_modules` and large binaries behave the way you'd expect on the first run. The CLI itself is written against Commander with a custom help formatter that splits long documentation into sub-commands (`examples`, `ai-help`, `options`, `features`) to keep `--help` legible, and inquirer drives interactive mode for users who'd rather answer questions than read flag tables.

- Outcome

The tool is published open-source under MIT and installs globally via npm as `repodigest`. It reads almost any repository end-to-end, including private ones via `GITHUB_TOKEN`, and emits digests sized for real LLM context windows with an accurate token count up front. The AI path supports Gemini 2.5 Flash, 1.5 Pro, and 1.5 Flash, with environment-variable overrides for model, temperature, max tokens, and timeout. What started as a scratch-your-own-itch utility for feeding codebases to models became a useful piece of my own workflow and a small but complete example of how to build a Node CLI that doesn't feel like a weekend hack.