An LLM technical writer that tracks your git history

What it does

RepoAgent scans a Python repository and uses AST parsing to map objects and their bidirectional call relationships. It then dispatches an LLM to write Markdown documentation for each file, and integrates with git pre-commit hooks to detect added, deleted, or modified files. The hook regenerates only the affected docs, keeping the documentation folder in sync with the codebase.

The interesting bit

Instead of blindly regenerating everything, RepoAgent maintains a JSON record of the project hierarchy and performs targeted updates. The authors note it handled the roughly 270,000-line XAgent codebase using gpt-3.5-turbo, and it renders the output as a GitBook-style documentation site.

Key highlights

• Uses AST analysis to identify inter-object bidirectional invocation relationships before asking the LLM for prose.
• Updates docs incrementally via git pre-commit hooks, detecting file additions, deletions, and modifications.
• Supports local models (Llama, ChatGLM, Qwen, GLM4) in addition to OpenAI’s API.
• Generates a GitBook-compatible documentation book from the resulting Markdown.
• Multi-threaded execution for generating docs concurrently.

Caveats

• Currently supports only Python; Java, C, and C++ are listed as future work.
• The “Chat With Repo” feature is described as a preliminary prototype for Q&A and code explanation.
• Documentation quality depends heavily on the chosen LLM; the authors suggest GPT-4 variants over GPT-3.5 for better results.

Verdict

Python teams that already use git and struggle to keep internal docs current are the obvious audience. Everyone else—especially non-Python shops—should wait for the promised multi-language support.