llm-coding-agent

About

The plugin exists because Willison wanted a minimal, inspectable coding agent on top of the llm CLI's model-agnostic plugin system; the first alpha was itself written by prompting Claude Code through a spec-and-TDD workflow. Read-only tools (numbered read_file, gitignore-aware list_files, ripgrep-backed search) run freely, while write_file, edit_file, and execute_command require approval, with chain_limit bounding tool rounds per run and every session recorded for resume via -c or --cid. File access is confined to the session root, with traversal via .., absolute paths, or symlinks rejected as errors the model can correct. A Python API (CodingAgent, CodingTools) exposes the same loop programmatically, with a pause/resume approval protocol for non-terminal applications.