Vomit: Clean up Claude 5's token output with a separate LLM

Vomit addresses the problem of Claude 5 emitting incoherent 'token vomit' by translating that raw output into English via a separate local LLM. The tool is fully local (no telemetry) and has no external dependencies. The author notes it's vibe-coded, only tested on Mac, and that the local LLM can only see what Claude communicates (not files or actions), so hallucination is possible. There's also a risk of completely missing Claude's message; the author suggests using AgentsView to retrieve original messages, since Vomit doesn't touch runtime data except for writing files to TMPDIR.

Installation uses `go install github.com/zachahn/vomit@latest`, followed by `vomit init` to set up LLM connection details, and `vomit scrub -claude` for instructions on replacing Claude's output via hooks. For non-invasive use, `vomit list` shows session IDs, `vomit tail [<session_identifier>]` translates tokens for a specific session or follows the latest, and `vomit help` lists all commands.

The tool works with Llama.app, Ollama, or anything using the OpenAI API; the author recommends Llama.app with GPT-OSS 20B if no local LLM is set up. The project is licensed under GNU GPLv3, and a blog post about it is available at https://zachahn.com/posts/1787191554.