sqz
Optimizes AI coding agent context by compressing tool output and deduplicating repeated content before LLM injection.
AI coding agents frequently consume a large portion of their token budget on verbose and often redundant tool outputs, such as build logs, test results, or repeated file reads. sqz addresses this by transparently intercepting and compressing this output before it reaches the language model. It employs per-command formatters to extract and retain only the agent-critical information, like failing tests or assertions, and replaces content the model already has with concise 13-token references. Built in Rust, sqz operates deterministically without making any LLM calls, ensuring offline functionality and byte-exact recoverability of all compressed data.