Paritok-4B Compresses Coding Agent Context, Saves Tokens

Coding agents frequently resend large code files and tool outputs to their underlying Large Language Models (LLMs) in every turn, leading to high token costs. Existing general-purpose prompt compressors are often unsuitable for code, as they paraphrase identifiers or drop exact spans critical for code editing. Paritok-4B is a 4B LoRA compressor specifically developed for coding-agent trajectories. It operates on two core principles: it is extractive, meaning it selects exact code spans rather than rewriting them, preserving critical identifiers and numbers. Secondly, it is intent-conditioned, using the agent's current task to intelligently select which lines to retain, ensuring relevance. Trained on real OpenHands trajectories, Paritok-4B compresses agent context to about 25.7% of its original size, outperforming larger GPT models in compression efficiency. Crucially, it retains 86.5% of the uncompressed solve quality on SWE-bench Lite, with no statistically significant reduction in solve rate when fed line-numbered input. This 264 MB adapter can self-host on a single 24 GB GPU, offering significant cost savings compared to using larger, more expensive compressors.