I've been running coding agents heavily for the past year or so using frontier model APIs, open weight model APIs and, most recently, local models (Qwen family models on a Strix Halo). Starting to run local inference has highlighted something I've been aware for longer: just running tests output shedloads of text into the context window that is there for good until compaction or starting afresh. For example, a single `cargo test` dumping 8KB into the agent's context just to communicate "47 test passed." The agent reads all of it, learns nothing useful, and the context window fills with noise. Makes LLM prefill slower as well as costs more when using per token APIs. I created a small program that sits between the command output and the LLM: oo, or double-o ... yes, sad play on words. Double-o, the agent's best friend :) oo wraps commands and classifies their output: No agent modification needed: add "prefix commands with oo" to your system prompt. Single Rust binary, 197 tests, Apache-2.0. The classification engine works using regex-based pattern matching with per-command failure strategies (tail, head, grep, between) and automatic command categorization (status/content/data/unknown) that determines what happens with unrecognized commands. Content commands like git diff always pass through; data commands like git log get indexed when large. Especially noticeable with local models & wall-clock time. Helps with frontier models too ... cleaner context, fewer confused follow-ups. |