tough

hx is a terminal coding agent that improves its own working environment from its own failures. Coding agents keep repeating the same mistakes across sessions, such as calling python on a machine that only has python3 or trying to run interactive programs without a terminal, because nothing about their environment changes after they recover. Putting every past incident into the prompt is not a fix either, since it bloats the context. When a real failure happens during normal work, hx reviews its own trace, finds the root cause, and decides whether to store a short general rule, file a specific case note, or store nothing. A new rule starts as a trial and is only promoted after it actually prevents the failure on a later task. Cases are loaded only when a quick relevance check says they apply, so the context stays small. A local web console makes all of this visible: you can watch each step of the agent loop live, see the exact prompt the model received, and read every rule and case it has written. In our tests with a deliberately weak model, real failures dropped from ten on the first task to one on later, different tasks, after the agent learned a single rule and promoted it.