This message was deleted.

Of course, i'm hitting the context limit on almost every run; my goal is to set the context limit to 30k because imo performance is better on individual iters when provided context is shorter. it sounds like 1, 2, will take a long time. Can we just add an option (off by default) for a naive context truncater which is optional where i can cut context at 30k? maybe this will reduce performance on whatever arbitrary metric you guys are optimizing on or be more expensive, but the way i have set up my code with very short files and very broad initial prompts i just notice the openhands agent always starts off strong, and then gets worse over time. my guess is that it will only be maybe touching up the tests since my initial PR and adjusting 1 config, so very quick, and off by default naturally. the initial context-cut when there was an error because the context was too long is all I need, except for the user manually set value, not at the point of error. https://github.com/All-Hands-AI/OpenHands/pull/4977/files