PR that clears max_input_tokens on mode and model changes in addition to base_url change (for example switch from gemini to gpt now resets to meaningful max tokens): https://github.com/OpenHands/OpenHands-CLI/pull/697 Running targeted model change tests (instead of using make to run all) was corrupting ~/.openhands/agent_settings.json so moved prevention from Makefile to test framework: https://github.com/OpenHands/OpenHands-CLI/pull/699