Currently, the judge agent only gets the compressed history as context for its decision whether to terminate the execution agent. This could become problematic if the history context gets compressed multiple times and loses information, leading the judge making the wrong decision due to lack of information.
Consider whether we want to preserve the entire conversation history and provide it to the judge. While this could lead to more input token usage, we could strip the tool output and the agent output should be reasonable small.
Currently, the judge agent only gets the compressed history as context for its decision whether to terminate the execution agent. This could become problematic if the history context gets compressed multiple times and loses information, leading the judge making the wrong decision due to lack of information.
Consider whether we want to preserve the entire conversation history and provide it to the judge. While this could lead to more input token usage, we could strip the tool output and the agent output should be reasonable small.