Currently, when a tool (such as run_shell_command) produces a very large output, it can cause the Gemini API to return a 429 Resource Exhausted error due to the context window being exceeded. This can lead to unexpected failures and reduced reliability of the agent. A mechanism is needed to gracefully handle such large outputs, guiding the model to process them in a more efficient manner, possibly by redirecting output to temporary files and reading them in chunks.