Instrumentation

[gewenyu99]

Created a tracker that is called through:

tracker?.onMessage(message);

and runs parallel to the agent to benchmark things like:

• Input and output size
• Context size
• Log compactions
• Turns per phase (that's what I'm calling each of the prompts we have in llm-prompts
• Time per turn
• Cost

What's cool is that we can use a similar pattern to build all kinds of plugins and instrumentation, so I might generalize this in the future to an interface...