tester-army/unbox-ai ? reverse-engineered prompt

Build me a tool called unbox ai that lets me open an AI agent trace file and actually understand what happened.

I want to run one command like npx unbox-ai trace.json and have it start a local viewer in the browser, showing how the run used tokens, which messages were sent each step, what tools were called, how long each generation took, and where the cost went. The viewer should make repeated context easier to read by showing only what is new each step, and it should also let me compare two runs side by side.

Please also make it work as a safe, read only CLI for agents, so it can print summaries, events, messages, tools, and comparisons without accidentally starting a server when it is being used in a script. It should support the trace formats mentioned in the README, and feel polished enough that I can use it to inspect AI SDK devtools data as well. If you need current docs or examples for the trace formats, look them up online.