DasDigitaleMomentum/opencode-agent-evaluator ? reverse-engineered prompt

Build me a CLI tool that can benchmark AI coding agents while they work through a project in OpenCode.

I want to point it at a test project, give it one or more prompts, and have it record how many tokens the lead agent and any sub agents used, including input, output, reasoning, and cache tokens. It should label sessions clearly so I can see which parts came from the main agent versus sub agents, and it should save the results in a structured YAML file plus a full text log.

Also include a simple analysis script that can read multiple benchmark runs and compare them side by side, so I can see totals, billable tokens, cache usage, context usage, tool calls, and how many sub agent sessions happened.

Make it easy to run from the command line with a config file, and include example benchmark projects and scripts to create, reset, snapshot, and inspect testbeds. If you need to check current OpenCode docs or APIs, look them up online first.