feat: orchestrate evals with Claude Code and VS Code subagents #141
Description
instead of using hardcoded logic, just have Claude Code read the yaml files and pass the input to subagents. Then the main agent evaluates and uses the same scoring method as the llm judge.
The agent judge will need to look at the file system. You can still use llm judge but we need a script to pass the code diff to the LLM. Also create a new workspace clone to avoid changing existing workspace.