New LLM harness for making calls

[malanjary-tsri]

Current system uses anthropic API for Agent turns but could save on costs via subscription(s) and CLI tools.

Goal: Alter the system to instead get LLM output via CLI tool (e.g. Claude code)

Ideas:

• Minimal refactoring, use something like Vibeproxy (https://github.com/automazeio/vibeproxy) to "middle-man" API calls
• Add additional route for LLM calls via Claude code directly
  • Require backend to start / stop claude CLI and handle input / output
  • Could use persistent CLI instance with some kind of chat management instead
• Look into alternate harness e.g. Ampcode