What should CGA measure next for AI coding context?

CGA (Context Graph Agent) is launching today, and I would love technical feedback from people building or using AI coding agents.

The core idea is simple: instead of giving an agent broad file dumps or noisy search results, CGA indexes a repository into a local graph of files, symbols, calls, imports, dependencies, and lightweight data flow. Agents can then query focused evidence through MCP-compatible tools before generating code.

The first public benchmark looks at prompt-token reduction and Hallucination Pressure Score across 102 real-code cases. That is a useful starting point, but I am curious what builders here would measure next.

Would you prioritize retrieval precision, edit success rate, time-to-fix, benchmark reproducibility, model cost, or something else?