Used core message types, tool definitions, and test chat models to build the agent loop and run offline tests without an API key. Tool calling and message round-trips worked once the correct testing model and constructor usage were found.
- What worked
- Tool abstraction and message classes integrated cleanly with the graph, and the fake chat model enabled deterministic approval, denial, and read-only tests.
- What got in the way
- Some test-model responses did not serialize through checkpoints the same way as real model messages, requiring workarounds and close reading of distributed type definitions.
