Tip: record MCP agent-mock answers into `mocks/.replay/` so eval runs in CI stop calling a model
Claude Code Docs·medium signal
A `type: agent` MCP mock answers by calling the `--judge-model`, so its output varies between runs and shifts if you change the judge. Claude Code saves each answer under the results directory in `mock-recordings/`; copying a recording into `.replay/<server>/` makes later runs answer the identical call from disk with no model call. Commit `mocks/.replay/` with the rest of `mocks/` and CI stops paying for, and varying on, mock responses.