vally init scaffolds a starter eval.yaml so you don’t have to write your
first spec from scratch. The generated eval asks the agent to write a
hello.json file and grades it with deterministic graders (file-exists and
file-matches), so grading needs no judge model or extra runtime dependencies;
only vally eval itself needs an executor (the built-in one uses the Copilot
SDK). It also includes commented-out “level-up” graders (running a command to
parse the JSON, and an LLM judge with a rubric), an environment block for
loading a skill or MCP server, constraints, and a pointer to experiments.
By default it writes to evals/hello/eval.yaml. Pass a directory to write
somewhere else.