Variables

A prompt body carries {{name}} placeholders. The SDK fills them from vars in your process — with the same rules the server uses — and returns the finished text. The values are never sent.

// "Reply to {{customerName}} about {{topic}}."
await client.prompts.resolve(SUPPORT_REPLY_ID, {
  vars: { customerName: "Ada", topic: "billing" },
});

Rules

Rule	Why
A missing variable fails with EvalwrightClientError — one given as undefined too	Substituting an empty string ships a prompt with a hole in it, and you find out from the model's output instead of from the error
Unknown keys are ignored	Lets you hand one context object to several prompts, and lets a prompt drop a variable without breaking callers
Values are substituted literally	Not parsed, not re-scanned, not trimmed — a value containing {{x}} stays that text
Everything is a string — anything else fails with EvalwrightClientError	Format numbers and dates yourself, where the decision belongs. null used to become an empty string, an object [object Object]

BRAID graphs

A version can carry a BRAID graph instead of, or alongside, classic text. Node labels use the same syntax and the same rules — one grammar definition is shared by the app, the server and the SDK, so what the editor accepts is what the server resolves.

The values are not written into a graph's labels: a value carrying ] would close the label it landed in and change the graph. They are listed beside it in <value> blocks, the way the benchmark gave them to its model. In a value containing <value> or </value>, that < is written as <, so it cannot end its block. A benchmark refuses such a value before it starts.