Tool Calling

Operator defines a tool with `parameters: { city: 'string' }` (plain object) instead of `parameters: z.object({ city: z.string() })`.

tools[].parameters must be a Zod schema (or jsonSchema()-wrapped JSON Schema) — the SDK converts it to a provider tool-spec at call time. Plain objects are rejected with InvalidArgumentError. Use z.object({...}).describe('...') to give each field a description the model can see.

Tool's execute is `(args) => fetch(...)` returning a Response object directly (not awaited, not serialized).

execute({input, toolCallId, messages}) must be async and return a JSON-serializable value — that value becomes the toolResult passed back to the model. Await any IO inside execute; resolve to a plain object (or string). Throwing inside execute surfaces as a tool-error part on the fullStream.

generateText with two tools and no maxSteps; the model keeps calling tools in a loop, exhausting the operator's token budget.

Default maxSteps is 1 — generateText returns after a single step (which may include tool calls but does not auto-resume). Set maxSteps to a sane cap (e.g., 5) for agentic loops. For finer control, use stopWhen with a custom predicate. Always cap; never set maxSteps=Infinity in production.