Custom Tools And Processing

An operator defines a custom tool wrapping their internal API but gives it no input schema, so the model passes free-form arguments.

Define the custom action with an explicit input schema (typed, required fields, descriptions) so the model produces valid arguments and the handler can validate them. A schemaless custom tool yields unvalidated, ambiguous calls.

A custom action's raw response includes an internal access token field. A post-processor returns the response to the model.

Use the post-processor to strip secrets/credentials and trim large fields before the result reaches the model. The post-processor must remove the token field, not just rename it. Confirm secret fields are gone from the model-visible payload.

A pre-processor should inject a tenant_id and normalize a date before a custom action executes, but it instead mutates unrelated fields.

Use a pre-processor (schema/request processor) to deterministically transform inputs before execution — inject server-derived context (tenant_id), normalize formats — without altering the model's intended arguments. Keep the transform narrow and idempotent.