Tool Schemas

Tool schemas describe the arguments a model provider can pass to a HarnessTool. They should be precise, small, and serializable to JSON Schema when possible.

import { HarnessTool } from "@harness-kernel/core/agent/tool";
import { s, type InferInput } from "@harness-kernel/core/schema";

const searchSchema = s.object({
  query: s.string().min(1).describe("Search query."),
  limit: s.number().int().positive().max(20).default(5),
});

type SearchInput = InferInput<typeof searchSchema>;

class SearchDocsTool extends HarnessTool<SearchInput> {
  name = "search_docs";
  description = "Search project documentation.";
  schema = searchSchema;
  risk = "read" as const;

  async execute(args) {
    const input = searchSchema.parse(args);
    return {
      content: `Search ${input.query} with limit ${input.limit}`,
      data: input,
    };
  }
}

Defaults And Inputs

InferInput represents what callers may pass. InferOutput represents the parsed result after defaults are applied. Tools commonly use InferInput for the class generic and call schema.parse(args) inside execute().

Approval Metadata

Schemas validate inputs. Tool risk, permissions, and requiresApproval describe execution risk:

class WriteFileTool extends HarnessTool<WriteInput> {
  risk = "write" as const;
  permissions = [{ kind: "filesystem" as const, access: "write" as const, path: "." }];
  requiresApproval = true;
}

The runtime host still owns approval policy.