tool-use-validation

Tool Use Validation: Schema-Based Input and Output Checking

An agent generates a tool call to fetch weather data, but the arguments might be malformed (missing required city field, wrong type for units, invalid enum value). Even if the input is correct, the tool's output might be incomplete or out of range. Without validation at both ends, bad data silently propagates through the pipeline and produces unreliable results.

This workflow validates tool calls at both the input and output boundaries, using JSON-schema-style checks before execution and after.

Pipeline Architecture

userRequest, toolName
         |
         v
  tv_generate_tool_call  (toolArgs, inputSchema, outputSchema)
         |
         v
  tv_validate_input      (isValid, validatedArgs, checks list)
         |
         v
  tv_execute_tool        (rawOutput with weather data, executionTimeMs=245)
         |
         v
  tv_validate_output     (isValid, validatedOutput, validationReport)
         |
         v
  tv_deliver             (formattedResult string, validationSummary)

Worker: GenerateToolCall (tv_generate_tool_call)

Produces structured toolArgs: city: "London", country: "UK", units: "metric", include: ["temperature", "humidity", "wind", "forecast"]. Also generates an inputSchema map defining the expected structure: required: ["city", "country"], with units constrained to enum: ["metric", "imperial"] and include typed as array of strings. Separately generates an outputSchema requiring temperature (number), humidity (number), and conditions (string).

Worker: ValidateInput (tv_validate_input)

Runs four validation checks returned as List<Map<String, Object>>: required_fields ("All required fields present: city, country"), type_validation, enum_validation ("units value 'metric' is in allowed enum values"), and array_items. All checks pass. Returns isValid: "true", the validatedArgs (pass-through), and an empty validationErrors list.

Worker: ExecuteTool (tv_execute_tool)

Returns raw weather data: temperature: 14.2, humidity: 72, conditions: "partly cloudy", windSpeed: 18.5, windDirection: "SW", and a 3-day forecast list with maps containing day, high, low, and conditions for Monday (cloudy, 16/11), Tuesday (sunny, 18/12), and Wednesday (rain, 15/10). Includes provider: "weather_api_v3" and executionTimeMs: 245.

Worker: ValidateOutput (tv_validate_output)

Runs four output checks: required_fields (temperature, humidity, conditions present), type_validation, range_validation ("values within expected ranges"), and format_validation (timestamp and provider properly formatted). Returns isValid: "true" and a validationReport with the checks list, empty errors list, and timestamp.

Worker: Deliver (tv_deliver)

Formats the validated output into a human-readable string: "Weather in London: 14.2C, partly cloudy. Humidity: 72%. Wind: 18.5 km/h SW. 3-day forecast available." Includes a validationSummary: "All input and output validations passed. 4 input checks and 4 output checks completed with zero errors."

Tests

5 tests cover tool call generation, input validation, execution, output validation, and delivery.

Further Reading

• RUNNING.md -- how to build and run this example