Wrap a securer_tool with input and output guardrails

Composes a tool from securer::securer_tool() with guardrails from secureguard. The returned object is itself a securer::securer_tool() (same schema, same IPC contract) whose closure runs each invocation through the supplied input guardrails, executes the underlying tool, then runs the result through the output guardrails.

Usage

guarded_tool(tool, input_guards = list(), output_guards = list())

Arguments

tool: A securer_tool object (typically from one of the tool_*() factories in this package, but any securer_tool works).
input_guards: A list of secureguard input guardrails (type "input" or "code"). Each receives the stringified tool args and must pass for the call to proceed.
output_guards: A list of secureguard output guardrails (type "output"). Each receives the tool's return value (coerced to text via secureguard::output_to_text) and must pass for the result to be returned.

Value

A new securer_tool object with guardrails applied.

Details

Guardrail failures are translated into errors raised from the tool closure; inside a securer session these surface as tool-call errors and are returned to the LLM via ellmer's ContentToolResult(error =) shape.

Guardrails are applied lazily: if secureguard is not installed, calling guarded_tool() errors with a clear installation hint rather than silently skipping enforcement.

Examples

if (FALSE) { # \dontrun{
  calc <- tool_calculator()
  injection <- secureguard::guard_prompt_injection()
  secrets <- secureguard::guard_output_secrets(action = "block")
  guarded <- guarded_tool(
    calc,
    input_guards = list(injection),
    output_guards = list(secrets)
  )
} # }