Example sandboxed scratch tool to improve LLM response accuracy (PAL Pattern)

citation: PAL: Program-aided Language Models

(macher-agent-make-tool macher-agent-eval-deno-tool
    "Evaluate JavaScript or TypeScript code to perform maths operations, string manipulation, or data generation."
  :category "compute"
  :args '((:name "code" :type string :description "The JavaScript or TypeScript code to evaluate. Use console.log() to output the final result."))
  :command-fn (lambda (payload)
                (let* ((code (plist-get payload :code))
                       (temp-file (make-temp-file "macher-eval-" nil ".ts"))
                       (result nil))
                  
                  (with-temp-file temp-file
                    (insert code))
                  
                  (unwind-protect
                      (with-temp-buffer
                        (call-process "deno" nil t nil "run" 
                                      temp-file)
                        (setq result (string-trim (buffer-string))))
                    (delete-file temp-file))
                  
                  (make-macher-agent-lisp-result-response
                   :payload (format "EVALUATION RESULT:\n%s" result)))))