An application executes a model-requested function and receives a result. What should it supply back to the model for continuation?

1. Only the tool's original description again
2. The application's secret credentials instead of the result
3. The function output associated with the corresponding tool call
4. A fabricated result if the actual output is inconvenient