Support case when model returns empty Actions list

[andrewfrench]

This PR addresses a reported issue when parsing a list of Actions while initializing an ActionsSubtask.

In some cases, Models include an empty list of Actions in their responses. We currently raise a SchemaError this happens:

[05/11/24 12:58:44] INFO     ToolkitTask 9f94e60e40d04516b9cce029eab11a25                                                                                     
                             Input: hi my name is jason                                                                                                       
[05/11/24 12:58:47] ERROR    Subtask 9f94e60e40d04516b9cce029eab11a25                                                                                         
                             Invalid action JSON: Array item count 0 is less than minimum count of 1.                                                         
                    INFO     Subtask 3352ba0de32041969bbb7d9781a8c236                                                                                         
                             Thought: The user has introduced himself as Jason. There is no specific question or request, so I will acknowledge the           
                             introduction and ask how I can assist him.                                                                                       
                             Actions: []                                                                                                                      
                             Answer: Hi Jason! How can I assist you today?                                                                                    
                    INFO     Subtask 3352ba0de32041969bbb7d9781a8c236                                                                                         
                             Response: Action JSON validation error: Array item count 0 is less than minimum count of 1.                                      

If the model does not return an Actions list at all, we do not attempt to parse actions and do not raise an exception. In both cases the model is electing to take no actions. If "no actions" is a valid response, allowing one case and disallowing another is inconsistent behavior.

[vasinov]

Did you test it outside of unit tests with a real LLM? I remember there potentially being some quirk for an empty actions list.

[andrewfrench]

We're currently testing which branch to take based on whether or not actions_matches has elements or not. If an empty list is provided, there's a match and we try to parse the actions. If, while parsing, we decide there actually aren't any actions, we continue execution without any actions or output. This is interpreted by ToolkitTask as a deviation from the ReAct prompt and the Subtask's raw input is set as its output. We don't want that, we want to extract the answer just like we would if there were no actions_matches.

This change updates the logic to first try parsing actions. If there are no actions_matches, or if an empty set is provided by the model, this results in an empty self.actions list. If the model chooses to take no actions, we set the output to the answer as we would expect, standardizing behavior.

[vasinov]

Thanks for refactoring this out. Long overdue!