It is currently difficult to tell vyb what to focus on, because the system message is generic about their role and kind of task, and the user message is just the collection of files in scope for changing and the context of their modules. Even when using the TODO(vyb): notation within the code, it will many times decide to focus on something else.
It would be good if we could persist a task description in the .vyb folder, and include that in the user message to direct the LLM's action.
To implement this, we will need:
- A new command to initialize and plan a task (there may be the need for a whole command family here, we need to refine this requirement better). A task should have a description of the expected final state of the system, as well as a list of individual actions to reach that. Each action would map to a single
vyb execution and git commit;
- Either a new command or a change to existing commands so the task description is included in the user message in a way that becomes unambiguous to the LLM what the next step is;
- Possibly additional commands or sub-commands to mark a task completed, abandon a task, etc;
It is currently difficult to tell
vybwhat to focus on, because the system message is generic about their role and kind of task, and the user message is just the collection of files in scope for changing and the context of their modules. Even when using theTODO(vyb):notation within the code, it will many times decide to focus on something else.It would be good if we could persist a task description in the
.vybfolder, and include that in the user message to direct the LLM's action.To implement this, we will need:
vybexecution and git commit;