faint night
Hi friends!
I'm reviewing logs of a task that failed, and wondering if using always 'new_task' and never 'switch_mode' couldn't bring more structured executions.
I'm even trying to do this in my local kilocode repo, but wasn't able yet.
What's are your thoughts?
sly condor
Can you provide more details (what model you're using)
faint night
I gave a task to 'architect' and noticed that when it started switching modes to execute the task, it started messing the thread. Task wasn't successfully completed and it became very hard to debug the reason. So I'm wondering that, if all agents just use new_task instead of switch_mode, we would have more organized execution.
Is it what you're referring to?
coral terrace
Orchestrator is the one that creates a new_task, architect don't do it, maybe if u modify the prompt if architect to use new_task instead of just switching modes
faint night
Yes, I'm trying to do that
But noticing that there is an implicit opposite instruction included by kilocode about this
If I put my own instructions, the agent will read them only after this 'introductory' prompt
I'm trying to create other layers like:
But all of these attempts are only 'models best efforts', not deterministic.
That's why I've considered to update kilocode's code to remove the 'switch_mode' tool from the 'ALWAYS_ALLOWED_TOOLS' in the project's source code, but not so safe on the hell it could cause further lol
dry summit
i have worked on a system to enable/disable tools on granular level so you can disable these tools from it, its still not approved because ive added alot more things in the feature request if they accept it i'll put up. a PR