Expose raw Assistant trace (tool calls, model I/O, system prompt, etc.)

kslays

When the Assistant breaks, there's no way for me to figure out why. It could be a wrong tool call, bad parameter, silent retry, an empty reply, or the tool returning something the model then ignores. The UI only shows the formatted answer.

Add a debug feature that shows everything in its raw unformatted form: system prompt, user message, every tool call with its JSON, every tool response, raw model output. Perhaps the best ways is a toggle to show the complete unformatted log. (Syntax highlighting would be nice.)

Local LLM tools already make this available. Without it, debugging Assistant failures is guesswork, and it's hard to know whether to report problems, ignore them, or just fix my own custom prompt.

It could be really helpful for technical users and easily ignored for everyone else.

andrewski

I'd love that too, but consider that maybe Kagi doesn't want to expose how their assistant works behind the scenes for one reason or another.

kslays

andrewski Yes, I considered that, but they do share the system prompt and allow it to be disabled via slash commands, and are generally quite transparent and helpful to their technical customers, so it's certainly worth asking.

kslays

Perhaps if people at Kagi are concerned about misuse, good approach to mitigate that might be to make this feature only available upon application (a "developer mode"). Or if they want to restrict further, upon application by contributor level users, or something like that.