Component
CLI
Summary
I have to disable all MCP tools and ask_user_question when I set Large 3 (2512) as the model. If I don't, Vibe spends ages (often several minutes) doing what appears to be pre-processing of the tool schemas before any context is sent to the model. And then it repeats this for every response in the conversation. I assume this is some sort of JSON flattening perhaps to meet the expectations of Large 3? Using a skill causes the same problem.
I like to use large-3 in an analyst/explainer role when I'm researching new stuff. If this is what's happening then surely tool schemas should be pre-processed and stored in a cache?
Reproduction steps
- Set Large 3 (2512) as the model and have
ask_user_question and some MCP tool enabled.
- Ask a basic question, e.g. "what model are you"
Versions / environment
v2.7.2 / Windows 11
Logs & screenshots
Component
CLI
Summary
I have to disable all MCP tools and
ask_user_questionwhen I set Large 3 (2512) as the model. If I don't, Vibe spends ages (often several minutes) doing what appears to be pre-processing of the tool schemas before any context is sent to the model. And then it repeats this for every response in the conversation. I assume this is some sort of JSON flattening perhaps to meet the expectations of Large 3? Using a skill causes the same problem.I like to use large-3 in an analyst/explainer role when I'm researching new stuff. If this is what's happening then surely tool schemas should be pre-processed and stored in a cache?
Reproduction steps
ask_user_questionand some MCP tool enabled.Versions / environment
v2.7.2 / Windows 11
Logs & screenshots