The repeat_penalty helps but won't fully solve it — infinite tool call loops are a fundamental issue with reasoning models that don't have a hard stopping condition outside the model itself.

Beyond sampling params, worth adding an external loop guard: a max tool call count per run, or a budget cap that kills the run if it exceeds N steps. That way it can't spiral regardless of how the model is behaving.

We built SupraWall for exactly this kind of enforcement — hard caps on tool call counts, execution budgets, and blocked categories before they execute. Works as a wrapper around local agent setups like llama.cpp-based servers: github.com/wiserautomation/SupraWall