response context limit randomly ignored #6265
Description
App Version
3.23.19
API Provider
OpenAI Compatible
Model Used
gemma 3 27b
Roo Code Task Links (Optional)
No response
🔁 Steps to Reproduce
on linux, I am running VSC roo code extension, with koboldcpp as the backend, sometimes roo code will follow the setting in providers for maximum output, other times it will be ignored. I can not find a reasoning why. every api request should have that, but due to it not being sent every time, the koboldcpp fallback of 512 is used, which is not enough to provide a complete diff, erroring over and over on the same response until it gets a warning.
in koboldcpp I have a context limit of 32k and other applications that properly send a requested length with each request are setting the max context in kcpp fine.
💥 Outcome Summary
expected a diff, got 1/3 of one.