Jani Woo
The software version used is v1.19.0-beta.2, Windows desktop
(I received both beta.1 and beta.2 updates in one day. The update frequency was well received, but more often the automatic update mechanism didn't work properly. I wonder if this requires scientific internet access)
  1. When the automatic compression switch is turned on, there is a situation where the calculated threshold does not match the set threshold. For example, for Deepseek-Reasoner (official context 128k), when the threshold value is 90%, the first automatic context compression starts when the actual number of tokens used is far less than 90% (about 30k-50k)
  2. What is the mechanism for automatic context compression? If LLM is used for generalization, can more customization options be provided in the future, such as the model used and the length of output (for example, to avoid losing too much contextual information by compression)
  3. After the context is automatically compressed, by default, all message records before the current output are compressed, making it easy for the text generation style to change (especially during the writing process). Is it possible to keep a few recent original text records uncompressed (or take the previous deepseek-reasoner as an example, if 128k context is reached when 120 records are reached, whether to only compress the first 90 or top 100 records, and keep the original text of some of the latest generated records as a reference for style and details)
·