Description
In previous versions crush executes a summarization then cleanup the context and injects this summarization to continue working.
Currently, every time I reach 80%> context is completely lost, summarization never executed, model guesses (normally wrong) and made up assumptions and try to resolve a problem I didnt asked for.
Is hard to bring more information, if anyone has any way to test this and bring more accurate descriptions, let me know.
This is happening near v0.75, I didnt changed model or model settings in the process.
- Qwen3.6 27b on LMStudio using context of 192000, YoLo active.
EDIT: One interesting behavior I didnt notice early, If I call summarization manually, agent picks up old context, and obtains the correct information to keep working.
Version
v0.76
Environment
Manjaro OS Up to date, Wezterm / Wezterm.
Description
In previous versions crush executes a summarization then cleanup the context and injects this summarization to continue working.
Currently, every time I reach 80%> context is completely lost, summarization never executed, model guesses (normally wrong) and made up assumptions and try to resolve a problem I didnt asked for.
Is hard to bring more information, if anyone has any way to test this and bring more accurate descriptions, let me know.
This is happening near v0.75, I didnt changed model or model settings in the process.
EDIT: One interesting behavior I didnt notice early, If I call summarization manually, agent picks up old context, and obtains the correct information to keep working.
Version
v0.76
Environment
Manjaro OS Up to date, Wezterm / Wezterm.