When the context size reaches roughly 280 – 300 k tokens, both models become unusable: • Code-Supernova stalls for 5 – 10 minutes (“Planning next step”), then another long pause before any output; an hour of waiting produced only ~80 lines of SQL. • Gemini 2.5 Pro stops the task once the context hits ~280 k. If I ask it to continue, it simply repeats the previous code block and stops again. The issue is reproducible and makes working with large repos impractical.
Steps to Reproduce
Open a large repository (≈ 3 k files).2. Ask the model to do a non-trivial refactor, e.g. “apply SQL migrations and update all ORM models.”3. Wait until the status bar shows Context 286 k / 800 k. 4. Observe: Code-Supernova hangs for > 20 min per step; Gemini stops and echoes the same code.
Expected Behavior
The model should keep generating without multi-minute stalls or duplicate messages until the true model limit is reached.
Hey, thanks for the report. It looks like you’re seeing two separate issues:
AI hanging/repeating around ~280k context: this behavior at 286k tokens is unusual since you’re within stated limits.
Try:
Starting a fresh chat to see if performance improves
Confirming that Max Mode is enabled
Reducing context by being more selective with @-mentions
Chat history loss after reboot: this is a known bug affecting multiple users. The team is investigating. As a workaround, periodically export important chats (three dots at the top of the chat → Export Chat) or use the SpecStory extension.
For the AI performance issue: does reducing context and starting a new chat stop the stalling? That’ll help us tell if it’s context-size related or not.
I’ll confirm your assumptions: All models were used in “MAX MODE”.
Regarding new chats and reducing the context window:
I’ll admit that I haven’t used Gemini 2.5 Pro outside of MAX MODE - because even at 200k there are neural networks that work much better than it.
As for Gemini 2.5 Pro and Code-Supernova in new chats: They work normally. Gemini 2.5 Pro in a new chat becomes responsive and doesn’t freeze. Code-Supernova works quickly and the same requests that took it 20 minutes or longer can be completed in 2-5 minutes.
Thanks for confirming that fresh chats resolve the performance issue, that’s helpful data.
Since the slowdowns are clearly context-related, I’ll pass the Code-Supernova performance issue to the team. They’ll investigate why both models degrade around 280k tokens even though this is within the stated limits.
I also noticed an issue with Gemini 2.5pro. When reaching around the 230k token mark, or 280k+ tokens, it starts having problems with editing project files. It literally writes that there was a file editing error, buuut it says everything is ready, go grab your code. 6c167db4-637f-49af-a972-eda3c7594ae5
