Claude Code in the terminal summarises a long conversation while you take a break
Since version 2.1.299, released on Wednesday 14 October, Claude Code in the terminal summarises a long conversation by itself when you step away from it for a while. The part of the conversation the model has already read goes into its prompt cache, and on a Claude subscription that cache lasts an hour. After a break of more than an hour the cache has expired, so Claude Code processes the whole conversation history again, and that counts towards your plan’s limit. Now, when a conversation has reached 200,000 tokens and about 54 minutes have passed since the last request to the model, Claude Code summarises it just before the cache expires. Your next message then sends the model the summary instead of the whole conversation.
According to the documentation, summarising during a break only works in an interactive Claude Code session in the terminal. It also depends on the one-hour cache, which Claude Code uses on a subscription only while you’re within your plan’s limit. Once the limit runs out and you carry on with usage credits – paying separately for work beyond the limit – the main conversation’s cache drops to five minutes and there’s no summarising during breaks. It also won’t happen if your computer is asleep when the 54 minutes are up. Until now a conversation could go past 200,000 tokens without any summary at all: Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 run in Claude Code with a million-token context window by default, and Auto-compact only steps in once the conversation starts to fill the window.
Summarising during a break is a good default for anyone who keeps one long conversation going all day and leaves it for lunch or a client meeting. The summary itself costs something too, but according to the documentation, when the cache is fresh, the cost is only a fraction of what the conversation’s size would suggest. Details can get lost in a summary, though – something you agreed on first thing in the morning, for instance. If you’d rather come back to the full history, you can turn it off in /config with the Idle compaction option – it has been available since version 2.1.302, released on Friday 16 October, and is on by default. Turning off Auto-compact turns off summarising during breaks as well.
What Auto-compact is and why it’s worth leaving on is covered in the lesson First-run settings in Claude Code.
The VS Code extension shows maths formulas in responses
Since version 2.1.301, released on Thursday 15 October, the Claude Code extension for VS Code shows LaTeX formulas in the chat panel as fractions and maths symbols. LaTeX is a text notation for formulas, full of backslashes, curly braces and dollar signs, and until now that raw notation was exactly what you got in the response. It comes in handy when you ask Claude to work out a formula for the margin in a price list or for an instalment in a quote: you see a fraction as a fraction, which makes it easier to check before you carry it over to Excel.
The panel tries to tell a formula apart from other uses of the same characters, so a price written as “$5” stays a price. You won’t see formulas in the extension’s terminal mode, that is, when the Use Terminal option is on in your VS Code settings – there LaTeX stays raw text.
To finish: FLUX 3 Image Klein – an open image model for an ordinary graphics card
On Thursday 15 October, Black Forest Labs (BFL) released FLUX 3 Image Klein, an image generation and editing model with open weights – you can download it from Hugging Face and run it on your own computer. Klein creates images from a description, edits existing graphics and can work from up to ten reference images. It has 5.6 billion parameters, and according to BFL its INT8 version runs on NVIDIA GeForce RTX cards with 8 GB of memory, for example using ComfyUI.
The FLUX Kommunity License lets individuals and companies with annual revenue under $5 million use Klein for free, commercially too. So under this licence, a sole trader or a small business below that threshold can use Klein to make illustrations for quotes or graphics for company posts without paying licence fees.
If you don’t have an NVIDIA graphics card with 8 GB of memory, you can try Klein in the free demo on Hugging Face Spaces, right in your browser. Just type in, say, a description of the graphic for your next company post.