Claude Code in the terminal warns you when resuming an old conversation will eat into your limit
When you go back to a saved conversation in Claude Code in the terminal on a Pro or Max plan, with the /resume command for example, a Resume this conversation? dialog with an estimate of your limit usage may now appear. The dialog arrived in version 2.1.304 on Monday 19 October. The estimate shows how much of your five-hour limit loading the saved conversation alone will use, and leaves out whatever you have already used. The dialog appears when a long conversation is resumed after a long break or on a different model. You then get two options: Resume, which takes you back into the old conversation, or Start a new conversation.
Resuming costs something because Claude Code sends the whole conversation to the model again, and reads from the prompt cache only the part that hasn’t changed and hasn’t expired yet. After a long break the cache has already expired, and after switching models there is none, because every model keeps its own – so the entire history of a long conversation then counts against your limit. That usage comes off your limit with the first message you send after choosing Resume, not at the moment you click it. If you go for a new conversation, you can still open the old one later with /resume. If there’s a note next to Start a new conversation saying that Claude can look at the previous conversation, just mention the earlier work in the new one and Claude will find it.
Before version 2.1.304, resuming an old conversation cost exactly the same – Claude Code just didn’t warn you. We recommend a new conversation when all you need from the old one is a decision or two, not the whole thread. When you leave a session open for a while instead, Claude Code handles it differently – it summarises a long conversation while you take a break.
We go through when to summarise a conversation and when to start a new one in our tip Where tokens go – and how not to burn through your limit.
/usage applies fixes from its recommendations with the f key
The /usage command in Claude Code can now carry out some of the recommendations it shows, starting with version 2.1.305 from Tuesday 20 October. /usage already showed a usage breakdown and recommendations in its Recommendations rows. Version 2.1.299 from 14 October added warnings about, among other things, the max effort level and large tool results. What’s new is that each recommendation is now a single line: what to fix and where the usage comes from – a specific CLAUDE.md file, for instance. When a hint such as f to fix 2 of 3 appears at the bottom of the window, press f and Claude Code will deal with the fixes it is able to make. In the VS Code extension, a button below the recommendations in the Account & usage window does the same thing – Fix both, for example.
Once you press f, Claude Code waits for the current response to finish and then changes the effort level without asking wherever the recommendation is an /effort command. For the other fixes it can handle, Claude gets a single instruction with a separate step for each recommendation, and has to show you the change before making it – trimming an overly long CLAUDE.md file, say. Some recommendations trigger nothing at all, such as the advice to type /compact before a break. If Claude doesn’t show you a change for a recommendation, that one is yours to carry out by hand.
The percentage next to a recommendation tells you what share of your usage it concerns, not how much you will save – 40% next to a recommendation to change the effort level doesn’t mean you’ll save 40% of your limit. The figures are approximate and come from the session history on this computer, so conversations on claude.ai and on other devices aren’t counted. You also won’t see the f hint while Remote Control is connected. We think /usage is worth a look for anyone whose limit regularly runs out before it resets.
What else /usage shows, and how to spread your limit across the whole day, is in our tip Where tokens go – and how not to burn through your limit.
VS Code extension: background subagents on one line, and why dictation stopped
The Claude Code extension for VS Code, from version 2.1.304 released on Monday 19 October, shows a subagent working in the background as a single status line under the subagent’s Agent card, instead of printing every step it takes into the chat. When Claude Code hands a background subagent a job such as going through a dozen price-list files, the chat keeps showing your conversation with Claude Code rather than dozens of the subagent’s steps. You can still check what the subagent did: type /tasks in the message box and click the subagent in the window that opens, and its card shows the tool calls and a button that opens the full record of its work.
Version 2.1.307 of the extension, out on Wednesday 21 October, also makes the microphone button tell you why dictation stopped. Dictation already ended before, after 15 seconds with no recognised words or after two minutes of recording. Now a message by the button gives the reason and tells you that the text recognised so far stays in the message box, unsent. To dictate the rest, click the microphone again. When you’re dictating a longer task description, two minutes go by faster than you’d think. Thanks to the message, you know straight away that the microphone works and the recording simply ended.
The Code tab in Claude Desktop – what it does better than the terminal, and what’s new this week
Starting this week, our News covers not only Claude Code in the terminal and in VS Code, but also the Code tab in the Claude Desktop app. The app gets new versions several times a week, and the Code tab has picked up quite a few things the terminal doesn’t have.
The Code tab in Claude Desktop is Claude Code in a desktop app window. It runs on the same engine and reads the same configuration files as the terminal, so it works with your CLAUDE.md files, skills, hooks, MCP servers and the same models from the start. A conversation you began in the terminal moves over to the app with the /desktop command. The tab runs on Windows and macOS, and on Linux in beta.
In our view the Code tab has caught up with the terminal and pulled ahead of it in a few places – at least for someone who works with documents rather than writing scripts:
- You attach a PDF or a photo to a request with the + button next to the message box, or by dragging the file onto the message box.
- A file Claude Code is changing right now opens next to the conversation in the file panel, and you can edit it there yourself. In the browser panel you can check a web page without switching windows.
- A list at the side of the window holds several conversations at once, and the app sends you a notification when one of them finishes its task.
- Recurring tasks, such as a morning summary from your calendar and inbox, are set up in the Scheduled tasks window, and plugins and connectors are added from a menu, with no commands.
- A mod that Claude Code writes at your request can redraw parts of the conversation or add a panel of its own. In the app, a mod like that can also display an SVG graphic, which the terminal can’t show. We explained what mods are in our News from 5 October.
The Code tab has its own copy of Claude Code with its own version number, and updating the terminal doesn’t update the app, nor the other way round. So what’s new in the terminal sometimes reaches the app later – version 3.6556.1 of the app from 22 October contained Claude Code 2.1.308, while the terminal got 2.1.310 a day later. To see the version in the app, type /status in a local session in the Code tab and look at the Claude Code row.
Version 3.1623.0 of the app, from Tuesday 20 October, added a Storage section to the settings. It shows how much disk space Claude takes up and lets you free some of it by deleting inactive sessions from the Code tab and tasks from Cowork, the neighbouring part of the app. Handy on a laptop with a small drive. You can also open it by typing /storage in the Code tab. The same version no longer loses messages you add in a local session while Claude is still working: if the app closes, updates or crashes, the message comes back the next time it starts and waits for you to send it. Version 3.6556.1 from Thursday 22 October improved dictation across the whole app – a click or Enter meant to end dictation no longer sends the message when the recording has just failed.
To finish: AI agents helped two mathematicians prove the 31-year-old KLS conjecture
Xinzhi Zhang of Microsoft Research and Zhao Song presented a proof of the Kannan–Lovász–Simonovits (KLS) conjecture, which had been waiting 31 years to be settled. The conjecture comes from high-dimensional geometry and probability theory, and asks whether, when you split a high-dimensional probability distribution into two parts, the best flat cut is always almost as good as the best cut of any shape. The answer is yes, whatever the number of dimensions. Microsoft described the result on its blog on Wednesday 21 October, and the authors posted the preprint to arXiv on 1 October. The dimension-independent result only appeared in the version updated on 4 October.
Zhang and Song tried more than a hundred approaches with AI agents: many agents at once checked variants of a single idea, while the researchers assessed the intermediate results, dropped weak directions and passed the promising ones on to the next round. The two researchers chose the direction and the ideas and checked the final proof themselves; according to Microsoft, the recursive method and the argument that bounds the error were their own contribution. Separately, several leading AI models made attempt after attempt with no human guidance, using a similar number of tokens in total – and didn’t get there. Microsoft doesn’t say which models or tools were used.
For now the proof is a preprint on arXiv. After a final review, the authors intend to produce a shorter proof and check it in Lean, a tool for verifying proofs by computer – as of today, that is only an announcement.