Idle Compaction, Permission Defaults, and a Sixteen Point Eight Billion Dollar Quarter

I'm Shannon, and this is the Claude Notes Brief -- your weekly rundown of Claude Code updates and Anthropic news for the week of Monday, October nineteenth. Long idle sessions now compact themselves before the cache expires. Headless sessions start with standard permissions. And Anthropic reported sixteen point eight billion dollars in quarterly revenue.

Let's start with Claude Code, where four releases landed this week and the headline is idle compaction -- if you leave a long conversation sitting untouched, Claude Code now compacts it for you before the prompt cache expires. Here's the problem it solves. Prompt caching lasts about an hour. So if you've got a two hundred thousand token conversation going and you step away for lunch, your next message pays to reprocess that entire history from scratch.

Now, at around fifty-four minutes of idle time, Claude Code compacts the session just ahead of that expiry. You can tune the token threshold with an environment variable, switch it off in your settings file, or toggle it from a new idle compaction row in the configuration screen. That theme of smarter context spending shows up elsewhere too. There's a new setting that stops at-mentions from inlining file contents -- instead of dumping the whole file into your prompt, Claude gets the path and reads selectively with its file tools.

If you regularly mention large files, that's a meaningful difference in how much context you burn. The changes most likely to surprise you, though, are around permissions. Headless sessions -- that's running Claude Code with the print flag, or using the Agent SDK -- now start in standard permission mode instead of auto, specifically on third-party providers, with telemetry off, or on a fresh install. If you were relying on the old behavior, you'll want to pass the permission mode flag explicitly.

And MCP allow rules got stricter. A rule naming one server no longer leaks across to servers with similar names that share a prefix. You'll want to write out the server's full normalized name with a wildcard instead, and the same applies to an agent's tools list. If you have MCP rules with underscores in the server name, check them.

Fork subagents also became the default outside the terminal -- they're now on in headless mode, the Agent SDK, the desktop app, and Visual Studio Code, matching terminal behavior. The desktop app and Visual Studio Code now always run subagents in the background rather than letting Claude decide. In Visual Studio Code, Claude's responses now render LaTeX math, both display blocks and inline, and there's a new setting to stop editor selections from being appended to your messages. Over in the terminal, vim mode picked up the percent motion for jumping to a matching bracket, on its own or combined with delete, change, and yank.

And for anyone wiring Claude Code into observability tooling, OpenTelemetry traces now carry standard GenAI semantic convention attributes -- provider name, conversation ID, error type -- and cache creation tokens are split out by duration. There's a new tracing cookbook that walks through the spans attribute by attribute, and we'll link it in the show notes.

On the under the hood side, this was a security-heavy week. Plan mode no longer executes writes that a pre-tool hook had allowed through. On macOS, sandboxed commands can no longer reach a denied path by way of a symlink, or write into a plugin directory folder. And feature flag, startup, and fast mode requests no longer carry a third-party gateway's API key back to Anthropic.

The Read tool also now refuses image sources over one hundred twenty-eight megabytes rather than loading them whole into memory -- which bridges nicely into the performance work. Plugin command and agent files now load with far fewer filesystem calls, and that speeds up startup noticeably on Windows. The Visual Studio Code panel stopped re-parsing the full thinking text on every update. And two memory leaks are fixed: unbounded growth under debug logging, and growth after resuming sessions with a lot of subagents.

Now to the broader Anthropic news, and the financial picture is the place to start. The Wall Street Journal reports Anthropic disclosed sixteen point eight billion dollars in third-quarter revenue, up forty-five percent, ahead of a planned IPO. A large share of that comes from developer and API usage. At the same time, Axios has a scoop that the Federal Trade Commission has subpoenaed both Anthropic and OpenAI, demanding documents over consumer harms tied to recent safety incidents.

That regulatory pressure sits alongside a warning from Anthropic itself, reported in The Guardian, that models may be months away from being able to copy themselves onto everyday networked devices without data center access -- which is useful context for the steady stream of sandbox and permission hardening we just walked through. Anthropic also published two posts worth a look: an update on how independent outside groups evaluate its models, and an outline of election safeguards ahead of the twenty twenty-six US midterms. And one partnership change -- TechCrunch reports that Replit is no longer working with Anthropic, according to its CEO. If you build on third-party coding platforms, model sourcing decisions like that one can change what's sitting behind your tooling.

That's it for the brief. I'm Shannon, and we'll see you next week.