Claude Code

Walk away from a long chat and Claude Code tidies it up. Terminal sessions over 200K tokens compact after ~54 minutes idle, so the next message does not reprocess everything. CLAUDE_CODE_DISABLE_FILE_MENTIONS passes @-mention paths instead of file contents. OpenTelemetry now splits cache writes into 5m and 1h buckets.

Subagents get their own memory budget and model. Each subagent can set its own auto-compact window, workflow subagents can pick a different model, and overloaded retries wait longer before trying again.

Keep Claude Code’s file edits inside your project folders. A new setting, permissions.blockWritesOutsideWorkingDirectories, makes the Write, Edit and NotebookEdit tools refuse paths outside your working directories. The release came out late on Oct 29, 2026.

Action needed: Claude Code deprecated fast mode for Opus 4.8 as of Oct 30, 2026, and removes it on Nov 30, 2026. After that, fast mode requests on Opus 4.8 return an error and don’t fall back to standard Opus 4.8.

Your headless Claude Code jobs now get a heads-up before the stored sign-in runs out. The latest release adds a one-line warning to claude -p text output starting 10 days before a stored claude.ai sign-in ends, and shows the end date in claude auth status.

A scheduled task that quietly picks up connectors now gets called out. Claude Code’s latest release warns in /schedule when a new cloud scheduled task gets connectors you didn’t ask for, and shows how to remove them.

Claude Code quietly shortens the prompt for custom model names. On Bedrock, Vertex, Foundry and Mantle, names it doesn’t recognize, even ones that point at Sonnet or Haiku, now get the shorter system prompt. Set CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 to keep the full prompt.

Update now: long Claude Code sessions stop dead-ending on a credits error. Standard-context sessions stopped with “Usage credits required for 1M context” when a request ran over 200K tokens. The release out Oct 26, 2026 compacts the conversation and keeps going instead.

Bedrock and Vertex users, update now: Claude Code’s failing requests are fixed. The previous release could fail every request on Vertex, Bedrock and other providers when the endpoint or a gateway rejected tool search. Claude Code now retries without it.

Back to an old chat after a break? On Pro and Max plans, Claude Code now warns you when resuming an inactive conversation would use a noticeable share of your 5-hour usage limit.

Resume anyway, or start a new conversation and Claude can look up the previous one if you refer to it.

Want long chats left alone while you’re away? A new Idle compaction row in /config turns that feature off or back on, and the notice after an idle compaction now says how.

Also changed: fork subagents are on by default in claude -p, the Agent SDK, the desktop app and VS Code, as they already were in the terminal.

Switch to a model with a smaller context window and Claude Code makes the chat fit. A conversation too long for the new model is now summarized by the previous model first, and a fix covers a case where that summary could spend usage credits without asking.

Unattended retry mode can now also give up on a usage limit that resets too far away.

A broken safety hook no longer waves the action through. The latest Claude Code release adds onFailure: "block" for command and HTTP hooks: a hook that can’t start, times out or exits with an unexpected code now blocks the action.

Your deny rules stop slipping through side doors. The latest Claude Code release applies Bash deny and ask rules behind environment variable prefixes, holds them on nested parts of compound commands on managed machines, and applies Read deny rules to files reached through a symlink.

Proxy users, update now: Claude Code stops failing every request. Update to 2.1.276 or later: it fixes the 400 advisor tag error when ANTHROPIC_BASE_URL points at a proxy.