Claude Code v2.1.257 Changelog

This release is a big deal for developers using Claude Code, especially those who run complex tasks in the background or manage teams. It introduces a highly efficient new AI model and tightens up security so automated tasks don't accidentally touch files or systems they shouldn't. It also fixes a ton of annoying glitches with background sessions, making the whole tool feel much more reliable and smooth.

What's new

• New AI Model: Claude Fable 5.1 is now the default "Fable" model. It can remember up to 1 million tokens (about 750,000 words) at once, making it amazing for huge projects. It also costs less to run and is cheaper to read from its memory cache.
• Custom Time Settings: You can now choose how the clock displays time. Pick from 12-hour, 24-hour, or UTC time, or even use a custom code pattern (like strftime). This applies to the clock at the end of a turn and in your history logs.
• Stricter Auto-Mode Security: When running in "auto mode" (where Claude acts without asking you first), it will no longer automatically approve risky actions like grabbing cloud security keys, sneaking past network blocks, or reaching into other users' spaces. It will only do these if your system specifically expects them.
• Force Subagent Models: A new setting (CLAUDE_CODE_SUBAGENT_MODEL_FORCE) lets you force all "subagents" (smaller AI helpers Claude spawns to do specific tasks) to use the exact same AI model as the main agent, ignoring any other instructions.
• Temporary Effort Changes: Typing s in the /effort menu now changes how hard the AI works *just for this session*, matching how the /model menu works.
• Sandbox Warnings: A new /doctor check warns you if a "sandbox mask" (a temporary security file) gets left behind by a session that crashed or was killed.
• File Read Prompts: In auto mode, Claude will now ask for your permission the first time it tries to read a file outside your main working folder. You can even set a rule to block these reads completely.
• Gateway Descriptions: If you use a network gateway to find models, it can now send a description for each model in the /model menu. If there's no description, it just says "From gateway."

Improvements

• Background sessions and subagents now update their token counters live while they are thinking, so you can watch the usage go up in real-time instead of freezing.
• The "Completed" list in claude agents now sorts by the latest finish time, so sessions you just re-prompted won't stay buried at the bottom.
• Subagents are now smarter about network drops: if a connection drops, your computer goes to sleep, or the server glitches mid-response, the subagent will automatically continue instead of just stopping with a broken, half-finished answer.

Fixes

• Settings Loading: Settings put in a .claude/ folder after the app started weren't recognized until you restarted. Now they load automatically.
• Agent View Permissions: Sessions started from the agent view (opened with the ← key) used to ignore the folder's specific permission rules. They now respect them.
• Keyboard Shortcuts: Custom keyboard shortcuts for Ctrl+G were being ignored in claude agents. Shortcuts for Ctrl+S and Ctrl+T are also now fully customizable.
• Background Session Crashes: Fixed issues where background sessions wouldn't start during a self-update on Mac, or on Windows when a leftover lock file confused the system.
• UI Glitches: Fixed the loading spinner freezing while a command ran, and fixed a weird glitch where pasting colored text (like logs from a CI machine) into feedback boxes would crash the app.
• Memory Leaks: Fixed massive memory spikes that happened when piping non-standard data into the app, or when dealing with broken project files. It now fails fast with a clear error instead of eating all your RAM.
• Security Holes: Fixed a bug where plugins could use symlinks (shortcuts) to sneakily read files outside their allowed folders. Fixed an issue where network security rules didn't properly block websites with a trailing dot (like example.com.).
• Remote Control: Fixed a bug where hitting "Escape" or "No" on a Remote Control permission prompt was accidentally counted as a "Yes," letting the next request connect without asking. Also fixed it ignoring the model you picked.
• Scheduled Tasks: Fixed /schedule routines that had nothing to do because their prompts were saved incorrectly.
• Lost Prompts: Fixed a bug where a prompt you saved with Ctrl+S inside a background session would disappear if the session went idle or was stopped and reopened.
• Teammate Requests: Fixed an issue where a teammate's permission request was accidentally answered twice if the leader's mailbox was briefly locked.
• Token Counters: Fixed the token counter freezing or crawling after switching to another subagent's history.
• Fullscreen Mode: You can now click on ! shell command outputs to expand them while in fullscreen mode.
• Old Background Sessions: Fixed old background sessions piling up and not retiring after the app auto-updates.
• Terminal Glitches: Fixed claude agents --json messing up your terminal settings when it exits.
• Proactive Output: Fixed a bug where "Proactive" style sessions would spam filler messages and repeatedly read logs instead of waiting patiently for a background command to finish.
• Resuming Subagents: The main agent is now properly notified when you resume a subagent you had previously stopped.
• Adding Directories: Fixed /add-dir rejecting folders that were inside your current working folder. It now loads their skills and commands properly.
• API Keys: Fixed several bugs where leftover API keys or authorization headers were sent to the wrong cloud services (like Foundry, Vertex, or Bedrock), causing authentication conflicts.
• Bash Rules: Fixed a bug where rules to block certain Bash commands (like Read() or Edit()) weren't applied to file redirects (like < file) or reader commands like tac and egrep.
• Resuming Sessions: Fixed --resume listing a backgrounded conversation twice, and --continue reopening a broken, stalled copy. --continue now properly opens finished background sessions.
• Idle Loops: Fixed claude -p exiting 5 seconds early while a background monitor it started was still running. It now waits for the monitor to finish or time out.
• Auto-Mode Permissions: Fixed a bug where a rule asking for permission was skipped in auto mode if the command was run inside a compound command or subshell, letting it run without asking.
• MCP Servers: Fixed /mcp reconnecting to servers that should have been blocked by your security rules, and fixed claude mcp remove leaving behind stored passwords for remote servers.
• Dropped Tools: Fixed a bug where backgrounding a turn while a tool was running made the system think you rejected the tool, instead of just pausing and re-running it later.
• Cache Misses: Fixed sessions using an "advisor model" missing their memory cache on background requests, causing them to resend the whole conversation every time, which was slow and expensive.
• Left Arrow in Panels: Fixed the ← key doing nothing in the /btw panel inside an agent session. It now returns you to the agents list, and the panel comes back when you reopen the session.
• Timer Limits: Fixed a bug where setting a timer value too high (over 2.1 billion milliseconds) caused the system to fail or run a task every single millisecond. It now caps the value safely.
• Deleted Folders: Fixed claude --bg claiming it "backgrounded" a session when it was actually started from a folder you just deleted. It now prints the error and exits.

Breaking changes

No breaking changes in this release.