Rejected MCP tool schemas
Claude Code v2.1.300: one rejected MCP tool no longer fails every turn of your conversation
When the API rejects an MCP tool’s input schema
- Before
- every turn failed with a 400 error
- Now
- that tool is sent without parameters for the rest of the conversation
When the API rejected one MCP tool's input schema, every turn of your conversation failed with a four hundred error. Claude Code now sends that one tool without parameters, and the conversation keeps going.
Exit stops background commands
>Exiting Claude Code now stops the background commands it started
- After 1.5 seconds
- a background command still running is force-killed
- Fullscreen mode
- could leave those commands running before
- SessionEnd hooks
- run even when a background command is still going
Quit Claude Code with a background command still going, and it asks that command to stop, then force-kills it a second and a half later. Your SessionEnd hooks also run on that exit.
Hook input is checked
>Input a hook rewrites is now checked like Claude’s own
- updatedInput
- from PreToolUse and PermissionRequest hooks: unknown keys are dropped
- Input the tool rejects
- denies the call
- Agent hooks
- fall back to the caller’s model when theirs is unavailable, instead of allowing the action
When a PreToolUse or PermissionRequest hook rewrites a tool's input, Claude Code checks it like Claude's own input. Unknown keys are dropped, and input the tool rejects denies the call.
Mods: prompt text and agent views
>Mods can draw text in the prompt box and open an agent’s transcript
- virtualText
- drawn in the prompt box, never part of the draft or the submitted prompt
- $.ui.view({ agentId })
- shows that agent’s transcript in the main view
- prompt.change
- fires on a history recall, a clear, an undo or a plugin’s fill
Mods can now draw unsent text in the prompt box and open an agent's transcript in the main view. A new prompt change hook fires when undo or history recall rewrites the prompt.
Remote Control server
>claude remote-control drops its sandbox flags and retries after a sign-in error
- --sandbox, --no-sandbox
- had no effect; passing either now exits with an error
- Brief sign-in errors
- such as a locked keychain: the server retries, and warns when you need to log in
- Several sessions at once
- no longer marked stopped while still running
Claude Code's remote control server drops its sandbox flags, which never did anything. Passing either one now exits with an error. The server also retries after a brief sign-in error, such as a locked keychain.
Advisor on AWS
>/advisor comes to Claude Platform on AWS
- Who gets it
- Claude Code on Claude Platform on AWS
- Default
- no advisor runs until you pick one
If you run Claude Code on Claude Platform on AWS, the /advisor command is there now too. It stays off until you pick an advisor yourself.
Claude in Slack
>Claude in Slack answers in Slack Connect channels and asks before a connector action
- Slack Connect
- off by default; an Owner or Claude Tag admin turns it on
- Allow once or Deny
- a private card in the thread; others only see that Claude is waiting on you
- Voice clips
- Claude reads Slack’s transcript when Slack has one
Claude in Slack can answer in channels shared with another organization, once an admin turns it on. When a connector action needs your approval, Claude now asks you privately instead of refusing.
Fixes you may have hit
>Fixed: vanishing subagents, reftable worktrees, VS Code fast mode
- /tasks
- subagents you messaged stay listed after they finish
- --worktree
- reopens worktrees in repositories using git’s reftable ref storage
- VS Code fast mode
- turns off when usage credits run out, instead of re-sending refused requests
Subagents you messaged stay in the tasks list after they finish. Worktrees reopen in repositories that use git's reftable storage. And in VS Code, fast mode turns off when your usage credits run out.
The full changelog
Every change is in the Claude Code changelog
https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md
The Claude Code changelog on GitHub lists the rest, including new OpenTelemetry fields for cache writes and more fixes for VS Code and cloud sessions.
