Claude Code 2.1.76: MCP Elicitation, Voice Fixes

TL;DR

• MCP servers can now request structured input mid-task via interactive dialogs
• New elicitation hooks let you intercept and override MCP responses
• Fixed deferred tools losing input schemas after compaction, breaking array/number parameters
• Voice mode, slash commands, and Remote Control session handling all improved

New

• MCP elicitation support — MCP servers can request structured input mid-task through interactive dialogs with form fields or browser URLs.
• Elicitation hooks — New Elicitation and ElicitationResult hooks intercept and override responses before sending them back.
• Session display name flag — Use -n / --name <name> to set a display name at startup.
• Sparse checkout for monorepos — worktree.sparsePaths setting lets claude --worktree check out only needed directories via git sparse-checkout.
• PostCompact hook — New hook fires after compaction completes.
• /effort slash command — Set model effort level directly in chat.
• Session quality survey — Enterprise admins can configure sample rate via feedbackSurveyRate setting.

Fixed

• Deferred tools losing input schemas after conversation compaction, causing array and number parameters to be rejected.
• Slash commands showing "Unknown skill" error.
• Plan mode asking for re-approval after plan was already accepted.
• Voice mode swallowing keypresses when permission dialog or plan editor was open.
• /voice not working on Windows when installed via npm.
• Spurious "Context limit reached" error when invoking skills with model: frontmatter on 1M-context sessions.
• "Adaptive thinking is not supported on this model" error with non-standard model strings.
• Bash(cmd:*) permission rules not matching when quoted arguments contain #.
• "Don't ask again" in Bash permission dialog showing full raw command for pipes and compound commands.
• Auto-compaction retrying indefinitely after failures — circuit breaker now stops after 3 attempts.
• MCP reconnect spinner persisting after successful reconnection.
• LSP plugins not registering servers when LSP Manager initialized before marketplaces were reconciled.
• Clipboard copying in tmux over SSH — now attempts both direct terminal write and tmux clipboard integration.
• /export showing only filename instead of full file path in success message.
• Transcript not auto-scrolling to new messages after selecting text.
• Escape key not working to exit login method selection screen.
• Remote Control sessions silently dying when server reaps idle environment, rapid messages queued one-at-a-time instead of batched, and stale work items causing redelivery after JWT refresh.
• Bridge sessions failing to recover after extended WebSocket disconnects.
• Slash commands not found when typing exact name of soft-hidden command.
• Gitignore patterns containing commas silently excluding entire filetypes from @-mention file picker (VSCode).

Improved

• --worktree startup performance — now reads git refs directly and skips redundant git fetch when remote branch is already available locally.
• Background agent behavior — killing a background agent now preserves its partial results in conversation context.
• Model fallback notifications — always visible with human-friendly model names instead of hidden behind verbose mode.
• Blockquote readability on dark terminal themes — now italic with left bar instead of dim text.
• Stale worktree cleanup — worktrees left behind after interrupted parallel runs are now automatically cleaned up.
• Remote Control session titles — now derived from your first prompt instead of showing "Interactive session".
• /voice — now shows your dictation language on enable and warns when your language setting isn't supported for voice input.
• --plugin-dir — updated to accept one path per flag to support subcommands; use repeated --plugin-dir for multiple directories.

Update via npm install -g @anthropic-ai/claude-code or your package manager.

Source: Claude Code