Claude Code 2.1.141: MCP Fixes, Permission Dialogs, Agent Cleanup

TL;DR

• Major MCP server fixes: auth failures, config parsing, HTTP/SSE reconnection logic
• Permission dialogs now auto-dismiss when mode changes, explain why they appeared
• Background agents clean up properly; idle sessions auto-retire after 5 minutes
• 50+ bug fixes across terminal rendering, keybindings, voice mode, and Bedrock auth

New

• terminalSequence field in hooks — emit desktop notifications, window titles, and bells without a controlling terminal
• CLAUDE_CODE_PLUGIN_PREFER_HTTPS — clone GitHub plugin sources over HTTPS instead of SSH for environments without SSH keys
• ANTHROPIC_WORKSPACE_ID — scope minted tokens to specific workspaces under workload identity federation
• claude agents --cwd — filter session list to a specific directory
• Rewind "Summarize up to here" — compress earlier context while keeping recent turns intact
• Improved spinner feedback — spinner warms to amber after 10 seconds to signal Claude is still thinking
• /feedback recent sessions — include last 24 hours or 7 days of sessions for multi-session issues

Fixed

• Permission dialogs now auto-dismiss when permission mode changes and the new setting permits the tool
• Auto mode permission dialog explains when a permissions.ask rule caused the prompt
• "View diff in your IDE" option restored on file-edit permission prompts when IDE is connected
• Background agents preserve current permission mode instead of reverting to default
• Agents that finish work but leave a background shell now move to Completed instead of staying under Working
• Background side-queries on Bedrock/Vertex/Foundry no longer send unavailable Haiku model ID — falls back to main-loop model
• claude daemon status and /doctor on Windows now show underlying errors instead of opaque failures
• claude agents no longer shows agent-type list when launched through wrapper with flags
• Background jobs on custom ANTHROPIC_BASE_URL gateway now get auto-named correctly
• /model in one session no longer silently changes autocompact threshold in other concurrent sessions
• Pressing Enter while permission/dialog prompt is open no longer also submits text in input box
• Hooks no longer receive non-existent transcript_path after EnterWorktree switches directory
• Markdown tables with cell wrapping now render as bordered grid instead of falling back to vertical layout
• Cancelled prompts no longer removed from Up-arrow history, avoiding duplicate entries
• Ctrl+C now interrupts running turn while in vim INSERT/VISUAL mode
• Alternative chat:submit keybindings (meta+enter, ctrl+enter) work when enter is rebound to chat:newline
• Prompt suggestions no longer silently disabled when output style is configured
• spinnerVerbs setting now honored in turn-completion messages
• AskUserQuestion popup no longer hides last line of preceding chat content
• Web Search status shows correct search count instead of "Did 0 searches" on errors
• Multi-line statusline output no longer drops or corrupts rows exceeding terminal width
• Light-ansi theme now uses black for diff context lines on light backgrounds instead of invisible white
• Error overlay no longer dumps minified bundle source hiding original error message
• Pressing Enter after typing feedback survey rating digit now submits rating instead of chat message
• Pressing x on selected subagent now stops agent instead of typing into prompt
• Session title no longer derived from plugin monitor notifications before user's first prompt
• /tui no longer silently drops running background shells and subagents — now refuses and asks to wait
• Welcome banner now shows provider name on Bedrock, Vertex, Foundry instead of "API Usage Billing"
• /mcp server list keeps focused server visible in short terminals in fullscreen mode
• Redaction in /feedback bundles no longer produces invalid JSON for quoted values
• Desktop and third-party provider sessions no longer incorrectly inherit apiKeyHelper/ANTHROPIC_AUTH_TOKEN
• Early analytics events no longer silently dropped before logger initialization
• claude plugin install no longer fails for plugins with missing marketplace ref when sha is pinned
• Plugin details pane now shows correct MCP server count for plugins declaring via .mcp.json
• Plugin MCP servers with unset config variables show "config issue" message with fix-it hint; malformed .mcp.json no longer drops other servers
• MCP server configs using POSIX shell parameter expansions no longer incorrectly flagged as missing variables
• MCP HTTP/SSE servers returning 403 on connect now show as "needs auth" instead of "failed"
• Remote MCP servers no longer disconnect unnecessarily when optional server-events stream fails — tool calls continue over POST
• Remote Control MCP connectors no longer fail with 401 when worker session token rotates mid-session
• Remote Control no longer re-enrolls trusted device on stale token rejection — avoids /login loop
• Early OTel spans no longer silently dropped in SDK/headless mode with beta tracing enabled
• Custom voice:pushToTalk keybindings and "space": null unbinds no longer silently ignored
• Windows Alt+V image paste no longer reports "no image found" for clipboard screenshots
• SDK "Claude Code native binary not found" error fixed on Linux with both glibc and musl packages installed
• Bedrock: awsCredentialExport now always runs when configured, fixing cross-account access auth
• VSCode: in-chat mic now shows feedback when microphone produces only silence — displays "No audio detected"
• VSCode voice mode: WSL error now suggests installing sox libsox-fmt-pulse for WSLg users
• claude agents launch no longer fails when pre-warmed background worker is unhealthy — falls back to fresh launch
• claude agents no longer shows empty placeholder sessions from backgrounding fresh REPL; shows onboarding text when entered via ← with no other agents
• Empty idle background sessions from ← now automatically retired by daemon after 5 minutes

Update: npm install -g @anthropic-ai/claude-code@latest or check your IDE extension marketplace for Claude Code updates.

Source: Claude Code