Version 2.1.20

New /keybindings skill enables full customization of keyboard shortcuts via ~/.claude/keybindings.json configuration file.

/keybindings

• Create custom keybindings organized by context (Global, Chat, Autocomplete, etc.)
• Support for chord bindings (multi-keystroke shortcuts like ctrl+k ctrl+s)
• Validation integrated into /doctor command with detailed error/warning messages
• Unbind default shortcuts by setting them to null
• Schema validation with JSON Schema support for editor autocomplete
• Documentation at https://code.claude.com/docs/en/keybindings

Skill registration at line 748555 (name: "keybindings-help", contains "Use when the user wants to customize keyboard shortcuts")

New --debug-file CLI flag writes debug logs to a specific file path.

claude --debug-file /path/to/debug.log
claude --debug-file=/path/to/debug.log

• Implicitly enables debug mode when specified
• Useful for persistent debugging and troubleshooting

Argument parser at line 2836 (if (K.startsWith("--debug-file="))) and help text at line 577087 ("--debug-file <path>")

New --hard-fail CLI flag causes Claude Code to exit immediately on certain errors rather than attempting recovery.

claude --hard-fail

• Useful for CI/CD pipelines and automated testing
• Ensures failures are detected rather than silently handled

Detection at line 5582 (return process.argv.includes("--hard-fail"))

New --who-owns integration for Alpine Linux (apk) package manager to verify file ownership.

Internal feature - automatically used when verifying Claude Code installation on Alpine Linux.

• Detects apk-managed installations with logging: "Detected apk installation: ..."
• Also detects deb, rpm, and pacman installations
• Enables better package manager integration and update guidance

APK query at line 419998 (OP1("apk", ["info", "--who-owns", K]) and detection logs at lines 420004, 420011, 420018, 420025

New SendMessageTool replaces previous teammate messaging implementation with clearer protocol separation.

Internal tool used by agents in swarm mode.

• Handles message, broadcast, request, and response types
• Protocol subtypes for shutdown and plan approval workflows
• Better error messages for missing parameters
• Improved in-process teammate abort handling

Tool definition at line 448126 (SendMessageTool: () => tT2) with logging prefix [SendMessageTool] at lines 448249, 448275, 448284

New ctrl+k ctrl+k chord keybinding to trigger /compact command.

ctrl+k ctrl+k

• Quick access to conversation compaction
• Uses the new chord keybinding system
• Replaces removed ctrl+k ctrl+c for /commit

Keybinding definition at line 486662 ("ctrl+k ctrl+k": "command:compact")

New feature or mode referenced in codebase (details unclear from minified code).

Internal implementation detail.

String literal at line 388067 ("Newspapering")

Significantly improved Unix domain socket security with ownership verification and stale socket cleanup.

• Socket directory ownership check: warns if directory owned by different user
• Stale socket cleanup: removes sockets for non-running PIDs
• Better seccomp filter path discovery for global @anthropic-ai/sandbox-runtime installations
• Added .npm-global to seccomp binary search paths
• Improved error messages with actionable installation instructions
• Warning when seccomp binaries unavailable: "[Sandbox Linux] Seccomp binaries not available - unix socket blocking disabled. Install @anthropic-ai/sandbox-runtime globally for full protection."

Socket ownership check at line 523583 ("Socket directory not owned by current user"), stale socket cleanup messages, seccomp search paths at line 95936 (".npm-global"), warning at line 96398

MCP browser tool's accessibility tree output limit is now configurable.

• Default limit remains 50,000 characters
• New parameter allows setting custom limits for clients that can handle larger outputs
• More flexible description: "Output is limited to 50000 characters by default" (was: "Output is limited to 50000 characters")
• New parameter: "Maximum characters for output (default: 50000). Set to a higher value if your client can handle large outputs."

Description change at line 522359 and parameter at line 522387

Compaction notification now shows "✻ Conversation compacted" with visual symbol instead of plain text.

• Old: Conversation compacted · {keybinding} for history
• New: ✻ Conversation compacted ({keybinding} for history)
• More visually distinct in transcript

Notification text at line 410281 ("✻ Conversation compacted (")

Refined safety check messaging for workspace trust.

• Slightly more concise wording in safety prompt
• Old: "Quick safety check: ... If not, take a moment to review what's in this folder first."
• New: "Quick safety check: ... If not, take a moment to review what"
• Message continues with context about Claude Code's file access capabilities

Safety message at line 570322

Clearer guest pass sharing notification in UI.

• More prominent formatting: "You have free guest passes to share · /passes"
• Encourages user engagement with referral system

Notification at line 562731

More detailed logging for task claiming and agent operations.

• New logs: [inProcessRunner] Claimed task #N: {subject}
• Task failure logging: [inProcessRunner] Failed to claim task #N
• Task list check errors: [inProcessRunner] Error checking task list:
• New logs for task assignment and completion

Logs at lines 422622, 522622, 522628 with prefix [inProcessRunner]

Explicit layout engine logging for debugging rendering issues.

• Logs: "Layout engine: yoga" and "[render] initLayout starting/complete"
• Helps diagnose UI rendering problems

Engine log at line 174577, render logs at lines 182436-182438

Automatic repair of malformed tool result pairings in conversation history.

• Detects and repairs missing tool_result blocks
• Logs: ensureToolResultPairing: repaired missing tool_result blocks (X -> Y messages)
• Prevents conversation corruption from improperly paired tool uses

Repair logging at line 469512

Improved teammate communication and protocol handling.

• Operation descriptions now mention "discoverTeams to list available teams to join"
• Better error messages for missing parameters in teammate operations
• Clearer separation between message types (message, broadcast, request, response)

Operation description updates in SendMessageTool schema, error messages for missing recipient, request_id, etc.

Removed automatic git commit trailer attribution system.

• Removed all "Attribution hook:" logging
• Removed Claude-Generated-By: git trailers
• Removed Claude-Escapes:, Claude-Permission-Prompts:, Claude-Session:, Claude-Steers: trailers
• No longer tracks file contributions or calculates Claude percentage
• Simplifies git workflow and removes potentially confusing commit metadata

All Attribution: and Attribution hook: log strings removed from codebase (verified absent in v2.1.20)

Removed ctrl+k ctrl+k keybinding for /commit command.

• Previous binding: ctrl+k ctrl+c → /commit
• Now unbound by default
• Users can re-add via custom keybindings if desired

Binding "ctrl+k ctrl+c": "command:commit" existed in v2.1.19 line 483280, not present in v2.1.20

Simplified permission request UI removed some verbose text and interactive elements.

• Removed: "Do you trust the files in this folder?"
• Removed: "(shift+↑/↓ to select)" navigation hints
• Removed: Several verbose permission dialog messages
• Streamlined to clearer, more concise safety checks

Strings like "Do you trust the files in this folder?", "shift+↑/↓ to select" removed from v2.1.20

GitHub pull request status displayed in footer area.

Feature-flagged behind tengu_pr_status_cli

• Settings option: "Show PR status footer" (enabled by default when flag is on)
• Infrastructure exists in settings menu at line 480230
• Will display PR status information in CLI footer when enabled

Feature flag check at line 480226 (i4("tengu_pr_status_cli", !1)) with settings UI at line 480230 (label: "Show PR status footer")