Version 2.1.78

A new hook event type that fires when a turn ends due to an API error (rate limit, auth failure, billing error, etc.), distinct from the normal Stop event.

{
  "hooks": {
    "StopFailure": [
      {
        "type": "command",
        "command": "echo 'API error: $CLAUDE_HOOK_ERROR'"
      }
    ]
  }
}

• Fires instead of Stop when an API error ended the turn
• Fire-and-forget — hook output and exit codes are ignored
• Supports matching on error type: rate_limit, authentication_failed, billing_error, invalid_request, server_error, max_output_tokens, unknown

Hook event definition (search for "StopFailure") — added to the hook event schema with matcherMetadata on "error" field

Plugins now get a dedicated persistent data directory at ~/.claude/plugins/data/{id}/, accessible via the CLAUDE_PLUGIN_DATA environment variable and the ${CLAUDE_PLUGIN_DATA} template in plugin configs.

• Each plugin receives its own data directory, created automatically
• The directory persists across plugin updates and sessions
• On plugin uninstall, users are prompted: "Delete it along with the plugin?" with the data directory path shown
• Use --keep-data flag with uninstall to preserve the data directory

Plugin data path helper (search for "CLAUDE_PLUGIN_DATA") — new env var injected into plugin processes; --keep-data CLI option (search for "--keep-data")

Claude Code now proactively detects when sandboxing is configured but cannot actually run, and displays a clear warning at startup instead of silently falling back.

• Checks for missing dependencies (bubblewrap, socat on Linux; ripgrep), unsupported platforms, and WSL1 incompatibility
• Displays a warning: ⚠ Sandbox disabled: {reason} followed by Commands will run WITHOUT sandboxing. Network and filesystem restrictions will NOT be enforced.
• The /sandbox diagnostic UI now also shows ripgrep status and uses "seatbelt: built-in (macOS)" labeling on macOS

Sandbox check function (search for "getSandboxUnavailableReason") and startup warning (search for "Commands will run WITHOUT sandboxing")

Administrators can inject a custom model option into the model picker via environment variables, useful for enterprise deployments with custom-routed models.

export ANTHROPIC_CUSTOM_MODEL_OPTION="my-custom-model-id"
export ANTHROPIC_CUSTOM_MODEL_OPTION_NAME="My Custom Model"
export ANTHROPIC_CUSTOM_MODEL_OPTION_DESCRIPTION="Custom routed model for our org"

• The custom model appears as an additional option in the model selection UI
• If no name is provided, the model ID is used as the label
• If no description is provided, defaults to Custom model ({id})

Model picker injection (search for "ANTHROPIC_CUSTOM_MODEL_OPTION")

A new background task type monitor_mcp for MCP-based monitors that can run alongside the main conversation.

• Appears in the background tasks panel with status indicators ("done", "unread" for completed but unviewed)
• Displayed as "1 monitor" / "N monitors" in task counts

Task type definition (search for "monitor_mcp")

Voice mode now probes the arecord audio device before attempting to record on Linux, providing clearer error messages when audio hardware is unavailable. On WSL specifically, the error message now explains WSL2 with WSLg (Windows 11) requirements:

> Voice mode could not access an audio device in WSL. WSL2 with WSLg (Windows 11) provides audio via PulseAudio — if you are on Windows 10 or WSL1, run Claude Code in native Windows instead.

Audio probe function (search for "arecord probe failed") and improved WSL message (search for "WSLg")

The auto mode classifier setting field deny has been renamed to soft_deny in the settings schema. This is a semantic clarification — existing deny rules in your config should be updated to soft_deny.

Schema change (search for "soft_deny")

You can now press q in addition to Escape or Ctrl+C to exit the transcript viewer.

Keybinding addition (search for "transcript:exit")

The bridge reconnection logic now handles up to 2 consecutive failed /bridge/reconnect calls before falling through to a fresh session, improving reliability on flaky connections. Empty message arrays from resumed sessions are now also handled gracefully.

Reconnect logic (search for "Next 2 POST /bridge/reconnect") and empty-session guard (search for "z.messages.length === 0")

The terminal renderer now re-asserts mouse tracking and raw mode settings when stdin resumes after a gap, preventing issues where terminal state was lost during long idle periods or after backgrounding the process.

Terminal mode fix (search for "reassertTerminalModes")

The model picker label for Opus 4.6 no longer shows the [NEW] tag, reflecting that it's now the established default model.

String change — old: "Opus 4.6 with 1M context [NEW]" → new: "Opus 4.6 with 1M context"

In remote/CCR environments, OAuth tokens and API keys are now persisted to well-known file paths (/home/claude/.claude/remote/.oauth_token, /home/claude/.claude/remote/.api_key) so that subprocesses can access credentials without inheriting file descriptors.

Remote credential paths (search for "/.oauth_token") and persistence function (search for "Failed to persist")

A new environment variable that lets you inject additional metadata (as a JSON object) into session telemetry. Useful for enterprise environments that need to tag sessions with org-specific identifiers.

Metadata injection (search for "CLAUDE_CODE_EXTRA_METADATA")

The sandbox initialization now detects and scrubs "planted" bare-repo files (directories like HEAD, objects, refs, etc.) that could be mistakenly treated as git repository internals.

Scrubbing function (search for "scrubbed planted bare-repo file")

Code highlighting for diffs, code blocks, and tool use previews has been refactored to use React Suspense with lazy loading. The highlighter module is loaded asynchronously, showing an unhighlighted fallback until ready. This should improve perceived startup time.

Suspense-based highlighting wrapper (search for "Suspense" near highlight components)

Commands that aren't usable in Remote Control mode now return a clear message: /{command} isn't available over Remote Control.

Availability check (search for "isn't available over Remote Control")

• Fixed bridge session resumption failing silently when the resumed session has zero messages — now correctly falls through to a fresh session (search for "z.messages.length === 0")
• Fixed ANTHROPIC_BETAS environment variable not being applied in all request contexts — the beta header filter was incorrectly gated on a condition that could prevent user-specified betas from being sent (search for "ANTHROPIC_BETAS")
• Telemetry additional_metadata is now base64-encoded before transmission, fixing potential encoding issues with JSON payloads (search for "base64")

A background process that automatically reviews recent session transcripts and consolidates insights into your memory files, similar to running a manual memory consolidation pass but triggered automatically.

Feature-flagged via tengu_onyx_plover — disabled by default (enabled: !1).

• When enabled, triggers after a configurable minimum hours (default: 24h) and minimum sessions (default: 5) since last consolidation
• Uses a .consolidate-lock file with PID-based locking to prevent concurrent consolidation
• Runs as a forked background query with querySource: "auto_dream" that reads session transcripts and updates memory files
• Includes throttling to prevent excessive scanning (minimum gap between scans)
• The dream prompt instructs the model to orient by reading existing memories, gather signal from recent sessions, consolidate into topic files, and prune the index
• Telemetry events: tengu_auto_dream_fired, tengu_auto_dream_completed, tengu_auto_dream_failed

Dream infrastructure (search for "autoDream") — gated by tengu_onyx_plover flag (search for "tengu_onyx_plover")