Version 2.1.108

A new built-in tool that lets Claude execute JavaScript code in a persistent VM context with programmatic access to Claude Code tools (Read, Write, Edit, Glob, Grep, Bash).

Usage: Claude can now run JavaScript directly to perform file operations, chain tool calls, and compute results. Variables persist across multiple REPL invocations within the same conversation.

• Supports top-level await for async operations
• State persists across calls — variables defined in one REPL call remain accessible in subsequent calls
• Provides registerTool(), unregisterTool(), listTools(), and getTool() for dynamic tool management
• Includes a haiku(prompt, schema?) function for lightweight model sampling
• Includes shQuote(s) for safe shell string quoting
• import/require are not available (VM context is sealed); use the built-in tool functions for file and shell operations
• REPL VM state is cleared on /compact — a notification informs you: "Your REPL VM state has been cleared as part of this compaction"
• Includes a replay mechanism for validating REPL execution determinism across compaction boundaries

REPL tool registration with "Execute JavaScript code with access to Claude Code tools" description (search for "Execute JavaScript")

Infrastructure for replaying REPL code blocks after compaction to validate that execution produces consistent results.

• Extracts code blocks and tool calls from conversation history for replay
• Caches tool call results to replay against
• Detects nondeterminism: "likely nondeterminism (Date.now, Math.random) took a different branch"
• Throws ReplayCacheExhausted when replayed code makes more tool calls than the original
• Reports replay status: "N blocks replayed cleanly"

Replay infrastructure (search for "REPL replay" and "ReplayCacheExhausted")

A new settings panel that analyzes your local sessions and shows what's contributing to your rate limit usage, with actionable recommendations.

• Available to Pro and Max subscribers
• Shows approximate usage breakdown based on local sessions (does not include other devices or claude.ai)
• Tracks five cost behavior categories:
• Cache misses: "X% of your usage hit a >100k-token cache miss" — recommendation: /compact before stepping away
• Long context: "X% of your usage was at >150k context" — recommendation: /compact mid-task, /clear when switching
• Subagent-heavy sessions: "X% of your usage came from subagent-heavy sessions" — recommendation: configure cheaper models for simpler subagents
• High parallelism: "X% of your usage was while 4+ sessions ran in parallel" — recommendation: queue sessions for more even usage
• Long-running sessions: "X% of your usage came from sessions active for 8+ hours" — recommendation: ensure continuous usage is intentional
• Toggle between 24-hour and 7-day views with d and w keyboard shortcuts
• Notes: "these are independent characteristics of your usage, not a breakdown"
• Warns about oversized session files (>200MB) that would skew analysis

Feature gated by tengu_birch_compass flag (search for "tengu_birch_compass" and "What's contributing to your limits usage")

A new confirmation dialog warns you when switching models mid-conversation that the cache will be invalidated.

• Dialog title: "Switch model?"
• Message: "This conversation is cached for the current model. Switching to [model] means the full history gets re-read on your next message."
• Subtitle: "Your next response will be slower and use more tokens"
• Options: "Yes, switch to [model]" or "No, go back"
• Only shown when the conversation has cached content for the current model

Model switch confirmation (search for "Switch model?" and "This conversation is cached")

When you mistype a slash command or skill name, Claude Code now suggests the closest match.

• For slash commands: Unknown command: /stauts. Did you mean /status?
• For skills: Unknown skill: comit. Did you mean commit?
• Uses fuzzy string matching to find the best candidate from available commands and skills
• Tracks suggestion availability for analytics

Fuzzy matcher for commands and skills (search for "Did you mean /" and "Did you mean ")

When the model tries to invoke a built-in CLI command as a skill, it now receives a clear error explaining the distinction.

• Error message: "[name] is a [type] command, not a skill. Ask the user to run /[name] themselves — it cannot be invoked via the [tool] tool."
• Prevents confusion between built-in slash commands and skill invocations

Disambiguation error message (search for "command, not a skill")

Failed remote MCP servers now automatically reconnect using exponential backoff instead of staying disconnected.

• Logs: [MCP] Retry: N failed remote server(s) after Xms backoff
• Stops retrying when all servers reconnect: [MCP] Retry: all remote servers connected, stopping
• Reports persistent failures: remote server(s) still failed after all retries:

MCP reconnection logic (search for "[MCP] Retry" and "failed remote server")

The wildcard syntax for permission rules has changed from colon-asterisk to space-asterisk. For example:

• Old: Bash(git:), Bash(npm run:)
• New: Bash(git ), Bash(npm run )

This affects settings files, autocomplete suggestions, onboarding examples, and documentation. The old : syntax continues to be recognized alongside the new syntax for backward compatibility.

Permission rule syntax change (search for "Bash(git )" and "Bash(npm )")

The /clear command description has been updated from "Clear conversation history and free up context" to "Start fresh: discard the current conversation and context" for clearer communication.

Clear command description (search for "Start fresh: discard the current conversation and context")

The /rewind command (also aliased as /checkpoint) now also responds to /undo, making it more discoverable.

Undo alias addition (search for "undo" near "checkpoint")

When Claude responds in non-English languages, it now receives explicit instructions to maintain full orthographic correctness, including all required diacritical marks, accents, and special characters. For example, it will no longer substitute "nao" for "não" or "fur" for "für".

Orthographic instruction (search for "orthographic" and "diacritical marks")

Rate limit and server error messages now include dynamic status information and direct users to the updated status page at status.claude.com (changed from status.anthropic.com).

• 429 errors now show: "Request rejected (429)" with "Server is temporarily limiting requests (not your usage limit)"
• Status URL changed to status.claude.com
• Error guidance suggests: "this may be a temporary capacity issue — check status.claude.com"

Status URL change (search for "status.claude.com")

The guidance for sleep commands and blocking behavior has been substantially rewritten to emphasize Monitor as the preferred alternative.

• Old: sleep N as first command with N ≥ 2 is blocked with brief explanation
• New: "Long leading sleep commands are blocked. To poll until a condition is met, use Monitor with an until-loop (e.g. until <check>; do sleep 2; done)"
• Additional guidance: "Do not chain shorter sleeps to work around the block"
• General sleep advice simplified from "keep the duration short (1-5 seconds)" to "keep the duration short"

Blocking guidance improvements (search for "Long leading" and "until-loop")

Multiline paste operations now normalize Windows (CRLF) and classic Mac (CR) line endings to Unix (LF) format, preventing issues with mixed line endings in pasted content.

Paste normalization (search for "\\r\\n|\\r" in paste handler)

File reading now properly handles UTF-8 Byte Order Marks with improved byte count accounting and comprehensive line ending normalization after BOM removal.

BOM detection (search for "65279")

Success checkmarks (✓), warning icons, and error indicators throughout the CLI have been replaced with a consistent component-based icon system (N4 component), improving visual consistency across status messages, plugin marketplace results, GitHub Actions workflow confirmations, and iTerm2 setup output.

45+ instances of the new component replacing text-based icons (search for status: "success" near N4)

Plugin installations now track a resolvedVersion field that records the tag-derived semver version resolved during installation. This improves version verification and deduplication, especially when upstream plugins forget to bump their plugin.json version.

• Description: "Tag-derived semver this install resolved to (when fetched via a version constraint). Used by verifyAndDemote in preference to manifest.version"
• Plugin cache is invalidated when resolved version changes

Plugin version tracking (search for "resolvedVersion")

Both ExitWorktree and EnterWorktree now validate that they are not being called from a subagent with a cwd override (e.g., isolation: "worktree" or explicit cwd). This prevents subagents from mutating the parent session's working directory.

• Error: "ExitWorktree cannot be called from a subagent with a cwd override — it would mutate the parent session's process-wide working directory."
• Directs subagents to use Bash with cd for directory changes

Worktree validation (search for "ExitWorktree cannot be called from a subagent")

Ultrareview has been updated with a clearer cloud launch confirmation flow.

• Now prompts: "Run ultrareview in the cloud?" with estimated time "~10–20 min"
• Requires Claude.ai authentication: "Ultrareview requires a Claude.ai account. Run /login to authenticate."
• Unavailable during essential-traffic-only mode
• Description updated: "Finds and verifies bugs in your branch using a multi-agent review fleet"
• REPL tool execution is blocked during ultrareview with access validation

Ultrareview flow (search for "Run ultrareview in the cloud?" and "ultrareview-launch")

The monolithic settings schema has been refactored into a modular feature gate system. Features like autoMode, deepLink, voice, assistant, and briefView are now individually defined with their own buildGate(), shape(), and permissionsShape() methods, making the settings system more maintainable and enabling conditional feature compilation.

Feature gate definitions (search for "autoMode" near "buildGate")

Plugin installations via the SDK now support progress callbacks, enabling SDK consumers to track installation stages (started, completed, per-marketplace installed/failed).

Progress callback (search for "Headless plugin installation progress")

REPL tool execution now respects PreToolUse hooks, showing a clear "Blocked by PreToolUse hook" message when execution is blocked by a hook.

Hook blocking message (search for "Blocked by PreToolUse hook")

• Fixed paste handling to normalize CRLF and CR line endings to LF in multiline mode, preventing mixed line endings from confusing editors (search for "\\r\\n|\\r/g")
• Fixed attribution snapshot handling to clear the map before storing new snapshots, preventing stale data accumulation (search for "D.clear()")
• Fixed dependency version resolution to check both resolvedVersion and manifest version for more accurate plugin deduplication (search for "resolvedVersion")
• Fixed ANSI escape sequence parsing to correctly handle CSI sequences, OSC sequences, and other terminal control codes (search for "Tq4")
• Fixed a variable reference bug in a slice operation where the wrong variable was being used (search for "O6.slice(1)")
• Improved validation of at-mentions to require content presence alongside title checks, preventing empty mentions from being processed (search for "contentLength")
• Fixed React component state management by correcting dependency arrays and initial state values (search for "lzA")
• Improved skill source filtering logic to correctly include builtin sources instead of excluding them (search for "q.source !== 'builtin'")

While the dashboard infrastructure is fully built, it is gated behind the tengu_birch_compass feature flag with a default of false, meaning it is not yet available to all users.

Feature-flagged with tengu_birch_compass (default: disabled)

• Full UI, cost analysis engine, and 5 behavior categories are implemented
• Requires Pro or Max subscription even when enabled
• Keyboard shortcuts (d/w) for period switching are wired up
• Session scanning and file size validation are complete

Feature gated with default false (search for tengu_birch_compass — b8("tengu_birch_compass", !1))

A new coachMode configuration option has been added to the settings schema and tracked configuration attributes, but its behavioral implementation is not yet visible.

Infrastructure only — configuration plumbing exists but no observable behavioral changes

• Added to tracked configuration attributes alongside existing settings
• No user-facing documentation or tips found for this feature yet

Configuration property (search for "coachMode")

The feature gate system includes an assistant feature with buildGate: () => !1 (always false), indicating planned but disabled functionality.

Disabled — build gate returns false, shape returns empty object

• Included in the feature gate list alongside autoMode, deepLink, voice, and briefView
• Shape returns Md5 (an empty object), indicating the settings schema is not yet defined

Assistant gate disabled (search for assistant near buildGate — returns !1)

The permission rule wildcard syntax change from : to is the most impactful change for existing users. If you have custom permission rules in your settings files using the old syntax (e.g., Bash(git:)), the old format continues to work but new autocomplete suggestions and documentation now use the space-asterisk format (e.g., Bash(git )). Consider updating your rules to the new syntax for consistency.

The removal of the bundled Chalk template library and HTML5 tokenizer (parse5) from the CLI reduces the package size. These were replaced by lighter-weight alternatives or removed where unused.