Version 2.1.70

Bash tool results can now send a compressed version of output to the model while the UI still displays the full uncompressed stdout, reducing token usage on large command outputs.

• A new tokenSaverOutput field is added to the Bash tool result schema
• When token-saver is active, the model receives a compressed representation while users see full output
• Helps conserve context window on commands with verbose output

New schema field with description "Compressed output sent to model when token-saver is active (UI still uses stdout)" (search for "tokenSaverOutput")

Claude Code now reads NODE_EXTRA_CA_CERTS from your config and user settings files, applying them automatically at startup — no need to set environment variables manually.

• On startup, if NODE_EXTRA_CA_CERTS is not already set in the environment, Claude Code checks globalConfig.env and userSettings.env for the path
• Supports --use-system-ca flag to load system CA certificates via the Node.js tls.getCACertificates API
• Custom CA certificates are now also applied to WebSocket connections (both Bun and Node.js)
• Cache clearing added for CA certificates, mTLS configuration, and proxy agents

CA certificate loader (search for "CA certs: Applied NODE_EXTRA_CA_CERTS from config to process.env")

A visual indicator in the input box border now shows the current reasoning effort level (low, medium, high, or max) using colored block characters.

• Displays 1–3 filled ▪ blocks based on effort level (1 for low, 2 for medium, 3 for high/max)
• Shows alongside model name and /fast mode indicator when present
• Controlled by existing effort level settings

Effort-to-dots mapping function (search for "▪") and effort indicator builder $Qq/Cgz at line ~258963

The /color command now accepts default as a valid argument to reset session color.

• Argument hint changed from <color> to <color|default>
• Resetting displays "Session color reset to default"
• Color values "default", "reset", "none", "gray", and "grey" are treated as reset values

Updated argument hint (search for "<color|default>") and reset color list (search for "Session color reset to default")

Remote session reconnection now supports git worktrees by writing and discovering bridge-pointer.json files across worktree directories.

• A bridge-pointer.json file stores sessionId, environmentId, and source for reconnection
• When reconnecting, Claude Code searches across git worktrees to find an active session pointer
• Pointers older than 4 hours are automatically cleared
• Fanout search is capped to prevent performance issues with many worktrees

Bridge pointer system (search for "bridge-pointer.json") and worktree fanout (search for "[bridge:pointer] fanout found pointer in worktree")

Session transcripts now have companion .meta.json files that store additional metadata about each session.

• Metadata files are stored alongside session .jsonl transcripts
• Written and read via dedicated helpers
• Errors on inaccessible files are handled gracefully (ENOENT, EACCES, EPERM)

Metadata file handler (search for ".meta.json")

A new CCR v2 worker registration flow enables Claude Code to register workers with session servers via a /worker/register endpoint.

• Workers register and receive a worker_epoch from the server
• Epoch mismatch (409) handling is now callback-based instead of hard process.exit(1)
• Enables reconnection flows and transport handoff in CCR v2 sessions
• Worker registration is also available for standalone session creation (not just environment-based)

Worker registration function (search for "/worker/register") and epoch handling (search for "epoch superseded")

Background agent management now uses softer language throughout the UI:

• "Press ctrl+f again to kill background agents" → "Press ctrl+f again to stop background agents"
• "Background agents were killed by the user" → "Background agents were stopped by the user"
• Agent status label changed from "Killed" to "Stopped"
• Command keywords: kill agents → stop agents, kill all agents → stop all agents

Terminology change (search for "stop agents") replacing old (search for "kill agents")

Clipboard copy failure messages now suggest both PowerShell and clip as alternatives:

• Windows: "Make sure powershell or clip is available on your system"
• WSL: "Make sure powershell.exe or clip.exe is available in your WSL environment"

Updated error messages (search for "powershell.exe")

The Task tool's run_in_background description was simplified from "The tool result will include an output_file path — use Read tool or Bash tail to check on output" to "You will be notified when it completes."

Updated background description (search for "You will be notified when it completes")

The Task tool's system prompt now documents that agents can be resumed using the resume parameter, that agent outputs should generally be trusted, and includes guidance on distinguishing research vs. code-writing tasks.

New prompt text (search for "Agents can be resumed using the \resume\ parameter")

Voice recording now distinguishes between "no audio detected from microphone" (hardware/permission issue) and "no speech detected" (audio received but no speech recognized), providing more actionable error messages.

Improved error message (search for "No audio detected from microphone")

MCP server instructions are now delivered incrementally as a new mcp_instructions_delta attachment type, sending only added/removed server instructions instead of the full set each turn. Falls back to the previous full-set approach when the tengu_basalt_3kr feature flag is disabled.

Delta attachment handler (search for "mcp_instructions_delta")

The full IANA MIME type database (~4000+ entries) has been replaced with stubs that throw descriptive errors, significantly reducing bundle size. The stubs warn against relying on axios auto-multipart serialization.

Stub implementation (search for "mime-types." and "is stubbed in this build")

Both async and sync mkdir with recursive: true now catch EEXIST errors, fixing race conditions where parallel operations could fail when two processes try to create the same directory simultaneously.

EEXIST handling in mkdir (search for "EEXIST" in the filesystem module at line ~3495)

When running in CLAUDE_CODE_REMOTE, only acceptEdits, plan, and default are allowed as defaultMode values. Other modes are silently ignored with a warning logged.

Remote mode restriction (search for "is not supported in CLAUDE_CODE_REMOTE")

Optimistic ToolSearch is now properly disabled when ANTHROPIC_BASE_URL points to a non-first-party Anthropic host, preventing unnecessary tool search requests that would fail.

Host check (search for "[ToolSearch:optimistic] disabled: ANTHROPIC_BASE_URL=")

After marketplace plugins are installed in the background, Claude Code now auto-refreshes the plugin list immediately instead of just setting a needsRefresh flag, reducing the need for manual refresh.

Auto-refresh logic (search for "Auto-refreshing plugins after")

• CLAUDE_CODE_ALWAYS_ENABLE_EFFORT: Force-enables the effort/reasoning level selector for all models
• CLAUDE_CODE_EFFORT_LEVEL: Override the effort level (set to "unset" to disable)
• Third-party models no longer unconditionally get effort level support; it's now checked against first-party host detection

New env vars (search for "CLAUDE_CODE_ALWAYS_ENABLE_EFFORT" and "CLAUDE_CODE_EFFORT_LEVEL")

Terminal rendering now tracks whether the previous frame was "contaminated" (had selection overlays or image clears), preventing unnecessary full-screen redraws and reducing flickering in alt-screen mode.

New prevFrameContaminated property (search for "prevFrameContaminated")

The text input editor now strips trailing \r from single-line inputs more precisely and handles multi-line paste with trailing carriage returns, which improves behavior when pasting from Windows clipboard sources.

CR handling in input editor (structural change in function Uy1 at line ~452368)

The git file watcher now sets this.initialized = true after (instead of before) calling gitDir(), preventing a race where the watcher could report as initialized before knowing the git directory.

Initialization reorder in ZFA class (search for "this.initialized" near "this.gitDir")

• Fixed clipboard copy on Windows/WSL to suggest PowerShell as an alternative when clip/clip.exe is unavailable (search for "powershell")
• Fixed AskUserQuestion tool input schema being re-parsed on every render by wrapping in useMemo (search for "inputSchema.safeParse")
• Fixed remote hook (useRemoteMode) return value not being memoized, causing unnecessary re-renders (search for "useMemo" in vQq)
• Fixed destructive command warning and sandboxing check not being memoized in permission prompts (search for "destructiveWarning")
• Fixed permission animation timer starting before component is visible by using null initial delay when K is true (search for "lc6")
• Fixed CLAUDE_CODE_EXTRA_BODY env var to shallow-copy parsed object, preventing mutations to the cached parse result (search for "CLAUDE_CODE_EXTRA_BODY")
• Fixed WebFetch HTML-to-Markdown converter (Turndown) being loaded eagerly; it's now lazy-loaded via dynamic import (search for "turndown")
• Fixed Bash command argument parser to use case-insensitive comparison for command names like python/Python3 (search for "python3")
• Fixed early return from onSubmit when bridge reconnection is active, preventing accidental message submission (search for "[onSubmit] early return")
• Fixed deferred context attachments (tips, changed files) leaking empty arrays when no items are found by adding early-return checks (search for "if (K.length === 0) return []")

A new "fork" agent type for the Task tool that creates a copy of the current conversation context to process a task in the background.

Feature-flagged — only triggers when the fork experiment is active.

• When the fork experiment is active, omitting subagent_type from the Task tool creates a fork instead of using the general-purpose agent
• Forks inherit the full conversation context from the parent
• Fork workers cannot create nested forks ("Fork is not available inside a forked worker")
• Uses permissionMode: "bubble" to propagate permission requests to the parent
• System prompt indicates: "You are a forked worker process" with worktree context awareness
• Maximum 200 turns per fork

Fork agent definition (search for "You are a forked worker process") and fork config (search for "Implicit fork")

The BTW side-question feature, previously hardcoded as disabled (w1(void 0) which was always false), is now properly gated behind the tengu_marble_whisper feature flag.

Feature-flagged via tengu_marble_whisper (default false)

• All three references to the BTW feature (tip display, input parser, spinner tip) now check gz6() which reads tengu_marble_whisper
• This means the feature can be enabled server-side without a client update

Feature gate function (search for "tengu_marble_whisper")

• The coral_reef_sonnet client data flag was added, checked via YgA() — this may gate future model-related behavior (search for "coral_reef_sonnet")
• A kairosActive flag was added to session telemetry, tracked as is_assistant_mode in analytics (search for "kairosActive")
• The telemetry event tengu_session_quality_classification was replaced by tengu_session_file_read (search for "tengu_session_file_read")
• The "server" subcommand infrastructure (including WebSocket server, session management, auth tokens, direct-connect protocol, and cc+unix:// URL scheme) has been removed, suggesting the standalone server mode has been deprecated in favor of the CCR bridge architecture