Claude Code v2.1.70

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

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"

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."

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.

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.

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.

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.

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.

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.

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.

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.

• 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

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.

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.

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.

• 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 []")

• 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