Claude Code 2.1.72: Effort Levels, Voice Fixes, Bundle Shrink

TL;DR

• Effort levels simplified to low/medium/high with new visual symbols (○ ◐ ●)
• Voice mode fixes: transcription accuracy improved, input lag and false "no speech" errors gone
• Bundle size cut by ~510 KB; prompt cache fix reduces input tokens up to 12x
• SSH-friendly `/copy w` writes directly to file, bypassing clipboard
• Worktree isolation, permission rules, and hooks issues resolved

New

• `/copy w` flag — Write focused selection directly to file, bypassing clipboard (useful over SSH)
• `/plan` description argument — Pass a description like `/plan fix the auth bug` to enter plan mode immediately
• `ExitWorktree` tool — Leave an `EnterWorktree` session cleanly
• `CLAUDE_CODE_DISABLE_CRON` env var — Stop scheduled cron jobs mid-session
• Bash auto-approval allowlist expanded — Added `lsof`, `pgrep`, `tput`, `ss`, `fd`, `fdfind` for read-only operations without permission prompts
• Agent `model` parameter restored — Per-invocation model overrides now work again
• Effort levels redesigned — Simplified to low/medium/high (removed max) with symbols ○ ◐ ● and brief notification instead of persistent icon; use `/effort auto` to reset
• `/config` UX improvements — Escape cancels, Enter saves and closes, Space toggles settings
• VSCode URI handler — `vscode://anthropic.claude-code/open` opens a new Claude Code tab with optional `prompt` and `session` parameters

Fixed

• Tool search now activates with `ANTHROPIC_BASE_URL` when `ENABLE_TOOL_SEARCH` is set
• Voice transcription accuracy improved for repo names and dev terms (regex, OAuth, JSON)
• Bash command parsing switched to native module — faster initialization, no memory leak
• Slow exits when background tasks or hooks were unresponsive
• Agent task progress stuck on "Initializing…"
• Skill hooks firing twice per event
• Voice mode: input lag, false "No speech detected" errors, stale transcripts re-filling prompt
• `--continue` not resuming from most recent point after `--compact`
• Bash security parsing edge cases
• Marketplace git URLs without `.git` suffix now supported (Azure DevOps, AWS CodeCommit)
• Marketplace clone failure messages now show diagnostic info
• Plugin issues: Windows `EEXIST` errors in OneDrive, marketplace blocking user-scope installs, `CLAUDE_CODE_PLUGIN_CACHE_DIR` creating literal `~` directories, `plugin.json` with marketplace-only fields failing to load
• Feedback survey appearing too frequently in long sessions
• `--effort` CLI flag being reset by unrelated settings writes
• Backgrounded Ctrl+B queries losing transcript or corrupting conversation after `/clear`
• `/clear` killing background agent/bash tasks — only foreground tasks cleared now
• Worktree isolation: Task tool resume not restoring cwd, background task notifications missing `worktreePath` and `worktreeBranch`
• `/model` not displaying results when Claude is working
• Digit keys selecting menu options instead of typing in plan mode permission prompt
• Sandbox permission issues: certain file writes incorrectly allowed without prompting, output redirections to allowlisted directories prompting unnecessarily
• Prompt cache invalidation in SDK `query()` calls — reduces input token costs up to 12x
• Escape key becoming unresponsive after cancelling a query
• Double Ctrl+C not exiting with background agents or tasks running
• Team agents not inheriting leader's model
• "Always Allow" saving permission rules that never match again
• Hooks issues: `transcript_path` pointing to wrong directory for resumed/forked sessions, agent `prompt` silently deleted from settings.json, PostToolUse block reason displaying twice, async hooks not receiving stdin with bash `read -r`, validation error example failing validation
• Session crashes in Desktop/SDK with Read returning files containing U+2028/U+2029 characters
• Terminal title cleared on exit despite `CLAUDE_CODE_DISABLE_TERMINAL_TITLE` being set
• Permission rule matching: wildcard rules not matching commands with heredocs/embedded newlines/no arguments, `sandbox.excludedCommands` failing with env var prefixes, "always allow" suggesting overly broad prefixes, deny rules not applying to all command forms
• Oversized and truncated images from Bash data-URL output
• Crash when resuming sessions with Bedrock API errors
• Intermittent "expected boolean, received string" validation errors on Edit, Bash, Grep tool inputs
• Multi-line session titles when forking from conversations with newlines in first message
• Queued messages not showing attached images, images lost when pressing ↑ to edit
• Parallel tool calls where failed Read/WebFetch/Glob cancelled siblings — only Bash errors cascade now
• VSCode: Scroll speed in integrated terminals not matching native terminals
• VSCode: Shift+Enter submitting input instead of inserting newline for users with older keybindings

Improvements

• Up-arrow history shows current session's messages first when running multiple concurrent sessions
• Bundle size reduced by ~510 KB
• CPU utilization improved in long sessions
• CLAUDE.md HTML comments hidden from Claude when auto-injected (visible when read with Read tool)
• VSCode: Effort level indicator added to input border

Update with: npm install -g @anthropic-ai/claude-code@latest or check your IDE's extension marketplace.

Source: Claude Code