This comprehensive changelog tracks Claude Code’s rapid evolution since its preview launch in February 2025. Each version brings significant improvements based on user feedback and real-world usage patterns.
Cloud defaults, Auto Mode rollout, and security hardening
Amazon Bedrock, Google Cloud’s Agent Platform, and Claude Platform on AWS now use Claude Opus 4.8 as the runtime default. Microsoft Foundry remains on Sonnet 4.5 unless configured otherwise.
Auto mode no longer requires CLAUDE_CODE_ENABLE_AUTO_MODE on Bedrock, Vertex/Agent Platform, or Foundry. Administrators can disable it with permissions.disableAutoMode; autoMode is ignored in repo-local .claude/settings.local.json.
Fixed non-interactive runs permanently recording remote managed settings as consented without showing the security dialog, and stopped project-level settings from supplying pluginConfigs values.
Shell-form plugin hooks, monitors, and MCP headersHelper commands now reject ${user_config.*} substitution to close a shell-injection path; use exec-form arguments or read values inside the script.
Added an auto mode rule protecting transcript files and fixed benign system updates triggering false prompt-injection warnings, alongside updater, Windows/AWS, worktree, Remote Control, and background-session reliability fixes.
v2.1.203–206 - July 7–10, 2026
Agent reliability, /doctor, MCP auth, and safer automation
/doctor is now a full setup checkup that can diagnose and fix issues; it also recommends trimming checked-in CLAUDE.md content that Claude can derive from the codebase.
Background agents survive upgrades more reliably, inherit dispatch-session environment overrides, recover stale auth, and expose clearer states, headlines, PR links, and exact input requests in claude agents.
Added directory suggestions to /cd, broader push-remote support for /commit-push-pr, MCP per-server request_timeout_ms support, and confirmation before entering worktrees outside .claude/worktrees/.
Hardened auto mode against transcript tampering and unresolved destructive paths; fixed Windows junction-safe worktree removal and numerous background-agent, Remote Control, LSP, OAuth, resume, and rendering failures.
v2.1.199–202 - July 2–6, 2026
Dynamic-workflow sizing, manual permissions, stacked skills, and partial-result recovery
/config now exposes advisory small/medium/large dynamic workflow size; workflow-spawned agents emit workflow.run_id and workflow.name telemetry.
Manual is the default permission mode across CLI and IDEs; AskUserQuestion no longer auto-continues unless an idle timeout is configured.
Up to five leading slash skills can be stacked in one prompt; subagents preserve partial output on rate-limit/server failures instead of reporting false success.
/review <pr> is again a fast single-pass review; use /code-review <level> <pr#> for multi-agent review.
v2.1.198 - July 1, 2026
Background subagents by default and Claude in Chrome GA
Subagents now run in the background by default; the parent keeps working and receives completion/input notifications through the Notification hook.
Background code agents that isolate changes in Claude-created worktrees can commit, push their own branch, and open a draft PR; they never push main/master, force-push, or merge, and other checkout modes still follow normal permission prompts. Explore inherits the main session model (capped at Opus), and subagents inherit extended-thinking configuration.
Added the /dataviz skill, Claude in Chrome reached GA, and the legacy /agents wizard was removed in favor of asking Claude or editing .claude/agents/ directly.
Claude Sonnet 5 + Fable 5 restored - June 30–July 1, 2026
Claude Sonnet 5 became Claude Code’s default model in v2.1.197, with native 1M context and launch pricing of $2 / $10 per MTok through August 31 (then $3 / $15).
Global access to Claude Fable 5 returned July 1; after July 7 it uses usage credits on eligible subscription plans.
Added claude mcp login/logout, Skills in the installed-plugin view, sandbox.credentials, organization model restrictions/defaults, /rewind across /clear, and autoMode.classifyAllShell.
Background subagents surface permission prompts to the main session; long-running agents, commands, and workflows survive process restarts and upgrades, with a five-minute stream watchdog enabled by default.
MCP and OAuth retries, remote tool idle timeouts, model picker policy, worktree cleanup, accessibility, voice, rendering, and Remote Control reliability were substantially improved.
New claude_code.assistant_response OTel logging can include response text; set OTEL_LOG_ASSISTANT_RESPONSES=0 if your deployment must keep prompt-only logging.
Artifacts turns a Claude Code session into an interactive page — a PR walkthrough, a living project dashboard — shared with your team at a private link.
Available in beta on Team and Enterprise plans.
v2.1.172–183 - June 2026
Nested Sub-Agents, Implicit Agent Teams & Auto-Mode Guardrails
Nested sub-agents up to 5 levels deep (v2.1.172) — agents can spawn agents, with the depth capped so fan-out stays under control.
Implicit agent teams (v2.1.178) — the TeamCreate / TeamDelete tools are removed; teams now form by spawning teammates directly. The same release adds Tool(param:value) permission rules that match specific parameters, and closest-directory-wins for nested .claude/ name collisions.
language setting & model allowlists (v2.1.175–176) — session titles generate in your conversation’s language, and enforceAvailableModels constrains the Default model to an approved allowlist. Hook path conditions like Edit(src/**) now match correctly.
/config key=value, Apple Events sandbox & Bun 1.4 (v2.1.181) — set configuration from the prompt, opt into macOS Apple Events with sandbox.allowAppleEvents, and the bundled Bun runtime moves to 1.4.
Auto-mode guardrails (v2.1.183) — auto mode now blocks destructive git commands and terraform / pulumi destroy without explicit approval, and attribution.sessionUrl keeps claude.ai session links out of your commits and PRs.
Dynamic Workflows GA & Scheduled Agents - June 10, 2026
From Code with Claude, Tokyo
Dynamic workflows in Claude Code are now generally available — trigger multi-agent runs with the ultracode keyword (see Dynamic Workflows).
Claude Managed Agents add scheduled deployments and environment variables in vaults (public beta) — agents run on a schedule and use your tools securely.
v2.1.170 - June 9, 2026
Claude Fable 5 — a new model tier above Opus
Claude Fable 5 is now available — “a Mythos-class model that we’ve made safe for general use. Fable’s capabilities exceed those of any model we’ve ever made generally available.” Update to version 2.1.170 for access; switch with /model fable (model ID claude-fable-5).
$10 / $50 per million tokens — exactly 2× Opus 4.8 — with a 1M context window, 128K max output, effort levels up to max (including xhigh), and adaptive thinking only.
Included on Pro, Max, Team, and seat-based Enterprise plans from June 9 through June 22, 2026; from June 23, 2026 further use requires usage credits.
Fixed sessions not saving transcripts (and not appearing in --resume) when launched from the VS Code integrated terminal.
v2.1.169 - June 8, 2026
--safe-mode, /cd & Bundled-Skill Controls
--safe-mode flag + CLAUDE_CODE_SAFE_MODE env var — start with all customizations (CLAUDE.md, plugins, skills, hooks, MCP servers) disabled, for troubleshooting
/cd — move the session to a new working directory without breaking the prompt cache
disableBundledSkills setting + CLAUDE_CODE_DISABLE_BUNDLED_SKILLS env var — hide bundled skills, workflows, and built-in slash commands
Self-hosted runner gains a post-session lifecycle hook — runs after the session ends, before workspace deletion, so you can snapshot uncommitted work or export logs
fallbackModel setting (and --fallback-model now in interactive sessions) — configure up to three fallback models tried in order when the primary is overloaded or unavailable; Claude Code also retries a turn once on the fallback when the API returns an unexpected non-retryable error (auth, rate-limit, request-size, and transport errors still surface immediately)
Glob patterns in deny rules — the tool-name position now accepts globs, so "*" denies all tools; allow rules reject non-MCP globs, and unknown tool names in deny rules warn at startup
Hardened cross-session messaging — messages relayed via SendMessage from another Claude session no longer carry user authority; receivers refuse relayed permission requests and auto mode blocks them
MAX_THINKING_TOKENS=0, --thinking disabled, and the per-model toggle now disable thinking on models that think by default (via the Claude API)
claude update announces the target version before downloading; claude agents filters the list as you type a session URL
v2.1.167–168 followed with reliability fixes (JetBrains terminal flicker on 2026.1+, Kitty-protocol Shift keys, voice mode, managed settings)
/plugin list with --enabled/--disabled filters; requiredMinimumVersion / requiredMaximumVersion managed settings refuse startup outside an approved version range and point users to a sanctioned build
claude agents --json now reports waitingFor (what a blocked session is waiting on); dispatching from the state-grouped view starts in the directory the view was opened from
/effort confirms when your chosen level will persist as the default for new sessions
Clicking a slash command in autocomplete now fills it into the prompt instead of running it immediately — press Enter to run
Remote Control shows as a persistent footer pill; Windsurf renamed to Devin Desktop across /ide, /terminal-setup, and /scroll-speed
/btw gains a “c to copy” shortcut (raw markdown); Stop and SubagentStop hooks can return hookSpecificOutput.additionalContext
v2.1.160–161 - June 2, 2026
ultracode Trigger Keyword & Config-Write Guards
The dynamic-workflow trigger keyword is now ultracode (renamed from workflow) and is highlighted in violet in the prompt input — the bare word “workflow” no longer triggers a run, though asking for one in your own words still does. See Dynamic Workflows.
Config-write guards — a prompt now appears before writing to shell startup files (.zshenv, .zlogin, .bash_login) and ~/.config/git/; acceptEdits mode also prompts before writing execution-granting build configs (.npmrc, .yarnrc*, bunfig.toml, .bazelrc, .pre-commit-config.yaml, .devcontainer/)
Edit no longer needs a separate Read after a single-file grep/egrep/fgrep — the grep satisfies the read-before-edit check
/mcp collapses claude.ai connectors you’ve never signed into; a failed Bash call in a parallel batch no longer cancels the others
OTEL_RESOURCE_ATTRIBUTES values are emitted as metric labels, so you can slice usage by custom dimensions like team or repo
.claude/skills Plugin Auto-Load & Auto Mode on Cloud Providers
Plugins in .claude/skills load automatically — no marketplace required; claude plugin init <name> scaffolds a new one, and /plugin arguments autocomplete
EnterWorktree can switch between Claude-managed worktrees mid-session; the agent field in settings.json is now honored for dispatched claude agents sessions
Auto mode on Amazon Bedrock, Google Vertex, and Microsoft Foundry for Opus 4.7 and Opus 4.8 — at this release it required CLAUDE_CODE_ENABLE_AUTO_MODE=1; v2.1.207 removed that opt-in
Fixed an Opus 4.8 issue where modified thinking blocks led to API errors
v2.1.154 - May 28, 2026
Claude Opus 4.8, Dynamic Workflows & ultracode
Claude Opus 4.8 is now available — Anthropic’s new flagship model, around four times less likely than Opus 4.7 to leave flaws in its own code unflagged. Pricing is unchanged at $5 / $25 per million tokens, and xhigh effort is the new default.
Dynamic workflows (research preview) — ask Claude to create a workflow and it orchestrates work across tens to hundreds of subagents in the background, verifying outputs before reporting back. Watch active runs with the new /workflows command.
ultracode — run /effort ultracode to combine xhigh reasoning with automatic workflow orchestration; Claude decides when a task warrants a workflow. It lasts the session — drop back with /effort high for routine work.
Fast mode on Opus 4.8 is now far cheaper: 2× the standard rate for 2.5× the speed.
Lean system prompt is now the default for every model except Haiku, Sonnet, and Opus 4.7 and earlier.
/simplify redesigned — runs a cleanup-only review (reuse, simplification, efficiency, altitude) and applies the fixes.
Effort slider relabeled from “Speed” / “Intelligence” to “Faster” / “Smarter”.
v2.1.152 - May 27, 2026
/code-review --fix & Skill Tool Controls
/code-review --fix applies the review’s findings to your working tree after the review — surfacing reuse, simplification, and efficiency cleanups.
Skills and slash commands can set disallowed-tools in frontmatter to remove tools from the model for that command.
Added /reload-skills; SessionStart hooks can now return reloadSkills: true.
v2.1.147 - May 21, 2026
/simplify → /code-review with Effort Levels
Renamed /simplify to /code-review and added effort levels (e.g. /code-review high).
--strict-mcp-config improvements for subagent MCP servers.
Streaming tool execution is now enabled by default across all platforms.
v2.1.145 - May 19, 2026
claude agents --json, Plugin Discovery
Added claude agents --json for scripting and automation.
The Agent tool now includes agent_id and parent_agent_id in OpenTelemetry spans.
/plugin Discover shows a plugin’s commands, agents, skills, hooks, and MCP/LSP servers before you install it.
v2.1.144 - May 19, 2026
/resume for Background Sessions
/resume now supports background sessions — sessions started with claude --bg or from the agent view are marked with a bg label.
Background sessions show their elapsed duration on completion (e.g. “Agent completed · 3h 2m 5s”).
Supervisor now reconnects to detached sessions after auto-update without losing in-flight work
Peek panel: paste an image to attach a screenshot to the reply you send to a blocked session
New claude respawn --all flag restarts every stopped session in one shot — pairs well with morning login after laptop sleep
Filter syntax in the dispatch input: s:blocked matches everything waiting on you (needs input + ready for review PRs)
Fixed ← shortcut occasionally backgrounding the wrong session when multiple peek panels were open
Fixed claude rm <id> failing to clean up empty .claude/worktrees/ parents
Fixed agent view ignoring disableAgentView: true set only in managed settings
v2.1.139 - May 11, 2026
/goal, Agent View Research Preview, Background Sessions
/goal sets a completion condition; Claude keeps working across turns until a small fast model confirms it. Aliases stop/off/reset/none/cancel for /goal clear. Up to 4,000-char conditions. Works in -p and Remote Control.
Agent view (claude agents) — research preview of a full-terminal dashboard for every background session: dispatch, peek, attach, detach, all sessions hosted by a per-user supervisor process
New shell commands: claude --bg "<prompt>", claude attach <id>, claude logs <id>, claude stop <id>, claude respawn <id>, claude rm <id>
New in-session commands: /bg / /background (background the current session), /stop (end a session from inside it), /tasks (list in-flight work that would block backgrounding)
← on an empty prompt in any Claude Code session backgrounds it and opens agent view with that row pre-selected (toggleable in /config)
Fixed credential save crash on Linux/Windows corrupting ~/.claude/.credentials.json
Fixed /fork writing full parent conversation per fork (now uses pointer + lazy hydrate)
v2.1.117 - April 22, 2026
Native bfs/ugrep on macOS/Linux, Parallel MCP Startup
Native macOS/Linux builds replace Glob/Grep tools with embedded bfs and ugrep available through the Bash tool — faster searches without the tool round-trip (Windows and npm builds unchanged)
Improved /model: selections persist across restarts even when the project pins a different model; startup header shows when the active model comes from a project/managed-settings pin
Faster startup when both local and claude.ai MCP servers are configured (concurrent connect default)
/resume on stale, large sessions now offers to summarize before re-reading
OpenTelemetry: user_prompt events include command_name/command_source; cost.usage, token.usage, api_request, api_error include effort when supported
Advisor Tool (experimental): dialog now carries an “experimental” label; no more stuck “result content could not be processed” errors
cleanupPeriodDays retention sweep now covers ~/.claude/tasks/, shell-snapshots/, and backups/
Default effort for Pro/Max on Opus 4.6 and Sonnet 4.6 is now high (was medium)
Fixed Opus 4.7 sessions showing inflated /context percentages — Claude Code was computing against 200K instead of native 1M
Fixed Bedrock application-inference-profile requests failing with 400 when backed by Opus 4.7 with thinking disabled
Fixed WebFetch hanging on very large HTML pages by truncating input before HTML→markdown conversion
v2.1.116 - April 20, 2026
Faster /resume, Thinking Spinner Progress
/resume on large sessions is significantly faster (up to 67% on 40MB+ sessions)
Faster MCP startup with multiple stdio servers; resources/templates/list is deferred to first @-mention
Smoother fullscreen scrolling in VS Code / Cursor / Windsurf terminals — /terminal-setup now configures editor scroll sensitivity
/doctor can now be opened while Claude is responding
Security: sandbox auto-allow no longer bypasses dangerous-path check for rm/rmdir on /, $HOME, or critical system directories
Claude Code now uses https://downloads.claude.ai/claude-code-releases
Fixed Devanagari and other Indic scripts rendering with broken column alignment
Fixed Ctrl+- not triggering undo in Kitty keyboard protocol terminals (iTerm2, Ghostty, WezTerm, Windows Terminal)
Fixed /branch rejecting conversations with transcripts larger than 50MB
Fixed intermittent API 400 “cache control TTL ordering” errors
v2.1.114 - April 18, 2026
Permission Dialog Crash Fix
Fixed a crash in the permission dialog when an agent teams teammate requested tool permission
v2.1.113 - April 17, 2026
Native CLI Binary, Sandbox Hardening
CLI now spawns a native Claude Code binary (via per-platform optional dependency) instead of bundled JavaScript
Added sandbox.network.deniedDomains to block specific domains even when a broader allowedDomains wildcard permits them
Fullscreen: Shift+↑/↓ scrolls the viewport when extending a selection past visible edge
Ctrl+A / Ctrl+E now move to start/end of current logical line in multiline input (readline behavior)
Windows: Ctrl+Backspace now deletes the previous word
Improved /loop: Esc cancels pending wakeups
Improved /ultrareview: faster launch, parallelized checks, diffstat in launch dialog
Subagents that stall mid-stream now fail with a clear error after 10 minutes instead of hanging
Security: macOS /private/{etc,var,tmp,home} paths treated as dangerous removal targets under Bash(rm:*)
Security: Bash deny rules now match commands wrapped in env/sudo/watch/ionice/setsid
Security: Bash(find:*) allow rules no longer auto-approve find -exec/-delete
Fixed thinking.type.enabled is not supported 400 error when using Opus 4.7 via Bedrock Application Inference Profile ARN
Fixed Bash dangerouslyDisableSandbox running commands outside the sandbox without a permission prompt
v2.1.112 - April 16, 2026
Opus 4.7 Availability Fix
Fixed “claude-opus-4-7 is temporarily unavailable” for auto mode
v2.1.111 - April 16, 2026
Claude Opus 4.7 + xhigh effort + /ultrareview
Claude Opus 4.7 xhigh effort level now available — tune speed vs. intelligence via /effort, --effort, or the model picker; other models fall back to high
Auto mode now available for Max subscribers when using Opus 4.7
/effort opens an interactive slider when called without arguments
Fixed claude -w <name> failing with “already exists” after stale worktree cleanup
Fixed subagents in isolated worktrees being denied Read/Edit access to files inside their own worktree
Fixed RemoteTrigger tool’s run action sending an empty body
v2.1.100 - April 10, 2026
Paste Improvements, Plugin Dependency Resolution
Faster paste handling and better CRLF handling in long pastes
Plugin dependency auto-install during marketplace add
Performance fixes across SDK subprocess cleanup
v2.1.98 - April 9, 2026
Google Vertex Setup Wizard, Perforce Mode, Monitor Tool
Added interactive Google Vertex AI setup wizard from the login screen — guides through GCP auth, project/region config, credential verification, model pinning
Added CLAUDE_CODE_PERFORCE_MODE env var — Edit/Write/NotebookEdit fail on read-only files with a p4 edit hint instead of silently overwriting
Added Monitor tool for streaming events from background scripts
Added subprocess sandboxing with PID namespace isolation on Linux via CLAUDE_CODE_SUBPROCESS_ENV_SCRUB, plus CLAUDE_CODE_SCRIPT_CAPS to cap per-session script invocations
Added --exclude-dynamic-system-prompt-sections flag to print mode for cross-user cache improvement
Added workspace.git_worktree to the status line JSON when inside a linked git worktree
Added W3C TRACEPARENT env var to Bash tool subprocesses when OTEL tracing is enabled
Fixed a Bash tool permission bypass where backslash-escaped flags could be auto-allowed as read-only leading to arbitrary code execution
Fixed --dangerously-skip-permissions being silently downgraded to accept-edits mode
Fixed managed-settings allow rules remaining active after admin removed them until restart
Improved /agents with tabbed layout: Running tab + Library tab with Run agent / View running instance
v2.1.97 - April 8, 2026
NO_FLICKER Focus View, Cedar Highlighting
Added focus view toggle (Ctrl+O) in NO_FLICKER mode showing prompt, one-line tool summary with edit diffstats, and final response
Added refreshInterval status line setting to re-run the command every N seconds
Added ● N running indicator in /agents next to agent types with live subagent instances
Added syntax highlighting for Cedar policy files (.cedar, .cedarpolicy)
Fixed MCP HTTP/SSE connections accumulating ~50 MB/hr of unreleased buffers on reconnect
Improved image handling: pasted and attached images compressed to same token budget as Read-tool images
Improved sandbox: sandbox.network.allowMachLookup now takes effect on macOS
Improved Bridge sessions to show local git repo, branch, and cwd on claude.ai session card
Fixed CJK/@-mention completion after punctuation; Japanese/Chinese input no longer requires space before / or @
v2.1.96 - April 8, 2026
Bedrock Auth Hotfix
Fixed Bedrock requests failing with 403 "Authorization header is missing" when using AWS_BEARER_TOKEN_BEDROCK or CLAUDE_CODE_SKIP_BEDROCK_AUTH (regression in 2.1.94)
v2.1.94 - April 7, 2026
Mantle-backed Bedrock, Higher Default Effort
Added support for Amazon Bedrock powered by Mantle — set CLAUDE_CODE_USE_MANTLE=1
Changed default effort level from medium to high for API-key, Bedrock/Vertex/Foundry, Team, and Enterprise users
Added compact Slacked #channel header with clickable channel link for Slack MCP send-message calls
Added keep-coding-instructions frontmatter for plugin output styles
Added hookSpecificOutput.sessionTitle to UserPromptSubmit hooks
Plugin skills declared via "skills": ["./"] now use the skill’s frontmatter name
Fixed agents stuck after 429 with long Retry-After
Fixed Console login on macOS silently failing with “Not logged in” when keychain is locked
Improved --resume to resume sessions from other worktrees of the same repo directly
v2.1.92 - April 4, 2026
forceRemoteSettingsRefresh, Bedrock Setup Wizard
Added forceRemoteSettingsRefresh policy setting — CLI blocks startup until remote managed settings are freshly fetched (fail-closed)
Added interactive Bedrock setup wizard from the login screen’s “3rd-party platform” option
Added per-model and cache-hit breakdown to /cost for subscription users
/release-notes is now an interactive version picker
Remote Control session names use your hostname as the default prefix (e.g. myhost-graceful-unicorn)
Removed /tag command
Removed /vim command (toggle vim mode via /config → Editor mode)
Improved Write tool diff computation speed for large files (60% faster on files with tabs/&/$)
Linux sandbox now ships the apply-seccomp helper in both npm and native builds
Fixed subagent spawning permanently failing with “Could not determine pane count” after tmux window churn
v2.1.91 - April 2, 2026
MCP Result Size Override, Plugin bin/
Added MCP tool result persistence override via _meta["anthropic/maxResultSizeChars"] (up to 500K)
Added disableSkillShellExecution setting to disable inline shell execution in skills and slash commands
Plugins can now ship executables under bin/ and invoke them as bare commands from the Bash tool
Added support for multi-line prompts in claude-cli://open?q= deep links (encoded newlines %0A no longer rejected)
Fixed transcript chain breaks on --resume that could lose history when async transcript writes fail silently
Fixed plan mode in remote sessions losing track of the plan file after a container restart
Conditional hooks: Added if field for hooks using permission rule syntax (e.g., Bash(git *)) to filter when they run, reducing process spawning overhead
MCP OAuth RFC 9728: MCP OAuth now follows Protected Resource Metadata discovery to find the authorization server
Added CLAUDE_CODE_MCP_SERVER_NAME and CLAUDE_CODE_MCP_SERVER_URL env vars to MCP headersHelper scripts
Added .jj and .sl to VCS directory exclusion lists so Grep and file autocomplete skip Jujutsu/Sapling metadata
MEMORY.md index now truncates at 25KB as well as 200 lines
Read tool uses compact line-number format and deduplicates unchanged re-reads, reducing token usage
Plugins blocked by organization policy can no longer be installed or enabled
Fixed --resume failing with orphan tool_result blocks on sessions created before v2.1.85
Fixed /compact failing when the conversation has grown too large for the compact request itself
VSCode: Fixed extension defaulting Max plan users to Sonnet after OAuth token refresh
v2.1.83-84 - March 25-26, 2026
Transcript Search, Plugin Options & PowerShell
Transcript search: Press / in transcript mode (Ctrl+O) to search, n/N to step through matches
managed-settings.d/ drop-in directory: Separate teams can deploy independent policy fragments that merge alphabetically
Plugin options: manifest.userConfig allows plugins to prompt for configuration at enable time, with sensitive: true values stored in keychain
PowerShell tool for Windows as an opt-in preview
Added CwdChanged and FileChanged hook events for reactive environment management
Added TaskCreated hook that fires when a task is created via TaskCreate
Added sandbox.failIfUnavailable setting to exit with error when sandbox cannot start
Pasted images now insert [Image #N] chip at cursor for positional reference
Rules and skills paths: frontmatter now accepts a YAML list of globs
MCP tool descriptions and server instructions capped at 2KB to prevent context bloat
Token counts ≥1M now display as “1.5m” instead of “1512.6k”
Added idle-return prompt nudging users after 75+ minutes to /clear
Agents can declare initialPrompt in frontmatter to auto-submit a first turn
VSCode: Added rate limit warning banner with usage percentage and reset time
v2.1.81 - March 22, 2026
Bare Mode & Channels
--bare flag for scripted -p calls — skips hooks, LSP, plugin sync, and skill directory walks; requires ANTHROPIC_API_KEY; auto-memory fully disabled
--channels permission relay — channel servers that declare the permission capability can forward tool approval prompts to your phone
Resuming a session that was in a worktree now switches back to that worktree
MCP read/search tool calls collapse into a single Queried <server> line (expand with Ctrl+O)
Improved plugin freshness — ref-tracked plugins now re-clone on every load
v2.1.37: Fixed an issue where /fast was not immediately available after enabling /extra-usage
v2.1.36: Fast mode is now available for Opus 4.6 — same model, faster output. Toggle with /fast
v2.1.34 - February 17, 2026
Agent Teams & Sandbox Fixes
Fixed a crash when agent teams setting changed between renders
Fixed a bug where commands excluded from sandboxing (via sandbox.excludedCommands or dangerouslyDisableSandbox) could bypass the Bash ask permission rule when autoAllowBashIfSandboxed was enabled
v2.1.33 - February 15, 2026
Agent Hooks & VSCode Remote Sessions
Added TeammateIdle and TaskCompleted hook events for multi-agent workflows
Added memory frontmatter field for agents — persistent memory with user, project, or local scope
Added support for restricting sub-agents via Task(agent_type) syntax in agent “tools” frontmatter
Added plugin name to skill descriptions and /skills menu for better discoverability
VSCode: Added support for remote sessions, allowing OAuth users to browse and resume sessions from claude.ai
VSCode: Added git branch and message count to the session picker, with search by branch name
Fixed multiple issues: thinking interruption, API proxy 404 fallback, proxy env var propagation, /resume XML markup display
v2.1.32 - February 12, 2026
Claude Opus 4.6 & Agent Teams
Claude Opus 4.6 is now available!
Agent Teams research preview — multi-agent collaboration (set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1)
Automatic Memories — Claude now records and recalls memories as it works
Summarize from here — partial conversation summarization via message selector
Skills from .claude/skills/ in additional directories (--add-dir) now loaded automatically
Skill character budget scales with context window (2% of context)