Version 2.1.139

claude plugin details <name> shows what a plugin contributes and estimates how much context its components add.

claude plugin details my-plugin
claude plugin details my-plugin --json

• Lists component inventory for skills, commands, agents, hooks, and MCP servers.
• Shows estimated always-on token cost added to each session.
• Breaks down per-component always-on and on-invoke costs when available.
• Supports JSON output for scripting.

Plugin details subcommand and report output (search for "Show a plugin's component inventory and projected token cost", "Component inventory", "Projected token cost")

A new built-in Proactive output style makes Claude act more autonomously: execute immediately, minimize interruptions, and prefer action over planning.

claude
# Open /config and set Output style to Proactive

• Adds Proactive alongside existing built-in output styles.
• Uses a dedicated style prompt, not just a generic reminder.
• Adds a shorter per-turn reminder: Execute autonomously, minimize interruptions, prefer action over planning.
• This is a fully built-in style; no plugin is required.

Built-in output style definition (search for "Proactive Style Active", "Claude executes immediately, minimizes interruptions, and prefers action over planning")

/terminal-setup now includes an interactive scroll-speed control for tuning mouse wheel behavior in the terminal.

claude
# Run /terminal-setup, then use left/right to adjust scroll speed

• Shows current terminal/editor scroll behavior.
• Lets users adjust with left/right arrows, reset with r, save with Enter, or cancel with Esc.
• Saves the selected value to user settings as CLAUDE_CODE_SCROLL_SPEED.
• Detects terminal context such as VS Code, Cursor, Windows Terminal, xterm.js, and high-rate wheel events.

Interactive scroll-speed UI (search for "Scroll speed", "Scroll to feel it · ←/→ adjust · r reset to auto · Enter save · Esc cancel", "CLAUDE_CODE_SCROLL_SPEED")

Claude Code now has first-class plumbing for an enterprise Cloud gateway backend.

claude auth login
# Select Cloud gateway when that option is available for your environment

• Adds Cloud gateway as a distinct API provider.
• Stores gateway URL/status in account and diagnostics views.
• Refreshes gateway JWTs through the gateway OAuth token endpoint.
• Pins and verifies TLS fingerprints on restore, warning users to re-login if trust changes.
• Loads managed settings through the gateway and surfaces actionable re-authentication messages when they fail.

Cloud gateway provider and login flow (search for "Cloud gateway", "Gateway URL", "Connected to Cloud gateway.", "[gateway-refresh] refreshed gateway JWT")

/goal is not new in this release, but its UX is clearer. Active goals now show a richer panel with elapsed time, turns, token count, last check, and /goal clear to stop early. Completed goals show Goal achieved and the stop-hook instruction now says the goal auto-clears after success, instead of telling the user to clear it manually.

Goal status panel and updated hook instruction (search for "Goal active", "/goal clear to stop early", "It auto-clears once the condition is met")

Hook command definitions can now use an explicit args array for exec-style invocation. Claude Code validates the common mistake of putting both whitespace in command and also supplying args, and explains how to split executable and arguments correctly.

{
  "command": "node",
  "args": ["script.js"]
}

• command is resolved as an executable when args is present.
• Arguments avoid shell parsing for quotes, $, and backticks.
• Plugin and hook substitutions now include ${CLAUDE_PROJECT_DIR} in more places, including monitor hooks and plugin command paths.

Exec-form hook schema and validation (search for "Argument list for exec form", "Exec form treats \"command\" as a single executable name", "${CLAUDE_PROJECT_DIR}")

Remote Control now tells users why it is unavailable in more cases, especially when the session is using API-key auth or a non-first-party backend.

• Distinguishes ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, apiKeyHelper, and ANTHROPIC_UNIX_SOCKET cases.
• Explains that Remote Control requires claude.ai subscription auth.
• Adds a specific message when the active backend is not api.anthropic.com.

Remote Control auth diagnostics (search for "Remote Control requires claude.ai subscription auth", "ANTHROPIC_API_KEY is set", "Remote Control is only available when using Claude via api.anthropic.com")

Claude.ai MCP connector setup now reports clearer disabled states for third-party providers, API-key precedence, and missing user:mcp_servers scope.

• Clarifies that claude.ai org connectors are disabled when the inference token lacks MCP scope.
• Explicitly says locally configured MCP servers in managed-mcp.json, .claude.json, and .mcp.json are not affected.
• Adds clearer provider-specific messages for third-party backends.

MCP eligibility messages (search for "[claudeai-mcp] Disabled on third-party provider", "inference token lacks user:mcp_servers scope")

The managed setting/environment-variable wording changed from “background-agents fleet” to “agent view.”

• New setting field: disableAgentView.
• New environment variable: CLAUDE_CODE_DISABLE_AGENT_VIEW.
• Old wording referenced CLAUDE_CODE_DISABLE_AGENTS_FLEET; administrators should update managed settings and environment overrides.

Setting rename (search for "Disable agent view", "CLAUDE_CODE_DISABLE_AGENT_VIEW", and compare old "CLAUDE_CODE_DISABLE_AGENTS_FLEET")

Claude Code now reads CLAUDE_CODE_MAX_TURNS as a default max-turn override when max turns are not otherwise supplied.

CLAUDE_CODE_MAX_TURNS=50 claude

• The value must be a positive integer.
• Invalid values fail early with a clear error.

Max-turn environment parsing (search for "CLAUDE_CODE_MAX_TURNS must be a positive integer")

When CLAUDE_CODE_SUPERVISED is set, uncaught exceptions and unhandled rejections now cause the process to exit after logging. This is useful for service managers or wrapper processes that should restart a crashed CLI instead of letting it limp along.

CLAUDE_CODE_SUPERVISED=1 claude

Supervised fatal-error handling (search for "Uncaught exception under CLAUDE_CODE_SUPERVISED — exiting", "Unhandled rejection under CLAUDE_CODE_SUPERVISED — exiting")

• Bash safety analysis now detects ambiguous heredoc terminators, including <<- tab-prefixed delimiters and heredoc body lines that can be mistaken for terminators with shell metacharacters. Evidence: heredoc safety checks (search for "ambiguous heredoc terminator", "Heredoc uses <<- with a tab-prefixed delimiter")

• Plugin ZIP loading now handles archives with a wrapper directory instead of failing or choosing the wrong root. Evidence: inline plugin ZIP handling (search for "Inline plugin zip had wrapper directory; using")

• Missing plugin errors are more actionable and now suggest claude plugin list or --plugin-dir <path>. Evidence: plugin lookup error (search for "not found. Run claude plugin list to see installed plugins")

• Marketplace plugin cache misses now explain that the plugin may have been removed and tell users to disable it via /plugin. Evidence: plugin source cache-miss message (search for "the plugin may have been removed. Disable it via /plugin to clear this warning")

• Settings-file watching now tracks symlink targets so atomic-save edits to the target file are detected. Evidence: symlink target watcher (search for "also watching", "so atomic-save edits to the target are detected")

• Plugin directory copying now handles cyclic symlinks, non-regular symlink targets, and symlinks escaping the containment root more carefully. Evidence: symlink materialization safeguards (search for "copyDir: skipping cyclic symlink target", "copyDir: skipping symlink escaping containment root")

• HTTP/SSE transport now aborts streams that exceed the body limit without an SSE event boundary, preventing unbounded memory growth when a server returns non-protocol data. Evidence: SSE body overflow guard (search for "without an SSE event boundary", "HttpBodyOverflowError")

• Proxy and TLS connection setup now has explicit timeout/failure errors for gateway-style connections. Evidence: connection error strings (search for "TLS connection timed out", "proxy CONNECT failed", "proxy CONNECT timed out")

The main migration item is for administrators or wrapper authors: update disableAgentView / CLAUDE_CODE_DISABLE_AGENT_VIEW if you previously depended on the older background-agents fleet setting name. For normal interactive users, this release is backward-compatible.