Version 2.1.129

You can now load a session-only plugin directly from a remote .zip URL.

claude --plugin-url https://example.com/my-plugin.zip
claude --plugin-url https://example.com/a.zip --plugin-url https://example.com/b.zip

• --plugin-url is repeatable, like --plugin-dir.
• Claude Code fetches the plugin archive, enforces a maximum archive size, extracts it, and loads it only for the current session.
• If re-fetching fails but a cached copy exists, Claude Code can reuse the cached archive.
• /doctor now labels these as Session-only plugins (--plugin-dir / --plugin-url).

New CLI option and help text (search for "Fetch a plugin .zip from a URL for this session only (repeatable: --plugin-url A --plugin-url B)")

Claude Code can now automatically run supported package-manager updates when a newer version is available.

CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE=1 claude

• When enabled, Claude Code can run update commands for supported package managers instead of only telling you what command to run.
• Homebrew uses brew upgrade --cask ....
• Windows winget uses winget.exe upgrade --id Anthropic.ClaudeCode --exact --silent --disable-interactivity.
• The UI reports progress with messages like Updating via ..., success with ✓ Update installed · Restart to apply, and falls back to manual instructions if auto-update fails.
• Manual update guidance now also recognizes mise upgrade claude.

Auto-update gate and commands (search for "CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE", "--disable-interactivity", and "✓ Update installed")

Remote Control now has a first-run enrollment path for organizations that require Trusted Devices.

claude remote-control
# or inside a session:
/remote-control

• If Remote Control preflight finds that the device must be enrolled, Claude Code starts a sign-in flow instead of only reporting that the device is blocked.
• If sign-in is canceled or enrollment does not complete, Claude Code now gives specific next steps.
• Organizations can temporarily disable proactive enrollment with a feature flag, in which case users see an explicit administrator-facing message.

Trusted Device enrollment flow (search for "Sign in to enroll this device for Remote Control." and "Your organization requires Trusted Devices for Remote Control, but enrollment is temporarily disabled")

Claude Code now warns plugin authors about several manifest and packaging mistakes that previously could be confusing.

• Experimental plugin components should be declared under experimental.<component> instead of at the top level.
• Invalid experimental values are ignored with a clearer warning.
• claude.md and claude.local.md at a plugin root now produce guidance because they are not loaded as project context.
• Plugin root context should move into a skill at skills/<name>/SKILL.md.

Plugin validation warnings (search for "is an experimental component; declare it under 'experimental.", "'experimental' must be an object containing component declarations", and "To ship context with your plugin, use a skill (skills/<name>/SKILL.md) instead.")

Claude Code now surfaces when the skill listing sent to Claude is being shortened or truncated to fit the context budget.

• The UI can show whether descriptions exceed the per-entry cap.
• It suggests /skills to disable skills or skillListingBudgetFraction to allocate more context.
• It estimates the extra token cost of opting into a larger skill listing.

Skill budget warning UI (search for "Skill listing will be truncated", "skillListingBudgetFraction", and "Opting in would cost ~")

Claude Code now refuses more unsafe write paths involving symlinks unless the caller explicitly allows symlink writes.

• Writes through symlinked files can now fail fast with guidance to resolve the symlink and pass the real path.
• Writes into symlinked parent directories are rejected.
• The fallback non-atomic write path also checks O_NOFOLLOW.

Symlink write protections (search for "Refusing to write through symlink:" and "Refusing to write into symlinked directory:")

Remote Control and third-party-provider messages are more specific.

• Remote session creation failures now surface as Remote session create failed: ....
• Remote Control diagnostics include clearer hints when OAuth token or organization UUID state is missing in a CCR environment.
• /feedback now explicitly reports that it is unavailable for Claude Platform on AWS and Amazon Bedrock Mantle.
• Channel support now reports Channels are not available on third-party providers instead of listing only specific providers.

Provider and Remote Control messages (search for "Remote session create failed:", "/feedback is not available when using Claude Platform on AWS", and "Channels are not available on third-party providers")

Fast Mode descriptions no longer hardcode Opus 4.6 in all paths. The displayed Fast Mode model can now be generated dynamically.

Dynamic Fast Mode copy (search for "Fast mode for Claude Code uses Claude" and "CLAUDE_CODE_ENABLE_OPUS_4_7_FAST_MODE")

• Daemon startup failures now fall back to transient spawn with a repair command instead of leaving the user at a dead daemon state. Evidence: search for "daemon service did not become reachable within 5s" and "falling back to transient spawn. Run 'claude daemon install' to repair."

• Renaming a non-responsive remote/background session now shows a specific error instead of silently reverting or using only the generic local state-file failure. Evidence: search for "Couldn't rename — that session isn't responding"

• Background startup can now mark a session as blocked when it appears stuck on a startup dialog, with the user-facing action open this session to continue setup. Evidence: search for "stuck on a startup dialog" and "CLAUDE_BG_STARTUP_WEDGE_MS"

• Background worktree cleanup now distinguishes a worktree kept because it is on a different branch from one kept because it has uncommitted changes. Evidence: search for "Worktree is on a different branch — kept at"

Features with infrastructure added but not yet broadly enabled. These are shipped behind environment variables or feature flags and may become available in future versions.

Claude Code has infrastructure to switch Fast Mode labeling and model selection from Opus 4.6 to Opus 4.7.

Environment-gated

• The new gate is CLAUDE_CODE_ENABLE_OPUS_4_7_FAST_MODE.
• When enabled, Fast Mode text can report Opus 4.7; otherwise it remains on Opus 4.6.
• Because this is gated by an environment variable and existing docs elsewhere still mention Opus 4.6 Fast behavior, treat this as experimental rather than broadly available.

Environment-gated model switch (search for "CLAUDE_CODE_ENABLE_OPUS_4_7_FAST_MODE" and "Opus 4.7")

Claude Code includes new autonomous loop behavior that keeps timer-invoked work alive more persistently when there is still useful work to pursue.

Feature-flagged

• The behavior is enabled by CLAUDE_CODE_LOOP_PERSISTENT or the tengu_kairos_loop_persistent feature flag.
• The new instructions tell the loop to broaden once before stopping when work appears quiet.
• This affects background/autonomous operation rather than normal interactive CLI usage.

Persistent loop gate (search for "CLAUDE_CODE_LOOP_PERSISTENT" and "tengu_kairos_loop_persistent")