Claude Code v2.1.121: MCP alwaysLoad, plugin prune, terminal scroll fixes, and memory leak patches

Anthropic shipped Claude Code v2.1.121 with a solid batch of practical improvements for daily use. Here's what's worth knowing.

MCP: alwaysLoad and hook output replacement

MCP server configs now support alwaysLoad. When set to true, all tools from that server skip the usual tool-search deferral and stay available at all times — no more lazy loading delays for frequently used tool sets.

PostToolUse hooks can now replace tool output for any tool via hookSpecificOutput.updatedToolOutput. Previously this was MCP-only; now it works across all tools.

Plugin management: claude plugin prune

Added claude plugin prune to remove orphaned auto-installed plugin dependencies. Running plugin uninstall --prune cascades cleanup. Handy for keeping the plugin directory tidy.

/skills search and dialog scroll improvements

The /skills command now includes a type-to-filter search box — no more scrolling through long lists. Dialogs that overflow the terminal are now scrollable with arrow keys, PgUp/PgDn, home/end, and mouse wheel in both fullscreen and non-fullscreen modes.

Fullscreen mode also fixed a long-standing annoyance: typing into the prompt no longer jumps the scroll back to the bottom if you've scrolled up to read earlier output. Clicking any line of a long URL that wraps across rows in fullscreen mode now opens the full URL.

SDK and non-interactive improvements

CLAUDE_CODE_FORK_SUBAGENT=1 now works in non-interactive sessions (SDK and claude -p). The mcp_authenticate SDK method now supports redirectUri for custom scheme completion and claude.ai connectors.

Permissions, terminal setup, and retries

--dangerously-skip-permissions no longer prompts for writes to .claude/skills/, .claude/agents/, and .claude/commands/. /terminal-setup now enables iTerm2's clipboard-access setting so /copy works, including from tmux. MCP servers that hit transient errors during startup now auto-retry up to 3 times.

Memory leak fixes

Several important memory issues addressed:

• Fixed unbounded memory growth (multi-GB RSS) when processing many images in a session
• Fixed /usage leaking up to ~2GB on machines with large transcript histories
• Fixed memory leak when long-running tools fail to emit a clear progress event
• Fixed Bash tool becoming permanently unusable when the starting directory is deleted or moved mid-session
• Fixed embedded grep/find/rg shell wrappers failing when the binary is deleted mid-session — now falls back to installed tools
• Reduced peak file descriptor usage during find in the Bash tool on large directory trees

Other notable fixes

• --resume crashing on startup in external builds fixed; large sessions with corrupted transcript lines now skip the corrupt line
• Fixed thinking.type.enabled is not supported error with Bedrock application inference profile ARNs
• Fixed Microsoft 365 MCP OAuth failing with duplicate or unsupported prompt parameter
• Fixed scrollback duplication on Ctrl+L in non-fullscreen mode on tmux, GNOME Terminal, Windows Terminal, and Konsole
• Fixed NO_PROXY not respected for all HTTP clients when set via managed-settings.json
• Fixed /usage returning "rate limited" after a stale OAuth token — now auto-refreshes
• Fixed /focus showing "Unknown command" when fullscreen renderer is off — now explains how to enable it

OpenTelemetry and Vertex AI

LLM request spans now include stop_reason, gen_ai.response.finish_reasons, and user_system_prompt (gated behind OTEL_LOG_USER_PROMPTS). Vertex AI adds support for X.509 certificate-based Workload Identity Federation (mTLS ADC).