Claude Code 2.1.128: 25+ Reliability Fixes, MCP Diagnostics, and Sub-Agent Cache Savings

Claude Code 2.1.128: Reliability, Diagnostics, and Multi-Agent Performance

Released May 4, 2026, Claude Code 2.1.128 is defined by depth rather than headline features. The release resolves more than 25 distinct issues — touching nearly every layer of the tool, from MCP orchestration and sub-agent coordination to terminal rendering and clipboard handling.

MCP Server Diagnostics

The /mcp command now displays the tool count for each connected server alongside server status. More usefully, it flags servers that connected but exposed zero tools — a condition that previously resulted in a silent, opaque failure where the server appeared connected but contributed nothing to the session. Developers managing complex MCP configurations can now diagnose misconfigured servers at a glance without inspecting logs or restarting with debug flags.

Additionally, reconnecting MCP servers no longer flood the conversation with a full list of tool names on every reconnect cycle. Re-announced tools are now summarized by server prefix, keeping the conversation context clean during transient disconnects.

Sub-Agent Prompt Cache Fix: ~3x Reduction in Cache Creation Overhead

One of the most impactful changes for multi-agent users is a fix to sub-agent progress summaries. Previously, those summaries were generated without referencing the prompt cache, causing cache_creation token counts to spike on every summary cycle. The fix ensures summaries correctly utilize cached prompt context, delivering approximately a 3x reduction in cache_creation overhead. For long-running workflows with many active sub-agents, this meaningfully reduces API cost.

A companion fix prevents sub-agent summaries from firing repeatedly while a sub-agent's transcript is static — previously, an idle sub-agent could trigger redundant summary generations, creating unnecessary cost on agents waiting for user input or external events.

Notable Bug Fixes

Sessions and context:

• Sessions on 1M-context models with a smaller autocompact window were being falsely blocked with "Prompt is too long" before reaching the actual API limit — now corrected.
• Stale "remote-control is active" status lines from prior sessions no longer appear after --resume or --continue.
• /rename without arguments no longer fails on resumed sessions whose last entry is a compact boundary.

Terminal and rendering:

• The terminal progress indicator (OSC 9;4) no longer flickers off between tool calls — it stays visible across the full turn.
• Long URLs are now individually clickable on every wrapped row in fullscreen mode.
• Focus mode no longer briefly dims the previous response when submitting a new prompt.
• A stray 4;0; desktop notification that appeared on every /exit in Kitty and compatible terminals has been removed.

MCP and tool calls:

• MCP tool results no longer drop images when the server returns both structured content and content blocks.
• MCP stdio servers no longer receive corrupted arguments when CLAUDE_CODE_SHELL_PREFIX is set and an argument contains spaces or shell metacharacters.
• Parallel shell tool calls: a failing read-only command (grep, git diff, ls) no longer cancels sibling calls — improving the reliability of any workflow running multiple shell commands simultaneously.

Other notable fixes:

• A crash loop triggered by piping very large input (over 10 MB) to claude -p via stdin is resolved.
• Vim mode: Space in NORMAL mode now moves the cursor right, matching standard vi/vim behavior.
• Drag-and-drop image uploads no longer hang on "Pasting text..." when the image read fails.
• Fenced code blocks inside list items no longer carry leading whitespace into the clipboard on copy-paste.
• /plugin update now correctly detects new versions of npm-sourced plugins.
• Stale installed_plugins.json entries pointing at deleted cache directories no longer pollute PATH.