Claude-Code v2.1.91 adds MCP result persistence, shell execution controls, and multi-line deep links
Claude-Code v2.1.91 Release Notes
Claude-Code v2.1.91 includes several new features, bug fixes, and performance improvements for developers using AI coding agents. The release focuses on enhancing tool integration, security controls, and user experience.
New Features
- MCP tool result persistence override: Added support for
_meta["anthropic/maxResultSizeChars"]annotation allowing results up to 500K characters to pass through without truncation. This enables larger data like database schemas to persist in tool results. - Shell execution control: Added
disableSkillShellExecutionsetting to disable inline shell execution in skills, custom slash commands, and plugin commands. - Multi-line deep links: Added support for multi-line prompts in
claude-cli://open?q=deep links. Encoded newlines (%0A) are no longer rejected. - Plugin executables: Plugins can now ship executables under
bin/and invoke them as bare commands from the Bash tool.
Bug Fixes
- Fixed transcript chain breaks on
--resumethat could lose conversation history when async transcript writes fail silently - Fixed
cmd+deletenot deleting to start of line on iTerm2, kitty, WezTerm, Ghostty, and Windows Terminal - Fixed plan mode in remote sessions losing track of the plan file after container restart, which caused permission prompts on plan edits and empty plan-approval modals
- Fixed JSON schema validation for
permissions.defaultMode: "auto"in settings.json - Fixed Windows version cleanup not protecting the active version's rollback copy
/feedbacknow explains why it's unavailable instead of disappearing from the slash menu
Improvements
- Improved
/claude-apiskill guidance for agent design patterns including tool surface decisions, context management, and caching strategy - Improved performance: faster
stripAnsion Bun by routing throughBun.stripANSI - Edit tool now uses shorter
old_stringanchors, reducing output tokens
📖 Read the full source: GitHub Claude-Code
👀 See Also
Open Source vs Frontier Models: Single-File Canvas Car Scene Benchmark
A developer tested 12 models including GPT-5.5, Claude Opus 4.7, and Qwen 3.6 Plus on a single-file HTML canvas car driving animation task, with results publicly compared.
Claude Opus 4.7 System Prompt Changes: Platform Renaming, Tool Integration, and Behavioral Updates
Anthropic updated the Claude Opus system prompt from version 4.6 (February 5, 2026) to 4.7 (April 16, 2026), renaming the 'developer platform' to 'Claude Platform', adding Claude in Powerpoint to the tools list, expanding child safety instructions, and implementing new behavioral guidelines for tool usage and response conciseness.
Decoupled DiLoCo: Resilient Distributed Training Across Data Centers with Low Bandwidth
Google DeepMind's Decoupled DiLoCo trains LLMs across distant data centers using 2-5 Gbps WAN, with self-healing islands of compute that isolate hardware failures without degrading ML performance.
OpenClaw API Costs Hit $275 in 5.5 Hours, Annualizing to Over $200K
A developer testing OpenClaw with OpenAI's GPT-5.4 API spent $275 between 11am and 4:30pm, which annualizes to over $200,000 per year at that usage rate.