Claude-Code v2.1.63 adds HTTP hooks, slash commands, and fixes memory leaks
Claude-Code v2.1.63 is a maintenance and feature release focused on stability improvements and new integration capabilities. The update addresses several memory leaks in long-running sessions while adding practical features for developers using AI coding assistants.
New Features
Added bundled slash commands: /simplify and /batch for streamlined workflows. The /model command now shows the currently active model directly in the slash command menu.
HTTP hooks enable external system integration by allowing JSON POST requests to URLs instead of shell commands. This provides a cleaner API-style interface for external tooling.
Project configurations and auto memory are now shared across git worktrees of the same repository, improving consistency in multi-branch development environments.
Memory and Performance Fixes
- Fixed listener leaks in bridge polling loop, MCP OAuth flow cleanup, and WebSocket transport reconnection
- Fixed memory leaks in git root detection cache, JSON parsing cache, bash command prefix cache, and MCP tool/resource cache
- Fixed memory leak where long-running teammates retained all messages in AppState even after conversation compaction
- Fixed MCP server fetch caches not clearing on disconnect, causing growing memory usage with frequently reconnecting servers
- Improved memory usage in long sessions with subagents by stripping heavy progress message payloads during context compaction
UI and Workflow Improvements
Fixed local slash command output like /cost appearing as user-sent messages instead of system messages in the UI. Added "Always copy full response" option to the /copy picker - when selected, future /copy commands skip the code block picker and copy the full response directly.
VSCode: Added session rename and remove actions to the sessions list. Fixed remote sessions not appearing in conversation history.
MCP Server Enhancements
Added ENABLE_CLAUDEAI_MCP_SERVERS=false environment variable to opt out from making claude.ai MCP servers available. Added manual URL paste fallback during MCP OAuth authentication - if the automatic localhost redirect doesn't work, users can paste the callback URL to complete authentication.
Bug Fixes
- Fixed file count cache ignoring glob ignore patterns
- Fixed IDE host IP detection cache incorrectly sharing results across ports
- Fixed a race condition in the REPL bridge where new messages could arrive interleaved with historical messages during initial connection flush
- Fixed
/clearnot resetting cached skills, which could cause stale skill content to persist in new conversations
📖 Read the full source: GitHub Claude-Code
👀 See Also
Nutrition MCP Server Built with Claude Code for Food Diary Export
A developer built a nutrition MCP server using Claude Code after commercial apps blocked data export, creating a tool that logs meals via Claude, generates summaries, and exports to Excel without switching apps.
Claude Code's Read Tool Silently Downscales Images, Causing Hallucinations
Claude Code's `read` tool silently downscales images before the model sees them, leading to degraded output and unrecognized hallucinations when extracting text from screenshots.
Managing AI Context with a SQLite Knowledge Store and MCP Tools
One developer built RunawayContext, an MIT-licensed system that stores project lessons in SQLite with FTS5 and optional sqlite-vec, keeping per-session context under 3K tokens via MCP query tools and hard-coded caps.
LLM Skirmish: A Real-Time Strategy Game Benchmark for AI Coding Agents
LLM Skirmish is a benchmark where AI agents write code to play 1v1 real-time strategy games against each other. It uses a modified Screeps API and tests in-context learning across five tournament rounds.