Claude Code v2.1.217: Subagent Depth Cap, Emoji Autocomplete, and Critical Fixes
Anthropic released Claude Code v2.1.217 with notable changes to subagent governance, a long-requested emoji autocomplete feature, and a dozen bug fixes addressing memory leaks, session isolation, and Windows auto-update corruption.
Subagent Controls
The release introduces two environment variables to prevent runaway background agents:
CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS(default: 20) — caps how many subagents can run in parallel from a single message.CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH— subagents no longer spawn nested agents by default. Set this to allow deeper nesting.
Also, --max-budget-usd now stops background subagents: once the budget cap is hit, new spawns are denied and running background agents are halted.
Emoji Autocomplete
You can now type :heart: to insert ❤️, or :hea for suggestions. Disable with the emojiCompletionEnabled setting.
Notable Bug Fixes
- Memory leak fixed — truncated MCP tool outputs no longer keep the full untruncated result in memory for the rest of the session.
- Windows auto-update corruption — failed updates that could leave
claude.exemissing now restore the preserved executable automatically. - Background session isolation — symlinked working directories are now canonicalized, preventing sessions from escaping their workspace folder.
- Claude Opus 4.8 on Bedrock — auto-compact now triggers properly, and
/compactno longer fails once over the limit. - Corporate settings — mTLS, TLS-verify, OAuth scope, and proxy settings are now respected in Claude Desktop sessions.
- Screen reader mode — startup announcement no longer cut off, and the thinking status row re-renders efficiently.
- Telemetry routing — managed
OTEL_EXPORTER_OTLP_ENDPOINTnow governs all signals; lower-scope overrides no longer redirect telemetry away. - Transcript handling —
--resumeand/resumeno longer crash with malformed attachment entries; warnings added when transcript writes fail (e.g., disk full). - Remote Control — permission prompt now shows to viewers that connect after the prompt appears.
- Brace expansion —
CLAUDE.md/SKILL.mdpaths with many brace groups no longer OOM-kill the CLI; expansion is now budget-bounded. - Footer links — PR badge links are clickable hyperlinks even over ssh/tmux; set
FORCE_HYPERLINK=0to opt out.
Other Changes
- Login-expiry warning moved from 5 to 3 days before expiry.
- Frontend-design plugin suggestion tip capped at 3 lifetime impressions.
- Transcript preview gap fixed when attaching to a starting background session.
📖 Read the full source: GitHub Claude-Code
👀 See Also
Claude Code Skills for Automated Project Scaffolding
A developer has built Claude Code skills that automate full-stack project setup with commands for React, Next.js, Node.js APIs, and Turborepo monorepos. The skills pull latest dependencies, support 50+ integrations, and are MIT licensed.
Kvaser: An Open-Source Local-First AI Orchestrator with Sub-Agent Routing and Wolfram Integration
Kvaser is a Man-in-the-Middle MCP proxy that orchestrates sub-agents with smart tool whitelisting, zero-embedding RAG via Kiwix, and Wolfram Engine integration for symbolic math. Built with Qwen 3.6 35B and sub-agent routing to different models/machines.
GLM-5-Turbo Shows Low Tool Call Error Rate in User Testing
The z-ai/glm-5-turbo model demonstrates a 0.57% average tool call error rate in testing, significantly lower than GLM-5's ~3% rate. A user reported successfully using it with a CLI tool to write a 97,000-word fantasy novel with minimal issues.
Google Surf MCP: Free Google Search MCP with PDF Handling and Tiered Extraction
Google Surf MCP is a free MCP server for Google search and URL extraction that handles PDFs and offers a tiered extraction mode (abstract/full) to save tokens.