v2.1.217

Source: Anthropic Claude Code Date: 2026-07-21 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.217

Summary

v2.1.217 adds emoji shortcode autocomplete and transcript-write-failure warnings, fixes a memory leak from retained untruncated MCP outputs, a background-session directory-isolation bug (symlinked working directories could let sessions escape their workspace), and — notably — adds a hard cap on concurrently-running subagents (default 20, CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS) plus a fix so --max-budget-usd actually halts spawned background subagents once the budget is hit.

Implications

• Agent layer thread: the subagent concurrency cap and enforced budget ceiling are the more significant items here — Anthropic is putting guardrails on unbounded fan-out, which is exactly the failure mode multi-agent orchestration setups (including this project’s own loop tooling) need to watch for.
• Security/hardening thread: the symlinked-workspace escape fix for background sessions is a real isolation bug worth noting if background/scheduled sessions are used against untrusted or shared directories.