By ByteBot
Claude Code’s July 17 /fork rework got the headlines. The follow-up drop — v2.1.215, shipped July 18-19 — is the one worth reading if you’re running agentic workflows at any real scale. Subagent spawn caps, MCP auto-background, a session resurrection picker, a kill switch for abusive sessions, and a batch of Windows security fixes: this is Claude Code treating itself like a production runtime instead of a clever demo.
Why This Update Exists: Runaway Loops Cost Real Money
Agentic loops have a billing problem. When Claude Code spawns subagents to parallelize work and something goes wrong — a hook recurses, a task gets stuck, a delegation chain has no exit condition — the session doesn’t crash. It loops. And every iteration costs tokens. Developers have reported $400 bills from a single overnight session; others have logged five-figure losses. Anthropic’s own April post-mortem documented a bug where a hook chain recursed without a timeout or depth limit and ran past its wall-clock budget. The caps in v2.1.215 are the direct response.
Subagent and WebSearch Caps
Claude Code now enforces a default limit of 200 subagent spawns per session. When the cap is hit, the Agent tool fails gracefully and Claude falls back to completing remaining work directly with its own tools — no silent looping, no crash. A companion limit of 200 also applies to WebSearch calls. Both limits reset when you run /clear.
You can tune both with environment variables:
# Raise the subagent cap for large parallel workflows
CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION=500
# Raise or lower the WebSearch cap
CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION=400200 is a reasonable default for most sessions. If your workflow legitimately exceeds 200 subagents, you almost certainly know it — raise the limit explicitly rather than discovering you need it mid-task.
MCP Auto-Background
Before this update, a slow MCP tool call blocked the entire session. Database migrations, file indexing operations, calls to third-party APIs that take three-plus minutes — all of them pinned your session to a spinner. Now, any MCP call still running after two minutes automatically moves to a background task. Claude receives the task ID immediately and keeps working. The result arrives as a task notification; view or stop background tasks at any point with /tasks.
The threshold is configurable via environment variable:
# Change the auto-background threshold (default: 120000ms = 2 min)
CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS=300000 # 5 minutes
# Disable auto-background entirely
CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS=0
# Or disable all background task features
CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1Per-call wall-clock and idle timeouts still apply while tasks run in the background. This isn’t a free pass to leave slow calls running indefinitely.
/resume Now Finds Sessions You Deleted
Typing /resume during any conversation now opens an interactive picker that includes sessions you’ve removed from the agent view — not just active ones. The picker surfaces session summaries, message counts, git branch names, and timestamps. Selected sessions resume as background sessions, so your current conversation stays intact. This fixes the classic problem: deleting a session to keep the UI clean, then realizing twenty minutes later you needed context from it.
The EndConversation Tool
Claude Code can now end sessions with abusive users or jailbreak attempts — the same capability that’s existed on claude.ai since 2025. The more interesting detail is what happens in background agents: EndConversation has no effect in a background fork. The system instructs background sessions to return only if welfare concerns require stopping the task, not to use the tool as an exit mechanism.
That framing — “welfare concerns” — is worth noting. It’s Anthropic explicitly carving out space for Claude to have something like interests even in an automated IDE context. Whether you read that as safety engineering or something more, it signals how Anthropic thinks about the system they’re building.
Windows PowerShell 5.1 Security Fix
Windows developers on PowerShell 5.1 — still the default in many enterprise environments — were exposed to a permission-check bypass that let commands execute without going through Claude Code’s approval flow. This is fixed in v2.1.215, along with Python encoding crashes, file output encoding problems, and tool commands that hung when a child process waited on stdin. If you use Claude Code on Windows and haven’t updated, update now.
Other Fixes in This Wave
Several smaller issues were resolved: plugins loaded via the --settings CLI flag now load correctly; feature flags no longer go stale in long-running sessions after OAuth token rotation; /ultrareview works again in repos without a merge base; and background sessions parked idle no longer keep the daemon alive indefinitely.
What This Update Signals
None of these features are for someone who occasionally asks Claude Code to explain a function. Subagent caps, MCP auto-background, progress heartbeats for long tasks, and EndConversation are features for teams running Claude Code in CI pipelines, in overnight research loops, at scale. The product is maturing to match its actual use case — and the guardrails are a sign that Anthropic has seen, firsthand, what happens without them.
The full diff is in the official Claude Code changelog and GitHub releases page. The sub-agents reference has been updated to reflect the new session caps.
