Multi-agent setup (6-8 windows) causing repeated system crashes on Windows -- v2.1.118

Summary

Running a multi-agent Claude Code setup (1 Opus PM + 6-8 Sonnet worker windows) on Windows causes repeated full system crashes. Crashes require full restart of all agent windows and interrupt overnight autonomous pipeline runs.

Environment

• OS: Windows
• Claude Code: v2.1.118
• Setup: 1 Opus window (PM/coordinator) + 6-8 Sonnet windows (workers)
• Coordination: filesystem handshake files (.claude/handshake/)
• Runtime: overnight autonomous runs (8+ hours)

Behavior

System crashes repeatedly during sustained multi-agent operation. Each crash requires:

• Full system restart
• Manual relaunch of all 6-8 agent windows
• Re-approval of permission prompts in each window (compounded by #37442)
• Manual check of which pipeline steps completed before crash

Suspected contributors

1. Orphaned subagent processes not releasing resources after task completion (related: #51860, #52016)
2. Sustained disk I/O from continuous parquet writes + agent context writes to ~/.claude/projects/
3. Memory pressure from 6-8 simultaneous large-context sessions (each loading ~200KB CLAUDE.md per turn)
4. CPU load from permission prompt polling when #37442 blocks workers

Impact

Overnight autonomous runs cannot complete reliably. Every crash loses 1-3 hours of pipeline progress and requires manual intervention -- defeating the purpose of autonomous operation.

Request

• Memory and CPU profiling of multi-window Claude Code setups on Windows
• Resource limits or warnings when total Claude Code process memory exceeds threshold
• Graceful checkpoint on crash so sessions can resume without full manual restart