[BUG] OOM crash during subagent execution (8GB heap)

### Environment

• Claude Code: 2.1.7
• Node.js: 25.2.1 (Homebrew)
• OS: macOS Darwin 24.6.0 (Apple Silicon)
• Project: ~56,000 files, 7 custom subagents in .claude/agents/

### Problem
Fresh session crashes with OOM during subagent execution after minimal operations.

### Reproduction

1. Fresh claude session (no --resume)
2. Simple code review request referencing one file
3. Main agent reads ~500 lines, delegates to custom subagent
4. Subagent performs ~25 tool uses (Search, Read operations)
5. Crash at 8.1GB heap after 67 seconds

### Session log
❯ check RunAssets in @app/queries/assets_query.py [simple question]
⎿ Read app/queries/assets_query.py (250 lines)
⏺ Read(app/domain/run.py) - 265 lines
⏺ Search (2x)
⏺ domain-engineer(Refactor RunAssets dataclass)
⎿ 25+ tool uses
· Ionizing… (1m 7s · ↓ 1.9k tokens)
--- CRASH ---

### Error
Mark-Compact 8118.2 (8221.8) -> 8116.0 (8169.3) MB
average mu = 0.21
FATAL ERROR: Ineffective mark-compacts near heap limit

### What I tried

• NODE_OPTIONS="--max-old-space-size=8192" → crashes at 8GB
• NODE_OPTIONS="--max-old-space-size=16384" → crashes at 16GB
• --dangerously-skip-permissions → same crash
• State file ~/.claude.json is only 48KB
• Added .claudeignore excluding .venv, __pycache__, node_modules

### Expected behavior
Reading 500 lines of code + 25 subagent tool calls should not require 8GB+ memory.

### Hypothesis
Memory leak during subagent execution - tool results not being garbage collected, or subagent context accumulating in parent process.

---

The issue description was made by claude code itself, but I am confirm the issue
The crashes started once the CC was updated 2.0.76 -> 2.1.6