[BUG] Memory leak: Single agent task consumes 15+ GB RAM when using custom orchestrator agent

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

When using a custom orchestrator agent via a slash command, a single claude process consumes all available RAM (15+ GB) within ~2 minutes, causing the system to become completely unresponsive.

The memory leak occurs in a SINGLE process - not from spawning multiple parallel tasks. The orchestrator task itself consumes all RAM before it even gets to delegate work to other agents.

Evidence from ps aux during the issue:

• Single claude process: 91.8% MEM, 15.4 GB RSS, 90 GB VSZ
• Process status: Dl+ (uninterruptible sleep)
• System: 16 GB RAM + 8 GB Swap completely exhausted

What Should Happen?

The orchestrator agent task should execute with reasonable memory usage (< 1 GB) and successfully delegate tasks to specialist agents without exhausting system memory.

Direct delegation to individual agents (e.g., /delegate @backend-developer) works fine with normal memory usage - only the orchestrator pattern triggers the leak.

Error Messages/Logs

No error messages - system becomes unresponsive before any error can be shown.

  Memory monitoring log captured before freeze:

  === Mon Dec 15 14:46:13 UTC 2025 ===
  USER       PID   %CPU %MEM    VSZ        RSS        STAT  COMMAND
  elzorro    29858 72.0 91.8 90395356   15417580    Rl+   claude
  elzorro    1127   6.5  2.7 75158276     466408    Sl+   claude

                total       used        free      shared  buff/cache   available
  Memory:       16Gi        15Gi       129Mi       3.1Mi       553Mi       680Mi
  Swap:         8.0Gi       1.3Gi      6.7Gi

  === Mon Dec 15 14:46:15 UTC 2025 ===
  Memory:       16Gi        15Gi       131Mi       3.1Mi       541Mi       670Mi
  Swap:         8.0Gi       1.5Gi      6.5Gi

  The problematic process (PID 29858) was the orchestrator task, consuming 15.4 GB while the main claude session (PID 1127) used only 466 MB.

Steps to Reproduce

1. Create custom agents in ~/.claude/agents/:

• orchestrator.md (with tools: Read, Grep, Glob, TodoWrite, Task, TaskOutput)
• Specialist agents (backend-developer.md, frontend-developer.md, etc.)

1. Create ~/.claude/commands/start.md that launches orchestrator:

Task(
subagent_type="orchestrator",
prompt="[orchestration instructions ~100 lines]",
description="Orchestrator"
)

1. Navigate to any project directory (even empty)

1. Run: /start Create a simple FastAPI app

1. Watch memory consumption - within 2 minutes RAM is exhausted

Workaround: Using direct /delegate @agent-name works fine - only the orchestrator pattern causes the leak.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.0.67

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Non-interactive/CI environment

Additional Information

• Claude Code Version: 2.0.67
• OS: Debian GNU/Linux 13 (trixie)
• Kernel: Linux 6.17.2-2-pve x86_64
• RAM: 16 GB + 8 GB Swap
• Environment: LXC Container (Proxmox)