release: v1.2.0 - Agent resilience and crash recovery

[williamzujkowski]

Summary

This release adds comprehensive agent process resilience features to prevent and recover from AI CLI crashes (particularly Claude Code memory leaks).

Key Features

Agent Resilience Phase (New Bootstrap Phase 8)

• Claude Memory Watchdog: systemd service that monitors Claude CLI memory usage
  • Warns at 8GB, kills at 13GB threshold
  • Prevents cascade failures across parallel clones
• cgroups Memory Limiting: run-claude-limited wrapper enforces hard limits
  • Default 12GB limit, customizable
  • Uses systemd cgroups v2
• Session Persistence: agent-session tmux wrapper
  • Auto-reattach on reconnect
  • Crash event logging
• Enhanced Health Check: vm-health-check with trend analysis
  • Memory trend prediction
  • OOM alerts 30-60 minutes before exhaustion

Resource Planning

• Default RAM increased: 8GB → 16GB
• Default swap increased: 4GB → 8GB
• --memory and --vcpus flags for setup_cloud.sh
• New RESOURCES.md with parallel deployment guide

Cleanup

• Removed .mcp.json from git (local MCP config)
• Removed vestigial guest/vm-health-check.sh (now embedded in bootstrap)
• Fixed README accuracy (swap size, directory structure)

New Commands (Inside VM)

vm-health-check              # Quick health with Claude memory
vm-health-check --watch      # Continuous monitoring
vm-health-check --trend      # Memory trend analysis
run-claude-limited           # Run Claude with 12GB limit
run-claude-limited 8G        # Custom limit
agent-session                # Persistent tmux session
systemctl status claude-watchdog  # Check watchdog

Test Plan

• make lint passes
• README accuracy verified
• CHANGELOG updated
• Directory structure accurate
• Bootstrap runs successfully in VM (manual test)
• Watchdog service starts and monitors correctly

🤖 Generated with Claude Code