Agent sdk setup

[akhatua2]

Description

Adds OpenHands Agent SDK as a new agent framework for CooperBench with full multi-agent collaboration support. The adapter runs agents in isolated Modal sandboxes while enabling communication via Redis messaging and Git code sharing.

Changes

New Agent: OpenHands SDK (openhands_sdk)

• Full integration of OpenHands Software Agent SDK with CooperBench
• Runs agent-server in Modal sandboxes layered on top of existing task images
• Uses SDK's native RemoteConversation and RemoteWorkspace - no custom agent loop needed
• Isolated Python 3.12 venv for agent while preserving task's original runtime
• Registers as oh shorthand for experiment naming (e.g., solo-oh-gemini-3-flash)

Multi-Agent Collaboration Infrastructure

• Messaging: New SendMessageTool and ReceiveMessageTool for Redis-based inter-agent communication
• Git Sharing: ModalGitServer provides shared bare repo with per-agent branches
• Self-Contained: Adapter manages its own Modal-based Redis/Git servers internally - orchestrator just passes flags
• Thread-safe infrastructure creation with module-level registry

Runner Improvements

• Extended AgentResult with sent_messages field for tracking tool-based agent communication
• Improved conversation extraction to support both bash-command and tool-based message formats
• Pass run_id to agents for infrastructure coordination
• Made GCP git server config more robust (reads from ConfigManager)

Bug Fixes

• Fixed messaging prompt issues
• Fixed messaging being broken in certain scenarios
• Default model changed to vertex_ai/gemini-3-flash-preview

Checklist

• Code follows project style
• Tests pass locally
• Documentation updated (if needed)