Foci Dashboard

A cloud-native collaborative studio for humans and AI to work together.

Comment

### 💡 Inspiration

  • Problem: Single-turn chat loses context, fails heavy tasks, lacks multi-agent coordination. Fragmented tools cause high friction.
  • Solution: Foci Dashboard for Track 2: Collaborative Partner—cloud workspace pairing human director with Lead Gemini and specialized background workers.

🚀 What It Does

  • Autonomous Task Delegation (delegate_worker): Lead Gemini plans and delegates sub-tasks (research, review, implement) to workers (gemini-worker, antigravity-cli).
  • Contextual Memory (rules.md / MEMORY.md): Workspace rules and project memory automatically discovered and inherited.
  • Session Management: Full serialization with auto-resumption across workspace switches.
  • Integrated Studio: Monaco editor, integrated terminal, Git diff viewer, live web preview.
  • High-Performance Compute: 200-turn adaptive loop, 16 GB Cloud Run environment for full builds and 1 m USGS LiDAR processing.

🛠️ How We Built It

  • Frontend: React, TypeScript, Vite, Tailwind CSS, Monaco Editor, WebSocket streams.
  • Backend: Node.js, WebSocket RPC, GCS FUSE mounts.
  • AI Engine: Google Gemini API (gemini-3.7-flash / gemini-2.5-pro with reasoning), Antigravity CLI.
  • System Integration: Custom Linux syscall shims, persistent OAuth, shell tooling.
  • Infrastructure: Google Cloud Run (16 GB RAM / 4 vCPUs) + persistent GCS volumes.

🧗 Key Challenges & Fixes

  • Cloud Storage POSIX: Standard tools failed chmod on GCS FUSE. Fix: Custom syscall shim (libfuse-chmod-shim.so via /etc/ld.so.preload) for 100% native Git support.
  • Agent Resilience: Loops stall on long tasks. Fix: Adaptive progress engine with duplicate-action guards; supports 100+ turns without infinite loops.
  • OAuth Persistence: Container rebuilds wiped tokens. Fix: Bidirectional sync between container and persistent cloud storage.

🏆 Accomplishments

  • Multi-Agent Workspace: Autonomous initiative, strict rule compliance, verified project output.
  • Zero-Loss Continuity: Total state and trajectory restoration across workspaces.
  • Browser-Native Cloud Studio: Zero local setup, high-memory capacity.

📚 Key Learnings

  • Division of labor: Lead orchestrator + bounded execution workers = optimal multi-agent performance.
  • Persistent context: MEMORY.md and rules.md drastically cut hallucinations and prompt repetition.

🔮 What's Next

  • Multi-User Sync: Real-time human-to-human collaboration in active agent workspaces.
  • Voice Control: Hands-free task delegation and status queries.
  • Visual Workflow Builder: Drag-and-drop DAG builder for custom agent pipelines.

Built With

Share this project:

Updates

Submission history