feat: initial AgentSandbox CrewAI orchestration app

[phoenixfire808]

What

Initial commit of the AgentSandbox CrewAI orchestration app — a private, local-first workflow execution layer over Drew's selected local Ollama LLM.

Key features

• CrewAI Flow — manager (plans) → specialist (executes) → reviewer (audits) pipeline
• FastAPI service on localhost:8090 with SQLite job persistence
• Dashboard UI at http://localhost:8090/dashboard for browser-based objective submission and job monitoring
• CLI via python cli.py for terminal-based job submission
• Git workspace for artifact commits
• Ollama LLM integration via LiteLLM — no cloud calls, no Qwen
• Local UNCENSORED LFM model — no Qwen or coder fallback
• WorkspaceWriteTool with path-traversal guards
• Async kickoff for FastAPI event-loop compatibility
• Startup recovery — interrupted jobs marked FAILED with RECOVERED event

Acceptance verified

• GET /health → 200, all services connected
• Full manager → specialist → reviewer pipeline executed end-to-end
• Git commit created in workspace repo
• Dashboard loads in browser at localhost:8090

Running

.venv\Scripts\activate
python start.py
start http://localhost:8090/dashboard

Fixmes logged

• Specialist writes to absolute container paths on Windows (path translation); tracked separately.