midsane-sui 0.1.0

AI Coding Agent Runtime

Entities:

Agent Task Execution Tool Sandbox Memory LLM Model Evaluation Reflection

            Agent
              │
    ┌─────────┴─────────┐
    │                   │
  Model            ToolPermission
    │
    │

Execution │ ┌──────┼─────────────┐ │ │ │ │ │ │ Logs ToolCalls Memory │ │ │ │ │ │ └──────┼─────────────┘ │ Evaluation │ Reflection

Database SQLAlchemy models:

Agent Execution Model Tool Memory ExecutionLog

Plain python classes:

Planner Loop Engine Context Builder Docker Executor LLM Client Reflection Engine Memory Retriever

app/ ├── agents/ │ ├── init.py │ ├── models.py │ ├── schemas.py │ ├── repository.py │ ├── service.py │ ├── exceptions.py │ └── api.py # optional later │ ├── execution/ │ ├── init.py │ ├── models.py │ ├── schemas.py │ ├── repository.py │ ├── service.py │ ├── exceptions.py │ └── api.py │ ├── models/ │ ├── init.py │ ├── models.py │ ├── schemas.py │ ├── repository.py │ ├── service.py │ └── providers.py │ ├── tools/ │ ├── init.py │ ├── models.py │ ├── schemas.py │ ├── repository.py │ ├── service.py │ └── registry.py │ ├── memory/ │ ├── init.py │ ├── models.py │ ├── repository.py │ ├── service.py │ └── retriever.py │ ├── sandbox/ │ ├── init.py │ ├── docker.py │ ├── filesystem.py │ └── executor.py │ ├── runtime/ │ ├── init.py │ ├── planner.py │ ├── loop.py │ ├── executor.py │ ├── context_builder.py │ └── reflection.py │ ├── llm/ │ ├── init.py │ ├── client.py │ └── providers/ │ ├── init.py │ ├── openai.py │ ├── anthropic.py │ ├── ollama.py │ └── openrouter.py

execution lifestyle
User Request
  │
  ▼

Create Execution │ ▼ Load Agent │ ▼ Load Model │ ▼ Build Context │ ▼ Planner │ ▼ Loop Engine │ ▼ Tool Calls │ ▼ Reflection │ ▼ Evaluation │ ▼ Complete Execution

This is the order I'd implement entities.

Phase 1 ✅ Agent

Done.

Phase 2 Model

Because agents need models.

Phase 3 Execution

Now agents can actually execute tasks.

Phase 4 Tool

Now executions can use tools.

Phase 5 ExecutionLog

Now we can observe what happened.

Phase 6 Memory

Now executions become stateful across runs.

Phase 7 Evaluation

Now we can score execution quality.

Phase 8 Reflection

Now the system can improve itself.

[ User Input ] │ ▼ ┌─────────────────────────┐ │ Orchestrator Router │ ──(Simple Q&A)──► [ Simple LLM Stream ] └─────────────────────────┘ │ (Task detected) ▼ ┌─────────────────────────┐ │ Planner Agent │ ◄──► [ Memory / Chat History ] │ (Defines Scope & Stack) │ └─────────────────────────┘ │ (User Confirms Plan) ▼ ┌─────────────────────────┐ │ Execution Agent (Loop) │ │ - ReAct Loop │ │ - Call MCP Tools │ ──► [ Sandbox / Terminal ] │ - Track Logs & Tokens │ └─────────────────────────┘