moldgen-agent-dev

name	moldgen-agent-dev
description	Develop and extend MoldGen's built-in Agent system. Use when creating new agents, adding tools to the registry, modifying agent routing, updating execution pipelines, or working with moldgen/ai/ files.

MoldGen Agent System Development

Architecture

moldgen/ai/
├── agent_base.py       # BaseAgent ABC, AgentConfig, AgentContext, AgentEvent, StepResult
├── execution_engine.py # AgentExecutionEngine, PlanStep, ExecutionPlan, PIPELINE_TEMPLATES
├── tool_registry.py    # ToolRegistry singleton, ToolDef, ToolParam, ToolResult
├── memory.py           # ShortTermMemory, LongTermMemory, AgentMemoryManager
├── service_manager.py  # AIServiceManager — unified AI API access
├── chat.py             # Chat completion logic
├── agents/
│   ├── master_agent.py  # MasterAgent — intent routing + LLM reasoning + self-reflection
│   ├── model_agent.py   # ModelAgent — mesh import/repair/edit
│   ├── mold_agent.py    # MoldDesignAgent — orientation/parting/shells
│   ├── insert_agent.py  # InsertAgent — support plate design
│   ├── simopt_agent.py  # SimOptAgent — simulation + optimization
│   └── creative_agent.py# CreativeAgent — AI image/3D generation
└── prompts/             # System prompt templates for each agent

Adding a New Agent

1. Create moldgen/ai/agents/your_agent.py:

   • Subclass BaseAgent from agent_base.py
   • Implement: name, description, system_prompt, get_available_tools(), execute()
   • Use self.call_tool(name, **kwargs) for tool execution (has built-in retry)
   • Use self.emit_event(type, data) for real-time event tracking

2. Add AgentRole enum value in agent_base.py

3. Register in moldgen/ai/agents/__init__.py

4. Register in moldgen/api/routes/ai_agent.py (add to _engine.register_agent(...))

5. Add routing keywords in master_agent.py KEYWORD_ROUTES

Adding a New Tool

Register in ToolRegistry._register_builtin_tools():

self.register(ToolDef(
    name="tool_name",
    description="工具描述",
    category="model|mold|insert|sim|export|ai",
    parameters=[
        ToolParam("param_name", "string|number|boolean|array|object", "描述", required=True),
    ],
    handler=actual_function,  # async or sync
    requires_confirmation=False,
))

Thinking Styles

• ThinkingStyle.FAST: Keyword matching only, no LLM calls
• ThinkingStyle.BALANCED: LLM classification with keyword fallback
• ThinkingStyle.DEEP: Chain-of-thought reasoning + self-reflection + fallback

Key Patterns

• All agents share AgentContext for cross-agent state (model_id, mold_id, etc.)
• AgentConfig controls per-agent behavior (retries, temperature, timeouts)
• AgentMemoryManager provides short-term (session) and long-term (persisted) memory
• Events (AgentEvent) enable real-time UI updates via WebSocket
• API routes at /api/v1/ai/agent/ expose config, memory, history, execution