feat: split xagent/core/agent.py into focused collaborative modules

[Copilot]

xagent/core/agent.py was a 1,128-line monolith mixing OpenAI API calls, tool execution, MCP server management, image captioning, and conversation orchestration.

New modules

Module	Responsibility
xagent/core/model_caller.py	ReplyType enum + ModelCaller — OpenAI API calls, retry, streaming event classification/extraction, input sanitization
xagent/core/tool_executor.py	ToolExecutor — tool registration, invalidation-aware spec cache, semaphore-controlled concurrent execution; make_http_agent_tool() HTTP sub-agent factory
xagent/core/mcp_manager.py	MCPManager — MCP server connections, TTL-based caching, tool discovery with tenacity retry
xagent/core/image_processor.py	ImageProcessor — vision-model caption generation with prompt-based fallback
agent.py (refactored)	Orchestration only: conversation loop, memory, message storage, background tasks, sub-agent conversion (~570 lines, down from 1,128)

All four components are composed inside Agent.__init__:

self._image_processor = ImageProcessor(self.client, ...)
self._tool_executor    = ToolExecutor(self.logger, self._image_processor)
self._mcp_manager      = MCPManager(servers=..., cache_ttl=..., logger=self.logger)
self._model_caller     = ModelCaller(self.client, self.model, self.logger)

Additional changes

• Lazy storage imports in agent.py: MemoryStorageLocal/MessageStorageLocal are deferred to __init__ body so importing the module no longer eagerly loads chromadb/langfuse. This incidentally fixed several pre-existing test failures (11 → 49 passing in the existing suite).
• xagent/core/__init__.py now uses __getattr__-based lazy loading (matching the top-level package pattern) and exports all new public classes.
• tests/test_modular_core.py: 32 new unit tests covering each module's public API, static helpers, cache-invalidation logic, TTL behaviour, captioning fallback, and Agent backward-compat proxies.

Backward compatibility

All existing Agent attributes (tools, mcp_tools, cached_tool_specs, mcp_servers) are preserved as property proxies to the internal components. _register_tools() delegates to ToolExecutor.register() so callers like Swarm continue to work unchanged.

Original prompt

建议：将 Agent 拆分为多个专注的协作模块：

xagent/core/
├── agent.py # 只负责对话编排（< 300 行）
├── model_caller.py # 封装 OpenAI API 调用和重试逻辑
├── tool_executor.py # 工具执行引擎（并发、信号量控制）
├── mcp_manager.py # MCP 服务器连接、缓存和工具发现
└── image_processor.py # 图像字幕生成

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.