📝 docs: update CLAUDE.md for v0.2.0

- Add version information and overview of new evaluation system features
- Expand model provider support documentation (SiliconFlow integration)
- Add comprehensive evaluation commands section with graph trajectory and multi-turn evaluations
- Update environment setup with SILICONFLOW_API_KEY configuration
- Reorganize model integrations documentation under src/common/models/ structure
- Add evaluation system file structure documentation (tests/evaluations/)
- Enhanced development guidelines with evaluation best practices and multi-model testing
- Update Python configuration with new evaluation dependencies (openevals, agentevals, langsmith)

Author: webup

CLAUDE.md

@@ -4,7 +4,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 ## Project Overview
 
-This is a LangGraph ReAct (Reasoning and Action) Agent template that implements an iterative reasoning agent using LangGraph framework. The agent processes user queries through a cycle of reasoning, action execution, and observation.
+This is a LangGraph ReAct (Reasoning and Action) Agent template that implements an iterative reasoning agent using LangGraph framework. The agent processes user queries through a cycle of reasoning, action execution, and observation. **Version 0.2.0** introduces comprehensive evaluation systems and expanded model provider support.
 
 ## Architecture
 
@@ -13,11 +13,12 @@ The core architecture follows a modular stateful graph pattern:
 - **Common Module**: Shared components in `src/common/` provide reusable functionality across agents
 - **State Management**: Uses `State` and `InputState` dataclasses (defined in `src/react_agent/state.py`) to track conversation messages and execution state
 - **Graph Structure**: The main graph is defined in `src/react_agent/graph.py` with two primary nodes:
-  - `call_model`: Handles LLM reasoning and tool selection  
+  - `call_model`: Handles LLM reasoning and tool selection
   - `tools`: Executes selected tools via ToolNode
 - **Execution Flow**: `call_model` → conditional routing → either `tools` (if tool calls needed) or `__end__` (if ready to respond) → back to `call_model` (creates the ReAct loop)
 - **Context System**: Runtime context defined in `src/common/context.py` provides model configuration, system prompts, and DeepWiki integration control
 - **Dynamic Tools**: Runtime tool loading with MCP integration for external documentation sources (DeepWiki MCP server)
+- **Model Providers**: Multi-provider support including Anthropic, OpenAI, Qwen, QwQ, QvQ, and SiliconFlow with regional endpoint configuration
 
 ## Development Commands
 
@@ -40,6 +41,22 @@ make test_watch_e2e         # Run e2e tests in watch mode
 make extended_tests         # Run extended test suite
 ```
 
+### Evaluations
+```bash
+# Comprehensive evaluation suite (NEW in v0.2.0)
+make evals                   # Run all evaluations (graph + multiturn)
+make eval_graph              # Run graph trajectory evaluations (LLM-as-judge)
+make eval_multiturn          # Run multi-turn chat evaluations (role-persona simulations)
+
+# Model-specific evaluations
+make eval_graph_qwen         # Test with Qwen/Qwen3-8B model
+make eval_graph_glm          # Test with THUDM/GLM-4-9B model
+
+# Persona-specific evaluations
+make eval_multiturn_polite   # Test with polite persona
+make eval_multiturn_hacker   # Test with hacker persona
+```
+
 ### Code Quality
 ```bash
 make lint                   # Run linters (ruff + mypy)  
@@ -57,8 +74,12 @@ make dev_ui                 # Start LangGraph development server with UI
 ### Environment Setup
 - Copy `.env.example` to `.env` and configure API keys
 - **Required**: `TAVILY_API_KEY` for web search functionality
-- **Model Providers**: `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, `DASHSCOPE_API_KEY` (for Qwen models)
-- **Optional**: `REGION` (set to `prc` or `international` for Qwen API endpoints)
+- **Model Providers**:
+  - `ANTHROPIC_API_KEY` for Anthropic models
+  - `OPENAI_API_KEY` for OpenAI models
+  - `DASHSCOPE_API_KEY` for Qwen/QwQ/QvQ models
+  - `SILICONFLOW_API_KEY` for SiliconFlow models (NEW in v0.2.0)
+- **Optional**: `REGION` (set to `prc`/`cn` or `international`/`en` for regional API endpoints)
 - **Optional**: `ENABLE_DEEPWIKI=true` to enable DeepWiki MCP documentation tools
 - **Default Model**: Uses `qwen:qwen-flash` as default model (configurable via `MODEL` environment variable)
 
@@ -72,14 +93,23 @@ make dev_ui                 # Start LangGraph development server with UI
 - `src/common/context.py`: Runtime context and configuration with environment variable support and DeepWiki integration
 - `src/common/tools.py`: Tool definitions including web search and dynamic MCP tool loading
 - `src/common/mcp.py`: MCP client management for external documentation sources (e.g. DeepWiki)
-- `src/common/models.py`: Custom model integrations (Qwen, QwQ, QvQ) with regional API support
+- `src/common/models/`: Model provider integrations with regional API support
+  - `qwen.py`: Qwen, QwQ, QvQ model integrations via DashScope
+  - `siliconflow.py`: SiliconFlow model integrations (NEW in v0.2.0)
 - `src/common/prompts.py`: System prompt templates
-- `src/common/utils.py`: Shared utility functions
+- `src/common/utils.py`: Shared utility functions including model loading
 
 ### Configuration
 - `langgraph.json`: LangGraph Studio configuration pointing to the main graph
 - `.env`: Environment variables for API keys and configuration
 
+### Evaluation System (NEW in v0.2.0)
+- `tests/evaluations/`: Comprehensive evaluation framework
+  - `graph.py`: Graph trajectory evaluation using AgentEvals with LLM-as-judge methodology
+  - `multiturn.py`: Multi-turn chat evaluation with role-persona simulations
+  - `config.py`: Evaluation configuration and test scenarios
+  - `utils.py`: Shared evaluation utilities and scoring functions
+
 ## LangGraph Studio Integration
 
 This project works seamlessly with LangGraph Studio. The `langgraph.json` config file defines:
@@ -105,13 +135,19 @@ This project integrates with Model Context Protocol (MCP) servers for dynamic ex
 ## Python Configuration
 
 - Python requirement: `>=3.11,<4.0`
-- Main dependencies: LangGraph, LangChain, provider-specific packages, langchain-mcp-adapters
-- Development tools: mypy, ruff, pytest
+- Main dependencies: LangGraph, LangChain, provider-specific packages, langchain-mcp-adapters, langchain-siliconflow
+- Development tools: mypy, ruff, pytest, langgraph-cli, langgraph-sdk
+- Evaluation dependencies: openevals, agentevals, langsmith (NEW in v0.2.0)
 - Package structure supports both standalone and LangGraph template usage
 
 ## Development Guidelines
 
 - **Research Tools**: Use `context7` and/or `deepwiki` to study unfamiliar projects or frameworks
 - **Code Quality**: Always run `make lint` after completing tasks
 - **Testing**: Comprehensive test suite includes unit, integration, and e2e tests with DeepWiki MCP integration coverage
-- **MCP Integration**: DeepWiki tools are dynamically loaded when `enable_deepwiki=True` in context configuration
+- **Evaluation**: Use `make evals` to run comprehensive agent evaluations (NEW in v0.2.0)
+  - Graph trajectory evaluation with LLM-as-judge methodology
+  - Multi-turn conversation evaluation with role-persona simulations
+  - Model-specific testing across different providers (Qwen, GLM, etc.)
+- **MCP Integration**: DeepWiki tools are dynamically loaded when `enable_deepwiki=True` in context configuration
+- **Multi-Model Support**: Test across different providers using evaluation commands for comprehensive coverage