feat: add MiniMax and Zhipu LLM provider support

Add MiniMax (MiniMax-M2.5) and Zhipu as new OpenAI-compatible LLM
providers with factory routing, API endpoint config, and env vars.
Also add CLAUDE.md for Claude Code guidance.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

.env.example

@@ -4,3 +4,5 @@ GOOGLE_API_KEY=
 ANTHROPIC_API_KEY=
 XAI_API_KEY=
 OPENROUTER_API_KEY=
+ZHIPU_API_KEY=
+MINIMAX_API_KEY=

CLAUDE.md

@@ -0,0 +1,142 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Project Overview
+
+TradingAgents is a multi-agent LLM financial trading framework built with LangGraph. It simulates a trading firm with specialized agents: analysts (market, social, news, fundamentals), researchers (bull/bear), trader, and risk management teams. The system supports multiple LLM providers (OpenAI, Anthropic, Google, xAI, OpenRouter, Ollama) and data vendors (Yahoo Finance, Alpha Vantage).
+
+## Essential Commands
+
+### Installation (requires Python >=3.10)
+```bash
+pip install -r requirements.txt
+# OR for editable install with CLI entry point
+pip install -e .
+```
+
+### CLI Usage
+```bash
+python -m cli.main
+# or if installed via pip install -e .
+tradingagents
+```
+
+### Python API Usage
+```python
+from tradingagents.graph.trading_graph import TradingAgentsGraph
+from tradingagents.default_config import DEFAULT_CONFIG
+
+ta = TradingAgentsGraph(debug=True, config=DEFAULT_CONFIG.copy())
+_, decision = ta.propagate("NVDA", "2026-01-15")
+print(decision)
+
+# After observing actual returns, memorize mistakes for future runs
+# ta.reflect_and_remember(returns)  # parameter is the position returns (e.g. 1000)
+```
+
+### Testing and Linting
+No formal test suite or linting configuration exists. `test.py` at root is a manual smoke test for Yahoo Finance data functions.
+
+### Environment Variables
+Set via `.env` file (loaded with `python-dotenv`) or shell exports. Only the provider you use is required:
+```bash
+OPENAI_API_KEY=...          # OpenAI (GPT)
+GOOGLE_API_KEY=...          # Google (Gemini)
+ANTHROPIC_API_KEY=...       # Anthropic (Claude)
+XAI_API_KEY=...             # xAI (Grok)
+OPENROUTER_API_KEY=...      # OpenRouter
+ZHIPU_API_KEY=...           # Zhipu AI
+ALPHA_VANTAGE_API_KEY=...   # Alpha Vantage (optional, for non-yfinance data)
+TRADINGAGENTS_RESULTS_DIR=... # Optional results directory override
+```
+
+## Architecture
+
+### Core Workflow (LangGraph State Machine)
+The trading graph follows a sequential multi-phase workflow:
+
+1. **Analyst Phase** - Each selected analyst runs sequentially, uses tools to gather data, clears messages after completion
+2. **Research Phase** - Bull/Bear researchers debate with analyst reports as context; Research Manager finalizes after `max_debate_rounds`
+3. **Trading Phase** - Trader synthesizes research into investment plan
+4. **Risk Phase** - Aggressive/Conservative/Neutral agents debate; Risk Judge finalizes after `max_risk_discuss_rounds`
+
+Key files:
+- `tradingagents/graph/trading_graph.py` - Main TradingAgentsGraph class
+- `tradingagents/graph/setup.py` - GraphSetup for building workflow
+- `tradingagents/graph/conditional_logic.py` - Routing logic
+- `tradingagents/graph/propagation.py` - State initialization
+
+### Directory Structure
+```
+cli/                    # Command-line interface (Typer-based)
+tradingagents/
+  agents/
+    analysts/           # Initial analysis agents
+    researchers/        # Bull/bear researchers + manager
+    managers/           # Research and risk managers
+    trader/            # Trading decision agent
+    risk_mgmt/         # Risk assessment agents
+    utils/             # Agent states, tools, memory (BM25)
+  graph/               # LangGraph workflow orchestration
+  dataflows/           # Data source abstraction (vendor routing)
+  llm_clients/         # Multi-provider LLM factory
+```
+
+### LLM Client Factory
+`tradingagents/llm_clients/factory.py` implements the factory pattern for creating LLM clients. Base client class in `base_client.py` with provider-specific implementations (OpenAI, Anthropic, Google). Provider-specific thinking configuration:
+- Google: `google_thinking_level` ("high", "minimal")
+- OpenAI: `openai_reasoning_effort` ("medium", "high", "low")
+
+### Data Vendor Routing
+`tradingagents/dataflows/interface.py` routes tool calls to vendor implementations. Configuration via `default_config.py`:
+- Category-level: `data_vendors["core_stock_apis"] = "yfinance"`
+- Tool-level override: `tool_vendors["get_stock_data"] = "alpha_vantage"`
+
+Supported vendors: Yahoo Finance (free, default), Alpha Vantage (requires API key).
+
+### Agent State Management
+Agent states are TypedDict classes in `agents/utils/agent_states.py`:
+- `AgentState` - Base state for analyst agents
+- `InvestDebateState` - Research team debate state
+- `RiskDebateState` - Risk management debate state
+
+Each agent function takes `state` and returns state updates. Messages are cleared between phases to manage context window.
+
+### Tool Definition
+All data tools use `@tool` decorator from `langchain_core.tools` and are defined in `agents/utils/`. Tools route via `dataflows.interface.route_to_vendor()` to vendor implementations.
+
+## Configuration
+
+`tradingagents/default_config.py` defines the default configuration:
+- `llm_provider` - "openai", "google", "anthropic", "xai", "openrouter", "ollama"
+- `deep_think_llm` - Model for complex reasoning (default: "gpt-5.2")
+- `quick_think_llm` - Model for quick tasks (default: "gpt-5-mini")
+- `backend_url` - Custom API endpoint (default: OpenAI's endpoint)
+- `max_debate_rounds` - Research team debate iterations (default: 1)
+- `max_risk_discuss_rounds` - Risk team debate iterations (default: 1)
+- `max_recur_limit` - LangGraph recursion limit (default: 100)
+- `data_vendors` - Per-category vendor selection (categories: core_stock_apis, technical_indicators, fundamental_data, news_data)
+- `tool_vendors` - Per-tool vendor overrides (takes precedence over category-level)
+
+Copy and modify `DEFAULT_CONFIG` to customize. Always use `.copy()` to avoid mutating the shared default.
+
+## Important Patterns
+
+1. **Agent Functions** - Each agent is a function `agent(state) -> dict` returning state updates
+2. **Message Clearing** - Analysts clear messages after completion to prevent context bloat
+3. **Tool Decorators** - Use `@tool` from `langchain_core.tools` for all data tools
+4. **Factory Pattern** - LLM clients via `create_llm_client()` in `llm_clients/factory.py`
+5. **Memory** - BM25-based lexical matching in `agents/utils/memory.py` (no API calls, offline-capable)
+6. **Caching** - Data cached in `tradingagents/dataflows/data_cache/`
+
+## Results and Output
+
+Analysis logs and reports are saved to `results/` directory (configurable via `TRADINGAGENTS_RESULTS_DIR` env var). The CLI provides real-time progress display using Rich library.
+
+## Key Entry Points
+
+- CLI: `cli/main.py` (Typer app)
+- Trading Graph: `tradingagents/graph/trading_graph.py` - `TradingAgentsGraph.propagate()`
+- LLM Factory: `tradingagents/llm_clients/factory.py` - `create_llm_client()`
+- Data Interface: `tradingagents/dataflows/interface.py` - `route_to_vendor()`

cli/utils.py

@@ -160,6 +160,9 @@ def select_shallow_thinking_agent(provider) -> str:
             ("GPT-OSS:latest (20B, local)", "gpt-oss:latest"),
             ("GLM-4.7-Flash:latest (30B, local)", "glm-4.7-flash:latest"),
         ],
+        "zhipu": [
+            ("GLM-4.7 (Latest, strong reasoning, optimized for Chinese)", "glm-4.7"),
+        ],
     }
 
     choice = questionary.select(
@@ -228,6 +231,9 @@ def select_deep_thinking_agent(provider) -> str:
             ("GPT-OSS:latest (20B, local)", "gpt-oss:latest"),
             ("Qwen3:latest (8B, local)", "qwen3:latest"),
         ],
+        "zhipu": [
+            ("GLM-4.7 (Latest, strong reasoning, optimized for Chinese)", "glm-4.7"),
+        ],
     }
 
     choice = questionary.select(
@@ -262,6 +268,7 @@ def select_llm_provider() -> tuple[str, str]:
         ("xAI", "https://api.x.ai/v1"),
         ("Openrouter", "https://openrouter.ai/api/v1"),
         ("Ollama", "http://localhost:11434/v1"),
+        ("zhipu","https://open.bigmodel.cn/api/paas/v4"),
     ]
     
     choice = questionary.select(

tradingagents/default_config.py

@@ -8,7 +8,7 @@ DEFAULT_CONFIG = {
         "dataflows/data_cache",
     ),
     # LLM settings
-    "llm_provider": "openai",
+    "llm_provider": "openai",  # openai, google, anthropic, xai, openrouter, ollama, zhipu, minimax
     "deep_think_llm": "gpt-5.2",
     "quick_think_llm": "gpt-5-mini",
     "backend_url": "https://api.openai.com/v1",

tradingagents/llm_clients/factory.py

@@ -15,7 +15,7 @@ def create_llm_client(
     """Create an LLM client for the specified provider.
 
     Args:
-        provider: LLM provider (openai, anthropic, google, xai, ollama, openrouter)
+        provider: LLM provider (openai, anthropic, google, xai, ollama, openrouter, zhipu, minimax)
         model: Model name/identifier
         base_url: Optional base URL for API endpoint
         **kwargs: Additional provider-specific arguments
@@ -40,4 +40,10 @@ def create_llm_client(
     if provider_lower == "google":
         return GoogleClient(model, base_url, **kwargs)
 
+    if provider_lower == "zhipu":
+        return OpenAIClient(model, base_url, provider="zhipu", **kwargs)
+
+    if provider_lower == "minimax":
+        return OpenAIClient(model, base_url, provider="minimax", **kwargs)
+
     raise ValueError(f"Unsupported LLM provider: {provider}")

tradingagents/llm_clients/openai_client.py

@@ -29,7 +29,7 @@ class UnifiedChatOpenAI(ChatOpenAI):
 
 
 class OpenAIClient(BaseLLMClient):
-    """Client for OpenAI, Ollama, OpenRouter, and xAI providers."""
+    """Client for OpenAI, Ollama, OpenRouter, xAI, Zhipu, and MiniMax providers."""
 
     def __init__(
         self,
@@ -58,6 +58,16 @@ class OpenAIClient(BaseLLMClient):
         elif self.provider == "ollama":
             llm_kwargs["base_url"] = "http://localhost:11434/v1"
             llm_kwargs["api_key"] = "ollama"  # Ollama doesn't require auth
+        elif self.provider == "zhipu":
+            llm_kwargs["base_url"] = "https://open.bigmodel.cn/api/paas/v4"
+            api_key = os.environ.get("ZHIPU_API_KEY")
+            if api_key:
+                llm_kwargs["api_key"] = api_key
+        elif self.provider == "minimax":
+            llm_kwargs["base_url"] = "https://api.minimaxi.chat/v1"
+            api_key = os.environ.get("MINIMAX_API_KEY")
+            if api_key:
+                llm_kwargs["api_key"] = api_key
         elif self.base_url:
             llm_kwargs["base_url"] = self.base_url
 

tradingagents/llm_clients/validators.py

@@ -69,11 +69,11 @@ VALID_MODELS = {
 def validate_model(provider: str, model: str) -> bool:
     """Check if model name is valid for the given provider.
 
-    For ollama, openrouter - any model is accepted.
+    For ollama, openrouter, zhipu - any model is accepted.
     """
     provider_lower = provider.lower()
 
-    if provider_lower in ("ollama", "openrouter"):
+    if provider_lower in ("ollama", "openrouter", "zhipu", "minimax"):
         return True
 
     if provider_lower not in VALID_MODELS: