Merge pull request #52 from octo-patch/feature/add-minimax-provider

feat: add MiniMax as first-class LLM provider

GETTING_STARTED.md

@@ -80,6 +80,16 @@ claude_config = {
 }
 ```
 
+```python
+# MiniMax 使用 OpenAI 兼容格式，变量名含 'oai' 即可
+# 温度自动修正为 (0, 1]，支持 M2.7 / M2.5 全系列，204K 上下文
+oai_minimax_config = {
+    'apikey': 'eyJh...',
+    'apibase': 'https://api.minimax.io/v1',
+    'model': 'MiniMax-M2.7',
+}
+```
+
 **使用标准工具调用格式（适合较弱模型）：**
 
 ```python
@@ -105,6 +115,7 @@ native_claude_config = {
 | `native` + `oai` | OpenAI 标准工具调用 | 较弱模型推荐，工具调用更规范 |
 
 > 例：用 Claude 模型，但 API 服务提供的是 OpenAI 兼容接口 → 变量名用 `oai_xxx`。
+> 例：用 MiniMax 模型 → 变量名用 `oai_minimax_config`，MiniMax 走 OpenAI 兼容接口。
 
 **`apibase` 填写规则**（会自动拼接端点路径）：
 

README.md

@@ -22,7 +22,7 @@ Every time GenericAgent solves a new task, it automatically crystallizes the exe
 - **Self-Evolving**: Automatically crystallizes each task into an skill. Capabilities grow with every use, forming your personal skill tree.
 - **Minimal Architecture**: ~3,300 lines of core code. Agent Loop is just 92 lines. No complex dependencies, zero deployment overhead.
 - **Strong Execution**: Injects into a real browser (preserving login sessions). 7 atomic tools take direct control of the system.
-- **High Compatibility**: Supports Claude / Gemini / Kimi and other major models. Cross-platform.
+- **High Compatibility**: Supports Claude / Gemini / Kimi / MiniMax and other major models. Cross-platform.
 
 
 ## 🧬 Self-Evolution Mechanism
@@ -199,7 +199,7 @@ MIT License — see [LICENSE](LICENSE)
 - **自我进化**: 每次任务自动沉淀 Skill，能力随使用持续增长，形成专属技能树
 - **极简架构**: ~3,300 行核心代码，Agent Loop 仅 92 行，无复杂依赖，部署零负担
 - **强执行力**: 注入真实浏览器（保留登录态），7 个原子工具直接接管系统
-- **高兼容性**: 支持 Claude / Gemini / Kimi 等主流模型，跨平台运行 
+- **高兼容性**: 支持 Claude / Gemini / Kimi / MiniMax 等主流模型，跨平台运行
 
 ## 🧬 自我进化机制
 

llmcore.py

@@ -20,7 +20,7 @@ def compress_history_tags(messages, keep_recent=10, max_len=800):
     compress_history_tags._cd = getattr(compress_history_tags, '_cd', 0) + 1
     if compress_history_tags._cd % 5 != 0: return messages
     _before = sum(len(json.dumps(m, ensure_ascii=False)) for m in messages)
-    _pats = {tag: re.compile(rf'(<{tag}>)([\s\S]*?)(</{tag}>)') for tag in ('thinking', 'tool_use', 'tool_result')}
+    _pats = {tag: re.compile(rf'(<{tag}>)([\s\S]*?)(</{tag}>)') for tag in ('thinking', 'think', 'tool_use', 'tool_result')}
     def _trunc(text):
         for pat in _pats.values(): text = pat.sub(lambda m: m.group(1) + m.group(2)[:max_len] + '...' + m.group(3) if len(m.group(2)) > max_len else m.group(0), text)
         return text
@@ -225,7 +225,9 @@ def _openai_stream(api_base, api_key, messages, model, api_mode='chat_completion
                    temperature=0.5, max_tokens=None, tools=None, reasoning_effort=None,
                    max_retries=0, connect_timeout=10, read_timeout=300, proxies=None):
     """Shared OpenAI-compatible streaming request with retry. Yields text chunks, returns list[content_block]."""
-    if 'kimi' in model.lower() or 'moonshot' in model.lower(): temperature = 1.0
+    ml = model.lower()
+    if 'kimi' in ml or 'moonshot' in ml: temperature = 1.0
+    elif 'minimax' in ml: temperature = max(0.01, min(temperature, 1.0))  # MiniMax requires temp in (0, 1]
     headers = {"Authorization": f"Bearer {api_key}", "Content-Type": "application/json", "Accept": "text/event-stream"}
     if api_mode == "responses":
         url = auto_make_url(api_base, "responses")
@@ -345,7 +347,9 @@ class ClaudeSession:
         return result[::-1] or raw_msgs[-2:]
     def raw_ask(self, messages, model=None, temperature=0.5, max_tokens=6144):
         model = model or self.default_model
-        if 'kimi' in model.lower() or 'moonshot' in model.lower(): temperature = 1.0  # kimi/moonshot only accepts temp 1.0
+        ml = model.lower()
+        if 'kimi' in ml or 'moonshot' in ml: temperature = 1.0  # kimi/moonshot only accepts temp 1.0
+        elif 'minimax' in ml: temperature = max(0.01, min(temperature, 1.0))  # MiniMax requires temp in (0, 1]
         headers = {"x-api-key": self.api_key, "Content-Type": "application/json", "anthropic-version": "2023-06-01", "anthropic-beta": "prompt-caching-2024-07-31"}
         payload = {"model": model, "messages": messages, "temperature": temperature, "max_tokens": max_tokens, "stream": True}
         if self.system: payload["system"] = [{"type": "text", "text": self.system, "cache_control": {"type": "persistent"}}]
@@ -517,7 +521,7 @@ class NativeOAISession:
                 tool_calls = [MockToolCall(b["name"], b.get("input", {}), id=b.get("id", "")) for b in raw if b.get("type") == "tool_use"]
                 content = content[:idx].strip()
             except: pass
-        think_pattern = r"<thinking>(.*?)</thinking>"; thinking = ''
+        think_pattern = r"<think(?:ing)?>(.*?)</think(?:ing)?>"; thinking = ''
         think_match = re.search(think_pattern, content, re.DOTALL)
         if think_match:
             thinking = think_match.group(1).strip()
@@ -733,7 +737,7 @@ class ToolClient:
 
     def _parse_mixed_response(self, text):
         remaining_text = text; thinking = ''
-        think_pattern = r"<thinking>(.*?)</thinking>"
+        think_pattern = r"<think(?:ing)?>(.*?)</think(?:ing)?>"
         think_match = re.search(think_pattern, text, re.DOTALL)
         
         if think_match:
@@ -884,10 +888,10 @@ class NativeToolClient:
         if resp:
             _write_llm_log('Response', resp.raw)
             text = resp.content
-            think_match = re.search(r'<thinking>(.*?)</thinking>', text, re.DOTALL)
+            think_match = re.search(r'<think(?:ing)?>(.*?)</think(?:ing)?>', text, re.DOTALL)
             if think_match:
                 resp.thinking = think_match.group(1).strip()
-                text = re.sub(r'<thinking>.*?</thinking>', '', text, flags=re.DOTALL)
+                text = re.sub(r'<think(?:ing)?>.*?</think(?:ing)?>', '', text, flags=re.DOTALL)
             resp.content = text.strip()
         if resp and hasattr(resp, 'tool_calls') and resp.tool_calls and isinstance(self.backend, NativeClaudeSession):
             self._pending_tool_ids = [tc.id for tc in resp.tool_calls]

mykey_template.py

@@ -68,6 +68,16 @@ native_oai_config = {
 # key命名含 'sider' 触发 SiderLLMSession（需安装 sider_ai_api 包）
 #sider_cookie = 'token=Bearer%20eyJhbGciOiJIUz...'
 
+# ── MiniMax (OpenAI-compatible) ─────────────────────────────────────────────────
+# MiniMax 使用 OpenAI 兼容接口，key命名含 'oai' 即可
+# 温度自动修正为 (0, 1]，支持 M2.7 / M2.5 全系列，204K 上下文
+# oai_minimax_config = {
+#     'apikey': 'eyJh...',                        # MiniMax API Key
+#     'apibase': 'https://api.minimax.io/v1',
+#     'model': 'MiniMax-M2.7',                    # MiniMax-M2.7-highspeed / MiniMax-M2.5 等
+#     'context_win': 50000,                       # M2.7 支持 204K context
+# }
+
 
 # If you need them
 # tg_bot_token = '84102K2gYZ...'

tests/conftest.py

@@ -0,0 +1,17 @@
+"""Conftest to allow importing llmcore without mykey.py/mykey.json."""
+import os
+import json
+import sys
+
+# Create a minimal mykey.json so llmcore can be imported in test environments
+_repo_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+_mykey_path = os.path.join(_repo_dir, 'mykey.json')
+if not os.path.exists(_mykey_path) and not os.path.exists(os.path.join(_repo_dir, 'mykey.py')):
+    with open(_mykey_path, 'w') as f:
+        json.dump({}, f)
+    import atexit
+    atexit.register(lambda: os.path.exists(_mykey_path) and os.unlink(_mykey_path))
+
+# Create temp/ dir needed by _write_llm_log
+_temp_dir = os.path.join(_repo_dir, 'temp')
+os.makedirs(_temp_dir, exist_ok=True)

tests/test_minimax.py

@@ -0,0 +1,290 @@
+"""Unit tests for MiniMax provider support in llmcore.py."""
+import json
+import re
+import sys
+import os
+import unittest
+from unittest.mock import patch, MagicMock
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+
+class TestMiniMaxTemperatureClamping(unittest.TestCase):
+    """Test MiniMax temperature clamping in _openai_stream."""
+
+    def _make_stream_call(self, model, temperature):
+        """Capture the payload sent by _openai_stream."""
+        from llmcore import _openai_stream
+
+        captured = {}
+
+        def fake_post(url, headers=None, json=None, stream=None, timeout=None, proxies=None):
+            captured['payload'] = json
+            captured['url'] = url
+            resp = MagicMock()
+            resp.status_code = 200
+            resp.iter_lines.return_value = iter([b'data: [DONE]'])
+            resp.__enter__ = lambda s: s
+            resp.__exit__ = MagicMock(return_value=False)
+            return resp
+
+        with patch('llmcore.requests.post', side_effect=fake_post):
+            gen = _openai_stream(
+                'https://api.minimax.io/v1', 'test-key', [{"role": "user", "content": "hi"}],
+                model, temperature=temperature
+            )
+            # Drain the generator
+            for _ in gen:
+                pass
+
+        return captured.get('payload', {})
+
+    def test_minimax_temp_zero_clamped(self):
+        """MiniMax rejects temperature=0, should be clamped to 0.01."""
+        payload = self._make_stream_call('MiniMax-M2.7', 0.0)
+        self.assertAlmostEqual(payload['temperature'], 0.01)
+
+    def test_minimax_temp_negative_clamped(self):
+        """Negative temperature should be clamped to 0.01."""
+        payload = self._make_stream_call('MiniMax-M2.5', -0.5)
+        self.assertAlmostEqual(payload['temperature'], 0.01)
+
+    def test_minimax_temp_normal_preserved(self):
+        """Normal temperature (0 < t <= 1) should be preserved."""
+        payload = self._make_stream_call('MiniMax-M2.7', 0.5)
+        self.assertAlmostEqual(payload['temperature'], 0.5)
+
+    def test_minimax_temp_one_preserved(self):
+        """Temperature=1.0 should be preserved."""
+        payload = self._make_stream_call('MiniMax-M2.7-highspeed', 1.0)
+        self.assertAlmostEqual(payload['temperature'], 1.0)
+
+    def test_minimax_temp_above_one_clamped(self):
+        """Temperature > 1.0 should be clamped to 1.0."""
+        payload = self._make_stream_call('MiniMax-M2.7', 1.5)
+        self.assertAlmostEqual(payload['temperature'], 1.0)
+
+    def test_minimax_case_insensitive(self):
+        """Model name matching should be case-insensitive."""
+        payload = self._make_stream_call('minimax-m2.7', 0.0)
+        self.assertAlmostEqual(payload['temperature'], 0.01)
+
+    def test_non_minimax_temp_zero_unchanged(self):
+        """Non-MiniMax models should not have temperature clamped."""
+        payload = self._make_stream_call('gpt-4o', 0.0)
+        self.assertAlmostEqual(payload['temperature'], 0.0)
+
+    def test_kimi_temp_still_forced(self):
+        """Kimi/Moonshot temp override should still work."""
+        payload = self._make_stream_call('kimi-2.0', 0.5)
+        self.assertAlmostEqual(payload['temperature'], 1.0)
+
+
+class TestMiniMaxThinkTagHandling(unittest.TestCase):
+    """Test <think>...</think> tag stripping for MiniMax M2.7 responses."""
+
+    def test_think_tag_stripped_from_response(self):
+        """<think> tags (used by MiniMax M2.7) should be stripped from content."""
+        from llmcore import ToolClient, LLMSession
+
+        mock_cfg = {
+            'apikey': 'test', 'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7',
+        }
+        with patch('llmcore._load_mykeys', return_value={}):
+            session = LLMSession(mock_cfg)
+
+        client = ToolClient(session)
+        text = '<think>Let me reason about this task.</think>\n\nHere is the answer.'
+        result = client._parse_mixed_response(text)
+        self.assertEqual(result.thinking, 'Let me reason about this task.')
+        self.assertEqual(result.content, 'Here is the answer.')
+
+    def test_thinking_tag_still_works(self):
+        """<thinking> tags (used by Claude) should still work."""
+        from llmcore import ToolClient, LLMSession
+
+        mock_cfg = {
+            'apikey': 'test', 'apibase': 'https://api.anthropic.com',
+            'model': 'claude-sonnet-4-20250514',
+        }
+        with patch('llmcore._load_mykeys', return_value={}):
+            session = LLMSession(mock_cfg)
+
+        client = ToolClient(session)
+        text = '<thinking>Let me analyze this.</thinking>\n\nThe result is 42.'
+        result = client._parse_mixed_response(text)
+        self.assertEqual(result.thinking, 'Let me analyze this.')
+        self.assertEqual(result.content, 'The result is 42.')
+
+    def test_think_tag_with_tool_use(self):
+        """<think> tags should be separated from tool_use blocks."""
+        from llmcore import ToolClient, LLMSession
+
+        mock_cfg = {
+            'apikey': 'test', 'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7',
+        }
+        with patch('llmcore._load_mykeys', return_value={}):
+            session = LLMSession(mock_cfg)
+
+        client = ToolClient(session)
+        text = '<think>I need to read the file first.</think>\n\n<summary>Reading config</summary>\n\n<tool_use>\n{"name": "file_read", "arguments": {"path": "/tmp/test.txt"}}\n</tool_use>'
+        result = client._parse_mixed_response(text)
+        self.assertEqual(result.thinking, 'I need to read the file first.')
+        self.assertTrue(len(result.tool_calls) > 0)
+        self.assertEqual(result.tool_calls[0].function.name, 'file_read')
+
+
+class TestMiniMaxCompressHistoryTags(unittest.TestCase):
+    """Test that <think> tags are compressed in history like <thinking> tags."""
+
+    def test_think_tag_compressed_in_old_messages(self):
+        """<think> tags in old messages should be truncated."""
+        from llmcore import compress_history_tags
+
+        long_think = "A" * 2000
+        messages = [
+            {"role": "assistant", "prompt": f"<think>{long_think}</think>\nShort answer."},
+            {"role": "user", "prompt": "Follow up"},
+        ] + [{"role": "user", "prompt": f"msg{i}"} for i in range(12)]
+
+        # Force compression (counter divisible by 5)
+        compress_history_tags._cd = 4
+        result = compress_history_tags(messages, keep_recent=10, max_len=800)
+        # The first message's <think> content should be truncated
+        first_content = result[0]["prompt"]
+        self.assertIn("<think>", first_content)
+        self.assertIn("...", first_content)
+        self.assertLess(len(first_content), len(f"<think>{long_think}</think>\nShort answer."))
+
+
+class TestMiniMaxAutoMakeUrl(unittest.TestCase):
+    """Test URL construction for MiniMax API base."""
+
+    def test_minimax_base_url(self):
+        from llmcore import auto_make_url
+        url = auto_make_url('https://api.minimax.io/v1', 'chat/completions')
+        self.assertEqual(url, 'https://api.minimax.io/v1/chat/completions')
+
+    def test_minimax_base_url_no_version(self):
+        from llmcore import auto_make_url
+        url = auto_make_url('https://api.minimax.io', 'chat/completions')
+        self.assertEqual(url, 'https://api.minimax.io/v1/chat/completions')
+
+    def test_minimax_full_url_preserved(self):
+        from llmcore import auto_make_url
+        url = auto_make_url('https://api.minimax.io/v1/chat/completions$', 'chat/completions')
+        self.assertEqual(url, 'https://api.minimax.io/v1/chat/completions')
+
+
+class TestMiniMaxNativeOAISessionThinkTag(unittest.TestCase):
+    """Test <think> tag handling in NativeOAISession."""
+
+    def test_think_tag_extracted_in_native_oai(self):
+        """NativeOAISession.ask should extract <think> tags from MiniMax M2.7 responses."""
+        from llmcore import NativeOAISession
+
+        cfg = {
+            'apikey': 'test-key',
+            'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7',
+        }
+        session = NativeOAISession(cfg)
+
+        # Mock the raw_ask to return content with <think> tag via generator
+        def mock_raw_ask(messages, tools=None, system=None, model=None, temperature=0.5, max_tokens=6144, **kw):
+            content_text = "<think>Planning the approach.</think>\n\nHere is the result."
+            yield content_text
+            return [{"type": "text", "text": content_text}]
+
+        session.raw_ask = mock_raw_ask
+
+        msg = {"role": "user", "content": [{"type": "text", "text": "test"}]}
+        gen = session.ask(msg)
+        # Drain generator
+        try:
+            while True:
+                next(gen)
+        except StopIteration as e:
+            resp = e.value
+
+        self.assertEqual(resp.thinking, 'Planning the approach.')
+        self.assertNotIn('<think>', resp.content)
+        self.assertIn('Here is the result.', resp.content)
+
+
+class TestMiniMaxLLMSessionConfig(unittest.TestCase):
+    """Test LLMSession configuration with MiniMax settings."""
+
+    def test_llm_session_init_with_minimax(self):
+        """LLMSession should initialize correctly with MiniMax config."""
+        from llmcore import LLMSession
+
+        cfg = {
+            'apikey': 'eyJhbGciOiJSUzI1NiIsInR5cCI6IkpXVCJ9',
+            'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7',
+            'context_win': 50000,
+            'max_retries': 2,
+            'connect_timeout': 10,
+            'read_timeout': 120,
+        }
+        session = LLMSession(cfg)
+        self.assertEqual(session.default_model, 'MiniMax-M2.7')
+        self.assertEqual(session.api_base, 'https://api.minimax.io/v1')
+        self.assertEqual(session.context_win, 50000)
+        self.assertEqual(session.max_retries, 2)
+
+    def test_llm_session_minimax_highspeed(self):
+        """LLMSession should work with MiniMax-M2.7-highspeed model."""
+        from llmcore import LLMSession
+
+        cfg = {
+            'apikey': 'test-key',
+            'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7-highspeed',
+        }
+        session = LLMSession(cfg)
+        self.assertEqual(session.default_model, 'MiniMax-M2.7-highspeed')
+
+
+class TestMiniMaxNativeToolClientThinkTag(unittest.TestCase):
+    """Test <think> tag handling in NativeToolClient.chat."""
+
+    def test_native_tool_client_think_tag(self):
+        """NativeToolClient should extract <think> tags from MiniMax responses."""
+        from llmcore import NativeToolClient, NativeOAISession, MockResponse
+
+        cfg = {
+            'apikey': 'test-key',
+            'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7',
+        }
+        session = NativeOAISession(cfg)
+        client = NativeToolClient(session)
+
+        # Mock the backend.ask to yield chunks and return a MockResponse with think tags
+        def mock_ask(msg, tools=None, model=None):
+            text = "<think>Analyzing the request.</think>\n\nResult: success"
+            yield text
+            return MockResponse('', text, [], text)
+
+        session.ask = mock_ask
+
+        messages = [{"role": "user", "content": "test query"}]
+        gen = client.chat(messages)
+        resp = None
+        try:
+            while True:
+                next(gen)
+        except StopIteration as e:
+            resp = e.value
+
+        self.assertIsNotNone(resp)
+        self.assertEqual(resp.thinking, 'Analyzing the request.')
+        self.assertEqual(resp.content, 'Result: success')
+
+
+if __name__ == '__main__':
+    unittest.main()

tests/test_minimax_integration.py

@@ -0,0 +1,229 @@
+"""Integration tests for MiniMax provider support.
+
+These tests verify end-to-end MiniMax integration by mocking the HTTP layer
+while exercising the full session → stream → parse pipeline.
+
+To run against a real MiniMax API, set MINIMAX_API_KEY in your environment.
+"""
+import json
+import os
+import sys
+import unittest
+from unittest.mock import patch, MagicMock
+from io import BytesIO
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+
+def _make_sse_response(chunks, finish_reason="stop"):
+    """Build a mock SSE HTTP response from a list of text chunks."""
+    lines = []
+    for chunk in chunks:
+        data = {
+            "choices": [{"delta": {"content": chunk}}],
+        }
+        lines.append(f"data: {json.dumps(data)}".encode())
+    # Final chunk with usage
+    usage_data = {
+        "choices": [{"delta": {}}],
+        "usage": {"prompt_tokens": 100, "completion_tokens": 50,
+                  "prompt_tokens_details": {"cached_tokens": 0}},
+    }
+    lines.append(f"data: {json.dumps(usage_data)}".encode())
+    lines.append(b"data: [DONE]")
+    return lines
+
+
+class TestMiniMaxEndToEnd(unittest.TestCase):
+    """End-to-end integration test: LLMSession + ToolClient + MiniMax streaming."""
+
+    def test_full_pipeline_with_think_tag(self):
+        """Full pipeline: LLMSession → _openai_stream → ToolClient parse with <think> tag."""
+        from llmcore import LLMSession, ToolClient
+
+        cfg = {
+            'apikey': 'test-integration-key',
+            'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7',
+            'context_win': 50000,
+        }
+        session = LLMSession(cfg)
+        client = ToolClient(session)
+
+        sse_lines = _make_sse_response([
+            "<think>Let me analyze this task step by step.\n",
+            "1. First, I need to understand the request.\n",
+            "2. Then, execute the appropriate action.</think>\n\n",
+            "<summary>Analyzing user request</summary>\n\n",
+            "I'll help you with that task.",
+        ])
+
+        mock_resp = MagicMock()
+        mock_resp.status_code = 200
+        mock_resp.iter_lines.return_value = iter(sse_lines)
+        mock_resp.__enter__ = lambda s: s
+        mock_resp.__exit__ = MagicMock(return_value=False)
+
+        with patch('llmcore.requests.post', return_value=mock_resp):
+            messages = [
+                {"role": "system", "content": "You are a helpful assistant."},
+                {"role": "user", "content": "Help me read a file."},
+            ]
+            gen = client.chat(messages=messages, tools=None)
+            chunks = []
+            try:
+                while True:
+                    chunks.append(next(gen))
+            except StopIteration as e:
+                response = e.value
+
+        self.assertIsNotNone(response)
+        self.assertIn("analyze this task", response.thinking)
+        self.assertIn("help you with that task", response.content)
+        # <think> should be stripped from content
+        self.assertNotIn("<think>", response.content)
+
+    def test_full_pipeline_with_tool_call(self):
+        """Full pipeline: MiniMax response with tool_use block."""
+        from llmcore import LLMSession, ToolClient
+
+        cfg = {
+            'apikey': 'test-key',
+            'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7',
+        }
+        session = LLMSession(cfg)
+        client = ToolClient(session)
+
+        sse_lines = _make_sse_response([
+            "<think>I need to read the config file.</think>\n\n",
+            "<summary>Reading config</summary>\n\n",
+            '<tool_use>\n{"name": "file_read", "arguments": {"path": "/etc/config.json"}}\n</tool_use>',
+        ])
+
+        mock_resp = MagicMock()
+        mock_resp.status_code = 200
+        mock_resp.iter_lines.return_value = iter(sse_lines)
+        mock_resp.__enter__ = lambda s: s
+        mock_resp.__exit__ = MagicMock(return_value=False)
+
+        with patch('llmcore.requests.post', return_value=mock_resp):
+            messages = [{"role": "user", "content": "Read the config file."}]
+            gen = client.chat(messages=messages, tools=None)
+            try:
+                while True:
+                    next(gen)
+            except StopIteration as e:
+                response = e.value
+
+        self.assertEqual(response.thinking, "I need to read the config file.")
+        self.assertEqual(len(response.tool_calls), 1)
+        self.assertEqual(response.tool_calls[0].function.name, "file_read")
+
+    def test_temperature_enforced_in_request(self):
+        """Verify the actual HTTP request has clamped temperature for MiniMax."""
+        from llmcore import LLMSession
+
+        cfg = {
+            'apikey': 'test-key',
+            'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7',
+        }
+        session = LLMSession(cfg)
+        captured = {}
+
+        def capture_post(url, headers=None, json=None, stream=None, timeout=None, proxies=None):
+            captured['json'] = json
+            captured['url'] = url
+            resp = MagicMock()
+            resp.status_code = 200
+            resp.iter_lines.return_value = iter([b'data: [DONE]'])
+            resp.__enter__ = lambda s: s
+            resp.__exit__ = MagicMock(return_value=False)
+            return resp
+
+        with patch('llmcore.requests.post', side_effect=capture_post):
+            session.raw_msgs = [{"role": "user", "prompt": "test", "image": None}]
+            gen = session.raw_ask(
+                [{"role": "user", "content": "test"}],
+                model='MiniMax-M2.7',
+                temperature=0.0,
+            )
+            for _ in gen:
+                pass
+
+        self.assertAlmostEqual(captured['json']['temperature'], 0.01)
+        self.assertIn('api.minimax.io', captured['url'])
+
+
+@unittest.skipUnless(
+    os.environ.get('MINIMAX_API_KEY'),
+    'Set MINIMAX_API_KEY to run live integration tests'
+)
+class TestMiniMaxLive(unittest.TestCase):
+    """Live integration tests against MiniMax API (requires MINIMAX_API_KEY)."""
+
+    def test_live_chat_completion(self):
+        """Send a real chat completion to MiniMax API."""
+        from llmcore import LLMSession
+
+        cfg = {
+            'apikey': os.environ['MINIMAX_API_KEY'],
+            'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7-highspeed',
+        }
+        session = LLMSession(cfg)
+
+        messages = [{"role": "user", "content": "Say 'hello' and nothing else."}]
+        gen = session.raw_ask(messages, temperature=0.1)
+        text = ''
+        for chunk in gen:
+            text += chunk
+
+        self.assertFalse(text.startswith('Error:'), f"API returned error: {text}")
+        self.assertIn('hello', text.lower())
+
+    def test_live_tool_client_pipeline(self):
+        """Full ToolClient pipeline with real MiniMax API."""
+        from llmcore import LLMSession, ToolClient
+
+        cfg = {
+            'apikey': os.environ['MINIMAX_API_KEY'],
+            'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7-highspeed',
+            'context_win': 50000,
+        }
+        session = LLMSession(cfg)
+        client = ToolClient(session)
+
+        messages = [{"role": "user", "content": "What is 2+2? Reply with just the number."}]
+        gen = client.chat(messages=messages, tools=None)
+        try:
+            while True:
+                next(gen)
+        except StopIteration as e:
+            response = e.value
+
+        self.assertIn('4', response.content)
+
+    def test_live_streaming_chunks(self):
+        """Verify streaming works with MiniMax API."""
+        from llmcore import LLMSession
+
+        cfg = {
+            'apikey': os.environ['MINIMAX_API_KEY'],
+            'apibase': 'https://api.minimax.io/v1',
+            'model': 'MiniMax-M2.7-highspeed',
+        }
+        session = LLMSession(cfg)
+
+        session.raw_msgs.append({"role": "user", "prompt": "Count from 1 to 5.", "image": None})
+        result = session.ask("Count from 1 to 5.", stream=False)
+        self.assertFalse(result.startswith('Error:'), f"API returned error: {result}")
+        # Should contain at least some numbers
+        for n in ['1', '2', '3']:
+            self.assertIn(n, result)
+
+
+if __name__ == '__main__':
+    unittest.main()