fix: switch default model to qwen3-coder-plus for better function calling

DeepSeek-chat has limited/partial function calling support. Qwen3-coder-plus
(DashScope) has robust OpenAI-compatible function calling.

Also added tool usage instructions to system prompt and enhanced logging
to trace tool propagation through the pipeline.

configs/llm_config.yaml

@@ -21,13 +21,13 @@ providers:
         cost_per_1k_output: 0.00028
 
 model_aliases:
-  default: "deepseek/deepseek-chat"
-  fast: "deepseek/deepseek-chat"
-  powerful: "deepseek/deepseek-chat"
+  default: "openai/qwen3-coder-plus"
+  fast: "openai/qwen3-coder-plus"
+  powerful: "openai/qwen3-coder-plus"
 
 fallbacks:
-  deepseek/deepseek-chat:
-    - "openai/qwen3-coder-plus"
+  openai/qwen3-coder-plus:
+    - "deepseek/deepseek-chat"
 
 # 上下文压缩配置 — 长会话自动压缩历史消息，保持 Token 在预算内
 # GEO Pipeline 启用后，工具输出（搜索结果、网页抓取等）会自动压缩

src/agentkit/chat/skill_routing.py

@@ -213,10 +213,17 @@ async def resolve_skill_routing(
     # Append available tools to system prompt so LLM knows what it can call
     if result.tools:
         tools_desc = _build_tools_description(result.tools)
+        tool_instruction = (
+            "\n\n## Tool Usage\n"
+            "You have access to the following tools. When you need to use a tool, "
+            "respond with a tool call in the format specified by the system.\n"
+            "Never make up information or guess answers when you can use a tool to find the answer.\n"
+            "Always prefer using tools over guessing.\n"
+        )
         if result.system_prompt:
-            result.system_prompt += f"\n\n## Available Tools\n{tools_desc}"
+            result.system_prompt += f"{tool_instruction}\n## Available Tools\n{tools_desc}"
         else:
-            result.system_prompt = f"## Available Tools\n{tools_desc}"
+            result.system_prompt = f"{tool_instruction}\n## Available Tools\n{tools_desc}"
 
     return result
 

src/agentkit/llm/providers/openai.py

@@ -184,8 +184,10 @@ class OpenAICompatibleProvider(LLMProvider):
         if request.tools:
             payload["tools"] = request.tools
             payload["tool_choice"] = request.tool_choice
-
-        logger.debug(f"Stream request to {url}: model={request.model}, messages={len(request.messages)}, tools={len(request.tools or [])}")
+            tool_names = [t.get("function", {}).get("name", "?") for t in request.tools]
+            logger.info(f"OpenAIProvider stream: model={request.model}, tools={len(request.tools)} {tool_names}")
+        else:
+            logger.info(f"OpenAIProvider stream: model={request.model}, NO tools")
 
         response_ctx = self._client.stream("POST", url, json=payload, headers=headers)
         response = await response_ctx.__aenter__()

src/agentkit/server/routes/chat.py

@@ -390,6 +390,10 @@ async def _handle_chat_message(
         session_id=session_id,
     )
 
+    # Debug: log tools that will be passed to ReActEngine
+    tool_names = [t.name for t in routing.tools]
+    logger.info(f"Chat {session_id}: resolved {len(routing.tools)} tools: {tool_names}, model={routing.model}, skill={routing.skill_name}")
+
     # Notify frontend about skill match
     if routing.matched:
         await websocket.send_json({