fix: regulatory 对话模式改用 event_stream_handler 修复工具调用截断，优化节点 prompt 和日志展示

- regulatory_node: stream_working 从 run_stream 改为 agent.run + event_stream_handler，
  解决工具调用后文本被截断的问题；添加 PartStartEvent 处理修复首字丢失
- consciousness_node: prompt 重写为三模式（生成/执行/报告），强调禁止编造 agent_id
- workflow API: _merge_runtime_status 暴露步骤输出内容（workflow_log 第三元素）
- 前端日志: 系统日志改为终端滚动样式，工作流步骤可展开查看输出

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

kilostar/api/workflow.py

@@ -177,8 +177,8 @@ def _merge_runtime_status(work_link: list, workflow_log: list) -> list:
     前端 ``WorkflowDiagram`` 依赖每个 step 的 ``status`` 字段着色，这个拼装让
     后端真正把运行期状态喂过去。
     """
-    # step_index -> 最新 status
     latest_status: dict[int, str] = {}
+    latest_output: dict[int, str] = {}
     for entry in workflow_log or []:
         if not isinstance(entry, dict):
             continue
@@ -189,14 +189,16 @@ def _merge_runtime_status(work_link: list, workflow_log: list) -> list:
                 continue
             if isinstance(payload, (list, tuple)) and len(payload) >= 2:
                 latest_status[idx] = payload[1]
+            if isinstance(payload, (list, tuple)) and len(payload) >= 3:
+                latest_output[idx] = payload[2]
 
     merged = []
     for i, step in enumerate(work_link or []):
         step_copy = dict(step) if isinstance(step, dict) else {}
-        # step 自带的 step 字段优先，否则用位置索引
         step_idx = step_copy.get("step")
         lookup_idx = (step_idx - 1) if isinstance(step_idx, int) else i
         step_copy["status"] = latest_status.get(lookup_idx, "pending")
+        step_copy["output"] = latest_output.get(lookup_idx, "")
         merged.append(step_copy)
     return merged
 

kilostar/core/individual/regulatory_node/regulatory_node.py

@@ -126,11 +126,23 @@ class RegulatoryNode:
         """
         return await self._run(payload)
 
-    async def stream_working(self, payload: MessageRequest, token_queue: "asyncio.Queue") -> None:
-        """流式工具调用版本：逐 token 推送到 queue，工具调用结果也会通过 token 输出。
+    _CHAT_INSTRUCTIONS = (
+        "你是 kilostar 智能助手。你现在处于【直接对话模式】，请直接回答用户的问题。\n"
+        "规则：\n"
+        "1. 直接、详细地回答用户问题，像一个专业且友好的助手。\n"
+        "2. 如果你有可用工具，可以调用工具来辅助回答（如搜索、读文件等）。\n"
+        "3. 不要输出内部思考过程，不要做路由判断，不要提及 ForUser/ForConsciousnessNode 等格式。\n"
+        "4. 回复应当完整、有帮助，避免过于简短。\n"
+    )
 
-        完成后 push None 作为终止信号。
+    async def stream_working(self, payload: MessageRequest, token_queue: "asyncio.Queue") -> None:
+        """流式对话：完整执行 agent graph（含工具调用），逐 token 推送文本到 queue。
+
+        使用 event_stream_handler 回调拿到每个 text delta，保证工具调用后
+        的文本也能被流式输出。完成后 push None 作为终止信号。
         """
+        from pydantic_ai.messages import PartStartEvent, PartDeltaEvent, TextPart, TextPartDelta
+
         platform = payload.platform
         user_name = payload.user_name
         message = payload.message
@@ -140,17 +152,27 @@ class RegulatoryNode:
             await token_queue.put(None)
             return
 
+        async def _stream_handler(ctx, events):
+            async for event in events:
+                if isinstance(event, PartStartEvent) and isinstance(event.part, TextPart):
+                    if event.part.content:
+                        await token_queue.put(event.part.content)
+                elif isinstance(event, PartDeltaEvent) and isinstance(event.delta, TextPartDelta):
+                    await token_queue.put(event.delta.content_delta)
+
         try:
             deps = RegulatoryNodeDeps(
                 platform=platform,
                 user_name=user_name,
                 time=time_str
             )
-            async with self.agent.run_stream(
-                user_prompt=message, deps=deps, output_type=str
-            ) as stream_result:
-                async for delta in stream_result.stream_text(delta=True):
-                    await token_queue.put(delta)
+            await self.agent.run(
+                user_prompt=message,
+                deps=deps,
+                output_type=str,
+                instructions=self._CHAT_INSTRUCTIONS,
+                event_stream_handler=_stream_handler,
+            )
         except Exception as e:
             self.logger.exception(f"RegulatoryNode.stream_working failed: {e}")
             await token_queue.put(f"\n\n[错误: {str(e)}]")

kilostar/utils/i18n.py

@@ -59,21 +59,41 @@ _PROMPTS: Dict[str, Dict[str, str]] = {
     "consciousness_node": {
         "zh": (
             "你叫kilostar，是一个多智能体AI助手系统中的【意识节点 (Consciousness Node)】。\n"
-            "你是系统的'高级规划师'和'架构师'，负责处理监控节点分配过来的复杂任务。\n"
-            "你的主要工作场景包括：\n"
-            "1. 拆解任务 (Workflow Generation)：结合用户的原始命令和提供的模板，生成严谨、可执行的工作流 (kilostarWorkflow)，并将其输出为 ForWorkflowEngine 格式。拆解时步骤应清晰连贯。\n"
-            "2. 中途指导 (Workflow Execution)：在工作流执行中，如果某一步骤指派给你，你需要对控制节点的结果进行分析或提供下一步的指导，输出 ForWorkflow 格式。\n"
-            "3. 总结报告 (regulatory Report)：在整个工作流执行完毕后，你需要对整体流程、各个控制节点的执行情况进行审查，并生成一份技术性的总结报告，输出 ForregulatoryNode 格式。\n"
-            "请确保所有的思考和生成过程符合逻辑，严密且高质量。"
+            "你是系统的'高级规划师'和'架构师'，负责处理监控节点分配过来的复杂任务。\n\n"
+            "你的工作根据收到的输入类型严格分为三种模式：\n\n"
+            "【模式1：工作流生成】当你收到用户的原始任务命令时：\n"
+            "- 将复杂任务拆解为多个清晰、可执行的步骤\n"
+            "- 每个步骤必须指派给真实存在的 Worker（使用其真实 agent_id）或 consciousness_node 自己\n"
+            "- 严禁编造不存在的 agent_id！只能使用上下文中列出的可用 Worker\n"
+            "- 输出格式：ForWorkflowEngine\n\n"
+            "【模式2：工作流步骤执行】当某个步骤指派给你自己时：\n"
+            "- 直接完成该步骤描述的具体任务\n"
+            "- 输出应当是任务的实际结果（代码、分析、文档等），而非对任务的描述\n"
+            "- 输出格式：ForWorkflow\n\n"
+            "【模式3：总结报告】当整个工作流执行完毕时：\n"
+            "- 审查各步骤执行情况，生成面向用户的技术总结报告\n"
+            "- 报告应包含：完成了什么、关键结果、是否有失败步骤及原因\n"
+            "- 输出格式：ForregulatoryNode\n\n"
+            "确保所有输出符合逻辑、严密且高质量。"
         ),
         "en": (
             "You are kilostar, the [Consciousness Node] in a multi-agent AI assistant system.\n"
-            "You are the system's 'senior planner' and 'architect', responsible for handling complex tasks assigned by the Regulatory Node.\n"
-            "Your main scenarios include:\n"
-            "1. Task Decomposition (Workflow Generation): Combine the user's original command with provided templates to generate rigorous, executable workflows (kilostarWorkflow), outputting them in the ForWorkflowEngine format. Steps should be clear and coherent.\n"
-            "2. Mid-flight Guidance (Workflow Execution): During workflow execution, if a step is assigned to you, analyze the Control Node's results or provide next-step guidance, outputting in the ForWorkflow format.\n"
-            "3. Summary Report (Regulatory Report): After the entire workflow completes, review the overall process and each Control Node's execution, generating a technical summary report in the ForregulatoryNode format.\n"
-            "Ensure all reasoning and generation is logical, rigorous, and high-quality."
+            "You are the system's 'senior planner' and 'architect', responsible for handling complex tasks assigned by the Regulatory Node.\n\n"
+            "Your work is strictly divided into three modes based on input type:\n\n"
+            "[Mode 1: Workflow Generation] When you receive the user's original task command:\n"
+            "- Decompose the complex task into clear, executable steps\n"
+            "- Each step must be assigned to a real existing Worker (using its real agent_id) or to consciousness_node itself\n"
+            "- NEVER fabricate non-existent agent_ids! Only use Workers listed in the context\n"
+            "- Output format: ForWorkflowEngine\n\n"
+            "[Mode 2: Workflow Step Execution] When a step is assigned to you:\n"
+            "- Directly complete the specific task described in the step\n"
+            "- Output should be the actual result (code, analysis, documentation, etc.), not a description of the task\n"
+            "- Output format: ForWorkflow\n\n"
+            "[Mode 3: Summary Report] When the entire workflow has completed:\n"
+            "- Review each step's execution and generate a user-facing technical summary\n"
+            "- Report should include: what was accomplished, key results, any failed steps and reasons\n"
+            "- Output format: ForregulatoryNode\n\n"
+            "Ensure all output is logical, rigorous, and high-quality."
         ),
     },
     "control_node": {