fix openclaw session problem

Author: binary-husky

ajet/task_rollout/async_llm_bridge.py

@@ -146,8 +146,6 @@ async def llm_chat_verl(
                 "completion_tokens": len(token_array), # type: ignore
                 "total_tokens": len(prompt_token_ids) + len(token_array), # type: ignore
             }
-            print("====----====usage", usage)
-            print("====----====finish_reason", finish_reason)
             return {
                 "role": "assistant",
                 "request_id": request_id,

ajet/utils/swarm_overwatch.py

@@ -455,6 +455,7 @@ def display_latest_llm_call(self):
         while True:
             response = httpx.post(f"{self.server_url}/replay_latest_llm_call", timeout=30.0)
             structured_response = response.json()
+            self.console.clear()
             if "input" not in structured_response or "output" not in structured_response:
                 self.console.print(f"[bold red]{structured_response}[/bold red]")
                 time.sleep(5)
@@ -464,6 +465,33 @@ def display_latest_llm_call(self):
                 output = structured_response["output"]
                 self.console.print(f"\n[bold green]Input:[/bold green]\n{input}")
                 self.console.print(f"\n[bold green]Output:[/bold green]\n{output}")
+                hide_when_more_than_n_line_break = 4
+                try:
+                    input_items = ""
+                    output_items = ""
+                    for item in input['messages']:
+                        role = item['role']
+                        content = item['content']
+                        if isinstance(content, list):
+                            content = content[0].get('text', '')
+                        if content.count('\n') >= hide_when_more_than_n_line_break:
+                            content = content.replace('\n',' ')[:200] + " ....."
+                        else:
+                            content = content.replace('\n',' ')
+                        input_items += f"[bold blue]@{role}:[/bold blue] {content}\n"
+                    for item in output['choices']:
+                        role = item['message']['role']
+                        content = item['message']['content']
+                        if content.count('\n') >= hide_when_more_than_n_line_break:
+                            content = content.replace('\n',' ')[:200] + " ....."
+                        else:
+                            content = content.replace('\n',' ')
+                        output_items += f"[bold red]@{role}:[/bold red] {content}\n"
+                    self.console.print(f"\n-------------------------------------------------------------")
+                    self.console.print(f"\n[bold green]Input Simlified:[/bold green]\n{input_items}")
+                    self.console.print(f"\n[bold green]Output Simlified:[/bold green]\n{output_items}")
+                except:
+                    pass
                 time.sleep(5)
 
     def choose_run(self) -> str:

docs/en/swarm_intro_blog_zh.md

@@ -88,6 +88,20 @@ AgentJet具备全分布式蜂群训练（Swarm Training）能力，
 
 接下来，用简单的几个case展示 AgentJet 蜂群模式的优势。
 
+## 蜂群 Agentic RL 框架核心优势
+
+| 特性 | 经典训练框架 | AgentJet 蜂群框架 |
+|------|-------------|------------------|
+| **多模型异构训练** | ❌ 所有智能体共享同一模型 | ✅ 支持多个不同规模模型同时训练 |
+| **训练推理解耦** | ❌ 采样与训练紧耦合 | ✅ Server 训练，Client 采样，完全解耦 |
+| **运行环境限制** | ❌ 受训练服务器环境限制 | ✅ Client 可在任意设备运行（笔记本/服务器） |
+| **动态节点管理** | ❌ 无法动态添加/移除节点 | ✅ 训练中随时添加/移除 Client 节点 |
+| **调试迭代速度** | ❌ 修改代码需重启训练（每次15分钟+） | ✅ 仅重启 Client（秒级），无需重载模型 |
+| **容错能力** | ❌ 外部依赖故障导致训练中断 | ✅ Client 崩溃不影响训练，自动恢复 |
+| **多任务混合训练** | ❌ 所有任务共享同一运行环境 | ✅ 不同 Client 运行不同任务环境 |
+| **Agent 框架兼容** | ❌ 需适配特定训练框架 | ✅ 兼容任何 OpenAI API 协议框架 |
+| **本地开发体验** | ❌ 必须在训练服务器上调试 | ✅ 本地 IDE 调试，连接远程训练 |
+
 ## 灵活的蜂群训练模式
 
 ### 用笔记本电脑全参训练Agentic LLM模型

tutorial/opencode_build_openclaw_agent/mock_user_request.py

@@ -15,11 +15,60 @@ def load_dataset(filepath: str = "extraversion_questions.json") -> List[Dict]:
     with open(filepath, 'r', encoding='utf-8') as f:
         return json.load(f)
 
-def send_openclaw_message(question: str) -> str:
-    """Send message via OpenClaw CLI."""
+
+def generate_agent_name() -> str:
+    """Generate a random agent name."""
+    adjectives = ["happy", "quick", "bright", "clever", "bold", "calm", "eager", "gentle"]
+    nouns = ["fox", "wolf", "bear", "eagle", "hawk", "lion", "tiger", "owl"]
+    return f"{random.choice(adjectives)}_{random.choice(nouns)}_{random.randint(1000, 9999)}"
+
+
+def create_agent(agent_name: str) -> bool:
+    """Create a new agent using OpenClaw CLI."""
+    try:
+        workspace = f"/root/.openclaw/workspace-{agent_name}"
+        result = subprocess.run(
+            ["openclaw", "agents", "add", agent_name, "--workspace", workspace, "--non-interactive"],
+            capture_output=True,
+            text=True,
+            timeout=60
+        )
+        if result.returncode == 0:
+            print(f"Created agent: {agent_name}")
+            return True
+        else:
+            print(f"Error creating agent {agent_name}: {result.stderr}")
+            return False
+    except Exception as e:
+        print(f"Error creating agent: {str(e)}")
+        return False
+
+
+def delete_agent(agent_name: str) -> bool:
+    """Delete an agent using OpenClaw CLI."""
+    try:
+        result = subprocess.run(
+            ["openclaw", "agents", "delete", agent_name, "--force"],
+            capture_output=True,
+            text=True,
+            timeout=60
+        )
+        if result.returncode == 0:
+            print(f"Deleted agent: {agent_name}")
+            return True
+        else:
+            print(f"Error deleting agent {agent_name}: {result.stderr}")
+            return False
+    except Exception as e:
+        print(f"Error deleting agent: {str(e)}")
+        return False
+
+
+def send_openclaw_message(agent_name: str, message: str) -> str:
+    """Send message via OpenClaw CLI to specific agent."""
     try:
         result = subprocess.run(
-            ["openclaw", "agent", "--agent", "main", "--message", question],
+            ["openclaw", "agent", "--agent", agent_name, "--message", message],
             capture_output=True,
             text=True,
             timeout=300
@@ -28,6 +77,7 @@ def send_openclaw_message(question: str) -> str:
     except Exception as e:
         return f"Error: {str(e)}"
 
+
 def main():
     """Main loop to send requests from dataset."""
     print("Starting OpenClaw mock user requests")
@@ -37,15 +87,33 @@ def main():
     random.shuffle(dataset)
     print(f"Loaded {len(dataset)} questions from dataset\n")
 
-    # Send requests via OpenClaw CLI
-    for i, item in enumerate(dataset):
-        question = item.get("Question", "")
-        print(f"[{i+1}/{len(dataset)}] Sending: {question[:80]}...")
+    # Process dataset in chunks of 5
+    for chunk_start in range(0, len(dataset), 5):
+        chunk = dataset[chunk_start:chunk_start + 5]
+
+        # Generate random agent name
+        agent_name = generate_agent_name()
+        print(f"\n=== Creating agent: {agent_name} ===\n")
+
+        # Create agent
+        if not create_agent(agent_name):
+            print(f"Failed to create agent, skipping chunk")
+            continue
+
+        # Send 5 messages
+        for i, item in enumerate(chunk):
+            question = item.get("Question", "")
+            print(f"[{agent_name}/{i+1}/5] Sending: {question[:80]}...")
+            response = send_openclaw_message(agent_name, question)
+            print(f"Response: {response[:200]}...\n")
+            time.sleep(2)
+
+        # Delete agent
+        delete_agent(agent_name)
+        print(f"\n=== Deleted agent: {agent_name} ===\n")
 
-        response = send_openclaw_message(question)
-        print(f"Response: {response[:200]}...\n")
+    print("\nAll agents processed successfully")
 
-        time.sleep(2)
 
 if __name__ == "__main__":
     main()