ljsysfurry commited on
Commit
80e1d19
·
verified ·
1 Parent(s): c6405fd

Upload agentframe_orchestrator.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. agentframe_orchestrator.py +333 -0
agentframe_orchestrator.py ADDED
@@ -0,0 +1,333 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ AgentFrame 编排层 (Orchestrator)
3
+ ===============================
4
+ DeepSeek 决策 ↔ OpenClaw 执行的桥接层
5
+
6
+ 核心循环:
7
+ 1. DeepSeek 分析用户意图 → 决定下一步
8
+ 2. 调用 OpenClaw 工具 (exec/browser/computer-use)
9
+ 3. 工具结果 → 分层量化存入 KV 缓存
10
+ 4. 循环直到任务完成
11
+
12
+ 架构:
13
+ Orchestrator
14
+ ├── Brain: DeepSeek 客户端 (本地 L40S / API 兜底)
15
+ ├── Hands: OpenClaw 工具桥 (exec/browser)
16
+ ├── Memory: 前缀感知缓存池 (复用 agentframe_core.py)
17
+ └── Loop: Agent 循环控制
18
+
19
+ 依赖:
20
+ - agentframe_core.py (前缀缓存池 + KV 量化)
21
+ - OpenClaw gateway (exec/browser 工具)
22
+ - DeepSeek 模型 (本地或 API)
23
+ """
24
+
25
+ import json
26
+ import time
27
+ import subprocess
28
+ from typing import Dict, List, Optional, Any, Callable
29
+
30
+ # ============================================================
31
+ # 1. Brain: DeepSeek 客户端抽象 (本地/API 可切换)
32
+ # ============================================================
33
+
34
+ class DeepSeekBrain:
35
+ """
36
+ DeepSeek 决策大脑
37
+ mode: 'local' (L40S 本地) | 'api' (DeepSeek 官方)
38
+ """
39
+
40
+ def __init__(self, mode: str = "api", model: str = "deepseek-chat",
41
+ api_key: str = "", base_url: str = "https://api.deepseek.com"):
42
+ self.mode = mode
43
+ self.model = model
44
+ self.api_key = api_key
45
+ self.base_url = base_url
46
+ self.system_prompt = ""
47
+ self.conversation: List[Dict] = []
48
+
49
+ def set_system_prompt(self, prompt: str):
50
+ """设置 Agent 系统提示 (含工具定义)"""
51
+ self.system_prompt = prompt
52
+ self.conversation = [{"role": "system", "content": prompt}]
53
+
54
+ def think(self, user_input: str, tools: List[Dict]) -> Dict:
55
+ """
56
+ DeepSeek 决策: 返回 JSON {action, tool, args} 或 {action: "reply", content}
57
+ """
58
+ if self.mode == "api":
59
+ return self._think_api(user_input, tools)
60
+ else:
61
+ return self._think_local(user_input, tools)
62
+
63
+ def _think_api(self, user_input: str, tools: List[Dict]) -> Dict:
64
+ """API 模式 (OpenAI 兼容)"""
65
+ import urllib.request
66
+
67
+ self.conversation.append({"role": "user", "content": user_input})
68
+ payload = {
69
+ "model": self.model,
70
+ "messages": self.conversation,
71
+ "tools": tools,
72
+ "tool_choice": "auto",
73
+ "stream": False,
74
+ }
75
+ req = urllib.request.Request(
76
+ f"{self.base_url}/chat/completions",
77
+ data=json.dumps(payload).encode(),
78
+ headers={
79
+ "Content-Type": "application/json",
80
+ "Authorization": f"Bearer {self.api_key}",
81
+ },
82
+ )
83
+ with urllib.request.urlopen(req, timeout=60) as resp:
84
+ data = json.loads(resp.read().decode())
85
+
86
+ msg = data["choices"][0]["message"]
87
+ self.conversation.append(msg)
88
+
89
+ # 解析工具调用
90
+ if msg.get("tool_calls"):
91
+ tc = msg["tool_calls"][0]
92
+ return {
93
+ "action": "tool",
94
+ "tool": tc["function"]["name"],
95
+ "args": json.loads(tc["function"]["arguments"] or "{}"),
96
+ }
97
+ return {"action": "reply", "content": msg.get("content", "")}
98
+
99
+ def _think_local(self, user_input: str, tools: List[Dict]) -> Dict:
100
+ """本地模式 (L40S + V2-Lite, 待实现真实推理)"""
101
+ # TODO: 接入 agentframe 的本地推理 (KV 优化)
102
+ # 目前返回占位, 等 GPU 开机实现
103
+ return {"action": "reply", "content": "[本地模式待实现 - 需 GPU]"}
104
+
105
+ def remember_tool_result(self, result: str):
106
+ """把工具结果追加到对话 (供下轮决策)"""
107
+ self.conversation.append({"role": "tool", "content": result})
108
+
109
+
110
+ # ============================================================
111
+ # 2. Hands: OpenClaw 工具桥
112
+ # ============================================================
113
+
114
+ class OpenClawHands:
115
+ """
116
+ OpenClaw 工具执行桥 (调用 gateway 的 exec/browser)
117
+ 通过 subprocess 调用 openclaw CLI, 或直接调用系统命令
118
+ """
119
+
120
+ def __init__(self, workspace: str = "/root/.openclaw/workspace"):
121
+ self.workspace = workspace
122
+
123
+ def exec(self, command: str, timeout: int = 30) -> Dict:
124
+ """执行 shell 命令 (OpenClaw exec 能力)"""
125
+ try:
126
+ result = subprocess.run(
127
+ command,
128
+ shell=True,
129
+ capture_output=True,
130
+ text=True,
131
+ timeout=timeout,
132
+ cwd=self.workspace,
133
+ )
134
+ return {
135
+ "success": result.returncode == 0,
136
+ "stdout": result.stdout[:2000],
137
+ "stderr": result.stderr[:500],
138
+ "exit_code": result.returncode,
139
+ }
140
+ except subprocess.TimeoutExpired:
141
+ return {"success": False, "stdout": "", "stderr": "timeout", "exit_code": -1}
142
+
143
+ def read_file(self, path: str) -> str:
144
+ """读文件"""
145
+ return self.exec(f"cat {path}")["stdout"]
146
+
147
+ def write_file(self, path: str, content: str) -> bool:
148
+ """写文件"""
149
+ import base64
150
+ b64 = base64.b64encode(content.encode()).decode()
151
+ result = self.exec(f"echo '{b64}' | base64 -d > {path}")
152
+ return result["success"]
153
+
154
+ def browser_open(self, url: str) -> Dict:
155
+ """浏览器打开网页"""
156
+ result = self.exec(f"openclaw browser --browser-profile openclaw open {url}")
157
+ return {"success": result["success"], "note": "browser opened"}
158
+
159
+ def browser_snapshot(self) -> str:
160
+ """浏览器快照"""
161
+ result = self.exec("openclaw browser --browser-profile openclaw snapshot")
162
+ return result["stdout"]
163
+
164
+ def list_tools(self) -> List[str]:
165
+ """可用工具清单"""
166
+ return ["exec", "read_file", "write_file", "browser_open", "browser_snapshot"]
167
+
168
+
169
+ # ============================================================
170
+ # 3. Memory: KV 记忆集成 (前缀缓存池)
171
+ # ============================================================
172
+
173
+ class AgentMemory:
174
+ """Agent 记忆: 复用 agentframe_core 的前缀池 + 分层量化"""
175
+
176
+ def __init__(self, system_prompt: str, prompt_tokens: int = 3000):
177
+ from agentframe_core import PrefixPool, SessionManager, AbsorbedMLAEncoder
178
+ self.pool = PrefixPool()
179
+ self.sessions = SessionManager(self.pool)
180
+ self.encoder = AbsorbedMLAEncoder()
181
+ self.system_prompt = system_prompt
182
+ # 创建主会话
183
+ self.session = self.sessions.create_session(
184
+ "agent-main", system_prompt, prompt_tokens
185
+ )
186
+ self.history: List[Dict] = []
187
+
188
+ def record(self, role: str, content: str, layer_idx: int = 0):
189
+ """记录对话/工具结果到记忆 (分层量化)"""
190
+ import numpy as np
191
+ # 模拟 KV 写入: 思考用 INT8, 工具结果用 INT4
192
+ fake_kv = np.random.randn(1, 8, self.encoder.kv_rank)
193
+ if role == "thought":
194
+ kv = self.encoder.encode_thought(fake_kv)
195
+ else:
196
+ kv = self.encoder.encode_tool_result(fake_kv)
197
+ self.sessions.append_tool_result(self.session, layer_idx, kv)
198
+ self.history.append({"role": role, "content": content[:500]})
199
+
200
+ def memory_report(self) -> Dict:
201
+ return self.sessions.session_memory(self.session)
202
+
203
+ def close(self):
204
+ self.sessions.close_session("agent-main")
205
+
206
+
207
+ # ============================================================
208
+ # 4. Loop: Agent 主循环
209
+ # ============================================================
210
+
211
+ class AgentLoop:
212
+ """Agent 执行循环: 思考 → 行动 → 观察 → 循环"""
213
+
214
+ def __init__(self, brain: DeepSeekBrain, hands: OpenClawHands, memory: AgentMemory):
215
+ self.brain = brain
216
+ self.hands = hands
217
+ self.memory = memory
218
+ self.max_steps = 10
219
+
220
+ def _tool_schemas(self) -> List[Dict]:
221
+ """给 DeepSeek 的工具定义 (OpenClaw 能力)"""
222
+ return [
223
+ {
224
+ "type": "function",
225
+ "function": {
226
+ "name": "exec",
227
+ "description": "执行 shell 命令",
228
+ "parameters": {
229
+ "type": "object",
230
+ "properties": {
231
+ "command": {"type": "string", "description": "要执行的命令"}
232
+ },
233
+ "required": ["command"],
234
+ },
235
+ },
236
+ },
237
+ {
238
+ "type": "function",
239
+ "function": {
240
+ "name": "browser_open",
241
+ "description": "打开网页",
242
+ "parameters": {
243
+ "type": "object",
244
+ "properties": {
245
+ "url": {"type": "string", "description": "网页地址"}
246
+ },
247
+ "required": ["url"],
248
+ },
249
+ },
250
+ },
251
+ ]
252
+
253
+ def run(self, task: str) -> str:
254
+ """执行一个任务"""
255
+ self.brain.set_system_prompt(self.memory.system_prompt)
256
+ self.memory.record("user", task)
257
+
258
+ for step in range(self.max_steps):
259
+ print(f"\n[Step {step+1}] 🧠 DeepSeek 思考中...")
260
+ self.memory.record("thought", f"step {step+1}")
261
+
262
+ decision = self.brain.think(task, self._tool_schemas())
263
+
264
+ if decision["action"] == "reply":
265
+ print(f" 💬 Agent: {decision['content']}")
266
+ return decision["content"]
267
+
268
+ if decision["action"] == "tool":
269
+ tool = decision["tool"]
270
+ args = decision["args"]
271
+ print(f" 🛠 调用 {tool}({args})")
272
+
273
+ # 执行工具
274
+ if tool == "exec":
275
+ result = self.hands.exec(args.get("command", ""))
276
+ elif tool == "browser_open":
277
+ result = self.hands.browser_open(args.get("url", ""))
278
+ else:
279
+ result = {"success": False, "stdout": f"未知工具 {tool}"}
280
+
281
+ # 记录结果 (工具结果 → INT4 压缩)
282
+ result_str = json.dumps(result, ensure_ascii=False)[:500]
283
+ self.memory.record("tool_result", result_str)
284
+ self.brain.remember_tool_result(result_str)
285
+ task = result_str # 下一轮基于结果继续
286
+
287
+ return "[达到最大步数,任务未完成]"
288
+
289
+
290
+ # ============================================================
291
+ # 5. 演示
292
+ # ============================================================
293
+
294
+ if __name__ == "__main__":
295
+ print("=" * 60)
296
+ print("AgentFrame 编排层 演示 (无 GPU 版)")
297
+ print("=" * 60)
298
+
299
+ # 1. 组装
300
+ sys_prompt = """你是智能助手,可以操控电脑完成任务。
301
+ 你有以下工具:
302
+ - exec: 执行 shell 命令
303
+ - browser_open: 打开网页
304
+ 请根据用户需求,一步步完成任务。每次只调用一个工具,观察结果后再决定下一步。"""
305
+
306
+ brain = DeepSeekBrain(mode="api", model="deepseek-chat", api_key="") # 无 key 时演示流程
307
+ hands = OpenClawHands()
308
+ memory = AgentMemory(sys_prompt, prompt_tokens=3000)
309
+
310
+ # 2. 测试手的能力 (无需 GPU)
311
+ print("\n🖐 测试 OpenClaw 手:")
312
+ r = hands.exec("echo 'AgentFrame 就绪!' && ls /root/.openclaw/workspace/*.py | head -3")
313
+ print(f" exec: {'✅' if r['success'] else '❌'}")
314
+ print(f" → {r['stdout'][:100]}")
315
+
316
+ # 3. 测试记忆 (前缀池)
317
+ print("\n🧠 测试记忆 (前缀缓存池):")
318
+ memory.record("thought", "分析任务")
319
+ memory.record("tool_result", "ls 输出: agentframe_core.py, agentframe_orchestrator.py")
320
+ rep = memory.memory_report()
321
+ print(f" 前缀引用数: {rep['prefix_refs']}")
322
+ print(f" 增量内存: {rep['incremental_bytes']/1024:.0f}KB")
323
+
324
+ # 4. Agent 循环 (无 API key 时走 reply 占位)
325
+ print("\n🤖 Agent 循环:")
326
+ loop = AgentLoop(brain, hands, memory)
327
+ # 用本地模式跑流程 (不真正调 API)
328
+ brain.mode = "local"
329
+ result = loop.run("查看当前目录有什么文件")
330
+ print(f" 结果: {result}")
331
+
332
+ print("\n✅ AgentFrame 编排层骨架验证完成")
333
+ print(" (真实 DeepSeek 推理需 GPU 开机后接入)")