Spaces:
Paused
Paused
Upload 2 files
Browse files- app.py +311 -0
- requirements.txt +5 -0
app.py
ADDED
|
@@ -0,0 +1,311 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Claude Web API Proxy - HuggingFace Spaces with ReAct
|
| 3 |
+
"""
|
| 4 |
+
from fastapi import FastAPI, Request
|
| 5 |
+
from fastapi.responses import StreamingResponse, JSONResponse
|
| 6 |
+
from fastapi.middleware.cors import CORSMiddleware
|
| 7 |
+
import httpx
|
| 8 |
+
import json
|
| 9 |
+
import asyncio
|
| 10 |
+
from datetime import datetime
|
| 11 |
+
from typing import Optional, Dict
|
| 12 |
+
import os
|
| 13 |
+
import uuid
|
| 14 |
+
import re
|
| 15 |
+
import pickle
|
| 16 |
+
from pathlib import Path
|
| 17 |
+
import logging
|
| 18 |
+
|
| 19 |
+
# 日志配置
|
| 20 |
+
logging.basicConfig(level=logging.INFO, format='%(asctime)s [%(levelname)s] %(message)s')
|
| 21 |
+
logger = logging.getLogger(__name__)
|
| 22 |
+
|
| 23 |
+
# 重试函数
|
| 24 |
+
async def retry_async(fn, max_retries=3, delay=1.0):
|
| 25 |
+
for i in range(max_retries):
|
| 26 |
+
try:
|
| 27 |
+
return await fn()
|
| 28 |
+
except Exception as e:
|
| 29 |
+
if i == max_retries - 1:
|
| 30 |
+
raise
|
| 31 |
+
logger.warning(f"Retry {i+1}/{max_retries}: {e}")
|
| 32 |
+
await asyncio.sleep(delay * (i + 1))
|
| 33 |
+
|
| 34 |
+
app = FastAPI(title="Claude Proxy")
|
| 35 |
+
app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
|
| 36 |
+
|
| 37 |
+
# 文件持久化
|
| 38 |
+
CACHE_FILE = Path("/tmp/conversation_cache.pkl")
|
| 39 |
+
conversation_cache: Dict[str, str] = {}
|
| 40 |
+
|
| 41 |
+
def load_cache():
|
| 42 |
+
global conversation_cache
|
| 43 |
+
if CACHE_FILE.exists():
|
| 44 |
+
try:
|
| 45 |
+
with open(CACHE_FILE, 'rb') as f:
|
| 46 |
+
conversation_cache = pickle.load(f)
|
| 47 |
+
except:
|
| 48 |
+
conversation_cache = {}
|
| 49 |
+
|
| 50 |
+
def save_cache():
|
| 51 |
+
try:
|
| 52 |
+
with open(CACHE_FILE, 'wb') as f:
|
| 53 |
+
pickle.dump(conversation_cache, f)
|
| 54 |
+
except:
|
| 55 |
+
pass
|
| 56 |
+
|
| 57 |
+
# 智能保活
|
| 58 |
+
last_request_time = datetime.now()
|
| 59 |
+
|
| 60 |
+
async def keep_alive():
|
| 61 |
+
while True:
|
| 62 |
+
# 根据最后请求时间动态调整保活间隔
|
| 63 |
+
idle_time = (datetime.now() - last_request_time).total_seconds()
|
| 64 |
+
if idle_time < 3600: # 1小时内有请求,频繁保活
|
| 65 |
+
await asyncio.sleep(1800) # 30分钟
|
| 66 |
+
else: # 空闲,降低频率
|
| 67 |
+
await asyncio.sleep(6 * 3600) # 6小时
|
| 68 |
+
|
| 69 |
+
try:
|
| 70 |
+
async with httpx.AsyncClient() as client:
|
| 71 |
+
url = os.getenv("SPACE_URL", "")
|
| 72 |
+
if url:
|
| 73 |
+
await client.get(f"{url}/health")
|
| 74 |
+
except:
|
| 75 |
+
pass
|
| 76 |
+
|
| 77 |
+
@app.on_event("startup")
|
| 78 |
+
async def startup():
|
| 79 |
+
load_cache()
|
| 80 |
+
asyncio.create_task(keep_alive())
|
| 81 |
+
|
| 82 |
+
@app.get("/health")
|
| 83 |
+
async def health():
|
| 84 |
+
return {"status": "ok", "time": datetime.now().isoformat()}
|
| 85 |
+
|
| 86 |
+
@app.get("/")
|
| 87 |
+
async def root():
|
| 88 |
+
return {"name": "Claude Proxy", "version": "2.0-hf-react"}
|
| 89 |
+
|
| 90 |
+
async def get_org_id(key: str) -> Optional[str]:
|
| 91 |
+
try:
|
| 92 |
+
async with httpx.AsyncClient() as client:
|
| 93 |
+
r = await client.get('https://claude.ai/api/organizations',
|
| 94 |
+
headers={'Cookie': f'sessionKey={key}', 'User-Agent': 'Mozilla/5.0'}, timeout=30.0)
|
| 95 |
+
if r.status_code == 200:
|
| 96 |
+
data = r.json()
|
| 97 |
+
return data[0]['uuid'] if data else None
|
| 98 |
+
except:
|
| 99 |
+
pass
|
| 100 |
+
return None
|
| 101 |
+
|
| 102 |
+
async def create_conv(key: str, org_id: str) -> Optional[str]:
|
| 103 |
+
try:
|
| 104 |
+
conv_id = str(uuid.uuid4())
|
| 105 |
+
async with httpx.AsyncClient() as client:
|
| 106 |
+
r = await client.post(f'https://claude.ai/api/organizations/{org_id}/chat_conversations',
|
| 107 |
+
headers={'Content-Type': 'application/json', 'Cookie': f'sessionKey={key}', 'User-Agent': 'Mozilla/5.0'},
|
| 108 |
+
json={'uuid': conv_id, 'name': '', 'include_conversation_preferences': True}, timeout=30.0)
|
| 109 |
+
if r.status_code == 200:
|
| 110 |
+
return r.json().get('uuid', conv_id)
|
| 111 |
+
except:
|
| 112 |
+
pass
|
| 113 |
+
return None
|
| 114 |
+
|
| 115 |
+
def extract_prompt(messages):
|
| 116 |
+
result = []
|
| 117 |
+
for m in messages:
|
| 118 |
+
role = m['role'].title()
|
| 119 |
+
content = m['content']
|
| 120 |
+
|
| 121 |
+
# 支持图片
|
| 122 |
+
if isinstance(content, list):
|
| 123 |
+
parts = []
|
| 124 |
+
for part in content:
|
| 125 |
+
if part.get('type') == 'text':
|
| 126 |
+
parts.append(part['text'])
|
| 127 |
+
elif part.get('type') == 'image_url':
|
| 128 |
+
parts.append(f"[Image: {part['image_url']['url']}]")
|
| 129 |
+
result.append(f"{role}: {' '.join(parts)}")
|
| 130 |
+
else:
|
| 131 |
+
result.append(f"{role}: {content}")
|
| 132 |
+
|
| 133 |
+
return '\n\n'.join(result)
|
| 134 |
+
|
| 135 |
+
def extract_attachments(messages):
|
| 136 |
+
attachments = []
|
| 137 |
+
for m in messages:
|
| 138 |
+
if isinstance(m['content'], list):
|
| 139 |
+
for part in m['content']:
|
| 140 |
+
if part.get('type') == 'image_url':
|
| 141 |
+
attachments.append({
|
| 142 |
+
'extracted_content': '',
|
| 143 |
+
'file_name': 'image.png',
|
| 144 |
+
'file_size': 0,
|
| 145 |
+
'file_type': 'image/png',
|
| 146 |
+
'url': part['image_url']['url']
|
| 147 |
+
})
|
| 148 |
+
return attachments
|
| 149 |
+
|
| 150 |
+
def enhance_prompt(prompt: str, messages: list) -> str:
|
| 151 |
+
has_system = any(m['role'] == 'system' for m in messages)
|
| 152 |
+
if not has_system:
|
| 153 |
+
return prompt
|
| 154 |
+
return prompt + "\n\n[Environment: OpenClaw | Tools: web_search, artifacts, repl available]"
|
| 155 |
+
|
| 156 |
+
def inject_react_prompt(prompt: str, tools: list) -> str:
|
| 157 |
+
if not tools:
|
| 158 |
+
return prompt
|
| 159 |
+
tool_defs = '\n\n'.join([f"{t['name']}: {t.get('description', '')}" for t in tools])
|
| 160 |
+
return prompt + f"""
|
| 161 |
+
|
| 162 |
+
[Tool Calling Protocol]
|
| 163 |
+
You have access to these tools:
|
| 164 |
+
{tool_defs}
|
| 165 |
+
|
| 166 |
+
Use this format:
|
| 167 |
+
Thought: [reasoning]
|
| 168 |
+
Action: [tool_name]
|
| 169 |
+
Action Input: [JSON]
|
| 170 |
+
|
| 171 |
+
After result:
|
| 172 |
+
Observation: [result]
|
| 173 |
+
|
| 174 |
+
When done:
|
| 175 |
+
Final Answer: [response]
|
| 176 |
+
"""
|
| 177 |
+
|
| 178 |
+
def parse_react_output(text: str) -> Optional[dict]:
|
| 179 |
+
action_match = re.search(r'Action:\s*(\w+)', text)
|
| 180 |
+
input_match = re.search(r'Action Input:\s*(\{[^}]*\})', text)
|
| 181 |
+
if action_match and input_match:
|
| 182 |
+
try:
|
| 183 |
+
return {'type': 'tool_use', 'id': f'tool_{int(datetime.now().timestamp())}',
|
| 184 |
+
'name': action_match.group(1), 'input': json.loads(input_match.group(1))}
|
| 185 |
+
except:
|
| 186 |
+
return None
|
| 187 |
+
return None
|
| 188 |
+
|
| 189 |
+
@app.post("/v1/chat/completions")
|
| 190 |
+
async def chat(request: Request):
|
| 191 |
+
global last_request_time
|
| 192 |
+
last_request_time = datetime.now()
|
| 193 |
+
start_time = datetime.now()
|
| 194 |
+
|
| 195 |
+
try:
|
| 196 |
+
body = await request.json()
|
| 197 |
+
auth = request.headers.get('Authorization', '')
|
| 198 |
+
|
| 199 |
+
logger.info(f"Request received: model={body.get('model', 'claude-sonnet-4-6')}")
|
| 200 |
+
|
| 201 |
+
if not auth.startswith('Bearer '):
|
| 202 |
+
return JSONResponse({"error": {"message": "Missing auth", "type": "auth_error"}}, 401)
|
| 203 |
+
|
| 204 |
+
key = auth.replace('Bearer ', '')
|
| 205 |
+
org_id = await retry_async(lambda: get_org_id(key))
|
| 206 |
+
if not org_id:
|
| 207 |
+
logger.error("Invalid sessionKey")
|
| 208 |
+
return JSONResponse({"error": {"message": "Invalid key", "type": "auth_error"}}, 401)
|
| 209 |
+
|
| 210 |
+
conv_id = conversation_cache.get(key)
|
| 211 |
+
if not conv_id:
|
| 212 |
+
conv_id = await retry_async(lambda: create_conv(key, org_id))
|
| 213 |
+
if conv_id:
|
| 214 |
+
conversation_cache[key] = conv_id
|
| 215 |
+
save_cache()
|
| 216 |
+
if not conv_id:
|
| 217 |
+
logger.error("Failed to create conversation")
|
| 218 |
+
return JSONResponse({"error": {"message": "Conv error", "type": "api_error"}}, 500)
|
| 219 |
+
|
| 220 |
+
messages = body.get('messages', [])
|
| 221 |
+
tools = body.get('tools', [])
|
| 222 |
+
prompt = extract_prompt(messages)
|
| 223 |
+
prompt = enhance_prompt(prompt, messages)
|
| 224 |
+
if tools:
|
| 225 |
+
prompt = inject_react_prompt(prompt, tools)
|
| 226 |
+
|
| 227 |
+
attachments = extract_attachments(messages)
|
| 228 |
+
model = body.get('model', 'claude-sonnet-4-6')
|
| 229 |
+
payload = {
|
| 230 |
+
'prompt': prompt,
|
| 231 |
+
'timezone': body.get('timezone', 'Asia/Shanghai'),
|
| 232 |
+
'locale': body.get('locale', 'zh-CN'),
|
| 233 |
+
'model': model,
|
| 234 |
+
'rendering_mode': 'messages',
|
| 235 |
+
'attachments': attachments,
|
| 236 |
+
'files': [],
|
| 237 |
+
'tools': [
|
| 238 |
+
{"type": "web_search_v0", "name": "web_search"},
|
| 239 |
+
{"type": "artifacts_v0", "name": "artifacts"},
|
| 240 |
+
{"type": "repl_v0", "name": "repl"}
|
| 241 |
+
]
|
| 242 |
+
}
|
| 243 |
+
|
| 244 |
+
logger.info(f"Sending to Claude: attachments={len(attachments)}")
|
| 245 |
+
|
| 246 |
+
async def generate():
|
| 247 |
+
try:
|
| 248 |
+
full_text = ''
|
| 249 |
+
async with httpx.AsyncClient() as client:
|
| 250 |
+
async with client.stream('POST',
|
| 251 |
+
f'https://claude.ai/api/organizations/{org_id}/chat_conversations/{conv_id}/completion',
|
| 252 |
+
headers={'Content-Type': 'application/json', 'Cookie': f'sessionKey={key}', 'User-Agent': 'Mozilla/5.0'},
|
| 253 |
+
json=payload, timeout=120.0) as response:
|
| 254 |
+
|
| 255 |
+
async for line in response.aiter_lines():
|
| 256 |
+
if line.startswith('data: '):
|
| 257 |
+
try:
|
| 258 |
+
data = json.loads(line[6:])
|
| 259 |
+
if data.get('type') == 'content_block_delta' and data.get('delta', {}).get('text'):
|
| 260 |
+
text = data['delta']['text']
|
| 261 |
+
full_text += text
|
| 262 |
+
|
| 263 |
+
# ReAct: 检测工具调用
|
| 264 |
+
react_result = parse_react_output(full_text)
|
| 265 |
+
if react_result and react_result['type'] == 'tool_use':
|
| 266 |
+
chunk = {
|
| 267 |
+
"id": f"chatcmpl-{int(datetime.now().timestamp())}",
|
| 268 |
+
"object": "chat.completion.chunk",
|
| 269 |
+
"created": int(datetime.now().timestamp()),
|
| 270 |
+
"model": model,
|
| 271 |
+
"choices": [{
|
| 272 |
+
"index": 0,
|
| 273 |
+
"delta": {
|
| 274 |
+
"tool_calls": [{
|
| 275 |
+
"id": react_result['id'],
|
| 276 |
+
"type": "function",
|
| 277 |
+
"function": {
|
| 278 |
+
"name": react_result['name'],
|
| 279 |
+
"arguments": json.dumps(react_result['input'])
|
| 280 |
+
}
|
| 281 |
+
}]
|
| 282 |
+
},
|
| 283 |
+
"finish_reason": "tool_calls"
|
| 284 |
+
}]
|
| 285 |
+
}
|
| 286 |
+
yield f"data: {json.dumps(chunk)}\n\n"
|
| 287 |
+
yield "data: [DONE]\n\n"
|
| 288 |
+
return
|
| 289 |
+
|
| 290 |
+
# 正常文本
|
| 291 |
+
chunk = {
|
| 292 |
+
"id": f"chatcmpl-{int(datetime.now().timestamp())}",
|
| 293 |
+
"object": "chat.completion.chunk",
|
| 294 |
+
"created": int(datetime.now().timestamp()),
|
| 295 |
+
"model": model,
|
| 296 |
+
"choices": [{"index": 0, "delta": {"content": text}, "finish_reason": None}]
|
| 297 |
+
}
|
| 298 |
+
yield f"data: {json.dumps(chunk)}\n\n"
|
| 299 |
+
except:
|
| 300 |
+
pass
|
| 301 |
+
yield "data: [DONE]\n\n"
|
| 302 |
+
except Exception as e:
|
| 303 |
+
logger.error(f"Stream error: {e}")
|
| 304 |
+
yield f"data: {json.dumps({'error': str(e)})}\n\n"
|
| 305 |
+
|
| 306 |
+
return StreamingResponse(generate(), media_type="text/event-stream")
|
| 307 |
+
|
| 308 |
+
except Exception as e:
|
| 309 |
+
duration = (datetime.now() - start_time).total_seconds()
|
| 310 |
+
logger.error(f"Request failed: {e}, duration={duration}s")
|
| 311 |
+
return JSONResponse({"error": {"message": str(e), "type": "api_error"}}, 500)
|
requirements.txt
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
fastapi>=0.115.2
|
| 2 |
+
uvicorn>=0.24.0
|
| 3 |
+
httpx>=0.25.1
|
| 4 |
+
sse-starlette>=1.8.2
|
| 5 |
+
python-multipart>=0.0.6
|