feat(tools): refactor migrate OpenClaw from custom tool to builtin tool
Create OpenClawTool class inheriting BuiltinTool with dedicated config Remove all x-openclaw special handling from CustomTool (~270 lines) Add multi-operation support (print_task, device_query, image_understand, general) Change ensure_builtin_tools_initialized to incremental mode for auto-provisioning Fix OperationTool and LangchainAdapter to support OpenClaw operation routing
This commit is contained in:
miao
@@ -31,7 +31,7 @@ class CustomTool(BaseTool):
|
||||
self.base_url = config.get("base_url", "")
|
||||
self.timeout = config.get("timeout", 30)
|
||||
|
||||
#===========OpenClaw特殊判断（取到OpenClaw特殊配置）==========
|
||||
# 解析 schema
|
||||
schema = self.schema_content
|
||||
if isinstance(schema, str):
|
||||
try:
|
||||
@@ -39,51 +39,7 @@ class CustomTool(BaseTool):
|
||||
self.schema_content = schema
|
||||
except json.JSONDecodeError:
|
||||
schema = {}
|
||||
|
||||
info = schema.get("info", {}) if isinstance(schema, dict) else {}
|
||||
self._is_openclaw = info.get("x-openclaw", False)
|
||||
|
||||
if self._is_openclaw:
|
||||
# 从扩展字段读取 OpenClaw 配置
|
||||
self._openclaw_agent_id = info.get("x-openclaw-agent-id", "main")
|
||||
self._openclaw_model = info.get("x-openclaw-default-model", "openclaw")
|
||||
self._openclaw_session_strategy = info.get(
|
||||
"x-openclaw-session-strategy", "by_user")
|
||||
self._openclaw_timeout = info.get("x-openclaw-timeout", 60)
|
||||
self._openclaw_input_mode = info.get("x-openclaw-input-mode", "text")
|
||||
self._openclaw_output_mode = info.get("x-openclaw-output-mode", "text")
|
||||
|
||||
# 从 servers 读取 base_url
|
||||
servers = schema.get("servers", [])
|
||||
if servers:
|
||||
self.base_url = servers[0].get("url", "")
|
||||
|
||||
# 从 auth_config 读取 token（兼容 api_key 和 bearer_token 两种认证方式）
|
||||
self._openclaw_token = (
|
||||
self.auth_config.get("api_key") # api_key 认证方式
|
||||
or self.auth_config.get("token") # bearer_token 认证方式
|
||||
or ""
|
||||
)
|
||||
|
||||
# 覆盖 timeout
|
||||
self.timeout = self._openclaw_timeout
|
||||
|
||||
# 运行时上下文（后续注入）
|
||||
self._user_id = "anonymous"
|
||||
self._conversation_id = None
|
||||
self._uploaded_files = [] # 新增：用户上传的文件
|
||||
|
||||
# 跳过 Schema 解析
|
||||
self._parsed_operations = {}
|
||||
|
||||
logger.info(
|
||||
f"检测到 OpenClaw 工具: agent_id={self._openclaw_agent_id}, "
|
||||
f"base_url={self.base_url}, "
|
||||
f"input_mode={self._openclaw_input_mode}, "
|
||||
f"output_mode={self._openclaw_output_mode}")
|
||||
else:
|
||||
# 解析schema
|
||||
self._parsed_operations = self._parse_openapi_schema()
|
||||
self._parsed_operations = self._parse_openapi_schema()
|
||||
|
||||
@property
|
||||
def name(self) -> str:
|
||||
@@ -109,31 +65,6 @@ class CustomTool(BaseTool):
|
||||
@property
|
||||
def parameters(self) -> List[ToolParameter]:
|
||||
"""工具参数定义"""
|
||||
# ========== OpenClaw 特判 根据输入模式解析是否需要image_url ==========
|
||||
if self._is_openclaw:
|
||||
params = [
|
||||
ToolParameter(
|
||||
name="message",
|
||||
type=ParameterType.STRING,
|
||||
description="发送给 OpenClaw Agent 的文本请求内容",
|
||||
required=True
|
||||
)
|
||||
]
|
||||
# 多模态输入模式下，增加 image_url 参数
|
||||
if self._openclaw_input_mode == "multimodal":
|
||||
params.append(ToolParameter(
|
||||
name="image_url",
|
||||
type=ParameterType.STRING,
|
||||
description=(
|
||||
"可选，附带的图片URL或base64 data URI"
|
||||
"（如 data:image/png;base64,...）。"
|
||||
"传入后 Agent 可以理解图片内容。"
|
||||
),
|
||||
required=False
|
||||
))
|
||||
return params
|
||||
# ========== 特判结束 ==========
|
||||
|
||||
params = []
|
||||
|
||||
# 添加操作选择参数
|
||||
@@ -166,10 +97,6 @@ class CustomTool(BaseTool):
|
||||
|
||||
async def execute(self, **kwargs) -> ToolResult:
|
||||
"""执行自定义工具"""
|
||||
# ========== OpenClaw 特判 ==========
|
||||
if self._is_openclaw:
|
||||
return await self._execute_openclaw(**kwargs)
|
||||
# ========== 特判结束 ==========
|
||||
start_time = time.time()
|
||||
|
||||
try:
|
||||
@@ -210,275 +137,6 @@ class CustomTool(BaseTool):
|
||||
execution_time=execution_time
|
||||
)
|
||||
|
||||
#=============openclaw执行函数开始===============
|
||||
async def _execute_openclaw(self, **kwargs) -> ToolResult:
|
||||
"""OpenClaw 专属执行逻辑（支持多模态输入）"""
|
||||
start_time = time.time()
|
||||
try:
|
||||
message = kwargs.get("message", "")
|
||||
# 从用户实际上传的文件中提取图片 URL
|
||||
image_url = None
|
||||
if self._uploaded_files:
|
||||
for f in self._uploaded_files:
|
||||
f_type = f.get("type", "")
|
||||
if f_type == "image":
|
||||
# Bedrock/Anthropic 格式：{"type": "image", "source": {"type": "base64", ...}}
|
||||
source = f.get("source", {})
|
||||
if source.get("type") == "base64":
|
||||
media_type = source.get("media_type", "image/jpeg")
|
||||
data = source.get("data", "")
|
||||
image_url = f"data:{media_type};base64,{data}"
|
||||
elif f.get("image"):
|
||||
# DashScope 格式：{"type": "image", "image": "url"}
|
||||
image_url = f.get("image")
|
||||
elif f.get("url"):
|
||||
# 其他格式：{"type": "image", "url": "https://..."}
|
||||
image_url = f.get("url")
|
||||
break
|
||||
elif f_type == "image_url":
|
||||
# OpenAI/Volcano 格式：{"type": "image_url", "image_url": {"url": "..."}}
|
||||
image_url = f.get("image_url", {}).get("url", "")
|
||||
break
|
||||
|
||||
# 如果 image_url 是服务器中转 URL，直接下载图片转 base64
|
||||
# 避免 OSS 签名 URL 在重定向解析过程中被破坏
|
||||
if image_url and not image_url.startswith("data:"):
|
||||
try:
|
||||
import base64
|
||||
from io import BytesIO
|
||||
from PIL import Image
|
||||
|
||||
MAX_RAW_SIZE = 4 * 1024 * 1024 # 超过 4MB 则压缩
|
||||
|
||||
async with aiohttp.ClientSession() as _session:
|
||||
async with _session.get(image_url, allow_redirects=True, timeout=aiohttp.ClientTimeout(total=30)) as _resp:
|
||||
if _resp.status == 200:
|
||||
content_type = _resp.headers.get("Content-Type", "image/jpeg")
|
||||
if content_type.startswith("image/"):
|
||||
img_bytes = await _resp.read()
|
||||
original_size = len(img_bytes)
|
||||
logger.info(f"OpenClaw 下载图片: size={original_size} bytes, type={content_type}")
|
||||
|
||||
if original_size > MAX_RAW_SIZE:
|
||||
img = Image.open(BytesIO(img_bytes))
|
||||
if img.mode in ("RGBA", "P", "LA"):
|
||||
img = img.convert("RGB")
|
||||
max_side = 2048
|
||||
if max(img.size) > max_side:
|
||||
img.thumbnail((max_side, max_side), Image.LANCZOS)
|
||||
buf = BytesIO()
|
||||
img.save(buf, format="JPEG", quality=75, optimize=True)
|
||||
img_bytes = buf.getvalue()
|
||||
content_type = "image/jpeg"
|
||||
logger.info(f"OpenClaw 图片已压缩: {original_size} -> {len(img_bytes)} bytes")
|
||||
|
||||
b64_data = base64.b64encode(img_bytes).decode("utf-8")
|
||||
image_url = f"data:{content_type};base64,{b64_data}"
|
||||
logger.info(f"OpenClaw 图片已转为 base64, size={len(img_bytes)} bytes")
|
||||
else:
|
||||
logger.warning(f"OpenClaw 图片 URL 返回非图片类型: {content_type}")
|
||||
else:
|
||||
logger.warning(f"OpenClaw 下载图片失败: HTTP {_resp.status}")
|
||||
except Exception as e:
|
||||
logger.warning(f"OpenClaw 下载图片失败，使用原始 URL: {e}")
|
||||
|
||||
|
||||
if not message:
|
||||
return ToolResult.error_result(
|
||||
error="message 参数不能为空",
|
||||
error_code="OPENCLAW_INVALID_INPUT",
|
||||
execution_time=time.time() - start_time)
|
||||
|
||||
url = f"{self.base_url.rstrip('/')}/v1/responses"
|
||||
#请求头
|
||||
headers = {
|
||||
"Authorization": f"Bearer {self._openclaw_token}",
|
||||
"Content-Type": "application/json",
|
||||
"x-openclaw-agent-id": self._openclaw_agent_id
|
||||
}
|
||||
|
||||
# session 路由
|
||||
if (self._openclaw_session_strategy == "by_conversation"
|
||||
and self._conversation_id):
|
||||
user_field = f"conv-{self._conversation_id}"
|
||||
else:
|
||||
user_field = f"user-{self._user_id}"
|
||||
|
||||
# 根据 input_mode 和是否有图片构造 input
|
||||
input_field = self._build_openclaw_input(message, image_url)
|
||||
#请求体
|
||||
body = {
|
||||
"model": self._openclaw_model,
|
||||
"user": user_field,
|
||||
"input": input_field,
|
||||
"stream": False
|
||||
}
|
||||
|
||||
logger.info(f"OpenClaw 请求体: {json.dumps(body, ensure_ascii=False)[:1000]}")
|
||||
|
||||
timeout_config = aiohttp.ClientTimeout(total=self.timeout)
|
||||
#请求
|
||||
async with aiohttp.ClientSession(timeout=timeout_config) as session:
|
||||
async with session.post(url, json=body, headers=headers) as resp:
|
||||
execution_time = time.time() - start_time
|
||||
|
||||
if resp.status >= 400:
|
||||
error_text = await resp.text()
|
||||
_img_preview2 = (image_url[:100] + "...") if image_url and len(image_url) > 100 else image_url
|
||||
logger.error(
|
||||
f"OpenClaw 调用失败: HTTP {resp.status}, "
|
||||
f"url={url}, agent_id={self._openclaw_agent_id}, "
|
||||
f"has_image={bool(image_url)}, image_url={_img_preview2}, "
|
||||
f"input_type={'multimodal' if isinstance(input_field, list) else 'text'}, "
|
||||
f"error_response={error_text[:1000]}"
|
||||
)
|
||||
return ToolResult.error_result(
|
||||
error=f"OpenClaw HTTP {resp.status}: {error_text[:500]}",
|
||||
error_code="OPENCLAW_HTTP_ERROR",
|
||||
execution_time=execution_time)
|
||||
|
||||
data = await resp.json()
|
||||
|
||||
# 根据 output_mode 解析响应
|
||||
result = self._extract_openclaw_response(
|
||||
data, self._openclaw_output_mode)
|
||||
display_text = self._format_openclaw_result(result)
|
||||
|
||||
logger.info(
|
||||
"OpenClaw 调用成功",
|
||||
extra={
|
||||
"tool_id": self.tool_id,
|
||||
"agent_id": self._openclaw_agent_id,
|
||||
"has_images": len(result["images"]) > 0,
|
||||
"execution_time": execution_time
|
||||
})
|
||||
return ToolResult.success_result(
|
||||
data=display_text, execution_time=execution_time)
|
||||
|
||||
except aiohttp.ClientError as e:
|
||||
return ToolResult.error_result(
|
||||
error=f"OpenClaw 网络连接失败: {str(e)}",
|
||||
error_code="OPENCLAW_NETWORK_ERROR",
|
||||
execution_time=time.time() - start_time)
|
||||
except Exception as e:
|
||||
return ToolResult.error_result(
|
||||
error=f"OpenClaw 调用失败: {str(e)}",
|
||||
error_code="OPENCLAW_EXECUTION_ERROR",
|
||||
execution_time=time.time() - start_time)
|
||||
|
||||
def _build_openclaw_input(self, message: str, image_url: str = None):
|
||||
"""根据 input_mode 和是否有图片构造 OpenClaw input 字段
|
||||
|
||||
纯文本模式或无图片 → 返回字符串
|
||||
多模态模式且有图片 → 返回结构化 item 数组
|
||||
"""
|
||||
if not image_url or self._openclaw_input_mode != "multimodal":
|
||||
return message
|
||||
|
||||
# 构造多模态 content 数组
|
||||
content_parts = [
|
||||
{"type": "input_text", "text": message}
|
||||
]
|
||||
|
||||
if image_url.startswith("data:"):
|
||||
# base64 data URI: data:image/png;base64,iVBORw0KGgo...
|
||||
try:
|
||||
header, data = image_url.split(",", 1)
|
||||
media_type = header.split(":")[1].split(";")[0]
|
||||
content_parts.append({
|
||||
"type": "input_image",
|
||||
"source": {
|
||||
"type": "base64",
|
||||
"media_type": media_type,
|
||||
"data": data
|
||||
}
|
||||
})
|
||||
except (ValueError, IndexError):
|
||||
logger.warning("无法解析 base64 data URI，回退为纯文本输入")
|
||||
return message
|
||||
else:
|
||||
# URL 引用
|
||||
content_parts.append({
|
||||
"type": "input_image",
|
||||
"source": {
|
||||
"type": "url",
|
||||
"url": image_url
|
||||
}
|
||||
})
|
||||
|
||||
return [{
|
||||
"type": "message",
|
||||
"role": "user",
|
||||
"content": content_parts
|
||||
}]
|
||||
|
||||
@staticmethod
|
||||
def _extract_openclaw_response(response_data: Dict[str, Any],
|
||||
output_mode: str = "text") -> Dict[str, Any]:
|
||||
"""从 OpenClaw 响应中提取文本和图片
|
||||
|
||||
响应格式：
|
||||
{"output": [{"type": "message", "content": [
|
||||
{"type": "output_text", "text": "..."},
|
||||
{"type": "output_image", "image_url": "..."}
|
||||
]}]}
|
||||
|
||||
返回:
|
||||
{"text": "文本内容", "images": [{"url": "...", "media_type": "image/png"}]}
|
||||
"""
|
||||
output = response_data.get("output", [])
|
||||
texts = []
|
||||
images = []
|
||||
|
||||
for item in output:
|
||||
if item.get("type") == "message":
|
||||
for content in item.get("content", []):
|
||||
content_type = content.get("type")
|
||||
|
||||
if content_type == "output_text":
|
||||
text = content.get("text", "")
|
||||
if text:
|
||||
texts.append(text)
|
||||
|
||||
elif content_type == "output_image" and output_mode == "multimodal":
|
||||
image_url = content.get("image_url", "")
|
||||
if image_url:
|
||||
images.append({
|
||||
"url": image_url,
|
||||
"media_type": content.get("media_type", "image/png")
|
||||
})
|
||||
|
||||
text_result = "\n".join(texts) if texts else ""
|
||||
|
||||
# text 模式下只返回文本（向后兼容）
|
||||
if output_mode == "text":
|
||||
return {"text": text_result or str(response_data), "images": []}
|
||||
|
||||
return {"text": text_result, "images": images}
|
||||
|
||||
@staticmethod
|
||||
def _format_openclaw_result(result: Dict[str, Any]) -> str:
|
||||
"""将解析结果格式化为返回给 LLM 的字符串
|
||||
|
||||
纯文本 → 直接返回
|
||||
有图片 → 将图片以 Markdown 格式嵌入文本
|
||||
"""
|
||||
text = result.get("text", "")
|
||||
images = result.get("images", [])
|
||||
|
||||
if not images:
|
||||
return text or "（OpenClaw 返回了空内容）"
|
||||
|
||||
parts = []
|
||||
if text:
|
||||
parts.append(text)
|
||||
for i, img in enumerate(images):
|
||||
parts.append(f"")
|
||||
|
||||
return "\n\n".join(parts)
|
||||
|
||||
|
||||
#=============openclaw执行函数结束================
|
||||
def _parse_openapi_schema(self) -> Dict[str, Any]:
|
||||
"""解析OpenAPI schema"""
|
||||
operations = {}
|
||||
|
||||
Reference in New Issue
Block a user