Files
2026-09-24 10:40:57 +08:00

189 lines
7.8 KiB
Python

"""多轮agent循环版本的理解层:不再是一次性吐出整份JSON,而是像一个真正在动手
建模的人一样,一步一步调用能力、时不时看一眼当前状态/渲染图、自己发现问题自己
改,最后调用finish提交、执行器自动核验,核验不通过可以在同一个会话里继续修正
——不需要像原来的三级兜底那样整个从头重来。
这是参考同类项目(ai-mechanical-3dcad-design-agent)"编码agent驱动CAD、边做边看、
自动验证再修正"的思路,把它内建到我们自己的后端服务里,不需要用户额外装一个
编码agent——这正是我们跟那个项目最大的架构差异,也是补上的那一课。
"""
import base64
import json
from pathlib import Path
from openai import OpenAI
from .agent_session import AgentSession
from .agent_tools import CONTROL_TOOLS, build_capability_tools
from .llm_client import MAX_IMAGES, MODEL
from .prompt_builder import build_agent_system_prompt
from .registry import load_registry
from .images import normalize_images, response_content
from .validation import safe_name
MAX_STEPS = 60
class AgentGenerationError(Exception):
"""多轮agent循环始终无法收尾(比如超过步数上限)时抛出。"""
def _client() -> OpenAI:
import os
api_key = os.environ.get("OPENAI_API_KEY")
if not api_key:
raise AgentGenerationError("未找到 OPENAI_API_KEY,请检查 .env 文件")
return OpenAI(api_key=api_key)
def _to_input_item(item):
"""responses API返回的输出项(reasoning/function_call等)要原样喂回下一轮
input里维持会话连续性;SDK对象需要先转成普通dict才能重新序列化提交。"""
if hasattr(item, "model_dump"):
return item.model_dump(exclude_none=True)
return item
def generate_from_description_agentic(
user_text: str, base_dir: Path, image_base64: str | list[str] | None = None
) -> dict:
client = _client()
registry = load_registry(base_dir / "capabilities")
tools = build_capability_tools(registry) + CONTROL_TOOLS
system_prompt = build_agent_system_prompt(base_dir / "capabilities", user_text)
# 参考图可能不止一张(正视图/侧视图/俯视图分开拍),最多给MAX_IMAGES张,
# 跟一次性生成模式的build_user_content用的是同一条上限,行为保持一致。
images = normalize_images(image_base64)
user_content = response_content(user_text, images)
input_items = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_content},
]
session = AgentSession(registry)
default_name = "agentic_part"
print(f"[agent] 开始多轮建模会话...")
for step in range(1, MAX_STEPS + 1):
resp = client.responses.create(
model=MODEL,
input=input_items,
tools=tools,
reasoning={"effort": "medium"},
)
input_items += [_to_input_item(item) for item in resp.output]
function_calls = [item for item in resp.output if item.type == "function_call"]
if not function_calls:
print(f"[agent] 第{step}步:模型只说话没调用工具,提醒继续")
input_items.append(
{
"role": "user",
"content": "请继续调用工具完成建模,或者调用finish提交结果,不要只输出文字说明。",
}
)
continue
for call in function_calls:
name = call.name
try:
args = json.loads(call.arguments or "{}")
except json.JSONDecodeError as e:
input_items.append(
{
"type": "function_call_output",
"call_id": call.call_id,
"output": json.dumps({"ok": False, "error": f"ARGS_JSON_ERROR: {e}"}),
}
)
continue
if name == "finish":
safe_name(args.get("name", default_name))
out_dir = base_dir / "output" / args.get("name", default_name)
result = session.finish(
output_id=args.get("output_id"),
claims=args.get("claims", []),
expected_solid_count=args.get("expected_solid_count"),
out_dir=out_dir,
)
print(f"[agent] 第{step}步:调用finish -> {'成功' if result.get('ok') else '未通过: ' + result.get('error','')[:150]}")
if result.get("ok"):
result["part_name"] = args.get("name", default_name)
result["llm_assumptions"] = args.get("assumptions", [])
result["llm_model"] = MODEL
result["source_description"] = user_text
result["agentic"] = True
result["step_count"] = step
(out_dir / "report.json").write_text(
json.dumps(result, ensure_ascii=False, indent=2), encoding="utf-8"
)
return result
input_items.append(
{
"type": "function_call_output",
"call_id": call.call_id,
"output": json.dumps(result, ensure_ascii=False),
}
)
continue
if name == "inspect_state":
result = session.get_state()
elif name == "describe_shape":
result = session.describe(args.get("id", ""))
elif name == "render_snapshot":
png = session.render_png(args.get("id", ""), view=args.get("view", "iso"))
if png is None:
result = {"ok": False, "error": f"UNKNOWN_ID: '{args.get('id')}' 不存在"}
input_items.append(
{
"type": "function_call_output",
"call_id": call.call_id,
"output": json.dumps(result, ensure_ascii=False),
}
)
else:
input_items.append(
{
"type": "function_call_output",
"call_id": call.call_id,
"output": "已生成渲染图,见下一条图片消息。",
}
)
input_items.append(
{
"role": "user",
"content": [
{
"type": "input_image",
"image_url": f"data:image/png;base64,{base64.b64encode(png).decode()}",
}
],
}
)
continue
elif name.startswith("op_"):
op_name = name[len("op_") :]
target = args.pop("target", None)
op_id = args.pop("id", None)
result = session.apply_op(op_name, args, target=target, op_id=op_id)
tag = "OK" if result.get("ok") else "FAIL:" + result.get("error", "")[:100]
print(f"[agent] 第{step}步: {op_name} -> {tag}")
else:
result = {"ok": False, "error": f"UNKNOWN_TOOL: {name}"}
input_items.append(
{
"type": "function_call_output",
"call_id": call.call_id,
"output": json.dumps(result, ensure_ascii=False),
}
)
raise AgentGenerationError(f"超过{MAX_STEPS}步仍未调用finish完成建模")