Files
xiaoyin-gateway/app/adapters/qwenpaw_engine.py
T
api-gateway c0706d1846 feat(agents): 新增「智能体」功能模块(路线 A)+ 版本 2.5.0
老板要求「在 ai 网关里做智能体 agent,新的一个菜单」,一期仅管理员、成熟后下发。
完整设计(27 条决策,4 轮评审)见 docs/agents-dev-plan.md。

架构(路线 A):界面与权限在网关,执行在 QwenPaw(独立 profile ai-gateway)。
网关当门面(鉴权/计量/灰度/审计),引擎提供工具能力(文件/命令/浏览器)。

Step -1 技术验证(实测,全部改变或确认了设计):
- 请求级指定模型【无效】:传 model 被忽略,turn_usage 回 provider_id=opencode、
  model_name=deepseek-v4.1-flash;AgentRequest 无 model 字段。模型是 profile 级。
  → 甲案:所有智能体共用 ai-gateway 一个 profile ⇒ 共用同一模型,配置页模型栏只读。
- role:"system"【被接受且生效】:塞入「只用英文回答」后模型真的用英文
  → params.system_suffix 走注入 system 消息。
- SSE 事件谱已录制 → tests/fixtures/sse-tool-flow.sse + docs/qwenpaw-agent-api-sse.md

后端:
- app/agents.py:AgentStore(data/agents.yaml,600/原子写/mtime 热载)+ 严格校验
  (slug 三重查重:模型 id/别名/QwenPaw profile;params 4 键白名单;suffix 限长 2000;
   max_agents=5 软上限,归档不计入;restore 可恢复)
- app/adapters/qwenpaw_engine.py:OpenAI 形状 ↔ AgentScope Agent API + SSE 翻译器
  (text 累积、delta:false 快照忽略、plugin_call→轨迹、call_id 配对 output、turn_usage→计量)
- app/agents_routes.py:GET/POST/PUT/DELETE /agents、/restore、/chat(SSE)、/stop、
  /runs、/sessions、/test + 新增 GET /auth/me(灰度通道,features.agents 后端权威);
  权限唯一入口 can_use_agent();防自指(引擎模型不得指向 agent/*)
- app/history.py:会话类别扩展支持 agent_<slug>(复用现有存储)
- data/agents.yaml(配置,入库);data/agent_runs.json(运行时,gitignore)

前端:
- views/Agents.vue(列表 + 对话/配置/运行记录)+ lib/agents.ts(SSE 解析)
- 菜单核心组第 5 位 + 图标 bot;router meta.admin 双层守卫;i18n 中英对称 871/871

上线实测:真实对话跑通 —— 让智能体 read_file 读 /etc/hostname,前端正确显示
「执行了 1 步」+ 工具参数、正文、计量(wjd · 47196 tokens · 缓存 13696)、终态;
引擎调用经 wangjinding 过网关(计量闭环)。

测试 209 → 217 全过(新增 8 条,含用真实夹具验证 SSE 翻译器);ci_check 全绿。
2026-09-29 13:22:12 +08:00

321 lines
13 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""QwenPaw 引擎适配器(第三十八轮 · 路线 A)。
把网关的对话请求转成 QwenPaw 的 **AgentScope Agent API** 形状,再把它的 SSE 事件流
翻译回网关前端认识的形式。
协议事实全部来自实测 —— 见 `docs/qwenpaw-agent-api-sse.md` 与
`tests/fixtures/sse-tool-flow.sse`。要点:
- `POST {base}/api/console/chat`,body `{input: [Message], session_id, stream: true}`
- 头 `X-Agent-Id: <profile>` 选择 QwenPaw 侧用哪个智能体 profile
- 回 SSE(前缀固定 `data: `,**没有 event: 行**)
- 事件:`text` / `plugin_call` / `plugin_call_output` / `data` / `message` / `response` / `turn_usage`
- `role: "system"` 的输入消息**被接受且生效**(实测)→ `system_suffix` 走这条
- `model` 字段**被忽略**(实测)→ 模型是 profile 级,不在请求里指定
网关对前端的输出形状(自定义、简单):
```jsonc
{"type": "text", "text": "增量"} // 正文增量
{"type": "step", "id": "...", "name": "read_file", "args": {...}, "status": "running"}
{"type": "step", "id": "...", "status": "done", "ms": 320, "output": "..."}
{"type": "usage", "prompt": 1, "completion": 2, "cached": 0, "model": "wjd", "context_ratio": 0.02}
{"type": "done", "status": "completed"}
{"type": "error", "message": "..."}
```
这样前端不必理解 AgentScope 的事件细节。
"""
from __future__ import annotations
import json
import logging
from typing import Any, AsyncIterator
import httpx
log = logging.getLogger("api-gateway.agents")
#: 单轮对话的超时(秒)。引擎可能跑很久(多步工具),给足。
DEFAULT_TIMEOUT = 300.0
#: 引擎不可达时的兜底文案。
ENGINE_DOWN_MSG = "引擎不可达(QwenPaw 未运行或未就绪)"
class EngineError(Exception):
"""引擎层错误(映射为 502)。"""
# ── 请求构造 ──────────────────────────────────────────────────────────
def build_input(messages: list[dict[str, Any]], system_suffix: str = "") -> list[dict[str, Any]]:
"""把网关侧的消息列表转成 AgentScope 的 `input`。
- 网关历史里的 role/文本 → `{role, content:[{type:"text", text}]}`
- `system_suffix` → 在最前面插一条 **system 消息**(实测被接受且生效)
- `system` 角色的历史消息也照传(引擎接受)
"""
out: list[dict[str, Any]] = []
if system_suffix:
out.append({
"role": "system",
"content": [{"type": "text", "text": system_suffix}],
})
for m in messages:
role = (m.get("role") or "user").lower()
if role not in ("user", "assistant", "system"):
role = "user"
content = m.get("content")
# content 允许是字符串或 [{type:text,text}] 形状
if isinstance(content, str):
parts = [{"type": "text", "text": content}]
elif isinstance(content, list):
parts = []
for c in content:
if isinstance(c, str):
parts.append({"type": "text", "text": c})
elif isinstance(c, dict) and c.get("type") == "text":
parts.append({"type": "text", "text": str(c.get("text") or "")})
# 图片等一期不支持,静默跳过(计划「一期不做图片/附件」)
else:
parts = [{"type": "text", "text": str(content or "")}]
out.append({"role": role, "content": parts})
return out
def build_body(messages: list[dict[str, Any]], session_id: str,
system_suffix: str = "") -> dict[str, Any]:
return {
"input": build_input(messages, system_suffix),
"session_id": session_id,
"stream": True,
}
# ── SSE 解析与翻译 ────────────────────────────────────────────────────
def parse_sse_line(line: str) -> dict[str, Any] | None:
"""把一行 SSE 解析成事件 dict;非数据行返回 None。"""
if not line.startswith("data:"):
return None
payload = line[5:].strip()
if not payload or payload == "[DONE]":
return None
try:
obj = json.loads(payload)
except json.JSONDecodeError:
log.debug("引擎 SSE 非 JSON 行(忽略):%.120s", payload)
return None
return obj if isinstance(obj, dict) else None
class StreamTranslator:
"""把引擎的事件序列翻译成网关前端的消息序列(有状态)。
状态:
- `_text_buf`: 已累积的正文(用来在结束时兜底补全)
- `_steps`: call_id → 步骤信息(配对 plugin_call / plugin_call_output)
"""
def __init__(self) -> None:
self.text_parts: list[str] = []
self.steps: dict[str, dict[str, Any]] = {}
self.usage: dict[str, Any] | None = None
self.final_status: str = "completed"
self._step_started: dict[str, float] = {}
# ── 单事件 → 若干前端消息 ──
def feed(self, ev: dict[str, Any], now: float | None = None) -> list[dict[str, Any]]:
import time
now = now or time.time()
t = ev.get("type")
out: list[dict[str, Any]] = []
if t == "text":
# delta:true 是增量;delta:false 是整段快照 —— 忽略,否则会重复
if ev.get("delta") is False:
return out
piece = ev.get("text")
if isinstance(piece, str) and piece:
self.text_parts.append(piece)
out.append({"type": "text", "text": piece})
return out
if t == "plugin_call":
info = _extract_call(ev, key="arguments")
if info:
cid = info.get("call_id") or info.get("name") or "?"
self.steps[cid] = info
self._step_started[cid] = now
out.append({
"type": "step", "id": cid, "name": info.get("name"),
"args": info.get("args"), "status": "running",
})
return out
if t == "plugin_call_output":
info = _extract_call(ev, key="output")
if info:
cid = info.get("call_id") or info.get("name") or "?"
started = self._step_started.pop(cid, None)
ms = int((now - started) * 1000) if started else None
if cid in self.steps:
self.steps[cid]["status"] = "done"
msg: dict[str, Any] = {"type": "step", "id": cid, "status": "done"}
if ms is not None:
msg["ms"] = ms
if info.get("output") is not None:
msg["output"] = _clip(info["output"])
out.append(msg)
return out
if t == "turn_usage":
u = ev.get("usage") or {}
self.usage = u
ctx = ev.get("context_usage") or {}
out.append({
"type": "usage",
"prompt": int(u.get("prompt_tokens") or 0),
"completion": int(u.get("completion_tokens") or 0),
"cached": int(u.get("cache_read_tokens") or 0),
"model": u.get("model_name") or "",
"provider": u.get("provider_id") or "",
"context_ratio": ctx.get("context_usage_ratio"),
})
return out
if t == "response":
st = ev.get("status")
if st:
self.final_status = st
if st in ("failed", "cancelled"):
out.append({"type": "done", "status": st})
return out
# data / message 等其余事件:忽略(分片由快照覆盖)
return out
def finish(self) -> list[dict[str, Any]]:
"""流结束时收尾。"""
out: list[dict[str, Any]] = []
if self.final_status not in ("failed", "cancelled"):
out.append({"type": "done", "status": self.final_status})
elif self.final_status == "cancelled":
out.append({"type": "done", "status": "cancelled"})
return out
@property
def text(self) -> str:
return "".join(self.text_parts)
def _extract_call(ev: dict[str, Any], *, key: str) -> dict[str, Any] | None:
"""从 plugin_call / plugin_call_output 里取 {call_id, name, args|output}。
结构(实测):`content[0].data = {call_id, name, arguments|output}`
注意 `arguments` 是 **JSON 字符串**,要二次解析。
"""
content = ev.get("content")
if not isinstance(content, list):
return None
for c in content:
if not isinstance(c, dict):
continue
data = c.get("data")
if not isinstance(data, dict) or key not in data:
continue
info: dict[str, Any] = {
"call_id": data.get("call_id"),
"name": data.get("name"),
}
raw = data.get(key)
if key == "arguments":
info["args"] = _maybe_json(raw)
else:
info["output"] = raw
return info
return None
def _maybe_json(v: Any) -> Any:
"""`arguments` 是 JSON 字符串 → 解析成对象;解析不了就原样返回字符串。"""
if not isinstance(v, str):
return v
try:
return json.loads(v)
except (json.JSONDecodeError, TypeError):
return v
def _clip(v: Any, limit: int = 4000) -> Any:
"""工具输出可能很长 —— 转发给前端前裁一刀(避免 SSE 帧过大)。"""
if isinstance(v, str) and len(v) > limit:
return v[:limit] + "…(已截断)"
return v
# ── 调用引擎 ──────────────────────────────────────────────────────────
async def stream_chat(base_url: str, agent_id: str, messages: list[dict[str, Any]],
session_id: str, *, system_suffix: str = "",
timeout: float = DEFAULT_TIMEOUT) -> AsyncIterator[dict[str, Any]]:
"""打引擎并产出"已翻译"的前端消息。
用法:
async for msg in stream_chat(...): # msg 就是 {type: text|step|usage|done|error}
yield sse(msg)
"""
url = base_url.rstrip("/") + "/api/console/chat"
body = build_body(messages, session_id, system_suffix)
tr = StreamTranslator()
headers = {
"Content-Type": "application/json",
"Accept": "text/event-stream",
"X-Agent-Id": agent_id,
}
try:
async with httpx.AsyncClient(timeout=httpx.Timeout(timeout, connect=10.0)) as client:
async with client.stream("POST", url, headers=headers, json=body) as resp:
if resp.status_code >= 400:
raw = (await resp.aread()).decode("utf-8", "replace")[:300]
yield {"type": "error",
"message": f"引擎返回 {resp.status_code}:{raw}"}
return
async for line in resp.aiter_lines():
ev = parse_sse_line(line)
if ev is None:
continue
for msg in tr.feed(ev):
yield msg
for msg in tr.finish():
yield msg
except httpx.ConnectError:
yield {"type": "error", "message": ENGINE_DOWN_MSG}
except httpx.ReadTimeout:
yield {"type": "error", "message": "引擎响应超时(本轮未完成)"}
except httpx.RemoteProtocolError:
# 引擎中途断开 —— 保留已收到的内容(前端自己累积着),只提示中断
for msg in tr.finish():
pass
yield {"type": "error", "message": "引擎连接中断,本轮未完成"}
except Exception as exc: # noqa: BLE001 —— 兜底,别让流把异常抛给客户端
log.exception("引擎调用失败")
yield {"type": "error", "message": f"引擎调用失败:{exc}"}
async def check_engine(base_url: str, *, timeout: float = 5.0) -> tuple[bool, str]:
"""探活:引擎是否可达。返回 (ok, 说明)。
不引 `/api/healthz`(那是 QwenPaw 的)—— 用最轻的请求确认端口活着。
"""
url = base_url.rstrip("/") + "/api/healthz"
try:
async with httpx.AsyncClient(timeout=timeout) as client:
r = await client.get(url)
if r.status_code < 500:
return True, "就绪"
return False, f"引擎异常(HTTP {r.status_code})"
except httpx.ConnectError:
return False, ENGINE_DOWN_MSG
except Exception as exc: # noqa: BLE001
return False, f"引擎不可达:{exc}"