mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 04:31:58 +00:00
ThearchyHelios
7255ae98a5
- 「不指定接入」交给设置（核查模型、裁判模型）时：启用的接入全是演示类，设置这一级不管写的是什么，执行时落到的都是 演示模型，说明照实带上「（演示用，不调用真实模型）」、候选带 demo 标记；一个启用的接入都没有时不算落得稳（同别处 「换成内置演示模型」不算）。以前 land() 在判演示之前就把交给设置的一处当成落稳返回了； - 检查器里没有下拉框、发布前检查又给不出修复时，提示末尾补「无法改用其他接入：」加第一处落不稳的地方（Landing.why， 比如「成员「研究员」指定的模型「小模型」所属的模型接入「旧接入」已停用」），不再只剩「核对名称」让人去设置里找一个 找不到的名字； - 点名的接入名恰好是某个启用接入的模型：画布、发布前检查按这一处在检查器里有哪几栏写（provider_refs._model_note）—— 模型调用一类在节点的「模型」中选择它，接入在发布前检查里改用这个接入或选择「不指定接入」（修复里有哪项说哪项）； 协作成员、全图默认只说可在发布前检查里改用；核查节点、报告自己的结论句裁判两栏都在，照设置页写这两栏的叫法。 不再叫人去一个没有的「接入」栏留空。运行时报错不分节点类型，「填写」改为「选择它」； - 修复候选修完以后落到演示类接入的带 demo 标记，发布弹窗据此在选项名旁加醒目的标签； - 性能：改一个节点的配置只影响它自己取模型的那几处，修复规划（autofix._ProviderPlan）每试一个候选只重算这个节点 （provider_refs.node_uses），不再深拷贝、重算整张图；问题提示只问给不给得出（provider_fixable，找到第一个落得稳的 就停）；接入问题的引用整张图只算一遍。60 个节点都点名同一个停用接入时，画布校验里这一部分约 0.35 秒降到约 0.002 秒， 发布前检查约 0.72 秒降到约 0.01 秒，一次应用 30 条修复约 2.7 秒降到约 0.09 秒； - 助手：GET /api/copilot/model 另给 panel_error（写在助手栏「助手使用的模型」那一格下面，去向写「上方」，不把人指回这个 下拉框本身）；接入名恰好是模型名时并成一句，不再两次提「助手使用的模型」；PUT /api/settings 里助手那一项的 422 用 助手报错那一句（那里只有一个选模型的下拉框，没有「换一个接入」的地方）； - 裁判设置的格式检查写「留空表示跟随助手的模型」，同设置页空选项（engine/judge.py 一行，同核查那条的改法）；「交给助手」 后只剩接入问题时写「助手无法处理」，有修复时另起半句「也可在各条问题下方选择改用的模型接入」。 测试：test_provider_ref_wording 加演示落在设置上、demo 标记、给不出修复时卡在哪里（含一个启用接入都没有）、各处的 接入名恰好是模型名、助手栏那一格的说法、助手那一项的 422、裁判格式检查、规划开销随节点数线性增长（数试算时算了多少处 取模型的地方）；其余几个文件跟上新说法。新加的核心用例都做过去掉实现、测试变红的验证。
861 lines
40 KiB
Python
861 lines
40 KiB
Python
from __future__ import annotations
|
||
|
||
import math
|
||
import os
|
||
from typing import Any
|
||
|
||
from fastapi import APIRouter, Depends, HTTPException
|
||
from pydantic import BaseModel, Field
|
||
from sqlalchemy import select
|
||
from sqlalchemy.ext.asyncio import AsyncSession
|
||
|
||
from app.api import health
|
||
from app.api.coded import CodedHTTPException
|
||
from app.core.errors import explain, not_configured, raw
|
||
from app.core.config import settings as app_settings
|
||
from app.core.crypto import encrypt, mask
|
||
from app.db.base import get_session
|
||
from app.db.models import Provider, Setting
|
||
from app.engine.judge import JUDGE_DEFAULTS, MAX_TOKENS as JUDGE_MAX_TOKENS
|
||
from app.engine.judge import SETTING_GROUP as JUDGE_GROUP, SETTING_LABEL as JUDGE_LABEL, SPEND_KEY as JUDGE_SPEND
|
||
from app.engine.judge import daily_spend, judge_model_spec, judge_settings
|
||
from app.engine.judge import sanitize_settings as judge_values, settings_problem as judge_problem
|
||
from app.engine.spend import SPEND_KEY as VERIFY_SPEND, daily_spend as verify_daily_spend
|
||
from app.providers import catalog
|
||
from app.providers.factory import (
|
||
ModelSpec, ProviderNotConfigured, _resolve_key, assistant_unavailable, build_chat_model,
|
||
list_providers as all_providers, model_note, named_provider,
|
||
)
|
||
|
||
router = APIRouter(prefix="/api", tags=["settings"])
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# Provider
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
class ProviderIn(BaseModel):
|
||
name: str = Field(min_length=1, max_length=100)
|
||
kind: str = "anthropic"
|
||
base_url: str | None = None
|
||
api_key: str | None = None
|
||
models: list[dict[str, Any]] = Field(default_factory=list)
|
||
default_model: str | None = None
|
||
enabled: bool = True
|
||
extra: dict[str, Any] = Field(default_factory=dict)
|
||
|
||
|
||
class ProviderPatch(BaseModel):
|
||
name: str | None = None
|
||
base_url: str | None = None
|
||
api_key: str | None = None # 传空字符串表示清空；不传表示不动
|
||
models: list[dict[str, Any]] | None = None
|
||
default_model: str | None = None
|
||
enabled: bool | None = None
|
||
extra: dict[str, Any] | None = None
|
||
|
||
|
||
class ProviderOut(BaseModel):
|
||
id: str
|
||
name: str
|
||
kind: str
|
||
base_url: str | None
|
||
default_model: str | None
|
||
models: list[dict[str, Any]]
|
||
enabled: bool
|
||
extra: dict[str, Any]
|
||
# 只回掩码，明文 key 永远不出后端
|
||
api_key_masked: str = ""
|
||
has_key: bool = False
|
||
#: 最近一次测连接（见 app/api/health.py）。没测过、或者连接配置改过之后都是 None
|
||
last_checked_at: str | None = None
|
||
last_check_ok: bool | None = None
|
||
last_latency_ms: int | None = None
|
||
last_error: str | None = None
|
||
|
||
|
||
#: extra 里这一项是后端记的最近一次测连接结果，不是用户填的配置：不回显、
|
||
#: 不接受客户端写入，保存 extra 时原样留着
|
||
LAST_CHECK = "last_check"
|
||
|
||
|
||
def _user_extra(extra: dict[str, Any] | None) -> dict[str, Any]:
|
||
return {k: v for k, v in (extra or {}).items() if k != LAST_CHECK}
|
||
|
||
|
||
def _with_saved_headers(extra: dict[str, Any], saved: dict[str, Any] | None) -> dict[str, Any]:
|
||
"""请求头的值回显的是 ***：原样带回来的，换回存着的真值。
|
||
|
||
设置页保存时会把读到的 extra 整块带回来。以前照单全收，于是改个名字、
|
||
勾一下启用，自定义请求头就全变成了字面上的 ***，接口从此 401。
|
||
"""
|
||
headers = extra.get("headers")
|
||
if not isinstance(headers, dict):
|
||
return extra
|
||
old = (saved or {}).get("headers") or {}
|
||
return {**extra, "headers": {k: (old.get(k, v) if v == "***" else v) for k, v in headers.items()}}
|
||
|
||
|
||
def _connection(provider: Provider, model: str | None = None) -> str:
|
||
"""连接配置的指纹：类型、地址、实际用的钥匙、请求头这些、测的哪个模型。"""
|
||
return health.fingerprint(
|
||
"provider", provider.kind, (provider.base_url or "").strip(), _resolve_key(provider) or "",
|
||
_user_extra(provider.extra), model or provider.default_model or "",
|
||
)
|
||
|
||
|
||
def _set_last_check(row: Provider, last: dict[str, Any] | None) -> None:
|
||
row.extra = {**_user_extra(row.extra), **({LAST_CHECK: last} if last else {})}
|
||
|
||
|
||
def _to_out(row: Provider) -> ProviderOut:
|
||
extra = _user_extra(row.extra)
|
||
return ProviderOut(
|
||
id=row.id,
|
||
name=row.name,
|
||
kind=row.kind,
|
||
base_url=row.base_url,
|
||
default_model=row.default_model,
|
||
models=row.models or [],
|
||
enabled=row.enabled,
|
||
extra={k: v for k, v in extra.items() if k != "headers"} | (
|
||
{"headers": {k: "***" for k in extra.get("headers", {})}}
|
||
if extra.get("headers")
|
||
else {}
|
||
),
|
||
api_key_masked=mask(row.api_key),
|
||
has_key=bool(row.api_key),
|
||
**health.fields((row.extra or {}).get(LAST_CHECK)),
|
||
)
|
||
|
||
|
||
@router.get("/providers", response_model=list[ProviderOut])
|
||
async def list_providers(session: AsyncSession = Depends(get_session)) -> list[ProviderOut]:
|
||
rows = (await session.execute(select(Provider).order_by(Provider.created_at))).scalars()
|
||
return [_to_out(r) for r in rows]
|
||
|
||
|
||
@router.get("/providers/catalog")
|
||
async def provider_catalog() -> dict[str, Any]:
|
||
"""内置的 provider 预设和模型目录，前端用它渲染"添加 Provider"表单。"""
|
||
return {
|
||
"kinds": catalog.PROVIDER_KINDS,
|
||
"env_detected": {
|
||
"ANTHROPIC_API_KEY": bool(os.environ.get("ANTHROPIC_API_KEY")),
|
||
"ANTHROPIC_AUTH_TOKEN": bool(os.environ.get("ANTHROPIC_AUTH_TOKEN")),
|
||
"ANTHROPIC_BASE_URL": os.environ.get("ANTHROPIC_BASE_URL") or None,
|
||
"OPENAI_API_KEY": bool(os.environ.get("OPENAI_API_KEY")),
|
||
"TAVILY_API_KEY": bool(os.environ.get("TAVILY_API_KEY")),
|
||
},
|
||
}
|
||
|
||
|
||
@router.post("/providers", response_model=ProviderOut, status_code=201)
|
||
async def create_provider(
|
||
payload: ProviderIn, session: AsyncSession = Depends(get_session)
|
||
) -> ProviderOut:
|
||
problem = _config_problem(payload.kind, payload.base_url)
|
||
if problem:
|
||
raise HTTPException(400, problem)
|
||
exists = (
|
||
await session.execute(select(Provider).where(Provider.name == payload.name))
|
||
).scalar_one_or_none()
|
||
if exists:
|
||
raise HTTPException(409, f"已存在名为「{payload.name}」的模型接入，请换一个名称")
|
||
|
||
row = Provider(
|
||
name=payload.name,
|
||
kind=payload.kind,
|
||
base_url=payload.base_url or None,
|
||
api_key=encrypt(payload.api_key),
|
||
models=payload.models,
|
||
default_model=payload.default_model,
|
||
enabled=payload.enabled,
|
||
extra=_user_extra(payload.extra),
|
||
)
|
||
# 表单里刚测过这份配置：保存下来就带着那次结果
|
||
_set_last_check(row, health.draft_result(_connection(row)))
|
||
session.add(row)
|
||
await session.commit()
|
||
await session.refresh(row)
|
||
return _to_out(row)
|
||
|
||
|
||
@router.patch("/providers/{provider_id}", response_model=ProviderOut)
|
||
async def update_provider(
|
||
provider_id: str, payload: ProviderPatch, session: AsyncSession = Depends(get_session)
|
||
) -> ProviderOut:
|
||
row = await session.get(Provider, provider_id)
|
||
if not row:
|
||
raise HTTPException(404, "模型接入不存在，可能已被删除")
|
||
|
||
data = payload.model_dump(exclude_unset=True)
|
||
before, last = _connection(row), (row.extra or {}).get(LAST_CHECK)
|
||
if "api_key" in data:
|
||
row.api_key = encrypt(data.pop("api_key")) if data["api_key"] else None
|
||
data.pop("api_key", None)
|
||
if data.get("extra") is not None:
|
||
data["extra"] = _with_saved_headers(_user_extra(data["extra"]), row.extra)
|
||
else:
|
||
data.pop("extra", None)
|
||
for key, value in data.items():
|
||
setattr(row, key, value)
|
||
after = _connection(row)
|
||
if after != before:
|
||
# 测过的已经不是这份配置了：旧结果作废，表单里测过这一份的话换成那一次
|
||
last = health.draft_result(after)
|
||
_set_last_check(row, last)
|
||
await session.commit()
|
||
await session.refresh(row)
|
||
return _to_out(row)
|
||
|
||
|
||
@router.delete("/providers/{provider_id}", status_code=204)
|
||
async def delete_provider(
|
||
provider_id: str, session: AsyncSession = Depends(get_session)
|
||
) -> None:
|
||
row = await session.get(Provider, provider_id)
|
||
if not row:
|
||
raise HTTPException(404, "模型接入不存在，可能已被删除")
|
||
await session.delete(row)
|
||
await session.commit()
|
||
|
||
|
||
class TestIn(BaseModel):
|
||
model: str | None = None
|
||
prompt: str = "用一句话介绍你自己"
|
||
|
||
|
||
def _config_problem(kind: str, base_url: str | None) -> str | None:
|
||
"""保存或测试前就能看出来的配置错误。"""
|
||
if kind not in {k["kind"] for k in catalog.PROVIDER_KINDS}:
|
||
return f"不支持「{kind}」接入类型"
|
||
if kind == "openai_compatible" and not (base_url or "").strip():
|
||
return "「OpenAI 兼容」必须填写 Base URL，例如 https://api.deepseek.com/v1"
|
||
return None
|
||
|
||
|
||
#: 测试连接最多等多久。真正跑节点用的是 model_timeout_seconds（默认 5 分钟），
|
||
#: 弹窗里转 5 分钟等于没有反馈
|
||
_TEST_TIMEOUT = 30
|
||
|
||
|
||
async def _try_model(provider: Provider, model: str | None, prompt: str) -> dict[str, Any]:
|
||
"""真发一次请求。失败也是正常结果，照实返回原因，不抛。"""
|
||
import asyncio
|
||
import time
|
||
|
||
from app.engine.state import message_text
|
||
|
||
model = model or provider.default_model
|
||
if not model:
|
||
return {"ok": False, "error": "没有可测试的模型",
|
||
"hint": "请先在「可选模型」中添加一个，或填写默认模型", "detail": ""}
|
||
try:
|
||
chat = build_chat_model(provider, ModelSpec(model=model, max_tokens=256,
|
||
timeout=_TEST_TIMEOUT))
|
||
except ProviderNotConfigured as e:
|
||
if e.code == "missing_key":
|
||
return {"ok": False, "error": "尚未填写 API Key",
|
||
"hint": "请填写 API Key 后再测试，也可在服务端的环境变量中配置", "detail": str(e)}
|
||
return {"ok": False, "error": not_configured(e, with_hint=False), "hint": "", "detail": str(e)}
|
||
|
||
started = time.perf_counter()
|
||
try:
|
||
async with asyncio.timeout(_TEST_TIMEOUT):
|
||
reply = await chat.ainvoke(prompt)
|
||
except Exception as e: # noqa: BLE001
|
||
reason, hint = explain(e)
|
||
return {"ok": False, "error": f"测试未通过：{reason}", "hint": hint, "detail": raw(e),
|
||
"model": model}
|
||
|
||
usage = getattr(reply, "usage_metadata", None) or {}
|
||
return {
|
||
"ok": True,
|
||
"latency_ms": int((time.perf_counter() - started) * 1000),
|
||
"model": model,
|
||
"reply": message_text(reply)[:500],
|
||
"usage": usage,
|
||
}
|
||
|
||
|
||
@router.post("/providers/{provider_id}/test")
|
||
async def test_provider(
|
||
provider_id: str, payload: TestIn, session: AsyncSession = Depends(get_session)
|
||
) -> dict[str, Any]:
|
||
"""真发一次请求验证配置。设置页的"测试连接"按钮调它。"""
|
||
row = await session.get(Provider, provider_id)
|
||
if not row:
|
||
raise HTTPException(404, "模型接入不存在，可能已被删除")
|
||
result = await _try_model(row, payload.model, payload.prompt)
|
||
# 只记默认模型的结果：拿别的模型测出 invalid_model，不等于这个接入连不上
|
||
if (payload.model or row.default_model) == row.default_model:
|
||
_set_last_check(row, _test_record(result))
|
||
await session.commit()
|
||
return result
|
||
|
||
|
||
def _test_record(result: dict[str, Any]) -> dict[str, Any]:
|
||
return health.record(result.get("ok"), result.get("latency_ms"), result.get("error"))
|
||
|
||
|
||
class ProviderDraftIn(BaseModel):
|
||
"""一份还没保存（或改了还没保存）的接入配置。带 id 表示在编辑已有的那个。"""
|
||
|
||
id: str | None = None
|
||
name: str = "草稿"
|
||
kind: str = "anthropic"
|
||
base_url: str | None = None
|
||
api_key: str | None = None # 编辑时留空=沿用已保存的
|
||
models: list[dict[str, Any]] = Field(default_factory=list)
|
||
default_model: str | None = None
|
||
extra: dict[str, Any] = Field(default_factory=dict)
|
||
#: 这次测哪个模型，不填用 default_model
|
||
model: str | None = None
|
||
prompt: str = "用一句话介绍你自己"
|
||
|
||
|
||
def _draft_provider(payload: ProviderDraftIn, saved: Provider | None) -> Provider:
|
||
"""拼一个不进会话的临时接入点，只拿来试一下。
|
||
|
||
编辑时 Key 框留空表示不改、请求头回显的是 ***——测试都得换回存着的真值，
|
||
否则改个 Base URL 想测一下，只会得到一句"密钥不对"。
|
||
"""
|
||
extra = _user_extra(payload.extra)
|
||
if saved:
|
||
extra = _with_saved_headers(extra, saved.extra)
|
||
return Provider(
|
||
name=payload.name or (saved.name if saved else "草稿"),
|
||
kind=payload.kind,
|
||
base_url=(payload.base_url or "").strip() or None,
|
||
api_key=encrypt(payload.api_key) if payload.api_key else (saved.api_key if saved else None),
|
||
models=payload.models,
|
||
default_model=payload.default_model,
|
||
enabled=True,
|
||
extra=extra,
|
||
)
|
||
|
||
|
||
async def _draft(payload: ProviderDraftIn, session: AsyncSession) -> Provider:
|
||
saved = await session.get(Provider, payload.id) if payload.id else None
|
||
return _draft_provider(payload, saved)
|
||
|
||
|
||
@router.post("/providers/test")
|
||
@router.post("/settings/providers/test", include_in_schema=False)
|
||
async def test_provider_draft(
|
||
payload: ProviderDraftIn, session: AsyncSession = Depends(get_session)
|
||
) -> dict[str, Any]:
|
||
"""测一份没保存的配置，不落库。
|
||
|
||
以前只能测已保存的：Base URL 填错要保存→关弹窗→点测试→看提示→再打开改。
|
||
接内网网关这种填错一个字就连不上的东西，来回四五趟。
|
||
"""
|
||
problem = _config_problem(payload.kind, payload.base_url)
|
||
if problem:
|
||
return {"ok": False, "error": problem, "hint": "", "detail": ""}
|
||
saved = await session.get(Provider, payload.id) if payload.id else None
|
||
draft = _draft_provider(payload, saved)
|
||
result = await _try_model(draft, payload.model, payload.prompt)
|
||
last, key = _test_record(result), _connection(draft, payload.model)
|
||
health.remember_draft(key, last)
|
||
if saved is not None and key == _connection(saved):
|
||
# 编辑框里没改连接、只是再测一次：测的就是已保存的那份
|
||
_set_last_check(saved, last)
|
||
await session.commit()
|
||
return result
|
||
|
||
|
||
@router.post("/providers/models")
|
||
async def probe_provider_models(
|
||
payload: ProviderDraftIn, session: AsyncSession = Depends(get_session)
|
||
) -> dict[str, Any]:
|
||
"""问这个接入点有哪些模型，省得手敲模型 id。
|
||
|
||
知识库那边的探测不带鉴权，接需要 Key 的网关时必然 401，所以这里单独做一份：
|
||
Key、自定义请求头都按真正调用时的方式带上。
|
||
"""
|
||
import httpx
|
||
|
||
problem = _config_problem(payload.kind, payload.base_url)
|
||
if problem:
|
||
return {"ok": False, "error": problem, "hint": "", "detail": "", "models": []}
|
||
provider = await _draft(payload, session)
|
||
if provider.kind == "mock":
|
||
return {"ok": True, "models": [m["id"] for m in catalog.MOCK_MODELS], "url": ""}
|
||
|
||
extra = dict(provider.extra or {})
|
||
# 和真正调用时取同一把钥匙：只有官方 OpenAI / Anthropic 才退回环境变量。
|
||
# 兼容网关是别人家的服务，Key 框空着就不带——否则点一下「拉取模型」，
|
||
# 环境里那把 OpenAI 的钥匙就发给了表单上随便填的一个地址
|
||
key = _resolve_key(provider) or ""
|
||
headers: dict[str, str] = {}
|
||
if provider.kind == "anthropic":
|
||
base = provider.base_url or os.environ.get("ANTHROPIC_BASE_URL") or "https://api.anthropic.com"
|
||
url = base.rstrip("/") + "/v1/models"
|
||
headers["anthropic-version"] = "2023-06-01"
|
||
if key:
|
||
if extra.get("auth_style") == "bearer":
|
||
headers["Authorization"] = f"Bearer {key}"
|
||
else:
|
||
headers["x-api-key"] = key
|
||
else:
|
||
url = (provider.base_url or "https://api.openai.com/v1").rstrip("/") + "/models"
|
||
if key:
|
||
headers["Authorization"] = f"Bearer {key}"
|
||
headers.update({k: str(v) for k, v in (extra.get("headers") or {}).items()})
|
||
|
||
try:
|
||
async with httpx.AsyncClient(timeout=10) as client:
|
||
resp = await client.get(url, headers=headers)
|
||
resp.raise_for_status()
|
||
data = resp.json()
|
||
except Exception as e: # noqa: BLE001
|
||
reason, hint = explain(e)
|
||
return {"ok": False, "error": f"未能获取模型列表：{reason}", "hint": hint,
|
||
"detail": raw(e), "models": [], "url": url}
|
||
models = [m.get("id", "") for m in (data.get("data") or []) if isinstance(m, dict) and m.get("id")]
|
||
return {"ok": True, "models": models, "url": url}
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 核查：双盲复算、反方审稿、审批质询共用的一组设置
|
||
# --------------------------------------------------------------------------
|
||
|
||
#: 设置里 verify 这一组。三项都是让一个只读 agent 查库核对（engine/verify_agent.py），共用核查模型和每日金额上限；
|
||
#: 各自的开关和上限分开放。上限写 null 表示不限（和裁判一组的约定一样）
|
||
VERIFY_GROUP = "verify"
|
||
VERIFY_DEFAULTS: dict[str, Any] = {
|
||
"provider": None, # 核查模型：空着沿用证据裁判模型的取值链（verify_model_spec）
|
||
"model": None,
|
||
# 双盲复算：tolerance 是允许的相对误差，两边的值差在这个比例以内算一致
|
||
"dual_check": {"enabled": True, "tolerance": 1e-6, "max_queries": 6, "timeout_s": 120},
|
||
# 反方审稿：max_claims 是每份报告最多挑几句关键结论来审，查询次数、时长按每句算
|
||
"challenge": {"enabled": True, "max_claims": 5, "max_queries": 4, "timeout_s": 90},
|
||
# 审批质询：查询次数、时长按每次提问算
|
||
"inquiry": {"enabled": True, "max_queries": 6, "timeout_s": 120},
|
||
"daily_max_usd": 2.0, # 全局每日核查金额上限（按本地日期），和裁判的每日上限分开记账（engine/spend.py）
|
||
}
|
||
#: 三项各自那一组的键
|
||
_VERIFY_PARTS = ("dual_check", "challenge", "inquiry")
|
||
_VERIFY_COUNTS = frozenset({"max_queries", "max_claims"})
|
||
|
||
#: 设置页「核查」一节各项的叫法：报错时写它，不写键名。嵌套的项写成「组.项」
|
||
VERIFY_SETTING_LABEL = {
|
||
"provider": "核查模型 · 接入", "model": "核查模型 · 模型",
|
||
"daily_max_usd": "核查每日金额上限（美元）",
|
||
"dual_check": "双盲复算", "dual_check.enabled": "双盲复算 · 启用",
|
||
"dual_check.tolerance": "双盲复算 · 允许的相对误差", "dual_check.max_queries": "双盲复算 · 每项最多查询次数",
|
||
"dual_check.timeout_s": "双盲复算 · 每项时长上限（秒）",
|
||
"challenge": "反方审稿", "challenge.enabled": "反方审稿 · 启用",
|
||
"challenge.max_claims": "反方审稿 · 每份报告最多审几句", "challenge.max_queries": "反方审稿 · 每句最多查询次数",
|
||
"challenge.timeout_s": "反方审稿 · 每句时长上限（秒）",
|
||
"inquiry": "审批质询", "inquiry.enabled": "审批质询 · 启用",
|
||
"inquiry.max_queries": "审批质询 · 每次最多查询次数", "inquiry.timeout_s": "审批质询 · 每次时长上限（秒）",
|
||
}
|
||
|
||
|
||
def _positive(value: Any, *, whole: bool) -> bool:
|
||
"""上限：null（不限），或大于 0 的有限数；whole 时还得是整数（次数、句数）。True / False 不算数。"""
|
||
if value is None:
|
||
return True
|
||
if isinstance(value, bool) or not isinstance(value, (int, float)) or not math.isfinite(value) or value <= 0:
|
||
return False
|
||
return not (whole and int(value) != value)
|
||
|
||
|
||
def _tolerance_ok(value: Any) -> bool:
|
||
"""允许的相对误差：不小于 0 的有限数（0 表示必须完全一致）。不能是 null：没有「不限」的说法。"""
|
||
return not isinstance(value, bool) and isinstance(value, (int, float)) and math.isfinite(value) and value >= 0
|
||
|
||
|
||
def _verify_item_ok(key: str, value: Any) -> bool:
|
||
if key == "enabled":
|
||
return isinstance(value, bool)
|
||
if key == "tolerance":
|
||
return _tolerance_ok(value)
|
||
return _positive(value, whole=key in _VERIFY_COUNTS)
|
||
|
||
|
||
def _verify_item_problem(part: str, key: str, value: Any) -> str | None:
|
||
label = VERIFY_SETTING_LABEL[f"{part}.{key}"]
|
||
if _verify_item_ok(key, value):
|
||
return None
|
||
if key == "enabled":
|
||
return f"「{label}」需要选择开启或关闭，当前为「{value}」"
|
||
if key == "tolerance":
|
||
return f"「{label}」需要填写不小于 0 的数（例如 0.000001 表示百万分之一），当前为「{value}」"
|
||
words = "正整数（次数）" if key == "max_queries" else "正整数（句数）" if key == "max_claims" else "大于 0 的秒数"
|
||
return f"「{label}」需要填写{words}，留空表示不限，当前为「{value}」"
|
||
|
||
|
||
def verify_problem(value: Any) -> str | None:
|
||
"""设置页写进来的 verify 这一组有什么问题，没问题返回 None。说明写设置页上的叫法。"""
|
||
if not isinstance(value, dict):
|
||
return f"「核查」设置格式有误，需要是一个对象，当前为「{value}」"
|
||
for key, item in value.items():
|
||
if key not in VERIFY_DEFAULTS:
|
||
return (f"「核查」设置中不支持「{key}」，可设置的项："
|
||
+ "、".join(f"「{VERIFY_SETTING_LABEL[k]}」" for k in VERIFY_DEFAULTS))
|
||
label = VERIFY_SETTING_LABEL[key]
|
||
if key in ("provider", "model"):
|
||
if item is not None and not isinstance(item, str):
|
||
# 「跟随证据裁判模型」是设置页接入下拉框空选项的原话
|
||
return f"「{label}」需要填写模型接入的名称或模型 ID，留空表示跟随证据裁判模型，当前为「{item}」"
|
||
elif key == "daily_max_usd":
|
||
if not _positive(item, whole=False):
|
||
return f"「{label}」需要填写大于 0 的金额（美元），留空表示不限，当前为「{item}」"
|
||
elif not isinstance(item, dict):
|
||
return f"「{label}」设置格式有误，需要是一个对象，当前为「{item}」"
|
||
else:
|
||
allowed = VERIFY_DEFAULTS[key]
|
||
for sub, sub_value in item.items():
|
||
if sub not in allowed:
|
||
return (f"「{label}」设置中不支持「{sub}」，可设置的项："
|
||
+ "、".join(f"「{VERIFY_SETTING_LABEL[f'{key}.{k}']}」" for k in allowed))
|
||
if problem := _verify_item_problem(key, sub, sub_value):
|
||
return problem
|
||
return None
|
||
|
||
|
||
def verify_values(stored: Any) -> dict[str, Any]:
|
||
"""库里存的 verify 一组 → 补齐默认的完整设置。存坏了的项（旧版本、手改）按默认，存了 null 的上限就是不限；
|
||
三项各自那一组逐项补，只存了一项的其余照默认。"""
|
||
stored = stored if isinstance(stored, dict) else {}
|
||
out: dict[str, Any] = {k: dict(v) if isinstance(v, dict) else v for k, v in VERIFY_DEFAULTS.items()}
|
||
for key in ("provider", "model"):
|
||
value = stored.get(key)
|
||
out[key] = value.strip() or None if isinstance(value, str) else None
|
||
if "daily_max_usd" in stored and _positive(stored["daily_max_usd"], whole=False):
|
||
out["daily_max_usd"] = stored["daily_max_usd"]
|
||
for part in _VERIFY_PARTS:
|
||
saved = stored.get(part)
|
||
if not isinstance(saved, dict):
|
||
continue
|
||
for key in VERIFY_DEFAULTS[part]:
|
||
if key in saved and _verify_item_ok(key, saved[key]):
|
||
out[part][key] = saved[key]
|
||
return out
|
||
|
||
|
||
async def verify_settings(session: AsyncSession) -> dict[str, Any]:
|
||
"""设置里的 verify 一组，没存过的键补默认（读法同 judge_settings）。节点、质询接口都从这里读。"""
|
||
row = await session.get(Setting, VERIFY_GROUP)
|
||
return verify_values(row.value if row else None)
|
||
|
||
|
||
def _model_text(value: Any) -> str | None:
|
||
return value.strip() or None if isinstance(value, str) else None
|
||
|
||
|
||
async def verify_model_spec(session: AsyncSession, node_config: dict[str, Any] | None = None, *,
|
||
settings: dict[str, Any] | None = None) -> ModelSpec:
|
||
"""核查用哪个模型：节点配置里的 provider / model > 设置里的核查模型 > 证据裁判模型的取值链（裁判设置 >
|
||
助手模型 > 默认接入，见 judge.judge_model_spec）。哪一级写了就整组用哪一级，不跨级拼接入和模型。
|
||
|
||
和裁判一样关掉 thinking：核查最后要按结构交结果（强制工具调用），和思考模式不兼容。额度也和裁判一样（judge.MAX_TOKENS，
|
||
16384）：经 OpenAI 兼容接入的推理模型收不到 thinking=off，思考算在额度里，4096 常常想完就没额度查库、交结果了；
|
||
服务方嫌大时核查 agent 退回 judge.FALLBACK_MAX_TOKENS（engine/verify_agent.py）。
|
||
建议和被核查的模型不同，否则等于自己核自己；比较、提醒由调用方做（它知道被核查的是哪个模型）。
|
||
"""
|
||
node = node_config if isinstance(node_config, dict) else {}
|
||
provider, model = _model_text(node.get("provider")), _model_text(node.get("model"))
|
||
if not (provider or model):
|
||
chosen = settings if settings is not None else await verify_settings(session)
|
||
provider, model = _model_text(chosen.get("provider")), _model_text(chosen.get("model"))
|
||
if not (provider or model):
|
||
return await judge_model_spec(session, None)
|
||
return ModelSpec(provider=provider, model=model, max_tokens=JUDGE_MAX_TOKENS, thinking="off")
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 通知：定时任务推送 Webhook 时拼运行链接用的对外访问地址
|
||
# --------------------------------------------------------------------------
|
||
|
||
#: 设置里 notify 这一组。public_base_url 是别人打开 AgentLab 用的地址（https://agentlab.example.com），通知里的
|
||
#: 运行链接 = 它 + /runs/{运行 id}。服务端不知道自己在反向代理后面叫什么，只能由人填；留空的通知里不附链接
|
||
NOTIFY_GROUP = "notify"
|
||
NOTIFY_DEFAULTS: dict[str, Any] = {"public_base_url": None}
|
||
NOTIFY_SETTING_LABEL = {"public_base_url": "对外访问地址"}
|
||
|
||
|
||
def _base_url_problem(value: Any) -> str | None:
|
||
if value is None or (isinstance(value, str) and not value.strip()):
|
||
return None
|
||
label = NOTIFY_SETTING_LABEL["public_base_url"]
|
||
hint = (f"「{label}」需要是 http:// 或 https:// 开头的完整地址（例如 https://agentlab.example.com），"
|
||
f"留空表示通知里不附运行链接，当前为「{value}」")
|
||
if not isinstance(value, str):
|
||
return hint
|
||
from urllib.parse import urlsplit
|
||
|
||
try:
|
||
parts = urlsplit(value.strip())
|
||
except ValueError:
|
||
return hint
|
||
if parts.scheme not in ("http", "https") or not parts.hostname or " " in value.strip():
|
||
return hint
|
||
if parts.query or parts.fragment:
|
||
return f"「{label}」不能带 ? 或 # 之后的部分，运行链接会接在地址后面，当前为「{value}」"
|
||
return None
|
||
|
||
|
||
def notify_problem(value: Any) -> str | None:
|
||
"""设置页写进来的 notify 这一组有什么问题，没问题返回 None。"""
|
||
if not isinstance(value, dict):
|
||
return f"「通知」设置格式有误，需要是一个对象，当前为「{value}」"
|
||
for key, item in value.items():
|
||
if key not in NOTIFY_DEFAULTS:
|
||
return (f"「通知」设置中不支持「{key}」，可设置的项："
|
||
+ "、".join(f"「{NOTIFY_SETTING_LABEL[k]}」" for k in NOTIFY_DEFAULTS))
|
||
if problem := _base_url_problem(item):
|
||
return problem
|
||
return None
|
||
|
||
|
||
def notify_values(stored: Any) -> dict[str, Any]:
|
||
"""库里存的（或刚写进来的）notify 一组 → 补齐默认、规范化：去掉首尾空白和末尾的 /，空串当没填；存坏了的按没填。"""
|
||
stored = stored if isinstance(stored, dict) else {}
|
||
url = stored.get("public_base_url")
|
||
if not isinstance(url, str) or not url.strip() or _base_url_problem(url):
|
||
return {"public_base_url": None}
|
||
return {"public_base_url": url.strip().rstrip("/")}
|
||
|
||
|
||
async def notify_settings(session: AsyncSession) -> dict[str, Any]:
|
||
"""设置里的 notify 一组，没存过的补默认。定时任务推送 Webhook 时从这里读。"""
|
||
row = await session.get(Setting, NOTIFY_GROUP)
|
||
return notify_values(row.value if row else None)
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 用户设置
|
||
# --------------------------------------------------------------------------
|
||
|
||
DEFAULT_SETTINGS: dict[str, Any] = {
|
||
"ui": {"theme": "system", "language": "zh", "canvas_snap": True, "autosave": True},
|
||
"run": {
|
||
"default_memory_scope": "default",
|
||
"default_collection": "default",
|
||
"stream_tokens": True,
|
||
"confirm_dangerous_tools": True,
|
||
# 「始终允许 · 门控把关」的工具每次调用前问的那个模型（tools/trust.py）。
|
||
# 空着就用默认 provider 的默认模型
|
||
"tool_gate_provider": None,
|
||
"tool_gate_model": None,
|
||
# agent 护栏（engine/guards.py）：兜底步数、每个节点的令牌 / 金额预算。预算为 null 表示不限
|
||
"agent_max_steps": 100,
|
||
"agent_budget_tokens": 2_000_000,
|
||
"agent_budget_usd": None,
|
||
},
|
||
"limits": {
|
||
"max_graph_steps": app_settings.max_graph_steps,
|
||
"max_agent_steps": app_settings.max_agent_steps,
|
||
"max_run_seconds": app_settings.max_run_seconds,
|
||
},
|
||
# 结论句裁判（engine/judge.py）：证据裁判模型和各项上限，上限存 null 表示不限。不放进 limits：
|
||
# 那一组是环境变量定的、只读，这一组是用户自己定的
|
||
JUDGE_GROUP: dict(JUDGE_DEFAULTS),
|
||
# 核查（双盲复算、反方审稿、审批质询）：核查模型、三项的开关和上限、每日金额上限。和裁判一组一样由用户定
|
||
VERIFY_GROUP: verify_values(None),
|
||
# 通知：定时任务的 Webhook 里拼运行链接用的对外访问地址
|
||
NOTIFY_GROUP: notify_values(None),
|
||
}
|
||
|
||
|
||
async def run_defaults(session: AsyncSession) -> dict[str, Any]:
|
||
"""设置里 run 这一组，没存过的键补内置默认。
|
||
|
||
发起运行和引擎都从这里读。各拼各的默认值，两边迟早会对「没存过」理解得
|
||
不一样——以前这三项就是只存不读，设置页上改了等于没改。
|
||
"""
|
||
row = await session.get(Setting, "run")
|
||
return {**DEFAULT_SETTINGS["run"], **((row.value if row else None) or {})}
|
||
|
||
|
||
def tool_approval_default(run: dict[str, Any]) -> str:
|
||
"""「危险工具默认需要人工确认」换算成节点 approval 的取值，节点自己配了的不看它。
|
||
|
||
只有明确存了 False 才关：值缺了、存坏了都按开着算。关掉审批得是一次明确的
|
||
选择，不能是某个字段没写上的副作用。
|
||
"""
|
||
return "never" if run.get("confirm_dangerous_tools") is False else "dangerous"
|
||
|
||
|
||
async def resolve_run_scope(
|
||
session: AsyncSession, memory_scope: str | None, collection: str | None
|
||
) -> tuple[str, str]:
|
||
"""请求没带的记忆域和知识库，取设置里的运行默认值。
|
||
|
||
空串等于没填：设置页把输入框清空再保存，存下来的就是空串，拿它当记忆域
|
||
写进去的东西谁也读不回来。
|
||
"""
|
||
defaults = await run_defaults(session)
|
||
|
||
def pick(explicit: str | None, key: str) -> str:
|
||
for value in (explicit, defaults.get(key)):
|
||
if isinstance(value, str) and value.strip():
|
||
return value.strip()
|
||
return "default"
|
||
|
||
return pick(memory_scope, "default_memory_scope"), pick(collection, "default_collection")
|
||
|
||
|
||
@router.get("/settings")
|
||
async def get_settings(session: AsyncSession = Depends(get_session)) -> dict[str, Any]:
|
||
rows = (await session.execute(select(Setting))).scalars()
|
||
stored = {r.key: r.value for r in rows}
|
||
# limits 是服务端配置（环境变量），只读：库里可能存着旧版设置页写进去的一份
|
||
#（见过 max_agent_steps: 25），它从来不生效，却会盖住真实的上限，界面照着错的数许诺
|
||
stored.pop("limits", None)
|
||
# 裁判、核查的每日计数是引擎记的账，不是用户设置：设置页整组回写时带着一份旧的，会把今天的花费清零
|
||
stored.pop(JUDGE_SPEND, None)
|
||
stored.pop(VERIFY_SPEND, None)
|
||
merged = {k: {**v, **(stored.get(k) or {})} for k, v in DEFAULT_SETTINGS.items()}
|
||
# 裁判、核查设置按引擎实际生效的样子给：存坏了的项显示默认，存了 null 的上限显示不限
|
||
merged[JUDGE_GROUP] = judge_values(stored.get(JUDGE_GROUP))
|
||
merged[VERIFY_GROUP] = verify_values(stored.get(VERIFY_GROUP))
|
||
merged[NOTIFY_GROUP] = notify_values(stored.get(NOTIFY_GROUP))
|
||
for key, value in stored.items():
|
||
merged.setdefault(key, value)
|
||
return merged
|
||
|
||
|
||
class SettingsIn(BaseModel):
|
||
values: dict[str, Any]
|
||
|
||
|
||
#: 点名的模型接入不存在或已停用时 PUT /api/settings 回 422 的机读码，field 是出错的那一项（verify.provider……）
|
||
PROVIDER_INVALID = "settings_provider_invalid"
|
||
|
||
|
||
def _provider_fields() -> dict[tuple[str, str], tuple[str, str | None]]:
|
||
"""设置里点名模型接入的项：(组, 键) → (这一项在界面上的叫法, 同一组模型那一项的叫法)。门控、裁判、核查的叫法和
|
||
设置页一致；助手的模型由「编排」页助手栏的「助手使用的模型」单独设定（PUT /api/copilot/model），只有这一个下拉框
|
||
（模型那一项为 None），这里只防经设置接口写进去的那一份。"""
|
||
from app.api.copilot import COPILOT_SETTING_KEY # 在这里导入：copilot 模块很大，设置模块被引擎各处导入
|
||
from app.providers.factory import ASSISTANT_MODEL_FIELD
|
||
|
||
return {
|
||
("run", "tool_gate_provider"): ("门控模型 · 接入", "门控模型 · 模型"),
|
||
(JUDGE_GROUP, "provider"): (JUDGE_LABEL["provider"], JUDGE_LABEL["model"]),
|
||
(VERIFY_GROUP, "provider"): (VERIFY_SETTING_LABEL["provider"], VERIFY_SETTING_LABEL["model"]),
|
||
(COPILOT_SETTING_KEY, "provider"): (ASSISTANT_MODEL_FIELD, None),
|
||
}
|
||
|
||
|
||
async def _provider_problem(session: AsyncSession, values: dict[str, Any]) -> tuple[str, str] | None:
|
||
"""这次写进来的设置里点名的模型接入有没有问题：(字段, 说明)，没问题返回 None。说明和运行时报的原因、提示同一句
|
||
（factory.named_provider），主语写这一项在设置页上的叫法；下拉框里选的，写「指定的」。助手那一项只有「编排」页
|
||
助手栏一个下拉框（只选模型，选了就清掉接入），没有「换一个接入」的地方：说法同助手取不到模型时那一句
|
||
（factory.assistant_unavailable），哪里报都一样。
|
||
|
||
以前点名的接入不存在时运行悄悄换成演示模型，名字写错一个字就由演示模型写报告、做核查，现在运行时照实报错；保存时
|
||
先拦一道。留空照旧表示沿用上一级——说明里不再讲留空是什么意思：设置页那一格的空选项写着（「默认接入」「跟随助手
|
||
的模型」「跟随证据裁判模型」……），这里另写一遍只会对不上。只查这次改了的：存着的值没改（写好之后接入才被删、
|
||
停用）不挡这次保存别的项——设置页在那一项下标着「不存在」「已停用」，运行时照实报错，同设置页「只校验改过的」。
|
||
点名的名字恰好是某个启用接入的模型时，补一句在这一组的哪一栏用它（factory.model_note）。
|
||
|
||
名字不是字符串（接口写进来的数字一类）：核查、裁判两组自己的格式检查在前面先报；门控、助手两组没有自己的检查，
|
||
在这里报，免得存下去之后读它的地方（查助手当前用哪个模型）出错。"""
|
||
providers = None
|
||
for (group, key), (label, model_label) in _provider_fields().items():
|
||
sent = values.get(group)
|
||
if not isinstance(sent, dict):
|
||
continue
|
||
if (raw := sent.get(key)) is not None and not isinstance(raw, str):
|
||
return f"{group}.{key}", f"「{label}」需要填写模型接入的名称，当前为「{raw}」"
|
||
if (name := _model_text(raw)) is None:
|
||
continue
|
||
row = await session.get(Setting, group)
|
||
saved = row.value if row is not None and isinstance(row.value, dict) else {}
|
||
if _model_text(saved.get(key)) == name:
|
||
continue
|
||
if providers is None:
|
||
providers = await all_providers(session)
|
||
try:
|
||
named_provider(providers, name)
|
||
except ProviderNotConfigured as e:
|
||
if model_label is None:
|
||
return f"{group}.{key}", assistant_unavailable(e)
|
||
said = f"「{label}」指定的{e.reason}。{e.hint}"
|
||
if e.model_owner:
|
||
# 这几组的模型一栏是输入框：写「填写」
|
||
said += f"。{model_note(e.name, e.model_owner, f'「{model_label}」', f'「{label}」', '填写')}"
|
||
return f"{group}.{key}", said
|
||
return None
|
||
|
||
|
||
@router.put("/settings")
|
||
async def put_settings(
|
||
payload: SettingsIn, session: AsyncSession = Depends(get_session)
|
||
) -> dict[str, Any]:
|
||
from app.tools.trust import SETTING_KEY as TRUST_KEY
|
||
|
||
# 先查后写：一组里有一项不对就整个请求不落库，免得存下半组
|
||
if JUDGE_GROUP in payload.values and (problem := judge_problem(payload.values[JUDGE_GROUP])):
|
||
raise HTTPException(422, problem)
|
||
if VERIFY_GROUP in payload.values and (problem := verify_problem(payload.values[VERIFY_GROUP])):
|
||
raise HTTPException(422, problem)
|
||
if NOTIFY_GROUP in payload.values:
|
||
if problem := notify_problem(payload.values[NOTIFY_GROUP]):
|
||
raise HTTPException(422, problem)
|
||
# 存规范化之后的：末尾带 / 的地址拼链接会多出一个斜杠
|
||
payload.values[NOTIFY_GROUP] = notify_values(payload.values[NOTIFY_GROUP])
|
||
if found := await _provider_problem(session, payload.values):
|
||
raise CodedHTTPException(422, found[1], PROVIDER_INVALID, field=found[0])
|
||
for key, value in payload.values.items():
|
||
if key in (TRUST_KEY, "limits", JUDGE_SPEND, VERIFY_SPEND):
|
||
# 信任三档只走 PUT /api/tools/trust：设置页整组回写时带着一份旧的，
|
||
# 会把刚在工具页改的档位冲掉。limits 是环境变量定的，写进库也不生效。
|
||
# 裁判、核查的每日计数是引擎记的账，只有引擎自己写
|
||
continue
|
||
row = await session.get(Setting, key)
|
||
if row:
|
||
row.value = value
|
||
else:
|
||
session.add(Setting(key=key, value=value))
|
||
await session.commit()
|
||
return await get_settings(session)
|
||
|
||
|
||
@router.get("/settings/judge/spend")
|
||
async def get_judge_spend(session: AsyncSession = Depends(get_session)) -> dict[str, Any]:
|
||
"""今天（本地日期）结论句裁判花了多少，设置页放在每日上限旁边。只读：计数只有裁判自己记。
|
||
|
||
{date, usd, calls, unpriced_calls, daily_max_usd}。unpriced_calls 是价格目录里没有的模型的调用次数，
|
||
这些调用估不出金额、没有算进 usd。daily_max_usd 为 null 表示不限。
|
||
"""
|
||
return {**await daily_spend(), "daily_max_usd": (await judge_settings(session))["daily_max_usd"]}
|
||
|
||
|
||
@router.get("/settings/verify/spend")
|
||
async def get_verify_spend(session: AsyncSession = Depends(get_session)) -> dict[str, Any]:
|
||
"""今天（本地日期）核查（双盲复算、反方审稿、审批质询）花了多少，设置页放在核查的每日上限旁边。只读。
|
||
|
||
形状同 /settings/judge/spend：{date, usd, calls, unpriced_calls, daily_max_usd}，daily_max_usd 为 null 表示不限。
|
||
"""
|
||
return {**await verify_daily_spend(), "daily_max_usd": (await verify_settings(session))["daily_max_usd"]}
|
||
|
||
|
||
@router.get("/system")
|
||
async def system_info() -> dict[str, Any]:
|
||
"""给设置页顶部展示的运行环境概览。"""
|
||
import sys
|
||
|
||
from app.sandbox.manager import sandbox_manager
|
||
|
||
return {
|
||
"python": sys.version.split()[0],
|
||
"data_dir": str(app_settings.data_dir),
|
||
"sandbox": await sandbox_manager.health(),
|
||
"limits": {
|
||
"max_graph_steps": app_settings.max_graph_steps,
|
||
"max_agent_steps": app_settings.max_agent_steps,
|
||
"max_concurrent_runs": app_settings.max_concurrent_runs,
|
||
"max_run_seconds": app_settings.max_run_seconds,
|
||
"model_timeout_seconds": app_settings.model_timeout_seconds,
|
||
},
|
||
}
|