| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165 |
- """System-owned web tools over Hermes's native providers and safety pipeline.
- Called under hermes_service's configuration lock. Bot homes never execute these
- operations; Yaovi stores the returned full text as a session-scoped tool result.
- """
- import json
- import asyncio
- import os
- import stat
- from pathlib import Path
- def _run(value):
- return asyncio.run(value) if asyncio.iscoroutine(value) else value
- def _system():
- if os.environ.get("YAOVI_HERMES_ROOT"):
- raise ValueError("搜索与网页配置仅支持系统层")
- def _extract_backend():
- from tools import web_tools
- from tools.tool_backend_helpers import read_web_capability_selection
- backend = web_tools._get_extract_backend()
- if read_web_capability_selection("extract") is None:
- # Native shared autodetection can choose search-only ddgs/Brave. For
- # an unpinned extract capability use Hermes's capability-aware registry.
- web_tools._ensure_web_plugins_loaded()
- from agent.web_search_registry import get_provider, get_active_extract_provider
- candidate = get_provider(backend)
- if candidate is not None and not candidate.supports_extract():
- candidate = get_active_extract_provider()
- if candidate is not None:
- backend = candidate.name
- return backend
- def config():
- _system()
- from hermes_cli.web_routers.tools import get_toolset_config
- from hermes_cli.config import read_raw_config
- result = _run(get_toolset_config("web"))
- web = (read_raw_config() or {}).get("web") or {}
- from tools.tool_backend_helpers import read_selection
- result["selection"] = {cap: web.get(f"{cap}_backend") or read_selection("web")
- for cap in ("search", "extract")}
- result["active_extract_backend"] = _extract_backend()
- return result
- def provider(args):
- _system()
- from hermes_cli.config import read_raw_config, save_config
- from hermes_cli.web_routers.tools import select_toolset_provider
- from hermes_cli.web_models import ToolsetProviderSelect
- capability = args["capability"]
- if capability not in ("search", "extract"):
- raise ValueError("未知网页工具能力")
- if args.get("provider") is None:
- # Restore this capability to native autodetection. A legacy shared pin
- # must not keep it pinned; preserve the other capability explicitly.
- raw = read_raw_config() or {}
- web = raw.setdefault("web", {})
- from tools.tool_backend_helpers import read_selection
- shared = read_selection("web")
- web.pop("backend", None)
- web.pop("use_gateway", None)
- other = "extract" if capability == "search" else "search"
- if shared and not web.get(f"{other}_backend"):
- web[f"{other}_backend"] = shared
- web.pop(f"{capability}_backend", None)
- save_config(raw)
- else:
- _run(select_toolset_provider("web", ToolsetProviderSelect(
- provider=args["provider"], capability=capability)))
- return config()
- def env(args):
- _system()
- from hermes_cli.web_routers.tools import save_toolset_env
- from hermes_cli.web_models import ToolsetEnvUpdate
- _run(save_toolset_env("web", ToolsetEnvUpdate(env=args["env"])))
- return config()
- def _public_result(result):
- # Vendor error envelopes are successful RPC replies; redact them too.
- from hermes_cli.config import load_env, OPTIONAL_ENV_VARS
- secrets = [v for k, v in (load_env() or {}).items()
- if len(v) >= 8 and OPTIONAL_ENV_VARS.get(k, {}).get("password", True)]
- def clean(value):
- if isinstance(value, str):
- for secret in secrets:
- value = value.replace(secret, "[已隐藏授权]")
- return value
- if isinstance(value, list):
- return [clean(v) for v in value]
- if isinstance(value, dict):
- return {k: clean(v) for k, v in value.items()}
- return value
- result = clean(result)
- failed = result.get("success") is False or bool(result.get("error")) or (
- "results" in result and not any(r.get("content") and not r.get("error") for r in result["results"]))
- if failed:
- result["isError"] = True
- result["configuration_hint"] = "请在我的设置 → 模型服务 → 搜索与网页中检查供应商、凭据和连接。"
- return result
- def search(args):
- _system()
- from tools.web_tools import web_search_tool, _get_search_backend
- result = json.loads(web_search_tool(args["query"], limit=args.get("limit", 5)))
- return _public_result({**result, "provider": _get_search_backend()})
- def extract(args):
- _system()
- from hermes_constants import get_hermes_dir
- from tools import web_tools_truncate as truncation
- from tools import web_tools
- created = {}
- backend = _extract_backend()
- original_backend = web_tools._get_extract_backend
- original = truncation._store_full_text
- def record(url, content):
- path = original(url, content)
- if path:
- created[url] = Path(path)
- return path
- # Record only paths generated by this native call, never parse a path from
- # provider text or accept one from the model. Restore even on failure.
- truncation._store_full_text = record
- web_tools._get_extract_backend = lambda: backend
- try:
- result = json.loads(_run(web_tools.web_extract_tool(args["urls"], char_limit=args.get("char_limit"))))
- finally:
- truncation._store_full_text = original
- web_tools._get_extract_backend = original_backend
- cache = get_hermes_dir("cache/web", "web_cache").resolve()
- for row in result.get("results", []):
- path = created.get(row.get("url"))
- if path is None:
- continue
- try:
- if path.parent.resolve() != cache:
- raise ValueError("网页缓存目录无效")
- fd = os.open(path, os.O_RDONLY | os.O_NOFOLLOW)
- with os.fdopen(fd, "r", encoding="utf-8") as stream:
- info = os.fstat(stream.fileno())
- if not stat.S_ISREG(info.st_mode) or info.st_size > 8 * 1024 * 1024:
- raise ValueError("网页缓存超过读取上限")
- row["content"] = stream.read(2 * 1024 * 1024)
- except (OSError, ValueError):
- # Do not expose the native read_file footer if storage/read failed.
- row["content"] = row.get("content", "").split("──────── [TRUNCATED]", 1)[0]
- row["note"] = "全文缓存无法读取,可提高 char_limit 重新提取。"
- return _public_result({**result, "provider": backend})
- def test(args):
- result = search({"query": args.get("query") or "Hermes web search", "limit": 1}) if args["capability"] == "search" else extract({"urls": [args.get("url") or "https://example.com"], "char_limit": 2000})
- return {"ok": not result.get("isError", False), "provider": result.get("provider"), "result": result}
|