feat: 报告结构支持检索方向/来源与智能新增,并修复协助卡停轮与宿主跳转

为深度研究报告结构增加检索方向与技能来源、对话式智能新增页与详情排版;澄清中间件在 after_model 可靠结束本轮,避免弱模型出卡后继续生成;任务按钮补充 host 跳转与相关深链能力。

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
董宏凯andCursor committed 2026-08-22 17:46:06 +08:00
1 parent a5a9c952e8
commit 51afebb394
67 files changed
+4511 -699

No files matched your search

@@ -4,13 +4,13 @@ import json
import logging
from collections.abc import Callable
from hashlib import sha256
from typing import override
from typing import Any, override
from deerflow.tools.builtins.clarification_utils import resolve_allow_multiple
from langchain.agents import AgentState
from langchain.agents.middleware import AgentMiddleware
from langchain_core.messages import ToolMessage
from langchain.agents.middleware import AgentMiddleware, hook_config
from langchain_core.messages import AIMessage, ToolMessage
from langgraph.graph import END
from langgraph.prebuilt.tool_node import ToolCallRequest
from langgraph.types import Command
@@ -197,6 +197,89 @@ class ClarificationMiddleware(AgentMiddleware[ClarificationMiddlewareState]):
goto=END,
)
def _tool_field(self, tool_call: Any, key: str, default: Any = None) -> Any:
if isinstance(tool_call, dict):
return tool_call.get(key, default)
getter = getattr(tool_call, "get", None)
if callable(getter):
return getter(key, default)
return getattr(tool_call, key, default)
def _clarification_tool_message(self, tool_call_id: str, args: dict[str, Any]) -> ToolMessage:
formatted_message = self._format_clarification_message(args)
return ToolMessage(
id=self._stable_message_id(tool_call_id, formatted_message),
content=formatted_message,
tool_call_id=tool_call_id,
name="ask_clarification",
additional_kwargs={"clarification": self._build_payload(args)},
)
def _after_model_intercept(self, last: AIMessage) -> dict[str, Any] | None:
"""Stop the run as soon as the model emits ask_clarification.
``wrap_tool_call`` + ``Command(goto=END)`` is unreliable with parallel
tool calls (weak models often emit two clarifications, or clarification
plus another tool): the Command mixes with other tool results and the
graph loops back to the model, which then treats the question ToolMessage
as "the user still hasn't answered" and keeps generating.
"""
tool_calls = list(last.tool_calls or [])
clarification_calls = [
tc for tc in tool_calls if self._tool_field(tc, "name") == "ask_clarification"
]
if not clarification_calls:
return None
messages: list[ToolMessage] = []
for tc in tool_calls:
name = str(self._tool_field(tc, "name") or "")
tool_call_id = str(self._tool_field(tc, "id") or "")
args = self._tool_field(tc, "args") or {}
if not isinstance(args, dict):
args = {}
if name == "ask_clarification":
messages.append(self._clarification_tool_message(tool_call_id, args))
else:
messages.append(
ToolMessage(
id=f"clarification-skip:{tool_call_id or name}",
content="已暂停:需要先等待用户回答澄清问题,本工具未执行。",
tool_call_id=tool_call_id,
name=name or "unknown",
)
)
logger.info(
"Intercepted %s clarification tool call(s) in after_model; ending turn",
len(clarification_calls),
)
return {"messages": messages, "jump_to": "end"}
@hook_config(can_jump_to=["end"])
@override
def after_model(
self,
state: ClarificationMiddlewareState,
runtime: object,
) -> dict[str, Any] | None:
messages = state.get("messages") or []
if not messages:
return None
last = messages[-1]
if not isinstance(last, AIMessage):
return None
return self._after_model_intercept(last)
@hook_config(can_jump_to=["end"])
@override
async def aafter_model(
self,
state: ClarificationMiddlewareState,
runtime: object,
) -> dict[str, Any] | None:
return self.after_model(state, runtime)
@override
def wrap_tool_call(
self,
@@ -0,0 +1,113 @@
"""Approval middleware for the report-structure setup tool.
与 ``SetupWritingApprovalMiddleware`` 同构:模型在「智能新增报告结构」里调用
``setup_report_structure`` 时,于 ``after_model`` 拦截,附上一条结构化审批 ToolMessage
(建议字段打包在 ``additional_kwargs["setup_report_structure_approval"]``)并
``jump_to="end"`` 立即结束本轮——前端据此渲染确认卡,用户点「应用到表单」后回填右侧表单。
"""
from __future__ import annotations
import logging
from hashlib import sha256
from typing import Any, override
from langchain.agents import AgentState
from langchain.agents.middleware import AgentMiddleware, hook_config
from langchain_core.messages import AIMessage, ToolMessage
from langgraph.runtime import Runtime
from deerflow.persistence.report_structures.directions import normalize_retrieval_directions
from deerflow.persistence.report_structures.skills import normalize_retrieval_skills
logger = logging.getLogger(__name__)
class SetupReportStructureApprovalMiddlewareState(AgentState):
"""Compatible with the DeerFlow thread state schema."""
pass
class SetupReportStructureApprovalMiddleware(AgentMiddleware[SetupReportStructureApprovalMiddlewareState]):
"""Stop before ``setup_report_structure`` and ask the user to apply the proposed config."""
state_schema = SetupReportStructureApprovalMiddlewareState
def _is_approved(self, runtime: Runtime) -> bool:
context = getattr(runtime, "context", None) or {}
if not isinstance(context, dict):
return False
return context.get("setup_report_structure_approved") is True
def _stable_message_id(self, tool_call_id: str, args: dict[str, Any]) -> str:
if tool_call_id:
return f"setup-report-structure-approval:{tool_call_id}"
digest = sha256(repr(sorted(args.items())).encode("utf-8")).hexdigest()[:16]
return f"setup-report-structure-approval:{digest}"
def _normalize_mode(self, raw: Any) -> str:
text = str(raw or "").strip().lower()
if text in {"fixed", "固定", "固定大纲"}:
return "fixed"
return "adaptive"
def _build_payload(self, args: dict[str, Any]) -> dict[str, Any]:
return {
"title": str(args.get("title") or "").strip()[:256],
"name": str(args.get("name") or "").strip()[:256],
"content": str(args.get("content") or "").strip(),
"structure_mode": self._normalize_mode(args.get("structure_mode")),
"retrieval_directions": normalize_retrieval_directions(args.get("retrieval_directions")),
"retrieval_skills": normalize_retrieval_skills(args.get("retrieval_skills")),
}
def _intercept(self, last: AIMessage) -> dict[str, Any] | None:
tool_calls = list(last.tool_calls or [])
setup_call = next((tc for tc in tool_calls if tc.get("name") == "setup_report_structure"), None)
if setup_call is None:
return None
args = setup_call.get("args") or {}
tool_call_id = setup_call.get("id") or ""
payload = self._build_payload(args)
logger.info("Setup report-structure approval requested before applying config")
approval_msg = ToolMessage(
id=self._stable_message_id(tool_call_id, args),
content="请确认是否把这份报告结构应用到表单。",
tool_call_id=tool_call_id,
name="setup_report_structure",
additional_kwargs={"setup_report_structure_approval": payload},
)
return {
"messages": [approval_msg],
"jump_to": "end",
}
@hook_config(can_jump_to=["end"])
@override
def after_model(
self,
state: SetupReportStructureApprovalMiddlewareState,
runtime: Runtime,
) -> dict[str, Any] | None:
if self._is_approved(runtime):
return None
messages = state.get("messages") or []
if not messages:
return None
last = messages[-1]
if not isinstance(last, AIMessage):
return None
return self._intercept(last)
@hook_config(can_jump_to=["end"])
@override
async def aafter_model(
self,
state: SetupReportStructureApprovalMiddlewareState,
runtime: Runtime,
) -> dict[str, Any] | None:
return self.after_model(state, runtime)
@@ -24,25 +24,32 @@ class WritingSetupRoundCapMiddleware(AgentMiddleware[AgentState]):
无视提示,``tool_choice`` 也会逼它产出一次 ``setup_writing`` 调用。
"""
def __init__(self, max_rounds: int = 3, tool_name: str = "setup_writing") -> None:
def __init__(
self,
max_rounds: int = 3,
tool_name: str = "setup_writing",
force_instruction: str | None = None,
) -> None:
super().__init__()
self._max_rounds = max(1, max_rounds)
self._tool_name = tool_name
self._force_instruction = force_instruction
def _should_force(self, messages: list) -> bool:
human_count = sum(1 for m in messages if isinstance(m, HumanMessage))
return human_count >= self._max_rounds
def _reminder(self) -> HumanMessage:
content = self._force_instruction or (
f"【系统强制】意图澄清已达 {self._max_rounds} 轮上限。"
"现在必须立即调用 setup_writing 提交写作配置,禁止再向用户提任何问题。"
"缺失或不确定的字段一律用合理默认值:"
"word_count_target 默认 800、target_audience 默认「普通读者」、"
"writing_mode 默认 loose、material_source 默认 knowledge_base、"
"article_type 不确定就留空。"
)
return HumanMessage(
content=(
f"【系统强制】意图澄清已达 {self._max_rounds} 轮上限。"
"现在必须立即调用 setup_writing 提交写作配置,禁止再向用户提任何问题。"
"缺失或不确定的字段一律用合理默认值:"
"word_count_target 默认 800、target_audience 默认「普通读者」、"
"writing_mode 默认 loose、material_source 默认 knowledge_base、"
"article_type 不确定就留空。"
),
content=content,
additional_kwargs={"hide_from_ui": True, "lc_source": "writing_setup_round_cap"},
)