代码合并

This commit is contained in:
董宏凯 committed 2026-09-08 09:40:31 +08:00
1 parent b0970de677
commit 8859c13cfb
26 files changed
+1707 -334

No files matched your search

@@ -0,0 +1,422 @@
"""规划师智能体(PlannerAgent)——真智能体,不是规则路由 + 模板文案。
设计 v2 的需求对话与工作流生成主路径:
- ``converse``:与用户多轮自由对话确认需求。回复文案(含开场)完全由
智能体生成并直接落库展示 —— 「收到您的报告需求。为了给出…」这类
死模板文案已从该路径移除。
- ``design_workflow``:带着**完整对话上下文**(规划师问了什么、用户
答了什么)生成受词表约束的自定义工作流(``StrategyRoleSet``),
与确定性模板走同一条 装配 → 校验 → 绑定 链路。
规则路由与模板只作为模型彻底不可用时的最后防线,不出现在用户面前。
"""
from __future__ import annotations
import json
import logging
import re
from typing import Any, Callable
from pydantic import BaseModel, ConfigDict, Field, ValidationError
from app.report_collaboration.agentscope_runtime.model_resolver import ROLE_MODEL_FIELD, RoleModelSnapshot, resolve_role_model
from app.report_collaboration.contracts.plans import RoleDemand
from app.report_collaboration.contracts.requirements import ReportRequirementSnapshot
from app.report_collaboration.planning.controller import RoleSlot, StrategyRoleSet
from app.report_collaboration.requirement.intent_router import _merge_snapshot, apply_default_assumptions
from deerflow.config.app_config import AppConfig, get_app_config
logger = logging.getLogger(__name__)
# ── 词表(结构化输出的安全边界,LLM 不受约束的自由只在文案层) ─────────────
MEMBER_ROLES: tuple[str, ...] = tuple(key for key in ROLE_MODEL_FIELD if key != "coordinator")
ROLE_ARTIFACTS: dict[str, tuple[str, ...]] = {
"researcher": ("EvidenceBundle",),
"verifier": ("VerificationResult",),
"analyst": ("AngleAnalysis",),
"synthesizer": ("SynthesisOutline",),
"writer": ("ReportSectionDraft",),
"reviewer": ("ReviewDecision",),
"reviser": ("ReportSectionDraft",),
}
QUALITY_GATES: tuple[str, ...] = ("cite_required", "dual_verification", "independent_review", "coverage_matrix")
# ── 智能体输出契约 ─────────────────────────────────────────────────────────
class PlannerTurn(BaseModel):
"""规划师一轮对话的完整产出:自由文案 + 结构化状态。"""
model_config = ConfigDict(extra="ignore")
action: str = "clarify"
reply: str = ""
questions: list[str] = Field(default_factory=list)
impact: list[str] = Field(default_factory=list)
assumptions: list[str] = Field(default_factory=list)
requirement: dict[str, Any] = Field(default_factory=dict)
class StrategyNodeDraft(BaseModel):
model_config = ConfigDict(extra="ignore")
slot_id: str
role_key: str
display_name: str
angle: str | None = None
mission: str
depends_on: list[str] = Field(default_factory=list)
output_artifact_type: str
suggested_tools: list[str] = Field(default_factory=list)
class StrategyDraft(BaseModel):
model_config = ConfigDict(extra="ignore")
reply: str = ""
strategy: str = "custom"
title: str
summary: str
rationale: str
estimated_duration_seconds: int = 900
estimated_cost_level: str = "medium"
quality_gate_ids: list[str] = Field(default_factory=lambda: ["cite_required"])
nodes: list[StrategyNodeDraft] = Field(default_factory=list)
LlmFactory = Callable[[], Any]
def planner_model_snapshot(app_config: AppConfig | None = None) -> RoleModelSnapshot:
"""规划师身份沿用角色解析链:coordinator_model → planner_model → 默认模型。"""
return resolve_role_model("coordinator", app_config=app_config or get_app_config())
def build_planner_llm(app_config: AppConfig | None = None) -> Any:
from deerflow.models.factory import create_chat_model
snapshot = planner_model_snapshot(app_config)
return create_chat_model(snapshot.config_name, app_config=app_config or get_app_config())
_THINK_BLOCK_RE = re.compile(r"<think>.*?</think>", re.DOTALL | re.IGNORECASE)
def _extract_json(text: str) -> dict[str, Any] | None:
if not text:
return None
# 思考模型(MiniMax-M2.7 等)回复可能带 <think>…</think>;JSON 在思考块之后。
# 先剥思考块提取;剥完没有再回退用全文(极端情况 JSON 混在思考内)。
for candidate in (_THINK_BLOCK_RE.sub("", text).strip(), text.strip()):
if not candidate:
continue
cleaned = candidate
if cleaned.startswith("```"):
cleaned = cleaned.strip("`")
if cleaned.lower().startswith("json"):
cleaned = cleaned[4:]
cleaned = cleaned.strip()
start = cleaned.find("{")
end = cleaned.rfind("}")
if start < 0 or end <= start:
continue
try:
parsed = json.loads(cleaned[start : end + 1])
except Exception:
continue
if isinstance(parsed, dict):
return parsed
return None
async def _invoke_json(llm: Any, messages: list[Any]) -> tuple[dict[str, Any] | None, str]:
"""调用模型并提取 JSON。返回 (解析结果, 原始文本);解析失败结果为 None。"""
try:
response = await llm.ainvoke(messages)
raw = getattr(response, "content", response)
if isinstance(raw, list):
raw = "".join(str(getattr(block, "text", block) if not isinstance(block, str) else block) for block in raw)
raw = str(raw)
return _extract_json(raw), raw
except Exception as exc: # noqa: BLE001 — 智能体故障必须显式失败
logger.warning("rc planner agent llm call failed: %s: %s", exc.__class__.__name__, exc)
return None, ""
async def _invoke_json_with_retry(llm: Any, system: str, user_prompt: str) -> dict[str, Any] | None:
"""带一次格式重试:模型输出不含 JSON 时,附原始输出要求只出 JSON 再问一次。"""
parsed, raw = await _invoke_json(llm, [("system", system), ("human", user_prompt)])
if parsed is None:
logger.warning("rc planner reply without JSON (raw head: %.200s)", raw[:200])
retry_prompt = (
user_prompt
+ "\n\n【格式提醒】你上一轮的输出没有包含合法 JSON。这次必须只输出一个 JSON 对象,"
+ f"不要在 JSON 之外输出任何文字。你上一轮的输出是:{raw[:500]}"
)
parsed, _ = await _invoke_json(llm, [("system", system), ("human", retry_prompt)])
return parsed
# ── 规划师智能体 ───────────────────────────────────────────────────────────
class PlannerAgent:
"""需求对话与工作流设计的唯一主路径(真智能体)。"""
def __init__(self, llm: Any, *, app_config: AppConfig | None = None) -> None:
self._llm = llm
self._app_config = app_config
# ── 需求对话 ───────────────────────────────────────────────────────────
async def converse(
self,
*,
user_text: str,
history: list[dict[str, str]],
previous_snapshot: dict[str, Any] | None,
previous_revision: int,
clarification_round: int,
max_rounds: int,
) -> PlannerTurn | None:
"""一轮规划对话。返回 None 表示智能体不可用(调用方走最后防线)。"""
forced = clarification_round >= max_rounds
system = _CONVERSE_SYSTEM.format(
roles=", ".join(MEMBER_ROLES),
forced_note=(
"\n\n【重要】澄清轮次已达上限,本轮你必须判定需求就绪(action=ready):"
"自然地告诉用户你将按当前理解开始规划,把未能确认的点写进 assumptions。"
if forced
else ""
),
)
known = {k: v for k, v in (previous_snapshot or {}).items() if v not in (None, "", [])}
known.pop("revision", None)
history_text = "\n".join(
f"{'用户' if item.get('role') == 'human' else '规划师'}: {str(item.get('content') or '')[:400]}"
for item in history[-10:]
)
user_prompt = (
f"当前已确认的需求快照(JSON):{json.dumps(known, ensure_ascii=False)}\n\n"
f"最近的对话:\n{history_text or '(无)'}\n\n"
f"用户最新消息:{user_text}"
)
parsed = await _invoke_json_with_retry(self._llm, system, user_prompt)
if parsed is None:
return None
try:
turn = PlannerTurn.model_validate(parsed)
except ValidationError:
logger.warning("rc planner turn invalid: %s", str(parsed)[:200])
return None
if not (turn.reply or "").strip():
return None
if forced and turn.action != "ready":
turn.action = "ready"
if turn.action == "ready" and not turn.assumptions:
turn.assumptions = ["按对话中已确认的信息直接规划"]
return turn
def merge_requirement(
self,
turn: PlannerTurn,
*,
previous_snapshot: dict[str, Any] | None,
previous_revision: int,
) -> ReportRequirementSnapshot:
"""把智能体本轮带回的需求字段合并进既有快照(revision 语义同规则路由)。"""
known = {k: v for k, v in (previous_snapshot or {}).items() if k != "revision"}
# LLM 契约是「只填能确认的字段」;再过滤一次空值,避免空串覆盖已确认信息。
patch = {
k: v
for k, v in (turn.requirement or {}).items()
if k != "revision" and v not in (None, "", [])
}
revision = (
max(1, int(previous_revision) + 1)
if (previous_snapshot or patch)
else max(1, int(previous_revision) or 1)
)
if not previous_snapshot:
revision = 1
merged = _merge_snapshot(known, patch, revision=revision)
if turn.action == "ready":
merged = apply_default_assumptions(merged)[0]
return merged
# ── 工作流设计 ─────────────────────────────────────────────────────────
async def design_workflow(
self,
*,
requirement: ReportRequirementSnapshot,
history: list[dict[str, str]],
max_team_members: int = 8,
) -> tuple[StrategyRoleSet, str] | None:
"""带着完整对话上下文设计工作流;返回 (策略, 展示文案) 或 None。"""
fields = {k: v for k, v in requirement.model_dump().items() if v not in (None, "", [])}
fields.pop("revision", None)
history_text = "\n".join(
f"{'用户' if item.get('role') == 'human' else '规划师'}: {str(item.get('content') or '')[:300]}"
for item in history[-12:]
)
system = _DESIGN_SYSTEM.format(
roles=", ".join(MEMBER_ROLES),
artifacts="\n".join(f" - {role}: {' / '.join(arts)}" for role, arts in ROLE_ARTIFACTS.items()),
gates=", ".join(QUALITY_GATES),
max_nodes=min(max_team_members, 8),
)
user_prompt = (
f"需求快照(JSON):{json.dumps(fields, ensure_ascii=False)}\n\n"
f"需求确认阶段的对话(你与用户):\n{history_text or '(无)'}\n\n请设计工作流。"
)
parsed = await _invoke_json_with_retry(self._llm, system, user_prompt)
if parsed is None:
return None
try:
draft = StrategyDraft.model_validate(parsed)
except ValidationError:
logger.warning("rc planner strategy draft invalid: %s", str(parsed)[:200])
return None
strategy = _validate_strategy_dag(draft)
if strategy is None:
logger.warning("rc planner strategy failed vocabulary/DAG validation: %s", str(parsed)[:200])
return None
reply = (draft.reply or "").strip() or f"已按需求设计好协作工作流「{draft.title}」。"
return strategy, reply
def workflow_reply_fallback(self, strategy: StrategyRoleSet) -> str:
return (
f"我按确认的需求设计了「{strategy.title}」:{strategy.summary}\n"
"可以直接采用开始执行;想调整就告诉我(例如「加一条竞品对比线」)。"
)
# ── 提示词 ─────────────────────────────────────────────────────────────────
_CONVERSE_SYSTEM = """你是「报告协作工作台」的规划师智能体。用户想让你组织一个多智能体团队协作产出一份报告,你的任务是先通过自然对话把需求弄清楚,然后(由你自己)设计协作工作流。
你的对话风格:
- 像一位资深顾问:先回应用户刚说的话,再追问。回复里**不要**出现「收到您的报告需求」这类机械开场——用你自己的话自然表达。
- 每轮最多追问 2~3 个真正影响方案设计的问题:关注对象与范围、给谁看、必须覆盖的角度、篇幅与深度、时间范围、特殊约束。用户已说清的绝不重复问。
- 用户回复后先简短确认你的理解(一两句),再决定是否继续问。
- 信息足够时,自然地总结你对需求的把握并判定就绪(action=ready),告诉用户下一步你会为他们设计工作流。
可用的团队角色(供你构思,不必对用户罗列):{roles}{forced_note}
每轮严格输出一个 JSON 对象(不要输出 JSON 以外的文字):
{{
"reply": "给用户看的完整回复(自然对话,直接展示)",
"action": "clarify" 或 "ready",
"questions": ["追问的问题(与 reply 一致,供系统记录)"],
"impact": ["为什么需要问(与 questions 对应)"],
"assumptions": ["ready 时:你采用的默认假设"],
"requirement": {{"topic": "...", "objective": "...", "audience": "...", "required_angles": ["..."], "time_range": "...", "length_target": "...", "tone": "..."}}
}}
requirement 只填你从对话中能确认的字段(含此前轮次已确认的),其余省略。reply 用中文。"""
_DESIGN_SYSTEM = """你是「报告协作工作台」的规划师智能体。需求已经与你确认完毕,现在请设计多智能体协作工作流。
词表约束(结构化部分必须严格遵守):
- role_key 只能取:{roles}
- 各角色的 output_artifact_type 只能取:
{artifacts}
- quality_gate_ids 只能取:{gates}
结构要求:
- 2~{max_nodes} 个节点;至少 1 个 writer;正式报告建议加 reviewer。
- 研究线(researcher)按角度并行;research → verify → analyze → write → review 是常见主干,可按需求裁剪。
- depends_on 引用其他节点的 slot_id;禁止环;至少一个节点 depends_on 为空(起点)。
- slot_id 用简短英文(如 research-market / write-main)。mission 写清楚该节点做什么、产出什么。
严格输出 JSON(不要输出 JSON 以外的文字):
{{
"reply": "给用户看的介绍:说明你为什么这样设计(对应需求的哪些要求),一两句即可",
"strategy": "custom",
"title": "方案标题",
"summary": "一句话概述团队构成与链路",
"rationale": "设计理由",
"estimated_duration_seconds": 1500,
"estimated_cost_level": "low|medium|high",
"quality_gate_ids": ["cite_required"],
"nodes": [
{{"slot_id": "research-market", "role_key": "researcher", "display_name": "研究员", "angle": "市场份额", "mission": "...", "depends_on": [], "output_artifact_type": "EvidenceBundle", "suggested_tools": ["web_search"]}}
]
}}"""
# ── 工作流草稿校验(词表 + DAG) ──────────────────────────────────────────
def _validate_strategy_dag(draft: StrategyDraft) -> StrategyRoleSet | None:
nodes = draft.nodes
if not 2 <= len(nodes) <= 8:
return None
slot_ids = {node.slot_id for node in nodes}
if len(slot_ids) != len(nodes):
return None
if not any(node.role_key == "writer" for node in nodes):
return None
for node in nodes:
if node.role_key not in MEMBER_ROLES:
return None
if node.output_artifact_type not in ROLE_ARTIFACTS.get(node.role_key, ()):
return None
for dep in node.depends_on:
if dep not in slot_ids or dep == node.slot_id:
return None
indegree = {node.slot_id: len(node.depends_on) for node in nodes}
children: dict[str, list[str]] = {node.slot_id: [] for node in nodes}
for node in nodes:
for dep in node.depends_on:
children[dep].append(node.slot_id)
queue = [slot for slot, degree in indegree.items() if degree == 0]
seen = 0
while queue:
current = queue.pop()
seen += 1
for child in children[current]:
indegree[child] -= 1
if indegree[child] == 0:
queue.append(child)
if seen != len(nodes):
return None
gates = [gate for gate in draft.quality_gate_ids if gate in QUALITY_GATES] or ["cite_required"]
return StrategyRoleSet(
strategy="custom" if draft.strategy == "custom" else f"custom_{draft.strategy[:24]}",
title=draft.title[:80],
summary=draft.summary[:200],
rationale=draft.rationale[:300],
estimated_duration_seconds=max(300, min(int(draft.estimated_duration_seconds or 1500), 7200)),
estimated_cost_level=draft.estimated_cost_level if draft.estimated_cost_level in {"low", "medium", "high"} else "medium",
recommended=True,
slots=[
RoleSlot(
slot_id=node.slot_id,
demand=RoleDemand(
role_key=node.role_key,
display_name=node.display_name[:32] or node.role_key,
angle=(node.angle or None),
mission=node.mission[:400],
depends_on_roles=[],
output_artifact_type=node.output_artifact_type,
suggested_tools=[tool for tool in node.suggested_tools if tool][:4],
),
depends_on=list(node.depends_on),
)
for node in nodes
],
quality_gate_ids=gates,
)
@@ -1,17 +1,18 @@
"""Fulfill a durable plan_request: catalog → select → assemble → validate → persist.
Does not start a run. Rule-based until the BE-7 model adapter is wired.
Does not start a run. P1.2 起优先由规划师 LLM 生成定制策略,模板为回退。
"""
from __future__ import annotations
import logging
from typing import Any
from uuid import uuid4
from app.report_collaboration.contracts.requirements import ReportRequirementSnapshot
from app.report_collaboration.planning.assembler import PlanAssembler
from app.report_collaboration.planning.catalog import project_agent_catalog
from app.report_collaboration.planning.controller import PlanController
from app.report_collaboration.planning.planner_agent import LlmFactory
from app.report_collaboration.planning.validator import apply_validation, validate_assembled_plan
from app.report_collaboration.requirement.defaults import COORDINATOR_NAME, DEFAULT_TITLE
from deerflow.config.app_config import get_app_config
@@ -22,6 +23,8 @@ from deerflow.persistence.report_collaboration import (
ReportCollaborationValidationError,
)
logger = logging.getLogger(__name__)
_BUSY = frozenset({"running", "awaiting_input", "reviewing", "completed", "failed", "cancelled"})
@@ -51,12 +54,15 @@ class PlanProposalService:
*,
agent_store: Any | None = None,
config: ReportCollaborationConfig | None = None,
llm_factory: LlmFactory | None = None,
) -> None:
self._store = store
self._agent_store = agent_store
self._config = config
self._controller = PlanController()
self._assembler = PlanAssembler()
self._llm_factory = llm_factory
self._llm: Any | None = None
self._llm_tried = False
def _cfg(self) -> ReportCollaborationConfig:
return self._config or get_app_config().report_collaboration
@@ -110,12 +116,29 @@ class PlanProposalService:
requirement = requirement_from_session(record)
cfg = self._cfg()
catalog = await project_agent_catalog(user_id=user_id, config=cfg, agent_store=self._agent_store)
strategies = self._controller.propose(
requirement,
max_parallel_tasks=cfg.max_parallel_tasks,
group_id = f"rpg_{uuid4().hex[:12]}"
# P1-R:规划师智能体是唯一路径 —— 带着完整对话上下文设计工作流。
# 模型不可用即无法继续(无模板兜底)。
from app.report_collaboration.planning.planner_agent import PlannerAgent
llm = self._planner_llm()
history = [
{"role": "human" if item.get("role") == "human" else "ai", "content": str(item.get("content") or "")}
for item in await self._store.list_messages(session_id)
]
designed = await PlannerAgent(llm).design_workflow(
requirement=requirement,
history=history,
max_team_members=cfg.max_team_members,
)
group_id = f"rpg_{uuid4().hex[:12]}"
if designed is None:
raise ReportCollaborationValidationError(
"PLANNER_UNAVAILABLE", "规划师智能体未能生成有效工作流,请重试或换个说法再试"
)
strategies: list[Any] = [designed[0]]
planner_reply: str | None = designed[1]
assembled_list = [
self._assembler.assemble(
strategy,
@@ -140,13 +163,44 @@ class PlanProposalService:
for candidate in persisted:
saved.append(await self._store.insert_plan(session_id, candidate.model_dump()))
await self._store.update_command_status(command_id, status="completed")
content = planner_reply or "工作流已生成,可开始执行。"
await self._store.create_message(
session_id,
role="ai",
content="已根据当前需求生成 3 个候选协作方案。请选择其中一个后再开始执行;选择方案不会自动开跑。",
content=content,
idempotency_key=f"{idempotency_key}:plans",
name=COORDINATOR_NAME,
metadata={"plan_request_command_id": command_id, "proposal_group_id": group_id},
metadata={"plan_request_command_id": command_id, "proposal_group_id": group_id, "plan_source": "planner_agent"},
completed=True,
)
return saved
def _planner_llm(self) -> Any:
"""规划师模型:注入优先(测试);否则按 planner_llm_enabled 懒构建。不可用即抛错。"""
if self._llm_factory is not None:
llm = self._llm_factory()
if llm is None:
raise ReportCollaborationValidationError("PLANNER_UNAVAILABLE", "规划师智能体不可用,请检查模型配置")
return llm
try:
if not (self._config or get_app_config().report_collaboration).planner_llm_enabled:
raise ReportCollaborationValidationError(
"PLANNER_UNAVAILABLE", "规划师智能体未启用(report_collaboration.planner_llm_enabled)"
)
except ReportCollaborationValidationError:
raise
except Exception as exc:
raise ReportCollaborationValidationError("PLANNER_UNAVAILABLE", f"规划师智能体配置不可读:{exc}") from None
if not self._llm_tried:
self._llm_tried = True
try:
from app.report_collaboration.planning.planner_agent import build_planner_llm
self._llm = build_planner_llm()
except Exception as exc:
logger.warning("rc planner agent llm build failed: %s", exc)
self._llm = None
if self._llm is None:
raise ReportCollaborationValidationError("PLANNER_UNAVAILABLE", "规划师智能体不可用,请检查模型配置")
return self._llm