From afb78a6943a2415462c62692cc62882d8150e4d3 Mon Sep 17 00:00:00 2001 From: fengxiangdebaozi <10845145+fengxiangdebaozi@user.noreply.gitee.com> Date: Wed, 24 Jun 2026 22:19:18 +0800 Subject: [PATCH] =?UTF-8?q?feat(ai=E5=86=99=E4=BD=9C):=20=E7=B4=A0?= =?UTF-8?q?=E6=9D=90=E6=94=B6=E9=9B=86=E4=B8=93=E5=AE=B6=E6=8C=89=E7=9C=9F?= =?UTF-8?q?=E5=AE=9E=E6=99=BA=E8=83=BD=E4=BD=93=E8=BF=90=E8=A1=8C=E6=97=B6?= =?UTF-8?q?=E8=B7=91=E6=8A=80=E8=83=BD=20+=20=E6=AD=A5=E9=AA=A4=E6=9D=A1?= =?UTF-8?q?=E5=AF=B9=E9=BD=90?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - 新增 run_skill_via_real_agent：复用 SubagentExecutor（与 lead agent / 子智能体 同一套运行时），加载技能 SKILL.md + 全套工具 + 懒加载沙箱，让技能的真实检索逻辑 （读文件 / 调内网接口 / MCP）真正执行，再按 JSON 契约收割结构化素材；researcher 的 _execute_skill 在 researcher_skill_agent 开启时切到它，失败/空结果回退直连检索。 解决内网技能（在别的智能体里好使）在 AI 写作里检不出东西的问题。 - 步骤条轻量对齐别的智能体：技能调用拆成「读取 SKILL.md → 执行检索 → 检索到 N 条」 三段（前两段新增，结果段沿用 skill_result）；前端 SkillStepsCard 零改动。 - 含 roundtable-planning（业务链选择 / Step2 编排）的进行中改动。 Co-Authored-By: Claude Opus 4.8 (1M context) --- .../components/BusinessChainPicker.tsx | 2 +- .../hooks/useStep2Orchestration.ts | 2 +- .../roundtable-planning/lib/business-chain.ts | 2 +- .../pages/RoundtablePlanningPage.tsx | 12 +- .../agents/ai_writing/nodes/_skill_agent.py | 142 ++++++++++++++++++ .../agents/ai_writing/nodes/researcher.py | 21 ++- 6 files changed, 164 insertions(+), 17 deletions(-) diff --git a/frontend-web/src/roundtable-planning/components/BusinessChainPicker.tsx b/frontend-web/src/roundtable-planning/components/BusinessChainPicker.tsx index 88de95a1..dbdfcc4a 100644 --- a/frontend-web/src/roundtable-planning/components/BusinessChainPicker.tsx +++ b/frontend-web/src/roundtable-planning/components/BusinessChainPicker.tsx @@ -140,7 +140,7 @@ export function BusinessChainPicker({ const hasCoord = (plan.coordinatorPrompt?.trim() ?? "") !== ""; return hasParallel || hasGoal || hasCoord ? plan : null; }, [selectedChain, survivingSeats]); - const [dagExecMode, setDagExecMode] = useState<"fast" | "sandbox">("fast"); + const [dagExecMode, setDagExecMode] = useState<"fast" | "sandbox">("sandbox"); const handleUse = () => { if (!canUse || !selectedChain) return; diff --git a/frontend-web/src/roundtable-planning/hooks/useStep2Orchestration.ts b/frontend-web/src/roundtable-planning/hooks/useStep2Orchestration.ts index 4f843cfc..edd770ba 100644 --- a/frontend-web/src/roundtable-planning/hooks/useStep2Orchestration.ts +++ b/frontend-web/src/roundtable-planning/hooks/useStep2Orchestration.ts @@ -460,7 +460,7 @@ export function useStep2Orchestration( gatherFirst = false, orchestrationMode, orchestrationPlan, - dagExecMode = "fast", + dagExecMode = "sandbox", buildInitialIntent, triggerToast, setRoles, diff --git a/frontend-web/src/roundtable-planning/lib/business-chain.ts b/frontend-web/src/roundtable-planning/lib/business-chain.ts index fc3918c8..bfec7ac6 100644 --- a/frontend-web/src/roundtable-planning/lib/business-chain.ts +++ b/frontend-web/src/roundtable-planning/lib/business-chain.ts @@ -192,7 +192,7 @@ export function buildChainSeatSelection( const agents = plan.stages .flatMap((st) => st.agentIds) .map((id) => ({ agent_id: id, name: nameOf(id), model: modelOf(id), seat_skill_directive: skillOf(id), seatMode: seatModeOf(id) })); - return { agents, mode: "dag", chain: chainMeta, plan, dagExecMode: "fast", seatMode: chain.seatMode, gatherFirst: chain.gatherFirst ?? false }; + return { agents, mode: "dag", chain: chainMeta, plan, dagExecMode: "sandbox", seatMode: chain.seatMode, gatherFirst: chain.gatherFirst ?? false }; } } diff --git a/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx b/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx index fb488f64..fb40a2ea 100644 --- a/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx +++ b/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx @@ -360,7 +360,7 @@ export default function RoundtablePlanningPage() { const [orchestrationPlan, setOrchestrationPlan] = useState(null); // dag 执行子模式：fast（并行双列禁写）/ sandbox（并行单列可写、沙箱跟随）。 // 由推荐弹窗的「执行方式」选择决定（仅 dag 有意义）。 - const [dagExecMode, setDagExecMode] = useState<"fast" | "sandbox">("fast"); + const [dagExecMode, setDagExecMode] = useState<"fast" | "sandbox">("sandbox"); const [step2SidebarExpanded, setStep2SidebarExpanded] = useState(true); const [step2Runs, setStep2Runs] = useState([]); const [activeStep2RunId, setActiveStep2RunId] = useState(null); @@ -684,7 +684,7 @@ export default function RoundtablePlanningPage() { setOrchestrationMode(run.orchestrationMode ?? "recommend"); setChainMeta(run.chain ?? null); setOrchestrationPlan(run.orchestrationPlan ?? null); - setDagExecMode(run.dagExecMode ?? "fast"); + setDagExecMode(run.dagExecMode ?? "sandbox"); setStep3Report(run.step3 ?? null); step2.hydrateFromDraft(run as unknown as Step2Snapshot); markRestoredConsensusAsEmitted(Boolean(run.hasConsensus)); @@ -791,7 +791,7 @@ export default function RoundtablePlanningPage() { setOrchestrationMode(snapshot.orchestrationMode ?? "recommend"); setChainMeta(snapshot.chain ?? null); setOrchestrationPlan(snapshot.orchestrationPlan ?? null); - setDagExecMode(snapshot.dagExecMode ?? "fast"); + setDagExecMode(snapshot.dagExecMode ?? "sandbox"); step2.hydrateFromDraft(snapshot as unknown as Step2Snapshot); markRestoredConsensusAsEmitted(Boolean(snapshot.hasConsensus)); }, @@ -945,7 +945,7 @@ export default function RoundtablePlanningPage() { // chain / dag 都来自业务链条，保留来源链名；recommend 无。 chain: result.mode === "recommend" ? null : result.chain ?? null, orchestrationPlan: result.mode === "dag" ? result.plan ?? null : null, - dagExecMode: result.mode === "dag" ? result.dagExecMode ?? "fast" : undefined, + dagExecMode: result.mode === "dag" ? result.dagExecMode ?? "sandbox" : undefined, step3: null, }; setStep2Runs([...syncedRuns, newRun]); @@ -998,7 +998,7 @@ export default function RoundtablePlanningPage() { setGatherFirst(result.gatherFirst ?? false); setChainMeta(result.mode === "recommend" ? null : result.chain ?? null); setOrchestrationPlan(result.mode === "dag" ? result.plan ?? null : null); - setDagExecMode(result.mode === "dag" ? result.dagExecMode ?? "fast" : "fast"); + setDagExecMode(result.mode === "dag" ? result.dagExecMode ?? "sandbox" : "sandbox"); setIsRecommendDialogOpen(false); setCurrentStep(2); triggerToast( @@ -1527,7 +1527,7 @@ export default function RoundtablePlanningPage() { setOrchestrationMode("recommend"); setChainMeta(null); setOrchestrationPlan(null); - setDagExecMode("fast"); + setDagExecMode("sandbox"); setStep3Report(null); // 停掉页面层的 Step3 三个生成流（新任务不该带着旧任务的流跑）+ 复位进入闸门， // 并停止后台作业实时跟随（作业本身继续在后台跑）。 diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/_skill_agent.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/_skill_agent.py index 42ee395e..524dea79 100644 --- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/_skill_agent.py +++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/_skill_agent.py @@ -41,6 +41,148 @@ def use_skill_search_agent() -> bool: return bool(cfg.get("researcher_skill_agent", False)) if isinstance(cfg, dict) else False +# ── 用真实智能体运行时跑技能（和「别的智能体调用技能」完全一致）──────────────────── +# +# 旧的 run_skill_search_agent 只给模型绑了一个写死路由到 DDG 的通用 search 工具，技能 +# SKILL.md 里要用的真实工具（read_file / bash / 内网检索工具 / MCP）根本不存在，所以 +# 内网技能跑不出东西。本函数改为复用 **SubagentExecutor**（与 lead agent / 子智能体同一套 +# 运行时）：加载技能 SKILL.md（skills 白名单）+ 全套工具 + 懒加载沙箱，让技能按其说明真正 +# 检索，再让模型把检索到的真实素材按 JSON 契约输出，我们解析成结构化素材。 +# +# 单次工具调用循环上限（防跑飞）。 +_REAL_AGENT_MAX_TURNS = 16 + +_MATERIAL_OUTPUT_CONTRACT = ( + "你的任务：严格按上面挂载的检索技能（SKILL.md）的说明，用相应工具检索与「写作意图」相关的**真实素材**。\n" + "- 必须按技能说明实际调用工具完成检索（读取知识库文件 / 调用内网检索接口 / MCP 等），**不得凭记忆编造**。\n" + "- 检索完成后，**只输出一个 JSON 对象**（不要解释、不要 markdown 代码块、用英文双引号），把检索到的真实素材列出：\n" + '{"materials": [{"title": "标题", "content": "素材正文摘录(尽量保留原文, <=500字)", "source": "来源", "url": "原文链接(可空)", "published_at": "日期(可空)"}]}\n' + "- content 必须来自检索到的真实内容；若确实没检索到任何素材，输出 {\"materials\": []}。" +) + + +async def run_skill_via_real_agent( + *, + skill, + intent: str, + keywords: list[str], + user_query: str, + config: RunnableConfig, + state: dict, + model_name_fallback: str | None, + max_results: int, + timeout_s: int, + emit_query=None, +) -> tuple[list, str]: + """用真实智能体运行时（SubagentExecutor）跑一个检索技能，收割结构化素材。 + + 与「别的智能体调用技能」一致：技能 SKILL.md + 全套工具 + 沙箱都齐全，技能的真实检索 + 逻辑（读文件 / 调内网接口 / MCP）能真正执行。模型最终按 JSON 契约吐出素材，解析成 + SearchResult。返回 ``(results, error_detail)``：results 为空且 error 非空 = 该技能失败 + （researcher 据此回退直连检索）。 + """ + from deerflow.config import get_app_config + from deerflow.subagents.config import SubagentConfig + from deerflow.subagents.executor import SubagentExecutor, SubagentStatus + from deerflow.tools import get_available_tools + + from ..search.base import SearchResult + from ._agent_runner import RESEARCHER_AGENT_ID, WritingAgentUnavailable, load_writing_agent + from ._utils import parse_json + + skill_name = getattr(skill, "name", "") or "技能" + + try: + agent_cfg, soul = load_writing_agent(RESEARCHER_AGENT_ID) + except WritingAgentUnavailable: + agent_cfg, soul = None, "" + model_name = resolve_agent_model_name( + agent_cfg, state, config, explicit_fallback=model_name_fallback, strict=False + ) + + try: + app_config = get_app_config() + tools = get_available_tools( + model_name=model_name, subagent_enabled=False, include_mcp=True, app_config=app_config + ) + except Exception as e: # noqa: BLE001 + logger.warning("skill real-agent[%s]: build tools failed: %s", skill_name, e) + return [], f"工具初始化失败: {e}" + + system_prompt = (soul + "\n\n" if soul else "") + _MATERIAL_OUTPUT_CONTRACT + + task_lines = [f"写作意图：{intent or '(无)'}"] + if user_query: + task_lines.append(f"用户补充检索意图：{user_query}") + if keywords: + task_lines.append(f"可参考的检索方向：{', '.join(keywords[:8])}") + task_lines.append(f"目标素材数量约 {max_results} 条。") + task = "\n".join(task_lines) + + sub_cfg = SubagentConfig( + name=f"ai-writing-skill-{skill_name}", + description="AI 写作素材检索", + system_prompt=system_prompt, + tools=None, # 继承全套工具，技能按 SKILL.md 自行取用 + disallowed_tools=["task"], # 素材检索不需要再派生子任务 + skills=[skill_name], # 仅加载这个技能的 SKILL.md + model=model_name or "inherit", + max_turns=_REAL_AGENT_MAX_TURNS, + timeout_seconds=max(30, timeout_s), + ) + + thread_id = (config.get("configurable") or {}).get("thread_id") + try: + executor = SubagentExecutor( + config=sub_cfg, tools=tools, app_config=app_config, + parent_model=model_name, thread_id=thread_id, + ) + except Exception as e: # noqa: BLE001 + logger.warning("skill real-agent[%s]: executor init failed: %s", skill_name, e) + return [], f"运行时初始化失败: {e}" + + # 步骤条「轻量对齐」别的智能体：读取 SKILL.md → 执行检索 两段（结果段由 researcher + # 的 _emit_skill_outcome 以 skill_result 收尾）。这两步在图的事件循环里同步发出， + # 写到时间线没问题；技能真正执行在下面的线程里跑。 + if emit_query: + emit_query(f"读取技能「{skill_name}」的 SKILL.md…") + emit_query(f"技能「{skill_name}」按 SKILL.md 说明执行检索…") + + # execute() 内部用「沙箱/MCP 友好」的隔离事件循环跑；放进线程，避免阻塞图的事件循环。 + try: + result = await asyncio.to_thread(executor.execute, task) + except Exception as e: # noqa: BLE001 + logger.warning("skill real-agent[%s]: execute failed: %s", skill_name, e, exc_info=True) + return [], f"技能检索失败: {e}" + + if result.status != SubagentStatus.COMPLETED: + return [], result.error or f"技能检索未完成（{result.status.value}）" + + parsed = parse_json(result.result or "", {"materials": []}) + raw_materials = parsed.get("materials") + if not isinstance(raw_materials, list): + return [], "" + + results: list = [] + for m in raw_materials[:max_results]: + if not isinstance(m, dict): + continue + content = str(m.get("content") or "").strip() + title = str(m.get("title") or "").strip() + if not content and not title: + continue + results.append(SearchResult( + title=title or "（无标题）", + content=content[:500], + source=str(m.get("source") or skill_name).strip() or skill_name, + published_at=str(m.get("published_at") or "").strip(), + url=str(m.get("url") or "").strip(), + relevance_score=0.8, + source_type="skill", + )) + return results, "" + + def _read_skill_body(skill) -> str: """读技能 SKILL.md 全文；失败回退 name+description。""" try: diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py index aaa4458a..786b0567 100644 --- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py +++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py @@ -40,7 +40,7 @@ from ._agent_runner import ( run_writing_agent, use_builtin_agent, ) -from ._skill_agent import run_skill_search_agent, use_skill_search_agent +from ._skill_agent import run_skill_via_real_agent, use_skill_search_agent from ._utils import get_ai_writing_config, keyword_outline_section, llm_json, safe_get_stream_writer logger = logging.getLogger(__name__) @@ -715,17 +715,20 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict error_detail = "；".join(dict.fromkeys(errs))[:200] return added, error_detail else: - # ── 智能体式技能调用（config ai_writing.researcher_skill_agent 打开时）── - # 把技能 SKILL.md 注入 + 绑定 search 工具，让模型按技能说明驱动检索， - # 收割工具返回的 SearchResult 作为素材（与「智能体调用技能」逻辑一致）。 + # ── 用真实智能体运行时跑技能（config ai_writing.researcher_skill_agent 打开时）── + # 复用 SubagentExecutor：技能 SKILL.md + 全套工具 + 沙箱齐全，技能真实检索逻辑 + # （读文件 / 调内网接口 / MCP）能真正执行，再按 JSON 契约收割结构化素材——和 + # 「别的智能体调用技能」完全一致，解决内网技能检不出东西的问题。 if use_skill_search_agent(): - def _emit_q(q: str, _name=sk.name) -> None: + def _emit_q(msg: str, _name=sk.name) -> None: + # run_skill_via_real_agent 传入的是完整步骤文案（读取 SKILL.md / + # 执行检索…），原样透传到步骤条，与别的智能体的技能调用步骤对齐。 _emit_skill_step({ "type": "invoke_skill", - "message": f"素材收集专家用技能「{_name}」检索：{q[:60]}", + "message": msg, "skill_name": _name, }) - agent_results, agent_err = await run_skill_search_agent( + agent_results, agent_err = await run_skill_via_real_agent( skill=sk, intent=intent, keywords=keywords, @@ -739,7 +742,9 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict ) if agent_results: return _collect(agent_results), "" - # 弱模型不会工具调用 / 智能体检索失败 → 回退直连检索，保证素材不空。 + # 技能运行时失败 / 没检索到 → 回退直连检索，保证素材不空。 + if agent_err: + logger.info("skill real-agent[%s] returned no materials: %s", sk.name, agent_err) # 定向技能检索（默认路径 / 智能体模式回退）：user_query 模式优先用用户原话； # 非知识库技能再附上 description 让 query 贴合该技能的领域。