fix(assistant): 工具卡片不再永久转圈

src/backend/bisheng/tool/domain/langchain/knowledge.py

@@ -1,9 +1,9 @@
-from typing import Any, Type, List, Optional
+from typing import Any
 
 from langchain_classic.chains.combine_documents import create_stuff_documents_chain
-from langchain_core.documents import Document, BaseDocumentCompressor
+from langchain_core.documents import BaseDocumentCompressor, Document
 from langchain_core.language_models import BaseChatModel
-from langchain_core.prompts import SystemMessagePromptTemplate, HumanMessagePromptTemplate, ChatPromptTemplate
+from langchain_core.prompts import ChatPromptTemplate, HumanMessagePromptTemplate, SystemMessagePromptTemplate
 from langchain_core.retrievers import BaseRetriever
 from langchain_core.tools import BaseTool
 from pydantic import BaseModel, Field
@@ -34,17 +34,17 @@ class ToolInputSchema(BaseModel):
 class KnowledgeRetrieverTool(BaseTool):
     name: str = "knowledge_retriever_tool"
     description: str = "在知识库中检索与查询相关的文档内容。"
-    args_schema: Type[BaseModel] = ToolInputSchema
+    args_schema: type[BaseModel] = ToolInputSchema
 
-    vector_retriever: Optional[BaseRetriever] = None
-    elastic_retriever: Optional[BaseRetriever] = None
-    rerank: Optional[BaseDocumentCompressor] = None
+    vector_retriever: BaseRetriever | None = None
+    elastic_retriever: BaseRetriever | None = None
+    rerank: BaseDocumentCompressor | None = None
     max_content: int = Field(default=15000, description="The max length of the combined document content.")
     sort_by_source_and_index: bool = Field(default=False, description="Sort by document name & chunk index.")
-    rrf_weights: List[float] = Field(default=None)
+    rrf_weights: list[float] = Field(default=None)
     rrf_remove_zero_score: bool = Field(default=False)
 
-    def _run(self, query: str, **kwargs: Any) -> List[Document]:
+    def _run(self, query: str, **kwargs: Any) -> list[Document]:
         milvus_docs, es_docs = [], []
         if self.vector_retriever:
             milvus_docs = self.vector_retriever.invoke(query)
@@ -57,7 +57,7 @@ def _run(self, query: str, **kwargs: Any) -> List[Document]:
             finally_docs = self.rerank.compress_documents(finally_docs, query)
         return finally_docs
 
-    async def _arun(self, query: str, **kwargs: Any) -> List[Document]:
+    async def _arun(self, query: str, **kwargs: Any) -> list[Document]:
         milvus_docs, es_docs = [], []
         if self.vector_retriever:
             milvus_docs = await self.vector_retriever.ainvoke(query)
@@ -70,7 +70,7 @@ async def _arun(self, query: str, **kwargs: Any) -> List[Document]:
             finally_docs = await self.rerank.acompress_documents(finally_docs, query)
         return finally_docs
 
-    def _rrf_rerank(self, milvus_docs: List[Document], es_docs: List[Document], query: str) -> List[Document]:
+    def _rrf_rerank(self, milvus_docs: list[Document], es_docs: list[Document], query: str) -> list[Document]:
         if not milvus_docs and not es_docs:
             return []
         rrf_rerank = RRFRerank(
@@ -103,16 +103,16 @@ def _rrf_rerank(self, milvus_docs: List[Document], es_docs: List[Document], quer
 class KnowledgeRagTool(BaseTool):
     name: str
     description: str
-    args_schema: Type[BaseModel] = ToolInputSchema
+    args_schema: type[BaseModel] = ToolInputSchema
 
     llm: BaseChatModel
-    chat_prompt: Optional[ChatPromptTemplate] = CHAT_PROMPT
+    chat_prompt: ChatPromptTemplate | None = CHAT_PROMPT
 
-    vector_retriever: Optional[BaseRetriever] = None
-    elastic_retriever: Optional[BaseRetriever] = None
+    vector_retriever: BaseRetriever | None = None
+    elastic_retriever: BaseRetriever | None = None
     max_content: int = Field(default=15000, description="The max length of the combined document content.")
     sort_by_source_and_index: bool = Field(default=False, description="Sort by document name & chunk index.")
-    rrf_weights: List[float] = Field(default=None)
+    rrf_weights: list[float] = Field(default=None)
     rrf_remove_zero_score: bool = Field(default=False)
 
     knowledge_retriever_tool: KnowledgeRetrieverTool = None
@@ -121,7 +121,10 @@ class KnowledgeRagTool(BaseTool):
     def init_knowledge_rag_tool(cls, name: str, description: str, **kwargs) -> BaseTool:
         llm = kwargs.pop("llm")
         chat_prompt = kwargs.pop("chat_prompt", CHAT_PROMPT)
-        # cancel assistant deep callback
+        # The retriever is an internal step of this tool, not a tool call of its
+        # own — it must stay invisible to the caller's callbacks. Dropping them
+        # here is not enough on its own (LangChain also inherits handlers from
+        # the ambient run context), so `_run`/`_arun` bypass `invoke` as well.
         kwargs.pop("callbacks", None)
         knowledge_retriever_tool = KnowledgeRetrieverTool(**kwargs)
         return cls(
@@ -134,14 +137,18 @@ def init_knowledge_rag_tool(cls, name: str, description: str, **kwargs) -> BaseT
         )
 
     def _run(self, query: str) -> Any:
-        # 1. retrieve documents
-        retrieval_result = self.knowledge_retriever_tool.invoke({"query": query})
+        # 1. retrieve documents — called directly rather than through `invoke`,
+        # which would open a nested tool run. The chat then showed a second card
+        # per search, named after the retriever instead of the knowledge base
+        # ("知识库已被删除"), and it never closed because the retriever answers
+        # with Document objects that the run-log frame could not serialize.
+        retrieval_result = self.knowledge_retriever_tool._run(query)
         llm_inputs = self._get_llm_inputs(query, retrieval_result)
         qa_chain = create_stuff_documents_chain(llm=self.llm, prompt=self.chat_prompt)
         return qa_chain.invoke(llm_inputs)
 
     async def _arun(self, query: str) -> Any:
-        retrieval_result = await self.knowledge_retriever_tool.ainvoke({"query": query})
+        retrieval_result = await self.knowledge_retriever_tool._arun(query)
         llm_inputs = self._get_llm_inputs(query, retrieval_result)
         qa_chain = create_stuff_documents_chain(llm=self.llm, prompt=self.chat_prompt)
         return await qa_chain.ainvoke(llm_inputs)

src/backend/test/api/test_assistant_runlog_frames.py

@@ -0,0 +1,61 @@
+"""A tool card in the chat closes only when its end frame arrives.
+
+Two ways it never did. The knowledge tool ran its retriever through `invoke`,
+which opened a *nested* tool run: a second card per search, named after the
+retriever rather than the knowledge base ("知识库已被删除"), whose result is a
+list of ``Document`` objects that plain ``json.dumps`` refuses — so the end frame
+raised inside the callback and the card spun forever, past the end of the
+session, with nothing persisted to recover it from.
+"""
+
+from __future__ import annotations
+
+import inspect
+import json
+
+from langchain_core.documents import Document
+
+from bisheng.api.v1.callback import AsyncGptsDebugCallbackHandler, _dump_run_log
+from bisheng.tool.domain.langchain.knowledge import KnowledgeRagTool
+
+
+def test_a_document_answer_still_serializes() -> None:
+    """The retriever's own answer shape must not be able to suppress a frame."""
+
+    payload = {"tool_key": "4138", "output": [Document(page_content="hello")]}
+
+    decoded = json.loads(_dump_run_log(payload))
+
+    assert decoded["tool_key"] == "4138"
+    assert "hello" in decoded["output"][0]
+
+
+def test_chinese_survives_the_frame() -> None:
+    # ensure_ascii stays off: the card shows this text verbatim.
+    assert "知识库" in _dump_run_log({"output": "知识库内容"})
+
+
+def test_a_nameless_end_callback_is_not_fatal() -> None:
+    """`on_tool_end` reads the name from kwargs, which is not always populated."""
+
+    assert AsyncGptsDebugCallbackHandler.parse_tool_category(None) == ("", "tool")
+    assert AsyncGptsDebugCallbackHandler.parse_tool_category("") == ("", "tool")
+
+
+def test_a_knowledge_tool_is_still_recognised_by_its_id() -> None:
+    name, category = AsyncGptsDebugCallbackHandler.parse_tool_category("knowledge_4138")
+    assert (name, category) == ("4138", "knowledge")
+
+
+def test_the_retriever_is_an_internal_step_not_a_tool_call() -> None:
+    """Going through `invoke` re-opens the callback machinery for the inner tool.
+
+    That is what produced the duplicate, wrongly-named card; the direct call
+    keeps the retrieval invisible to whoever is watching the outer tool.
+    """
+
+    for source in (inspect.getsource(KnowledgeRagTool._run), inspect.getsource(KnowledgeRagTool._arun)):
+        assert "knowledge_retriever_tool.invoke" not in source
+        assert "knowledge_retriever_tool.ainvoke" not in source
+    assert "knowledge_retriever_tool._run(query)" in inspect.getsource(KnowledgeRagTool._run)
+    assert "knowledge_retriever_tool._arun(query)" in inspect.getsource(KnowledgeRagTool._arun)

src/frontend/client/src/locales/en/translation.json

@@ -698,6 +698,7 @@
   "com_runlog_done": "Done",
   "com_runlog_flow_deleted": "The skill has been deleted and the name cannot be retrieved",
   "com_runlog_knowledge_deleted": "The knowledge base has been deleted and the name cannot be retrieved",
+  "com_runlog_interrupted": "The tool call returned no result (the session ended)",
   "com_runlog_offline": "Offline",
   "com_runlog_searched": "Searched",
   "com_runlog_searching": "Searching",

src/frontend/client/src/locales/ja/translation.json

@@ -679,6 +679,7 @@
   "com_runlog_done": "完了",
   "com_runlog_flow_deleted": "スキルが削除されたため、スキル名を取得できません",
   "com_runlog_knowledge_deleted": "ナレッジベースが削除されたため、名称を取得できません",
+  "com_runlog_interrupted": "ツール呼び出しは結果を返しませんでした（セッション終了）",
   "com_runlog_offline": "オフライン",
   "com_runlog_searched": "検索済み",
   "com_runlog_searching": "検索中",

src/frontend/client/src/locales/zh-Hans/translation.json

@@ -682,6 +682,7 @@
   "com_runlog_done": "完成",
   "com_runlog_flow_deleted": "技能已被删除，无法获取技能名",
   "com_runlog_knowledge_deleted": "知识库已被删除，无法获取知识库名",
+  "com_runlog_interrupted": "工具调用未返回结果（会话已结束）",
   "com_runlog_offline": "已下线",
   "com_runlog_searched": "已搜索",
   "com_runlog_searching": "正在搜索",

src/frontend/client/src/pages/appChat/appUtils/skillMethod.ts

@@ -1,6 +1,6 @@
 import { formatDate } from "~/utils";
 
-const runLogsTypes = ['tool', 'flow', 'knowledge']
+export const runLogsTypes = ['tool', 'flow', 'knowledge']
 // 兼容处理技能和助手
 export const SkillMethod = {
     /** 获取input发送参数 */

src/frontend/client/src/pages/appChat/components/MessageRunlog.tsx

@@ -16,6 +16,9 @@ export default function MessageRunlog({ data }) {
     }, [_chatsState, data])
 
     const [title, lost] = useMemo(() => {
+        // Settled by the session closing rather than by its own end frame — say so
+        // instead of showing a success tick the call never earned.
+        if (data.interrupted) return [t('com_runlog_interrupted'), true]
         let lost = false
         let title = ''
         const status = data.end ? t('com_runlog_used') : t('com_runlog_using')

src/frontend/client/src/pages/appChat/useChatHelpers.ts

@@ -8,7 +8,7 @@ import { Chat } from "~/@types/chat"
 import { baseMsgItem } from "~/api/apps"
 import { formatDate, generateUUID } from "~/utils"
 import { FLOW_TYPES } from "."
-import { SkillMethod } from "./appUtils/skillMethod"
+import { runLogsTypes, SkillMethod } from "./appUtils/skillMethod"
 import { bishengConfState, chatApiVersionState, chatIdState, chatsState, currentChatState, currentRunningState, runningState } from "./store/atoms"
 import { emitAreaTextEvent, EVENT_TYPE } from "./useAreaText"
 
@@ -341,6 +341,20 @@ export default function useChatHelpers() {
             )
         },
         skillCloseMsg: () => {
+            // A tool card only closes when its `end` frame arrives. Lose one — a
+            // serialization failure, a dropped socket — and it spins forever with
+            // nothing to recover it: the round is over and nothing was persisted.
+            // Settle what is still open, marked interrupted rather than wearing a
+            // success tick it never earned.
+            setChats((prev) =>
+                updateChatMessages(prev, chatId, (messages) =>
+                    messages.map((msg) =>
+                        runLogsTypes.includes(msg.category) && !msg.end
+                            ? { ...msg, end: true, interrupted: true }
+                            : msg
+                    )
+                )
+            )
             setRunningState((prev) => {
                 return {
                     ...prev,

src/frontend/platform/public/locales/en-US/bs.json

@@ -557,7 +557,8 @@
       "flowOffline": "{{name}} is offline",
       "flowDeleted": "The skill has been deleted, its name is unavailable",
       "toolDeleted": "The tool has been deleted, its name is unavailable",
-      "knowledgeDeleted": "The knowledge base has been deleted, its name is unavailable"
+      "knowledgeDeleted": "The knowledge base has been deleted, its name is unavailable",
+      "interrupted": "The tool call returned no result (the session ended)"
     }
   },
   "importLinsight": {

src/frontend/platform/public/locales/ja/bs.json

@@ -547,7 +547,8 @@
       "flowOffline": "{{name}} はオフラインです",
       "flowDeleted": "スキルが削除されているため、スキル名を取得できません",
       "toolDeleted": "ツールが削除されているため、ツール名を取得できません",
-      "knowledgeDeleted": "ナレッジベースが削除されているため、名称を取得できません"
+      "knowledgeDeleted": "ナレッジベースが削除されているため、名称を取得できません",
+      "interrupted": "ツール呼び出しは結果を返しませんでした（セッション終了）"
     }
   },
   "model": {

src/frontend/platform/public/locales/zh-Hans/bs.json

@@ -552,7 +552,8 @@
       "flowOffline": "{{name}} 已下线",
       "flowDeleted": "技能已被删除，无法获取技能名",
       "toolDeleted": "工具已被删除，无法获取工具名",
-      "knowledgeDeleted": "知识库已被删除，无法获取知识库名"
+      "knowledgeDeleted": "知识库已被删除，无法获取知识库名",
+      "interrupted": "工具调用未返回结果（会话已结束）"
     }
   },
   "model": {

src/frontend/platform/src/components/bs-comp/chatComponent/ChatInput.tsx

@@ -29,7 +29,7 @@ export default function ChatInput({ clear, form, questions, inputForm, wsUrl, on
 
     const { isLoading: audioOpening } = useAudioStore()
 
-    const { messages, hisMessages, chatId, createSendMsg, createWsMsg, updateCurrentMessage, destory, setShowGuideQuestion } = useMessageStore()
+    const { messages, hisMessages, chatId, createSendMsg, createWsMsg, updateCurrentMessage, closeDanglingRunLogs, destory, setShowGuideQuestion } = useMessageStore()
     const currentChatIdRef = useRef(null)
     const inputRef = useRef(null)
     const continueRef = useRef(false)
@@ -256,6 +256,7 @@ export default function ChatInput({ clear, form, questions, inputForm, wsUrl, on
 
             if (!msgClosedRef.current) msgClosedRef.current = true
         } else if (data.type === "close") {
+            closeDanglingRunLogs()
             setStop({ show: false, disable: false })
             setInputLock((prev) => (prev.reason ? prev : { locked: false, reason: '' }))
         }

src/frontend/platform/src/components/bs-comp/chatComponent/RunLog.tsx

@@ -20,6 +20,9 @@ export default function RunLog({ data }) {
     const assistantState = useAssistantStore(state => state.assistantState)
 
     const [title, lost] = useMemo(() => {
+        // Settled by the session closing rather than by its own end frame — say so
+        // instead of showing a success tick the call never earned.
+        if (data.interrupted) return [t('chat.runLog.interrupted'), true]
         let lost = false
         let title = ''
         const status = data.end ? t('chat.runLog.used') : t('chat.runLog.using')

src/frontend/platform/src/components/bs-comp/chatComponent/messageStore.ts

@@ -41,6 +41,7 @@ type Actions = {
     insetSystemMsg: (text: string) => void;
     insetBsMsg: (text: string) => void;
     setShowGuideQuestion: (text: boolean) => void;
+    closeDanglingRunLogs: () => void;
     clearMsgs: () => void;
 }
 
@@ -152,6 +153,24 @@ export const useMessageStore = create<State & Actions>((set, get) => ({
             set({ historyEnd: true })
         }
     },
+    /**
+     * A tool card only closes when its `end` frame arrives. Lose one — a
+     * serialization failure, a dropped socket — and the card spins forever, with
+     * nothing to recover it: the round is over and nothing was persisted. So on
+     * session close, settle whatever is still open and mark it interrupted
+     * rather than leaving a success tick it never earned.
+     */
+    closeDanglingRunLogs() {
+        const messages = get().messages
+        if (!messages.some(msg => runLogsTypes.includes(msg.category) && !msg.end)) return
+        set({
+            messages: messages.map(msg =>
+                runLogsTypes.includes(msg.category) && !msg.end
+                    ? { ...msg, end: true, interrupted: true }
+                    : msg
+            )
+        })
+    },
     clearMsgs() {
         setTimeout(() => {
             set({ hisMessages: [], messages: [], historyEnd: true })

src/frontend/platform/src/test/assistantRunLogClosing.test.ts

@@ -0,0 +1,58 @@
+import { useMessageStore } from "@/components/bs-comp/chatComponent/messageStore";
+import { ChatMessageType } from "@/types/chat";
+import { beforeEach, describe, expect, it } from "vitest";
+
+type RunLogCard = ChatMessageType & { interrupted?: boolean };
+
+const asMessages = (rows: unknown[]) => rows as ChatMessageType[];
+const readMessages = () => useMessageStore.getState().messages as RunLogCard[];
+
+/**
+ * A tool card closes only when its own end frame arrives. One got lost — the
+ * knowledge retriever answered with objects the frame could not serialize — and
+ * the card kept spinning after the session had ended, with nothing persisted to
+ * recover it from. Closing the session must settle whatever is still open.
+ */
+describe("dangling run-log cards at session close", () => {
+  const runLog = (id: string, end: boolean) => ({
+    id,
+    category: "knowledge",
+    end,
+    message: { tool_key: "4138" },
+    thought: "",
+  });
+
+  beforeEach(() => {
+    useMessageStore.setState({ messages: [], hisMessages: [] });
+  });
+
+  it("settles a tool card whose end frame never arrived", () => {
+    useMessageStore.setState({ messages: asMessages([runLog("a", true), runLog("b", false)]) });
+
+    useMessageStore.getState().closeDanglingRunLogs();
+
+    const [settled, interrupted] = readMessages();
+    expect(settled.interrupted).toBeUndefined();
+    // Marked, not silently ticked: it never earned a success icon.
+    expect(interrupted.end).toBe(true);
+    expect(interrupted.interrupted).toBe(true);
+  });
+
+  it("leaves the streaming answer alone", () => {
+    const answer = { id: "answer", category: "answer", end: false, message: "", thought: "" };
+    useMessageStore.setState({ messages: asMessages([answer]) });
+
+    useMessageStore.getState().closeDanglingRunLogs();
+
+    expect(readMessages()[0].end).toBe(false);
+  });
+
+  it("does not touch the list when every card is already closed", () => {
+    const messages = asMessages([runLog("a", true)]);
+    useMessageStore.setState({ messages });
+
+    useMessageStore.getState().closeDanglingRunLogs();
+
+    expect(useMessageStore.getState().messages).toBe(messages);
+  });
+});