fix: preserve agent debug conversation context

ai_knowledge_base_v2/apps/admin-web/src/components/AgentManagementView.vue

@@ -275,6 +275,10 @@ async function debugAgent() {
   if (!agentForm.question.trim()) return ElMessage.error("请输入调试问题");
   if (!promptContent.value.trim()) return ElMessage.error("主提示词不能为空");
   const question = agentForm.question.trim();
+  const conversationHistory = agentPreviewMessages.value
+    .slice(1)
+    .filter((message) => ["user", "assistant"].includes(message.role) && message.content.trim() && !message.streaming)
+    .map((message) => ({ role: message.role, content: message.content.trim() }));
   agentPreviewMessages.value.push({ role: "user", content: question });
   agentPreviewMessages.value.push({ role: "assistant", content: "", streaming: true });
   const assistantIndex = agentPreviewMessages.value.length - 1;
@@ -291,6 +295,7 @@ async function debugAgent() {
         modelId: agentForm.modelId,
         knowledgeIds: agentForm.knowledgeIds,
         question,
+        history: conversationHistory,
         temperature: agentForm.temperature,
         topP: agentForm.topP,
         topK: agentForm.topK,

ai_knowledge_base_v2/apps/backend/app/schemas/admin.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 from datetime import datetime
+from typing import Literal
 
 from pydantic import BaseModel, Field
 
@@ -77,12 +78,18 @@ class PromptSaveRequest(BaseModel):
     promptContent: str = Field(min_length=1)
 
 
+class AgentDebugHistoryMessage(BaseModel):
+    role: Literal["user", "assistant"]
+    content: str = Field(min_length=1, max_length=20000)
+
+
 class AgentDebugRequest(BaseModel):
     promptContent: str = Field(min_length=1)
     modelId: int = Field(gt=0)
     knowledgeIds: list[int] = Field(default_factory=list)
     knowledgeVersions: dict[int, int] = Field(default_factory=dict)
     question: str = Field(min_length=1, max_length=2000)
+    history: list[AgentDebugHistoryMessage] = Field(default_factory=list, max_length=100)
     temperature: float | None = Field(default=None, ge=0, le=2)
     topP: float | None = Field(default=None, ge=0, le=1)
     topK: int | None = Field(default=None, ge=1, le=1000)

ai_knowledge_base_v2/apps/backend/app/services/agent_debug_service.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import asyncio
 from collections.abc import AsyncIterator
+from types import SimpleNamespace
 
 from sqlalchemy.orm import Session
 
@@ -12,31 +13,33 @@ from app.services.admin_service import OperationLogService
 from app.services.knowledge_agent_service import KnowledgeAgentService
 from app.services.model_stream_service import ModelStreamService
 from app.services.reasoning_policy_service import ReasoningPolicyService
-from app.services.rag_service import PromptService, RagResult
+from app.services.rag_service import RagResult
 
 
 class AgentDebugService:
     @staticmethod
     async def build_result(db: Session, payload: AgentDebugRequest) -> RagResult:
+        history = [
+            SimpleNamespace(id=index + 1, role=item.role, content=item.content)
+            for index, item in enumerate(payload.history)
+        ]
         rag_result = await KnowledgeAgentService.build_result(
             db,
             question=payload.question,
+            history=history,
             version_overrides=payload.knowledgeVersions,
             preview_knowledge_ids=payload.knowledgeIds,
+            prompt_override=payload.promptContent,
         )
-        debug_messages = [
-            {"role": "system", "content": payload.promptContent.strip()},
-            *(rag_result.messages or []),
-        ]
         return RagResult(
             question=rag_result.question,
             knowledge_scopes=rag_result.knowledge_scopes,
             chunks=rag_result.chunks,
-            prompt=PromptService.render_messages(debug_messages),
+            prompt=rag_result.prompt,
             allow_general_knowledge=rag_result.allow_general_knowledge,
             retrieval_log_id=rag_result.retrieval_log_id,
             tool_trace=rag_result.tool_trace,
-            messages=debug_messages,
+            messages=rag_result.messages,
         )
 
     @staticmethod

ai_knowledge_base_v2/apps/backend/app/services/knowledge_agent_service.py

@@ -83,6 +83,7 @@ class KnowledgeAgentService:
         version_overrides: dict[int, int] | None = None,
         preview_knowledge_ids: list[int] | None = None,
         context_trace: list[dict] | None = None,
+        prompt_override: str | None = None,
     ) -> RagResult:
         started = perf_counter()
         catalog = cls.get_knowledge_catalog(
@@ -205,6 +206,7 @@ class KnowledgeAgentService:
                 history,
                 session_summary,
                 summary_up_to_message_id,
+                prompt_override=prompt_override,
             )
             return RagResult(
                 question=question,
@@ -309,15 +311,23 @@ class KnowledgeAgentService:
         text = question.strip()
         if len(text) <= 16:
             return True
+        if re.search(r"(上面|前面|之前|刚才|此前|前一个|原来).{0,8}(问题|内容|回答|练习|流程)", text):
+            return True
+        if re.search(r"(再|重新|继续).{0,6}(回答|解释|说明|分析|看看)", text):
+            return True
         return bool(re.search(r"(这个|那个|上述|上面|前面|其中|第二|继续|然后呢|怎么办|怎么做|什么意思|为什么)$", text))
 
     @staticmethod
     def _fallback_rewrite(question: str, history) -> str:
-        previous_user = next(
+        previous_users = [
-            (message.content.strip() for message in reversed(history) if message.role == "user" and message.content.strip()),
+            message.content.strip()
-            "",
+            for message in history
-        )
+            if message.role == "user" and message.content.strip()
-        return f"关于“{previous_user}”的追问：{question.strip()}" if previous_user else question.strip()
+        ][-3:]
+        if not previous_users:
+            return question.strip()
+        context = "\n".join(f"历史问题{index + 1}：{content}" for index, content in enumerate(previous_users))
+        return f"结合以下历史问题回答当前追问：\n{context}\n当前追问：{question.strip()}"
 
     @staticmethod
     def get_knowledge_catalog(

ai_knowledge_base_v2/apps/backend/app/services/rag_service.py

@@ -118,8 +118,9 @@ class PromptService:
         history: list[ChatMessage] | None = None,
         session_summary: str | None = None,
         summary_up_to_message_id: int | None = None,
+        prompt_override: str | None = None,
     ) -> list[dict[str, str]]:
-        prompt = cls._load_active_prompt(db)
+        prompt = prompt_override.strip() if prompt_override and prompt_override.strip() else cls._load_active_prompt(db)
 
         # 知识库上下文
         context = "\n\n".join(cls._format_chunk(index, chunk) for index, chunk in enumerate(chunks, start=1))

ai_knowledge_base_v2/apps/backend/tests/test_agent_runtime_config.py

@@ -1,7 +1,7 @@
 import asyncio
 from decimal import Decimal
 import json
-from unittest.mock import patch
+from unittest.mock import AsyncMock, patch
 
 import pytest
 from sqlalchemy import create_engine
@@ -164,6 +164,37 @@ def test_debug_stream_setting_overrides_model_without_changing_it():
     assert AgentDebugService.overrides(payload)["stream_enabled"] == 0
 
 
+def test_debug_preview_passes_conversation_history_and_replaces_saved_prompt():
+    rag_result = RagResult(
+        question="再回答上面的问题",
+        knowledge_scopes=[],
+        chunks=[],
+        prompt="调试提示词渲染结果",
+        allow_general_knowledge=True,
+        messages=[{"role": "system", "content": "调试主提示词"}],
+    )
+    payload = AgentDebugRequest(
+        promptContent="调试主提示词",
+        modelId=1,
+        question="再回答上面的问题",
+        history=[
+            {"role": "user", "content": "最开始的问题"},
+            {"role": "assistant", "content": "第一次回答"},
+        ],
+    )
+
+    with _database() as db:
+        build_result = AsyncMock(return_value=rag_result)
+        with patch("app.services.agent_debug_service.KnowledgeAgentService.build_result", build_result):
+            result = asyncio.run(AgentDebugService.build_result(db, payload))
+
+    kwargs = build_result.await_args.kwargs
+    assert [item.content for item in kwargs["history"]] == ["最开始的问题", "第一次回答"]
+    assert kwargs["prompt_override"] == "调试主提示词"
+    assert result.messages == rag_result.messages
+    assert result.prompt == "调试提示词渲染结果"
+
+
 @pytest.mark.parametrize("reasoning_visible", [True, False])
 def test_agent_debug_stream_respects_reasoning_visibility(reasoning_visible):
     async def chunks():

ai_knowledge_base_v2/apps/backend/tests/test_knowledge_agent.py

@@ -269,10 +269,30 @@ def test_contextual_follow_up_is_rewritten_before_agent_decision():
         rewrite = next(item for item in result.tool_trace if item["tool"] == "rewrite_contextual_question")
         decision = next(item for item in result.tool_trace if item["tool"] == "agent_decision")
 
-        assert rewrite["rewrittenQuestion"] == "关于“课程退款条件有哪些？”的追问：那第二种情况呢？"
+        assert rewrite["rewrittenQuestion"] == (
+            "结合以下历史问题回答当前追问：\n"
+            "历史问题1：课程退款条件有哪些？\n"
+            "当前追问：那第二种情况呢？"
+        )
         assert decision["request"]["question"] == rewrite["rewrittenQuestion"]
 
 
+def test_long_follow_up_reference_is_detected_and_keeps_multiple_user_questions():
+    history = [
+        ChatMessage(id=1, session_id=1, user_id=1, role="user", content="觉式练习和慧氏练习分别是什么，区别是什么？"),
+        ChatMessage(id=2, session_id=1, user_id=1, role="assistant", content="暂时无法确认。"),
+        ChatMessage(id=3, session_id=1, user_id=1, role="user", content="阴影人格练习步骤是什么？"),
+        ChatMessage(id=4, session_id=1, user_id=1, role="assistant", content="先稳定状态，再进行觉察。"),
+    ]
+    question = "你能根据这个练习的基本方向再来回答一下我上面的问题吗？"
+
+    assert KnowledgeAgentService._needs_contextual_rewrite(question) is True
+    rewritten = KnowledgeAgentService._fallback_rewrite(question, history)
+    assert "觉式练习和慧氏练习" in rewritten
+    assert "阴影人格练习步骤" in rewritten
+    assert question in rewritten
+
+
 def test_homework_overview_expands_practice_terms_and_section_limit():
     terms = KnowledgeAgentService._query_terms("合一的作业是什么？")