feat(agent): stream admin debug preview

ai_knowledge_base_v2/apps/backend/tests/test_agent_runtime_config.py

@@ -1,17 +1,22 @@
 import asyncio
 from decimal import Decimal
+import json
 from unittest.mock import patch
 
 from sqlalchemy import create_engine
 from sqlalchemy.orm import Session
 from sqlalchemy.pool import StaticPool
 
-from app.api.admin_settings import get_agent_runtime_config, save_agent_runtime_config
+from app.api.admin_settings import _debug_agent_stream, get_agent_runtime_config, save_agent_runtime_config
 from app.models import Base
 from app.models.admin import Admin
 from app.models.ai_config import ModelConfig
-from app.schemas.admin import AgentRuntimeConfigSaveRequest
-from app.services.model_stream_service import _openai_stream_payload, _stream_configured_model_async
+from app.schemas.admin import AgentDebugRequest, AgentRuntimeConfigSaveRequest
+from app.services.model_stream_service import (
+    AsyncStreamingModelResponse,
+    _openai_stream_payload,
+    _stream_configured_model_async,
+)
 from app.services.model_service import ModelClientService, _max_output_tokens
 from app.services.rag_service import RagResult
 
@@ -128,6 +133,66 @@ def test_disabled_stream_returns_one_complete_chunk():
     assert chunks == [answer]
 
 
+def test_agent_debug_stream_emits_status_content_and_trace():
+    async def chunks():
+        yield "<think>内部思考</think>"
+        yield "- **第一步**：停一下"
+
+    rag_result = RagResult(
+        question="怎么冷静",
+        knowledge_scopes=[],
+        chunks=[],
+        prompt="怎么冷静",
+        allow_general_knowledge=True,
+        retrieval_log_id=9,
+        tool_trace=[{"tool": "KnowledgeSearch", "status": "success"}],
+        messages=[{"role": "user", "content": "怎么冷静"}],
+    )
+    model_response = AsyncStreamingModelResponse(
+        model_id=1,
+        model_name="production-model",
+        input_token=10,
+        chunks=chunks(),
+    )
+
+    async def collect(payload, db, admin):
+        return [event async for event in _debug_agent_stream(payload, db, admin)]
+
+    with _database() as db:
+        admin = _admin()
+        model = _model()
+        db.add_all([admin, model])
+        db.commit()
+        payload = AgentDebugRequest(
+            promptContent="你是测试助手",
+            modelId=model.id,
+            knowledgeIds=[],
+            question="怎么冷静",
+            maxToken=8192,
+        )
+        with (
+            patch(
+                "app.services.agent_debug_service.KnowledgeAgentService.build_result",
+                return_value=rag_result,
+            ),
+            patch(
+                "app.services.agent_debug_service.ModelStreamService.debug_stream_async",
+                return_value=model_response,
+            ),
+        ):
+            events = asyncio.run(collect(payload, db, admin))
+
+    decoded = [
+        json.loads(event.removeprefix("data: "))
+        for event in events
+        if event != "data: [DONE]\n\n"
+    ]
+    assert [event["type"] for event in decoded] == ["status", "content", "content", "complete"]
+    assert decoded[1]["content"].startswith("<think>")
+    assert decoded[2]["content"] == "- **第一步**：停一下"
+    assert decoded[3]["retrievalTrace"][0]["tool"] == "KnowledgeSearch"
+
+
 def test_agent_debug_does_not_truncate_long_answer(monkeypatch):
     monkeypatch.setattr(
         "app.services.model_service._call_configured_model",