fix: report LLM failures and preserve complete history

.github/workflows/test.yaml

@@ -1,7 +1,13 @@
 name: Python Tests
 on:
   pull_request:
-    paths: ['app/**', 'tests/**', 'pyproject.toml']
+    paths:
+      - 'app/**'
+      - 'tests/**'
+      - 'Dockerfile'
+      - 'pyproject.toml'
+      - 'uv.lock'
+      - '.github/workflows/test.yaml'
 jobs:
   test:
     runs-on: ubuntu-latest
@@ -21,3 +27,17 @@ jobs:
         run: uv run mypy app --ignore-missing-imports
       - name: Lint
         run: uv run ruff check app
+
+  container:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: docker/setup-buildx-action@v3
+      - name: Build runtime image
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          load: true
+          tags: app-poly-gitops-fastapi:test
+      - name: Verify offline runtime startup
+        run: tests/test_container_runtime.sh app-poly-gitops-fastapi:test

Dockerfile

@@ -17,6 +17,7 @@ COPY ./app ./app
 # Expose port
 EXPOSE 8000
 
-# Run with uv
-CMD ["uv", "run", "uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8000"]
+# Run the environment assembled at build time. Runtime startup must not resolve,
+# download, or install dependencies.
+CMD [".venv/bin/uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8000"]
 # Trigger rebuild

README.md

@@ -6,7 +6,7 @@ FastAPI микросервис для LLM-чатов с управлением 
 
 - 🗨️ **Chat API** — standalone и project-scoped чаты
 - 📁 **Проекты** — группировка чатов по проектам
-- 📜 **История** — сохранение истории с trace_id/span_id
+- 📜 **История** — сохранение успешного transcript (входные сообщения и ответ модели) с trace_id/span_id
 - 🔗 **Notion** — загрузка behavior-схем агентов
 - 📊 **Observability** — OpenTelemetry traces/logs/metrics
 - 🐳 **Docker** — контейнеризация с uv
@@ -23,6 +23,9 @@ FastAPI микросервис для LLM-чатов с управлением 
 | `GET` | `/api/v1/behavior/schema` | Текущая схема поведения |
 | `GET` | `/health` | Health check |
 
+Ошибки LLM-провайдера возвращаются как HTTP `502`; они не маскируются под
+успешный `200` и не создают запись истории проекта.
+
 ## Запуск
 
 ### Локально (с uv)

app/api.py

@@ -10,6 +10,7 @@
 from app.logger import enrich_context
 
 from .core.db import get_session
+from .core.llm_client import LLMProviderError
 from .core.project_memory import ProjectMemory
 from .integrations.behavior_manager import BehaviorManager
 from .models import ChatHistory, ChatHistoryDB, Project, StoredChatMessage
@@ -18,6 +19,7 @@
     ChatResponse,
     CreateProjectRequest,
     ProjectInfo,
+    Role,
 )
 from .services.chat_service import ChatService
 
@@ -64,6 +66,12 @@ async def chat_endpoint(
         )
         return ChatResponse(reply=reply)
 
+    except LLMProviderError as e:
+        log.bind(event="chat_provider_error", error=str(e)).error(
+            "LLM provider request failed"
+        )
+        raise HTTPException(status_code=502, detail="AI provider unavailable")
+
     except Exception as e:
         log.bind(event="chat_api_error", error=str(e)).error("Standard chat request failed")
         traceback.print_exc()
@@ -123,10 +131,15 @@ async def chat_in_project(
         if not project:
             raise ValueError(f"Проект {project_id} не найден")
 
+        reply = await chat_service.get_ai_reply(
+            req.messages, req.user_api_key, project_id=project_id, trace_id=trace_id
+        )
+
         chat_messages = [
             StoredChatMessage(role=m.role, content=m.content)
             for m in req.messages
         ]
+        chat_messages.append(StoredChatMessage(role=Role.assistant, content=reply))
         history_db = ChatHistoryDB(
             project_id=project_id,
             messages=json.dumps([m.model_dump() for m in chat_messages], default=json_serial),
@@ -135,16 +148,18 @@ async def chat_in_project(
         )
         session.add(history_db)
         session.commit()
-
-        reply = await chat_service.get_ai_reply(
-            req.messages, req.user_api_key, project_id=project_id, trace_id=trace_id
-        )
         return ChatResponse(reply=reply)
 
     except ValueError as e:
         log.bind(event="project_not_found").warning("Project not found")
         raise HTTPException(status_code=404, detail=str(e))
 
+    except LLMProviderError as e:
+        log.bind(event="project_chat_provider_error", error=str(e)).error(
+            "LLM provider request failed"
+        )
+        raise HTTPException(status_code=502, detail="AI provider unavailable")
+
     except Exception as e:
         log.bind(event="project_chat_error", error=str(e)).error("Project chat failed")
         traceback.print_exc()

app/core/llm_client.py

@@ -33,6 +33,10 @@
     unit="1"
 )
 
+
+class LLMProviderError(RuntimeError):
+    """The configured LLM provider could not produce a valid reply."""
+
 class OpenRouterClient:
     def __init__(self):
         # Из секрета
@@ -240,10 +244,20 @@ async def generate_reply(
 
             # Проверяем на ошибки
             if "error" in response:
-                return f"Ошибка API: {response['error']}"
+                raise LLMProviderError(str(response["error"]))
+
+            # A transport-level success does not guarantee a valid provider
+            # payload.  Keep malformed upstream responses on the same 502 path
+            # as explicit provider errors instead of leaking an internal 500.
+            try:
+                content = response["choices"][0]["message"]["content"]
+            except (KeyError, IndexError, TypeError) as exc:
+                raise LLMProviderError("Malformed response from LLM provider") from exc
+
+            if not isinstance(content, str) or not content:
+                raise LLMProviderError("Malformed response from LLM provider")
 
-            # Извлекаем текст ответа
-            return response["choices"][0]["message"]["content"]
+            return content
 
         finally:
             # Восстанавливаем оригинальный ключ

tests/test_api.py

@@ -1,5 +1,7 @@
 """API endpoint tests."""
 
+from unittest.mock import AsyncMock
+
 from fastapi.testclient import TestClient
 
 from app.main import create_app
@@ -40,6 +42,40 @@ def test_chat_endpoint(client, mock_openrouter):
     assert "reply" in response.json()
 
 
+def test_chat_endpoint_reports_provider_failure_as_bad_gateway(client, monkeypatch):
+    """A provider failure must not look like a successful chat reply."""
+    monkeypatch.setattr(
+        "app.core.llm_client.OpenRouterClient.chat_completion",
+        AsyncMock(return_value={"error": "provider unavailable"}),
+    )
+
+    response = client.post(
+        "/api/v1/chat",
+        json={"messages": [{"role": "user", "content": "test"}]},
+    )
+
+    assert response.status_code == 502
+    assert response.json() == {"detail": "AI provider unavailable"}
+
+
+def test_chat_endpoint_reports_malformed_provider_response_as_bad_gateway(
+    client, monkeypatch
+):
+    """A malformed successful response is still a provider failure."""
+    monkeypatch.setattr(
+        "app.core.llm_client.OpenRouterClient.chat_completion",
+        AsyncMock(return_value={"choices": []}),
+    )
+
+    response = client.post(
+        "/api/v1/chat",
+        json={"messages": [{"role": "user", "content": "test"}]},
+    )
+
+    assert response.status_code == 502
+    assert response.json() == {"detail": "AI provider unavailable"}
+
+
 def test_chat_endpoint_empty_messages(client, mock_openrouter):
     """Test POST /api/v1/chat with empty messages list."""
     response = client.post("/api/v1/chat", json={
@@ -106,7 +142,34 @@ def test_project_history(client, mock_openrouter):
     # Get history
     response = client.get(f"/api/v1/projects/{project_id}/history")
     assert response.status_code == 200
-    assert isinstance(response.json(), list)
+    assert [
+        (message["role"], message["content"])
+        for message in response.json()[0]["messages"]
+    ] == [
+        ("user", "test message"),
+        ("assistant", "Mocked response from AI"),
+    ]
+
+
+def test_failed_project_chat_does_not_create_history(client, monkeypatch):
+    """Provider failures must not leave a request that looks completed."""
+    project = client.post(
+        "/api/v1/projects", json={"name": "Failed chat"}
+    ).json()
+    monkeypatch.setattr(
+        "app.core.llm_client.OpenRouterClient.chat_completion",
+        AsyncMock(return_value={"error": "provider unavailable"}),
+    )
+
+    response = client.post(
+        f"/api/v1/projects/{project['id']}/chat",
+        json={"messages": [{"role": "user", "content": "test"}]},
+    )
+
+    assert response.status_code == 502
+    history = client.get(f"/api/v1/projects/{project['id']}/history")
+    assert history.status_code == 200
+    assert history.json() == []
 
 
 def test_behavior_schema_not_loaded(client):

tests/test_container_runtime.sh

@@ -0,0 +1,28 @@
+#!/bin/sh
+set -eu
+
+image="${1:-app-poly-gitops-fastapi:test}"
+container_id="$(docker run -d --network none "$image")"
+
+cleanup() {
+  docker logs "$container_id" 2>/dev/null || true
+  docker rm -f "$container_id" >/dev/null 2>&1 || true
+}
+trap cleanup EXIT
+
+for _ in $(seq 1 30); do
+  if [ "$(docker inspect -f '{{.State.Running}}' "$container_id")" != true ]; then
+    echo "container exited before becoming healthy" >&2
+    exit 1
+  fi
+
+  if docker exec "$container_id" .venv/bin/python -c \
+    'import json, urllib.request; assert json.load(urllib.request.urlopen("http://127.0.0.1:8000/health", timeout=1)) == {"status": "ok"}' \
+    >/dev/null 2>&1; then
+    exit 0
+  fi
+  sleep 1
+done
+
+echo "container did not become healthy without runtime network access" >&2
+exit 1