Merge pull request 'revert: sync production to verified test version' (#34) from codex/sync-production-to-test-20260916 into main

revert: sync production to verified test version

digital-avatar-app/Dockerfile

@@ -1,12 +1,16 @@
 # 构建阶段：安装依赖并打包 H5
 FROM node:18-alpine AS build
 
+ARG APP_GIT_SHA=unknown
+ARG APP_BUILD_TIME=unknown
+
 WORKDIR /app
 
 COPY package*.json ./
 RUN npm ci
 
 COPY . .
+RUN printf '{"gitSha":"%s","buildTime":"%s"}\n' "$APP_GIT_SHA" "$APP_BUILD_TIME" > public/version.json
 RUN npm run build
 
 # 运行阶段：nginx 托管静态资源并反向代理 /api 到后端
@@ -14,6 +18,11 @@ RUN npm run build
 # 新版 nginx(>=1.31) 用 pwrite 写 pid 文件会被拦导致致命退出；1.28 用 write() 可正常启动。
 FROM nginx:1.28-alpine
 
+ARG APP_GIT_SHA=unknown
+ARG APP_BUILD_TIME=unknown
+LABEL org.opencontainers.image.revision=${APP_GIT_SHA} \
+      org.opencontainers.image.created=${APP_BUILD_TIME}
+
 COPY --from=build /app/dist /usr/share/nginx/html
 # 覆盖 nginx 默认主配置（含唯一可写的 pid /tmp/nginx.pid，规避受限容器内 /run 不可写导致反复重启）
 COPY nginx.conf /etc/nginx/nginx.conf

digital-avatar-app/backend/Dockerfile

@@ -7,6 +7,13 @@ WORKDIR /app
 COPY requirements.txt .
 RUN pip install --no-cache-dir --timeout 120 --retries 10 -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt
 
+ARG APP_GIT_SHA=unknown
+ARG APP_BUILD_TIME=unknown
+ENV APP_GIT_SHA=${APP_GIT_SHA} \
+    APP_BUILD_TIME=${APP_BUILD_TIME}
+LABEL org.opencontainers.image.revision=${APP_GIT_SHA} \
+      org.opencontainers.image.created=${APP_BUILD_TIME}
+
 COPY . .
 
 # 后端使用 SQLite（avatar.db 落在 /app 内），平铺结构以 `uvicorn main:app` 启动

digital-avatar-app/backend/main.py

@@ -1,13 +1,14 @@
 from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
 
-import os
+import importlib.util
 import logging
+import os
 
 from apscheduler.schedulers.asyncio import AsyncIOScheduler
 from apscheduler.triggers.interval import IntervalTrigger
 
-from database import init_db, SessionLocal
+from database import engine, init_db, SessionLocal
 from models import Avatar, Authorization, Organization, TokenAccount, TokenPlan, User
 from fastapi.staticfiles import StaticFiles
 import routers.avatars
@@ -54,7 +55,31 @@ app.mount("/api/files", StaticFiles(directory=UPLOAD_DIR), name="knowledge-files
 
 @app.get("/api/health")
 def health():
-    return ok({"status": "ok"})
+    checks = _runtime_checks()
+    return ok({
+        "status": "ok" if all(checks.values()) else "degraded",
+        "gitSha": os.getenv("APP_GIT_SHA", "unknown"),
+        "buildTime": os.getenv("APP_BUILD_TIME", "unknown"),
+        "checks": checks,
+    })
+
+
+def _runtime_checks():
+    return {
+        "database": _database_is_ready(),
+        "uploads": os.path.isdir(UPLOAD_DIR) and os.access(UPLOAD_DIR, os.W_OK),
+        "pdfOcr": importlib.util.find_spec("pymupdf") is not None,
+    }
+
+
+def _database_is_ready():
+    try:
+        with engine.connect() as connection:
+            connection.exec_driver_sql("SELECT 1")
+        return True
+    except Exception:
+        logger.exception("Database readiness check failed")
+        return False
 
 
 def seed():

digital-avatar-app/backend/routers/chat.py

@@ -79,34 +79,6 @@ _WRITING_SYSTEM_PATTERNS = {
 _JAPANESE_KANA = re.compile(r"[\u3040-\u30ff]")
 _KOREAN_HANGUL = re.compile(r"[\uac00-\ud7af\u1100-\u11ff]")
 
-_LATIN_LANGUAGE_MARKERS = {
-    "English": re.compile(
-        r"\b(?:i|you|we|they|he|she|have|has|had|friend|who|what|where|when|why|how|"
-        r"symptoms?|disease|please|can|could|would|should|is|are|was|were|the|this|that)\b",
-        re.IGNORECASE,
-    ),
-    "French": re.compile(
-        r"\b(?:je|tu|vous|nous|ils|elle|une|des|avec|pour|pourquoi|comment|bonjour|est|sont)\b",
-        re.IGNORECASE,
-    ),
-    "Spanish": re.compile(
-        r"\b(?:yo|tu|usted|nosotros|ellos|ella|una|con|para|por que|como|hola|esta|son)\b",
-        re.IGNORECASE,
-    ),
-    "German": re.compile(
-        r"\b(?:ich|du|sie|wir|eine|mit|fur|warum|wie|hallo|ist|sind|haben)\b",
-        re.IGNORECASE,
-    ),
-    "Portuguese": re.compile(
-        r"\b(?:eu|voce|nos|eles|ela|uma|com|para|porque|como|ola|esta|sao|tenho)\b",
-        re.IGNORECASE,
-    ),
-    "Italian": re.compile(
-        r"\b(?:io|tu|voi|noi|loro|una|con|per|perche|come|ciao|sono|avere)\b",
-        re.IGNORECASE,
-    ),
-}
-
 
 class ChatMessage(BaseModel):
     model_config = ConfigDict(populate_by_name=True)
@@ -515,67 +487,15 @@ def _qa_requires_per_turn_rendering(
     return bool(history) or _qa_requires_language_adaptation(question, answer)
 
 
-def _latin_language_name(value: str) -> str:
-    scores = {
-        language: len(pattern.findall(value or ""))
-        for language, pattern in _LATIN_LANGUAGE_MARKERS.items()
-    }
-    language, score = max(scores.items(), key=lambda item: item[1])
-    return language if score else "the same natural language as the latest user message"
-
-
-def _turn_language_name(value: str) -> str:
-    writing_system = _dominant_writing_system(value)
-    return {
-        "han": "Chinese",
-        "japanese": "Japanese",
-        "korean": "Korean",
-        "cyrillic": "the same Cyrillic-script language as the latest user message",
-        "arabic": "the same Arabic-script language as the latest user message",
-        "hebrew": "Hebrew",
-        "devanagari": "the same Devanagari-script language as the latest user message",
-        "thai": "Thai",
-        "greek": "Greek",
-        "latin": _latin_language_name(value),
-    }.get(writing_system, "the same natural language as the latest user message")
-
-
-def _per_turn_language_instruction(question: str = "") -> str:
-    language = _turn_language_name(question)
+def _per_turn_language_instruction() -> str:
     return (
-        f"MANDATORY OUTPUT LANGUAGE FOR THIS TURN: {language}. "
-        "Write the entire answer only in that language. This instruction overrides the languages used by "
-        "conversation history, profile data, standard answers, retrieved documents, and custom prompts. "
-        "Translate grounded source material faithfully when necessary. Do not mention language detection, "
-        "translation, or this instruction."
+        "本轮语言覆盖指令：只根据紧随其后的最新用户消息判断本轮回答语言。"
+        "即使此前整段对话一直使用另一种语言，只要最新消息切换了语言，本轮就必须立即切换到相同语言；"
+        "不要沿用上一轮语言。若最新消息明确指定回答语言，以该指定为准；若混用多种语言，使用其中占主导的"
+        "自然语言。不要说明你检测、切换或翻译了语言。"
     )
 
 
-def _answer_requires_language_repair(question: str, answer: str) -> bool:
-    question_system = _dominant_writing_system(question)
-    answer_system = _dominant_writing_system(answer)
-    return (
-        question_system != "unknown"
-        and answer_system != "unknown"
-        and question_system != answer_system
-    )
-
-
-def _language_repair_messages(question: str, answer: str) -> list[dict]:
-    return [
-        {"role": "system", "content": _per_turn_language_instruction(question)},
-        {
-            "role": "system",
-            "content": (
-                "Rewrite the supplied draft in the mandatory output language. Preserve every grounded fact, "
-                "number, proper noun, uncertainty, and safety qualification. Add no new information and output "
-                "only the rewritten answer."
-            ),
-        },
-        {"role": "user", "content": answer.strip()},
-    ]
-
-
 def _canonicalize_question(value: str) -> str:
     value = _normalize_question(value)
     replacements = (
@@ -808,7 +728,7 @@ def _build_prompt(
     for item in history[-MAX_HISTORY_MESSAGES:]:
         messages.append({"role": item.role, "content": item.content} if hasattr(item, "role") else item)
     # Keep the language instruction adjacent to the current turn so long histories cannot override it.
-    messages.append({"role": "system", "content": _per_turn_language_instruction(question)})
+    messages.append({"role": "system", "content": _per_turn_language_instruction()})
     messages.append({"role": "user", "content": question.strip()})
     return messages
 
@@ -871,41 +791,6 @@ def _call_qwen(
     return {"answer": answer.strip(), "usage": data.get("usage") or {}}
 
 
-def _call_billed_qwen(
-    db: Session,
-    avatar: Avatar,
-    messages: list[dict],
-    temperature: float,
-    usage_source: str,
-    model_config: ChatModelConfig,
-) -> tuple[str, dict]:
-    reservation = reserve_avatar_tokens(
-        db,
-        avatar,
-        usage_source,
-        model_config.model,
-        messages,
-        model_config.max_tokens,
-    )
-    try:
-        model_result = _call_qwen(
-            messages=messages,
-            temperature=temperature,
-            model_config=model_config,
-        )
-        answer = model_result["answer"]
-        token_usage = settle_reservation(
-            db,
-            reservation,
-            model_result.get("usage"),
-            fallback_total=estimate_fallback_usage(messages, answer),
-        )
-        return answer, token_usage
-    except Exception as exc:
-        release_reservation(db, reservation, str(exc))
-        raise
-
-
 def _iter_qwen_stream(
     messages: list[dict], temperature: float, model_config: ChatModelConfig | None = None
 ):
@@ -1014,36 +899,32 @@ def _resolve_reply(
     token_usage = None
     if model_client is not None:
         answer = model_client(messages=messages, temperature=temperature)
-        if _answer_requires_language_repair(question, str(answer or "")):
-            answer = model_client(
-                messages=_language_repair_messages(question, str(answer)),
-                temperature=0.0,
-            )
     else:
         model_config = get_chat_model_config()
-        answer, token_usage = _call_billed_qwen(
+        reservation = reserve_avatar_tokens(
             db,
             avatar,
-            messages,
-            temperature,
             usage_source,
-            model_config,
+            model_config.model,
+            messages,
+            model_config.max_tokens,
         )
-        if _answer_requires_language_repair(question, answer):
-            logger.warning(
-                "chat response language mismatch avatar=%s source=%s expected=%s",
-                avatar.id,
-                usage_source,
-                _turn_language_name(question),
+        try:
+            model_result = _call_qwen(
+                messages=messages,
+                temperature=temperature,
+                model_config=model_config,
             )
-            answer, token_usage = _call_billed_qwen(
+            answer = model_result["answer"]
+            token_usage = settle_reservation(
                 db,
-                avatar,
-                _language_repair_messages(question, answer),
-                0.0,
-                f"{usage_source}_language_repair",
-                model_config,
+                reservation,
+                model_result.get("usage"),
+                fallback_total=estimate_fallback_usage(messages, answer),
             )
+        except Exception as exc:
+            release_reservation(db, reservation, str(exc))
+            raise
     answer = str(answer or "").strip()
     if image_contexts and _answer_denies_available_image(answer):
         logger.warning(

digital-avatar-app/backend/tests/test_chat_orchestration.py

@@ -6,7 +6,6 @@ from fastapi import HTTPException
 
 from models import Avatar, User
 from routers.chat import (
-    _answer_requires_language_repair,
     _build_prompt,
     _iter_text_chunks,
     _match_standard_qa,
@@ -15,7 +14,6 @@ from routers.chat import (
     _qa_requires_per_turn_rendering,
     _require_owned_avatar,
     _resolve_reply,
-    _turn_language_name,
 )
 
 
@@ -109,8 +107,8 @@ class ChatOrchestrationTests(unittest.TestCase):
         messages = fake_model.call_args.kwargs["messages"]
         self.assertEqual(messages[-1], {"role": "user", "content": "Quelle est votre adresse ?"})
         self.assertEqual(messages[-2]["role"], "system")
-        self.assertIn("MANDATORY OUTPUT LANGUAGE", messages[-2]["content"])
-        self.assertIn("French", messages[-2]["content"])
+        self.assertIn("本轮语言覆盖指令", messages[-2]["content"])
+        self.assertIn("不要沿用上一轮语言", messages[-2]["content"])
 
     def test_latest_user_message_has_an_adjacent_language_override(self):
         history = [
@@ -121,37 +119,8 @@ class ChatOrchestrationTests(unittest.TestCase):
 
         self.assertEqual(messages[-1], {"role": "user", "content": "What can you help me with?"})
         self.assertEqual(messages[-2]["role"], "system")
-        self.assertIn("MANDATORY OUTPUT LANGUAGE", messages[-2]["content"])
-        self.assertIn("English", messages[-2]["content"])
-
-    def test_reported_alzheimer_question_is_explicitly_english(self):
-        question = "I have a friend who has symptoms of Alzheimer's disease"
-
-        self.assertEqual(_turn_language_name(question), "English")
-        messages = _build_prompt(self.avatar, [], question, [])
-        self.assertIn("MANDATORY OUTPUT LANGUAGE FOR THIS TURN: English", messages[-2]["content"])
-
-    def test_non_stream_reply_repairs_a_wrong_writing_system_before_sending(self):
-        question = "I have a friend who has symptoms of Alzheimer's disease"
-        fake_model = Mock(side_effect=["建议尽快就医评估。", "Please arrange a medical assessment soon."])
-
-        result = _resolve_reply(
-            None,
-            self.avatar,
-            question,
-            [],
-            qa_pairs=[],
-            search_fn=lambda *_args, **_kwargs: [],
-            model_client=fake_model,
-            usage_source="takeover",
-        )
-
-        self.assertEqual(result["answer"], "Please arrange a medical assessment soon.")
-        self.assertEqual(fake_model.call_count, 2)
-        repair_messages = fake_model.call_args.kwargs["messages"]
-        self.assertIn("English", repair_messages[0]["content"])
-        self.assertIn("建议尽快就医评估", repair_messages[-1]["content"])
-        self.assertTrue(_answer_requires_language_repair(question, "建议尽快就医评估。"))
+        self.assertIn("最新用户消息", messages[-2]["content"])
+        self.assertIn("立即切换到相同语言", messages[-2]["content"])
 
     def test_conversational_paraphrase_matches_standard_qa(self):
         for question in ("请问一下，你们公司在哪里呀？", "请问去你们那边怎么走"):

digital-avatar-app/backend/tests/test_health.py

@@ -0,0 +1,33 @@
+import main
+
+
+def test_health_reports_release_and_runtime_capabilities(monkeypatch):
+    monkeypatch.setenv("APP_GIT_SHA", "test-sha")
+    monkeypatch.setenv("APP_BUILD_TIME", "2026-09-09T00:00:00Z")
+    monkeypatch.setattr(main, "_runtime_checks", lambda: {
+        "database": True,
+        "uploads": True,
+        "pdfOcr": True,
+    })
+
+    response = main.health()
+
+    assert response["code"] == 200
+    assert response["data"]["status"] == "ok"
+    assert response["data"]["gitSha"] == "test-sha"
+    assert response["data"]["buildTime"] == "2026-09-09T00:00:00Z"
+    assert response["data"]["checks"] == {
+        "database": True,
+        "uploads": True,
+        "pdfOcr": True,
+    }
+
+
+def test_health_is_degraded_when_a_required_capability_is_missing(monkeypatch):
+    monkeypatch.setattr(main, "_runtime_checks", lambda: {
+        "database": True,
+        "uploads": True,
+        "pdfOcr": False,
+    })
+
+    assert main.health()["data"]["status"] == "degraded"

digital-avatar-app/docker-compose.yml

@@ -1,42 +1,56 @@
 # 会会数字分身 —— Docker 测试实例（独立端口，不干扰现有 :8088 huihui 部署）
 services:
   avatar-backend:
-    build: ./backend
-    image: avatar-test-backend:latest
+    build:
+      context: ./backend
+      args:
+        APP_GIT_SHA: ${APP_GIT_SHA:?APP_GIT_SHA must be the full release commit}
+        APP_BUILD_TIME: ${APP_BUILD_TIME:?APP_BUILD_TIME must be set}
+    image: avatar-test-backend:${APP_GIT_SHA}
     container_name: avatar-test-backend
     restart: unless-stopped
     env_file:
       - .env
     environment:
-      DATABASE_URL: sqlite:////data/avatar.db
+      DATABASE_URL: sqlite:////data/db/avatar.db
       UPLOAD_DIR: /data/uploads
       CHAT_MODEL_CONFIG_URL: http://host.docker.internal:8000/api/ai-models/runtime/digital-avatar
     extra_hosts:
       - "host.docker.internal:host-gateway"
     volumes:
-      - avatar-data:/data
+      # Mount the directory, not only avatar.db: SQLite WAL/SHM files must survive recreation.
+      - ${AVATAR_DB_DIR:?AVATAR_DB_DIR must contain the persistent avatar.db}:/data/db
+      - ${AVATAR_UPLOAD_DIR:?AVATAR_UPLOAD_DIR must point to persistent uploads}:/data/uploads
     expose:
       - "8000"
     ports:
       - "8011:8000"   # 仅用于直接调试 API；前端经内部网络访问，不走 host 端口
+    healthcheck:
+      test: ["CMD", "python", "-c", "import json,urllib.request; d=json.load(urllib.request.urlopen('http://127.0.0.1:8000/api/health', timeout=5))['data']; assert d['status']=='ok' and all(d['checks'].values())"]
+      interval: 10s
+      timeout: 8s
+      retries: 12
+      start_period: 20s
     networks:
       - avatar-net
 
   avatar-frontend:
-    build: .
-    image: avatar-test-frontend:latest
+    build:
+      context: .
+      args:
+        APP_GIT_SHA: ${APP_GIT_SHA:?APP_GIT_SHA must be the full release commit}
+        APP_BUILD_TIME: ${APP_BUILD_TIME:?APP_BUILD_TIME must be set}
+    image: avatar-test-frontend:${APP_GIT_SHA}
     container_name: avatar-test-frontend
     restart: unless-stopped
     ports:
       - "8099:80"     # 浏览器访问 http://<host>:8099
     depends_on:
-      - avatar-backend
+      avatar-backend:
+        condition: service_healthy
     networks:
       - avatar-net
 
 networks:
   avatar-net:
     driver: bridge
-
-volumes:
-  avatar-data:

digital-avatar-app/docs/H5_PRODUCTION_DEPLOYMENT.md

@@ -61,7 +61,9 @@ WECHAT_VIRTUAL_PRODUCT_2=<100元套餐商品ID>
 WECHAT_VIRTUAL_PRODUCT_3=<1000元套餐商品ID>
 WECHAT_VIRTUAL_PRODUCT_4=<10000元套餐商品ID>
 
-DATABASE_URL=sqlite:////data/avatar.db
+AVATAR_DB_DIR=/srv/digital-avatar/data/db
+AVATAR_UPLOAD_DIR=/srv/digital-avatar/data/uploads
+DATABASE_URL=sqlite:////data/db/avatar.db
 UPLOAD_DIR=/data/uploads
 CHAT_MODEL_CONFIG_URL=http://<huihuisquare-api>/api/ai-models/runtime/digital-avatar
 EMBEDDING_API_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
@@ -74,13 +76,17 @@ VISION_OCR_MODEL=qwen-vl-ocr
 VISION_MAX_OUTPUT_TOKENS=2048
 VISION_TIMEOUT_SECONDS=90
 VISION_TOKEN_RESERVE=12000
+APP_GIT_SHA=<本次发布的完整提交SHA>
+APP_BUILD_TIME=<UTC ISO-8601构建时间>
 CHAT_IMAGE_MAX_BYTES=8388608
 CHAT_IMAGE_MAX_PIXELS=16000000
 CHAT_ATTACHMENT_RETENTION_HOURS=24
 CHAT_ATTACHMENT_CLEANUP_MINUTES=60
 ```
 
-如生产 AI 配置中心不可用，还应提供当前项目支持的 `OPENAI_API_KEY`、`OPENAI_BASE_URL`、`CHAT_MODEL` 等兜底配置。`/data` 必须挂载持久卷，数据库与知识库文件不可存放在容器临时层。
+如生产 AI 配置中心不可用，还应提供当前项目支持的 `OPENAI_API_KEY`、`OPENAI_BASE_URL`、`CHAT_MODEL` 等兜底配置。数据库文件与上传目录必须从宿主机显式挂载，不能存放在容器临时层。
+
+`AVATAR_DB_DIR` 和 `AVATAR_UPLOAD_DIR` 必须是已备份的宿主机绝对路径，编排缺少任一变量都会直接拒绝构建或启动，防止误挂空卷造成用户、分身或知识库“丢失”的假象。SQLite 必须挂载整个数据库目录，不能只挂载 `avatar.db` 单文件，否则 `avatar.db-wal` 和 `avatar.db-shm` 会留在容器临时层，换容器后可能出现数据状态回退。
 
 `EMBEDDING_API_URL` 同时支持 OpenAI 兼容基础地址（如上面的 `/v1`）和完整的 `/v1/embeddings` 地址，后端会统一请求 `/embeddings`。发布后必须在后端容器内执行一次最小向量探针，确认返回向量数量和维度，而不能只检查 `/api/health`。
 
@@ -97,7 +103,7 @@ App 与 H5 积分充值使用会会支付体系的 `payment-v3/payment/pay`，
 ```bash
 BACKUP_DIR="backups/$(date +%Y%m%d-%H%M%S)"
 mkdir -p "$BACKUP_DIR"
-cp /srv/digital-avatar/data/avatar.db "$BACKUP_DIR/"
+cp /srv/digital-avatar/data/db/avatar.db "$BACKUP_DIR/"
 tar -C /srv/digital-avatar/data -czf "$BACKUP_DIR/uploads.tgz" uploads
 ```
 
@@ -107,13 +113,22 @@ tar -C /srv/digital-avatar/data -czf "$BACKUP_DIR/uploads.tgz" uploads
 git fetch origin
 git checkout <已验收的提交SHA>
 cd digital-avatar-app
-docker compose build --pull avatar-backend avatar-frontend
-docker compose up -d avatar-backend avatar-frontend
+export APP_GIT_SHA="$(git rev-parse HEAD)"
+export APP_BUILD_TIME="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
+docker compose build --pull --no-cache avatar-backend avatar-frontend
+docker compose up -d --force-recreate --wait avatar-backend avatar-frontend
 docker compose ps
-curl -fsS http://127.0.0.1:8099/api/health
+python3 scripts/verify-deployment.py \
+  https://digital.99hui.com "$APP_GIT_SHA" \
+  --backend-container avatar-backend \
+  --frontend-container avatar-frontend \
+  --expected-db-source /srv/digital-avatar/data/db \
+  --expected-upload-source /srv/digital-avatar/data/uploads
 docker compose exec avatar-backend python -c 'import embeddings; v=embeddings.embed(["部署向量探针"]); print(len(v), len(v[0]))'
 ```
 
+Jenkins 必须以 `verify-deployment.py` 返回成功作为发布成功条件，不能只以镜像构建或容器启动成功作为条件。脚本会同时核对公网前后端 Git SHA、数据库可读、上传目录可写、PDF OCR 依赖和宿主机数据挂载；任意一项不一致都会返回非零状态并阻止发布标绿。镜像使用 Git SHA 标签，不再依赖可被旧缓存覆盖的 `latest`。
+
 生产编排应把示例中的测试端口改为内网暴露，由统一 HTTPS 网关接入。后端暂时使用 SQLite，必须保持单实例写入；若扩展为多后端实例，应先迁移到 PostgreSQL，并把延迟接管任务改为共享队列。
 
 ## 4. 网关要求
@@ -153,7 +168,7 @@ location /api/ {
 5. 使用过期或伪造 token 时进入登录页并显示凭证失效，不得继续访问旧用户数据。
 6. 分身聊天 SSE 逐段输出正常，Markdown 正常渲染，知识库优先级和积分扣费正常。
 7. 开启 BOXIM 主动接管后保持在线，默认三分钟回复、自定义等待时间、已读回执、分身防回环和主人发言暂停均正常。
-8. 重建容器后数据库、头像、知识库文档仍存在，`/api/health` 返回成功。
+8. 重建容器后数据库、头像、知识库文档仍存在，`/api/health` 的 `gitSha` 与发布 SHA 一致，`database`、`uploads`、`pdfOcr` 三项检查均为 `true`。
 9. `https://digital.99hui.com/api/health` 可访问，证书域名和有效期正确，HTTP 自动跳转 HTTPS。
 10. 微信和支付宝各创建一笔最小套餐订单，未付款时积分不变；支付成功后回调到账一次，重复回调积分不重复增加。
 11. 微信虚拟支付在沙箱环境完成下单、支付回调、查单兜底和退款回调；错误 OpenID、商品、环境或金额均被拒绝。

digital-avatar-app/scripts/verify-deployment.py

@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+"""Fail a deployment unless frontend and backend run the expected release."""
+
+import argparse
+import json
+import os
+import subprocess
+import sys
+import urllib.request
+
+
+def fetch_json(url):
+    with urllib.request.urlopen(url, timeout=20) as response:
+        if response.status != 200:
+            raise RuntimeError(f"{url} returned HTTP {response.status}")
+        return json.load(response)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("base_url", help="Public site URL, for example https://digital.99hui.com")
+    parser.add_argument("expected_sha", help="Full Git commit SHA being deployed")
+    parser.add_argument("--backend-container", help="Backend container name for image and mount checks")
+    parser.add_argument("--frontend-container", help="Frontend container name for image checks")
+    parser.add_argument("--expected-db-source", help="Required host directory mounted for SQLite and its WAL files")
+    parser.add_argument("--expected-upload-source", help="Required host source mounted as the upload directory")
+    args = parser.parse_args()
+
+    base_url = args.base_url.rstrip("/")
+    errors = []
+    try:
+        health = fetch_json(f"{base_url}/api/health").get("data") or {}
+    except Exception as exc:
+        errors.append(f"cannot read backend release metadata: {exc}")
+        health = {}
+    try:
+        frontend = fetch_json(f"{base_url}/version.json")
+    except Exception as exc:
+        errors.append(f"cannot read frontend release metadata: {exc}")
+        frontend = {}
+
+    if health.get("status") != "ok":
+        errors.append(f"backend status is {health.get('status')!r}")
+    failed_checks = [name for name, passed in (health.get("checks") or {}).items() if not passed]
+    if failed_checks:
+        errors.append("backend checks failed: " + ", ".join(failed_checks))
+    if health.get("gitSha") != args.expected_sha:
+        errors.append(f"backend SHA is {health.get('gitSha')!r}")
+    if frontend.get("gitSha") != args.expected_sha:
+        errors.append(f"frontend SHA is {frontend.get('gitSha')!r}")
+
+    if args.backend_container:
+        backend = inspect_container(args.backend_container, errors)
+        check_container_revision(backend, args.expected_sha, "backend", errors)
+        check_mount(backend, args.expected_db_source, "database", errors)
+        check_mount(backend, args.expected_upload_source, "uploads", errors)
+    elif args.expected_db_source or args.expected_upload_source:
+        errors.append("--backend-container is required when checking data mounts")
+
+    if args.frontend_container:
+        frontend_container = inspect_container(args.frontend_container, errors)
+        check_container_revision(frontend_container, args.expected_sha, "frontend", errors)
+
+    if errors:
+        print("Deployment verification failed:", file=sys.stderr)
+        for error in errors:
+            print(f"- {error}", file=sys.stderr)
+        return 1
+
+    print(f"Deployment verified: {args.expected_sha}")
+    print("Backend checks: database, uploads, pdfOcr")
+    return 0
+
+
+def inspect_container(name, errors):
+    try:
+        output = subprocess.check_output(
+            ["docker", "inspect", name], universal_newlines=True
+        )
+        return json.loads(output)[0]
+    except Exception as exc:
+        errors.append(f"cannot inspect container {name!r}: {exc}")
+        return {}
+
+
+def check_container_revision(container, expected_sha, label, errors):
+    actual = ((container.get("Config") or {}).get("Labels") or {}).get(
+        "org.opencontainers.image.revision"
+    )
+    if actual != expected_sha:
+        errors.append(f"{label} container image SHA is {actual!r}")
+
+
+def check_mount(container, expected_source, label, errors):
+    if not expected_source:
+        return
+    expected = os.path.realpath(expected_source)
+    sources = {
+        os.path.realpath(mount.get("Source", ""))
+        for mount in container.get("Mounts") or []
+    }
+    if expected not in sources:
+        errors.append(f"{label} mount source {expected!r} is not attached")
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())