{
  "app_id": "meeting_client",
  "version": "3.2.1",
  "build_number": "20240115.1",
  "os": "windows",
  "os_version": "10.0.19045",
  "arch": "x64",
  "crash_type": "native_crash",
  "signal": "SIGSEGV",
  "address": "0x7ff812345678",
  "stack_trace": [...],
  "device_id": "hash(uuid)",
  "session_id": "meeting_abc123",
  "timestamp": 1705315200000,
  "custom_context": { "meeting_id": "12345", "user_role": "host" }
}

┌─────────────┐     ┌──────────────┐     ┌─────────────┐
│  CI/CD      │────▶│  对象存储    │◀───▶│  CDN 边缘   │
│  上传符号表  │     │  (版本/架构) │     │  就近下载   │
└─────────────┘     └──────────────┘     └─────────────┘
                           │
                           ▼
                    ┌──────────────┐
                    │  元数据索引  │
                    │  (Redis/ES)  │
                    └──────────────┘

graph LR
    A[崩溃原始数据] --> B{解析器}
    B -->|Native| C[addr2line / llvm-symbolizer]
    B -->|Java/Kotlin| D[Retrace / ReTrace]
    B -->|JS/TS| E[source-map-support]
    B -->|.NET| F[dotnet-symbol]
    C & D & E & F --> G[统一堆栈帧模型]
    G --> H[去重聚类 Key 生成]
    H --> I[入库 & 告警]

message StackFrame {
  string module = 1;           // 模块名
  string function = 2;         // 还原后函数签名
  string file = 3;             // 源文件路径
  int32 line = 4;              // 行号
  int32 column = 5;            // 列号
  uint64 address = 6;          // 原始地址
  uint64 offset_in_module = 7; // 模块内偏移
  bool inlined = 8;            // 是否内联帧
  repeated StackFrame inlined_frames = 9;
}

// 统一接口定义 (crash_adapter.h)
class ICrashAdapter {
public:
    virtual bool Initialize(const Config& cfg) = 0;
    virtual void SetUserContext(const std::string& key, const std::string& value) = 0;
    virtual void AddLua/JS/StackFrame(const Frame& frame) = 0; // 跨语言栈桥接
    virtual void TriggerTestCrash() = 0; // 供自动化测试调用
};

// 各平台实现
// Windows: CrashAdapterWin.cpp (SEH + MiniDumpWriteDump)
// macOS/iOS: CrashAdapterApple.mm (Mach Exception + PLCrashReporter 兼容层)
// Android: CrashAdapterAndroid.cpp (JNI + unwindstack + ART 栈走查)
// Web: crash-adapter-web.ts (ErrorEvent + PromiseRejectionEvent + WASM 栈解析)

-- 明细表：按天分区，按 (app_id, crash_group_id) 排序，TTL 30 天
CREATE TABLE crash_details (
    event_id UUID,
    app_id LowCardinality(String),
    version LowCardinality(String),
    build_number UInt32,
    os LowCardinality(String),
    arch LowCardinality(String),
    crash_type LowCardinality(String),
    crash_group_id UUID,          -- 聚类 ID
    crash_hash String,            -- Top-N Frames Hash
    stack_frames Array(Tuple(module String, function String, file String, line UInt32)),
    device_id_hash String,
    session_id String,
    custom_context Map(String, String), -- 业务扩展字段
    occur_time DateTime64(3),
    receive_time DateTime64(3),
    symbolicated Bool,
    symbolicate_latency_ms UInt32
) ENGINE = MergeTree()
PARTITION BY toDate(occur_time)
ORDER BY (app_id, crash_group_id, occur_time)
TTL occur_time + INTERVAL 30 DAY
SETTINGS index_granularity = 8192, compress_codes = ['ZSTD(3)'];

-- 物化视图：分钟级聚合，支撑大盘秒级查询
CREATE MATERIALIZED VIEW crash_agg_1m
ENGINE = SummingMergeTree()
PARTITION BY toDate(minute_bucket)
ORDER BY (app_id, version, crash_group_id, minute_bucket)
AS SELECT
    app_id, version, crash_group_id,
    toStartOfMinute(occur_time) AS minute_bucket,
    count() AS crash_count,
    uniqExact(device_id_hash) AS affected_users,
    quantileTDigest(0.5)(symbolicate_latency_ms) AS p50_symbolicate_latency
FROM crash_details
GROUP BY app_id, version, crash_group_id, minute_bucket;

# 离线构建流程 (每日增量)
def build_code_rag_index():
    # 1. 克隆单仓/多仓，按模块切片 (Tree-sitter AST 级切片，保留类/函数完整性)
    chunks = ast_chunking(repo_path, max_tokens=1500, overlap=200)
    
    # 2. 元数据注入：文件路径、Git Blame 最近提交者、模块归属、依赖图邻居
    for chunk in chunks:
        chunk.metadata = {
            "file": chunk.filepath,
            "owner": git_blame(chunk.filepath),
            "module": infer_module(chunk.filepath),
            "deps": get_call_graph_neighbors(chunk.function_name)
        }
    
    # 3. 嵌入向量化 (bge-m3 / text-embedding-3-small) -> 写入 Milvus/PGVector
    vector_store.upsert(chunks)

## System Prompt
你是资深客户端稳定性专家。输入：崩溃堆栈、寄存器快照、相关代码片段、近期变更记录。
输出：JSON 格式根因分析报告，包含：
1. root_cause: 一句话核心原因
2. evidence: 关键证据链 (堆栈帧 -> 代码行 -> 逻辑漏洞)
3. fix_suggestion: 可直接应用的代码修改建议 (Unified Diff 格式)
4. risk_level: P0/P1/P2
5. regression_test: 回归用例描述
6. confidence: 0.0-1.0

## Few-shot Examples (精选 20 个历史高质量复盘案例)
...

崩溃入库 → 判定是否「高频新崩溃/Top10/连续 3 版本未修复」 → 推送 Kafka Topic `crash_rca_request`
      ↓
RCA Worker (异步消费) → 拼装上下文 (堆栈 + 向量检索 Top-5 代码片段 + Git Log 最近 20 条) → 调用 LLM API (流式输出)
      ↓
结构化解析 JSON → 校验 Diff 语法正确性 (clang-format / black) → 写入 `crash_rca_result` 表
      ↓
飞书/钉钉机器人推送卡片 → 研发一键创建 Jira/工单 → 关联 PR Review 自动填充修复建议

{
  "config_version": "2024.01.15.10",
  "enable_native_capture": true,
  "enable_js_capture": true,
  "enable_anr_watchdog": true,
  "anr_threshold_ms": 5000,
  "max_local_cache_mb": 20,
  "upload_policy": {
    "wifi_charging": "realtime",
    "wifi_only": "batch_10min",
    "cellular": "batch_30min_sample_0.5"
  },
  "circuit_breakers": {
    "daily_crash_limit_per_device": 50,
    "global_qps_limit": 50000,
    "symbolicate_failure_rate_threshold": 0.3
  },
  "privacy": {
    "strip_pii_regex": ["(?i)password|token|auth", "meeting_id:\d+"],
    "max_context_kv_count": 20
  }
}

def calculate_health_score(version: str, window_hours: int = 24) -> float:
    """
    综合评分 = 0.4 * 无崩溃用户占比 + 0.3 * 核心流程崩溃率反向 + 0.2 * 新增崩溃组数反向 + 0.1 * 符号化成功率
    """
    metrics = query_clickhouse(f"""
        SELECT 
            uniqIf(device_id, crash_count=0) / uniq(device_id) AS crash_free_user_ratio,
            sumIf(crash_count, is_core_flow) / sum(crash_count) AS core_flow_crash_ratio,
            uniq(crash_group_id) AS new_groups,
            avg(symbolicated) AS symbol_rate
        FROM crash_agg_1m
        WHERE app_id='meeting' AND version='{version}' 
          AND minute_bucket >= now() - INTERVAL {window_hours} HOUR
    """)
    
    score = (0.4 * metrics.crash_free_user_ratio +
             0.3 * (1 - metrics.core_flow_crash_ratio) +
             0.2 * max(0, 1 - metrics.new_groups / 50) +
             0.1 * metrics.symbol_rate)
    return round(score * 100, 2)  # 0-100