diff --git a/packages/README.md b/packages/README.md
new file mode 100755
index 0000000..31e0c36
--- /dev/null
+++ b/packages/README.md
@@ -0,0 +1,46 @@
+# `packages/` — 业务域与能力包
+
+架构 §3.2：**单体不等于混乱**。这里的每个子目录都是一个边界，越界会被 CI 拦下。
+
+## 目录含义
+
+| 包 | 类型 | 内容 |
+| --- | --- | --- |
+| `domain/` | 纯逻辑层 | 领域模型与业务规则。**无 IO、无框架、可纯单测** |
+| `shared/` | 基础设施 | 配置、日志、错误码、统一响应、DB、Redis、可观测 |
+| `llm/` | AI 能力 | LLM 网关（已实现） |
+| `retrieval/` `rules/` `memory/` | AI 能力 | 混合检索 / 规则引擎 / 记忆系统（占位） |
+| `compliance/` | 业务域 | 统一合规出口（占位，CR-05/CR-06） |
+| `identity/` `billing/` `task/` `agent/` | 业务域 | 已有实现 |
+| `profile/` `report/` `tryon/` `kb/` `notify/` `wardrobe/` | 业务域 | 占位 |
+
+每个业务域包的标准结构：
+
+```
+packages/<域>/
+├─ __init__.py
+├─ README.md      这个域负责什么、不负责什么
+├─ service.py     对外唯一入口。其他域只能调这里
+└─ repo.py        SQL。只有本域可以 import
+```
+
+## 硬性规则（由 `make importlint` 机器校验）
+
+1. **禁止跨域直接查表**。要用别的域的数据，调它的 `service`，不要 import 它的
+   `repo`，更不要在自己的 SQL 里 JOIN 别人的表。表名的域前缀（`id_` / `bl_` /
+   `tk_` / …）就是为将来拆库预留的边界。
+2. **`packages/domain/` 必须保持纯净**。不许 import `fastapi`、`sqlalchemy`、
+   `httpx`、`redis`、`celery`，也不许 import `packages/shared`。放不进去说明
+   它不是领域逻辑，是应用逻辑，应该待在 `service.py`。
+3. **`compliance` 不得依赖任何业务域**。它必须能被任何模块安全调用。
+4. **`identity` 不得依赖任何其他业务域**。登录链路不能被业务拖累。
+5. **AI 能力包（`llm` / `retrieval` / `rules`）不得依赖业务域**，也不得 import
+   `apps`。它们要能在 V2 原样拆成独立服务。
+
+## 加一个新业务域时
+
+1. 建目录，写 `README.md` 说明职责与不负责的事。
+2. 在 `pyproject.toml` 的 `[tool.importlinter]` 契约里把它加进相应的
+   `layers` / `forbidden_modules` 列表 —— 新域默认没有约束，忘了加就等于没有边界。
+3. 领域规则先写进 `packages/domain/<域>/` 并配单测，再写 `service.py`。
+4. 表加域前缀，走 Alembic，迁移必须向后兼容（NFR-MAINT-06）。
diff --git a/packages/__init__.py b/packages/__init__.py
new file mode 100644
index 0000000..096fbb3
--- /dev/null
+++ b/packages/__init__.py
@@ -0,0 +1 @@
+"""业务域与能力包。域边界由 import-linter 强制（架构文档 §4.2）。"""
diff --git a/packages/domain/README.md b/packages/domain/README.md
new file mode 100755
index 0000000..fbf1f7f
--- /dev/null
+++ b/packages/domain/README.md
@@ -0,0 +1,39 @@
+# `packages/domain/` — 纯领域逻辑
+
+架构 §3.2 硬性规则 2：**领域逻辑放在这里，不依赖 FastAPI、不依赖数据库会话，
+可纯单元测试。**
+
+## 能放什么
+
+- 数据类与值对象（配额状态、任务状态、诊断维度…）
+- 判定与计算（配额是否够、状态能否流转、置信度冲突消解、色差计算…）
+- 纯函数、不可变数据结构
+
+## 绝对不能放什么
+
+- `import fastapi` / `sqlalchemy` / `httpx` / `redis` / `celery`
+- `import packages.shared`（连日志和配置都不行）
+- 任何 IO：读文件、发请求、查库、读环境变量
+- `async def` 里真的 await 了 IO 的东西
+
+违反会被 import-linter 的 `packages.domain is pure` 契约在 CI 拦下。
+
+## 为什么这么严
+
+这是 AI 大量写代码时唯一还能快速验证正确性的地方 —— 没有 mock、没有
+fixture、没有容器，几毫秒跑完几百个用例。一旦让 IO 渗进来，这个性质就没了，
+后面所有业务规则的测试都会退化成集成测试。
+
+## mypy strict
+
+本目录与 `packages/llm/` 一起，是 mypy **strict** 模式覆盖的范围
+（架构 §16.2）。所有函数必须有完整类型注解。
+
+## 现有内容
+
+| 模块 | 内容 |
+| --- | --- |
+| `ids.py` | 雪花 ID 生成（§6.1 原则 2） |
+| `errors.py` | 纯领域异常，适配层负责翻译成 `AppError` |
+| `billing/quota.py` | 配额三段式的算术与周期键（§7.4） |
+| `task/state.py` | `tk_task` 状态机（§6.2） |
diff --git a/packages/domain/__init__.py b/packages/domain/__init__.py
new file mode 100644
index 0000000..7091c6e
--- /dev/null
+++ b/packages/domain/__init__.py
@@ -0,0 +1 @@
+"""纯领域逻辑：无 IO、无框架、可纯单测（架构文档 §3.2）。"""
diff --git a/packages/domain/billing/__init__.py b/packages/domain/billing/__init__.py
new file mode 100644
index 0000000..77e3c69
--- /dev/null
+++ b/packages/domain/billing/__init__.py
@@ -0,0 +1 @@
+"""配额与账本的纯逻辑（架构文档 §7.4）。"""
diff --git a/packages/domain/billing/quota.py b/packages/domain/billing/quota.py
new file mode 100755
index 0000000..6086803
--- /dev/null
+++ b/packages/domain/billing/quota.py
@@ -0,0 +1,140 @@
+"""Quota arithmetic — pure logic, no IO (架构文档 §7.4 V1 配额三段式).
+
+The three-phase protocol lives here as data transformations so it can be unit
+tested without a database:
+
+    reserve  → 冻结额度并写 bl_ledger(status=reserved)
+    commit   → 确认扣除 (status=committed)
+    refund   → 释放额度并写反向流水 (tx_type=refund)
+
+V1 doesn't involve money, but 用户对「次数被白扣」的敏感度不比扣钱低，and V1
+has no customer-service refund path — the mechanism is the only guarantee.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, replace
+from datetime import UTC, datetime, timedelta, timezone
+from typing import Literal
+
+from packages.domain.errors import QuotaExceededError
+
+PeriodType = Literal["day", "month", "lifetime"]
+LedgerStatus = Literal["reserved", "committed", "refunded"]
+TxType = Literal["system_grant", "consume", "refund", "purchase", "gift"]
+
+#: FR-BILL-01: 跨时区按 UTC+8 统一。
+DEFAULT_TZ_OFFSET_HOURS = 8
+
+
+def period_key(
+    now: datetime,
+    period_type: PeriodType,
+    tz_offset_hours: int = DEFAULT_TZ_OFFSET_HOURS,
+) -> str:
+    """Bucket key that makes quota reset naturally, with no cron job.
+
+    ``day`` → ``d:2026-08-13``; ``month`` → ``m:2026-08``; ``lifetime`` → ``life``.
+    """
+    if period_type == "lifetime":
+        return "life"
+    local = now.astimezone(UTC) + timedelta(hours=tz_offset_hours)
+    if period_type == "day":
+        return f"d:{local:%Y-%m-%d}"
+    if period_type == "month":
+        return f"m:{local:%Y-%m}"
+    msg = f"unknown period_type: {period_type}"
+    raise ValueError(msg)
+
+
+def period_reset_at(
+    now: datetime,
+    period_type: PeriodType,
+    tz_offset_hours: int = DEFAULT_TZ_OFFSET_HOURS,
+) -> datetime | None:
+    """When the current period rolls over (used for the `reset_at` hint)."""
+    if period_type == "lifetime":
+        return None
+    tz = timezone(timedelta(hours=tz_offset_hours))
+    local = now.astimezone(tz)
+    if period_type == "day":
+        nxt = (local + timedelta(days=1)).replace(hour=0, minute=0, second=0, microsecond=0)
+    else:
+        year, month = (local.year + 1, 1) if local.month == 12 else (local.year, local.month + 1)
+        nxt = local.replace(
+            year=year, month=month, day=1, hour=0, minute=0, second=0, microsecond=0
+        )
+    return nxt.astimezone(UTC)
+
+
+@dataclass(frozen=True, slots=True)
+class QuotaState:
+    """Current period counters for one (user, quota_key)."""
+
+    used: int
+    base_limit: int
+    bonus: int = 0
+
+    @property
+    def effective_limit(self) -> int:
+        return self.base_limit + self.bonus
+
+    @property
+    def remaining(self) -> int:
+        return max(self.effective_limit - self.used, 0)
+
+
+@dataclass(frozen=True, slots=True)
+class QuotaDecision:
+    allowed: bool
+    required: int
+    remaining: int
+    effective_limit: int
+
+
+def evaluate(state: QuotaState, amount: int) -> QuotaDecision:
+    """Server-side hard check (FR-BILL-06). Client hints are UX only."""
+    if amount <= 0:
+        msg = "amount must be positive"
+        raise ValueError(msg)
+    return QuotaDecision(
+        allowed=state.used + amount <= state.effective_limit,
+        required=amount,
+        remaining=state.remaining,
+        effective_limit=state.effective_limit,
+    )
+
+
+def reserve(state: QuotaState, amount: int) -> QuotaState:
+    """Freeze ``amount`` units; raises :class:`QuotaExceededError` when over the limit."""
+    decision = evaluate(state, amount)
+    if not decision.allowed:
+        raise QuotaExceededError(required=amount, remaining=decision.remaining)
+    return replace(state, used=state.used + amount)
+
+
+def refund(state: QuotaState, amount: int) -> QuotaState:
+    """Release a previously reserved amount. Never drives ``used`` below zero."""
+    if amount <= 0:
+        msg = "amount must be positive"
+        raise ValueError(msg)
+    return replace(state, used=max(state.used - amount, 0))
+
+
+def balance_after(state: QuotaState) -> int:
+    """Value written to ``bl_ledger.balance_after``.
+
+    V1 has no wallet balance, so the ledger records *remaining quota* — that is
+    the number the user actually sees ("今日剩余 N 次"). When V2 introduces
+    ``bl_wallet`` this becomes the wallet balance without a schema change.
+    """
+    return state.remaining
+
+
+def reserve_idempotency_key(*, user_id: int, quota_key: str, request_key: str) -> str:
+    return f"rsv:{user_id}:{quota_key}:{request_key}"
+
+
+def refund_idempotency_key(reserve_key: str) -> str:
+    """FR-BILL-05: 退还必须幂等。Derived, so a retry cannot double-refund."""
+    return f"rfd:{reserve_key}"
diff --git a/packages/domain/errors.py b/packages/domain/errors.py
new file mode 100755
index 0000000..3dda900
--- /dev/null
+++ b/packages/domain/errors.py
@@ -0,0 +1,26 @@
+"""Pure-domain exceptions.
+
+This layer must not import :mod:`packages.shared` (enforced by import-linter),
+so it defines its own error types. Adapters translate them into
+:class:`packages.shared.errors.AppError` at the boundary.
+"""
+
+from __future__ import annotations
+
+
+class DomainError(Exception):
+    """Base class for all pure-domain rule violations."""
+
+
+class QuotaExceededError(DomainError):
+    def __init__(self, *, required: int, remaining: int) -> None:
+        self.required = required
+        self.remaining = remaining
+        super().__init__(f"quota exceeded: required={required} remaining={remaining}")
+
+
+class InvalidTaskTransitionError(DomainError):
+    def __init__(self, *, current: str, target: str) -> None:
+        self.current = current
+        self.target = target
+        super().__init__(f"invalid task transition: {current} -> {target}")
diff --git a/packages/domain/ids.py b/packages/domain/ids.py
new file mode 100755
index 0000000..d6efd42
--- /dev/null
+++ b/packages/domain/ids.py
@@ -0,0 +1,85 @@
+"""Snowflake IDs (架构文档 §6.1 原则 2).
+
+Auto-increment keys invite enumeration attacks and collide when the database is
+later sharded; UUID primary keys have poor index locality. A 64-bit snowflake
+avoids both.
+
+Layout: 41 bits ms since epoch | 10 bits node id | 12 bits sequence.
+"""
+
+from __future__ import annotations
+
+import threading
+import time
+from collections.abc import Callable
+
+#: 2025-01-01T00:00:00Z in milliseconds. Do not change once IDs are issued.
+DEFAULT_EPOCH_MS = 1_735_689_600_000
+
+_NODE_BITS = 10
+_SEQUENCE_BITS = 12
+_MAX_NODE_ID = (1 << _NODE_BITS) - 1
+_SEQUENCE_MASK = (1 << _SEQUENCE_BITS) - 1
+
+
+def _wall_clock_ms() -> int:
+    return int(time.time() * 1000)
+
+
+class SnowflakeGenerator:
+    """Thread-safe monotonic 64-bit id generator."""
+
+    __slots__ = ("_clock", "_epoch_ms", "_last_ms", "_lock", "_node_id", "_sequence")
+
+    def __init__(
+        self,
+        node_id: int = 1,
+        epoch_ms: int = DEFAULT_EPOCH_MS,
+        clock: Callable[[], int] | None = None,
+    ) -> None:
+        if not 0 <= node_id <= _MAX_NODE_ID:
+            msg = f"node_id must be within 0..{_MAX_NODE_ID}"
+            raise ValueError(msg)
+        self._node_id = node_id
+        self._epoch_ms = epoch_ms
+        self._sequence = 0
+        self._last_ms = -1
+        self._lock = threading.Lock()
+        self._clock = clock or _wall_clock_ms
+
+    def _now_ms(self) -> int:
+        return self._clock()
+
+    def next_id(self) -> int:
+        with self._lock:
+            now = self._now_ms()
+            if now < self._last_ms:
+                # Clock moved backwards: wait it out rather than issue duplicates.
+                now = self._last_ms
+            if now == self._last_ms:
+                self._sequence = (self._sequence + 1) & _SEQUENCE_MASK
+                if self._sequence == 0:
+                    while now <= self._last_ms:
+                        now = self._now_ms()
+            else:
+                self._sequence = 0
+            self._last_ms = now
+            elapsed = now - self._epoch_ms
+            return (
+                (elapsed << (_NODE_BITS + _SEQUENCE_BITS))
+                | (self._node_id << _SEQUENCE_BITS)
+                | self._sequence
+            )
+
+
+_default_generator = SnowflakeGenerator()
+
+
+def next_id() -> int:
+    """Process-wide default generator."""
+    return _default_generator.next_id()
+
+
+def timestamp_ms_of(snowflake_id: int, epoch_ms: int = DEFAULT_EPOCH_MS) -> int:
+    """Extract the creation timestamp from an id (useful for partition pruning)."""
+    return (snowflake_id >> (_NODE_BITS + _SEQUENCE_BITS)) + epoch_ms
diff --git a/packages/domain/task/__init__.py b/packages/domain/task/__init__.py
new file mode 100644
index 0000000..a5598e1
--- /dev/null
+++ b/packages/domain/task/__init__.py
@@ -0,0 +1 @@
+"""统一任务中心的状态机（架构文档 §6.2）。"""
diff --git a/packages/domain/task/state.py b/packages/domain/task/state.py
new file mode 100755
index 0000000..2aa4c3e
--- /dev/null
+++ b/packages/domain/task/state.py
@@ -0,0 +1,76 @@
+"""tk_task state machine — pure logic (架构文档 §6.2 task 域, FR-TASK-01).
+
+统一任务中心 is the only long-running-task view in the system; every long task
+(report / tryon / hairstyle / matting / render) moves through these states.
+"""
+
+from __future__ import annotations
+
+from enum import StrEnum
+
+from packages.domain.errors import InvalidTaskTransitionError
+
+
+class TaskStatus(StrEnum):
+    PENDING = "pending"  # created, quota reserved, not yet queued
+    QUEUED = "queued"  # handed to Celery
+    RUNNING = "running"
+    SUCCEEDED = "succeeded"
+    FAILED = "failed"
+    CANCELED = "canceled"
+
+
+class TaskType(StrEnum):
+    REPORT = "report"
+    TRYON = "tryon"
+    HAIRSTYLE = "hairstyle"
+    MATTING = "matting"
+    RENDER = "render"
+    PING = "ping"  # scaffold-only: proves the async path end to end
+
+
+TERMINAL_STATUSES: frozenset[TaskStatus] = frozenset(
+    {TaskStatus.SUCCEEDED, TaskStatus.FAILED, TaskStatus.CANCELED}
+)
+
+#: Refunds are owed for every terminal state other than success (FR-TRY-05).
+REFUNDABLE_STATUSES: frozenset[TaskStatus] = frozenset({TaskStatus.FAILED, TaskStatus.CANCELED})
+
+ALLOWED_TRANSITIONS: dict[TaskStatus, frozenset[TaskStatus]] = {
+    TaskStatus.PENDING: frozenset({TaskStatus.QUEUED, TaskStatus.FAILED, TaskStatus.CANCELED}),
+    TaskStatus.QUEUED: frozenset({TaskStatus.RUNNING, TaskStatus.FAILED, TaskStatus.CANCELED}),
+    TaskStatus.RUNNING: frozenset(
+        {TaskStatus.SUCCEEDED, TaskStatus.FAILED, TaskStatus.CANCELED, TaskStatus.QUEUED}
+    ),
+    TaskStatus.SUCCEEDED: frozenset(),
+    TaskStatus.FAILED: frozenset({TaskStatus.QUEUED}),  # retry re-queues
+    TaskStatus.CANCELED: frozenset(),
+}
+
+
+def can_transition(current: TaskStatus, target: TaskStatus) -> bool:
+    return target in ALLOWED_TRANSITIONS[current]
+
+
+def assert_transition(current: TaskStatus, target: TaskStatus) -> None:
+    if not can_transition(current, target):
+        raise InvalidTaskTransitionError(current=current.value, target=target.value)
+
+
+def is_terminal(status: TaskStatus) -> bool:
+    return status in TERMINAL_STATUSES
+
+
+def needs_refund(status: TaskStatus) -> bool:
+    return status in REFUNDABLE_STATUSES
+
+
+def progress_for(status: TaskStatus) -> int:
+    return {
+        TaskStatus.PENDING: 0,
+        TaskStatus.QUEUED: 5,
+        TaskStatus.RUNNING: 50,
+        TaskStatus.SUCCEEDED: 100,
+        TaskStatus.FAILED: 100,
+        TaskStatus.CANCELED: 100,
+    }[status]
diff --git a/packages/shared/README.md b/packages/shared/README.md
new file mode 100755
index 0000000..9cc5fb0
--- /dev/null
+++ b/packages/shared/README.md
@@ -0,0 +1,40 @@
+# `packages/shared/` — 横切基础设施
+
+配置、日志、错误码、统一响应、数据库、Redis、可观测性。架构 §16.1。
+
+| 模块 | 内容 |
+| --- | --- |
+| `config.py` | pydantic-settings 类型化配置，全部来自环境变量 |
+| `errors.py` | 统一错误码（§9.2：4xxxx 客户端 / 5xxxx 服务端 / 6xxxx 外部依赖） |
+| `response.py` | 统一响应信封 `{code, message, data, request_id, hint}` |
+| `handlers.py` | FastAPI 异常处理器，保证任何异常都返回信封且不泄露堆栈 |
+| `middleware.py` | 请求 ID / 追踪 ID、访问日志、HTTP 指标、`/v1` 路径别名 |
+| `logging.py` | structlog JSON 日志 + **脱敏处理器** |
+| `context.py` | 请求上下文（request_id / trace_id / user_id / task_id） |
+| `db.py` | 异步 SQLAlchemy 引擎与会话 |
+| `models.py` | ORM 表定义 |
+| `redis_client.py` | Redis 连接 |
+| `observability.py` | OTel tracing + Prometheus 指标注册表 |
+
+## 能放什么
+
+跨所有域复用、且不含任何业务判断的东西。
+
+## 不能放什么
+
+- **任何业务规则**。「配额够不够」属于 `packages/domain/billing`，
+  「任务能不能取消」属于 `packages/domain/task`。放进来就没人能纯单测了。
+- 对具体业务域的 import。`shared` 在依赖层级的底部，反向依赖会立刻形成环。
+
+## 日志脱敏不是可选项
+
+`logging.py` 的脱敏是 structlog **处理器**，作用于每一条日志，没有 opt-in。
+新增敏感字段时加进 `SENSITIVE_KEYS`，或用 `LOG_EXTRA_REDACT_KEYS` 环境变量。
+NFR-SEC-03 / 架构 §11.2：手机号、UnionID、图片地址、诊断结论、令牌一律脱敏。
+
+## 新增错误码
+
+在 `ErrorCode` 加值，**同时**在 `ERROR_META` 加 message 和 hint。
+测试 `test_every_error_code_has_metadata` 会强制这一点。
+FR-BILL-07：配额相关文案里不能出现价格、购买、会员、升级字样，
+`test_quota_hint_never_mentions_price_or_purchase` 会扫描全部文案。
diff --git a/packages/shared/__init__.py b/packages/shared/__init__.py
new file mode 100644
index 0000000..ab08ba6
--- /dev/null
+++ b/packages/shared/__init__.py
@@ -0,0 +1 @@
+"""横切基础设施：配置、日志、错误、响应、DB、Redis、可观测。"""
diff --git a/packages/shared/config.py b/packages/shared/config.py
new file mode 100755
index 0000000..e20532b
--- /dev/null
+++ b/packages/shared/config.py
@@ -0,0 +1,148 @@
+"""Typed application configuration.
+
+Every value comes from the environment (架构文档 §11.2: 密钥不进代码/镜像).
+`.env` is only a developer convenience; dev/staging/prod inject via SSM/KMS.
+"""
+
+from __future__ import annotations
+
+import json
+from decimal import Decimal
+from functools import lru_cache
+from typing import Any, Literal
+
+from pydantic import Field, ValidationInfo, field_validator
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+Environment = Literal["local", "dev", "staging", "prod"]
+CachePrefixMode = Literal["strict", "warn", "off"]
+
+_PLACEHOLDER_JWT_SECRET = "change-me-local-only-do-not-use-outside-local"  # noqa: S105
+
+
+class Settings(BaseSettings):
+    """Root settings object. Read it through :func:`get_settings`."""
+
+    model_config = SettingsConfigDict(
+        env_file=".env",
+        env_file_encoding="utf-8",
+        extra="ignore",
+        case_sensitive=False,
+    )
+
+    # --- Application -------------------------------------------------------
+    app_env: Environment = "local"
+    app_name: str = "meida-backend"
+    api_prefix: str = "/api/v1"
+    debug: bool = False
+    log_level: str = "INFO"
+    log_json: bool = True
+    log_extra_redact_keys: str = ""
+
+    # --- Database ----------------------------------------------------------
+    database_url: str = "postgresql+asyncpg://meida:meida@localhost:55432/meida"
+    db_pool_size: int = 10
+    db_max_overflow: int = 20
+    db_echo: bool = False
+
+    # --- Redis / Celery ----------------------------------------------------
+    redis_url: str = "redis://localhost:56379/0"
+    celery_broker_url: str = "redis://localhost:56379/1"
+    celery_result_backend: str = "redis://localhost:56379/2"
+
+    # --- Auth --------------------------------------------------------------
+    jwt_secret: str = _PLACEHOLDER_JWT_SECRET
+    jwt_algorithm: str = "HS256"
+    jwt_access_ttl_seconds: int = 900
+    jwt_issuer: str = "meida"
+    dev_auth_enabled: bool = True
+
+    # --- LLM routing -------------------------------------------------------
+    llm_primary_provider: str = "deepseek"
+    llm_backup_provider: str = "zhipu"
+    llm_structured_max_retries: int = 2
+    llm_circuit_failure_threshold: int = 3
+    llm_circuit_cooldown_seconds: float = 30.0
+    llm_scene_overrides: dict[str, dict[str, Any]] = Field(
+        default_factory=dict, alias="LLM_SCENE_OVERRIDES_JSON"
+    )
+
+    # --- DeepSeek (primary, OpenAI-compatible) -----------------------------
+    deepseek_api_key: str = ""
+    deepseek_base_url: str = "https://api.deepseek.com/v1"
+    deepseek_model: str = "deepseek-v4-flash"
+    deepseek_timeout_seconds: float = 30.0
+
+    # --- 智谱 GLM (backup slot, not implemented in V1) ---------------------
+    zhipu_api_key: str = ""
+    zhipu_base_url: str = "https://open.bigmodel.cn/api/paas/v4"
+    zhipu_model: str = "glm-4-flash"
+    zhipu_timeout_seconds: float = 30.0
+
+    # --- Prompt caching discipline (架构文档 §7.2) -------------------------
+    llm_cache_prefix_mode: CachePrefixMode = "strict"
+    prompts_dir: str = "prompts"
+
+    # --- Cost attribution (架构文档 §13) -----------------------------------
+    llm_cost_rates: dict[str, dict[str, Decimal]] = Field(
+        default_factory=dict, alias="LLM_COST_RATES_JSON"
+    )
+    llm_default_cost_rate: dict[str, Decimal] = Field(
+        default_factory=lambda: {
+            "input_cents_per_mtok": Decimal("100"),
+            "cached_input_cents_per_mtok": Decimal("20"),
+            "output_cents_per_mtok": Decimal("200"),
+        },
+        alias="LLM_DEFAULT_COST_RATE_JSON",
+    )
+
+    # --- Global daily cost circuit breaker ---------------------------------
+    llm_daily_budget_enabled: bool = True
+    llm_daily_budget_cents: Decimal = Decimal("100000")
+    llm_budget_timezone_offset_hours: int = 8
+
+    # --- Observability -----------------------------------------------------
+    otel_enabled: bool = False
+    otel_service_name: str = "meida-api"
+    otel_exporter_otlp_endpoint: str = ""
+    prometheus_enabled: bool = True
+
+    @field_validator(
+        "llm_scene_overrides", "llm_cost_rates", "llm_default_cost_rate", mode="before"
+    )
+    @classmethod
+    def _parse_json_object(cls, value: Any) -> Any:
+        if isinstance(value, str):
+            stripped = value.strip()
+            if not stripped:
+                return {}
+            return json.loads(stripped)
+        return value
+
+    @field_validator("jwt_secret")
+    @classmethod
+    def _reject_placeholder_secret_outside_local(cls, value: str, info: ValidationInfo) -> str:
+        env = info.data.get("app_env", "local")
+        if env in ("staging", "prod") and (value == _PLACEHOLDER_JWT_SECRET or len(value) < 32):
+            msg = "JWT_SECRET must be set to a >=32 char random value outside local/dev"
+            raise ValueError(msg)
+        return value
+
+    @property
+    def is_local(self) -> bool:
+        return self.app_env in ("local", "dev")
+
+    @property
+    def extra_redact_keys(self) -> tuple[str, ...]:
+        return tuple(k.strip() for k in self.log_extra_redact_keys.split(",") if k.strip())
+
+    @property
+    def sync_database_url(self) -> str:
+        """psycopg-style URL, used by tooling that cannot speak asyncpg."""
+        return self.database_url.replace("+asyncpg", "")
+
+
+@lru_cache(maxsize=1)
+def get_settings() -> Settings:
+    """Process-wide settings singleton."""
+    return Settings()
diff --git a/packages/shared/context.py b/packages/shared/context.py
new file mode 100755
index 0000000..5bd8623
--- /dev/null
+++ b/packages/shared/context.py
@@ -0,0 +1,75 @@
+"""Request-scoped context propagated into logs and traces (架构文档 §15)."""
+
+from __future__ import annotations
+
+import uuid
+from contextvars import ContextVar, Token
+from dataclasses import dataclass
+
+_request_id: ContextVar[str | None] = ContextVar("meida_request_id", default=None)
+_trace_id: ContextVar[str | None] = ContextVar("meida_trace_id", default=None)
+_user_id: ContextVar[int | None] = ContextVar("meida_user_id", default=None)
+_task_id: ContextVar[int | None] = ContextVar("meida_task_id", default=None)
+
+
+@dataclass(slots=True)
+class ContextTokens:
+    request_id: Token[str | None]
+    trace_id: Token[str | None]
+    user_id: Token[int | None]
+    task_id: Token[int | None]
+
+
+def new_request_id() -> str:
+    return uuid.uuid4().hex
+
+
+def bind_context(
+    *,
+    request_id: str | None = None,
+    trace_id: str | None = None,
+    user_id: int | None = None,
+    task_id: int | None = None,
+) -> ContextTokens:
+    return ContextTokens(
+        request_id=_request_id.set(request_id if request_id is not None else _request_id.get()),
+        trace_id=_trace_id.set(trace_id if trace_id is not None else _trace_id.get()),
+        user_id=_user_id.set(user_id if user_id is not None else _user_id.get()),
+        task_id=_task_id.set(task_id if task_id is not None else _task_id.get()),
+    )
+
+
+def reset_context(tokens: ContextTokens) -> None:
+    _request_id.reset(tokens.request_id)
+    _trace_id.reset(tokens.trace_id)
+    _user_id.reset(tokens.user_id)
+    _task_id.reset(tokens.task_id)
+
+
+def get_request_id() -> str | None:
+    return _request_id.get()
+
+
+def get_trace_id() -> str | None:
+    return _trace_id.get()
+
+
+def get_user_id() -> int | None:
+    return _user_id.get()
+
+
+def get_task_id() -> int | None:
+    return _task_id.get()
+
+
+def context_snapshot() -> dict[str, str | int]:
+    snapshot: dict[str, str | int] = {}
+    for key, value in (
+        ("request_id", _request_id.get()),
+        ("trace_id", _trace_id.get()),
+        ("user_id", _user_id.get()),
+        ("task_id", _task_id.get()),
+    ):
+        if value is not None:
+            snapshot[key] = value
+    return snapshot
diff --git a/packages/shared/db.py b/packages/shared/db.py
new file mode 100755
index 0000000..babf7e8
--- /dev/null
+++ b/packages/shared/db.py
@@ -0,0 +1,73 @@
+"""Async SQLAlchemy engine and session management (架构文档 §5, §12.2).
+
+The engine is created lazily so that importing this module never opens a
+connection — important for unit tests and for Celery workers that fork.
+"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncIterator
+from contextlib import asynccontextmanager
+
+from sqlalchemy.ext.asyncio import (
+    AsyncEngine,
+    AsyncSession,
+    async_sessionmaker,
+    create_async_engine,
+)
+
+from packages.shared.config import Settings, get_settings
+
+_engine: AsyncEngine | None = None
+_session_factory: async_sessionmaker[AsyncSession] | None = None
+
+
+def create_engine(settings: Settings | None = None) -> AsyncEngine:
+    cfg = settings or get_settings()
+    return create_async_engine(
+        cfg.database_url,
+        echo=cfg.db_echo,
+        pool_size=cfg.db_pool_size,
+        max_overflow=cfg.db_max_overflow,
+        pool_pre_ping=True,
+        pool_recycle=1800,
+    )
+
+
+def get_engine() -> AsyncEngine:
+    global _engine
+    if _engine is None:
+        _engine = create_engine()
+    return _engine
+
+
+def get_session_factory() -> async_sessionmaker[AsyncSession]:
+    global _session_factory
+    if _session_factory is None:
+        _session_factory = async_sessionmaker(
+            bind=get_engine(),
+            expire_on_commit=False,
+            autoflush=False,
+        )
+    return _session_factory
+
+
+@asynccontextmanager
+async def session_scope() -> AsyncIterator[AsyncSession]:
+    """Transactional scope: commit on success, rollback on error."""
+    factory = get_session_factory()
+    async with factory() as session:
+        try:
+            yield session
+            await session.commit()
+        except Exception:
+            await session.rollback()
+            raise
+
+
+async def dispose_engine() -> None:
+    global _engine, _session_factory
+    if _engine is not None:
+        await _engine.dispose()
+    _engine = None
+    _session_factory = None
diff --git a/packages/shared/errors.py b/packages/shared/errors.py
new file mode 100755
index 0000000..608567a
--- /dev/null
+++ b/packages/shared/errors.py
@@ -0,0 +1,138 @@
+"""Unified error codes (架构文档 §9.2, FR-SYS-02).
+
+Segments:
+    ``4xxxx`` client error, ``5xxxx`` server error, ``6xxxx`` external dependency.
+
+Splitting 5xxxx from 6xxxx lets monitoring separate "our bug" from "vendor
+outage" without parsing messages.
+
+Every code carries a user-readable ``message`` and an optional ``hint`` telling
+the user what to do next — required by FR-SYS-02 and NFR-UX-02.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from enum import IntEnum
+from typing import Any
+
+
+class ErrorCode(IntEnum):
+    OK = 0
+
+    # --- 4xxxx client -----------------------------------------------------
+    BAD_REQUEST = 40000
+    VALIDATION_FAILED = 40001
+    IDEMPOTENCY_CONFLICT = 40002
+    UNAUTHENTICATED = 40100
+    TOKEN_EXPIRED = 40101
+    TOKEN_INVALID = 40102
+    FORBIDDEN = 40300
+    QUOTA_EXHAUSTED = 40301
+    CONSENT_REQUIRED = 40302
+    MINOR_RESTRICTED = 40303
+    NOT_FOUND = 40400
+    STATE_CONFLICT = 40900
+    RATE_LIMITED = 42900
+
+    # --- 5xxxx server -----------------------------------------------------
+    INTERNAL_ERROR = 50000
+    DB_ERROR = 50001
+    TASK_STATE_INVALID = 50002
+    CONFIG_ERROR = 50003
+    NOT_IMPLEMENTED = 50100
+
+    # --- 6xxxx external dependency ---------------------------------------
+    UPSTREAM_ERROR = 60000
+    LLM_ALL_PROVIDERS_FAILED = 60001
+    LLM_STRUCTURED_OUTPUT_FAILED = 60002
+    LLM_DAILY_BUDGET_EXCEEDED = 60003
+    LLM_PROMPT_CACHE_VIOLATION = 60004
+    MODERATION_UNAVAILABLE = 60005
+
+
+@dataclass(frozen=True, slots=True)
+class ErrorMeta:
+    http_status: int
+    message: str
+    hint: str | None = None
+
+
+ERROR_META: dict[ErrorCode, ErrorMeta] = {
+    ErrorCode.OK: ErrorMeta(200, "ok"),
+    ErrorCode.BAD_REQUEST: ErrorMeta(400, "请求参数有误", "请检查后重试"),
+    ErrorCode.VALIDATION_FAILED: ErrorMeta(422, "请求参数校验未通过", "请按提示修改后重试"),
+    ErrorCode.IDEMPOTENCY_CONFLICT: ErrorMeta(
+        409, "相同请求正在处理中", "请稍候查看任务中心，不要重复提交"
+    ),
+    ErrorCode.UNAUTHENTICATED: ErrorMeta(401, "请先登录", "登录后即可继续"),
+    ErrorCode.TOKEN_EXPIRED: ErrorMeta(401, "登录状态已过期", "请重新登录"),
+    ErrorCode.TOKEN_INVALID: ErrorMeta(401, "登录状态无效", "请重新登录"),
+    ErrorCode.FORBIDDEN: ErrorMeta(403, "没有访问权限"),
+    ErrorCode.QUOTA_EXHAUSTED: ErrorMeta(403, "今日额度已用完", "明天恢复，或完成任务多换几次"),
+    ErrorCode.CONSENT_REQUIRED: ErrorMeta(403, "需要先完成授权", "请阅读并同意后继续"),
+    ErrorCode.MINOR_RESTRICTED: ErrorMeta(403, "该功能未向未成年人开放"),
+    ErrorCode.NOT_FOUND: ErrorMeta(404, "内容不存在或已删除"),
+    ErrorCode.STATE_CONFLICT: ErrorMeta(409, "当前状态无法执行该操作"),
+    ErrorCode.RATE_LIMITED: ErrorMeta(429, "操作太频繁了", "请稍后再试"),
+    ErrorCode.INTERNAL_ERROR: ErrorMeta(500, "服务开小差了", "请稍后重试"),
+    ErrorCode.DB_ERROR: ErrorMeta(500, "服务开小差了", "请稍后重试"),
+    ErrorCode.TASK_STATE_INVALID: ErrorMeta(500, "任务状态异常", "请到任务中心重试"),
+    ErrorCode.CONFIG_ERROR: ErrorMeta(500, "服务配置异常", "请联系客服"),
+    ErrorCode.NOT_IMPLEMENTED: ErrorMeta(501, "该能力尚未开放"),
+    ErrorCode.UPSTREAM_ERROR: ErrorMeta(502, "外部服务暂时不可用", "请稍后重试"),
+    ErrorCode.LLM_ALL_PROVIDERS_FAILED: ErrorMeta(
+        502, "AI 服务暂时不可用", "请稍后重试，额度不会被扣除"
+    ),
+    ErrorCode.LLM_STRUCTURED_OUTPUT_FAILED: ErrorMeta(
+        502, "AI 返回结果格式异常", "请稍后重试，额度不会被扣除"
+    ),
+    ErrorCode.LLM_DAILY_BUDGET_EXCEEDED: ErrorMeta(
+        503, "AI 能力今日已达服务上限", "请明天再试，核心功能不受影响"
+    ),
+    ErrorCode.LLM_PROMPT_CACHE_VIOLATION: ErrorMeta(500, "服务配置异常", "请联系客服"),
+    ErrorCode.MODERATION_UNAVAILABLE: ErrorMeta(
+        503, "内容安全服务不可用", "为保证安全本次请求已中止，请稍后重试"
+    ),
+}
+
+
+class AppError(Exception):
+    """Base application error carrying a unified error code.
+
+    Never let a raw traceback reach the client (FR-SYS-02).
+    """
+
+    def __init__(
+        self,
+        code: ErrorCode,
+        *,
+        message: str | None = None,
+        hint: str | None = None,
+        data: dict[str, Any] | None = None,
+    ) -> None:
+        meta = ERROR_META.get(code, ERROR_META[ErrorCode.INTERNAL_ERROR])
+        self.code = code
+        self.message = message or meta.message
+        self.hint = hint if hint is not None else meta.hint
+        self.http_status = meta.http_status
+        self.data = data or {}
+        super().__init__(f"[{int(code)}] {self.message}")
+
+
+class QuotaExhaustedError(AppError):
+    def __init__(self, *, required: int, remaining: int, reset_at: str | None = None) -> None:
+        super().__init__(
+            ErrorCode.QUOTA_EXHAUSTED,
+            data={"required": required, "remaining": remaining, "reset_at": reset_at},
+        )
+
+
+class NotFoundError(AppError):
+    def __init__(self, message: str | None = None) -> None:
+        super().__init__(ErrorCode.NOT_FOUND, message=message)
+
+
+class ForbiddenError(AppError):
+    def __init__(self, message: str | None = None) -> None:
+        super().__init__(ErrorCode.FORBIDDEN, message=message)
diff --git a/packages/shared/handlers.py b/packages/shared/handlers.py
new file mode 100755
index 0000000..fecf74b
--- /dev/null
+++ b/packages/shared/handlers.py
@@ -0,0 +1,69 @@
+"""FastAPI exception handlers producing the unified envelope (架构文档 §9.2).
+
+No handler is allowed to leak a stack trace to the client (FR-SYS-02).
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from fastapi import FastAPI, Request
+from fastapi.exceptions import RequestValidationError
+from fastapi.responses import JSONResponse
+from starlette.exceptions import HTTPException as StarletteHTTPException
+
+from packages.shared.errors import ERROR_META, AppError, ErrorCode
+from packages.shared.logging import get_logger
+from packages.shared.response import fail
+
+logger = get_logger(__name__)
+
+_HTTP_STATUS_TO_CODE: dict[int, ErrorCode] = {
+    400: ErrorCode.BAD_REQUEST,
+    401: ErrorCode.UNAUTHENTICATED,
+    403: ErrorCode.FORBIDDEN,
+    404: ErrorCode.NOT_FOUND,
+    405: ErrorCode.BAD_REQUEST,
+    409: ErrorCode.STATE_CONFLICT,
+    422: ErrorCode.VALIDATION_FAILED,
+    429: ErrorCode.RATE_LIMITED,
+    501: ErrorCode.NOT_IMPLEMENTED,
+    502: ErrorCode.UPSTREAM_ERROR,
+    503: ErrorCode.UPSTREAM_ERROR,
+}
+
+
+def _json(status: int, body: dict[str, Any]) -> JSONResponse:
+    return JSONResponse(status_code=status, content=body)
+
+
+def install_exception_handlers(app: FastAPI) -> None:
+    @app.exception_handler(AppError)
+    async def _app_error(_request: Request, exc: AppError) -> JSONResponse:
+        logger.warning("app_error", code=int(exc.code), error_message=exc.message)
+        return _json(
+            exc.http_status,
+            fail(exc.code, message=exc.message, hint=exc.hint, data=exc.data or None),
+        )
+
+    @app.exception_handler(RequestValidationError)
+    async def _validation_error(_request: Request, exc: RequestValidationError) -> JSONResponse:
+        details = [
+            {"field": ".".join(str(p) for p in err.get("loc", ())), "reason": err.get("msg", "")}
+            for err in exc.errors()
+        ]
+        return _json(
+            422,
+            fail(ErrorCode.VALIDATION_FAILED, data={"errors": details}),
+        )
+
+    @app.exception_handler(StarletteHTTPException)
+    async def _http_error(_request: Request, exc: StarletteHTTPException) -> JSONResponse:
+        code = _HTTP_STATUS_TO_CODE.get(exc.status_code, ErrorCode.INTERNAL_ERROR)
+        message = exc.detail if isinstance(exc.detail, str) else ERROR_META[code].message
+        return _json(exc.status_code, fail(code, message=message))
+
+    @app.exception_handler(Exception)
+    async def _unhandled(_request: Request, exc: Exception) -> JSONResponse:
+        logger.exception("unhandled_error", error_type=type(exc).__name__)
+        return _json(500, fail(ErrorCode.INTERNAL_ERROR))
diff --git a/packages/shared/logging.py b/packages/shared/logging.py
new file mode 100755
index 0000000..ed80e6c
--- /dev/null
+++ b/packages/shared/logging.py
@@ -0,0 +1,165 @@
+"""Structured JSON logging with request/trace propagation and redaction.
+
+NFR-SEC-03 / 架构文档 §11.2: 手机号、UnionID、图片地址、诊断结论、令牌
+一律脱敏或不记录. Redaction runs as a structlog processor so it applies to
+every log line regardless of who wrote it — there is no opt-in to forget.
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+import sys
+from collections.abc import Iterable, MutableMapping
+from typing import Any
+
+import structlog
+
+from packages.shared.context import context_snapshot
+
+REDACTED = "***"
+
+#: Field names whose values must never be logged in clear text.
+SENSITIVE_KEYS: frozenset[str] = frozenset(
+    {
+        # credentials
+        "authorization",
+        "access_token",
+        "refresh_token",
+        "token",
+        "id_token",
+        "password",
+        "secret",
+        "api_key",
+        "apikey",
+        "jwt",
+        "purchase_token",
+        "refresh_token_hash",
+        # personal identifiers
+        "phone",
+        "phone_number",
+        "mobile",
+        "id_card",
+        "union_id",
+        "unionid",
+        "open_id",
+        "openid",
+        "device_fp",
+        "client_ip",
+        "nickname",
+        # face / image material (架构文档 §11.2, CR-02)
+        "face_image",
+        "face_image_key",
+        "face_key",
+        "avatar_key",
+        "image_key",
+        "image_url",
+        "photo_url",
+        "cos_key",
+        "presigned_url",
+        # diagnosis output is sensitive derived data
+        "diagnosis",
+        "dimensions",
+        "derived_features",
+    }
+)
+
+_PHONE_RE = re.compile(r"(?<!\d)(1[3-9]\d)\d{4}(\d{4})(?!\d)")
+_SK_RE = re.compile(r"\bsk-[A-Za-z0-9]{16,}\b")
+_BEARER_RE = re.compile(r"(?i)\bbearer\s+[A-Za-z0-9._\-]{8,}")
+
+
+def _scrub_text(value: str) -> str:
+    value = _SK_RE.sub("sk-" + REDACTED, value)
+    value = _BEARER_RE.sub("Bearer " + REDACTED, value)
+    return _PHONE_RE.sub(r"\1****\2", value)
+
+
+def _redact_value(value: Any, sensitive_keys: frozenset[str], depth: int = 0) -> Any:
+    if depth > 6:
+        return value
+    if isinstance(value, str):
+        return _scrub_text(value)
+    if isinstance(value, dict):
+        return {
+            k: (
+                REDACTED
+                if isinstance(k, str) and k.lower() in sensitive_keys
+                else _redact_value(v, sensitive_keys, depth + 1)
+            )
+            for k, v in value.items()
+        }
+    if isinstance(value, (list, tuple)):
+        return type(value)(_redact_value(v, sensitive_keys, depth + 1) for v in value)
+    return value
+
+
+def make_redaction_processor(
+    extra_keys: Iterable[str] = (),
+) -> Any:
+    sensitive = SENSITIVE_KEYS | {k.lower() for k in extra_keys}
+
+    def processor(
+        _logger: Any, _name: str, event_dict: MutableMapping[str, Any]
+    ) -> MutableMapping[str, Any]:
+        for key in list(event_dict):
+            if isinstance(key, str) and key.lower() in sensitive:
+                event_dict[key] = REDACTED
+            else:
+                event_dict[key] = _redact_value(event_dict[key], sensitive)
+        return event_dict
+
+    return processor
+
+
+def _context_processor(
+    _logger: Any, _name: str, event_dict: MutableMapping[str, Any]
+) -> MutableMapping[str, Any]:
+    for key, value in context_snapshot().items():
+        event_dict.setdefault(key, value)
+    return event_dict
+
+
+def configure_logging(
+    *,
+    level: str = "INFO",
+    json_output: bool = True,
+    extra_redact_keys: Iterable[str] = (),
+) -> None:
+    """Idempotently configure stdlib logging + structlog."""
+    logging.basicConfig(
+        format="%(message)s",
+        stream=sys.stdout,
+        level=getattr(logging, level.upper(), logging.INFO),
+        force=True,
+    )
+    for noisy in ("uvicorn.access", "httpx", "httpcore"):
+        logging.getLogger(noisy).setLevel(logging.WARNING)
+
+    renderer: Any = (
+        structlog.processors.JSONRenderer(ensure_ascii=False)
+        if json_output
+        else structlog.dev.ConsoleRenderer(colors=False)
+    )
+
+    structlog.configure(
+        processors=[
+            structlog.contextvars.merge_contextvars,
+            structlog.processors.add_log_level,
+            structlog.processors.TimeStamper(fmt="iso", utc=True),
+            _context_processor,
+            make_redaction_processor(extra_redact_keys),
+            structlog.processors.StackInfoRenderer(),
+            structlog.processors.format_exc_info,
+            renderer,
+        ],
+        wrapper_class=structlog.make_filtering_bound_logger(
+            getattr(logging, level.upper(), logging.INFO)
+        ),
+        logger_factory=structlog.PrintLoggerFactory(file=sys.stdout),
+        cache_logger_on_first_use=True,
+    )
+
+
+def get_logger(name: str | None = None) -> Any:
+    return structlog.get_logger(name)
diff --git a/packages/shared/middleware.py b/packages/shared/middleware.py
new file mode 100755
index 0000000..2453885
--- /dev/null
+++ b/packages/shared/middleware.py
@@ -0,0 +1,74 @@
+"""ASGI middleware: request id, access log, metrics, /v1 path alias."""
+
+from __future__ import annotations
+
+import time
+from collections.abc import Awaitable, Callable
+
+from fastapi import FastAPI, Request, Response
+from starlette.middleware.base import BaseHTTPMiddleware
+
+from packages.shared.context import bind_context, new_request_id, reset_context
+from packages.shared.logging import get_logger
+from packages.shared.observability import HTTP_LATENCY, HTTP_REQUESTS, current_trace_id
+
+REQUEST_ID_HEADER = "X-Request-Id"
+
+logger = get_logger(__name__)
+
+
+class RequestContextMiddleware(BaseHTTPMiddleware):
+    """Bind request/trace ids, emit the access log, and record HTTP metrics."""
+
+    async def dispatch(
+        self, request: Request, call_next: Callable[[Request], Awaitable[Response]]
+    ) -> Response:
+        request_id = request.headers.get(REQUEST_ID_HEADER) or new_request_id()
+        tokens = bind_context(request_id=request_id, trace_id=current_trace_id())
+        started = time.perf_counter()
+        status_code = 500
+        try:
+            response = await call_next(request)
+            status_code = response.status_code
+            response.headers[REQUEST_ID_HEADER] = request_id
+            return response
+        finally:
+            elapsed = time.perf_counter() - started
+            route = _route_template(request)
+            HTTP_REQUESTS.labels(request.method, route, str(status_code)).inc()
+            HTTP_LATENCY.labels(request.method, route).observe(elapsed)
+            logger.info(
+                "http_request",
+                method=request.method,
+                route=route,
+                status=status_code,
+                duration_ms=round(elapsed * 1000, 2),
+            )
+            reset_context(tokens)
+
+
+def _route_template(request: Request) -> str:
+    route = request.scope.get("route")
+    path = getattr(route, "path", None)
+    return str(path) if path else request.url.path
+
+
+def install_version_alias(app: FastAPI, api_prefix: str) -> None:
+    """Accept ``/v1/...`` as an alias of ``/api/v1/...``.
+
+    架构文档 §9.1 fixes the base path at ``/api/v1``; several internal specs and
+    smoke scripts refer to ``/v1/...``. Rewriting at the ASGI scope keeps a
+    single set of route definitions and one OpenAPI document.
+    """
+    legacy_prefix = "/v1"
+    if not api_prefix.endswith(legacy_prefix):
+        return
+
+    @app.middleware("http")
+    async def _rewrite(
+        request: Request, call_next: Callable[[Request], Awaitable[Response]]
+    ) -> Response:
+        path = request.scope.get("path", "")
+        if path == legacy_prefix or path.startswith(legacy_prefix + "/"):
+            request.scope["path"] = api_prefix + path[len(legacy_prefix) :]
+        return await call_next(request)
diff --git a/packages/shared/models.py b/packages/shared/models.py
new file mode 100755
index 0000000..76defcb
--- /dev/null
+++ b/packages/shared/models.py
@@ -0,0 +1,249 @@
+"""SQLAlchemy 2.0 ORM models.
+
+Scope of this scaffold: only the tables needed to prove the vertical slice
+(架构文档 §6.2 identity / billing / task / 观测与成本). All remaining domains
+(pf_ / rp_ / kb_ / ag_ / tr_ / cp_) are intentionally absent — see
+packages/*/README.md for what still has to be built.
+
+Conventions enforced here (架构文档 §6.1):
+  1. 表名带域前缀 (`id_` / `bl_` / `tk_` / `ob_`)，为将来拆库预留边界。
+  2. 主键统一雪花 ID (bigint)，不用自增、不用 UUID。
+  3. 涉及配额的表必须有幂等键与流水。
+  4. JSONB 只用于「结构会演进且不需要索引全部字段」的场景。
+"""
+
+from __future__ import annotations
+
+from datetime import datetime
+from decimal import Decimal
+from typing import Any
+
+from sqlalchemy import (
+    BigInteger,
+    Boolean,
+    DateTime,
+    Index,
+    Integer,
+    Numeric,
+    PrimaryKeyConstraint,
+    SmallInteger,
+    String,
+    Text,
+    UniqueConstraint,
+    func,
+)
+from sqlalchemy.dialects.postgresql import JSONB
+from sqlalchemy.orm import DeclarativeBase, Mapped, mapped_column
+
+
+class Base(DeclarativeBase):
+    """Declarative base; ``Base.metadata`` is Alembic's ``target_metadata``."""
+
+
+# ---------------------------------------------------------------------------
+# identity 域
+# ---------------------------------------------------------------------------
+class User(Base):
+    __tablename__ = "id_user"
+
+    user_id: Mapped[int] = mapped_column(BigInteger, primary_key=True, autoincrement=False)
+    status: Mapped[str] = mapped_column(String(16), nullable=False, default="active")
+    nickname: Mapped[str | None] = mapped_column(String(64))
+    avatar_key: Mapped[str | None] = mapped_column(String(255))
+    register_source: Mapped[str] = mapped_column(String(32), nullable=False, default="unknown")
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+    last_active_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+    deleted_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+
+    __table_args__ = (Index("ix_id_user_status", "status"),)
+
+
+class UserIdentity(Base):
+    """External identity mapping. UNIQUE(provider, union_id) per 架构文档 §6.2."""
+
+    __tablename__ = "id_user_identity"
+
+    identity_id: Mapped[int] = mapped_column(BigInteger, primary_key=True, autoincrement=False)
+    user_id: Mapped[int] = mapped_column(BigInteger, nullable=False)
+    provider: Mapped[str] = mapped_column(String(32), nullable=False)
+    union_id: Mapped[str] = mapped_column(String(128), nullable=False)
+    open_id: Mapped[str | None] = mapped_column(String(128))
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+
+    __table_args__ = (
+        UniqueConstraint("provider", "union_id", name="uq_id_user_identity_provider_union"),
+        Index("ix_id_user_identity_user", "user_id"),
+    )
+
+
+class Consent(Base):
+    """CR-01 留证：人脸的单独同意必须独立成行，不得混入总协议。"""
+
+    __tablename__ = "id_consent"
+
+    consent_id: Mapped[int] = mapped_column(BigInteger, primary_key=True, autoincrement=False)
+    user_id: Mapped[int] = mapped_column(BigInteger, nullable=False)
+    consent_type: Mapped[str] = mapped_column(String(32), nullable=False)
+    doc_version: Mapped[str] = mapped_column(String(32), nullable=False)
+    granted_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+    revoked_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+    client_ip: Mapped[str | None] = mapped_column(String(64))
+    ua: Mapped[str | None] = mapped_column(String(512))
+
+    __table_args__ = (Index("ix_id_consent_user_type", "user_id", "consent_type"),)
+
+
+# ---------------------------------------------------------------------------
+# billing 域（V1：配额与计量，无支付）
+# ---------------------------------------------------------------------------
+class QuotaRule(Base):
+    """FR-BILL-04 配额规则配置化：改配置不发版。"""
+
+    __tablename__ = "bl_quota_rule"
+
+    quota_key: Mapped[str] = mapped_column(String(64), primary_key=True)
+    period_type: Mapped[str] = mapped_column(String(16), nullable=False)  # day/month/lifetime
+    base_limit: Mapped[int] = mapped_column(Integer, nullable=False)
+    task_bonus_max: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    hard_cap: Mapped[int] = mapped_column(Integer, nullable=False)
+    description: Mapped[str | None] = mapped_column(String(255))
+    updated_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+
+
+class Quota(Base):
+    """周期配额计量：按 ``period_key`` 自然重置，无需定时任务清零。"""
+
+    __tablename__ = "bl_quota"
+
+    user_id: Mapped[int] = mapped_column(BigInteger, nullable=False)
+    quota_key: Mapped[str] = mapped_column(String(64), nullable=False)
+    period_key: Mapped[str] = mapped_column(String(32), nullable=False)
+    used: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    limit_value: Mapped[int] = mapped_column(Integer, nullable=False)
+    bonus_value: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    updated_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+
+    __table_args__ = (
+        PrimaryKeyConstraint("user_id", "quota_key", "period_key", name="pk_bl_quota"),
+    )
+
+
+class Ledger(Base):
+    """流水是真相，余额是缓存 (架构文档 §6.1 原则 4 / §6.2 bl_ledger).
+
+    ``tx_type`` 在 V1 只用 system_grant / consume / refund；V2 增加
+    purchase / gift —— 这是 V1→V2 免数据迁移的关键，不要拆表。
+    ``status`` 承载 §7.4 的预扣三段式：reserved → committed / refunded。
+    """
+
+    __tablename__ = "bl_ledger"
+
+    tx_id: Mapped[int] = mapped_column(BigInteger, primary_key=True, autoincrement=False)
+    user_id: Mapped[int] = mapped_column(BigInteger, nullable=False)
+    delta: Mapped[int] = mapped_column(Integer, nullable=False)
+    tx_type: Mapped[str] = mapped_column(String(24), nullable=False)
+    status: Mapped[str] = mapped_column(String(16), nullable=False, default="committed")
+    reason: Mapped[str | None] = mapped_column(String(128))
+    ref_type: Mapped[str | None] = mapped_column(String(32))
+    ref_id: Mapped[str | None] = mapped_column(String(64))
+    quota_key: Mapped[str | None] = mapped_column(String(64))
+    period_key: Mapped[str | None] = mapped_column(String(32))
+    balance_after: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    idempotency_key: Mapped[str] = mapped_column(String(128), nullable=False)
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+
+    __table_args__ = (
+        UniqueConstraint("idempotency_key", name="uq_bl_ledger_idempotency_key"),
+        Index("ix_bl_ledger_user_created", "user_id", "created_at"),
+    )
+
+
+# ---------------------------------------------------------------------------
+# task 域（统一任务中心）
+# ---------------------------------------------------------------------------
+class Task(Base):
+    """全系统唯一的长耗时任务视图，前端任务中心直接读它 (架构文档 §6.2)。"""
+
+    __tablename__ = "tk_task"
+
+    task_id: Mapped[int] = mapped_column(BigInteger, primary_key=True, autoincrement=False)
+    user_id: Mapped[int] = mapped_column(BigInteger, nullable=False)
+    task_type: Mapped[str] = mapped_column(String(32), nullable=False)
+    status: Mapped[str] = mapped_column(String(16), nullable=False)
+    progress: Mapped[int] = mapped_column(SmallInteger, nullable=False, default=0)
+    stage_text: Mapped[str | None] = mapped_column(String(128))
+    idempotency_key: Mapped[str] = mapped_column(String(128), nullable=False)
+    payload: Mapped[dict[str, Any]] = mapped_column(JSONB, nullable=False, default=dict)
+    result: Mapped[dict[str, Any] | None] = mapped_column(JSONB)
+    error_code: Mapped[int | None] = mapped_column(Integer)
+    error_message: Mapped[str | None] = mapped_column(String(512))
+    retry_count: Mapped[int] = mapped_column(SmallInteger, nullable=False, default=0)
+    points_reserved: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    reserve_tx_id: Mapped[int | None] = mapped_column(BigInteger)
+    submitted_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+    started_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+    finished_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+    eta_seconds: Mapped[int | None] = mapped_column(Integer)
+
+    __table_args__ = (
+        UniqueConstraint("idempotency_key", name="uq_tk_task_idempotency_key"),
+        Index("ix_tk_task_user_status", "user_id", "status"),
+    )
+
+
+# ---------------------------------------------------------------------------
+# 观测与成本
+# ---------------------------------------------------------------------------
+class AiCall(Base):
+    """成本归因的唯一数据源 (架构文档 §13, FR-BILL-08).
+
+    **不允许「先上线后补埋点」** —— V1 不收费，成本是唯一财务变量，且这张表
+    是 V2 定价的唯一数据来源。每一次 provider 调用（含失败与结构化重试）都写
+    一行。生产环境按月分区。
+    """
+
+    __tablename__ = "ob_ai_call"
+
+    call_id: Mapped[int] = mapped_column(BigInteger, primary_key=True, autoincrement=False)
+    user_id: Mapped[int | None] = mapped_column(BigInteger)
+    scene: Mapped[str] = mapped_column(String(48), nullable=False)
+    provider: Mapped[str] = mapped_column(String(32), nullable=False)
+    model: Mapped[str] = mapped_column(String(64), nullable=False)
+    # Prompt 版本：架构文档 §8.1「可复现」要求记录，§6.2 的列清单漏了，这里补上。
+    prompt_id: Mapped[str | None] = mapped_column(String(64))
+    prompt_version: Mapped[str | None] = mapped_column(String(16))
+    tokens_in: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    tokens_out: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    tokens_cached: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    latency_ms: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    cost_cents: Mapped[Decimal] = mapped_column(
+        Numeric(14, 6), nullable=False, default=Decimal("0")
+    )
+    success: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
+    error_code: Mapped[str | None] = mapped_column(String(64))
+    error_message: Mapped[str | None] = mapped_column(Text)
+    ref_type: Mapped[str | None] = mapped_column(String(32))
+    ref_id: Mapped[str | None] = mapped_column(String(64))
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+
+    __table_args__ = (
+        Index("ix_ob_ai_call_created", "created_at"),
+        Index("ix_ob_ai_call_scene_created", "scene", "created_at"),
+        Index("ix_ob_ai_call_user_created", "user_id", "created_at"),
+    )
diff --git a/packages/shared/observability.py b/packages/shared/observability.py
new file mode 100755
index 0000000..578985a
--- /dev/null
+++ b/packages/shared/observability.py
@@ -0,0 +1,116 @@
+"""OpenTelemetry tracing + Prometheus metrics (架构文档 §15, TD-24).
+
+The three dashboards required by §15 (业务漏斗 / AI 成本 / 任务健康) are built
+from the metrics declared here plus the ``ob_ai_call`` table.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+from prometheus_client import CollectorRegistry, Counter, Histogram
+
+if TYPE_CHECKING:  # pragma: no cover
+    from fastapi import FastAPI
+
+    from packages.shared.config import Settings
+
+REGISTRY = CollectorRegistry(auto_describe=True)
+
+HTTP_REQUESTS = Counter(
+    "meida_http_requests_total",
+    "HTTP requests handled by the API",
+    labelnames=("method", "route", "status"),
+    registry=REGISTRY,
+)
+HTTP_LATENCY = Histogram(
+    "meida_http_request_duration_seconds",
+    "HTTP request latency",
+    labelnames=("method", "route"),
+    buckets=(0.01, 0.05, 0.1, 0.2, 0.3, 0.5, 1.0, 2.0, 5.0, 8.0, 15.0),
+    registry=REGISTRY,
+)
+
+# --- AI 成本看板 ---------------------------------------------------------
+LLM_CALLS = Counter(
+    "meida_llm_calls_total",
+    "LLM provider calls (one per HTTP attempt, including retries and failures)",
+    labelnames=("scene", "provider", "model", "success"),
+    registry=REGISTRY,
+)
+LLM_TOKENS = Counter(
+    "meida_llm_tokens_total",
+    "LLM tokens by kind; `cached` drives the prompt-cache hit-rate panel (NFR-COST-03)",
+    labelnames=("scene", "provider", "model", "kind"),
+    registry=REGISTRY,
+)
+LLM_COST_CENTS = Counter(
+    "meida_llm_cost_cents_total",
+    "Accumulated LLM cost in cents, attributable to scene/provider/model",
+    labelnames=("scene", "provider", "model"),
+    registry=REGISTRY,
+)
+LLM_LATENCY = Histogram(
+    "meida_llm_latency_seconds",
+    "LLM provider call latency",
+    labelnames=("scene", "provider", "model"),
+    buckets=(0.2, 0.5, 1.0, 2.0, 4.0, 8.0, 15.0, 30.0, 60.0),
+    registry=REGISTRY,
+)
+
+# --- 业务漏斗 / 任务健康 -------------------------------------------------
+QUOTA_EVENTS = Counter(
+    "meida_quota_events_total",
+    "Quota three-phase events (reserve/commit/refund) and their outcome",
+    labelnames=("quota_key", "phase", "result"),
+    registry=REGISTRY,
+)
+TASK_TRANSITIONS = Counter(
+    "meida_task_transitions_total",
+    "tk_task state machine transitions",
+    labelnames=("task_type", "status"),
+    registry=REGISTRY,
+)
+
+
+def setup_tracing(app: FastAPI, settings: Settings) -> None:
+    """Install the OTel tracer provider and instrument FastAPI.
+
+    Disabled by default locally; enable with ``OTEL_ENABLED=true`` and an
+    OTLP endpoint.
+    """
+    if not settings.otel_enabled:
+        return
+
+    from opentelemetry import trace
+    from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
+    from opentelemetry.instrumentation.fastapi import FastAPIInstrumentor
+    from opentelemetry.sdk.resources import Resource
+    from opentelemetry.sdk.trace import TracerProvider
+    from opentelemetry.sdk.trace.export import BatchSpanProcessor
+
+    resource = Resource.create(
+        {"service.name": settings.otel_service_name, "deployment.environment": settings.app_env}
+    )
+    provider = TracerProvider(resource=resource)
+    if settings.otel_exporter_otlp_endpoint:
+        provider.add_span_processor(
+            BatchSpanProcessor(
+                OTLPSpanExporter(endpoint=f"{settings.otel_exporter_otlp_endpoint}/v1/traces")
+            )
+        )
+    trace.set_tracer_provider(provider)
+    FastAPIInstrumentor.instrument_app(app, tracer_provider=provider)
+
+
+def current_trace_id() -> str | None:
+    """Return the active W3C trace id, if any."""
+    try:
+        from opentelemetry import trace
+    except ImportError:  # pragma: no cover
+        return None
+    span = trace.get_current_span()
+    ctx: Any = span.get_span_context()
+    if not getattr(ctx, "is_valid", False):
+        return None
+    return format(ctx.trace_id, "032x")
diff --git a/packages/shared/redis_client.py b/packages/shared/redis_client.py
new file mode 100755
index 0000000..76ab303
--- /dev/null
+++ b/packages/shared/redis_client.py
@@ -0,0 +1,28 @@
+"""Shared Redis connection (cache / quota counters / cost budget counter)."""
+
+from __future__ import annotations
+
+from redis.asyncio import Redis, from_url
+
+from packages.shared.config import Settings, get_settings
+
+_client: Redis | None = None
+
+
+def create_redis(settings: Settings | None = None) -> Redis:
+    cfg = settings or get_settings()
+    return from_url(cfg.redis_url, encoding="utf-8", decode_responses=True)
+
+
+def get_redis() -> Redis:
+    global _client
+    if _client is None:
+        _client = create_redis()
+    return _client
+
+
+async def close_redis() -> None:
+    global _client
+    if _client is not None:
+        await _client.aclose()
+    _client = None
diff --git a/packages/shared/response.py b/packages/shared/response.py
new file mode 100755
index 0000000..34a194a
--- /dev/null
+++ b/packages/shared/response.py
@@ -0,0 +1,49 @@
+"""Unified response envelope (架构文档 §9.2).
+
+    {"code": 0, "message": "ok", "data": {...}, "request_id": "..."}
+    {"code": 40301, "message": "今日额度已用完", "data": {...},
+     "request_id": "...", "hint": "明天恢复，或完成任务多换 3 次"}
+
+Every endpoint returns this shape. Handlers in :mod:`packages.shared.handlers`
+guarantee it even for unhandled exceptions.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from pydantic import BaseModel, Field
+
+from packages.shared.context import get_request_id
+from packages.shared.errors import ERROR_META, ErrorCode
+
+
+class Envelope(BaseModel):
+    code: int = Field(default=int(ErrorCode.OK), description="0 表示成功；4/5/6 开头为错误段")
+    message: str = Field(default="ok", description="用户可读文案")
+    data: Any = Field(default=None)
+    request_id: str | None = Field(default=None)
+    hint: str | None = Field(default=None, description="告诉用户下一步怎么办")
+
+
+def ok(data: Any = None) -> dict[str, Any]:
+    return Envelope(
+        code=int(ErrorCode.OK), message="ok", data=data, request_id=get_request_id()
+    ).model_dump(exclude_none=True)
+
+
+def fail(
+    code: ErrorCode,
+    *,
+    message: str | None = None,
+    hint: str | None = None,
+    data: Any = None,
+) -> dict[str, Any]:
+    meta = ERROR_META.get(code, ERROR_META[ErrorCode.INTERNAL_ERROR])
+    return Envelope(
+        code=int(code),
+        message=message or meta.message,
+        data=data,
+        request_id=get_request_id(),
+        hint=hint if hint is not None else meta.hint,
+    ).model_dump(exclude_none=True)
diff --git a/packages/shared/security.py b/packages/shared/security.py
new file mode 100755
index 0000000..8d29285
--- /dev/null
+++ b/packages/shared/security.py
@@ -0,0 +1,76 @@
+"""JWT access tokens (架构文档 §9.1: Bearer JWT, 15 分钟).
+
+The revocable refresh-token half (`id_session`) is FR-ACC-05 and is deliberately
+NOT implemented in this scaffold — see packages/identity/README.md.
+"""
+
+from __future__ import annotations
+
+import time
+from dataclasses import dataclass
+from typing import Any
+
+import jwt
+
+from packages.shared.errors import AppError, ErrorCode
+
+
+@dataclass(frozen=True, slots=True)
+class AccessTokenClaims:
+    user_id: int
+    issued_at: int
+    expires_at: int
+    issuer: str
+
+
+def create_access_token(
+    *,
+    user_id: int,
+    secret: str,
+    algorithm: str = "HS256",
+    ttl_seconds: int = 900,
+    issuer: str = "meida",
+    now: int | None = None,
+) -> str:
+    issued_at = int(now if now is not None else time.time())
+    payload: dict[str, Any] = {
+        "sub": str(user_id),
+        "iat": issued_at,
+        "exp": issued_at + ttl_seconds,
+        "iss": issuer,
+        "typ": "access",
+    }
+    return jwt.encode(payload, secret, algorithm=algorithm)
+
+
+def decode_access_token(
+    token: str,
+    *,
+    secret: str,
+    algorithm: str = "HS256",
+    issuer: str = "meida",
+) -> AccessTokenClaims:
+    try:
+        payload = jwt.decode(
+            token,
+            secret,
+            algorithms=[algorithm],
+            issuer=issuer,
+            options={"require": ["exp", "iat", "sub"]},
+        )
+    except jwt.ExpiredSignatureError as exc:
+        raise AppError(ErrorCode.TOKEN_EXPIRED) from exc
+    except jwt.InvalidTokenError as exc:
+        raise AppError(ErrorCode.TOKEN_INVALID) from exc
+
+    try:
+        user_id = int(payload["sub"])
+    except (KeyError, TypeError, ValueError) as exc:
+        raise AppError(ErrorCode.TOKEN_INVALID) from exc
+
+    return AccessTokenClaims(
+        user_id=user_id,
+        issued_at=int(payload["iat"]),
+        expires_at=int(payload["exp"]),
+        issuer=str(payload.get("iss", issuer)),
+    )