# Backend Development — Rules & Guidelines

Rules and conventions every backend developer must follow. Read this before writing your first line of code.

---

## 1. Language & Typing

- **Python 3.12+** is the minimum version.
- **Every** function, method, and variable must have explicit type annotations — no exceptions.
- Use `str`, `int`, `float`, `bool`, `None` for primitives.
- Use `list[T]`, `dict[K, V]`, `set[T]`, `tuple[T, ...]` (lowercase, built-in generics) — never `typing.List`, `typing.Dict`, etc.
- Use `T | None` instead of `Optional[T]`.
- Use `TypeAlias`, `TypeVar`, `Protocol`, and `NewType` when they improve clarity.
- Return types are **mandatory** — including `-> None`.
- Never use `Any` unless there is no other option and a comment explains why.
- Run `mypy --strict` (or `pyright` in strict mode) — the codebase must pass with zero errors.

```python
# Good
def get_jail_by_name(name: str) -> Jail | None:
    ...

# Bad — missing types
def get_jail_by_name(name):
    ...
```

---

## 2. Core Libraries

| Purpose | Library | Notes |
|---|---|---|
| Web framework | **FastAPI** | Async endpoints only. |
| Data validation & settings | **Pydantic v2** | All request/response bodies and config models. |
| Async HTTP client | **aiohttp** (`ClientSession`) | For external calls (blocklists, IP lookups). |
| Scheduling | **APScheduler 4.x** (async) | Blocklist imports, periodic health checks. |
| Structured logging | **structlog** | Every log call must use structlog — never `print()` or `logging` directly. |
| Database | **aiosqlite** | Async SQLite access for the application database. |
| Testing | **pytest** + **pytest-asyncio** + **httpx** (`AsyncClient`) | Every feature needs tests. |
| Mocking | **unittest.mock** / **pytest-mock** | Isolate external dependencies. |
| Date & time | **datetime** (stdlib) — always timezone-aware | Use `datetime.datetime.now(datetime.UTC)`. Never naive datetimes. |
| IP / Network | **ipaddress** (stdlib) | Validate and normalise IPs and CIDR ranges. |
| Environment / config | **pydantic-settings** | Load `.env` and environment variables into typed models. |
| fail2ban integration | **fail2ban client** (bundled) | Use the local copy at [`./fail2ban-master`](../fail2ban-master). Import from [`./fail2ban-master/fail2ban/client`](../fail2ban-master/fail2ban/client) to communicate with the fail2ban socket. Do **not** install fail2ban as a pip package. |

### fail2ban Client Usage

The repository ships with a vendored copy of fail2ban located at `./fail2ban-master`.
All communication with the fail2ban daemon must go through the client classes found in `./fail2ban-master/fail2ban/client`.
Add the project root to `sys.path` (or configure it in `pyproject.toml` as a path dependency) so that `from fail2ban.client ...` resolves to the bundled copy.

```python
import sys
from pathlib import Path

# Ensure the bundled fail2ban is importable
sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "fail2ban-master"))

from fail2ban.client.csocket import CSSocket  # noqa: E402
```

### Libraries you must NOT use

- `requests` — use `aiohttp` (async).
- `flask` — we use FastAPI.
- `celery` — we use APScheduler.
- `print()` for logging — use `structlog`.
- `json.loads` / `json.dumps` on Pydantic models — use `.model_dump()` / `.model_validate()`.

### Timestamp Handling

Timestamp consistency is critical for accurate ban history queries across the dashboard and history endpoints. Follow these rules:

**Rule 1: Use consistent UTC timestamps**
- All timestamps in the database are stored as Unix epochs (seconds since 1970-01-01 UTC).
- fail2ban stores timestamps using `time.time()`, which is always UTC epoch seconds.
- When querying fail2ban's SQLite database by timestamp, use `app.utils.time_utils.since_unix()` (not manual datetime calculations).

**Rule 2: Time-range windows include a 60-second slack**
- The `since_unix()` function includes a 60-second slack window (`TIME_RANGE_SLACK_SECONDS` in `app.utils.constants`).
- This slack accommodates:
  - Clock drift between the local system and fail2ban.
  - Test seeding delays when timestamps are manually set to exact boundaries.
- The slack ensures that dashboard and history queries return consistent row counts for the same time range.

**Rule 3: Never duplicate timestamp calculation logic**
- All services that query by time range must import and use `since_unix()`.
- Do not recalculate timestamps locally using `datetime` or `time` modules in service code.
- If you need a timestamp for a time range, use `since_unix()`.

**Example:**
```python
from app.utils.time_utils import since_unix

# Get all bans from the last 24 hours (with 60-second slack)
since_ts: int = since_unix("24h")
rows = await db.execute(
    "SELECT * FROM bans WHERE timeofban >= ?",
    (since_ts,)
)
```

---

## 3. Project Structure

```
backend/
├── app/
│   ├── __init__.py
│   ├── main.py              # FastAPI app factory, lifespan
│   ├── config.py            # Pydantic settings
│   ├── dependencies.py      # FastAPI dependency providers
│   ├── models/              # Pydantic schemas (request, response, domain)
│   ├── routers/             # FastAPI routers grouped by feature
│   ├── services/            # Business logic — one service per domain
│   ├── repositories/        # Database access layer
│   ├── tasks/               # APScheduler jobs
│   └── utils/               # Helpers, constants, shared types
├── tests/
│   ├── conftest.py
│   ├── test_routers/
│   ├── test_services/
│   └── test_repositories/
├── pyproject.toml
└── .env.example
```

- **Routers** receive requests, validate input via Pydantic, and delegate to **services**.
- **Services** contain business logic and call **repositories** or external clients.
- **Repositories** handle raw database queries — nothing else.
- Never put business logic inside routers or repositories.

### Service Dependencies and Injection

Services should **never** directly import other services to avoid hidden coupling and make testing harder. Instead:

1. **Define clear service interfaces** using Protocol classes in `app/services/protocols.py`.
2. **Make dependencies explicit** by passing them as function parameters with optional defaults.
3. **Use lazy imports** for fallback singletons (not at module level).
4. **Inject services via FastAPI dependencies** when called from routers.

**Example:** The `history_service` depends on `Fail2BanMetadataService` to resolve the fail2ban database path:

```python
# Good — dependency passed as parameter
async def list_history(
    socket_path: str,
    fail2ban_metadata_service: Fail2BanMetadataService | None = None,
) -> HistoryListResponse:
    if fail2ban_metadata_service is None:
        # Lazy import fallback for backward compatibility
        from app.services.fail2ban_metadata_service import default_fail2ban_metadata_service
        fail2ban_metadata_service = default_fail2ban_metadata_service
    ...
```

Routers inject the service dependency explicitly:

```python
from app.dependencies import Fail2BanMetadataServiceDep

@router.get("/api/history")
async def get_history(
    fail2ban_metadata_service: Fail2BanMetadataServiceDep,
) -> HistoryListResponse:
    return await history_service.list_history(
        socket_path,
        fail2ban_metadata_service=fail2ban_metadata_service,
    )
```

This pattern prevents circular imports, makes services testable, and allows easy mocking in tests.

### Mutable Runtime State

All mutable runtime state (state that changes during the application's lifetime) **must** be stored in `RuntimeState` defined in `app/utils/runtime_state.py`. This centralizes state management, prevents accidental global mutable variables, and makes state management testable and synchronization-safe.

**Allowed locations for mutable state:**

1. **RuntimeState fields** — Core application state (e.g., `server_status`, `last_activation`, `pending_recovery`, `runtime_settings`). Managed through dedicated functions (e.g., `record_activation()`, `clear_pending_recovery()`).
2. **Nested service state** — Service-specific mutable state (e.g., `JailServiceState` for jail capability detection cache) is nested within `RuntimeState` as a field. Services receive their state via dependency injection.
3. **Controlled via dependencies** — State is injected into services and routers using FastAPI `Depends()`. This ensures single-source-of-truth and testability.

**Example — jail_service state management:**

```python
# Define service-specific state (in app/utils/runtime_state.py)
@dataclass
class JailServiceState:
    backend_cmd_supported: bool | None = None
    backend_cmd_lock: asyncio.Lock | None = None

# Nested in RuntimeState
@dataclass
class RuntimeState:
    jail_service_state: JailServiceState = field(default_factory=JailServiceState)
    ...

# Injected into services via dependency
async def list_jails(socket_path: str, state: JailServiceState) -> JailListResponse:
    backend_cmd_is_supported = await _check_backend_cmd_supported(client, name, state)
    ...

# Routers inject state through FastAPI dependencies
@router.get("/api/jails")
async def get_jails(state: JailServiceStateDep) -> JailListResponse:
    return await jail_service.list_jails(socket_path, state)
```

**Why:** Centralizing mutable state prevents race conditions, makes concurrency boundaries explicit, simplifies testing (each test gets a fresh state object), and prepares for multi-worker deployments (shared state would need to be extracted to Redis, database, or shared memory).

---

## 4. FastAPI Conventions

- Use **async def** for every endpoint — no sync endpoints.
- Every endpoint must declare explicit **response models** (`response_model=...`).
- Use **Pydantic models** for request bodies and query parameters — never raw dicts.
- Use **Depends()** for dependency injection (database sessions, services, auth).
- Group endpoints into routers by feature domain (`routers/jails.py`, `routers/bans.py`, …).
- Use appropriate HTTP status codes: `201` for creation, `204` for deletion with no body, `404` for not found, etc.
- Protected endpoints should return `401 Unauthorized` or `403 Forbidden` when the session is invalid or expired; the frontend treats these responses as a session-expiry event and redirects the user to `/login`.
- Use **HTTPException** or custom exception handlers — never return error dicts manually.
- **GET endpoints are read-only — never call `db.commit()` or execute INSERT/UPDATE/DELETE inside a GET handler.** If a GET path produces side-effects (e.g., caching resolved data), that write belongs in a background task, a scheduled flush, or a separate POST endpoint. Users and HTTP caches assume GET is idempotent and non-mutating.

    ```python
    # Good — pass db=None on GET so geo_service never commits
    result = await geo_service.lookup_batch(ips, http_session, db=None)

    # Bad — triggers INSERT + COMMIT per IP inside a GET handler
    result = await geo_service.lookup_batch(ips, http_session, db=app_db)
    ```

```python
from fastapi import APIRouter, Depends, HTTPException, status
from app.models.jail import JailResponse, JailListResponse
from app.services.jail_service import JailService

router: APIRouter = APIRouter(prefix="/api/jails", tags=["Jails"])

@router.get("/", response_model=JailListResponse)
async def list_jails(service: JailService = Depends()) -> JailListResponse:
    jails: list[JailResponse] = await service.get_all_jails()
    return JailListResponse(jails=jails)
```

### Dependency Layering: Enforcing the Repository Boundary

The **repository boundary** separates database-aware code from application logic. This is enforced through dependency injection:

| Layer | Responsibilities | Dependencies |
|---|---|---|
| **Routers** | Receive requests, validate input, return responses. | Service context dependencies (SessionServiceContextDep, BlocklistServiceContextDep), settings, auth. Never raw database connections. |
| **Services** | Contain business logic, orchestrate operations. | Other services, repositories. May receive `aiosqlite.Connection` for repository operations. |
| **Repositories** | Execute all SQL queries. All database knowledge lives here. | `aiosqlite.Connection` (from callers). |

**Rule: Routers must NOT depend on `DbDep` (raw database connections).**

Instead, routers should:
1. Depend on **service context dependencies** like `SessionServiceContextDep`, `BlocklistServiceContextDep`, etc.
2. These context dependencies combine the database connection and related repositories.
3. Pass the context to services, which internally orchestrate database operations.

**Service Context Dependencies Available:**
- `SessionServiceContextDep` — Contains `db` and `session_repo` for session operations.
- `BlocklistServiceContextDep` — Contains `db`, `blocklist_repo`, `import_log_repo`, `settings_repo`.
- `SettingsServiceContextDep` — Contains `db` and `settings_repo`.
- `BanServiceContextDep` — Contains `db` and `fail2ban_db_repo`.
- `HistoryServiceContextDep` — Contains `db`, `fail2ban_db_repo`, `history_archive_repo`.

**Why:**
- **Enforcement**: Not exporting `DbDep` from the dependencies module makes it impossible for routers to accidentally bypass repositories.
- **Clarity**: Service context dependencies explicitly declare which database operations a router needs.
- **Testability**: Services and routers are easier to test when they depend on repositories (which can be mocked) rather than raw connections.

**Example:**
```python
# ✅ GOOD — router depends on service context
@router.post("/login")
async def login(
    body: LoginRequest,
    response: Response,
    session_ctx: SessionServiceContextDep,  # Contains db + session_repo
    _auth: AuthDep,
) -> LoginResponse:
    return await auth_service.login(
        session_ctx.db,
        password=body.password,
        session_repo=session_ctx.session_repo,
        ...
    )

# ❌ BAD — router depends on raw db (DbDep is not exported for this reason)
@router.post("/login")
async def login(
    body: LoginRequest,
    db: DbDep,  # ← Cannot import DbDep in routers
    _auth: AuthDep,
) -> LoginResponse:
    return await auth_service.login(db, password=body.password, ...)
```

**DEPRECATED: DbDep**
- The `DbDep` type alias is provided for backward compatibility only.
- DO NOT use in new code. Use service context dependencies instead.
- See `backend/app/dependencies.py` for available service contexts.

---

## 5. Pydantic Models

- Every model inherits from `pydantic.BaseModel`.
- Use `model_config = ConfigDict(strict=True)` where appropriate.
- Field names use **snake_case** in Python, export as **camelCase** to the frontend via alias generators if needed.
- Validate at the boundary — once data enters a Pydantic model it is trusted.
- Use `Field(...)` with descriptions for every field to keep auto-generated docs useful.
- Separate **request models**, **response models**, and **domain (internal) models** — do not reuse one model for all three.

```python
from pydantic import BaseModel, Field
from datetime import datetime

class BanResponse(BaseModel):
    ip: str = Field(..., description="Banned IP address")
    jail: str = Field(..., description="Jail that issued the ban")
    banned_at: datetime = Field(..., description="UTC timestamp of the ban")
    expires_at: datetime | None = Field(None, description="UTC expiry, None if permanent")
    ban_count: int = Field(..., ge=1, description="Number of times this IP was banned")
```

### Using `Literal` Types for Constrained Strings

When a field should only accept a small set of predefined values, use `Literal` to enforce this at the type level:

```python
from typing import Literal
from pydantic import BaseModel, Field

LogLevel = Literal["CRITICAL", "ERROR", "WARNING", "NOTICE", "INFO", "DEBUG"]

class GlobalConfigUpdate(BaseModel):
    log_level: LogLevel | None = Field(
        default=None,
        description="Log level: CRITICAL, ERROR, WARNING, NOTICE, INFO, or DEBUG.",
    )
```

This provides:
- **Type safety** — IDEs and type checkers enforce valid values.
- **API documentation** — OpenAPI docs automatically list all allowed values.
- **Validation** — Pydantic rejects invalid values and provides a clear error message.

### Custom Field Validators

For fields that require complex validation (e.g., file paths that must be within allowed directories), use `@field_validator`:

```python
from pydantic import field_validator
from app.utils.path_utils import validate_log_path

class AddLogPathRequest(BaseModel):
    log_path: str = Field(..., description="Absolute path to the log file to monitor.")

    @field_validator("log_path", mode="after")
    @classmethod
    def validate_log_path_field(cls, value: str) -> str:
        """Validate that the log path is within allowed directories."""
        return validate_log_path(value)
```

**Path Validation Helper:**

For query parameters and other contexts where Pydantic validators cannot be used directly, use the `validate_log_path()` helper from `app.utils.path_utils`:

```python
from fastapi import HTTPException, status
from app.utils.path_utils import validate_log_path

@router.delete("/{name}/logpath")
async def delete_log_path(
    name: str,
    log_path: str = Query(...),
) -> None:
    try:
        validate_log_path(log_path)
    except ValueError as e:
        raise HTTPException(
            status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
            detail=str(e),
        ) from e
    # ... rest of handler
```

**Key points:**
- Use `mode="after"` in model validators to validate after Pydantic's basic type coercion.
- Raise `ValueError` if validation fails; Pydantic converts it to an HTTP 400 response.
- For query parameters that cannot use Pydantic validators, use the `validate_log_path()` helper and raise HTTP 422.
- **Never use string prefix matching** for path validation (e.g., `path.startswith("/var/log")`). The helper uses `Path.relative_to()` to prevent bypasses like `/var/log_evil/file.log`.
- Symlinks are resolved before validating to prevent symlink-based escapes.

### Model Type Usage by Layer

**Pydantic models** are mandatory for all **external-facing** data structures — anything that crosses layer boundaries or is serialized to HTTP responses. **TypedDict** may be used **only** for internal, layer-private data structures where they provide precise typing without runtime overhead.

**Rules:**

1. **Routers (HTTP boundary):** All request and response types **must be Pydantic models**. FastAPI uses these for validation, serialization, and OpenAPI documentation.
   - Use Pydantic request models for request bodies and query parameters.
   - Use Pydantic response models in the `response_model` parameter.

   ```python
   # Good — Pydantic models for router layer
   class JailStatsRequest(BaseModel):
       jail_name: str

   class JailStatsResponse(BaseModel):
       jail_name: str
       active_bans: int

   @router.post("/stats", response_model=JailStatsResponse)
   async def get_stats(req: JailStatsRequest) -> JailStatsResponse:
       ...
   ```

2. **Services (business logic):** Return types should be **Pydantic models** if the result is:
   - Returned to a router (likely — they become API responses).
   - Used across multiple services (shared interfaces).
   - Exposed to external consumers (even indirectly).

   If a service returns a purely internal intermediate result used by a single caller, TypedDict is acceptable but should be rare.

   ```python
   # Good — service returns Pydantic (may be used by multiple routers)
   async def get_jail_details(name: str) -> JailDetailResponse:
       ...

   # Acceptable — purely internal utility result
   def _parse_fail2ban_response(raw: str) -> ParsedResponse:
       """Internal helper—used only by this service."""
       ...
   ```

3. **Repositories (data access):** Return types may use **TypedDict** because they represent **raw database rows** that:
   - Are layer-private (only called by their own service).
   - Do not cross HTTP boundaries directly.
   - Benefit from lightweight typing without runtime validation.

   ```python
   # Good — TypedDict for raw repository rows
   class GeoRow(TypedDict):
       ip: str
       country_code: str | None

   async def load_all(db: aiosqlite.Connection) -> list[GeoRow]:
       ...
   ```

   If a repository result becomes part of a service's public interface (returned to routers or other services), convert it to a Pydantic model.

4. **Utilities and helpers:** Internal helper results may use TypedDict if they are not part of a public module interface.

**Migration path:** Existing internal TypedDicts (e.g., `GeoCacheRow`, `ImportLogRow`) may remain as TypedDicts so long as they stay within their layer. If a type needs to cross layer boundaries (repo → service → router), convert it to a Pydantic model incrementally as you refactor that data flow.

---

## 6. Async Rules

- **Never** call blocking / synchronous I/O in an async function — no `time.sleep()`, no synchronous file reads, no `requests.get()`.
- Use `aiohttp.ClientSession` for HTTP calls, `aiosqlite` for database access.
- Use `asyncio.TaskGroup` (Python 3.11+) when you need to run independent coroutines concurrently.
- Long-running startup/shutdown logic goes into the **FastAPI lifespan** context manager.
- **Never call `db.commit()` inside a loop.** With aiosqlite, every commit serialises through a background thread and forces an `fsync`. N rows × 1 commit = N fsyncs. Accumulate all writes in the loop, then issue a single `db.commit()` once after the loop ends. The difference between 5,000 commits and 1 commit can be seconds vs milliseconds.

    ```python
    # Good — one commit for the whole batch
    for ip, info in results.items():
        await db.execute(INSERT_SQL, (ip, info.country_code, ...))
    await db.commit()  # ← single fsync

    # Bad — one fsync per row
    for ip, info in results.items():
        await db.execute(INSERT_SQL, (ip, info.country_code, ...))
        await db.commit()  # ← fsync on every iteration
    ```
- **Prefer `executemany()` over calling `execute()` in a loop** when inserting or updating multiple rows with the same SQL template. aiosqlite passes the entire batch to SQLite in one call, reducing Python↔thread overhead on top of the single-commit saving.

    ```python
    # Good
    await db.executemany(INSERT_SQL, [(ip, cc, cn, asn, org) for ip, info in results.items()])
    await db.commit()
    ```
- Shared resources (DB connections, HTTP sessions) are created once during startup and closed during shutdown — never inside request handlers.

```python
from contextlib import asynccontextmanager
from collections.abc import AsyncGenerator
from fastapi import FastAPI
import aiohttp
import aiosqlite

@asynccontextmanager
async def lifespan(app: FastAPI) -> AsyncGenerator[None]:
    # Startup
    app.state.http_session = aiohttp.ClientSession()
    app.state.db = await aiosqlite.connect("bangui.db")
    yield
    # Shutdown
    await app.state.http_session.close()
    await app.state.db.close()
```

### Fire-and-Forget Background Tasks

When you need to spawn a background task that runs independently without waiting for the result, use `asyncio.create_task()` with the `logged_task()` helper from `app.utils.async_utils`. This ensures exceptions in background tasks are always logged and never silently discarded.

**Why this matters:** In Python 3.11+, unhandled exceptions in fire-and-forget tasks become silent `RuntimeWarning`s. Without logging, background errors (network failures, database writes, API timeouts) become invisible in structured logs and are extremely hard to debug.

**Pattern:**
```python
from app.utils.async_utils import logged_task

# Bad — exceptions are silently discarded
asyncio.create_task(some_background_work())

# Good — exceptions are logged
asyncio.create_task(
    logged_task(some_background_work(), "task_name"),
    name="task_name",
)
```

The `logged_task()` wrapper:
- Wraps your coroutine to catch any exception
- Logs the exception with `log.exception()` (structlog automatically captures the traceback)
- Adds `task_name` to the structured log context
- Never re-raises — it's safe to use with `asyncio.create_task()` 

Example:
```python
import asyncio
from app.utils.async_utils import logged_task
import structlog

log = structlog.get_logger()

async def geo_lookup_batch(ips: list[str]) -> None:
    """Look up geolocation data for IPs asynchronously."""
    try:
        for ip in ips:
            # May timeout, fail network call, or fail DB write
            location = await lookup_ip_location(ip)
            await db.execute(INSERT_GEO_SQL, (ip, location))
        await db.commit()
    except Exception:
        # All exceptions are automatically logged by logged_task() wrapper
        raise

# In your request handler or service:
asyncio.create_task(
    logged_task(geo_lookup_batch(uncached_ips), "geo_cache_batch"),
    name="geo_cache_batch",
)
```

---

## 6.1 Database Query Conventions

### LIKE Queries and Wildcard Escaping

SQLite's `LIKE` operator treats `%` (any sequence of characters) and `_` (any single character) as wildcards. When querying with user-supplied filters that may contain these characters, you must escape them to prevent unintended matches.

**The Problem:**
```python
# Bad — ip_filter="10.0.0_" matches "10.0.0.1", "10.0.0.2", etc.
ip_filter = "10.0.0_"
await db.execute(
    "SELECT * FROM bans WHERE ip LIKE ?",
    (f"{ip_filter}%",)  # ← wildcard characters not escaped
)
```

**The Solution:**

Use the `escape_like()` helper from `app.utils.fail2ban_db_utils`:

```python
from app.utils.fail2ban_db_utils import escape_like

# Good — wildcard characters are escaped
ip_filter = "10.0.0_"
await db.execute(
    "SELECT * FROM bans WHERE ip LIKE ? ESCAPE '\\'",
    (f"{escape_like(ip_filter)}%",)  # ← underscores escaped to literal
)
```

**How `escape_like()` works:**

The function escapes backslashes first, then `%` and `_` signs:
```python
def escape_like(s: str) -> str:
    return s.replace("\\", "\\\\").replace("%", "\\%").replace("_", "\\_")
```

**Key rules:**
1. **Backslash escapes first** — to prevent double-escaping when the input contains backslashes.
2. **Add `ESCAPE '\\'` to the SQL** — tells SQLite which character to use for escaping.
3. **Dots are not wildcards** — they do not need escaping; normal IP addresses pass through unchanged.

**Test example:**
```python
assert escape_like("10.0.0_") == "10.0.0\\_"
assert escape_like("10.0.0%test") == "10.0.0\\%test"
assert escape_like("10.0.0.1") == "10.0.0.1"  # Unchanged
```

---

## 6.2 Database Migrations

The application database schema is versioned and migrated automatically on startup via `app.db.init_db()`.

### Migration Design Principles

**Migrations must be atomic.** All schema changes for a single version (DDL statements) and the `schema_migrations` record insert must be wrapped in a single `BEGIN IMMEDIATE ... COMMIT` transaction. This prevents partial migrations if a process crashes mid-migration.

If a crash occurs between migration steps, the next startup will:
1. Detect the missing `schema_migrations` record.
2. Re-apply the entire migration in a single transaction (all-or-nothing).
3. Avoid data corruption or schema inconsistency.

### Writing a New Migration

1. **Add the DDL statements** to `_MIGRATIONS` dict in `app/db.py`:

```python
_MIGRATIONS: dict[int, str] = {
    1: _CREATE_INITIAL_SCHEMA,
    2: """
-- Migration 2: Add new_column to users table.
ALTER TABLE users ADD COLUMN new_column TEXT DEFAULT 'default_value';
CREATE INDEX idx_users_new_column ON users(new_column);
""",
}
```

2. **Update `_CURRENT_SCHEMA_VERSION`** to the new version number:

```python
_CURRENT_SCHEMA_VERSION: int = 2  # was 1
```

3. **Ensure idempotency where possible:**
   - Use `CREATE TABLE IF NOT EXISTS` and `CREATE INDEX IF NOT EXISTS`.
   - For `ALTER TABLE ADD COLUMN`, check if the column exists first using `PRAGMA table_info()` if re-applying the migration is a concern.

4. **Verify atomicity in tests:**

```python
async def test_migration_2_is_atomic(tmp_path: Path) -> None:
    """Verify migration 2 rolls back on failure."""
    db = await open_db(str(tmp_path / "test.db"))
    try:
        await db.execute("CREATE TABLE schema_migrations (version INTEGER PRIMARY KEY);")
        await db.commit()

        # Add a test migration that fails mid-way
        original = db_module._MIGRATIONS.copy()
        db_module._MIGRATIONS[99] = """
        CREATE TABLE test_table (id INTEGER PRIMARY KEY);
        INSERT INTO nonexistent_table VALUES (1);
        """

        try:
            with pytest.raises(Exception):
                await _apply_migration(db, 99)

            # Verify rollback: migration NOT recorded
            async with db.execute(
                "SELECT version FROM schema_migrations WHERE version = 99;"
            ) as cursor:
                assert await cursor.fetchone() is None

            # Verify rollback: table NOT created
            async with db.execute(
                "SELECT name FROM sqlite_master WHERE type='table' AND name='test_table';"
            ) as cursor:
                assert await cursor.fetchone() is None
        finally:
            db_module._MIGRATIONS = original
    finally:
        await db.close()
```

### Common Pitfalls

- **Non-idempotent statements** — `ALTER TABLE ADD COLUMN` without `IF NOT EXISTS` will fail on re-run. Use explicit checks if needed.
- **Comments containing semicolons** — the migration parser strips comments correctly, but avoid unusual comment syntax.
- **String literals with semicolons** — the parser handles these; no special escaping needed.
- **Multiple operations in one migration** — keep migrations focused. Combine related DDL but split unrelated changes.

---

## 7. Logging

- Use **structlog** for every log message.
- Bind contextual key-value pairs — never format strings manually.
- Log levels: `debug` for development detail, `info` for operational events, `warning` for recoverable issues, `error` for failures, `critical` for fatal problems.
- **Never log sensitive data** (passwords, tokens, session tokens, raw credentials, private keys).
  - For session correlation without leaking token material, use a one-way hash fragment: `hashlib.sha256(token.encode()).hexdigest()[:12]`.
  - Use numeric database IDs for entity correlation instead of raw identifiers: `session_id=session.id` instead of `token=session.token`.

```python
import structlog
import hashlib

log: structlog.stdlib.BoundLogger = structlog.get_logger()

async def ban_ip(ip: str, jail: str) -> None:
    log.info("banning_ip", ip=ip, jail=jail)
    try:
        await _execute_ban(ip, jail)
        log.info("ip_banned", ip=ip, jail=jail)
    except BanError as exc:
        log.error("ban_failed", ip=ip, jail=jail, error=str(exc))
        raise

async def logout_session(db: aiosqlite.Connection, token: str) -> None:
    # Use a one-way hash for token correlation in logs
    token_hash = hashlib.sha256(token.encode()).hexdigest()[:12]
    await session_repo.delete_session(db, token)
    log.info("session_terminated", token_hash=token_hash)
```

---

## 8. Error Handling

- Define **custom exception classes** for domain errors (e.g., `JailNotFoundError`, `BanFailedError`).
- Catch specific exceptions — never bare `except:` or `except Exception:` without re-raising.
- Map domain exceptions to HTTP status codes via FastAPI **exception handlers** registered on the app.
- Always log errors with context before raising.

```python
class JailNotFoundError(Exception):
    def __init__(self, name: str) -> None:
        self.name: str = name
        super().__init__(f"Jail '{name}' not found")

# In main.py
@app.exception_handler(JailNotFoundError)
async def jail_not_found_handler(request: Request, exc: JailNotFoundError) -> JSONResponse:
    return JSONResponse(status_code=404, content={"detail": f"Jail '{exc.name}' not found"})
```

### Routers and Exception Propagation

- **Routers must NOT construct `HTTPException` for domain errors** — let domain exceptions propagate.
- Routers should never have helper functions like `_bad_gateway()`, `_not_found()`, `_conflict()` etc. that convert domain exceptions to `HTTPException`.
- All domain exception types must have corresponding handlers registered in `main.py` via `app.add_exception_handler()`.
- Exception handlers are registered in order from most specific to least specific — FastAPI evaluates them in registration order.

```python
# ❌ BAD — routers constructing HTTPException for domain exceptions
@router.get("/{name}")
async def get_jail(name: str, socket_path: Fail2BanSocketDep) -> JailDetailResponse:
    try:
        return await jail_service.get_jail(socket_path, name)
    except JailNotFoundError:
        raise HTTPException(status_code=404, detail=f"Jail not found: {name!r}") from None

# ✅ GOOD — domain exception propagates to global handler
@router.get("/{name}")
async def get_jail(name: str, socket_path: Fail2BanSocketDep) -> JailDetailResponse:
    return await jail_service.get_jail(socket_path, name)
```

All domain exceptions raised by services propagate to handlers in `main.py`, ensuring:
1. Consistent error response format across the entire API.
2. No duplicated exception-to-HTTP-status mapping logic.
3. Easy to audit all error codes — they are all in one place.

### Error Message Hygiene

HTTP responses must never leak sensitive internal details that aid attackers or expose infrastructure:

- **Never include system paths** in HTTP error messages (e.g., `/var/run/fail2ban/fail2ban.sock`, `/etc/fail2ban/`).
- **Never include raw exception messages** that expose internal parsing or implementation logic.
- **Log full details server-side only** — exception handlers must log `error=str(exc)` with full exception context, but return generic user-friendly messages in the HTTP response.

```python
# ❌ BAD — leaks socket path and internal details to the client
async def _fail2ban_connection_handler(request: Request, exc: Fail2BanConnectionError) -> JSONResponse:
    return JSONResponse(
        status_code=502,
        content={"detail": f"Cannot reach fail2ban: {exc}"},  # exc includes socket path!
    )

# ✅ GOOD — generic message in response, full details in server logs
async def _fail2ban_connection_handler(request: Request, exc: Fail2BanConnectionError) -> JSONResponse:
    log.warning(
        "fail2ban_connection_error",
        path=request.url.path,
        method=request.method,
        error=str(exc),  # Full details logged server-side
    )
    return JSONResponse(
        status_code=502,
        content={"detail": "Cannot reach the fail2ban service. Check the server status page."},
    )
```

### Exception Taxonomy and HTTP Mapping

BanGUI uses a **standardized exception taxonomy** that maps domain errors to HTTP status codes consistently across all services. This allows routers to handle exceptions by category rather than by individual type, reducing code duplication and ensuring consistent client-facing error responses.

#### Exception Categories

All domain exceptions inherit from one of six base categories defined in `app.exceptions`:

| Base Exception | HTTP Status | Meaning | Example |
|---|---|---|---|
| `NotFoundError` | 404 | Requested domain entity not found | `JailNotFoundError`, `FilterNotFoundError` |
| `BadRequestError` | 400 | Invalid input, validation failure, or invalid identifier | `ConfigValidationError`, `JailNameError` |
| `ConflictError` | 409 | State conflict or resource constraint violation | `JailAlreadyActiveError`, `FilterAlreadyExistsError` |
| `OperationError` | 500 | Domain operation failure (write, update, delete) | `ConfigWriteError`, `ConfigFileWriteError` |
| `ServiceUnavailableError` | 503 | Infrastructure or external service unreachable | `Fail2BanConnectionError`, `ConfigDirError` |

#### Service Exception Mapping

Every service-specific exception inherits from exactly one category. This allows `main.py` to register just **5 exception handlers** instead of 25+:

```python
# In app/exceptions.py — define each exception once with its category
class JailNotFoundError(NotFoundError):
    def __init__(self, name: str) -> None:
        self.name = name
        super().__init__(f"Jail not found: {name!r}")

class JailAlreadyActiveError(ConflictError):
    def __init__(self, name: str) -> None:
        self.name = name
        super().__init__(f"Jail is already active: {name!r}")

# In app/main.py — register category handlers
app.add_exception_handler(NotFoundError, _not_found_handler)
app.add_exception_handler(BadRequestError, _bad_request_handler)
app.add_exception_handler(ConflictError, _conflict_handler)
app.add_exception_handler(OperationError, _domain_error_handler)
app.add_exception_handler(ServiceUnavailableError, _service_unavailable_handler)
```

#### Service Exception Reference

When writing a new service, determine which category each exception belongs to:

- **Not found**: Always `NotFoundError` (e.g., jail, filter, action, config file not found)
- **Invalid input**: Always `BadRequestError` (e.g., validation errors, invalid names, regex compile failure)
- **State conflicts**: Always `ConflictError` (e.g., already exists, already active, readonly resource)
- **Operation failures**: Always `OperationError` (e.g., write failed, update failed, command failed)
- **Infrastructure**: Always `ServiceUnavailableError` (e.g., config dir missing, socket unreachable, fail2ban protocol error)

#### Client Expectations

Clients should expect the following HTTP status codes and response format for all domain errors:

```json
HTTP 400 Bad Request
{
  "detail": "Jail name contains invalid characters"
}

HTTP 404 Not Found
{
  "detail": "Jail not found: 'sshd'"
}

HTTP 409 Conflict
{
  "detail": "Jail is already active: 'sshd'"
}

HTTP 500 Internal Server Error
{
  "detail": "Failed to write configuration: permission denied"
}

HTTP 503 Service Unavailable
{
  "detail": "Cannot reach the fail2ban service. Check the server status page."
}
```

The `detail` field always contains the exception's message (from `str(exc)`). Sensitive details (socket paths, file paths, internal error messages) are never included — they are logged server-side only.



- **Network I/O**: `TimeoutError`, `aiohttp.ClientError`, `asyncio.TimeoutError`
- **File I/O**: `OSError` (includes `IOError`, `FileNotFoundError`, `PermissionError`)
- **JSON parsing**: `json.JSONDecodeError`, `ValueError`
- **Database errors**: `aiosqlite.Error` and derivatives (caught as `OSError`)
- **Third-party libraries**: Specific exception classes (e.g., `geoip2.errors.GeoIP2Error`)

**When catching service-critical exceptions**:
1. Catch the specific exception types for the operation.
2. Log with the exception type and relevant context.
3. Return a safe fallback (empty dict, None, etc.) or re-raise if the service cannot function.

**When truly unavoidable broad catches are needed** (e.g., retrying transient network failures):
1. Place specific catches first.
2. Add one final `except Exception` **after** specific cases, with `error_type="unexpected"` logged to flag surprises.
3. Document why broad catching is necessary (e.g., "tests use mock objects that may raise arbitrary exceptions").

**Example:**
```python
async def lookup_batch(ips: list[str], http_session: aiohttp.ClientSession) -> dict[str, GeoInfo]:
    """Resolve multiple IPs, returning empty map on failure."""
    try:
        result = await http_session.post(url, json=payload, timeout=timeout)
    except (TimeoutError, aiohttp.ClientError) as exc:
        # Expected network failures — log and return empty result
        log.warning("geo_batch_http_failed", error=type(exc).__name__)
        return {}
    except Exception as exc:
        # Unexpected — log as error for investigation
        log.error("geo_batch_unexpected_error", error=type(exc).__name__)
        return {}
```

---

## 9. Testing

- **Every** new feature or bug fix must include tests.
- Tests live in `tests/` mirroring the `app/` structure.
- Use `pytest` with `pytest-asyncio` for async tests.
- Use `httpx.AsyncClient` to test FastAPI endpoints (not `TestClient` which is sync).
- Mock external dependencies (fail2ban socket, aiohttp calls) — tests must never touch real infrastructure.
- Aim for **>80 % line coverage** — critical paths (auth, banning, scheduling) must be 100 %.
- Test names follow `test_<unit>_<scenario>_<expected>` pattern.

```python
import pytest
from httpx import AsyncClient, ASGITransport
from app.main import create_app

@pytest.fixture
async def client() -> AsyncClient:
    app = create_app()
    transport: ASGITransport = ASGITransport(app=app)
    async with AsyncClient(transport=transport, base_url="http://test") as ac:
        yield ac

@pytest.mark.asyncio
async def test_list_jails_returns_200(client: AsyncClient) -> None:
    response = await client.get("/api/jails/")
    assert response.status_code == 200
    data: dict = response.json()
    assert "jails" in data
```

---

## 9.1 Background Tasks and Scheduler Architecture

BanGUI uses **APScheduler 4.x** (async mode) to manage background jobs that execute on a schedule without user interaction. This section documents how to write and register background tasks.

### Task Location and Structure

All background tasks live in `backend/app/tasks/` as separate modules. Each task:
- Exports a `register(app: FastAPI) -> None` or `async def register(app: FastAPI) -> None` function.
- Opens its own database connection using `app.db.open_db()` or the `task_db()` helper.
- Closes connections when work completes (use the async context manager pattern).
- Runs independently of the FastAPI request/response cycle.

### Example Task

```python
# backend/app/tasks/my_task.py
import structlog
from fastapi import FastAPI
from apscheduler.schedulers.asyncio import AsyncIOScheduler

log = structlog.get_logger()

async def my_background_job(app: FastAPI) -> None:
    """Do important work on a schedule."""
    log.info("my_background_job_started")
    try:
        db = await app.db.open_db(app.state.settings.database_path)
        try:
            # Do work...
            pass
        finally:
            await db.close()
    except Exception:
        log.error("my_background_job_failed", exc_info=True)

def register(app: FastAPI) -> None:
    """Register the job with the scheduler."""
    scheduler: AsyncIOScheduler = app.state.scheduler
    scheduler.add_job(
        my_background_job,
        args=(app,),
        trigger="interval",
        seconds=60,
        id="my_task",
        name="My Background Job",
    )
```

### Accessing Shared Resources in Tasks

Since tasks do not have access to `Depends(get_db)` (no request scope), they must:
1. **Open their own DB connection** via `app.state.db_factory.open_db(path)`.
2. **Access app-level state** — `app.state.http_session`, `app.state.geo_cache`, `app.state.settings`, etc.
3. **Use structlog** for all logging (never `print()`).

### Single-Worker Requirement

**The scheduler is bound to a single asyncio event loop and cannot be shared across multiple worker processes.** BanGUI enforces single-worker mode to prevent duplicate task execution.

- **Deployment constraint:** Set `BANGUI_WORKERS=1` (default).
- **Startup validation:** `startup_shared_resources()` raises `RuntimeError` if `BANGUI_WORKERS > 1`.
- See [Architekture.md § 9.2](Architekture.md) for full details.

---

## 10. Code Style & Tooling

| Tool | Purpose |
|---|---|
| **Ruff** | Linter and formatter (replaces black, isort, flake8). |
| **mypy** or **pyright** | Static type checking in strict mode. |
| **pre-commit** | Run ruff + type checker before every commit. |

- Line length: **120 characters** max.
- Strings: use **double quotes** (`"`).
- Imports: sorted by ruff — stdlib → third-party → local, one import per line.
- No unused imports, no unused variables, no `# type: ignore` without explanation.
- Docstrings in **Google style** on every public function, class, and module.

---

## 11. fail2ban Response Utilities

All services that interact with the fail2ban daemon must use the canonical response parsing utilities from `app.utils.fail2ban_response`. This ensures consistent error handling, type safety, and makes it easy to fix bugs in response handling across the entire codebase.

### Available Functions

**`ok(response: object) -> object`**
Extracts the payload from a fail2ban ``(return_code, data)`` response tuple.
- Raises `ValueError` if return code ≠ 0 or response shape is invalid.
- Use this on every response from `Fail2BanClient.send()`.

**`to_dict(pairs: object) -> dict[str, object]`**
Converts a list of ``(key, value)`` pairs (fail2ban's native response format) to a Python dict.
- Silently ignores malformed entries and non-list/tuple inputs.
- Always returns a dict (empty if input is invalid).

**`ensure_list(value: object | None) -> list[str]`**
Coerces fail2ban response values (which may be `None`, a single string, or a list) to a normalized list of strings.
- Handles all three cases consistently.
- Returns empty list for `None` or empty strings.

**`is_not_found_error(exc: Exception) -> bool`**
Checks if an exception indicates a jail does not exist.
- Checks for multiple error message patterns (case-insensitive).
- Use this to distinguish "jail not found" errors from other failures.

### Example Usage

```python
from app.utils.fail2ban_response import ok, to_dict, ensure_list, is_not_found_error
from app.utils.fail2ban_client import Fail2BanClient

client = Fail2BanClient(socket_path="/var/run/fail2ban/fail2ban.sock")

try:
    # Get jail status
    response = await client.send(["status", "sshd", "short"])
    status_dict = to_dict(ok(response))  # Extract payload and convert to dict
    
    # Get list of banned IPs
    ban_response = await client.send(["get", "sshd", "banip"])
    banned_ips = ensure_list(ok(ban_response))  # Normalize to list of strings
    
except ValueError as exc:
    if is_not_found_error(exc):
        raise JailNotFoundError("sshd") from exc
    raise
```

### Why This Matters

Before this utility module, every service implemented its own copy of these functions, leading to:
- Code duplication across 7+ service files.
- Subtle inconsistencies in error handling.
- Difficult maintenance — every bug fix required touching multiple files.

Now, all services import from a single authoritative source, making response handling consistent, maintainable, and type-safe.

---

## 12. Configuration & Secrets

- All configuration lives in **environment variables** loaded through **pydantic-settings**.
- Secrets (master password hash, session key) are **never** committed to the repository.
- Provide a `.env.example` with all keys and placeholder values.
- Validate config at startup — fail fast with a clear error if a required value is missing.

```python
from pydantic_settings import BaseSettings
from pydantic import Field

class Settings(BaseSettings):
    database_path: str = Field("bangui.db", description="Path to SQLite database")
    fail2ban_socket: str = Field("/var/run/fail2ban/fail2ban.sock", description="fail2ban socket path")
    session_secret: str = Field(..., description="Secret key for session signing")
    log_level: str = Field("info", description="Logging level")

    model_config = {"env_prefix": "BANGUI_", "env_file": ".env"}
```

### Session Secret Configuration

The `session_secret` is the HMAC key used to sign all session tokens. It must be at least 32 characters (256 bits) to provide sufficient cryptographic strength for HMAC-SHA256.

**Minimum Length:** 32 characters

**Why 32 characters?** Session tokens are signed using HMAC-SHA256. A secret shorter than 32 bytes (256 bits) significantly weakens the signature, potentially allowing attackers to forge valid tokens. The constraint is enforced at startup — the application will fail to start if `session_secret` is shorter than 32 characters.

**Generation:** Generate a secure secret using Python:

```bash
python -c "import secrets; print(secrets.token_hex(32))"
```

This produces a 64-character hexadecimal string (256 bits) suitable for production use.

**Environment Variable:**

```bash
BANGUI_SESSION_SECRET="your-32-character-minimum-secret-here"
```

**Never** commit the actual secret to the repository. Provide a `.env.example` with a placeholder:

```bash
# .env.example
BANGUI_SESSION_SECRET="set-this-to-a-32-character-minimum-secret"
```

### Session Cookie Security

The `session_cookie_secure` configuration controls the `Secure` flag on the session cookie. This flag prevents browsers from sending the session cookie over unencrypted HTTP.

**Default:** `true` — Production deployments are secure by default. Cookies are only sent over HTTPS.

**Local Development:** Set `BANGUI_SESSION_COOKIE_SECURE=false` in your compose file or `.env` to allow cookies over HTTP (required for `localhost:8000`).

```yaml
# Docker/compose.debug.yml
environment:
  BANGUI_SESSION_COOKIE_SECURE: "false"  # Allow HTTP during local development
```

**Important:** If `Secure=true` is set, browsers will reject the session cookie when the backend is served over HTTP. Ensure your nginx/reverse proxy terminates TLS and passes `X-Forwarded-Proto: https` so FastAPI knows the connection is secure.

### CSRF Protection Middleware

State-mutating endpoints (POST, PUT, DELETE, PATCH) authenticated via session cookies are protected by the `CsrfMiddleware`, which enforces a custom header check.

**How It Works:**

1. For every request using a mutating HTTP method, the middleware checks:
   - Is this request authenticated via session cookie (not Bearer token)?
   - If yes, require the custom header `X-BanGUI-Request: 1`.
   - If missing or incorrect, return `403 Forbidden`.

2. **Bearer token requests** (via `Authorization: Bearer` header) bypass the check because tokens are not CSRF-vulnerable — they are never automatically sent on cross-origin requests.

3. **Safe HTTP methods** (GET, HEAD, OPTIONS) bypass the check.

4. **Cross-site protection:** Cross-site JavaScript (`fetch()` calls from other origins) cannot set custom headers without CORS preflight, which the backend rejects for non-allowed origins. This provides defense-in-depth against subdomain attacks and XSS injection.

**Implementation Location:**
- Middleware: `backend/app/middleware/csrf.py`
- Registered in: `backend/app/main.py` via `app.add_middleware(CsrfMiddleware)`

**Example:**
```python
# ✓ Cookie-authenticated POST with CSRF header — allowed
POST /api/bans
Cookie: bangui_session=...
X-BanGUI-Request: 1

# ✗ Cookie-authenticated POST without CSRF header — rejected with 403
POST /api/bans
Cookie: bangui_session=...
(no X-BanGUI-Request header)

# ✓ Bearer token authentication without CSRF header — allowed
POST /api/bans
Authorization: Bearer <token>
(no X-BanGUI-Request header needed)

# ✓ Safe GET method without CSRF header — allowed
GET /api/jails
Cookie: bangui_session=...
(no X-BanGUI-Request header needed)
```

### fail2ban_start_command Configuration

The `fail2ban_start_command` setting specifies the shell command used to start the fail2ban daemon during recovery operations (e.g., after a rollback).

**Format & Parsing:**
- The command is split into arguments using `shlex.split()`, which respects shell quoting rules.
- Paths with spaces must be quoted. Example: `"/opt/my tools/fail2ban-client" start`.
- The command is **not** executed through a shell — no shell variables or globbing are interpreted.

**Validation:**
- The command is validated at startup using `shlex.split()`. Mismatched quotes will raise a `ValueError` with the problematic command in the error message.

**Environment Variables:**
```bash
BANGUI_FAIL2BAN_START_COMMAND="fail2ban-client start"           # Default
BANGUI_FAIL2BAN_START_COMMAND="systemctl start fail2ban"        # systemd
BANGUI_FAIL2BAN_START_COMMAND='"/opt/my tools/fail2ban" start'  # Quoted path
```

**Common Pitfall:**
Using `.split()` instead of `shlex.split()` would break commands with spaces in paths. Always use quoted strings for paths that contain whitespace.

### IP Geolocation Resolution

BanGUI resolves IP addresses to country codes and network organization information for ban analytics and geomapping. The geolocation system implements a **primary + fallback** resolution strategy to balance security and availability:

1. **Primary Resolver (MaxMind GeoLite2-Country):** All IP lookups first attempt resolution using a local MaxMind GeoLite2-Country MMDB database file (if available). The MMDB is downloaded offline and mounted into the container — no IP data is sent over the network.
2. **Fallback Resolver (ip-api.com HTTP):** If the MMDB is unavailable or returns no result, the system can fall back to the ip-api.com HTTP API. **This fallback must be explicitly enabled** and only sends unresolved IPs over HTTP. HTTP is disabled by default for security (to avoid sending IP addresses in cleartext).

**Download & Configure MaxMind GeoLite2:**

The MaxMind GeoLite2-Country MMDB requires a free account and license key. To set up the database:

1. **Create a free MaxMind account** at https://www.maxmind.com/en/geolite2/signup and download your license key.
2. **Download the GeoLite2-Country MMDB** using the provided script or manually from the MaxMind downloads page.
3. **Mount the MMDB into the BanGUI container** at a known path (e.g., `/data/GeoLite2-Country.mmdb`).
4. **Set `BANGUI_GEOIP_DB_PATH`** to the mounted path in your environment.

Example Docker Compose configuration:

```yaml
services:
  bangui:
    volumes:
      - ./GeoLite2-Country.mmdb:/data/GeoLite2-Country.mmdb:ro
    environment:
      BANGUI_GEOIP_DB_PATH: /data/GeoLite2-Country.mmdb
```

**Fallback to HTTP (Not Recommended):**

If the MMDB cannot be mounted (e.g., in restricted environments), you can enable the HTTP fallback:

```yaml
services:
  bangui:
    environment:
      BANGUI_GEOIP_ALLOW_HTTP_FALLBACK: "true"
```

**⚠️ Security Warning:** Enabling HTTP fallback causes unresolved IP addresses to be sent **unencrypted** to ip-api.com. This is a privacy and GDPR/CCPA concern. Only enable this if the MMDB absolutely cannot be provisioned, and understand the implications.

**Data Structure:**

The `GeoInfo` returned by the resolution system includes:
- `country_code` (str | None): ISO 3166-1 alpha-2 country code (e.g., `"US"`, `"DE"`).
- `country_name` (str | None): Human-readable country name (e.g., `"United States"`).
- `asn` (str | None): Autonomous System Number (e.g., `"AS3320"`). Only populated when using the HTTP API; local MMDB lookups return `None`.
- `org` (str | None): Organization name associated with the ASN. Only populated when using the HTTP API; local MMDB lookups return `None`.

**Environment Variables:**

```bash
BANGUI_GEOIP_DB_PATH=/data/GeoLite2-Country.mmdb           # Path to MaxMind MMDB (primary)
BANGUI_GEOIP_ALLOW_HTTP_FALLBACK="false"                   # Default: false (MMDB-only)
BANGUI_GEOIP_ALLOW_HTTP_FALLBACK="true"                    # Enable HTTP fallback (not recommended)
```

**Caching & Performance:**

- Resolved IPs are cached in-memory and persisted to SQLite for fast subsequent lookups.
- Failed lookups are cached for 5 minutes to avoid hammering external APIs.
- The background `geo_cache_flush` task (runs every 60 seconds) persists newly resolved entries to the database.
- The background `geo_re_resolve` task (configurable schedule) periodically re-resolves stale entries to keep data fresh.
- The background `geo_cache_cleanup` task (runs nightly) removes entries not referenced in the configured retention period (default: 90 days) to prevent unbounded database growth and maintain query performance.

**Retention & Cleanup:**

The `geo_cache` table tracks the last time each IP was referenced via a `last_seen` timestamp. Over time, as unique IPs accumulate, the table can grow very large, degrading query performance on every geo lookup. To manage this:

- The `geo_cache_cleanup` background task runs once per day (default: midnight UTC).
- It removes all entries where `last_seen` is older than the configured retention period (default: 90 days).
- If a purged IP is encountered again after cleanup, it will be re-resolved from the MaxMind database or ip-api.com (if configured).
- The retention period is controlled by the constant `GEO_CACHE_RETENTION_DAYS` in `backend/app/tasks/geo_cache_cleanup.py`.

### API Documentation Configuration

The `enable_docs` setting controls whether FastAPI serves interactive API documentation at `/api/docs` (Swagger UI) and `/api/redoc` (ReDoc).

**Default:** `false` — API documentation is disabled by default to prevent information disclosure in production.


**When to Enable:**
- Set `BANGUI_ENABLE_DOCS=true` in development and debugging environments only.
- Never enable in production. Exposed API documentation reveals all endpoints, request/response schemas, and allows direct API invocation from the browser.

**Environment Variables:**
```bash
BANGUI_ENABLE_DOCS="true"   # Enable docs in development
BANGUI_ENABLE_DOCS="false"  # Disable docs (default)
# Unset                     # Defaults to false (production)
```

**Debug Compose File:**
The `Docker/compose.debug.yml` sets `BANGUI_ENABLE_DOCS: "true"` for local development. Production compose files (`Docker/compose.prod.yml`) leave this unset, defaulting to `false`.

**Middleware Allowlist:**
The `SetupRedirectMiddleware` in `main.py` includes `/api/docs`, `/api/redoc`, and `/api/openapi.json` in its `_ALWAYS_ALLOWED` paths so documentation can be accessed before setup completes (if enabled).

### Log Path Validation & Allowlisting

Authenticated users can instruct fail2ban to monitor additional log files through the API endpoint `POST /api/config/jails/{name}/logpath`. To prevent path-traversal attacks and unauthorized reads of sensitive system files, all requested log paths must resolve to locations within a configurable allowlist of safe directories.

**Allowed Directories:**
- Configured via the `BANGUI_ALLOWED_LOG_DIRS` environment variable (comma-separated list).
- Defaults to: `["/var/log", "/config/log"]`.

**Path Validation Rules:**
1. The requested path is resolved to its canonical form using `Path(log_path).resolve()`, which:
   - Expands relative paths to absolute paths.
   - Resolves symbolic links to their real targets.
   - Normalizes `.` and `..` components.
2. The resolved path is checked using `Path.is_relative_to()` against each allowed directory prefix.
3. If the resolved path is not relative to any allowed directory, a `ValueError` is raised with a descriptive error message.

**Implementation:**
- Validation occurs in the Pydantic model `AddLogPathRequest` using a `@field_validator`.
- The validator runs at request time, before the service layer is invoked.
- Symlinks that escape allowed directories are rejected (see [symlink bypass tests](../../backend/tests/test_models.py)).

**Important:** Use `is_relative_to()`, not `startswith()` or string prefix matching. The latter is bypassable with paths like `/var/log_evil/file.log`.

**Environment Variables:**
```bash
BANGUI_ALLOWED_LOG_DIRS="/var/log,/config/log"                    # Default
BANGUI_ALLOWED_LOG_DIRS="/var/log,/config/log,/home/app/logs"     # Custom directory
```

### Log Target Validation (fail2ban)

The `log_target` field on the global config endpoint (`PUT /api/config/global`) is critical for security because fail2ban runs as root. Users can only set log targets to:

1. **Special values:** `STDOUT`, `STDERR`, `SYSLOG` (case-insensitive)
2. **File paths:** Must resolve to one of the configured allowed directories (same allowlist as log paths)

**Why This Matters:**
- fail2ban creates/opens files with root privileges. Without validation, an attacker could write to arbitrary system paths (e.g., `/etc/cron.d/malicious_script`).
- Validation occurs at **both** the Pydantic model layer (`GlobalConfigUpdate.validate_log_target()`) **and** the service layer (`update_global_config()`) for defense in depth.
- This prevents both HTTP and non-HTTP attack vectors.

**Implementation:**
```python
# Model layer: Automatic validation via @field_validator
update = GlobalConfigUpdate(log_target="/etc/passwd")  # Raises ValidationError → HTTP 422

# Service layer: Defense in depth
await config_service.update_global_config(socket_path, update)  # Validates again before sending to fail2ban
```

### Login Rate Limiting

The login endpoint (`POST /api/auth/login`) is protected against brute-force attacks using an in-memory rate limiter.

**Design:**
- Uses a `dict[str, deque[float]]` keyed by client IP, storing login attempt timestamps within a time window.
- Attempts outside the window are automatically removed during validation checks.
- Expired IP entries are cleaned up to prevent unbounded memory growth.

**Rate Limit Rules:**
- **5 attempts per 60 seconds** per IP address.
- Requests exceeding the limit return **HTTP 429 Too Many Requests** with a `Retry-After` header.
- Each failed login triggers a 10-second server-side delay (`asyncio.sleep`) to further slow attacks, on top of bcrypt hashing (~100ms).

**IP Extraction (Proxy Safety):**
- When behind nginx, the rate limiter reads the real client IP from `X-Forwarded-For` or `X-Real-IP` headers.
- Only trusts these headers when the immediate connection source is in a configured trusted proxy list.
- Prevents attackers from spoofing these headers to bypass rate limits.
- Falls back to the direct connection IP when proxy headers cannot be trusted.

**Process-Local Limitation:**
- The rate limiter is process-local (in-memory). In multi-worker deployments (e.g., Gunicorn with 4 workers), each worker maintains its own rate limit counter.
- This is acceptable because the single-worker constraint is enforced elsewhere. See [TASK-002/003 notes](Instructions.md) for details.

**Implementation:**
- Rate limiter: `app.utils.rate_limiter.RateLimiter`
- IP extraction: `app.utils.client_ip.get_client_ip()`
- Dependency: `LoginRateLimiterDep` in `app.dependencies`


---

## 12. Authentication Endpoints

#### Browser SPA (Cookie-Based)

The **primary** authentication flow for the frontend is **cookie-based** and protects the session token from JavaScript access:

1. **Login (`POST /api/auth/login`)**
   - Accepts `LoginRequest` (password field)
   - Returns `LoginResponse` containing **only** `expires_at` (ISO 8601 UTC timestamp)
   - **Crucially:** The session token is **not** included in the JSON response body
   - Instead, the token is set as an **HttpOnly** `SameSite=Lax` cookie named `bangui_session`
   - Frontend automatically includes this cookie in all requests via `credentials: "include"`

2. **Why not return token in response body?**
   - Third-party JavaScript (analytics, ads, XSS injections) can intercept `fetch()` response bodies
   - If the token were in the response, malicious code could extract and store it in `localStorage`
   - An attacker could then use it via the `Authorization: Bearer <token>` header, bypassing the HttpOnly cookie protection
   - By returning **only** the expiry timestamp, we ensure the token stays exclusively in the HttpOnly cookie

3. **Session Validation (`GET /api/auth/session`)**
   - Frontend calls this on app mount to verify the session is still valid on the server
   - Works with both cookie and Bearer token authentication
   - Returns `{"valid": true}` if the session exists and is not expired
   - Returns **401 Unauthorized** if the session is invalid or expired

4. **Logout (`POST /api/auth/logout`)**
   - Revokes the session in the database
   - Clears the `bangui_session` cookie via `Set-Cookie` header
   - Works with both cookie and Bearer token authentication
   - Idempotent — calling without a session returns 200 without error

#### Programmatic API Clients (Bearer Token)

For non-browser clients (CLI tools, batch scripts, automation) that cannot use cookies, use the **Bearer token authentication path** by sending:

```http
Authorization: Bearer <token>
```

The token can be obtained by parsing the cookie from a login response or, in a future implementation, via a dedicated `POST /api/auth/token` endpoint (currently, these clients extract the token from cookies or use Bearer directly from the signed token value).

**Note:** Bearer token authentication is not recommended for browser-based clients because:
- Tokens must be stored somewhere (localStorage, sessionStorage, or request body)
- All storage mechanisms are accessible to JavaScript and thus vulnerable to XSS
- HttpOnly cookies provide better protection

---

## 13. Password Hashing

The master password is hashed using **bcrypt** with an auto-generated salt. All password validation uses the models in `app.models.auth` and `app.models.setup`.

### The 72-Byte Bcrypt Limitation

**Important:** bcrypt silently truncates all input at **72 bytes** before hashing. This means:
- A user who sets a 100-character password is actually authenticated by only the first 72 bytes
- Extra characters beyond 72 bytes provide **zero additional security**
- An attacker who has reduced their search space to 72 bytes can brute-force the password more efficiently than intended

**Solution:** Both password fields enforce a **maximum length of 72 bytes**:
- `LoginRequest.password` — max 72 characters (enforced via Pydantic `Field(max_length=72)`)
- `SetupRequest.master_password` — max 72 characters (enforced via Pydantic `Field(max_length=72)`)

**Validation flow:**
1. Frontend → hashes password with SHA256 using `SubtleCrypto` before transmission
2. Backend receives SHA256 hash, validates length (≤ 72 bytes)
3. Backend → hashes with bcrypt using `run_blocking(bcrypt.hashpw)` to avoid event loop stall
4. Hash stored in SQLite `settings` table

**If a password exceeds 72 bytes:**
- Pydantic raises `ValidationError` with error code `string_too_long`
- The router returns **HTTP 422 Unprocessable Entity**
- The frontend should inform the user to choose a shorter password

**Implementation:** 
- Models: `app.models.auth.LoginRequest`, `app.models.setup.SetupRequest`
- Service layer: `app.services.auth_service._check_password()`, `app.services.setup_service.run_setup()`

---

## 15. File I/O Conventions

All file write operations to critical configuration files must be **atomic** to prevent corruption if the process is killed mid-write.

### Atomic File Writes

Configuration files (e.g., fail2ban jail configs in `jail.d/`) are essential for system operation. A truncated or corrupt config file can break fail2ban's ability to reload and may disable active protection.

**Rule: Always use write-to-temp + atomic rename**

Never use `Path.write_text()` or `file.write()` directly for critical files. Instead:

1. Create a temporary file in the **same directory** as the target (crucial for atomic `os.replace()`).
2. Write content to the temp file.
3. Atomically rename the temp file to replace the target.
4. Clean up the temp file if an error occurs.

**Implementation Pattern:**

```python
import os
import tempfile
from pathlib import Path

target = Path("/path/to/config/file.conf")

tmp_name: str | None = None
try:
    # Create temp file in target's directory (same filesystem = atomic)
    with tempfile.NamedTemporaryFile(
        mode="w",
        encoding="utf-8",
        dir=target.parent,
        delete=False,
        suffix=".tmp",
    ) as tmp:
        tmp.write(content)
        tmp_name = tmp.name
    # Atomic rename (single syscall on POSIX systems)
    os.replace(tmp_name, target)
except OSError as exc:
    # Clean up temp file on error
    with contextlib.suppress(OSError):
        if tmp_name is not None:
            os.unlink(tmp_name)
    raise ConfigWriteError(f"Cannot write config: {exc}") from exc
```

**Why this matters:**

- `Path.write_text()` overwrites in place. If the process dies mid-write, the file is left truncated or partially written.
- `os.replace()` is atomic on POSIX systems (single rename syscall) **only if source and target are on the same filesystem**.
- Creating the temp file in `target.parent` ensures atomicity.
- On Linux containers, this prevents config corruption and service degradation.

**Atomic write helper:**

A shared `atomic_write(path: Path, content: str)` helper is available in `app/services/config_file_helpers.py`. This is the preferred way to perform atomic writes — it handles all the temp file and cleanup logic:

```python
from app.services.config_file_helpers import atomic_write

atomic_write(path, updated_content)  # Atomic write, auto-cleanup on error
```

**Files requiring atomic writes:**

- All config files under `jail.d/` (created/modified by `_write_conf_file`, `_create_conf_file`, `set_jail_config_enabled`, and `write_jail_config_file`)
- Any critical state files that fail2ban relies on

**Examples in the codebase:**

- `app/services/config_file_helpers.py`: `_write_conf_file`, `_create_conf_file`, `atomic_write`
- `app/services/raw_config_io_service.py`: `set_jail_config_enabled`, `write_jail_config_file`
- `app/services/jail_config_service.py`: `_write_local_file_sync`, `_restore_local_file_sync`

---

## 16. Git & Workflow

- **Branch naming:** `feature/<short-description>`, `fix/<short-description>`, `chore/<short-description>`.
- **Commit messages:** imperative tense, max 72 chars first line (`Add jail reload endpoint`, `Fix ban history query`).
- Every merge request must pass: ruff, type checker, all tests.
- Do not merge with failing CI.
- Keep pull requests small and focused — one feature or fix per PR.

---

## 17. Coding Principles

These principles are **non-negotiable**. Every backend contributor must internalise and apply them daily.

### 15.1 Clean Code

- Write code that **reads like well-written prose** — a new developer should understand intent without asking.
- **Meaningful names** — variables, functions, and classes must reveal their purpose. Avoid abbreviations (`cnt`, `mgr`, `tmp`) unless universally understood.
- **Small functions** — each function does exactly one thing. If you need a comment to explain a block inside a function, extract it into its own function.
- **No magic numbers or strings** — use named constants.
- **Boy Scout Rule** — leave every file cleaner than you found it.
- **Avoid deep nesting** — prefer early returns (guard clauses) to keep the happy path at the top indentation level.

```python
# Good — guard clause, clear name, one job
async def get_active_ban(ip: str, jail: str) -> Ban:
    ban: Ban | None = await repo.find_ban(ip=ip, jail=jail)
    if ban is None:
        raise BanNotFoundError(ip=ip, jail=jail)
    if ban.is_expired():
        raise BanExpiredError(ip=ip, jail=jail)
    return ban

# Bad — nested, vague name
async def check(ip, j):
    b = await repo.find_ban(ip=ip, jail=j)
    if b:
        if not b.is_expired():
            return b
        else:
            raise Exception("expired")
    else:
        raise Exception("not found")
```

### 15.2 Separation of Concerns (SoC)

- Each module, class, and function must have a **single, well-defined responsibility**.
- **Routers** → HTTP layer only (parse requests, return responses).
- **Services** → business logic and orchestration.
- **Repositories** → data access and persistence.
- **Models** → data shapes and validation.
- **Tasks** → scheduled background jobs.
- Never mix layers — a router must not execute SQL, and a repository must not raise `HTTPException`.

### 15.3 Single Responsibility Principle (SRP)

- A class or module should have **one and only one reason to change**.
- If a service handles both ban management *and* email notifications, split it into `BanService` and `NotificationService`.

### 15.4 Don't Repeat Yourself (DRY)

- Extract shared logic into utility functions, base classes, or dependency providers.
- If the same block of code appears in more than one place, **refactor it** into a single source of truth.
- But don't over-abstract — premature DRY that couples unrelated features is worse than a little duplication (see **Rule of Three**: refactor when something appears a third time).

### 15.5 KISS — Keep It Simple, Stupid

- Choose the simplest solution that works correctly.
- Avoid clever tricks, premature optimisation, and over-engineering.
- If a standard library function does the job, prefer it over a custom implementation.

### 15.6 YAGNI — You Aren't Gonna Need It

- Do **not** build features, abstractions, or config options "just in case".
- Implement what is required **now**. Extend later when a real need emerges.

### 15.7 Dependency Inversion Principle (DIP)

- High-level modules (services) must not depend on low-level modules (repositories) directly. Both should depend on **abstractions** (protocols / interfaces).
- Use FastAPI's `Depends()` to inject implementations — this makes swapping and testing trivial.

```python
from typing import Protocol

class BanRepository(Protocol):
    async def find_ban(self, ip: str, jail: str) -> Ban | None: ...
    async def save_ban(self, ban: Ban) -> None: ...

class SqliteBanRepository:
    """Concrete implementation — depends on aiosqlite."""
    async def find_ban(self, ip: str, jail: str) -> Ban | None: ...
    async def save_ban(self, ban: Ban) -> None: ...
```

#### 13.7.1 Repository Module Pattern — Module-as-Protocol Structural Compatibility

BanGUI uses **module-level functions** for repository implementations, not classes. Each repository module (e.g., `session_repo.py`, `blocklist_repo.py`) exports async functions that match the signatures defined in the Protocol interface in `protocols.py`. This is a **structural typing pattern** — mypy accepts the module as a valid Protocol implementation because the function signatures match, *even though* the module is not explicitly annotated as implementing the Protocol.

This approach works correctly with FastAPI's dependency injection via `cast()`:

```python
# In app/repositories/session_repo.py
async def create_session(db: aiosqlite.Connection, token: str, created_at: str, expires_at: str) -> Session:
    """Insert a new session row."""
    ...

# In app/repositories/protocols.py
class SessionRepository(Protocol):
    async def create_session(
        self,
        db: aiosqlite.Connection,
        token: str,
        created_at: str,
        expires_at: str,
    ) -> Session:
        ...

# In app/dependencies.py
async def get_session_repo() -> SessionRepository:
    """Provide the concrete session repository implementation."""
    from app.repositories import session_repo
    return session_repo  # ← mypy accepts this because the module has matching functions
```

**Why this pattern is used:**
- **Simplicity** — no boilerplate class/instance wrapping.
- **Compatibility** — Python's **structural typing** (PEP 544) means the module automatically satisfies the Protocol interface if function signatures match.
- **Testability** — the same DIP principle applies; services depend on the Protocol, not the module directly, so tests can mock the Protocol.

**Risks and mitigations:**
- **Silent breakage if function signatures change** — If a parameter is added or removed from a module function, the module no longer satisfies the Protocol, but mypy does not flag this as an error because the module is loosely coupled. To prevent this, **Protocol signatures in `protocols.py` are the source of truth**. Always check that module functions match the Protocol definitions before merging changes. The CI/CD pipeline validates this compatibility at build time.

**How the validation works (CI check):**
- Before each deployment, run `mypy --strict` to ensure all dependency providers return values compatible with their Protocol types.
- The `cast()` calls in `dependencies.py` are a documented signal that structural compatibility is being verified externally, not via explicit class inheritance.

#### 13.7.2 Session Token Hashing — One-Way Protection Against Database Exposure

Session tokens must be protected against database exposure. **Session tokens are stored as one-way SHA256 hashes in the database** to ensure that if the database file is compromised (volume mount misconfiguration, backup leak, etc.), the session tokens themselves cannot be directly used to hijack sessions.

**Implementation pattern:**

```python
import hashlib
from typing import TYPE_CHECKING

if TYPE_CHECKING:
    import aiosqlite

from app.models.auth import Session

def _hash_token(token: str) -> str:
    """Return the SHA256 hash of a session token."""
    return hashlib.sha256(token.encode()).hexdigest()

async def create_session(
    db: "aiosqlite.Connection",
    token: str,
    created_at: str,
    expires_at: str,
) -> Session:
    """Insert a new session row with the token hash."""
    token_hash = _hash_token(token)
    cursor = await db.execute(
        "INSERT INTO sessions (token_hash, created_at, expires_at) VALUES (?, ?, ?)",
        (token_hash, created_at, expires_at),
    )
    await db.commit()
    # Return the Session with the ORIGINAL token (not the hash)
    # so the service layer can sign and return it to the client.
    return Session(
        id=int(cursor.lastrowid) if cursor.lastrowid else 0,
        token=token,  # ← raw token, not the hash
        created_at=created_at,
        expires_at=expires_at,
    )

async def get_session(
    db: "aiosqlite.Connection",
    token: str
) -> Session | None:
    """Look up a session by token hash."""
    token_hash = _hash_token(token)
    async with db.execute(
        "SELECT id, token_hash, created_at, expires_at FROM sessions WHERE token_hash = ?",
        (token_hash,),
    ) as cursor:
        row = await cursor.fetchone()
    
    if row is None:
        return None
    
    # Return the Session with the INCOMING token (the one the client sent).
    return Session(
        id=int(row[0]),
        token=token,  # ← the raw token passed in
        created_at=str(row[2]),
        expires_at=str(row[3]),
    )
```

**Key points:**

1. **Hash on write** — When inserting a session, hash the token before storage.
2. **Hash on read** — When validating a session, hash the incoming token before the database lookup.
3. **Never store raw tokens** — The `token_hash` column contains only hashes; raw tokens are never persisted.
4. **Return raw tokens to the service layer** — The `Session` model's `token` field contains the raw token (for signing and response), not the hash.
5. **Database schema** — Use `token_hash TEXT NOT NULL UNIQUE` instead of `token TEXT NOT NULL UNIQUE`, and create an index on `token_hash`.
6. **Migration strategy** — When upgrading from plaintext to hashed tokens, drop the old table and recreate it. This invalidates all existing sessions, which is acceptable because the database was exposed in plaintext.

**Why one-way hashing is safe:**
- If an attacker obtains a token hash from the database, they cannot reverse the SHA256 hash to recover the original token.
- The attacker cannot use the hash directly in a client request — they would need the original token to pass the hash check.
- This forces the attacker to either compromise the client (where they'd also get the raw token) or perform a brute-force attack against the hash space (infeasible for random 128-bit tokens).

**Never use symmetric encryption** — symmetric encryption stores a key in the database or environment, which merely shifts the exposure risk. A one-way hash is the correct choice for protecting tokens.

#### 13.7.2a Session Token Signing Format — HMAC-SHA256 Integrity Protection

**All session tokens sent to clients are signed using HMAC-SHA256.** The signed token format is:

```
<raw_token>.<signature>
```

where:
- `<raw_token>` is a 16-byte (128-bit) random hex string generated by `secrets.token_hex(16)`.
- `.` is the separator (defined in `app.utils.constants.SESSION_TOKEN_SIGNATURE_SEPARATOR`).
- `<signature>` is the HMAC-SHA256 hex digest of `<raw_token>` using the configured `session_secret`.

**Example:** `a1b2c3d4e5f6g7h8i9j0k1l2m3n4o5p6.f7e6d5c4b3a2918f7e6d5c4b3a29180`

**Signing and verification pattern:**

```python
import hashlib
import hmac

def _session_token_signature(token: str, secret: str) -> str:
    """Return the HMAC-SHA256 signature for a session token."""
    return hmac.new(secret.encode(), token.encode(), hashlib.sha256).hexdigest()

def sign_session_token(token: str, secret: str) -> str:
    """Return a signed session token string for the client."""
    return f"{token}.{_session_token_signature(token, secret)}"

def unwrap_session_token(token: str, secret: str) -> str:
    """Verify and return the raw token from a signed session token.
    
    Raises ValueError if the token lacks a signature or signature is invalid.
    """
    if "." not in token:
        raise ValueError("Invalid session token.")
    
    raw_token, signature = token.rsplit(".", 1)
    expected_signature = _session_token_signature(raw_token, secret)
    if not hmac.compare_digest(expected_signature, signature):
        raise ValueError("Invalid session token.")
    return raw_token
```

**Key points:**

1. **All tokens must be signed** — Tokens without a signature (no separator) are rejected immediately.
2. **Signature is mandatory** — The `unwrap_session_token()` function raises `ValueError` if the separator is absent.
3. **Use HMAC-SHA256** — Always use `hmac.compare_digest()` for signature verification to prevent timing attacks.
4. **Sign on login** — `login()` creates a raw token, stores it (hashed) in the database, then returns the signed token to the client.
5. **Verify on every request** — The `validate_session()` service verifies the signature by calling `unwrap_session_token()` with the `session_secret`, then looks up the raw token in the database.
6. **Session invalidation** — When upgrading from plaintext to signed tokens (TASK-022), all existing sessions must be invalidated because raw tokens will no longer be stored unencrypted.

**Why HMAC signing is necessary:**

- **Prevents token forgery** — An attacker cannot create a valid token without knowing the `session_secret`.
- **Works alongside hashed storage** — Even if the database is compromised (plaintext before hashing), the attacker gets only the raw token, not a signed token. A raw token without a valid signature is rejected by `unwrap_session_token()`.
- **Timing attack resistance** — `hmac.compare_digest()` compares signatures in constant time, preventing attackers from using timing differences to guess valid signatures.

#### 13.7.3 Session Cache Pluggability — Process-Local vs. Shared Backends

Session validation is expensive (SQLite lookup + password verification). To improve performance, **validated session tokens are cached** using the `SessionCache` interface (`app.utils.session_cache`). The default implementation, `InMemorySessionCache`, stores cached sessions in process-local memory.

**Current implementation (single-worker):**

```python
from app.utils.session_cache import SessionCache, InMemorySessionCache, NoOpSessionCache

class SessionCache(Protocol):
    """Interface for session token validation cache backends."""
    def get(self, token: str) -> Session | None: ...
    def set(self, token: str, session: Session, ttl_seconds: float) -> None: ...
    def invalidate(self, token: str) -> None: ...
    def clear(self) -> None: ...

# Default in-memory implementation — PROCESS-LOCAL
class InMemorySessionCache:
    def __init__(self) -> None:
        self._entries: dict[str, tuple[Session, float]] = {}
```

**Single-worker constraint:**

`InMemorySessionCache` is **process-local** — each worker process has its own dict. In single-worker mode (enforced by TASK-002), this is safe and improves performance. In multi-worker deployments:
- A logout by worker A clears the session from A's cache, but worker B still has it → logout doesn't work.
- Enabling/disabling the cache requires restarting all workers to take effect.

**Multi-worker solution:**

To support multiple workers (future enhancement), implement a shared backend behind the same `SessionCache` Protocol:

```python
# Example Redis implementation (not yet in codebase)
class RedisSessionCache:
    """Session cache backed by Redis."""
    def __init__(self, redis_url: str) -> None:
        self.client = aioredis.from_url(redis_url)
    
    async def get(self, token: str) -> Session | None:
        data = await self.client.get(f"session:{token}")
        return Session.model_validate_json(data) if data else None
    
    async def set(self, token: str, session: Session, ttl_seconds: float) -> None:
        await self.client.setex(
            f"session:{token}",
            int(ttl_seconds),
            session.model_dump_json()
        )
    
    async def invalidate(self, token: str) -> None:
        await self.client.delete(f"session:{token}")
    
    async def clear(self) -> None:
        await self.client.flushdb()
```

To adopt a Redis backend:
1. Create `RedisSessionCache` in `app.utils.session_cache`.
2. Update `app.utils.runtime_state.set_runtime_settings()` to instantiate `RedisSessionCache` when `REDIS_URL` env var is set.
3. Update `app.config.Settings` to accept optional `REDIS_URL`.
4. Tests continue to use `InMemorySessionCache` (no Redis dependency in dev).

**Implementation rules:**
- All cache methods must be `async` (even if the backend is sync).
- Never log session tokens or session data.
- TTL must be respected — expired entries must be removed on access.
- See `app/utils/session_cache.py` for the full Protocol definition and current implementations.

### 15.8 Composition over Inheritance

- Favour **composing** small, focused objects over deep inheritance hierarchies.
- Use mixins or protocols only when a clear "is-a" relationship exists; otherwise, pass collaborators as constructor arguments.

### 15.9 Fail Fast

- Validate inputs as early as possible — at the API boundary with Pydantic, at service entry with assertions or domain checks.
- Raise specific exceptions immediately rather than letting bad data propagate silently.

### 15.10 Law of Demeter (Principle of Least Knowledge)

- A function should only call methods on:
  1. Its own object (`self`).
  2. Objects passed as parameters.
  3. Objects it creates.
- Avoid long accessor chains like `request.state.db.cursor().execute(...)` — wrap them in a meaningful method.

### 15.11 Defensive Programming

- Never trust external input — validate and sanitise everything that crosses a boundary (HTTP request, file, socket, environment variable).
- Handle edge cases explicitly: empty lists, `None` values, negative numbers, empty strings.
- Use type narrowing and exhaustive pattern matching (`match` / `case`) to eliminate impossible states.

### 15.12 SSRF Prevention (Server-Side Request Forgery)

When user-supplied URLs are fetched by the backend, validate them before making any HTTP requests:

1. **Use Pydantic's `AnyHttpUrl` type** to restrict schemes to `http://` and `https://` only.
   - Rejects `file://`, `ftp://`, `gopher://`, and other non-http schemes at the model boundary.

2. **Validate resolved IP addresses** before fetching:
   - Parse the hostname and resolve it via DNS (using `socket.getaddrinfo()`).
   - Use `ipaddress.ip_address().is_private` to reject private/reserved ranges:
     - RFC 1918: `10.0.0.0/8`, `172.16.0.0/12`, `192.168.0.0/16`
     - Loopback: `127.0.0.0/8`, `::1/128`
     - Link-local: `169.254.0.0/16`, `fe80::/10`
     - IPv6 site-local, multicast, and reserved ranges.
   - Raise `ValueError` if validation fails; let the router convert it to HTTP 400.

3. **Guard against DNS rebinding**:
   - Validate DNS at URL creation/validation time (performed during request deserialization).
   - For additional safety, re-validate the connection IP at HTTP client time (e.g., custom `aiohttp.TCPConnector` can inspect the resolved address during connect).

4. **Example implementation** (see `backend/app/utils/ip_utils.py`):
   - `is_private_ip(ip_str: str) → bool`: Checks if IP is private/reserved/loopback/link-local.
   - `async validate_blocklist_url(url: AnyHttpUrl) → None`: Async DNS resolution + private IP check.
   - Service layer calls `await validate_blocklist_url(url)` before persisting; router catches `ValueError` and returns 400.

---

## 18. Quick Reference — Do / Don't

| Do | Don't |
|---|---|
| Type every function, variable, return | Leave types implicit |
| Use `async def` for I/O | Use sync functions for I/O |
| Validate with Pydantic at the boundary | Pass raw dicts through the codebase |
| Log with structlog + context keys | Use `print()` or format strings in logs |
| Write tests for every feature | Ship untested code |
| Use `aiohttp` for HTTP calls | Use `requests` |
| Handle errors with custom exceptions | Use bare `except:` |
| Keep routers thin, logic in services | Put business logic in routers |
| Use `datetime.now(datetime.UTC)` | Use naive datetimes |
| Run ruff + mypy before committing | Push code that doesn't pass linting |
| Keep GET endpoints read-only (no `db.commit()`) | Call `db.commit()` / INSERT inside GET handlers |
| Batch DB writes; issue one `db.commit()` after the loop | Commit inside a loop (1 fsync per row) |
| Use `executemany()` for bulk inserts | Call `execute()` + `commit()` per row in a loop |