Merge branch 'fix/edge-hardening'

.devcontainer/setup.sh

@@ -63,6 +63,11 @@ if [ ! -f "$REPO/.env" ]; then
   cp "$REPO/.env.example" "$REPO/.env"
   chown vscode:vscode "$REPO/.env" 2>/dev/null || true
 fi
+# Dev needs the Vite HMR server; the compose default is fail-closed preview.
+if ! grep -q "^FRONTEND_MODE=" "$REPO/.env" 2>/dev/null; then
+  echo "FRONTEND_MODE=dev" >> "$REPO/.env"
+  chown vscode:vscode "$REPO/.env" 2>/dev/null || true
+fi
 echo "--- Waiting for compose services ---"
 python3 - <<'EOF'
 import socket

.env.example

@@ -28,8 +28,10 @@ CALLBACK_BASE_URL=http://api:8000
 # X-Callback-Token). Long random value; empty disables callbacks.
 CALLBACK_TOKEN=
 
-# --- frontend mode: dev server with HMR (prod .env must set `preview`) ---
-FRONTEND_MODE=dev
+# --- frontend mode: unset/preview serves the production bundle ---
+# Dev sets FRONTEND_MODE=dev (HMR); .devcontainer/setup.sh provisions it.
+# Never set `dev` in prod: the Vite dev server must not face the internet.
+# FRONTEND_MODE=dev
 
 # --- access management (request + approval, invite via email) ---
 # Public frontend URL (magic links point here).

CHANGELOG.md

@@ -14,6 +14,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   removed (SSRF / cross-tenant / flow-chaining hardening);
   continuations now resolve from the admin-authored definition.
 - Frontend: `/app/*` routes require login; approver is the session user.
+- Edge hardening: interactive API docs only with debug, security headers
+  (nosniff/frame/Referrer-Policy), 2 MiB body cap, per-IP run throttle,
+  generic upstream-error responses, magic links never logged.
+- Containers run rootless (api/vite as uid 1000, lfx as `lfx` user);
+  frontend fails closed to preview mode unless `FRONTEND_MODE=dev`.
 ### Added
 - Landing page: status-quo problem block, 5-step process with quote,
   portal overview with admin areas, AI platform parts, exception tiers,

Dockerfile.lfx

@@ -13,9 +13,11 @@ ENV PYTHONUNBUFFERED=1 \
 ARG LFX_VERSION=1.12.0
 RUN pip install "lfx~=${LFX_VERSION}" requests
 
-RUN mkdir -p /flows
+# Non-root runtime: uploads persist into the mounted volume.
+RUN useradd -r -s /bin/false lfx && mkdir -p /flows && chown lfx /flows
 VOLUME /flows
 EXPOSE 6004
+USER lfx
 
 # Serve the /flows dir (uploaded flows persist there via --flow-dir).
 # No positional path: with one, lfx>=1.12.1 exits when the dir holds no

backend/Dockerfile

@@ -17,9 +17,10 @@ COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
 
 # Foundational, inert in prod: non-root user for the editor, sudo for setup,
 # openssh-client for the devcontainer post-start hook, curl for the opencode
-# installer in setup.sh, git for the VS Code source-control view.
+# installer in setup.sh, git for the VS Code source-control view, gosu to
+# drop privileges before serving (the server never runs as root).
 RUN apt-get update \
-    && apt-get install -y --no-install-recommends sudo openssh-client curl git ca-certificates \
+    && apt-get install -y --no-install-recommends sudo openssh-client curl git ca-certificates gosu \
     && rm -rf /var/lib/apt/lists/* \
     && useradd -m -s /bin/bash vscode \
     && echo "vscode ALL=(ALL) NOPASSWD:ALL" > /etc/sudoers.d/vscode \

backend/app/auth/mail.py

@@ -124,7 +124,9 @@ def send_magic_link(to_email: str, link: str) -> bool:
     never include tokens or session ids (see test_auth no-leak test).
     """
     if not settings.smtp_host:
-        logger.info("SMTP not configured; magic link for %s: %s", to_email, link)
+        # Never log the link itself: without SMTP this fallback fires in
+        # every environment, and logs are rarely as protected as mail.
+        logger.info("SMTP not configured; dropping magic link for %s", to_email)
         return False
     message = _branded_mail(
         subject="Dein Login-Link für Kirche Einfach",

backend/app/flows/api.py

@@ -10,6 +10,7 @@ shared X-Callback-Token instead.
 from __future__ import annotations
 
 import json
+import logging
 import secrets
 from collections.abc import AsyncIterator, Iterator
 from pathlib import Path
@@ -64,6 +65,8 @@ from .sync import lfx_id_for, sync_from_directory
 flows_router = APIRouter(prefix="/api/flows", tags=["flows"])
 approvals_router = APIRouter(prefix="/api/approvals", tags=["approvals"])
 
+logger = logging.getLogger(__name__)
+
 
 def get_db() -> Iterator[Session]:
     yield from get_session()
@@ -226,9 +229,11 @@ async def deploy_version(
     except LfxConflict as exc:
         raise HTTPException(status_code=409, detail=str(exc)) from exc
     except LfxUnavailable as exc:
-        raise HTTPException(status_code=503, detail=str(exc)) from exc
+        logger.warning("lfx unavailable: %s", exc)
+        raise HTTPException(status_code=503, detail="flow runtime unavailable") from exc
     except LfxError as exc:
-        raise HTTPException(status_code=502, detail=str(exc)) from exc
+        logger.warning("lfx error: %s", exc)
+        raise HTTPException(status_code=502, detail="flow runtime error") from exc
     return _read_version(store.mark_deployed(session, version, uploaded.id))
 
 
@@ -274,9 +279,11 @@ async def run_flow(
     except LfxNotFound as exc:
         raise HTTPException(status_code=502, detail=f"LFX drift: {exc}") from exc
     except LfxUnavailable as exc:
-        raise HTTPException(status_code=503, detail=str(exc)) from exc
+        logger.warning("lfx unavailable: %s", exc)
+        raise HTTPException(status_code=503, detail="flow runtime unavailable") from exc
     except LfxError as exc:
-        raise HTTPException(status_code=502, detail=str(exc)) from exc
+        logger.warning("lfx error: %s", exc)
+        raise HTTPException(status_code=502, detail="flow runtime error") from exc
     return _read_run(run)
 
 
@@ -358,9 +365,11 @@ async def list_langflow_flows(
     try:
         return await langflow.list_flows()
     except LangflowUnavailable as exc:
-        raise HTTPException(status_code=503, detail=str(exc)) from exc
+        logger.warning("langflow unavailable: %s", exc)
+        raise HTTPException(status_code=503, detail="authoring server unavailable") from exc
     except LangflowError as exc:
-        raise HTTPException(status_code=502, detail=str(exc)) from exc
+        logger.warning("langflow error: %s", exc)
+        raise HTTPException(status_code=502, detail="authoring server error") from exc
 
 
 @flows_router.post("/import", response_model=FlowImportRead, status_code=201)
@@ -385,9 +394,11 @@ async def import_flow(
     except LangflowNotFound as exc:
         raise HTTPException(status_code=404, detail=str(exc)) from exc
     except LangflowUnavailable as exc:
-        raise HTTPException(status_code=503, detail=str(exc)) from exc
+        logger.warning("langflow unavailable: %s", exc)
+        raise HTTPException(status_code=503, detail="authoring server unavailable") from exc
     except LangflowError as exc:
-        raise HTTPException(status_code=502, detail=str(exc)) from exc
+        logger.warning("langflow error: %s", exc)
+        raise HTTPException(status_code=502, detail="authoring server error") from exc
     except store.ConflictError as exc:
         raise HTTPException(status_code=409, detail=str(exc)) from exc
     except store.InvalidError as exc:
@@ -397,9 +408,11 @@ async def import_flow(
     except LfxConflict as exc:
         raise HTTPException(status_code=409, detail=str(exc)) from exc
     except LfxUnavailable as exc:
-        raise HTTPException(status_code=503, detail=str(exc)) from exc
+        logger.warning("lfx unavailable: %s", exc)
+        raise HTTPException(status_code=503, detail="flow runtime unavailable") from exc
     except LfxError as exc:
-        raise HTTPException(status_code=502, detail=str(exc)) from exc
+        logger.warning("lfx error: %s", exc)
+        raise HTTPException(status_code=502, detail="flow runtime error") from exc
     return FlowImportRead(definition=_read_definition(definition), version=_read_version(version))
 
 

backend/app/main.py

@@ -5,10 +5,12 @@ from contextlib import asynccontextmanager
 from pathlib import Path
 
 from fastapi import FastAPI
+from starlette.middleware.body_limit import RequestBodyLimitMiddleware
 
 from .auth.api import access_router, auth_router
 from .flows.api import approvals_router, flows_router
 from .flows.sync import run_sync_blocking, watch_forever
+from .middleware import RunThrottleMiddleware, SecurityHeadersMiddleware
 from .settings import settings
 
 logger = logging.getLogger(__name__)
@@ -58,12 +60,21 @@ app = FastAPI(
     title="Kirche Einfach API",
     description="KI-gestützte Verwaltungsplattform für Kirche",
     version="0.1.0",
-    docs_url="/api/docs",
-    redoc_url="/api/redoc",
-    openapi_url="/api/openapi.json",
+    # Interactive docs only in debug: /api/docs maps the whole surface.
+    docs_url="/api/docs" if settings.debug else None,
+    redoc_url="/api/redoc" if settings.debug else None,
+    openapi_url="/api/openapi.json" if settings.debug else None,
     lifespan=lifespan,
 )
 
+# TLS terminates at the reverse proxy: uvicorn --proxy-headers (see
+# entrypoint.sh) trusts its forwarding headers so Secure cookies, client
+# IPs (throttle) and absolute URLs are correct. The API is not directly
+# internet-reachable (proxy-only), so trusting them is safe.
+app.add_middleware(RequestBodyLimitMiddleware, max_body_size=settings.max_body_bytes)
+app.add_middleware(RunThrottleMiddleware)
+app.add_middleware(SecurityHeadersMiddleware)
+
 app.include_router(flows_router)
 app.include_router(approvals_router)
 app.include_router(access_router)

backend/app/middleware.py

@@ -0,0 +1,92 @@
+"""Edge hardening: run throttling, security headers.
+
+Body caps come from Starlette's RequestBodyLimitMiddleware (wired in
+main.py). All limits are settings-driven so tests can shrink them. State
+is in-process memory: correct for the single-replica SQLite deployment;
+revisit with Redis if the API ever scales horizontally.
+"""
+
+from __future__ import annotations
+
+import time
+from collections import defaultdict
+
+from starlette.datastructures import MutableHeaders
+from starlette.requests import Request
+from starlette.responses import JSONResponse
+from starlette.types import ASGIApp, Message, Receive, Scope, Send
+
+from app.settings import settings
+
+# POST .../run is the expensive endpoint (LFX execution, LLM cost):
+# per client IP and minute.
+RUN_PATH_SUFFIX = "/run"
+THROTTLE_WINDOW_SECONDS = 60.0
+
+_hits: dict[str, list[float]] = defaultdict(list)
+
+
+def reset_throttle() -> None:
+    """Drop all counters (tests)."""
+    _hits.clear()
+
+
+def _client_ip(request: Request) -> str:
+    # Behind the TLS proxy ProxyHeadersMiddleware has already rewritten
+    # scope["client"] from X-Forwarded-For.
+    if request.client is not None:
+        return request.client.host
+    forwarded = request.headers.get("x-forwarded-for", "")
+    return forwarded.split(",")[0].strip() or "unknown"
+
+
+class RunThrottleMiddleware:
+    """Per-IP sliding window on POST .../run (429 + Retry-After)."""
+
+    def __init__(self, app: ASGIApp) -> None:
+        self.app = app
+
+    async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
+        if (
+            scope["type"] != "http"
+            or scope.get("method") != "POST"
+            or not scope.get("path", "").endswith(RUN_PATH_SUFFIX)
+        ):
+            await self.app(scope, receive, send)
+            return
+        request = Request(scope)
+        now = time.monotonic()
+        window_start = now - THROTTLE_WINDOW_SECONDS
+        bucket = [t for t in _hits[_client_ip(request)] if t > window_start]
+        if len(bucket) >= settings.run_throttle_per_minute:
+            retry_after = max(1, int(bucket[0] + THROTTLE_WINDOW_SECONDS - now))
+            response = JSONResponse({"detail": "too many runs, try again later"}, status_code=429)
+            response.headers["Retry-After"] = str(retry_after)
+            await response(scope, receive, send)
+            return
+        bucket.append(now)
+        _hits[_client_ip(request)] = bucket
+        await self.app(scope, receive, send)
+
+
+class SecurityHeadersMiddleware:
+    """Baseline response headers (TLS/HSTS live on the proxy)."""
+
+    def __init__(self, app: ASGIApp) -> None:
+        self.app = app
+
+    async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
+        if scope["type"] != "http":
+            await self.app(scope, receive, send)
+            return
+
+        async def secured(message: Message) -> None:
+            if message["type"] == "http.response.start":
+                headers = MutableHeaders(raw=message.setdefault("headers", []))
+                headers["X-Content-Type-Options"] = "nosniff"
+                headers["X-Frame-Options"] = "SAMEORIGIN"
+                headers["Referrer-Policy"] = "strict-origin-when-cross-origin"
+                headers["Content-Security-Policy"] = "frame-ancestors 'self'"
+            await send(message)
+
+        await self.app(scope, receive, secured)

backend/app/settings.py

@@ -36,6 +36,9 @@ class Settings(BaseSettings):
     # it back as X-Callback-Token. Empty disables callbacks (503) — set a
     # long random value in prod (.env CALLBACK_TOKEN).
     callback_token: str = ""
+    # Edge hardening (see app/middleware.py).
+    max_body_bytes: int = 2 * 1024 * 1024
+    run_throttle_per_minute: int = 60
 
     # --- access management (magic-link auth, see app/auth) ---
     # Public frontend URL (magic links point at /login?token= there).

backend/entrypoint.sh

@@ -23,7 +23,7 @@ if id vscode >/dev/null 2>&1; then
 fi
 
 if [ "${APP_ENV:-prod}" = "dev" ]; then
-  exec /opt/venv/bin/uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload
+  exec gosu vscode /opt/venv/bin/uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload --proxy-headers
 else
-  exec /opt/venv/bin/uvicorn app.main:app --host 0.0.0.0 --port 8000
+  exec gosu vscode /opt/venv/bin/uvicorn app.main:app --host 0.0.0.0 --port 8000 --proxy-headers
 fi

backend/tests/test_middleware.py

@@ -0,0 +1,85 @@
+"""Edge hardening: docs off, headers, body cap, run throttle, mail scrub."""
+
+import pytest
+
+from app.middleware import reset_throttle
+
+
+@pytest.fixture(autouse=True)
+def _clean_throttle():
+    reset_throttle()
+    yield
+    reset_throttle()
+
+
+async def test_docs_disabled_outside_debug(client):
+    assert (await client.get("/api/docs")).status_code == 404
+    assert (await client.get("/api/redoc")).status_code == 404
+    assert (await client.get("/api/openapi.json")).status_code == 404
+
+
+async def test_security_headers_present(client):
+    response = await client.get("/health")
+    assert response.status_code == 200
+    assert response.headers["X-Content-Type-Options"] == "nosniff"
+    assert response.headers["X-Frame-Options"] == "SAMEORIGIN"
+    assert response.headers["Referrer-Policy"] == "strict-origin-when-cross-origin"
+    assert response.headers["Content-Security-Policy"] == "frame-ancestors 'self'"
+
+
+async def test_oversize_body_rejected(client):
+    # Default cap is 2 MiB (settings.max_body_bytes); the middleware binds
+    # it at startup, so this test exceeds the real default.
+    big = "x" * (3 * 1024 * 1024)
+    response = await client.post("/api/auth/magic/consume", json={"token": big})
+    assert response.status_code == 413
+
+
+async def test_run_throttle_returns_429(client, lfx_server, login, monkeypatch):
+    from app.settings import settings
+
+    monkeypatch.setattr(settings, "run_throttle_per_minute", 2)
+    await login()
+
+    flow = (await client.post("/api/flows", json={"name": "throttled"})).json()
+    await client.post(
+        f"/api/flows/{flow['id']}/versions",
+        json={
+            "langflow_json": {
+                "name": "t",
+                "data": {
+                    "nodes": [
+                        {
+                            "id": "ChatInput-x",
+                            "data": {
+                                "id": "ChatInput-x",
+                                "type": "ChatInput",
+                                "node": {"template": {}, "outputs": []},
+                            },
+                            "type": "genericNode",
+                        }
+                    ],
+                    "edges": [],
+                },
+            }
+        },
+    )
+    # Undeployed: 422s still pass the throttle (only 429 matters here).
+    for _ in range(2):
+        response = await client.post(f"/api/flows/{flow['id']}/run", json={"input": {}})
+        assert response.status_code == 422
+    response = await client.post(f"/api/flows/{flow['id']}/run", json={"input": {}})
+    assert response.status_code == 429
+    assert response.headers["Retry-After"]
+
+
+def test_magic_link_never_logged(caplog, monkeypatch):
+    from app.auth import mail
+    from app.settings import settings
+
+    monkeypatch.setattr(settings, "smtp_host", "")
+    with caplog.at_level("INFO", logger="app.auth.mail"):
+        assert mail.send_magic_link("a@example.com", "http://x/login?token=SECRET") is False
+    assert "SECRET" not in caplog.text
+    assert "http://x/login" not in caplog.text
+    assert "a@example.com" in caplog.text

docker-compose.yml

@@ -53,8 +53,9 @@ services:
       - path: .env
         required: false
     environment:
-      # Dev default (HMR); prod .env sets `preview` (build + serve bundle).
-      FRONTEND_MODE: ${FRONTEND_MODE:-dev}
+      # Fail closed: unset means the production preview server. Dev sets
+      # FRONTEND_MODE=dev in .env (provisioned by .devcontainer/setup.sh).
+      FRONTEND_MODE: ${FRONTEND_MODE:-preview}
     volumes:
       # No node_modules volume by design: deps live in the bind-mount and the
       # entrypoint runs `npm ci` on first start (a volume used to shadow the

frontend/Dockerfile

@@ -3,11 +3,17 @@
 # file. DevDeps are included: `npm run dev` must work with zero steps.
 FROM node:22-slim
 
+# gosu drops privileges before serving (node user, uid 1000 — same as the
+# vscode editor user, so the bind-mounted node_modules stays writable).
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends gosu \
+    && rm -rf /var/lib/apt/lists/*
+
 WORKDIR /srv/kirche-einfach/frontend
 COPY package.json package-lock.json entrypoint.sh ./
 RUN npm ci
 COPY . ./
-RUN chmod +x entrypoint.sh
+RUN chmod +x entrypoint.sh && chown -R node:node /srv/kirche-einfach/frontend
 
 EXPOSE 3000
 

frontend/entrypoint.sh

@@ -1,7 +1,7 @@
 #!/bin/sh
 # Frontend entrypoint (dev + prod).
-# FRONTEND_MODE=dev (compose default) runs the HMR dev server.
-# FRONTEND_MODE=preview (prod .env) builds once and serves the bundle.
+# Fail closed: unset means the production preview server; dev sets
+# FRONTEND_MODE=dev in .env (see .devcontainer/setup.sh).
 set -e
 
 cd /srv/kirche-einfach/frontend
@@ -9,12 +9,13 @@ cd /srv/kirche-einfach/frontend
 # node_modules lives in the repo bind-mount (no named volume by design):
 # install on first start after a fresh checkout (dev and prod).
 if [ ! -x "node_modules/.bin/vite" ]; then
-  npm ci
+  gosu node npm ci
 fi
 
-if [ "${FRONTEND_MODE:-dev}" = "dev" ]; then
-  exec npm run dev -- --host 0.0.0.0 --port 3000
+if [ "${FRONTEND_MODE:-preview}" = "dev" ]; then
+  exec gosu node npm run dev -- --host 0.0.0.0 --port 3000
 else
-  npm run build
-  exec npm run preview -- --host 0.0.0.0 --port 3000
+  # Build as node so dist/ stays editor-writable; preview serves as node.
+  gosu node npm run build
+  exec gosu node npm run preview -- --host 0.0.0.0 --port 3000
 fi