Notify a phone that has this installation installed

A `webpush` alert channel, and the PWA it needs to arrive. The payload is
encrypted to the subscription (RFC 8291) and the request signed with this
installation's own keypair (RFC 8292), both over `http-ece` — `pywebpush`
does the same in one call but brings `requests` and `aiohttp` with it, two
HTTP stacks beside httpx on a machine that may be a Raspberry Pi.

The manifest and the worker are hand-written rather than `vite-plugin-pwa`:
there is nothing worth precaching when the page carrying the credential is
`no-store`, so the worker handles `push` and `notificationclick` and nothing
else. `registration.scope` is the app's root in both places it runs, which is
why the payload carries no URL.

A run finishing in error is the first event worth waking someone for; `ok`
and `cancelled` describe to nothing, so a nightly batch that works stays
quiet. The events were already on the bus — only the filter changed.

`WEBPUSH_FILE` is a derived path, so the keypair lands on the data volume
with the alerts beside it. Off it, a rebuild would silently stop every phone
being notified: the key they subscribed against would be gone.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014EbeFPm6WNC3YD9vrqqT3a

backend/fluksio/api/routes/alerts.py

@@ -5,9 +5,11 @@ from typing import Any
 
 from fastapi import APIRouter, Depends, HTTPException
 from fastapi.concurrency import run_in_threadpool
+from pydantic import BaseModel
 
 from fluksio.api.deps import FlowControllerDep, get_current_user
 from fluksio.core.config import settings
+from fluksio.flow import webpush
 from fluksio.flow.alerts import Alert, AlertsConfig
 from fluksio.models import Message
 
@@ -63,6 +65,36 @@ async def save_alerts_config(body: AlertsConfig, controller: FlowControllerDep)
     return body
 
 
+class WebPushKey(BaseModel):
+    key: str
+
+
+class Unsubscribe(BaseModel):
+    endpoint: str
+
+
+@router.get("/webpush/key", response_model=WebPushKey)
+async def read_webpush_key() -> Any:
+    """The key a browser subscribes against. Made on the first ask."""
+    return WebPushKey(key=await run_in_threadpool(webpush.public_key))
+
+
+@router.post("/webpush/subscriptions", response_model=Message)
+async def add_webpush_subscription(body: webpush.Subscription) -> Any:
+    """Remember this browser, so a `webpush` channel can reach it."""
+    if not body.keys.get("p256dh") or not body.keys.get("auth"):
+        raise HTTPException(status_code=422, detail="The subscription has no keys")
+    await run_in_threadpool(webpush.add_subscription, body)
+    return Message(message="This browser will be notified")
+
+
+@router.post("/webpush/unsubscribe", response_model=Message)
+async def remove_webpush_subscription(body: Unsubscribe) -> Any:
+    """Forget this browser."""
+    await run_in_threadpool(webpush.remove_subscription, body.endpoint)
+    return Message(message="This browser will no longer be notified")
+
+
 @router.post("/test/{channel_name}", response_model=Message)
 async def test_channel(channel_name: str, controller: FlowControllerDep) -> Any:
     """Send one alert, to prove the channel works before relying on it."""

backend/fluksio/core/config.py

@@ -22,6 +22,7 @@ DERIVED_PATHS = {
     "FLOWS_DIR": "flows",
     "SECRETS_FILE": "secrets.enc",
     "ALERTS_FILE": "alerts.json",
+    "WEBPUSH_FILE": "webpush.json",
     "PANELS_FILE": "panels.json",
     "OAUTH_PRIVATE_KEY_FILE": "oauth-key.pem",
     "CLOUD_CONFIG_FILE": "cloud.json",
@@ -71,6 +72,10 @@ class Settings(BaseSettings):
     # Which failures reach which channel. Beside the flows, not in them:
     # alerting is the deployment's concern, not any one flow's.
     ALERTS_FILE: Path = Path("flow-data/alerts.json")
+    # This installation's web push keypair and the browsers subscribed to it.
+    # Beside the alerts it serves; deleting it makes every device subscribe
+    # again.
+    WEBPUSH_FILE: Path = Path("flow-data/webpush.json")
     # Where machines can be started from when a node needs one and nothing that
     # could take it is attached. Operator-authored, like the alerts beside it,
     # and absent on an installation that has nowhere to start one.

backend/fluksio/flow/alerts.py

@@ -36,6 +36,7 @@ ALERTING_EVENTS = {
     "engine_degraded",
     "cascade_dropped",
     "queue_unavailable",
+    "run_finished",
 }
 
 # The same fault repeating is the same alert.
@@ -52,10 +53,11 @@ class Channel(BaseModel):
     """Somewhere to send an alert."""
 
     name: str
-    kind: Literal["ntfy", "smtp", "webhook", "dashboard"]
+    kind: Literal["ntfy", "smtp", "webhook", "dashboard", "webpush"]
     enabled: bool = True
     # ntfy: server + topic + optional token. smtp: to. webhook: url.
     # dashboard: message — the record a notification widget reads.
+    # webpush: nothing — it goes to whichever browsers subscribed here.
     config: dict[str, Any] = Field(default_factory=dict)
 
 
@@ -144,6 +146,15 @@ def describe(event: dict[str, Any]) -> Alert | None:
         )
     if kind == "queue_unavailable":
         return Alert(title="The work queue is unreachable", body=error or "")
+    if kind == "run_finished":
+        # Every run ends; only the ones that ended badly are worth a phone.
+        if event.get("status") != "error":
+            return None
+        return Alert(
+            title=f"A run of '{flow}' failed",
+            body=f"Run {event.get('run')} finished with errors.",
+            flow=flow,
+        )
     return None
 
 
@@ -304,6 +315,8 @@ class AlertManager:
                 await self._send_webhook(config, alert)
             elif channel.kind == "dashboard":
                 await self._send_dashboard(config, alert)
+            elif channel.kind == "webpush":
+                await self._send_webpush(alert)
             else:
                 await self._send_email(config, alert)
         except Exception as exc:
@@ -344,6 +357,18 @@ class AlertManager:
         # Blocking: it reads the store to find the declared message.
         await asyncio.to_thread(self.publish, message, alert.model_dump())
 
+    async def _send_webpush(self, alert: Alert) -> None:
+        """Wake every browser that subscribed to this installation.
+
+        No settings of its own: a browser subscribes by pressing a button on
+        the alerts screen, and this channel goes to whichever ones did. Saying
+        so when none have is the point of the test button.
+        """
+        from fluksio.flow import webpush
+
+        if await webpush.send_to_all(alert.model_dump()) == 0:
+            raise ValueError("No browsers are subscribed")
+
     async def _send_webhook(self, config: dict[str, Any], alert: Alert) -> None:
         url = config.get("url")
         if not url:

backend/fluksio/flow/webpush.py

@@ -0,0 +1,241 @@
+"""Web Push: an alert reaching a phone that has this installation installed.
+
+The browser hands us a subscription — an endpoint URL at its own push service,
+plus two keys — and from then on the engine can wake that device without it
+holding a connection open. Two pieces of crypto are involved and both are
+specified: the payload is encrypted to the subscription's keys (RFC 8291,
+aes128gcm) so the push service carries something it cannot read, and the
+request is signed with this installation's own keypair (VAPID, RFC 8292) so the
+service knows who is sending.
+
+Only ``http-ece`` is new here; the signing is `pyjwt` and the request is
+`httpx`, both of which the engine already carries. ``pywebpush`` does all of
+this in one call, but brings `requests` *and* `aiohttp` with it — two more HTTP
+stacks on a machine that may well be a Raspberry Pi.
+
+The keypair is this installation's identity to the push services and lives with
+the subscriptions in one file. Losing it means every browser has to subscribe
+again; it is regenerated on the spot if the file goes missing.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import base64
+import json
+import logging
+import threading
+import time
+from pathlib import Path
+from typing import Any
+from urllib.parse import urlparse
+
+import http_ece
+import httpx
+import jwt
+from cryptography.hazmat.primitives import serialization
+from cryptography.hazmat.primitives.asymmetric import ec
+from pydantic import BaseModel, Field
+
+from fluksio.core.config import settings
+
+logger = logging.getLogger(__name__)
+
+#: How long the push service keeps a notification for a device that is offline.
+TTL_S = 86400
+#: A VAPID token is good for twelve hours; well under the 24 the RFC allows.
+TOKEN_LIFETIME_S = 12 * 60 * 60
+
+# ponytail: one lock over the whole file. Subscriptions change when somebody
+# presses a button, and pruning happens once per send — per-record locking
+# would be machinery for a file with single-digit rows in it.
+_lock = threading.Lock()
+
+
+class Subscription(BaseModel):
+    """What a browser handed us. Opaque apart from the endpoint's host."""
+
+    endpoint: str
+    #: The subscription's public key and its authentication secret, both
+    #: base64url as `PushSubscription.toJSON()` writes them.
+    keys: dict[str, str] = Field(default_factory=dict)
+
+
+class _Store(BaseModel):
+    #: This installation's VAPID private key: the raw P-256 scalar, base64url.
+    private_key: str = ""
+    subscriptions: list[Subscription] = Field(default_factory=list)
+
+
+def _b64url(raw: bytes) -> str:
+    return base64.urlsafe_b64encode(raw).rstrip(b"=").decode()
+
+
+def _unb64url(text: str) -> bytes:
+    # A browser writes these unpadded; base64 refuses them that way.
+    return base64.urlsafe_b64decode(text + "=" * (-len(text) % 4))
+
+
+def _path() -> Path:
+    return settings.WEBPUSH_FILE
+
+
+def _read() -> _Store:
+    """Blocking. Call under `_lock` when a write follows."""
+    path = _path()
+    if not path.exists():
+        return _Store()
+    try:
+        return _Store.model_validate_json(path.read_text())
+    except Exception:
+        # A file that no longer parses must not stop the engine alerting
+        # through its other channels.
+        logger.error("The web push store is unreadable; starting a new one")
+        return _Store()
+
+
+def _write(store: _Store) -> None:
+    path = _path()
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(store.model_dump_json(indent=2))
+
+
+def _private_key(store: _Store) -> ec.EllipticCurvePrivateKey:
+    return ec.derive_private_key(
+        int.from_bytes(_unb64url(store.private_key), "big"), ec.SECP256R1()
+    )
+
+
+def _public_bytes(key: ec.EllipticCurvePrivateKey) -> bytes:
+    return key.public_key().public_bytes(
+        serialization.Encoding.X962, serialization.PublicFormat.UncompressedPoint
+    )
+
+
+def public_key() -> str:
+    """This installation's VAPID public key, generating the pair on first ask.
+
+    Base64url of the uncompressed point, which is the shape
+    `pushManager.subscribe` wants for `applicationServerKey`. Blocking.
+    """
+    with _lock:
+        store = _read()
+        if not store.private_key:
+            key = ec.generate_private_key(ec.SECP256R1())
+            store.private_key = _b64url(
+                key.private_numbers().private_value.to_bytes(32, "big")
+            )
+            _write(store)
+            logger.info("Generated this installation's web push keypair")
+        return _b64url(_public_bytes(_private_key(store)))
+
+
+def add_subscription(subscription: Subscription) -> None:
+    """Remember a browser, replacing what we knew about that endpoint. Blocking."""
+    with _lock:
+        store = _read()
+        store.subscriptions = [
+            s for s in store.subscriptions if s.endpoint != subscription.endpoint
+        ]
+        store.subscriptions.append(subscription)
+        _write(store)
+
+
+def remove_subscription(endpoint: str) -> None:
+    """Blocking."""
+    with _lock:
+        store = _read()
+        store.subscriptions = [s for s in store.subscriptions if s.endpoint != endpoint]
+        _write(store)
+
+
+def subscriptions() -> list[Subscription]:
+    """Blocking."""
+    with _lock:
+        return _read().subscriptions
+
+
+def _vapid_headers(key: ec.EllipticCurvePrivateKey, endpoint: str) -> dict[str, str]:
+    """Prove to the push service which installation is sending (RFC 8292)."""
+    origin = urlparse(endpoint)
+    token = jwt.encode(
+        {
+            "aud": f"{origin.scheme}://{origin.netloc}",
+            "exp": int(time.time()) + TOKEN_LIFETIME_S,
+            # Who to shout at about this traffic. A push service will not read
+            # it unless something has gone wrong.
+            "sub": f"mailto:{settings.EMAILS_FROM_EMAIL or 'admin@fluksio.invalid'}",
+        },
+        key,
+        algorithm="ES256",
+    )
+    return {"Authorization": f"vapid t={token},k={_b64url(_public_bytes(key))}"}
+
+
+def _encrypt(payload: bytes, subscription: Subscription) -> bytes:
+    """Encrypt to this subscription's keys, so only that browser can read it."""
+    # Ephemeral, per message: the whole point of the exchange is that this key
+    # and the subscription's agree on a secret nobody else can derive.
+    ours = ec.generate_private_key(ec.SECP256R1())
+    encrypted: bytes = http_ece.encrypt(
+        payload,
+        private_key=ours,
+        dh=_unb64url(subscription.keys["p256dh"]),
+        auth_secret=_unb64url(subscription.keys["auth"]),
+        version="aes128gcm",
+    )
+    return encrypted
+
+
+async def send_to_all(payload: dict[str, Any]) -> int:
+    """Push one payload to every subscribed browser. Returns how many took it.
+
+    A subscription the push service has retired (404/410) is dropped here —
+    that is the only signal we get that a browser is never coming back, and
+    without acting on it the store grows a dead entry per reinstalled phone.
+    """
+    # The keypair exists by now — nothing can have subscribed without asking
+    # for it — but generating it here costs nothing and keeps this callable.
+    await asyncio.to_thread(public_key)
+    with _lock:
+        store = _read()
+    known = store.subscriptions
+    if not known:
+        return 0
+    key = _private_key(store)
+
+    body = json.dumps(payload).encode()
+    delivered = 0
+    gone: list[str] = []
+    async with httpx.AsyncClient(timeout=10) as client:
+        for subscription in known:
+            try:
+                encrypted = await asyncio.to_thread(_encrypt, body, subscription)
+                response = await client.post(
+                    subscription.endpoint,
+                    content=encrypted,
+                    headers={
+                        "Content-Encoding": "aes128gcm",
+                        "Content-Type": "application/octet-stream",
+                        "TTL": str(TTL_S),
+                        **_vapid_headers(key, subscription.endpoint),
+                    },
+                )
+            except Exception as exc:
+                logger.error("Could not push to a subscription: %s", exc)
+                continue
+            if response.status_code in (404, 410):
+                gone.append(subscription.endpoint)
+            elif response.is_success:
+                delivered += 1
+            else:
+                logger.error(
+                    "A push service refused the notification: %s %s",
+                    response.status_code,
+                    response.text[:200],
+                )
+
+    for endpoint in gone:
+        await asyncio.to_thread(remove_subscription, endpoint)
+        logger.info("Dropped a subscription the push service has retired")
+    return delivered

backend/pyproject.toml

@@ -72,6 +72,11 @@ dependencies = [
     "rich>=13",
     # `fluksio serve` opens a dashboard with it at a terminal.
     "textual>=1.0",
+    # The encrypted-payload half of web push (RFC 8291). It brings only
+    # cryptography, which is already here. `pywebpush` would do the signing and
+    # the request too, at the cost of `requests` *and* `aiohttp` — two HTTP
+    # stacks beside httpx, on a machine that may be a Raspberry Pi.
+    "http-ece>=1.2",
 ]
 
 [project.optional-dependencies]
@@ -143,6 +148,11 @@ implicit_reexport = true
 module = ["pyarrow", "pyarrow.*"]
 ignore_missing_imports = true
 
+# The web push payload encoder carries no annotations.
+[[tool.mypy.overrides]]
+module = ["http_ece"]
+ignore_missing_imports = true
+
 
 [tool.ruff]
 target-version = "py312"

backend/tests/flow/test_alerts.py

@@ -188,6 +188,13 @@ def test_alerting_can_be_switched_off():
             "The work queue is unreachable",
         ),
         ({"type": "engine_degraded", "reason": "lag"}, "The engine is struggling"),
+        # Every run ends. Only a failed one is worth a phone buzzing.
+        (
+            {"type": "run_finished", "flow": "nightly", "run": 7, "status": "error"},
+            "A run of 'nightly' failed",
+        ),
+        ({"type": "run_finished", "flow": "nightly", "status": "ok"}, None),
+        ({"type": "run_finished", "flow": "nightly", "status": "cancelled"}, None),
         ({"type": "node_health", "health": "ok"}, None),
         # The engine publishes `health`, not `status`: reading the wrong key
         # meant a device dropping never alerted anyone.

backend/tests/flow/test_webpush.py

@@ -0,0 +1,142 @@
+"""Web push: the crypto is the part that cannot be checked by reading it."""
+
+import asyncio
+import os
+
+import http_ece
+import httpx
+import jwt
+import pytest
+from cryptography.hazmat.primitives import serialization
+from cryptography.hazmat.primitives.asymmetric import ec
+
+from fluksio.core.config import settings
+from fluksio.flow import webpush
+
+
+@pytest.fixture(autouse=True)
+def store(tmp_path, monkeypatch):
+    """A store per test, so nothing writes into the real installation."""
+    monkeypatch.setattr(settings, "WEBPUSH_FILE", tmp_path / "webpush.json")
+    return tmp_path / "webpush.json"
+
+
+def browser(
+    name: str = "one",
+) -> tuple[ec.EllipticCurvePrivateKey, bytes, webpush.Subscription]:
+    """What a browser hands over when it subscribes."""
+    key = ec.generate_private_key(ec.SECP256R1())
+    auth = os.urandom(16)
+    point = key.public_key().public_bytes(
+        serialization.Encoding.X962, serialization.PublicFormat.UncompressedPoint
+    )
+    return (
+        key,
+        auth,
+        webpush.Subscription(
+            endpoint=f"https://push.example.com/{name}",
+            keys={"p256dh": webpush._b64url(point), "auth": webpush._b64url(auth)},
+        ),
+    )
+
+
+def test_only_the_subscribed_browser_can_read_the_payload():
+    """RFC 8291, and the reason the push service is not trusted with content."""
+    key, auth, subscription = browser()
+    payload = b'{"title":"A run failed"}'
+
+    sealed = webpush._encrypt(payload, subscription)
+
+    assert sealed != payload
+    assert (
+        http_ece.decrypt(sealed, private_key=key, auth_secret=auth, version="aes128gcm")
+        == payload
+    )
+
+
+def test_the_request_is_signed_for_the_push_service_that_gets_it():
+    """RFC 8292: the audience is the endpoint's origin, never its path."""
+    key = ec.generate_private_key(ec.SECP256R1())
+
+    headers = webpush._vapid_headers(key, "https://push.example.com/quite/a/long/path")
+
+    scheme, rest = headers["Authorization"].split(" ", 1)
+    parts = dict(part.split("=", 1) for part in rest.split(","))
+    assert scheme == "vapid"
+    claims = jwt.decode(
+        parts["t"],
+        key.public_key(),
+        algorithms=["ES256"],
+        audience="https://push.example.com",
+    )
+    assert claims["sub"].startswith("mailto:")
+    # The service looks this key up against the one the browser subscribed with.
+    assert webpush._unb64url(parts["k"]) == webpush._public_bytes(key)
+
+
+def test_the_keypair_survives_the_next_ask():
+    first = webpush.public_key()
+    assert webpush.public_key() == first
+
+
+def push_to(handler) -> int:
+    """Send one alert to two browsers, the push services answering `handler`."""
+    webpush.add_subscription(browser("gone")[2])
+    webpush.add_subscription(browser("here")[2])
+
+    transport = httpx.MockTransport(handler)
+    original = httpx.AsyncClient
+
+    def client(**kwargs):
+        kwargs["transport"] = transport
+        return original(**kwargs)
+
+    webpush.httpx.AsyncClient = client  # type: ignore[misc]
+    try:
+        return asyncio.run(webpush.send_to_all({"title": "t", "body": "b"}))
+    finally:
+        webpush.httpx.AsyncClient = original  # type: ignore[misc]
+
+
+def test_every_subscribed_browser_is_pushed_to():
+    seen: list[str] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        seen.append(str(request.url))
+        assert request.headers["content-encoding"] == "aes128gcm"
+        assert request.headers["authorization"].startswith("vapid t=")
+        return httpx.Response(201)
+
+    assert push_to(handler) == 2
+    assert len(seen) == 2
+
+
+def test_a_retired_subscription_is_dropped():
+    """410 is the only word we get that a browser is never coming back."""
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        return httpx.Response(410 if request.url.path.endswith("gone") else 201)
+
+    delivered = push_to(handler)
+
+    assert delivered == 1
+    assert [s.endpoint for s in webpush.subscriptions()] == [
+        "https://push.example.com/here"
+    ]
+
+
+def test_a_channel_with_nobody_subscribed_says_so():
+    """What the test button on the alerts screen reports."""
+    from fluksio.flow.alerts import Alert, AlertManager, Channel
+    from fluksio.flow.events import EventBus
+
+    alerts = AlertManager(EventBus())
+
+    with pytest.raises(ValueError, match="No browsers"):
+        asyncio.run(
+            alerts.send(
+                Channel(name="phones", kind="webpush"),
+                Alert(title="t", body="b"),
+                raise_on_error=True,
+            )
+        )