feat: realtime WebSocket layer for live events and streaming logs

- Backend: ConnectionManager, Redis-backed bus with in-process fallback,
  channel authorization, and typed emit helpers wired into workers
  (ai_runtime, sandbox_gateway), workforce/sandbox/billing/audit/error services.
- Frontend: single shared WS client with backoff + heartbeat, RealtimeProvider,
  useLiveChannel/useLiveEvent hooks, live indicators on task detail, manager
  chat, dashboard, traces, sandbox terminal, admin errors, and billing.
- Contracts: WsEvent/WsServerFrame/WsClientFrame types mirroring the port.

All 255 backend tests pass; web typecheck and build pass.
This commit is contained in:
mjalmoudy94
2026-10-13 12:56:36 +03:30
parent afdef4b444
commit f1e5df29ce
31 changed files with 2981 additions and 28 deletions
@@ -0,0 +1,20 @@
"""Realtime push layer — the WebSocket *interface* adapter.
The adapter is transport and plumbing only: who is connected to which channel
(:mod:`manager`), how an event travels between processes (:mod:`bus`), which
channels a socket may subscribe to (:mod:`authz`), and the wire protocol
(:mod:`router`).
The event contract and the typed publishers live one layer in, in
``app.application.ports.realtime`` — services depend on the port, this package
implements it. ``app.main`` is the composition root that installs the bus as
the process publisher and starts the cross-process relay.
"""
from app.interfaces.realtime.bus import get_bus, start_relay, stop_relay
__all__ = [
"get_bus",
"start_relay",
"stop_relay",
]
+112
View File
@@ -0,0 +1,112 @@
"""Channel authorization — the only gate on what a socket may receive.
Authorization happens once, at subscribe time. After that the connection holds
plain channel names and the manager routes on them, so there is no code path
that can deliver an event to a connection that was never checked.
Every rule resolves the channel's tenant through the repositories and then the
caller's membership in it — exactly like the HTTP layer's ``get_org_context``.
A tenant that does not own the object gets a rejection, not an empty stream:
the difference matters when debugging "why is my page not live?".
The caller's identity arrives as two primitives (id, platform-owner flag) —
what an authenticated socket actually knows — rather than a ``User`` row, so a
subscribe costs one fewer database read.
"""
from __future__ import annotations
from sqlalchemy.ext.asyncio import AsyncSession
from app.infrastructure.repositories.identity import MembershipRepository
from app.infrastructure.repositories.sandbox import (
SandboxSessionRepository,
WorkspaceRepository,
)
from app.infrastructure.repositories.workforce import (
RunRepository,
TaskRepository,
ThreadRepository,
)
async def authorize_channel(
session: AsyncSession,
*,
user_id: str,
is_platform_owner: bool,
channel: str,
) -> bool:
"""May this identity subscribe to ``channel``? Deny by default."""
kind, _, identifier = channel.partition(":")
if not identifier and kind != "admin":
return False
if kind == "admin":
return is_platform_owner
if kind == "user":
return identifier == user_id
if kind == "org":
return await _is_member(
session=session, user_id=user_id, organization_id=identifier
)
if kind == "task":
return await _task_is_visible(
session=session, user_id=user_id, task_id=identifier
)
if kind == "run":
return await _run_is_visible(
session=session, user_id=user_id, run_id=identifier
)
if kind == "thread":
return await _thread_is_visible(
session=session, user_id=user_id, thread_id=identifier
)
if kind == "session":
return await _session_is_visible(
session=session, user_id=user_id, session_id=identifier
)
return False
async def _is_member(*, session: AsyncSession, user_id: str, organization_id: str) -> bool:
membership = await MembershipRepository(session).get(organization_id, user_id)
return membership is not None and membership.is_active
async def _task_is_visible(*, session: AsyncSession, user_id: str, task_id: str) -> bool:
task = await TaskRepository(session).get(task_id)
if task is None:
return False
return await _is_member(
session=session, user_id=user_id, organization_id=task.organization_id
)
async def _run_is_visible(*, session: AsyncSession, user_id: str, run_id: str) -> bool:
run = await RunRepository(session).get(run_id)
if run is None:
return False
return await _task_is_visible(session=session, user_id=user_id, task_id=run.task_id)
async def _thread_is_visible(*, session: AsyncSession, user_id: str, thread_id: str) -> bool:
thread = await ThreadRepository(session).get(thread_id)
if thread is None:
return False
return await _is_member(
session=session, user_id=user_id, organization_id=thread.organization_id
)
async def _session_is_visible(*, session: AsyncSession, user_id: str, session_id: str) -> bool:
"""A sandbox session belongs to a workspace, which belongs to a tenant."""
sandbox_session = await SandboxSessionRepository(session).get(session_id)
if sandbox_session is None:
return False
workspace = await WorkspaceRepository(session).get(sandbox_session.workspace_id)
if workspace is None:
return False
return await _is_member(
session=session, user_id=user_id, organization_id=workspace.organization_id
)
+260
View File
@@ -0,0 +1,260 @@
"""The event bus: how a published event reaches connected browsers.
Two transports, in this order:
1. **In-process.** Always on. When the API itself records the event (a task run
driven by a button, a sandbox command, an error handler), the fan-out is a
direct call into the ``ConnectionManager`` — microseconds, no broker.
2. **Redis pub/sub.** Carries events published by *other* processes: the
standalone workers (``pnpm worker:ai-runtime``) run tasks in their own
process, and without a broker their events would die there. The API process
relays whatever arrives on the shared channel into the same manager.
Redis is a bonus, never a dependency:
* The bus tries it once per cooldown window. If nothing is listening, the
attempt fails quietly and in-process delivery continues — the product stays
fully live for every API-driven action, only cross-process events are lost.
* The relay task owns the subscription lifecycle; it starts with the app and
stops on shutdown, and reconnects on its own after a broker restart.
Nothing here ever raises. Publishing is called from request handlers, from
inside exception handlers and from worker passes; an exception in the feed
must not be able to fail any of those.
"""
from __future__ import annotations
import asyncio
import logging
import time
from collections.abc import Awaitable, Callable
from contextlib import suppress
from app.application.ports.realtime import RealtimeEvent, set_publisher
from app.core.config import settings
logger = logging.getLogger("agy.realtime")
#: Redis pub/sub channel shared by every process of one deployment.
DEFAULT_REDIS_CHANNEL = "agy:realtime"
#: After a failed broker attempt, wait this long before trying again. Keeps a
#: dev machine without Redis from adding a connection timeout to every event.
REDIS_RETRY_COOLDOWN_SECONDS = 10.0
#: How long a single connection attempt may take before it is abandoned.
REDIS_CONNECT_TIMEOUT_SECONDS = 1.0
#: Listener signature: ``await listener(event)``. The ConnectionManager
#: registers one; tests register fakes. The return value is ignored so a
#: listener may report how many connections it fanned out to.
EventListener = Callable[[RealtimeEvent], Awaitable[object]]
class RealtimeBus:
"""Fan-out hub shared by every publisher in this process."""
def __init__(
self,
*,
redis_url: str | None = None,
channel: str = DEFAULT_REDIS_CHANNEL,
) -> None:
self._listeners: set[EventListener] = set()
self._redis_url = redis_url if redis_url is not None else settings.redis_url
self._channel = channel
self._redis = None
self._pubsub = None
self._relay_task: asyncio.Task | None = None
self._next_redis_attempt = 0.0
self._redis_healthy = False
# -- listeners (in-process fan-out) --------------------------------------
def attach(self, listener: EventListener) -> None:
self._listeners.add(listener)
def detach(self, listener: EventListener) -> None:
self._listeners.discard(listener)
# -- publishing -----------------------------------------------------------
async def publish(self, event: RealtimeEvent) -> None:
"""Deliver locally, then mirror to other processes. Never raises."""
try:
await self._dispatch_local(event)
except Exception: # noqa: BLE001 - the feed must never fail a caller
logger.exception("realtime: local dispatch failed for %s", event.type)
if not self._redis_enabled():
return
try:
await self._publish_redis(event)
except Exception: # noqa: BLE001 - same rule
logger.exception("realtime: redis publish failed for %s", event.type)
async def _dispatch_local(self, event: RealtimeEvent) -> None:
for listener in list(self._listeners):
try:
await listener(event)
except Exception: # noqa: BLE001 - one bad listener is not fatal
logger.exception("realtime: listener failed for %s", event.type)
def _redis_enabled(self) -> bool:
# "local" pins the bus to in-process delivery — used by tests and by
# single-process deployments that have no broker on purpose.
return getattr(settings, "realtime_transport", "auto").lower() != "local"
async def _publish_redis(self, event: RealtimeEvent) -> None:
client = await self._redis_client()
if client is None:
return
await client.publish(self._channel, event.to_json())
async def _redis_client(self):
"""Lazily connect, with a cooldown after failure. ``None`` if down."""
if self._redis is not None:
return self._redis
now = time.monotonic()
if now < self._next_redis_attempt:
return None
self._next_redis_attempt = now + REDIS_RETRY_COOLDOWN_SECONDS
try:
import redis.asyncio as aioredis
client = aioredis.from_url(
self._redis_url,
encoding="utf-8",
decode_responses=True,
socket_connect_timeout=REDIS_CONNECT_TIMEOUT_SECONDS,
socket_timeout=REDIS_CONNECT_TIMEOUT_SECONDS,
)
await client.ping()
self._redis = client
self._redis_healthy = True
logger.info("realtime: redis pub/sub connected (%s)", self._channel)
except Exception as exc: # noqa: BLE001 - broker is optional
self._redis = None
self._redis_healthy = False
logger.debug("realtime: redis unavailable (%s); in-process only", exc)
return None
return self._redis
# -- relay (other processes -> this one's browsers) ----------------------
async def start_relay(self) -> None:
"""Subscribe to the shared channel and feed the local manager."""
if self._relay_task is not None and not self._relay_task.done():
return
self._relay_task = asyncio.create_task(self._relay_loop(), name="realtime-relay")
async def stop_relay(self) -> None:
task = self._relay_task
self._relay_task = None
if task is not None:
task.cancel()
with suppress(asyncio.CancelledError):
await task
await self._close_redis()
async def _relay_loop(self) -> None:
while True:
client = await self._redis_client()
if client is None:
# Broker absent (or down mid-flight): wait out the cooldown and
# try again — a Redis started after the API must still work.
await asyncio.sleep(REDIS_RETRY_COOLDOWN_SECONDS)
continue
try:
pubsub = client.pubsub()
await pubsub.subscribe(self._channel)
self._pubsub = pubsub
logger.info("realtime: relay subscribed to %s", self._channel)
async for message in pubsub.listen():
if message.get("type") != "message":
continue
event = _decode_message(message.get("data"))
if event is not None:
# Dedupe by id happens in the manager, so an event this
# process published itself is not delivered twice.
await self._dispatch_local(event)
except asyncio.CancelledError:
raise
except Exception as exc: # noqa: BLE001 - relay must survive
self._redis_healthy = False
logger.warning("realtime: relay interrupted (%s); reconnecting", exc)
finally:
if self._pubsub is not None:
with suppress(Exception):
await self._pubsub.unsubscribe(self._channel)
self._pubsub = None
self._redis = None
await asyncio.sleep(1.0)
async def _close_redis(self) -> None:
client, self._redis = self._redis, None
if client is not None:
with suppress(Exception):
await client.aclose()
self._redis_healthy = False
# -- introspection -------------------------------------------------------
@property
def redis_healthy(self) -> bool:
return self._redis_healthy
@property
def listener_count(self) -> int:
return len(self._listeners)
def _decode_message(raw: object) -> RealtimeEvent | None:
if not isinstance(raw, (str, bytes)):
return None
try:
return RealtimeEvent.from_json(raw)
except Exception: # noqa: BLE001 - a bad frame is skipped, not fatal
logger.warning("realtime: undecodable relay frame dropped")
return None
_bus: RealtimeBus | None = None
def get_bus() -> RealtimeBus:
"""Process-wide bus. Created lazily so importing never opens a socket."""
global _bus
if _bus is None:
_bus = RealtimeBus()
return _bus
def install() -> RealtimeBus:
"""Make this process's bus the application-layer publisher.
Called by the composition roots (``app.main`` for the API,
``app.workers.__main__`` for the workers). Idempotent, so a worker that
imports several modules pays nothing for calling it more than once.
"""
bus = get_bus()
set_publisher(bus)
return bus
async def publish(event: RealtimeEvent) -> None:
"""Publish through the shared bus. Never raises."""
bus = get_bus()
try:
await bus.publish(event)
except Exception: # noqa: BLE001 - belt and braces; publish is best-effort
logger.exception("realtime: publish failed for %s", event.type)
async def start_relay() -> None:
await get_bus().start_relay()
async def stop_relay() -> None:
await get_bus().stop_relay()
+217
View File
@@ -0,0 +1,217 @@
"""Connection bookkeeping: who is connected, and to which channels.
The manager is deliberately transport-agnostic — it does not know it is talking
to a WebSocket, only to an object with ``send_text``. That keeps it testable
without a socket and lets a future transport (SSE, push bridge) reuse it.
Two invariants:
* **A connection only ever receives channels it was authorized for.** The
authorizer runs at subscribe time; the routing table built afterwards can
only contain pre-approved channel names.
* **An event is delivered at most once.** ``dispatch`` de-duplicates by event
id, because the same event reaches a process twice in the normal case: once
in-process and once relayed from Redis.
"""
from __future__ import annotations
import asyncio
import logging
from collections import deque
from collections.abc import Awaitable, Callable
from contextlib import suppress
from dataclasses import dataclass, field
from starlette.websockets import WebSocket, WebSocketState
from app.application.ports.realtime import RealtimeEvent
logger = logging.getLogger("agy.realtime")
#: Bounded dedupe memory. Large enough that a relayed duplicate is always
#: caught, small enough that a chatty day costs a few kilobytes.
MAX_SEEN_EVENTS = 2048
#: Per-connection outbox. If a browser stops reading (a frozen tab), a
#: producer must not be able to grow this without bound. Overflow drops the
#: oldest frame — ordering of what follows stays meaningful.
MAX_QUEUE_DEPTH = 256
#: What a subscribe check looks like: ``(connection, channel) -> allowed``.
ChannelAuthorizer = Callable[["Connection", str], Awaitable[bool]]
#: ``eq=False`` on purpose: a connection is identified by object identity. The
#: default dataclass ``__eq__`` would also make it unhashable (a ``set`` field
#: cannot be compared safely), and the manager keys connections by identity in
#: both its connection set and its channel routing table.
@dataclass(slots=True, eq=False)
class Connection:
"""One authenticated browser socket."""
user_id: str
is_platform_owner: bool
websocket: WebSocket
channels: set[str] = field(default_factory=set)
_queue: asyncio.Queue[str] = field(
default_factory=lambda: asyncio.Queue(maxsize=MAX_QUEUE_DEPTH)
)
_sender: asyncio.Task | None = None
async def start(self) -> None:
"""Single writer task per connection.
One task owns the socket: concurrent ``send_text`` calls on one
WebSocket interleave frames and corrupt the stream, so everything goes
through this queue instead.
"""
if self._sender is None or self._sender.done():
self._sender = asyncio.create_task(self._drain(), name="realtime-sender")
async def _drain(self) -> None:
while True:
payload = await self._queue.get()
if payload is None: # shutdown sentinel
return
try:
if self.websocket.client_state == WebSocketState.CONNECTED:
await self.websocket.send_text(payload)
except Exception: # noqa: BLE001 - a dead socket ends the connection
logger.debug("realtime: send failed; closing connection", exc_info=True)
return
def offer(self, payload: str) -> None:
"""Queue a frame, never blocking the publisher.
A slow consumer loses the oldest frames rather than stalling every
other tenant's stream behind it.
"""
if self._queue.full():
with suppress(asyncio.QueueFull):
# Drop the oldest frame to make room: ordering of what follows
# stays meaningful, which is what the UI renders.
self._queue.get_nowait()
with suppress(asyncio.QueueFull): # pragma: no cover - defensive
self._queue.put_nowait(payload)
async def close(self) -> None:
if self._sender is not None:
self._sender.cancel()
self._sender = None
class ConnectionManager:
"""The routing table between events and open sockets."""
def __init__(self) -> None:
self._connections: set[Connection] = set()
self._by_channel: dict[str, set[Connection]] = {}
self._seen: deque[str] = deque(maxlen=MAX_SEEN_EVENTS)
self._seen_set: set[str] = set()
# -- lifecycle -----------------------------------------------------------
async def register(self, connection: Connection) -> None:
self._connections.add(connection)
await connection.start()
async def unregister(self, connection: Connection) -> None:
self._connections.discard(connection)
for channel in list(connection.channels):
subscribers = self._by_channel.get(channel)
if subscribers is not None:
subscribers.discard(connection)
if not subscribers:
self._by_channel.pop(channel, None)
connection.channels.clear()
await connection.close()
# -- subscriptions ---------------------------------------------------------
async def subscribe(
self,
connection: Connection,
channels: list[str],
authorizer: ChannelAuthorizer,
) -> tuple[list[str], list[str]]:
"""Authorize and subscribe. Returns ``(accepted, rejected)``."""
accepted: list[str] = []
rejected: list[str] = []
for channel in channels:
if channel in connection.channels:
accepted.append(channel)
continue
try:
allowed = await authorizer(connection, channel)
except Exception: # noqa: BLE001 - a broken check denies, never allows
logger.exception("realtime: channel authorization crashed")
allowed = False
if not allowed:
rejected.append(channel)
continue
connection.channels.add(channel)
self._by_channel.setdefault(channel, set()).add(connection)
accepted.append(channel)
return accepted, rejected
async def unsubscribe(self, connection: Connection, channels: list[str]) -> None:
for channel in channels:
connection.channels.discard(channel)
subscribers = self._by_channel.get(channel)
if subscribers is not None:
subscribers.discard(connection)
if not subscribers:
self._by_channel.pop(channel, None)
# -- fan-out ---------------------------------------------------------------
async def dispatch(self, event: RealtimeEvent) -> int:
"""Deliver to every subscriber of every channel. Returns fan-out count.
The id check is what makes "publish locally, then receive the same event
from Redis" deliver once instead of twice.
"""
if event.id in self._seen_set:
return 0
self._remember(event.id)
recipients: set[Connection] = set()
for channel in event.channels:
recipients.update(self._by_channel.get(channel, ()))
envelope = event.envelope()
# Control frames and event frames share one top-level discriminator
# (``type``), so the client switches on one field for everything.
payload = _dumps({"type": "event", "event": envelope})
delivered = 0
for connection in recipients:
connection.offer(payload)
delivered += 1
return delivered
def _remember(self, event_id: str) -> None:
if len(self._seen) == self._seen.maxlen:
evicted = self._seen[0]
self._seen_set.discard(evicted)
self._seen.append(event_id)
self._seen_set.add(event_id)
# -- introspection -------------------------------------------------------
@property
def connection_count(self) -> int:
return len(self._connections)
@property
def channel_count(self) -> int:
return len(self._by_channel)
def channel_load(self) -> dict[str, int]:
return {channel: len(subs) for channel, subs in self._by_channel.items()}
def _dumps(envelope: dict) -> str:
import json
return json.dumps(envelope, ensure_ascii=False, default=str)
+175
View File
@@ -0,0 +1,175 @@
"""The WebSocket endpoint — one socket per browser, every live feed on it.
``GET /api/ws?token=<jwt>``
The token travels as a query parameter because the browser WebSocket API cannot
set an ``Authorization`` header. That is a deliberate, contained exception to
the header-only rule: the token is short-lived, the socket is TLS-terminated in
any real deployment, and the alternative (a cookie) would drag CSRF into a
protocol that has no CSRF story.
Protocol — small, symmetric, JSON in both directions:
.. code-block:: text
client -> {"action": "subscribe", "channels": ["org:…", "task:…"]}
client -> {"action": "unsubscribe", "channels": ["task:…"]}
client -> {"action": "ping"}
server -> {"type": "hello", "user_id": "…"}
server -> {"type": "subscribed", "accepted": [...], "rejected": [...]}
server -> {"type": "event", "event": {…envelope…}}
server -> {"type": "pong"}
server -> {"type": "error", "message": "…"}
A rejected channel is reported, not hidden: "my task page is not live" must be
answerable from the browser console alone.
Authorization queries open their own short-lived session: a socket can stay
open for hours, and one SQLAlchemy session per open tab would quietly exhaust
the connection pool.
"""
from __future__ import annotations
import json
import logging
from fastapi import APIRouter, Query, WebSocket, WebSocketDisconnect
from app.application.ports.realtime import dedupe_channels
from app.core.database import SessionFactory
from app.core.errors import AuthenticationError
from app.core.security import decode_access_token
from app.infrastructure.repositories.identity import UserRepository
from app.interfaces.realtime.authz import authorize_channel
from app.interfaces.realtime.manager import Connection, ConnectionManager
logger = logging.getLogger("agy.realtime")
router = APIRouter()
#: The manager instance this process fans out with. Created at import: the
#: endpoint and the lifecycle hooks can share it without a wiring ceremony.
_manager = ConnectionManager()
def realtime_manager() -> ConnectionManager:
"""The process-wide manager — attached to the bus by ``app.main``."""
return _manager
@router.websocket("/ws")
async def realtime_socket(
websocket: WebSocket,
token: str | None = Query(default=None),
) -> None:
connection = await _authenticate(websocket, token)
if connection is None:
return # _authenticate already closed with a policy-violation code
await _manager.register(connection)
await _send(
websocket,
{"type": "hello", "user_id": connection.user_id, "server_time": _now_iso()},
)
try:
while True:
raw = await websocket.receive_text()
await _handle_message(websocket, connection, raw)
except WebSocketDisconnect:
pass
except Exception: # noqa: BLE001 - any read failure ends the socket cleanly
logger.debug("realtime: socket loop ended", exc_info=True)
finally:
await _manager.unregister(connection)
async def _authenticate(websocket: WebSocket, token: str | None) -> Connection | None:
"""Accept only a valid, active user. Closes 1008 otherwise."""
user = None
if token:
try:
payload = decode_access_token(token)
user_id = payload.get("sub")
if user_id:
# Own session: the socket must not depend on request scope.
async with SessionFactory() as session:
user = await UserRepository(session).get(str(user_id))
except AuthenticationError:
user = None
if user is None or not user.is_active:
# 1008 (policy violation) rather than a bare close: the client's
# reconnect logic can then tell "bad token" from "network dropped".
await websocket.close(code=1008, reason="unauthorized")
return None
await websocket.accept()
return Connection(
user_id=user.id,
is_platform_owner=user.is_platform_owner,
websocket=websocket,
)
async def _handle_message(websocket: WebSocket, connection: Connection, raw: str) -> None:
try:
message = json.loads(raw)
except (TypeError, ValueError):
await _send(websocket, {"type": "error", "message": "malformed JSON frame"})
return
if not isinstance(message, dict):
await _send(websocket, {"type": "error", "message": "frame must be an object"})
return
action = message.get("action")
if action == "ping":
await _send(websocket, {"type": "pong"})
return
channels = dedupe_channels(
str(channel) for channel in (message.get("channels") or []) if isinstance(channel, str)
)
if action == "subscribe":
accepted, rejected = await _subscribe(connection, channels)
await _send(
websocket, {"type": "subscribed", "accepted": accepted, "rejected": rejected}
)
return
if action == "unsubscribe":
await _manager.unsubscribe(connection, list(channels))
await _send(websocket, {"type": "unsubscribed", "channels": channels})
return
await _send(websocket, {"type": "error", "message": f"unknown action: {action!r}"})
async def _subscribe(
connection: Connection, channels: tuple[str, ...]
) -> tuple[list[str], list[str]]:
if not channels:
return [], []
async def _authorize(conn: Connection, channel: str) -> bool:
async with SessionFactory() as session:
return await authorize_channel(
session,
user_id=conn.user_id,
is_platform_owner=conn.is_platform_owner,
channel=channel,
)
return await _manager.subscribe(connection, list(channels), _authorize)
def _now_iso() -> str:
from app.domain.base import utcnow
return utcnow().isoformat()
async def _send(websocket: WebSocket, payload: dict) -> None:
await websocket.send_text(json.dumps(payload, ensure_ascii=False, default=str))