feat: add per-IP rate limiting to WS LLM triggers

Per-account limits alone don't stop one account script-hitting
Ollama from many source IPs. Adds matching per-IP limiters
alongside the existing per-account ones for fragment/question/
summon triggers, per spec's dual per-account-and-per-IP requirement.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013PphXq1s43DNRj1uWKGXof

backend/app/ws.py

@@ -53,6 +53,16 @@ fragment_limiter = RateLimiter(max_requests=30, window_seconds=60)
 question_limiter = RateLimiter(max_requests=6, window_seconds=60)
 summon_limiter = RateLimiter(max_requests=4, window_seconds=60)
 
+# Per-IP limiters for the same trigger points (spec §5). Ollama is a shared,
+# single-instance, CPU-only resource — per-account limits alone don't stop
+# one compromised/scripted account from hammering it across many source IPs,
+# nor do they protect the resource from many accounts sharing one IP. IP
+# thresholds are looser than the per-account ones since a single IP can
+# legitimately host multiple accounts on a shared network.
+fragment_ip_limiter = RateLimiter(max_requests=60, window_seconds=60)
+question_ip_limiter = RateLimiter(max_requests=12, window_seconds=60)
+summon_ip_limiter = RateLimiter(max_requests=8, window_seconds=60)
+
 AUDIO_DIR = Path(settings.data_dir) / "audio"
 
 
@@ -65,6 +75,7 @@ def _audio_dir() -> Path:
 class SeanceState:
     user_id: uuid.UUID
     session_id: uuid.UUID
+    client_ip: str
     send_queue: asyncio.Queue = field(default_factory=asyncio.Queue)
     mode: str = "unknown"
     language: str = "en"
@@ -91,6 +102,10 @@ def serialize_entity(entity: Entity) -> dict:
     }
 
 
+def _client_ip(websocket: WebSocket) -> str:
+    return websocket.client.host if websocket.client else "unknown"
+
+
 async def _authenticate(websocket: WebSocket) -> uuid.UUID | None:
     raw_token = websocket.cookies.get(SESSION_COOKIE_NAME)
     if raw_token is None:
@@ -227,7 +242,12 @@ async def _summon(state: SeanceState, channel: str) -> tuple[Entity, bool]:
 
 
 async def _handle_summon(state: SeanceState) -> None:
-    if not summon_limiter.allow(str(state.user_id)):
+    # Short-circuits: an account already over its own cap never gets far
+    # enough to spend from the IP budget too.
+    if not (
+        summon_limiter.allow(str(state.user_id))
+        and summon_ip_limiter.allow(state.client_ip)
+    ):
         await state.send_queue.put(
             {
                 "type": "error",
@@ -266,7 +286,10 @@ async def _handle_anomaly(state: SeanceState, message: dict) -> None:
             await state.send_queue.put({"type": "status", "state": "attuning"})
         return
 
-    if not fragment_limiter.allow(str(state.user_id)):
+    if not (
+        fragment_limiter.allow(str(state.user_id))
+        and fragment_ip_limiter.allow(state.client_ip)
+    ):
         return  # anomalies during a crowded veil just pass unheard
 
     try:
@@ -279,7 +302,10 @@ async def _handle_anomaly(state: SeanceState, message: dict) -> None:
 
 
 async def _handle_question(state: SeanceState, text: str) -> None:
-    if not question_limiter.allow(str(state.user_id)):
+    if not (
+        question_limiter.allow(str(state.user_id))
+        and question_ip_limiter.allow(state.client_ip)
+    ):
         await state.send_queue.put(
             {
                 "type": "error",
@@ -390,7 +416,11 @@ async def session_socket(websocket: WebSocket) -> None:
         await db.commit()
         await db.refresh(contact_session)
 
-    state = SeanceState(user_id=user_id, session_id=contact_session.id)
+    state = SeanceState(
+        user_id=user_id,
+        session_id=contact_session.id,
+        client_ip=_client_ip(websocket),
+    )
     sender = asyncio.create_task(_sender(state, websocket))
     await state.send_queue.put({"type": "session", "id": str(contact_session.id)})
 

backend/tests/test_ws_session.py

@@ -7,6 +7,7 @@ import app.ws
 from app.entities import fallback_profile
 from app.models.contact_session import ContactSession
 from app.models.event import Event
+from app.rate_limit import RateLimiter
 
 
 class FakeSpiritService:
@@ -175,3 +176,59 @@ async def test_same_signature_recontacts_same_entity(sync_client):
         assert second_frame["entity"]["name"] == first
         assert second_frame["is_new"] is False
         assert second_frame["entity"]["contact_count"] == 2
+
+
+@pytest.mark.asyncio
+async def test_summon_rate_limited_per_account(sync_client, monkeypatch):
+    # Swap in a tight, test-scoped limiter so this doesn't depend on (or
+    # pollute) the shared module-level budget other tests draw from.
+    monkeypatch.setattr(
+        app.ws, "summon_limiter", RateLimiter(max_requests=2, window_seconds=60)
+    )
+    _login(sync_client, "account-limited")
+
+    with _ws_connect(sync_client, sync_client.cookies.get("qm_session")) as ws:
+        _read_until(ws, "session")
+        for _ in range(2):
+            ws.send_json({"type": "summon"})
+            _read_until(ws, "entity")
+            _read_until(ws, "utterance", kind="greeting")
+
+        ws.send_json({"type": "summon"})
+        rejection = _read_until(ws, "error")
+        assert rejection["code"] == "rate_limited"
+        assert rejection["message"] == (
+            "The veil is crowded. The spirits need a moment before another summoning."
+        )
+
+
+@pytest.mark.asyncio
+async def test_summon_rate_limited_per_ip_even_with_fresh_account(
+    sync_client, monkeypatch
+):
+    # Starve only the IP bucket; the per-account limiter stays at its
+    # production default so each account below has plenty of its own budget
+    # left. This proves the IP limiter alone can reject a request — the
+    # gap the per-account-only limiters left open.
+    monkeypatch.setattr(
+        app.ws, "summon_ip_limiter", RateLimiter(max_requests=1, window_seconds=60)
+    )
+
+    _login(sync_client, "ip-limited-a")
+    with _ws_connect(sync_client, sync_client.cookies.get("qm_session")) as ws:
+        _read_until(ws, "session")
+        ws.send_json({"type": "summon"})
+        _read_until(ws, "entity")  # spends the single per-IP slot
+
+    # A different account — its own per-account budget is untouched — but
+    # every connection in this test shares the same (simulated) source IP,
+    # which is already spent.
+    _login(sync_client, "ip-limited-b")
+    with _ws_connect(sync_client, sync_client.cookies.get("qm_session")) as ws:
+        _read_until(ws, "session")
+        ws.send_json({"type": "summon"})
+        rejection = _read_until(ws, "error")
+        assert rejection["code"] == "rate_limited"
+        assert rejection["message"] == (
+            "The veil is crowded. The spirits need a moment before another summoning."
+        )