feat(cloud): enforce host governance budgets

tests/test_host_agent_internal_api.py

@@ -12,9 +12,15 @@ from cloud.internal_api.api import create_internal_router
 from cloud.pool import DevicePool, PooledDevice
 from cloud.scheduler import ScheduledTask, TaskConstraints, TaskScheduler
 from cloud.store import CloudStore
+from runtime.tool_calling_client import ToolCallDecision, ToolCallUsage
 
 
-def _build_client(tmp_path) -> tuple[TestClient, DevicePool]:
+def _build_client(
+    tmp_path,
+    *,
+    planner_client_factory=None,
+    planner_token_reservation_ceiling: int = 4096,
+) -> tuple[TestClient, DevicePool]:
     pool = DevicePool(
         CloudStore(tmp_path / "internal.sqlite3"),
         CloudConfig(stale_after_seconds=60),
@@ -39,6 +45,8 @@ def _build_client(tmp_path) -> tuple[TestClient, DevicePool]:
             pool=pool,
             auth_provider=auth_provider,
             scheduler=TaskScheduler(pool, pool.store, CloudConfig(stale_after_seconds=60)),
+            planner_client_factory=planner_client_factory,
+            planner_token_reservation_ceiling=planner_token_reservation_ceiling,
         )
     )
     return TestClient(app), pool
@@ -232,6 +240,58 @@ def test_heartbeat_and_self_submission_preserve_host_isolation(tmp_path) -> None
     assert foreign.status_code == 403
 
 
+def test_planner_proxy_reserves_and_enforces_host_daily_token_budget(tmp_path) -> None:
+    class FakePlannerClient:
+        calls = 0
+
+        def decide(self, **_kwargs):
+            self.calls += 1
+            return ToolCallDecision(
+                tool_name="tap",
+                arguments={"x": 1, "y": 2},
+                usage=ToolCallUsage(input_tokens=1, output_tokens=1, total_tokens=2),
+            )
+
+    planner = FakePlannerClient()
+    client, pool = _build_client(
+        tmp_path,
+        planner_client_factory=lambda: planner,
+        planner_token_reservation_ceiling=5,
+    )
+    headers = {"Authorization": "Bearer token-a"}
+    client.put(
+        "/internal/v1/hosts/host-a/heartbeat",
+        headers=headers,
+        json=_heartbeat_payload("host-a", "device-a"),
+    )
+    pool.store.upsert_host_governance_policy(
+        host_id="host-a",
+        self_submission_enabled=True,
+        max_active_tasks=None,
+        daily_token_budget=5,
+        updated_at=datetime.now(UTC),
+    )
+    payload = {
+        "host_id": "host-a",
+        "system_prompt": "system",
+        "user_prompt": "user",
+        "tools": [{"name": "tap", "description": "tap", "parameters": {}}],
+        "timeout_seconds": 1,
+    }
+
+    first = client.post(
+        "/internal/v1/hosts/host-a/planner/decide", headers=headers, json=payload
+    )
+    second = client.post(
+        "/internal/v1/hosts/host-a/planner/decide", headers=headers, json=payload
+    )
+
+    assert first.status_code == 200, first.text
+    assert first.json()["total_tokens"] == 2
+    assert second.status_code == 429
+    assert planner.calls == 1
+
+
 def test_long_poll_claim_returns_at_most_one_owned_assignment(tmp_path) -> None:
     client, pool = _build_client(tmp_path)
     now = datetime.now(UTC)