fix(agents): exclude discovery from invocation billing

litellm/proxy/agent_endpoints/auth/managed_authorization.py

@@ -240,7 +240,7 @@ async def prepare_agent_invocation(
     if not await AgentRequestHandler.is_agent_allowed(effective.agent_id, auth):
         raise_identity_failure(AgentIdentityFailure(message="The caller is not permitted to invoke this agent"))
     auth.invoked_agent_id = effective.agent_id
-    if auth.agent_id is None and (effective.identity_managed or effective.litellm_budget_table is not None):
+    if billable and auth.agent_id is None and (effective.identity_managed or effective.litellm_budget_table is not None):
         auth.billing_agent_policy = effective
     raw_fee: Final = (effective.litellm_params or MappingProxyType({})).get("cost_per_query", 0.0) if billable else 0.0
     try:

litellm/proxy/auth/user_api_key_auth.py

@@ -3244,7 +3244,8 @@ async def _authorize_authenticated_request(
                 user_api_key_auth_obj,
                 target_name,
                 store,
-                billable=request_data.get("method")
+                billable=request.method == "POST"
+                and request_data.get("method")
                 in (None, "message/send", "message/stream", "SendMessage", "SendStreamingMessage"),
             )
         await _run_centralized_common_checks(

tests/test_litellm/proxy/agent_endpoints/auth/test_managed_authorization.py

@@ -129,9 +129,11 @@ async def test_agent_budget_accumulates_across_credentials_and_denies_the_next_a
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("autonomous", (True, False))
+@pytest.mark.parametrize("billable", (True, False))
 async def test_invocation_prepares_target_fee_for_the_correct_agent(
     monkeypatch: pytest.MonkeyPatch,
     autonomous: bool,
+    billable: bool,
 ) -> None:
     from unittest.mock import AsyncMock, MagicMock
 
@@ -158,11 +160,14 @@ async def test_invocation_prepares_target_fee_for_the_correct_agent(
         caller: Final = agent(agent_id="caller", object_permission=permission.model_dump())
         auth.managed_agent_policy = caller
         auth.billing_agent_policy = caller
-    await prepare_agent_invocation(auth, "agent", AgentIdentityStore.from_client(database))
-    assert auth.agent_invocation_cost == pytest.approx(0.25)
+    await prepare_agent_invocation(auth, "agent", AgentIdentityStore.from_client(database), billable=billable)
+    assert auth.agent_invocation_cost == pytest.approx(0.25 if billable else 0.0)
     assert auth.invoked_agent_id == "agent"
-    assert auth.billing_agent_policy is not None
-    assert auth.billing_agent_policy.agent_id == ("caller" if autonomous else "agent")
+    if autonomous or billable:
+        assert auth.billing_agent_policy is not None
+        assert auth.billing_agent_policy.agent_id == ("caller" if autonomous else "agent")
+    else:
+        assert auth.billing_agent_policy is None
 
 
 @pytest.mark.asyncio