Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

test(inference): add budget wrapping test for turn models

Steven Enamakel433b98e81136
diff --git a/crates/openhuman-core/src/inference/turn_models_tests.rs b/crates/openhuman-core/src/inference/turn_models_tests.rs
index a5312bb75b..abc6f116c3 100644
--- a/crates/openhuman-core/src/inference/turn_models_tests.rs
+++ b/crates/openhuman-core/src/inference/turn_models_tests.rs
@@ -53,6 +53,86 @@ async fn native_turn_factory_applies_default_temperature_and_preserves_explicit_
assert_eq!(*seen.lock().unwrap(), vec![Some(0.3), Some(0.9)]);
}
+#[tokio::test]
+async fn native_budget_wrapping_keeps_owner_error_slots_for_primary_tiers_and_summary() {
+ use crate::agent::tinyagents::budget::{CallBudget, ModelBudget, SpendLimits};
+ use crate::core::runtime::{CoreContext, DispatchCtx, DomainSet};
+ use crate::core::slots::Extensions;
+
+ let _guard = crate::inference::inference_test_guard_async().await;
+ let extensions = Extensions::new();
+ crate::inference::registry::register_chat_model_in(
+ &extensions,
+ crate::inference::provider::factory::PROVIDER_OPENHUMAN,
+ scoped_model_entry("budget-owner-model", false),
+ );
+ let context = CoreContext::for_test_with_host_kind_and_extensions(
+ crate::core::types::HostKind::Library,
+ DomainSet::full(),
+ None,
+ extensions,
+ );
+ let ctx = DispatchCtx::new(context);
+ let config = Arc::new(crate::config::Config::default());
+ let budget = ModelBudget {
+ ledger: tinyinference_llm::model::budget::Budget::new(SpendLimits {
+ tokens: Some(0),
+ cost_micros: None,
+ }),
+ call: CallBudget {
+ input_tokens: 10_000,
+ output_tokens: 1,
+ cost_micros: 1,
+ },
+ };
+ let models = TurnModelSource::new_crate_native(&ctx, "chat", config)
+ .build("hint:chat", 0.0, None, None)
+ .expect("native model set")
+ .with_budget(Some(&budget));
+ let owner = models
+ .native_model_set
+ .as_ref()
+ .expect("native owner set retained");
+ let primary_slot = owner.primary_error_slot();
+ let tier_slot = owner
+ .workload_error_slot("hint:coding")
+ .expect("coding route owns its error slot");
+ let summary_slot = owner.summarizer_error_slot();
+ let resolver = models.resolver();
+
+ for (request, expected_slot) in [
+ (ModelResolveRequest::new("chat"), primary_slot),
+ (
+ ModelResolveRequest::new("worker").with_model_pin("hint:coding"),
+ tier_slot,
+ ),
+ ] {
+ let model = resolver.resolve(&request).await.expect("route resolves");
+ let result = model
+ .invoke(
+ &(),
+ ModelRequest::new(vec![tinyinference_llm::message::Message::user(
+ "budget this call",
+ )]),
+ )
+ .await;
+ assert!(result.is_err(), "zero token budget rejects each route");
+ assert!(expected_slot.lock().unwrap().is_some());
+ }
+
+ let result = owner
+ .summarizer()
+ .invoke(
+ &(),
+ ModelRequest::new(vec![tinyinference_llm::message::Message::user(
+ "budget this summary",
+ )]),
+ )
+ .await;
+ assert!(result.is_err(), "zero token budget rejects summary");
+ assert!(summary_slot.lock().unwrap().is_some());
+}
+
use super::TurnModels;
use crate::agent::tinyagents::TurnModelSource;
use tinyagents_harness::host::PrebuiltModelResolver as TurnModelResolver;