Skip to content

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

chore: files changed crates/openhuman-core/src/flows/caps_host/agent_tests.rs

Steven Enamakel7102cf4ee1d7
diff --git a/crates/openhuman-core/src/flows/caps_host/agent_tests.rs b/crates/openhuman-core/src/flows/caps_host/agent_tests.rs
index 21b4ba8331..be3f4b6700 100644
--- a/crates/openhuman-core/src/flows/caps_host/agent_tests.rs
+++ b/crates/openhuman-core/src/flows/caps_host/agent_tests.rs
@@ -36,13 +36,17 @@ async fn headless_flow_agent_includes_nested_grandchild_result() {
const CHILD_PROMPT: &str = "headless-flow-child-result-marker";
const GRANDCHILD_PROMPT: &str = "headless-flow-grandchild-result-marker";
+ const GREAT_GRANDCHILD_PROMPT: &str = "headless-flow-great-grandchild-result-marker";
struct FlowChildModel {
spawn_calls: std::sync::atomic::AtomicUsize,
child_calls: std::sync::atomic::AtomicUsize,
child_schema_names: std::sync::Mutex<Vec<Vec<String>>>,
+ manager_calls: std::sync::atomic::AtomicUsize,
+ leaf_schema_names: std::sync::Mutex<Vec<Vec<String>>>,
grandchild_calls: std::sync::atomic::AtomicUsize,
nested_spawn_calls: std::sync::atomic::AtomicUsize,
+ leaf_refusal_readable: std::sync::atomic::AtomicBool,
accepted_followups: std::sync::atomic::AtomicUsize,
child_started: std::sync::Mutex<Option<tokio::sync::oneshot::Sender<()>>>,
child_release: tokio::sync::Notify,
@@ -65,25 +69,64 @@ async fn headless_flow_agent_includes_nested_grandchild_result() {
_state: &(),
request: ModelRequest,
) -> tinyinference_llm::Result<ModelResponse> {
- let contains_child_prompt = request.messages.iter().any(|message| {
+ let contains_user_marker = |marker: &str| request.messages.iter().any(|message| {
matches!(message, tinyinference_llm::message::Message::User(_))
- && message.text().contains(CHILD_PROMPT)
+ && message.text().contains(marker)
});
- if !contains_child_prompt
- && request.messages.iter().any(|message| {
- matches!(message, tinyinference_llm::message::Message::User(_))
- && message.text().contains(GRANDCHILD_PROMPT)
- })
- {
- self.grandchild_calls
- .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
- if let Some(started) = self.child_started.lock().unwrap().take() {
- let _ = started.send(());
+ if contains_user_marker(GREAT_GRANDCHILD_PROMPT) {
+ self.leaf_schema_names.lock().unwrap().push(
+ request.tools.iter().map(|tool| tool.name.clone()).collect(),
+ );
+ let call = self.grandchild_calls.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+ if call == 0 {
+ if let Some(started) = self.child_started.lock().unwrap().take() {
+ let _ = started.send(());
+ }
+ self.child_release.notified().await;
+ let mut response = ModelResponse::assistant("");
+ response.message.tool_calls = vec![tinyinference_llm::tool::ToolCall::new(
+ "headless-flow-at-cap-spawn",
+ "spawn_subagent",
+ serde_json::json!({
+ "agent_id": "flow_final",
+ "prompt": "must not run beyond the native depth cap",
+ "task_key": "headless-flow-at-cap-task",
+ "context": "depth cap refusal regression",
+ "model": "test-model"
+ }),
+ )];
+ response.finish_reason = Some("tool_calls".into());
+ return Ok(response);
}
- self.child_release.notified().await;
+ let refusal_visible = request.messages.iter().any(|message| {
+ matches!(message, tinyinference_llm::message::Message::Tool(_))
+ && message.text().to_ascii_lowercase().contains("unknown tool")
+ });
+ self.leaf_refusal_readable.store(refusal_visible, std::sync::atomic::Ordering::SeqCst);
return Ok(ModelResponse::assistant("headless-flow-grandchild-result"));
}
- if contains_child_prompt {
+ if contains_user_marker(GRANDCHILD_PROMPT) {
+ let call = self.manager_calls.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+ if call > 0 {
+ return Ok(ModelResponse::assistant("headless-flow-manager-result"));
+ }
+ self.nested_spawn_calls.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+ let mut response = ModelResponse::assistant("");
+ response.message.tool_calls = vec![tinyinference_llm::tool::ToolCall::new(
+ "headless-flow-leaf-spawn",
+ "spawn_subagent",
+ serde_json::json!({
+ "agent_id": "flow_leaf",
+ "prompt": GREAT_GRANDCHILD_PROMPT,
+ "task_key": "headless-flow-leaf-task",
+ "context": "headless flow depth-cap regression",
+ "model": "test-model"
+ }),
+ )];
+ response.finish_reason = Some("tool_calls".into());
+ return Ok(response);
+ }
+ if contains_user_marker(CHILD_PROMPT) {
self.child_schema_names
.lock()
.unwrap()
@@ -155,8 +198,11 @@ async fn headless_flow_agent_includes_nested_grandchild_result() {
spawn_calls: std::sync::atomic::AtomicUsize::new(0),
child_calls: std::sync::atomic::AtomicUsize::new(0),
child_schema_names: std::sync::Mutex::new(Vec::new()),
+ manager_calls: std::sync::atomic::AtomicUsize::new(0),
+ leaf_schema_names: std::sync::Mutex::new(Vec::new()),
grandchild_calls: std::sync::atomic::AtomicUsize::new(0),
nested_spawn_calls: std::sync::atomic::AtomicUsize::new(0),
+ leaf_refusal_readable: std::sync::atomic::AtomicBool::new(false),
accepted_followups: std::sync::atomic::AtomicUsize::new(0),
child_started: std::sync::Mutex::new(None),
child_release: tokio::sync::Notify::new(),