feat(auto-review) short-circuit (#18890)

## Summary
Short circuit the convo if auto-review hits too many denials

## Testing
- [x] Added unit tests

---------

Co-authored-by: Codex <noreply@openai.com>
This commit is contained in:
Dylan Hurd
2026-04-22 13:34:15 -07:00
committed by GitHub
Unverified
parent b77791c228
commit ed4def8286
28 changed files with 645 additions and 22 deletions
+2
View File
@@ -695,6 +695,8 @@ impl Session {
models_manager: Arc::clone(&models_manager),
tool_approvals: Mutex::new(ApprovalStore::default()),
guardian_rejections: Mutex::new(HashMap::new()),
guardian_rejection_circuit_breaker: Mutex::new(Default::default()),
runtime_handle: tokio::runtime::Handle::current(),
skills_manager,
plugins_manager: Arc::clone(&plugins_manager),
mcp_manager: Arc::clone(&mcp_manager),
+123
View File
@@ -3093,6 +3093,8 @@ pub(crate) async fn make_session_and_context() -> (Session, TurnContext) {
models_manager: Arc::clone(&models_manager),
tool_approvals: Mutex::new(ApprovalStore::default()),
guardian_rejections: Mutex::new(std::collections::HashMap::new()),
guardian_rejection_circuit_breaker: Mutex::new(Default::default()),
runtime_handle: tokio::runtime::Handle::current(),
skills_manager,
plugins_manager,
mcp_manager,
@@ -4313,6 +4315,8 @@ pub(crate) async fn make_session_and_context_with_dynamic_tools_and_rx(
models_manager: Arc::clone(&models_manager),
tool_approvals: Mutex::new(ApprovalStore::default()),
guardian_rejections: Mutex::new(std::collections::HashMap::new()),
guardian_rejection_circuit_breaker: Mutex::new(Default::default()),
runtime_handle: tokio::runtime::Handle::current(),
skills_manager,
plugins_manager,
mcp_manager,
@@ -5545,6 +5549,125 @@ impl SessionTask for NeverEndingTask {
}
}
#[derive(Clone, Copy)]
struct GuardianDeniedApprovalTask;
impl SessionTask for GuardianDeniedApprovalTask {
fn kind(&self) -> TaskKind {
TaskKind::Regular
}
fn span_name(&self) -> &'static str {
"session_task.guardian_denied_approval"
}
async fn run(
self: Arc<Self>,
session: Arc<SessionTaskContext>,
ctx: Arc<TurnContext>,
_input: Vec<UserInput>,
cancellation_token: CancellationToken,
) -> Option<String> {
let session = session.clone_session();
for _ in 0..3 {
crate::guardian::record_guardian_denial_for_test(&session, &ctx, &ctx.sub_id).await;
}
cancellation_token.cancelled().await;
None
}
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn guardian_auto_review_interrupts_after_three_consecutive_denials() {
let (sess, tc, rx) = make_session_and_context_with_rx().await;
let input = vec![UserInput::Text {
text: "trigger guardian denials".to_string(),
text_elements: Vec::new(),
}];
sess.spawn_task(Arc::clone(&tc), input, GuardianDeniedApprovalTask)
.await;
let mut observed = Vec::new();
let aborted = tokio::time::timeout(std::time::Duration::from_secs(5), async {
loop {
let event = rx.recv().await.expect("event");
if let EventMsg::TurnAborted(event) = &event.msg {
let event = event.clone();
observed.push(EventMsg::TurnAborted(event.clone()));
break event;
}
observed.push(event.msg);
}
})
.await
.unwrap_or_else(|_| {
panic!(
"guardian denial circuit breaker should interrupt the turn; observed events: {observed:?}"
)
});
assert_eq!(aborted.reason, TurnAbortReason::Interrupted);
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn guardian_helper_review_interrupts_after_three_consecutive_denials() {
let (sess, tc, rx) = make_session_and_context_with_rx().await;
let input = vec![UserInput::Text {
text: "keep turn active for helper reviews".to_string(),
text_elements: Vec::new(),
}];
sess.spawn_task(
Arc::clone(&tc),
input,
NeverEndingTask {
kind: TaskKind::Regular,
listen_to_cancellation_token: true,
},
)
.await;
let session_for_review = Arc::clone(&sess);
let turn_for_review = Arc::clone(&tc);
let turn_id = tc.sub_id.clone();
let review_thread = std::thread::spawn(move || {
let runtime = tokio::runtime::Builder::new_current_thread()
.enable_all()
.build()
.expect("helper review runtime");
runtime.block_on(async move {
for _ in 0..3 {
crate::guardian::record_guardian_denial_for_test(
&session_for_review,
&turn_for_review,
&turn_id,
)
.await;
}
});
});
review_thread.join().expect("helper review thread");
let mut observed = Vec::new();
let aborted = timeout(StdDuration::from_secs(5), async {
loop {
let event = rx.recv().await.expect("event");
if let EventMsg::TurnAborted(event) = &event.msg {
let event = event.clone();
observed.push(EventMsg::TurnAborted(event.clone()));
break event;
}
observed.push(event.msg);
}
})
.await
.unwrap_or_else(|_| {
panic!(
"helper review circuit breaker should interrupt the turn; observed events: {observed:?}"
)
});
assert_eq!(aborted.reason, TurnAbortReason::Interrupted);
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
#[test_log::test]
async fn abort_regular_task_emits_turn_aborted_only() {
+1
View File
@@ -1484,6 +1484,7 @@ pub(super) fn realtime_text_for_event(msg: &EventMsg) -> Option<String> {
},
EventMsg::Error(_)
| EventMsg::Warning(_)
| EventMsg::GuardianWarning(_)
| EventMsg::RealtimeConversationStarted(_)
| EventMsg::RealtimeConversationSdp(_)
| EventMsg::RealtimeConversationRealtime(_)