fix(next-prompt): keep hidden usage out of context state

codex-rs/core/src/next_prompt_suggestion.rs

@@ -28,6 +28,7 @@ use codex_protocol::openai_models::ReasoningEffort;
 use codex_protocol::protocol::Event;
 use codex_protocol::protocol::EventMsg;
 use codex_protocol::protocol::TokenCountEvent;
+use codex_protocol::protocol::TokenUsageInfo;
 use codex_rollout_trace::InferenceTraceContext;
 use futures::StreamExt;
 use std::collections::HashSet;
@@ -181,11 +182,14 @@ pub(crate) async fn suggest_next_prompt(
                 token_usage,
                 ..
             } => {
-                let should_emit_token_count = token_usage.is_some() || latest_rate_limits.is_some();
-                sess.record_token_usage_info(&turn_context, token_usage.as_ref())
-                    .await;
+                let token_usage_info = TokenUsageInfo::new_or_append(
+                    &sess.token_usage_info().await,
+                    &token_usage,
+                    turn_context.model_context_window(),
+                );
+                let should_emit_token_count =
+                    token_usage_info.is_some() || latest_rate_limits.is_some();
                 if should_emit_token_count {
-                    let token_usage_info = sess.token_usage_info().await;
                     // App-server clients key usage updates by turn id. Attribute hidden
                     // suggestion sampling to the latest real turn instead of the ephemeral
                     // lightweight turn used to configure the request.

codex-rs/core/tests/suite/next_prompt_suggestion.rs

@@ -96,6 +96,32 @@ async fn suggest_next_prompt_samples_history_without_tools() -> Result<()> {
             .used_percent,
         42.0
     );
+    assert_eq!(test.codex.token_usage_info().await, None);
+
+    Ok(())
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn suggest_next_prompt_skips_early_history_without_request() -> Result<()> {
+    skip_if_no_network!(Ok(()));
+
+    let server = MockServer::start().await;
+    let first_turn_response = sse(vec![
+        ev_response_created("resp-1"),
+        ev_assistant_message("msg-1", "first turn complete"),
+        ev_completed("resp-1"),
+    ]);
+    let responses = mount_sse_sequence(&server, vec![first_turn_response]).await;
+
+    let test = test_codex().build(&server).await?;
+    test.submit_turn("first task").await?;
+
+    let suggestion = test
+        .codex
+        .suggest_next_prompt(CancellationToken::new())
+        .await?;
+    assert_eq!(suggestion, None);
+    assert_eq!(responses.requests().len(), 1);
 
     Ok(())
 }