fix(core): refine next-prompt suggestion requests

codex-rs/core/src/client.rs

@@ -798,7 +798,6 @@ impl ModelClient {
             store: provider.is_azure_responses_endpoint(),
             stream: true,
             include,
-            max_output_tokens: prompt.max_output_tokens,
             service_tier,
             prompt_cache_key,
             text,
@@ -1317,7 +1316,9 @@ impl ModelClientSession {
                 client_setup.api_auth,
             )
             .with_telemetry(Some(request_telemetry), Some(sse_telemetry));
-            let stream_result = client.stream_request(request, options).await;
+            let stream_result = client
+                .stream_request_with_max_output_tokens(request, options, prompt.max_output_tokens)
+                .await;
 
             match stream_result {
                 Ok(stream) => {
@@ -1491,7 +1492,11 @@ impl ModelClientSession {
                     ))
                 })?;
             let stream_result = websocket_connection
-                .stream_request(ws_request, self.websocket_session.connection_reused())
+                .stream_request_with_max_output_tokens(
+                    ws_request,
+                    self.websocket_session.connection_reused(),
+                    prompt.max_output_tokens,
+                )
                 .await
                 .map_err(|err| {
                     let response_debug_context =

codex-rs/core/src/client_common_tests.rs

@@ -22,7 +22,6 @@ fn serializes_text_verbosity_when_set() {
         store: false,
         stream: true,
         include: vec![],
-        max_output_tokens: None,
         prompt_cache_key: None,
         service_tier: None,
         text: Some(TextControls {
@@ -70,7 +69,6 @@ fn serializes_text_schema_with_strict_format() {
         store: false,
         stream: true,
         include: vec![],
-        max_output_tokens: None,
         prompt_cache_key: None,
         service_tier: None,
         text: Some(text_controls),
@@ -132,7 +130,6 @@ fn omits_text_when_not_set() {
         store: false,
         stream: true,
         include: vec![],
-        max_output_tokens: None,
         prompt_cache_key: None,
         service_tier: None,
         text: None,
@@ -156,7 +153,6 @@ fn serializes_flex_service_tier_when_set() {
         store: false,
         stream: true,
         include: vec![],
-        max_output_tokens: None,
         prompt_cache_key: None,
         service_tier: Some(ServiceTier::Flex.to_string()),
         text: None,

codex-rs/core/src/next_prompt_suggestion.rs

@@ -73,9 +73,6 @@ pub(crate) async fn suggest_next_prompt(
     let started_at = Instant::now();
     let mut turn_context = sess.new_lightweight_turn().await;
     prefer_fast_suggestion_profile(&mut turn_context);
-    if !suggestion_prompt_fits_context_window(sess, &turn_context).await {
-        return None;
-    }
 
     let history = sess.clone_history().await;
     let history_snapshot = HistorySnapshot::from_history(&history);
@@ -111,6 +108,9 @@ pub(crate) async fn suggest_next_prompt(
         output_schema_strict: true,
         max_output_tokens: Some(NEXT_PROMPT_SUGGESTION_MAX_OUTPUT_TOKENS),
     };
+    if !suggestion_prompt_fits_context_window(&prompt, &turn_context) {
+        return None;
+    }
     if !session_is_idle_for_suggestion(sess).await {
         return None;
     }
@@ -276,14 +276,19 @@ fn history_matches_snapshot(history: &ContextManager, snapshot: HistorySnapshot)
     history.history_version() == snapshot.version && history.raw_items().len() == snapshot.len
 }
 
-async fn suggestion_prompt_fits_context_window(sess: &Session, turn_context: &TurnContext) -> bool {
+fn suggestion_prompt_fits_context_window(prompt: &Prompt, turn_context: &TurnContext) -> bool {
     let Some(model_context_window) = turn_context.model_context_window() else {
         tracing::debug!("next prompt suggestion skipped without model context window");
         return false;
     };
-    if let Some(estimated_token_count) = sess.get_estimated_token_count(turn_context).await
-        && !suggestion_prompt_has_headroom(estimated_token_count, model_context_window)
-    {
+    let Ok(input) = serde_json::to_string(&prompt.input) else {
+        tracing::debug!("next prompt suggestion skipped without serializable prompt input");
+        return false;
+    };
+    let estimated_token_count = approx_token_count(&prompt.base_instructions.text)
+        .saturating_add(approx_token_count(&input));
+    let estimated_token_count = i64::try_from(estimated_token_count).unwrap_or(i64::MAX);
+    if !suggestion_prompt_has_headroom(estimated_token_count, model_context_window) {
         let suggestion_prompt_limit =
             model_context_window.saturating_sub(NEXT_PROMPT_SUGGESTION_TOKEN_HEADROOM);
         tracing::debug!(
@@ -503,6 +508,7 @@ fn filter_next_prompt_suggestion(raw: &str) -> Option<String> {
     ) || lower.starts_with("suggestion:")
         || lower.starts_with("next prompt:")
         || is_wrapped_meta(&suggestion)
+        || is_wrapped_quote(&suggestion)
         || starts_with_any(&lower, &["looks good", "thanks", "thank you"])
         || starts_with_any(&lower, &["let me", "i'll", "i will", "here's"])
     {
@@ -524,6 +530,11 @@ fn is_wrapped_meta(suggestion: &str) -> bool {
         || (suggestion.starts_with('[') && suggestion.ends_with(']'))
 }
 
+fn is_wrapped_quote(suggestion: &str) -> bool {
+    (suggestion.starts_with('"') && suggestion.ends_with('"'))
+        || (suggestion.starts_with('\'') && suggestion.ends_with('\''))
+}
+
 fn starts_with_any(value: &str, prefixes: &[&str]) -> bool {
     prefixes.iter().any(|prefix| value.starts_with(prefix))
 }

codex-rs/core/src/next_prompt_suggestion_tests.rs

@@ -22,6 +22,7 @@ fn filter_keeps_valid_prompts() {
         ("set CODEX_HOME", "set CODEX_HOME"),
         ("update Cargo.toml", "update Cargo.toml"),
         ("open app-server/README.md", "open app-server/README.md"),
+        ("don't run tests yet", "don't run tests yet"),
     ] {
         assert_eq!(
             filter_next_prompt_suggestion(suggestion),
@@ -235,6 +236,8 @@ fn filter_rejects_invalid_prompts() {
         "let me run tests",
         "what about tests?",
         "run tests.",
+        "\"run the tests\"",
+        "'run the tests'",
         "run\ntests",
         "continue with every possible next step in this project and explain every detail now",
     ] {