fix(core): bound next-prompt suggestion requests

codex-rs/codex-api/src/common.rs

@@ -180,6 +180,8 @@ pub struct ResponsesApiRequest {
     pub stream: bool,
     pub include: Vec<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
+    pub max_output_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
     pub service_tier: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub prompt_cache_key: Option<String>,
@@ -203,6 +205,7 @@ impl From<&ResponsesApiRequest> for ResponseCreateWsRequest {
             store: request.store,
             stream: request.stream,
             include: request.include.clone(),
+            max_output_tokens: request.max_output_tokens,
             service_tier: request.service_tier.clone(),
             prompt_cache_key: request.prompt_cache_key.clone(),
             text: request.text.clone(),
@@ -228,6 +231,8 @@ pub struct ResponseCreateWsRequest {
     pub stream: bool,
     pub include: Vec<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
+    pub max_output_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
     pub service_tier: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub prompt_cache_key: Option<String>,

codex-rs/codex-api/tests/clients.rs

@@ -331,6 +331,7 @@ async fn streaming_client_retries_on_transport_error() -> Result<()> {
         store: false,
         stream: true,
         include: Vec::new(),
+        max_output_tokens: None,
         service_tier: None,
         prompt_cache_key: None,
         text: None,
@@ -432,6 +433,7 @@ async fn azure_default_store_attaches_ids_and_headers() -> Result<()> {
         store: true,
         stream: true,
         include: Vec::new(),
+        max_output_tokens: None,
         service_tier: None,
         prompt_cache_key: None,
         text: None,

codex-rs/core/src/client.rs

@@ -798,6 +798,7 @@ impl ModelClient {
             store: provider.is_azure_responses_endpoint(),
             stream: true,
             include,
+            max_output_tokens: prompt.max_output_tokens,
             service_tier,
             prompt_cache_key,
             text,
@@ -1663,10 +1664,11 @@ impl ModelClientSession {
         }
     }
 
-    /// Permanently disables WebSockets for this Codex session and resets WebSocket state.
+    /// Disables WebSockets for this Codex session and resets WebSocket state.
     ///
-    /// This is used after exhausting the provider retry budget, to force subsequent requests onto
-    /// the HTTP transport.
+    /// Shared sessions permanently force subsequent requests onto HTTP after exhausting the
+    /// provider retry budget. Isolated sessions only reset their own WebSocket state, so hidden
+    /// best-effort requests cannot change transport behavior for real turns.
     ///
     /// Returns `true` if this call activated fallback, or `false` if fallback was already active.
     pub(crate) fn try_switch_fallback_transport(
@@ -1674,9 +1676,12 @@ impl ModelClientSession {
         session_telemetry: &SessionTelemetry,
         model_info: &ModelInfo,
     ) -> bool {
-        let activated = self
-            .client
-            .force_http_fallback(session_telemetry, model_info);
+        let activated = match self.websocket_session_cache_mode {
+            WebsocketSessionCacheMode::Shared => self
+                .client
+                .force_http_fallback(session_telemetry, model_info),
+            WebsocketSessionCacheMode::Isolated => false,
+        };
         self.websocket_session = WebsocketSession::default();
         activated
     }

codex-rs/core/src/client_common.rs

@@ -35,6 +35,9 @@ pub struct Prompt {
 
     /// Whether the Responses API should strictly validate `output_schema`.
     pub output_schema_strict: bool,
+
+    /// Optional cap for generated output tokens.
+    pub max_output_tokens: Option<u64>,
 }
 
 impl Default for Prompt {
@@ -47,6 +50,7 @@ impl Default for Prompt {
             personality: None,
             output_schema: None,
             output_schema_strict: true,
+            max_output_tokens: None,
         }
     }
 }

codex-rs/core/src/client_common_tests.rs

@@ -22,6 +22,7 @@ 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 {
@@ -69,6 +70,7 @@ 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),
@@ -130,6 +132,7 @@ 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,
@@ -153,6 +156,7 @@ 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/compact_remote.rs

@@ -199,6 +199,7 @@ async fn run_remote_compact_task_inner_impl(
         personality: turn_context.personality,
         output_schema: None,
         output_schema_strict: true,
+        max_output_tokens: None,
     };
     let window_id = sess.services.model_client.current_window_id();
     let turn_metadata_header = turn_context

codex-rs/core/src/compact_remote_v2.rs

@@ -216,6 +216,7 @@ async fn run_remote_compact_task_inner_impl(
         personality: turn_context.personality,
         output_schema: None,
         output_schema_strict: true,
+        max_output_tokens: None,
     };
 
     let window_id = sess.services.model_client.current_window_id();

codex-rs/core/src/next_prompt_suggestion.rs

@@ -29,6 +29,7 @@ use codex_protocol::protocol::Event;
 use codex_protocol::protocol::EventMsg;
 use codex_protocol::protocol::TokenCountEvent;
 use codex_rollout_trace::InferenceTraceContext;
+use codex_utils_output_truncation::approx_token_count;
 use futures::StreamExt;
 use std::collections::HashSet;
 use std::time::Duration;
@@ -40,6 +41,8 @@ use tokio_util::sync::CancellationToken;
 mod next_prompt_suggestion_tests;
 
 const NEXT_PROMPT_SUGGESTION_TOKEN_HEADROOM: i64 = 1_024;
+const NEXT_PROMPT_SUGGESTION_MAX_HISTORY_TOKENS: usize = 4_096;
+const NEXT_PROMPT_SUGGESTION_MAX_OUTPUT_TOKENS: u64 = 32;
 const NEXT_PROMPT_SUGGESTION_SAMPLE_TIMEOUT: Duration = Duration::from_secs(8);
 
 #[derive(Clone, Copy)]
@@ -80,7 +83,7 @@ pub(crate) async fn suggest_next_prompt(
         tracing::debug!("next prompt suggestion skipped for incomplete tool flow");
         return None;
     }
-    let mut prompt_input = history.for_prompt(&turn_context.model_info.input_modalities);
+    let prompt_input = history.for_prompt(&turn_context.model_info.input_modalities);
     if !history_ends_at_assistant_response(&prompt_input) {
         tracing::debug!("next prompt suggestion skipped before assistant boundary");
         return None;
@@ -88,6 +91,12 @@ pub(crate) async fn suggest_next_prompt(
     if final_assistant_message_count(&prompt_input) < 2 {
         return None;
     }
+    let Some(mut prompt_input) =
+        bounded_recent_prompt_history(prompt_input, NEXT_PROMPT_SUGGESTION_MAX_HISTORY_TOKENS)
+    else {
+        tracing::debug!("next prompt suggestion skipped without bounded recent history");
+        return None;
+    };
     prompt_input.push(ContextualUserFragment::into(
         NextPromptSuggestionInstructions,
     ));
@@ -100,6 +109,7 @@ pub(crate) async fn suggest_next_prompt(
         personality: turn_context.personality,
         output_schema: None,
         output_schema_strict: true,
+        max_output_tokens: Some(NEXT_PROMPT_SUGGESTION_MAX_OUTPUT_TOKENS),
     };
     if !session_is_idle_for_suggestion(sess).await {
         return None;
@@ -305,8 +315,38 @@ fn final_assistant_message_count(items: &[ResponseItem]) -> usize {
         .count()
 }
 
+fn bounded_recent_prompt_history(
+    items: Vec<ResponseItem>,
+    max_tokens: usize,
+) -> Option<Vec<ResponseItem>> {
+    let mut remaining_tokens = max_tokens;
+    let mut keep_from = items.len();
+    for (index, item) in items.iter().enumerate().rev() {
+        let serialized = serde_json::to_string(item).ok()?;
+        let item_tokens = approx_token_count(&serialized);
+        if item_tokens > remaining_tokens {
+            break;
+        }
+        keep_from = index;
+        remaining_tokens = remaining_tokens.saturating_sub(item_tokens);
+    }
+    let bounded = items[keep_from..].to_vec();
+    if bounded.is_empty()
+        || has_unpaired_tool_flow(&bounded)
+        || !history_ends_at_assistant_response(&bounded)
+        || final_assistant_message_count(&bounded) < 2
+    {
+        return None;
+    }
+    Some(bounded)
+}
+
 fn history_ends_at_assistant_response(items: &[ResponseItem]) -> bool {
-    matches!(items.last(), Some(ResponseItem::Message { role, .. }) if role == "assistant")
+    matches!(
+        items.last(),
+        Some(ResponseItem::Message { role, phase, .. })
+            if role == "assistant" && !matches!(phase, Some(MessagePhase::Commentary))
+    )
 }
 
 fn assistant_output_text(item: &ResponseItem) -> Option<String> {

codex-rs/core/src/next_prompt_suggestion_tests.rs

@@ -1,4 +1,5 @@
 use super::HistorySnapshot;
+use super::bounded_recent_prompt_history;
 use super::filter_next_prompt_suggestion;
 use super::has_unpaired_tool_flow;
 use super::history_ends_at_assistant_response;
@@ -76,6 +77,14 @@ fn history_boundary_requires_final_assistant_message() {
     )));
     for tail in [
         user,
+        ResponseItem::Message {
+            id: None,
+            role: "assistant".to_string(),
+            content: vec![ContentItem::OutputText {
+                text: "still working".to_string(),
+            }],
+            phase: Some(MessagePhase::Commentary),
+        },
         ResponseItem::FunctionCallOutput {
             call_id: "call-1".to_string(),
             output: FunctionCallOutputPayload::from_text("done".to_string()),
@@ -88,6 +97,37 @@ fn history_boundary_requires_final_assistant_message() {
     }
 }
 
+#[test]
+fn bounded_recent_prompt_history_keeps_recent_complete_turns() {
+    let message = |role: &str, text: &str| ResponseItem::Message {
+        id: None,
+        role: role.to_string(),
+        content: vec![if role == "assistant" {
+            ContentItem::OutputText {
+                text: text.to_string(),
+            }
+        } else {
+            ContentItem::InputText {
+                text: text.to_string(),
+            }
+        }],
+        phase: None,
+    };
+    let recent = vec![
+        message("user", "first recent task"),
+        message("assistant", "first recent answer"),
+        message("user", "second recent task"),
+        message("assistant", "second recent answer"),
+    ];
+    let mut history = vec![message("user", &"old context ".repeat(1_000))];
+    history.extend(recent.clone());
+
+    assert_eq!(
+        bounded_recent_prompt_history(history, /*max_tokens*/ 256),
+        Some(recent)
+    );
+}
+
 #[test]
 fn final_assistant_message_count_ignores_commentary() {
     let assistant_message = |phase| ResponseItem::Message {

codex-rs/core/src/session/turn.rs

@@ -970,6 +970,7 @@ pub(crate) fn build_prompt(
         output_schema_strict: !crate::guardian::is_guardian_reviewer_source(
             &turn_context.session_source,
         ),
+        max_output_tokens: None,
     }
 }
 

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

@@ -73,6 +73,7 @@ async fn suggest_next_prompt_samples_history_without_tools() -> Result<()> {
     let suggestion_body = suggestion_request.body_json();
     assert_eq!(suggestion_body["tools"], json!([]));
     assert_eq!(suggestion_body["parallel_tool_calls"], false);
+    assert_eq!(suggestion_body["max_output_tokens"], json!(32));
 
     let user_texts = suggestion_request.message_input_texts("user");
     let suggestion_prompt = user_texts