Reduce cloning when building Responses requests (#34825)

## Why

Preparing Responses API requests rebuilt tool definitions as a generic JSON
tree, and incremental WebSocket requests cloned their full item prefix for
comparison.

## What changed

- Serialize tool definitions into shared raw JSON that can be embedded directly
  in HTTP and WebSocket requests.
- Compare incremental request prefixes in place while still ignoring internal
  message metadata.

## Testing

- Verify raw tool JSON matches the existing value encoding.
- Preserve the serialized WebSocket request payload.

GitOrigin-RevId: 66e2921792c332e8904954dafa597f6d39abcf29

codex-rs/codex-api/Cargo.toml

@@ -19,7 +19,7 @@ http = { workspace = true }
 reqwest = { workspace = true, features = ["json", "stream"] }
 schemars = { workspace = true }
 serde = { workspace = true, features = ["derive"] }
-serde_json = { workspace = true }
+serde_json = { workspace = true, features = ["raw_value"] }
 thiserror = { workspace = true }
 tokio = { workspace = true, features = ["fs", "macros", "net", "rt", "sync", "time"] }
 tokio-tungstenite = { workspace = true }

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

@@ -12,8 +12,10 @@ use futures::Stream;
 use serde::Deserialize;
 use serde::Serialize;
 use serde_json::Value;
+use serde_json::value::RawValue;
 use std::collections::HashMap;
 use std::pin::Pin;
+use std::sync::Arc;
 use std::task::Context;
 use std::task::Poll;
 use tokio::sync::mpsc;
@@ -29,7 +31,7 @@ pub struct CompactionInput<'a> {
     #[serde(skip_serializing_if = "str::is_empty")]
     pub instructions: &'a str,
     #[serde(skip_serializing_if = "Option::is_none")]
-    pub tools: Option<Vec<Value>>,
+    pub tools: Option<ResponsesApiTools>,
     pub parallel_tool_calls: bool,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub reasoning: Option<Reasoning>,
@@ -212,6 +214,40 @@ impl From<VerbosityConfig> for OpenAiVerbosity {
     }
 }
 
+/// Serialized tool definitions for Responses API requests.
+///
+/// Keeping the tool list as raw JSON avoids rebuilding a generic JSON value
+/// tree, while the shared allocation keeps request clones cheap.
+#[derive(Debug, Clone)]
+pub struct ResponsesApiTools(Arc<RawValue>);
+
+impl ResponsesApiTools {
+    pub(crate) fn as_raw_value(&self) -> &RawValue {
+        &self.0
+    }
+}
+
+impl From<Arc<RawValue>> for ResponsesApiTools {
+    fn from(value: Arc<RawValue>) -> Self {
+        Self(value)
+    }
+}
+
+impl PartialEq for ResponsesApiTools {
+    fn eq(&self, other: &Self) -> bool {
+        Arc::ptr_eq(&self.0, &other.0) || self.0.get() == other.0.get()
+    }
+}
+
+impl Serialize for ResponsesApiTools {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        self.0.serialize(serializer)
+    }
+}
+
 #[derive(Debug, Serialize, Clone, PartialEq)]
 pub struct ResponsesApiRequest {
     pub model: String,
@@ -219,7 +255,7 @@ pub struct ResponsesApiRequest {
     pub instructions: String,
     pub input: Vec<ResponseItem>,
     #[serde(skip_serializing_if = "Option::is_none")]
-    pub tools: Option<Vec<serde_json::Value>>,
+    pub tools: Option<ResponsesApiTools>,
     pub tool_choice: String,
     pub parallel_tool_calls: bool,
     pub reasoning: Option<Reasoning>,
@@ -245,7 +281,7 @@ impl<'a> From<&'a ResponsesApiRequest> for ResponseCreateWsRequest<'a> {
             instructions: &request.instructions,
             previous_response_id: None,
             input: &request.input,
-            tools: request.tools.as_deref(),
+            tools: request.tools.as_ref().map(ResponsesApiTools::as_raw_value),
             tool_choice: &request.tool_choice,
             parallel_tool_calls: request.parallel_tool_calls,
             reasoning: request.reasoning.as_ref(),
@@ -271,7 +307,7 @@ pub struct ResponseCreateWsRequest<'a> {
     pub previous_response_id: Option<String>,
     pub input: &'a [ResponseItem],
     #[serde(skip_serializing_if = "Option::is_none")]
-    pub tools: Option<&'a [Value]>,
+    pub tools: Option<&'a RawValue>,
     pub tool_choice: &'a str,
     pub parallel_tool_calls: bool,
     pub reasoning: Option<&'a Reasoning>,

codex-rs/codex-api/src/endpoint/responses_websocket.rs

@@ -907,7 +907,10 @@ mod tests {
     use codex_protocol::models::ResponseItem;
     use pretty_assertions::assert_eq;
     use serde_json::json;
+    use serde_json::value::RawValue;
+    use serde_json::value::to_raw_value;
     use std::collections::HashMap;
+    use std::sync::Arc;
 
     #[test]
     fn direct_serialization_preserves_websocket_request_payload() {
@@ -923,11 +926,17 @@ mod tests {
                 phase: None,
                 internal_chat_message_metadata_passthrough: None,
             }],
-            tools: Some(vec![json!({
-                "type": "function",
-                "name": "lookup",
-                "parameters": {"type": "object"}
-            })]),
+            tools: Some(
+                Arc::<RawValue>::from(
+                    to_raw_value(&vec![json!({
+                        "type": "function",
+                        "name": "lookup",
+                        "parameters": {"type": "object"}
+                    })])
+                    .expect("serialize tools"),
+                )
+                .into(),
+            ),
             tool_choice: "auto".to_string(),
             parallel_tool_calls: true,
             reasoning: None,

codex-rs/codex-api/src/lib.rs

@@ -39,6 +39,7 @@ pub use crate::common::ResponseCreateWsRequest;
 pub use crate::common::ResponseEvent;
 pub use crate::common::ResponseStream;
 pub use crate::common::ResponsesApiRequest;
+pub use crate::common::ResponsesApiTools;
 pub use crate::common::ResponsesWsRequest;
 pub use crate::common::StreamOptions;
 pub use crate::common::TextControls;

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

@@ -28,6 +28,7 @@ use http::HeaderMap;
 use http::HeaderValue;
 use http::StatusCode;
 use pretty_assertions::assert_eq;
+use serde_json::value::RawValue;
 
 fn assert_path_ends_with(requests: &[Request], suffix: &str) {
     assert_eq!(requests.len(), 1);
@@ -38,6 +39,10 @@ fn assert_path_ends_with(requests: &[Request], suffix: &str) {
     );
 }
 
+fn empty_tools() -> Arc<RawValue> {
+    Arc::from(RawValue::from_string("[]".to_string()).expect("valid tool JSON"))
+}
+
 fn request_body_bytes(request: &Request) -> &[u8] {
     let Some(RequestBody::EncodedJson(body)) = request.body.as_ref() else {
         panic!("expected a prepared request body");
@@ -316,7 +321,7 @@ async fn responses_client_stream_request_preserves_item_ids() -> Result<()> {
             phase: None,
             internal_chat_message_metadata_passthrough: None,
         }],
-        tools: Some(Vec::new()),
+        tools: Some(empty_tools().into()),
         tool_choice: "auto".into(),
         parallel_tool_calls: false,
         reasoning: None,
@@ -403,7 +408,7 @@ async fn streaming_client_retries_on_transport_error() -> Result<()> {
         model: "gpt-test".into(),
         instructions: "Say hi".into(),
         input: Vec::new(),
-        tools: Some(Vec::new()),
+        tools: Some(empty_tools().into()),
         tool_choice: "auto".into(),
         parallel_tool_calls: false,
         reasoning: None,
@@ -523,7 +528,7 @@ async fn azure_store_sends_ids_and_headers() -> Result<()> {
             phase: None,
             internal_chat_message_metadata_passthrough: None,
         }],
-        tools: Some(Vec::new()),
+        tools: Some(empty_tools().into()),
         tool_choice: "auto".into(),
         parallel_tool_calls: false,
         reasoning: None,

codex-rs/core/src/client.rs

@@ -89,6 +89,7 @@ use codex_rollout_trace::CompactionTraceContext;
 use codex_rollout_trace::InferenceTraceAttempt;
 use codex_rollout_trace::InferenceTraceContext;
 use codex_tools::create_tools_json_for_responses_api;
+use codex_tools::create_tools_raw_json_for_responses_api;
 use eventsource_stream::Event;
 use eventsource_stream::EventStreamError;
 use futures::StreamExt;
@@ -358,6 +359,21 @@ fn responses_request_properties_match(
         && previous_text == current_text
 }
 
+fn response_items_equal_ignoring_internal_metadata(
+    previous: &ResponseItem,
+    current: &ResponseItem,
+) -> bool {
+    if previous == current {
+        return true;
+    }
+
+    let mut previous = previous.clone();
+    previous.clear_internal_chat_message_metadata_passthrough();
+    let mut current = current.clone();
+    current.clear_internal_chat_message_metadata_passthrough();
+    previous == current
+}
+
 impl WebsocketSession {
     fn set_connection_reused(&self, connection_reused: bool) {
         *self
@@ -836,8 +852,8 @@ impl ModelClient {
                 .iter_mut()
                 .for_each(ResponseItem::clear_internal_chat_message_metadata_passthrough);
         }
-        let tools = create_tools_json_for_responses_api(&prompt.tools)?;
         let (instructions, tools) = if model_info.use_responses_lite {
+            let tools = create_tools_json_for_responses_api(&prompt.tools)?;
             let mut prefix = vec![ResponseItem::AdditionalTools {
                 id: None,
                 role: "developer".to_string(),
@@ -857,7 +873,10 @@ impl ModelClient {
             input.splice(0..0, prefix);
             (String::new(), None)
         } else {
-            (prompt.base_instructions.text.clone(), Some(tools))
+            (
+                prompt.base_instructions.text.clone(),
+                Some(create_tools_raw_json_for_responses_api(&prompt.tools)?.into()),
+            )
         };
         let reasoning = Self::build_reasoning(model_info, effort, summary);
         let stream_options = (self.state.concurrent_reasoning_summaries_enabled
@@ -1167,29 +1186,25 @@ impl ModelClientSession {
             return None;
         }
 
-        // To compare the inputs, we concatenate the previous request items with the response items,
-        // then compare that against the equivalent slice of request items, ignoring metadata. If
-        // they match, we can consider the remaining items the incremental request.
-        let mut previous_items = previous_request.input.clone();
-        if let Some(response) = last_response {
-            previous_items.extend_from_slice(&response.items_added);
-        }
-        previous_items
-            .iter_mut()
-            .for_each(ResponseItem::clear_internal_chat_message_metadata_passthrough);
-
+        let response_items =
+            last_response.map_or(&[][..], |response| response.items_added.as_slice());
+        let previous_items_len = previous_request
+            .input
+            .len()
+            .checked_add(response_items.len())?;
         let Some((request_items_to_compare, incremental_items)) =
-            request.input.split_at_checked(previous_items.len())
+            request.input.split_at_checked(previous_items_len)
         else {
             trace!("incremental request failed, incompatible request length");
             return None;
         };
-        let mut request_prefix = request_items_to_compare.to_vec();
-        request_prefix
-            .iter_mut()
-            .for_each(ResponseItem::clear_internal_chat_message_metadata_passthrough);
-
-        if previous_items != request_prefix {
+        let previous_items = previous_request.input.iter().chain(response_items);
+        if !previous_items
+            .zip(request_items_to_compare)
+            .all(|(previous, current)| {
+                response_items_equal_ignoring_internal_metadata(previous, current)
+            })
+        {
             trace!("incremental request failed, items didn't match");
             return None;
         }

codex-rs/core/src/client_common_tests.rs

@@ -6,9 +6,15 @@ use codex_protocol::config_types::ServiceTier;
 use codex_protocol::models::FunctionCallOutputPayload;
 use codex_protocol::models::ImageDetail;
 use pretty_assertions::assert_eq;
+use serde_json::value::RawValue;
+use std::sync::Arc;
 
 use super::*;
 
+fn empty_tools() -> Arc<RawValue> {
+    Arc::from(RawValue::from_string("[]".to_string()).expect("valid tool JSON"))
+}
+
 fn prompt_with_image_outputs() -> Prompt {
     Prompt {
         input: vec![
@@ -105,12 +111,11 @@ fn responses_lite_request_copies_strip_image_details() {
 #[test]
 fn serializes_text_verbosity_when_set() {
     let input: Vec<ResponseItem> = vec![];
-    let tools: Vec<serde_json::Value> = vec![];
     let req = ResponsesApiRequest {
         model: "gpt-5.4".to_string(),
         instructions: "i".to_string(),
         input,
-        tools: Some(tools),
+        tools: Some(empty_tools().into()),
         tool_choice: "auto".to_string(),
         parallel_tool_calls: true,
         reasoning: None,
@@ -139,7 +144,6 @@ fn serializes_text_verbosity_when_set() {
 #[test]
 fn serializes_text_schema_with_strict_format() {
     let input: Vec<ResponseItem> = vec![];
-    let tools: Vec<serde_json::Value> = vec![];
     let schema = serde_json::json!({
         "type": "object",
         "properties": {
@@ -158,7 +162,7 @@ fn serializes_text_schema_with_strict_format() {
         model: "gpt-5.4".to_string(),
         instructions: "i".to_string(),
         input,
-        tools: Some(tools),
+        tools: Some(empty_tools().into()),
         tool_choice: "auto".to_string(),
         parallel_tool_calls: true,
         reasoning: None,
@@ -215,12 +219,11 @@ fn serializes_text_schema_with_non_strict_format() {
 #[test]
 fn omits_text_when_not_set() {
     let input: Vec<ResponseItem> = vec![];
-    let tools: Vec<serde_json::Value> = vec![];
     let req = ResponsesApiRequest {
         model: "gpt-5.4".to_string(),
         instructions: "i".to_string(),
         input,
-        tools: Some(tools),
+        tools: Some(empty_tools().into()),
         tool_choice: "auto".to_string(),
         parallel_tool_calls: true,
         reasoning: None,
@@ -244,7 +247,7 @@ fn serializes_flex_service_tier_when_set() {
         model: "gpt-5.4".to_string(),
         instructions: "i".to_string(),
         input: vec![],
-        tools: Some(vec![]),
+        tools: Some(empty_tools().into()),
         tool_choice: "auto".to_string(),
         parallel_tool_calls: true,
         reasoning: None,

codex-rs/core/src/tools/router.rs

@@ -72,7 +72,7 @@ impl ToolRouter {
         }
     }
 
-    pub fn model_visible_specs(&self) -> Vec<ToolSpec> {
+    pub(crate) fn model_visible_specs(&self) -> Vec<ToolSpec> {
         self.model_visible_specs.clone()
     }
 

codex-rs/tools/Cargo.toml

@@ -26,7 +26,7 @@ rmcp = { workspace = true, default-features = false, features = [
     "server",
 ] }
 serde = { workspace = true, features = ["derive"] }
-serde_json = { workspace = true }
+serde_json = { workspace = true, features = ["raw_value"] }
 thiserror = { workspace = true }
 tracing = { workspace = true }
 urlencoding = { workspace = true }

codex-rs/tools/src/lib.rs

@@ -105,3 +105,4 @@ pub use tool_spec::ResponsesApiWebSearchFilters;
 pub use tool_spec::ResponsesApiWebSearchUserLocation;
 pub use tool_spec::ToolSpec;
 pub use tool_spec::create_tools_json_for_responses_api;
+pub use tool_spec::create_tools_raw_json_for_responses_api;

codex-rs/tools/src/tool_spec.rs

@@ -9,6 +9,8 @@ use codex_protocol::config_types::WebSearchUserLocation as ConfigWebSearchUserLo
 use codex_protocol::config_types::WebSearchUserLocationType;
 use serde::Serialize;
 use serde_json::Value;
+use serde_json::value::RawValue;
+use std::sync::Arc;
 
 /// When serialized as JSON, this produces a valid "Tool" in the OpenAI
 /// Responses API.
@@ -87,6 +89,13 @@ pub fn create_tools_json_for_responses_api(
     Ok(tools_json)
 }
 
+/// Returns raw JSON that can be embedded directly in a Responses API request.
+pub fn create_tools_raw_json_for_responses_api(
+    tools: &[ToolSpec],
+) -> Result<Arc<RawValue>, serde_json::Error> {
+    serde_json::value::to_raw_value(tools).map(Arc::from)
+}
+
 #[derive(Debug, Clone, Serialize, PartialEq)]
 pub struct ResponsesApiWebSearchFilters {
     #[serde(skip_serializing_if = "Option::is_none")]

codex-rs/tools/src/tool_spec_tests.rs

@@ -9,6 +9,7 @@ use crate::JsonSchema;
 use crate::ResponsesApiNamespaceTool;
 use crate::ResponsesApiTool;
 use crate::create_tools_json_for_responses_api;
+use crate::create_tools_raw_json_for_responses_api;
 use codex_protocol::config_types::WebSearchContextSize;
 use codex_protocol::config_types::WebSearchFilters as ConfigWebSearchFilters;
 use codex_protocol::config_types::WebSearchUserLocation as ConfigWebSearchUserLocation;
@@ -143,6 +144,29 @@ fn create_tools_json_for_responses_api_includes_top_level_name() {
     );
 }
 
+#[test]
+fn raw_tool_json_matches_value_encoding() {
+    let specs = vec![ToolSpec::Function(ResponsesApiTool {
+        name: "demo".to_string(),
+        description: "A demo tool".to_string(),
+        strict: false,
+        defer_loading: None,
+        parameters: JsonSchema::object(
+            BTreeMap::new(),
+            /*required*/ None,
+            /*additional_properties*/ None,
+        ),
+        output_schema: None,
+    })];
+    let expected = create_tools_json_for_responses_api(&specs).expect("serialize tools");
+    let raw = create_tools_raw_json_for_responses_api(&specs).expect("serialize raw tools");
+
+    assert_eq!(
+        serde_json::from_str::<Vec<serde_json::Value>>(raw.get()).expect("parse raw tools"),
+        expected,
+    );
+}
+
 #[test]
 fn namespace_tool_spec_serializes_expected_wire_shape() {
     assert_eq!(