Use input items for Responses Lite tools

This commit is contained in:
Rohit Arunachalam
2026-06-21 23:52:55 -07:00
parent 98845e4840
commit dd04b8d36f
16 changed files with 139 additions and 46 deletions
+2 -1
View File
@@ -41,7 +41,8 @@ fn keep_forked_rollout_item(item: &RolloutItem, preserve_reference_context_item:
_ => false,
},
RolloutItem::ResponseItem(
ResponseItem::AgentMessage { .. }
ResponseItem::AdditionalTools { .. }
| ResponseItem::AgentMessage { .. }
| ResponseItem::Reasoning { .. }
| ResponseItem::LocalShellCall { .. }
| ResponseItem::FunctionCall { .. }
+22 -2
View File
@@ -73,6 +73,7 @@ use codex_otel::current_span_w3c_trace_context;
use codex_protocol::ThreadId;
use codex_protocol::config_types::ReasoningSummary as ReasoningSummaryConfig;
use codex_protocol::config_types::Verbosity as VerbosityConfig;
use codex_protocol::models::ContentItem;
use codex_protocol::models::ResponseItem;
use codex_protocol::openai_models::ModelInfo;
use codex_protocol::openai_models::ReasoningEffort as ReasoningEffortConfig;
@@ -778,12 +779,31 @@ impl ModelClient {
service_tier: Option<String>,
responses_metadata: &CodexResponsesMetadata,
) -> Result<ResponsesApiRequest> {
let instructions = &prompt.base_instructions.text;
let mut input = prompt.get_formatted_input_for_request(model_info.use_responses_lite);
if !self.state.provider.info().is_openai() {
input.iter_mut().for_each(ResponseItem::clear_metadata);
}
let tools = create_tools_json_for_responses_api(&prompt.tools)?;
let (instructions, tools) = if model_info.use_responses_lite {
let mut prefix = vec![ResponseItem::AdditionalTools {
role: "developer".to_string(),
tools,
}];
if !prompt.base_instructions.text.is_empty() {
prefix.push(ResponseItem::Message {
id: None,
role: "developer".to_string(),
content: vec![ContentItem::InputText {
text: prompt.base_instructions.text.clone(),
}],
phase: None,
});
}
input.splice(0..0, prefix);
(String::new(), None)
} else {
(prompt.base_instructions.text.clone(), Some(tools))
};
let reasoning = Self::build_reasoning(model_info, effort, summary);
let include = if reasoning.is_some() {
vec!["reasoning.encrypted_content".to_string()]
@@ -810,7 +830,7 @@ impl ModelClient {
let service_tier = model_info.service_tier_for_request(service_tier);
let request = ResponsesApiRequest {
model: model_info.slug.clone(),
instructions: instructions.clone(),
instructions,
input,
tools,
tool_choice: "auto".to_string(),
+2 -1
View File
@@ -83,7 +83,8 @@ fn strip_image_details(items: &mut [ResponseItem]) {
}
}
}
ResponseItem::Reasoning { .. }
ResponseItem::AdditionalTools { .. }
| ResponseItem::Reasoning { .. }
| ResponseItem::AgentMessage { .. }
| ResponseItem::LocalShellCall { .. }
| ResponseItem::FunctionCall { .. }
+7 -13
View File
@@ -104,13 +104,11 @@ 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,
input: vec![],
tools: Some(vec![]),
tool_choice: "auto".to_string(),
parallel_tool_calls: true,
reasoning: None,
@@ -137,8 +135,6 @@ 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": {
@@ -156,8 +152,8 @@ fn serializes_text_schema_with_strict_format() {
let req = ResponsesApiRequest {
model: "gpt-5.4".to_string(),
instructions: "i".to_string(),
input,
tools,
input: vec![],
tools: Some(vec![]),
tool_choice: "auto".to_string(),
parallel_tool_calls: true,
reasoning: None,
@@ -212,13 +208,11 @@ 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,
input: vec![],
tools: Some(vec![]),
tool_choice: "auto".to_string(),
parallel_tool_calls: true,
reasoning: None,
@@ -241,7 +235,7 @@ fn serializes_flex_service_tier_when_set() {
model: "gpt-5.4".to_string(),
instructions: "i".to_string(),
input: vec![],
tools: vec![],
tools: Some(vec![]),
tool_choice: "auto".to_string(),
parallel_tool_calls: true,
reasoning: None,
+2 -1
View File
@@ -352,7 +352,8 @@ pub(crate) fn should_keep_compacted_history_item(item: &ResponseItem) -> bool {
ResponseItem::AgentMessage { .. } => true,
ResponseItem::Compaction { .. } | ResponseItem::ContextCompaction { .. } => true,
ResponseItem::CompactionTrigger { .. } => false,
ResponseItem::Reasoning { .. }
ResponseItem::AdditionalTools { .. }
| ResponseItem::Reasoning { .. }
| ResponseItem::LocalShellCall { .. }
| ResponseItem::FunctionCall { .. }
| ResponseItem::ToolSearchCall { .. }
+6 -3
View File
@@ -362,7 +362,8 @@ impl ContextManager {
output: truncate_function_output_payload(output, policy_with_serialization_budget),
metadata: metadata.clone(),
},
ResponseItem::Message { .. }
ResponseItem::AdditionalTools { .. }
| ResponseItem::Message { .. }
| ResponseItem::AgentMessage { .. }
| ResponseItem::Reasoning { .. }
| ResponseItem::LocalShellCall { .. }
@@ -452,7 +453,8 @@ pub(crate) fn truncate_function_output_payload(
fn is_api_message(message: &ResponseItem) -> bool {
match message {
ResponseItem::Message { role, .. } => role.as_str() != "system",
ResponseItem::AgentMessage { .. }
ResponseItem::AdditionalTools { .. }
| ResponseItem::AgentMessage { .. }
| ResponseItem::FunctionCallOutput { .. }
| ResponseItem::FunctionCall { .. }
| ResponseItem::ToolSearchCall { .. }
@@ -698,7 +700,8 @@ fn is_model_generated_item(item: &ResponseItem) -> bool {
| ResponseItem::Compaction { .. }
| ResponseItem::ContextCompaction { .. } => true,
ResponseItem::CompactionTrigger { .. } => false,
ResponseItem::FunctionCallOutput { .. }
ResponseItem::AdditionalTools { .. }
| ResponseItem::FunctionCallOutput { .. }
| ResponseItem::ToolSearchOutput { .. }
| ResponseItem::CustomToolCallOutput { .. }
| ResponseItem::AgentMessage { .. }
+2 -1
View File
@@ -52,7 +52,8 @@ pub(crate) fn prepare_response_items(items: &mut [ResponseItem]) {
prepare_tool_output_content(content);
}
}
ResponseItem::Reasoning { .. }
ResponseItem::AdditionalTools { .. }
| ResponseItem::Reasoning { .. }
| ResponseItem::AgentMessage { .. }
| ResponseItem::LocalShellCall { .. }
| ResponseItem::FunctionCall { .. }
+2 -1
View File
@@ -2072,7 +2072,8 @@ async fn try_run_sampling_request(
}
ResponseItem::Reasoning { .. } => true,
ResponseItem::AgentMessage { .. } => false,
ResponseItem::LocalShellCall { .. }
ResponseItem::AdditionalTools { .. }
| ResponseItem::LocalShellCall { .. }
| ResponseItem::FunctionCall { .. }
| ResponseItem::ToolSearchCall { .. }
| ResponseItem::FunctionCallOutput { .. }
+2 -1
View File
@@ -380,7 +380,8 @@ fn response_item_records_turn_ttft(item: &ResponseItem) -> bool {
| ResponseItem::Compaction { .. }
| ResponseItem::ContextCompaction { .. } => true,
ResponseItem::CompactionTrigger { .. } => false,
ResponseItem::FunctionCallOutput { .. }
ResponseItem::AdditionalTools { .. }
| ResponseItem::FunctionCallOutput { .. }
| ResponseItem::CustomToolCallOutput { .. }
| ResponseItem::ToolSearchOutput { .. }
| ResponseItem::Other => false,
+69 -13
View File
@@ -54,6 +54,70 @@ fn has_hosted_tool(tools: &[Value], tool_type: &str) -> bool {
.any(|tool| tool.get("type").and_then(Value::as_str) == Some(tool_type))
}
fn additional_tools(body: &Value) -> Result<&[Value]> {
body["input"]
.as_array()
.context("Responses request input should be an array")?
.iter()
.find(|item| item.get("type").and_then(Value::as_str) == Some("additional_tools"))
.context("Responses request should contain additional_tools")?["tools"]
.as_array()
.map(Vec::as_slice)
.context("additional_tools tools should be an array")
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_lite_uses_input_items_for_instructions_and_tools() -> Result<()> {
skip_if_no_network!(Ok(()));
let server = responses::start_mock_server().await;
let response_mock = responses::mount_sse_once(
&server,
responses::sse(vec![
responses::ev_response_created("resp-1"),
responses::ev_completed("resp-1"),
]),
)
.await;
let mut builder = test_codex()
.with_model_info_override("gpt-5.4", |model_info| {
model_info.use_responses_lite = true;
})
.with_config(|config| {
config.base_instructions = Some("test instructions".to_string());
});
let test = builder.build(&server).await?;
test.submit_turn("hello").await?;
let body = response_mock.single_request().body_json();
assert!(body.get("instructions").is_none());
assert!(body.get("tools").is_none());
let input = body["input"]
.as_array()
.context("Responses request input should be an array")?;
assert_eq!(input[0]["type"], "additional_tools");
assert_eq!(input[0]["role"], "developer");
assert_eq!(
input[1],
serde_json::json!({
"type": "message",
"role": "developer",
"content": [{
"type": "input_text",
"text": "test instructions",
}],
})
);
let tools = additional_tools(&body)?;
assert!(!tools.is_empty());
Ok(())
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_lite_strips_data_image_detail_without_resize_all_images() -> Result<()> {
skip_if_no_network!(Ok(()));
@@ -144,17 +208,10 @@ async fn responses_lite_uses_standalone_web_search_and_image_generation() -> Res
request.header(RESPONSES_LITE_HEADER).as_deref(),
Some("true")
);
request
.tool_by_name("web", "run")
.context("Responses Lite should expose standalone web search")?;
request
.tool_by_name("image_gen", "imagegen")
.context("Responses Lite should expose standalone image generation")?;
let body = request.body_json();
let tools = body["tools"]
.as_array()
.context("Responses request tools should be an array")?;
assert!(body.get("tools").is_none());
let tools = additional_tools(&body)?;
assert!(!tools.is_empty());
assert!(!has_hosted_tool(tools, "web_search"));
assert!(!has_hosted_tool(tools, "image_generation"));
@@ -242,9 +299,8 @@ async fn responses_lite_omits_hosted_tools_without_standalone_extensions() -> Re
test.submit_turn("Do not use hosted tools").await?;
let body = response_mock.single_request().body_json();
let tools = body["tools"]
.as_array()
.context("Responses request tools should be an array")?;
assert!(body.get("tools").is_none());
let tools = additional_tools(&body)?;
assert!(!has_hosted_tool(tools, "web_search"));
assert!(!has_hosted_tool(tools, "image_generation"));