Skip to content

armgit / codex

Public Imported from GitHub
0

Lightweight coding agent that runs in your terminal

Rust
107 KB
Raw Blame
#![allow(clippy::unwrap_used)]
use codex_api::WS_REQUEST_HEADER_TRACEPARENT_CLIENT_METADATA_KEY;
use codex_api::WS_REQUEST_HEADER_TRACESTATE_CLIENT_METADATA_KEY;
use codex_core::CodexResponsesMetadata;
use codex_core::ModelClient;
use codex_core::ModelClientSession;
use codex_core::Prompt;
use codex_core::ResponseEvent;
use codex_core::TurnInputRequest;
use codex_core::X_CODEX_ROUTING_HINT_HEADER;
use codex_core::X_RESPONSESAPI_INCLUDE_TIMING_METRICS_HEADER;
use codex_core::test_support::EmptyUserInstructionsProvider;
use codex_core::test_support::with_parent_turn;
use codex_extension_api::ExtensionRegistryBuilder;
use codex_features::Feature;
use codex_http_client::OutboundProxyPolicy;
use codex_login::AuthManager;
use codex_login::CodexAuth;
use codex_login::auth::AgentIdentityAuthPolicy;
use codex_model_provider_info::ModelProviderInfo;
use codex_model_provider_info::WireApi;
use codex_otel::MetricsClient;
use codex_otel::MetricsConfig;
use codex_otel::SessionTelemetry;
use codex_otel::TelemetryAuthMode;
use codex_otel::current_span_w3c_trace_context;
use codex_protocol::ResponseItemId;
use codex_protocol::SessionId;
use codex_protocol::ThreadId;
use codex_protocol::account::PlanType;
use codex_protocol::config_types::ReasoningSummary;
use codex_protocol::config_types::ServiceTier;
use codex_protocol::models::BaseInstructions;
use codex_protocol::models::ContentItem;
use codex_protocol::models::ExecutedToolCall;
use codex_protocol::models::FunctionCallOutputPayload;
use codex_protocol::models::ResponseInputItem;
use codex_protocol::models::ResponseItem;
use codex_protocol::models::ToolResultMetadata;
use codex_protocol::models::ToolResultSource;
use codex_protocol::models::ToolResultSources;
use codex_protocol::openai_models::ModelInfo;
use codex_protocol::openai_models::ModelServiceTier;
use codex_protocol::openai_models::ReasoningEffort as ReasoningEffortConfig;
use codex_protocol::protocol::EventMsg;
use codex_protocol::protocol::SessionSource;
use codex_protocol::protocol::W3cTraceContext;
use codex_protocol::user_input::UserInput;
use codex_rollout_trace::ConversationPart;
use codex_rollout_trace::InferenceTraceContext;
use codex_rollout_trace::RawTraceEventPayload;
use codex_rollout_trace::TraceWriter;
use codex_rollout_trace::replay_bundle;
use core_test_support::TestCodexResponsesRequestKind;
use core_test_support::ThreadIdle;
use core_test_support::load_default_config_for_test;
use core_test_support::responses::WebSocketConnectionConfig;
use core_test_support::responses::WebSocketTestServer;
use core_test_support::responses::ev_assistant_message;
use core_test_support::responses::ev_completed;
use core_test_support::responses::ev_response_created;
use core_test_support::responses::start_websocket_server;
use core_test_support::responses::start_websocket_server_with_headers;
use core_test_support::responses_metadata as test_responses_metadata;
use core_test_support::skip_if_no_network;
use core_test_support::test_codex::RecordingUserInstructionsProvider;
use core_test_support::test_codex::test_codex;
use core_test_support::tracing::install_test_tracing;
use core_test_support::wait_for_event;
use futures::StreamExt;
use opentelemetry_sdk::metrics::InMemoryMetricExporter;
use pretty_assertions::assert_eq;
use serde_json::json;
use std::sync::Arc;
use std::time::Duration;
use tempfile::TempDir;
use tracing::Instrument;
use tracing_test::traced_test;

const MODEL: &str = "gpt-5.4";
const OPENAI_BETA_HEADER: &str = "OpenAI-Beta";
const USER_AGENT_HEADER: &str = "user-agent";
const WS_V2_BETA_HEADER_VALUE: &str = "responses_websockets=2026-02-06";
const X_CLIENT_REQUEST_ID_HEADER: &str = "x-client-request-id";
const WS_REQUEST_HEADER_RESPONSES_LITE_CLIENT_METADATA_KEY: &str =
    "ws_request_header_x_openai_internal_codex_responses_lite";
const TEST_INSTALLATION_ID: &str = "11111111-1111-4111-8111-111111111111";
const TEST_WINDOW_ID: &str = "test-thread:0";
const X_CODEX_WS_STREAM_REQUEST_START_MS_CLIENT_METADATA_KEY: &str =
    "x-codex-ws-stream-request-start-ms";

fn assert_request_trace_matches(body: &serde_json::Value, expected_trace: &W3cTraceContext) {
    let client_metadata = body["client_metadata"]
        .as_object()
        .expect("missing client_metadata payload");
    let actual_traceparent = client_metadata
        .get(WS_REQUEST_HEADER_TRACEPARENT_CLIENT_METADATA_KEY)
        .and_then(serde_json::Value::as_str)
        .expect("missing traceparent");
    let expected_traceparent = expected_trace
        .traceparent
        .as_deref()
        .expect("missing expected traceparent");

    assert_eq!(actual_traceparent, expected_traceparent);
    assert_eq!(
        client_metadata
            .get(WS_REQUEST_HEADER_TRACESTATE_CLIENT_METADATA_KEY)
            .and_then(serde_json::Value::as_str),
        expected_trace.tracestate.as_deref()
    );
    assert!(
        body.get("trace").is_none(),
        "top-level trace should not be sent"
    );
}

struct WebsocketTestHarness {
    codex_home: TempDir,
    auth_manager: Arc<AuthManager>,
    client: ModelClient,
    outbound_proxy_policy: OutboundProxyPolicy,
    session_id: SessionId,
    thread_id: ThreadId,
    model_info: ModelInfo,
    effort: Option<ReasoningEffortConfig>,
    summary: ReasoningSummary,
    session_telemetry: SessionTelemetry,
}

fn responses_metadata(
    harness: &WebsocketTestHarness,
    turn_id: Option<&str>,
    request_kind: TestCodexResponsesRequestKind,
) -> CodexResponsesMetadata {
    test_responses_metadata(
        TEST_INSTALLATION_ID,
        &harness.session_id.to_string(),
        &harness.thread_id.to_string(),
        turn_id,
        TEST_WINDOW_ID.to_string(),
        &SessionSource::Exec,
        /*parent_thread_id*/ None,
        request_kind,
    )
}

fn turn_metadata(harness: &WebsocketTestHarness, turn_id: Option<&str>) -> CodexResponsesMetadata {
    responses_metadata(harness, turn_id, TestCodexResponsesRequestKind::Turn)
}

fn prewarm_metadata(
    harness: &WebsocketTestHarness,
    turn_id: Option<&str>,
) -> CodexResponsesMetadata {
    responses_metadata(harness, turn_id, TestCodexResponsesRequestKind::Prewarm)
}

fn websocket_connection_metadata(harness: &WebsocketTestHarness) -> CodexResponsesMetadata {
    responses_metadata(
        harness,
        /*turn_id*/ None,
        TestCodexResponsesRequestKind::WebsocketConnection,
    )
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_preserves_credit_usage_metadata() {
    skip_if_no_network!();

    let mut completed = ev_completed("resp-1");
    completed["response"]["usage_metadata"] = json!({ "amount": "0.12345678901234567890" });
    let expected_metadata = completed["response"]["usage"].clone();
    let server =
        start_websocket_server(vec![vec![vec![ev_response_created("resp-1"), completed]]]).await;
    let harness = websocket_harness_for_codex_backend(&server).await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);
    let responses_metadata = turn_metadata(&harness, /*turn_id*/ None);
    let mut stream = client_session
        .stream(
            &prompt,
            &harness.model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            /*service_tier*/ None,
            &responses_metadata,
            &InferenceTraceContext::disabled(),
        )
        .await
        .expect("websocket stream failed");

    let mut usage_metadata = None;
    while let Some(event) = stream.next().await {
        if let ResponseEvent::Completed {
            usage_metadata: metadata,
            ..
        } = event.expect("websocket stream failed")
        {
            usage_metadata = metadata;
            break;
        }
    }
    assert_eq!(
        usage_metadata,
        Some(codex_protocol::ResponseUsageMetadata {
            amount: Some("0.12345678901234567890".to_string()),
            metadata: Some(expected_metadata),
        }),
    );
    assert_eq!(server.single_connection().len(), 1);
    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_preserves_raw_tool_metadata_for_openai_custom_endpoint() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;
    let harness = websocket_harness_for_codex_backend(&server).await;
    let mut call = ExecutedToolCall::new("test_tool".to_string(), json!({ "query": "keep" }));
    call.set_tool_result_sources(ToolResultSources::new(vec![ToolResultSource {
        r#type: "test_resource".to_string(),
        id: "R1".to_string(),
    }]));
    call.set_tool_result_metadata(ToolResultMetadata::new(&json!({
        "private": "raw-result-metadata",
    })));
    let mut output = ResponseItem::from(ResponseInputItem::FunctionCallOutput {
        call_id: "tool-call".to_string(),
        output: FunctionCallOutputPayload::from_text("unchanged tool result".to_string()),
    });
    output.append_executed_tool_calls(vec![call]);
    output.mark_tool_calls_complete();
    let prompt = prompt_with_input(vec![output.clone()]);
    let expected = serde_json::to_value(&output).unwrap();

    let mut client_session = harness.client.new_session();
    stream_until_complete_with_model_info(
        &mut client_session,
        &harness,
        &prompt,
        &harness.model_info,
        "resp-1",
    )
    .await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 1);
    let body = connection.first().expect("missing request").body_json();
    assert_eq!(body["type"], "response.create");
    assert_eq!(body["input"], json!([expected]));
    assert_eq!(prompt.input, vec![output]);
    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_streams_request() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness = websocket_harness_for_codex_backend(&server).await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);

    stream_until_complete(&mut client_session, &harness, &prompt).await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 1);
    let body = connection.first().expect("missing request").body_json();

    assert_eq!(body["type"].as_str(), Some("response.create"));
    assert_eq!(body["model"].as_str(), Some(MODEL));
    assert_eq!(body["stream"], serde_json::Value::Bool(true));
    assert_eq!(body["input"].as_array().map(Vec::len), Some(1));
    let handshake = server.single_handshake();
    assert_eq!(
        handshake.header(OPENAI_BETA_HEADER),
        Some(WS_V2_BETA_HEADER_VALUE.to_string())
    );
    assert_eq!(
        handshake.header(X_CODEX_ROUTING_HINT_HEADER),
        Some(format!("model={MODEL}"))
    );
    assert_eq!(
        handshake.header(X_CLIENT_REQUEST_ID_HEADER),
        Some(harness.thread_id.to_string())
    );
    assert_eq!(
        handshake.header("session-id"),
        Some(harness.session_id.to_string())
    );
    assert_eq!(
        handshake.header("thread-id"),
        Some(harness.thread_id.to_string())
    );
    assert_eq!(
        handshake.header(USER_AGENT_HEADER),
        Some(codex_login::default_client::get_codex_user_agent())
    );
    assert_eq!(
        body["client_metadata"]["x-codex-installation-id"].as_str(),
        Some(TEST_INSTALLATION_ID)
    );
    let stream_request_start_ms = body["client_metadata"]
        [X_CODEX_WS_STREAM_REQUEST_START_MS_CLIENT_METADATA_KEY]
        .as_str()
        .expect("missing websocket stream request start timestamp")
        .parse::<i64>()
        .expect("websocket stream request start timestamp should be an integer");
    assert!(stream_request_start_ms > 0);

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_omits_routing_hint_for_provider_with_own_credentials() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let mut provider = websocket_provider(&server);
    provider.name = ModelProviderInfo::create_openai_provider(/*base_url*/ None).name;
    provider.experimental_bearer_token = Some("provider-specific-token".into());
    let harness = websocket_harness_with_provider_options_and_auth(
        provider,
        /*runtime_metrics_enabled*/ false,
        /*concurrent_reasoning_summaries_enabled*/ false,
        /*enabled_features*/ &[],
        Some(CodexAuth::create_dummy_chatgpt_auth_for_testing()),
    )
    .await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);

    stream_until_complete(&mut client_session, &harness, &prompt).await;

    let handshake = server.single_handshake();
    assert_eq!(
        handshake.header("authorization"),
        Some("Bearer provider-specific-token".to_string())
    );
    assert_eq!(handshake.header(X_CODEX_ROUTING_HINT_HEADER), None);

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_omits_unprefixed_item_ids_without_mutating_prompt() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness = websocket_harness_with_provider_options(
        websocket_provider(&server),
        /*runtime_metrics_enabled*/ false,
        /*concurrent_reasoning_summaries_enabled*/ false,
        /*enabled_features*/ &[],
    )
    .await;
    let mut client_session = harness.client.new_session();
    let mut prefixed = message_item("prefixed");
    prefixed.set_id(Some(ResponseItemId::with_suffix("msg", "existing")));
    let unprefixed = serde_json::from_value(json!({
        "type": "message",
        "id": "018f9e15-7a6a-7000-8000-000000000001",
        "role": "user",
        "content": [{"type": "input_text", "text": "legacy message"}],
    }))
    .expect("legacy response item should deserialize");
    let empty = serde_json::from_value(json!({
        "type": "message",
        "id": "",
        "role": "user",
        "content": [{"type": "input_text", "text": "empty-id message"}],
    }))
    .expect("response item with an empty id should deserialize");
    let prompt = prompt_with_input(vec![prefixed, unprefixed, empty]);

    stream_until_complete(&mut client_session, &harness, &prompt).await;

    let connection = server.single_connection();
    let body = connection.first().expect("missing request").body_json();
    assert_eq!(body["input"].as_array().map(Vec::len), Some(3));
    assert_eq!(body["input"][0]["id"].as_str(), Some("msg_existing"));
    assert_eq!(body["input"][1].get("id"), None);
    assert_eq!(body["input"][2].get("id"), None);
    assert_eq!(
        serde_json::to_value(&prompt.input).expect("prompt input should serialize")[1]["id"]
            .as_str(),
        Some("018f9e15-7a6a-7000-8000-000000000001")
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_streams_without_feature_flag_when_provider_supports_websockets() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness = websocket_harness_with_options(&server, /*runtime_metrics_enabled*/ false).await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);

    stream_until_complete(&mut client_session, &harness, &prompt).await;

    assert_eq!(server.handshakes().len(), 1);
    assert_eq!(server.single_connection().len(), 1);

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_streams_with_system_proxy_feature() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness = websocket_harness_with_provider_options(
        websocket_provider(&server),
        /*runtime_metrics_enabled*/ false,
        /*concurrent_reasoning_summaries_enabled*/ false,
        /*enabled_features*/ &[Feature::RespectSystemProxy],
    )
    .await;
    assert_eq!(
        harness.outbound_proxy_policy,
        OutboundProxyPolicy::RespectSystemProxy
    );
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);

    stream_until_complete(&mut client_session, &harness, &prompt).await;

    assert_eq!(server.handshakes().len(), 1);
    assert_eq!(server.single_connection().len(), 1);

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_reuses_connection_with_per_turn_trace_payloads() {
    skip_if_no_network!();

    let _trace_test_context = install_test_tracing("client-websocket-test");

    let server = start_websocket_server(vec![vec![
        vec![ev_response_created("resp-1"), ev_completed("resp-1")],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness(&server).await;
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let prompt_two = prompt_with_input(vec![message_item("again")]);

    let first_trace = {
        let mut client_session = harness.client.new_session();
        async {
            let expected_trace =
                current_span_w3c_trace_context().expect("current span should have trace context");
            stream_until_complete(&mut client_session, &harness, &prompt_one).await;
            expected_trace
        }
        .instrument(tracing::info_span!("client.websocket.turn_one"))
        .await
    };

    let second_trace = {
        let mut client_session = harness.client.new_session();
        async {
            let expected_trace =
                current_span_w3c_trace_context().expect("current span should have trace context");
            stream_until_complete(&mut client_session, &harness, &prompt_two).await;
            expected_trace
        }
        .instrument(tracing::info_span!("client.websocket.turn_two"))
        .await
    };

    assert_eq!(server.handshakes().len(), 1);
    assert_eq!(
        server.single_handshake().header(USER_AGENT_HEADER),
        Some(codex_login::default_client::get_codex_user_agent())
    );
    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);

    let first_request = connection
        .first()
        .expect("missing first request")
        .body_json();
    let second_request = connection
        .get(1)
        .expect("missing second request")
        .body_json();
    assert_request_trace_matches(&first_request, &first_trace);
    assert_request_trace_matches(&second_request, &second_trace);

    let first_traceparent = first_request["client_metadata"]
        [WS_REQUEST_HEADER_TRACEPARENT_CLIENT_METADATA_KEY]
        .as_str()
        .expect("missing first traceparent");
    let second_traceparent = second_request["client_metadata"]
        [WS_REQUEST_HEADER_TRACEPARENT_CLIENT_METADATA_KEY]
        .as_str()
        .expect("missing second traceparent");
    assert_ne!(first_traceparent, second_traceparent);

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_preconnect_reuses_connection_without_replacing_turn_trace() {
    skip_if_no_network!();

    let _trace_test_context = install_test_tracing("client-websocket-test");

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness = websocket_harness(&server).await;
    let mut client_session = harness.client.new_session();
    let responses_metadata = websocket_connection_metadata(&harness);
    client_session
        .preconnect_websocket(
            &harness.model_info,
            /*service_tier*/ None,
            &harness.session_telemetry,
            &responses_metadata,
        )
        .await
        .expect("websocket preconnect failed");
    assert!(server.single_connection().is_empty());
    let handshake = server.single_handshake();
    assert_eq!(
        handshake.header(USER_AGENT_HEADER),
        Some(codex_login::default_client::get_codex_user_agent())
    );
    assert_eq!(
        handshake.header("x-codex-window-id"),
        Some(TEST_WINDOW_ID.to_string())
    );
    assert_eq!(handshake.header("x-codex-turn-metadata"), None);
    assert_eq!(handshake.header(X_CODEX_ROUTING_HINT_HEADER), None);
    assert_eq!(
        handshake.header(OPENAI_BETA_HEADER),
        Some(WS_V2_BETA_HEADER_VALUE.to_string())
    );
    let prompt = prompt_with_input(vec![message_item("hello")]);

    let expected_trace = async {
        let expected_trace =
            current_span_w3c_trace_context().expect("current span should have trace context");
        stream_until_complete(&mut client_session, &harness, &prompt).await;
        expected_trace
    }
    .instrument(tracing::info_span!("client.websocket.request"))
    .await;

    assert_eq!(server.handshakes().len(), 1);
    let connection = server.single_connection();
    assert_eq!(connection.len(), 1);
    let request = connection.first().expect("missing request").body_json();
    assert_request_trace_matches(&request, &expected_trace);
    let turn_metadata: serde_json::Value = serde_json::from_str(
        request["client_metadata"]["x-codex-turn-metadata"]
            .as_str()
            .expect("turn metadata"),
    )
    .expect("valid turn metadata");
    assert_eq!(turn_metadata["request_kind"], "turn");

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_resume_prewarm_reuses_and_repairs_connection() -> anyhow::Result<()> {
    skip_if_no_network!(Ok(()));

    let server = start_websocket_server_with_headers(vec![
        WebSocketConnectionConfig {
            requests: vec![
                vec![ev_response_created("warm-1"), ev_completed("warm-1")],
                vec![ev_response_created("resp-1"), ev_completed("resp-1")],
                vec![
                    ev_response_created("resp-2"),
                    ev_assistant_message("msg_2", "ready to continue"),
                    ev_completed("resp-2"),
                ],
            ],
            response_headers: Vec::new(),
            accept_delay: None,
            close_after_requests: true,
        },
        WebSocketConnectionConfig {
            requests: vec![
                vec![ev_response_created("warm-2"), ev_completed("warm-2")],
                vec![ev_response_created("resp-3"), ev_completed("resp-3")],
            ],
            response_headers: Vec::new(),
            accept_delay: Some(Duration::from_millis(100)),
            close_after_requests: false,
        },
    ])
    .await;
    let mut extensions = ExtensionRegistryBuilder::new();
    extensions.thread_lifecycle_contributor(Arc::new(ThreadIdle));
    let instructions = Arc::new(RecordingUserInstructionsProvider::new(Arc::new(
        EmptyUserInstructionsProvider,
    )));
    let test = test_codex()
        .with_extensions(Arc::new(extensions.build()))
        .with_user_instructions_provider(instructions.clone())
        .build_with_websocket_server(&server)
        .await?;
    test.submit_text_turn("hello").await?;
    ThreadIdle::wait(&test.codex).await;

    // A healthy warm resume skips preparation; only the user turn reloads instructions.
    let instruction_loads = instructions.load_count();
    test.codex.prewarm_with_history().await;
    test.codex.prewarm_with_history().await;
    test.submit_text_turn("continue").await?;

    assert_eq!(instructions.load_count(), instruction_loads + 1);
    assert_eq!(server.handshakes().len(), 1);
    assert_eq!(
        server.single_handshake().header(USER_AGENT_HEADER),
        Some(codex_login::default_client::get_codex_user_agent())
    );
    assert_eq!(
        server.single_handshake().header("x-codex-window-id"),
        Some(format!("{}:0", test.session_configured.thread_id))
    );
    let connection = server.single_connection();
    assert_eq!(connection.len(), 3);
    let prewarm = connection[0].body_json();
    assert_eq!(prewarm["input"].as_array().unwrap().len(), 1);
    assert_eq!(prewarm["input"][0]["role"], "developer");
    assert_eq!(connection[2].body_json()["previous_response_id"], "resp-1");
    let mut expected_history = connection
        .iter()
        .flat_map(|request| request.body_json()["input"].as_array().unwrap().clone())
        .collect::<Vec<_>>();
    expected_history.push(serde_json::to_value(assistant_message_item(
        "2",
        "ready to continue",
    ))?);

    // Turn idle does not synchronize with the reader observing the server's close.
    // Retry resume until it sees the close; pending attempts must still share one socket.
    ThreadIdle::wait(&test.codex).await;
    let instruction_loads = instructions.load_count();
    let warmup = tokio::time::timeout(Duration::from_secs(10), async {
        loop {
            test.codex.prewarm_with_history().await;
            test.codex.prewarm_with_history().await;
            tokio::select! {
                request = server.wait_for_request(/*connection_index*/ 1, /*request_index*/ 0) => break request,
                _ = tokio::time::sleep(Duration::from_millis(10)) => {}
            }
        }
    })
    .await?;
    assert!(instructions.load_count() > instruction_loads);
    assert_eq!(warmup.body_json()["generate"], false);
    assert!(warmup.body_json().get("previous_response_id").is_none());
    let mut actual_history: Vec<ResponseItem> =
        serde_json::from_value(warmup.body_json()["input"].clone())?;
    let mut expected_history: Vec<ResponseItem> = serde_json::from_value(json!(expected_history))?;
    for item in actual_history.iter_mut().chain(&mut expected_history) {
        item.clear_internal_chat_message_metadata_passthrough();
    }
    assert_eq!(actual_history, expected_history);
    assert!(warmup.body_json().get("prompt_cache_options").is_none());

    test.submit_text_turn("continue after reconnect").await?;
    assert_eq!(server.handshakes().len(), 2);
    let connections = server.connections();
    assert_eq!(connections[1].len(), 2);
    assert_eq!(
        connections[1][1].body_json()["previous_response_id"],
        "warm-2"
    );

    test.codex.shutdown_and_wait().await?;
    server.shutdown().await;
    Ok(())
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_request_prewarm_reuses_connection() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![ev_response_created("warm-1"), ev_completed("warm-1")],
        vec![ev_response_created("resp-1"), ev_completed("resp-1")],
    ]])
    .await;

    let mut provider = websocket_provider(&server);
    provider.name = ModelProviderInfo::create_openai_provider(/*base_url*/ None).name;
    let harness = websocket_harness_with_provider_options(
        provider,
        /*runtime_metrics_enabled*/ true,
        /*concurrent_reasoning_summaries_enabled*/ true,
        /*enabled_features*/ &[],
    )
    .await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);
    let responses_metadata = prewarm_metadata(&harness, /*turn_id*/ None);
    client_session
        .prewarm_websocket(
            &prompt,
            &harness.model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            /*service_tier*/ None,
            &responses_metadata,
        )
        .await
        .expect("websocket prewarm failed");
    stream_until_complete(&mut client_session, &harness, &prompt).await;

    assert_eq!(server.handshakes().len(), 1);
    let handshake = server.single_handshake();
    assert_eq!(
        handshake.header(USER_AGENT_HEADER),
        Some(codex_login::default_client::get_codex_user_agent())
    );
    assert_eq!(
        handshake.header(OPENAI_BETA_HEADER),
        Some(WS_V2_BETA_HEADER_VALUE.to_string())
    );
    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let warmup = connection
        .first()
        .expect("missing warmup request")
        .body_json();
    let follow_up = connection
        .get(1)
        .expect("missing follow-up request")
        .body_json();

    assert_eq!(warmup["type"].as_str(), Some("response.create"));
    assert_eq!(warmup["generate"].as_bool(), Some(false));
    assert_eq!(
        warmup["input"],
        serde_json::to_value(&prompt.input).unwrap()
    );
    assert_eq!(
        warmup["stream_options"]["reasoning_summary_delivery"].as_str(),
        Some("sequential_cutoff")
    );
    assert_eq!(warmup["tools"], serde_json::json!([]));
    let warmup_turn_metadata: serde_json::Value = serde_json::from_str(
        warmup["client_metadata"]["x-codex-turn-metadata"]
            .as_str()
            .expect("warmup turn metadata"),
    )
    .expect("valid warmup turn metadata");
    assert_eq!(
        warmup_turn_metadata["request_kind"].as_str(),
        Some("prewarm")
    );
    assert_eq!(follow_up["type"].as_str(), Some("response.create"));
    assert_eq!(follow_up["previous_response_id"].as_str(), Some("warm-1"));
    assert_eq!(follow_up["input"], serde_json::json!([]));
    let follow_up_turn_metadata: serde_json::Value = serde_json::from_str(
        follow_up["client_metadata"]["x-codex-turn-metadata"]
            .as_str()
            .expect("follow-up turn metadata"),
    )
    .expect("valid follow-up turn metadata");
    assert_eq!(follow_up_turn_metadata["request_kind"], "turn");
    assert_eq!(
        follow_up["stream_options"]["reasoning_summary_delivery"].as_str(),
        Some("sequential_cutoff")
    );

    assert_continuation_metrics(
        &harness.session_telemetry,
        &[
            (["full", "no_previous_request", "warmup"], 1),
            (["incremental", "incremental", "generation"], 1),
        ],
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_request_prewarm_uses_caller_supplied_metadata() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("warm-1"),
        ev_completed("warm-1"),
    ]]])
    .await;

    let harness = websocket_harness_with_options(&server, /*runtime_metrics_enabled*/ true).await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);
    let responses_metadata = turn_metadata(&harness, /*turn_id*/ None);
    client_session
        .prewarm_websocket(
            &prompt,
            &harness.model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            /*service_tier*/ None,
            &responses_metadata,
        )
        .await
        .expect("websocket prewarm failed");

    let warmup = server
        .single_connection()
        .first()
        .expect("missing warmup request")
        .body_json();
    let warmup_turn_metadata: serde_json::Value = serde_json::from_str(
        warmup["client_metadata"]["x-codex-turn-metadata"]
            .as_str()
            .expect("warmup turn metadata"),
    )
    .expect("valid warmup turn metadata");
    assert_eq!(warmup_turn_metadata["request_kind"].as_str(), Some("turn"));

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_request_prewarm_traces_logical_request() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![ev_response_created("warm-1"), ev_completed("warm-1")],
        vec![ev_response_created("resp-1"), ev_completed("resp-1")],
    ]])
    .await;

    let harness = websocket_harness_with_options(&server, /*runtime_metrics_enabled*/ true).await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);
    let prewarm_responses_metadata = prewarm_metadata(&harness, /*turn_id*/ None);

    client_session
        .prewarm_websocket(
            &prompt,
            &harness.model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            /*service_tier*/ None,
            &prewarm_responses_metadata,
        )
        .await
        .expect("websocket prewarm failed");

    let trace_dir = TempDir::new().expect("trace dir");
    let writer = Arc::new(
        TraceWriter::create(
            trace_dir.path(),
            "trace-1".to_string(),
            harness.session_id.to_string(),
            harness.thread_id.to_string(),
        )
        .expect("trace writer"),
    );
    writer
        .append(RawTraceEventPayload::ThreadStarted {
            thread_id: harness.thread_id.to_string(),
            agent_path: "/root".to_string(),
            metadata_payload: None,
        })
        .expect("thread started");
    writer
        .append(RawTraceEventPayload::CodexTurnStarted {
            codex_turn_id: "turn-1".to_string(),
            thread_id: harness.thread_id.to_string(),
        })
        .expect("turn started");

    let inference_trace = InferenceTraceContext::enabled(
        writer,
        harness.thread_id.to_string(),
        "turn-1".to_string(),
        harness.model_info.slug.clone(),
        "test-provider".to_string(),
    );

    let responses_metadata = turn_metadata(&harness, /*turn_id*/ None);
    let mut stream = client_session
        .stream(
            &prompt,
            &harness.model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            /*service_tier*/ None,
            &responses_metadata,
            &inference_trace,
        )
        .await
        .expect("websocket stream failed");

    while let Some(event) = stream.next().await {
        if matches!(event, Ok(ResponseEvent::Completed { .. })) {
            break;
        }
    }

    let connection = server.single_connection();
    let follow_up = connection
        .get(1)
        .expect("missing follow-up request")
        .body_json();
    assert_eq!(follow_up["previous_response_id"].as_str(), Some("warm-1"));
    assert_eq!(follow_up["input"], serde_json::json!([]));

    let rollout = replay_bundle(trace_dir.path()).expect("replay trace");
    let inference = rollout
        .inference_calls
        .values()
        .next()
        .expect("inference should be present");
    assert_eq!(inference.request_item_ids.len(), 1);
    assert_eq!(
        rollout.conversation_items[&inference.request_item_ids[0]]
            .body
            .parts,
        vec![ConversationPart::Text {
            text: "hello".to_string(),
        }],
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_reuses_connection_after_session_drop() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![ev_response_created("resp-1"), ev_completed("resp-1")],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness(&server).await;
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let prompt_two = prompt_with_input(vec![message_item("again")]);

    {
        let mut client_session = harness.client.new_session();
        stream_until_complete(&mut client_session, &harness, &prompt_one).await;
    }

    let mut client_session = harness.client.new_session();
    stream_until_complete(&mut client_session, &harness, &prompt_two).await;

    assert_eq!(server.handshakes().len(), 1);
    assert_eq!(server.single_connection().len(), 2);

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_reconnects_after_account_switch() {
    skip_if_no_network!();

    for new_turn_on_switch in [false, true] {
        // Keep the first socket open so only the account switch forces a reconnect.
        let server = start_websocket_server(
            [
                (1..=5, "first-account-state"),
                (3..=5, "second-account-state"),
            ]
            .into_iter()
            .map(|(responses, turn_state)| {
                responses
                    .map(|index| {
                        let id = format!("resp-{index}");
                        vec![
                            ev_response_created(&id),
                            json!({
                                "type": "response.metadata",
                                "headers": {"x-codex-turn-state": turn_state},
                            }),
                            ev_completed(&id),
                        ]
                    })
                    .collect()
            })
            .collect(),
        )
        .await;
        let harness = websocket_harness_for_codex_backend(&server).await;
        let mut client_session = harness.client.new_session();
        let mut input = Vec::new();

        for index in 1..=5 {
            if index == 3 {
                let mut tokens = harness
                    .auth_manager
                    .auth_cached()
                    .unwrap()
                    .get_token_data()
                    .unwrap();
                tokens.id_token.raw_jwt = "e30.e30.signature".into();
                tokens.account_id = Some("second-account".into());
                tokens.access_token = "second-account-token".into();
                std::fs::write(
                    harness.codex_home.path().join("auth.json"),
                    serde_json::to_vec(&json!({
                        "auth_mode": "chatgpt",
                        "tokens": tokens,
                        "last_refresh": chrono::Utc::now(),
                    }))
                    .unwrap(),
                )
                .unwrap();
                harness.auth_manager.reload().await;
            }
            if index == 5 || (index == 3 && new_turn_on_switch) {
                drop(client_session);
                client_session = harness.client.new_session();
                if index == 3 {
                    client_session
                        .preconnect_websocket(
                            &harness.model_info,
                            /*service_tier*/ None,
                            &harness.session_telemetry,
                            &websocket_connection_metadata(&harness),
                        )
                        .await
                        .expect("new account should establish its own socket during preconnect");
                }
            }
            input.push(message_item(&format!("request {index}")));
            stream_until_complete_with_model_info(
                &mut client_session,
                &harness,
                &prompt_with_input(input.clone()),
                &harness.model_info,
                &format!("resp-{index}"),
            )
            .await;
        }

        assert_eq!(
            server
                .handshakes()
                .iter()
                .map(|handshake| (
                    handshake.header("chatgpt-account-id"),
                    handshake.header("authorization"),
                ))
                .collect::<Vec<_>>(),
            vec![
                (
                    Some("account_id".into()),
                    Some("Bearer Access Token".into())
                ),
                (
                    Some("second-account".into()),
                    Some("Bearer second-account-token".into())
                ),
            ],
        );
        assert_eq!(
            server
                .connections()
                .iter()
                .map(|requests| requests
                    .iter()
                    .map(|request| {
                        let body = request.body_json();
                        (
                            body["previous_response_id"].clone(),
                            body["input"].as_array().unwrap().len(),
                            body["client_metadata"]["x-codex-turn-state"].clone(),
                        )
                    })
                    .collect::<Vec<_>>())
                .collect::<Vec<_>>(),
            vec![
                vec![
                    (json!(null), 1, json!(null)),
                    (json!("resp-1"), 1, json!("first-account-state")),
                ],
                vec![
                    (json!(null), 3, json!(null)),
                    (json!("resp-3"), 1, json!("second-account-state")),
                    (json!("resp-4"), 1, json!(null)),
                ],
            ],
            "new_turn_on_switch={new_turn_on_switch}",
        );
        assert_continuation_metrics(
            &harness.session_telemetry,
            &[
                (["full", "other", "generation"], 1),
                (["full", "no_previous_request", "generation"], 1),
                (["incremental", "incremental", "generation"], 3),
            ],
        );
        server.shutdown().await;
    }
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_sends_responses_lite_metadata_per_request() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![ev_response_created("normal-1"), ev_completed("normal-1")],
        vec![ev_response_created("lite-1"), ev_completed("lite-1")],
        vec![ev_response_created("normal-2"), ev_completed("normal-2")],
    ]])
    .await;

    let harness = websocket_harness(&server).await;
    let normal_model_info = harness.model_info.clone();
    let mut lite_model_info = normal_model_info.clone();
    lite_model_info.use_responses_lite = true;
    let mut session = harness.client.new_session();

    stream_until_complete_with_model_info(
        &mut session,
        &harness,
        &prompt_with_input(vec![message_item("normal one")]),
        &normal_model_info,
        "normal-1",
    )
    .await;
    stream_until_complete_with_model_info(
        &mut session,
        &harness,
        &prompt_with_input(vec![message_item("lite")]),
        &lite_model_info,
        "lite-1",
    )
    .await;
    stream_until_complete_with_model_info(
        &mut session,
        &harness,
        &prompt_with_input(vec![message_item("normal two")]),
        &normal_model_info,
        "normal-2",
    )
    .await;

    let connection = server.single_connection();
    assert_eq!(
        connection
            .iter()
            .map(|request| {
                let body = request.body_json();
                json!({
                    "responses_lite": body["client_metadata"]
                        .get(WS_REQUEST_HEADER_RESPONSES_LITE_CLIENT_METADATA_KEY),
                    "reasoning_context": body["reasoning"].get("context"),
                    "parallel_tool_calls": body["parallel_tool_calls"],
                })
            })
            .collect::<Vec<_>>(),
        vec![
            json!({
                "responses_lite": null,
                "reasoning_context": null,
                "parallel_tool_calls": false,
            }),
            json!({
                "responses_lite": "true",
                "reasoning_context": "all_turns",
                "parallel_tool_calls": false,
            }),
            json!({
                "responses_lite": null,
                "reasoning_context": null,
                "parallel_tool_calls": false,
            }),
        ]
    );

    server.shutdown().await;
}

#[test_case::test_case(MODEL, Some("ultrafast"), Some("ultrafast"); "matching hint")]
#[test_case::test_case(MODEL, None, Some("ultrafast"); "tier selected after handshake")]
#[test_case::test_case("gpt-5.2", Some("ultrafast"), None; "model and tier changed after handshake")]
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_prewarm_reuses_advisory_model_and_tier_routing_hint(
    warmup_model: &str,
    preconnect_tier: Option<&str>,
    warmup_tier: Option<&str>,
) {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness = websocket_harness_for_codex_backend(&server).await;
    let mut model_info = harness.model_info.clone();
    model_info.service_tiers.push(ModelServiceTier {
        id: "ultrafast".to_string(),
        name: "Ultrafast".to_string(),
        description: "Ultrafast processing".to_string(),
    });
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);
    let responses_metadata = prewarm_metadata(&harness, /*turn_id*/ None);
    client_session
        .preconnect_websocket(
            &model_info,
            preconnect_tier.map(str::to_string),
            &harness.session_telemetry,
            &responses_metadata,
        )
        .await
        .expect("eager handshake should connect");
    let handshake = server.single_handshake();
    assert_eq!(
        handshake.header(X_CODEX_ROUTING_HINT_HEADER),
        Some(match preconnect_tier {
            Some(tier) => format!("model={MODEL};tier={tier}"),
            None => format!("model={MODEL}"),
        })
    );
    assert_eq!(
        handshake.header(OPENAI_BETA_HEADER),
        Some(WS_V2_BETA_HEADER_VALUE.to_string())
    );
    model_info.slug = warmup_model.to_string();
    client_session
        .prewarm_websocket(
            &prompt,
            &model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            warmup_tier.map(str::to_string),
            &responses_metadata,
        )
        .await
        .expect("websocket prewarm failed");

    assert_eq!(server.handshakes().len(), 1);
    let connection = server.single_connection();
    assert_eq!(connection.len(), 1);
    let prewarm = connection
        .first()
        .expect("missing prewarm request")
        .body_json();
    assert_eq!(prewarm["type"], "response.create");
    assert_eq!(prewarm["generate"], false);
    assert_eq!(prewarm["model"], warmup_model);
    assert_eq!(prewarm["service_tier"].as_str(), warmup_tier);
    assert_eq!(prewarm.get("previous_response_id"), None);

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_v2_requests_use_v2_when_provider_supports_websockets() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-1"),
            ev_assistant_message("msg_1", "assistant output"),
            ev_completed("resp-1"),
        ],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness_with_options(&server, /*runtime_metrics_enabled*/ true).await;
    let mut client_session = harness.client.new_session();
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let prompt_two = prompt_with_input(vec![
        message_item("hello"),
        assistant_message_item("1", "assistant output"),
        message_item("second"),
    ]);

    stream_until_complete(&mut client_session, &harness, &prompt_one).await;
    stream_until_complete(&mut client_session, &harness, &prompt_two).await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let second = connection.get(1).expect("missing request").body_json();
    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second["previous_response_id"].as_str(), Some("resp-1"));
    assert_eq!(
        second["input"],
        serde_json::to_value(&prompt_two.input[2..]).unwrap()
    );

    let handshake = server.single_handshake();
    let openai_beta_header = handshake
        .header(OPENAI_BETA_HEADER)
        .expect("missing OpenAI-Beta header");
    assert!(
        openai_beta_header
            .split(',')
            .map(str::trim)
            .any(|value| value == WS_V2_BETA_HEADER_VALUE)
    );
    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_v2_incremental_requests_are_reused_across_turns() {
    skip_if_no_network!();

    let mut assistant_output_with_turn_id = ev_assistant_message("msg_1", "assistant output");
    assistant_output_with_turn_id["item"]["internal_chat_message_metadata_passthrough"] =
        json!({ "turn_id": "turn-1" });

    let assistant_output_without_turn_id = ev_assistant_message("msg_2", "second assistant output");

    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-1"),
            assistant_output_with_turn_id,
            ev_completed("resp-1"),
        ],
        vec![
            ev_response_created("resp-2"),
            assistant_output_without_turn_id,
            ev_completed("resp-2"),
        ],
        vec![ev_response_created("resp-3"), ev_completed("resp-3")],
    ]])
    .await;

    // use OpenAI provider to check metadata logic
    let mut provider = websocket_provider(&server);
    provider.name = ModelProviderInfo::create_openai_provider(/*base_url*/ None).name;
    let harness = websocket_harness_with_provider_options(
        provider,
        /*runtime_metrics_enabled*/ false,
        /*concurrent_reasoning_summaries_enabled*/ false,
        /*enabled_features*/ &[],
    )
    .await;

    // Turn one: initiate
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    {
        let mut client_session = harness.client.new_session();
        stream_until_complete(&mut client_session, &harness, &prompt_one).await;
    }

    // Turn two: the response and reconstructed history have matching metadata.
    let mut first_assistant_output = assistant_message_item("1", "assistant output");
    first_assistant_output.set_turn_id_if_missing("turn-1");

    let prompt_two = prompt_with_input(vec![
        message_item("hello"),
        first_assistant_output.clone(),
        message_item("second"),
    ]);

    {
        let mut client_session = harness.client.new_session();
        stream_until_complete(&mut client_session, &harness, &prompt_two).await;
    }

    // Turn three: the reconstructed history has metadata that the response omitted.
    let mut second_assistant_output = assistant_message_item("2", "second assistant output");
    second_assistant_output.set_turn_id_if_missing("turn-2");

    let prompt_three = prompt_with_input(vec![
        message_item("hello"),
        first_assistant_output,
        message_item("second"),
        second_assistant_output,
        message_item("third"),
    ]);

    {
        let mut client_session = harness.client.new_session();
        stream_until_complete(&mut client_session, &harness, &prompt_three).await;
    }

    assert_eq!(server.handshakes().len(), 1);
    let connection = server.single_connection();
    assert_eq!(connection.len(), 3);

    // validate second turn
    let second = connection.get(1).expect("missing request").body_json();
    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second["previous_response_id"].as_str(), Some("resp-1"));
    assert_eq!(
        second["input"],
        serde_json::to_value(&prompt_two.input[2..]).unwrap()
    );

    // validate third turn
    let third = connection.get(2).expect("missing request").body_json();
    assert_eq!(third["type"].as_str(), Some("response.create"));
    assert_eq!(third["previous_response_id"].as_str(), Some("resp-2"));
    assert_eq!(
        third["input"],
        serde_json::to_value(&prompt_three.input[4..]).unwrap()
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_v2_wins_when_both_features_enabled() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-1"),
            ev_assistant_message("msg_1", "assistant output"),
            ev_completed("resp-1"),
        ],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness_with_options(&server, /*runtime_metrics_enabled*/ false).await;
    let mut client_session = harness.client.new_session();
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let prompt_two = prompt_with_input(vec![
        message_item("hello"),
        assistant_message_item("1", "assistant output"),
        message_item("second"),
    ]);

    stream_until_complete(&mut client_session, &harness, &prompt_one).await;
    stream_until_complete(&mut client_session, &harness, &prompt_two).await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let second = connection.get(1).expect("missing request").body_json();
    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second["previous_response_id"].as_str(), Some("resp-1"));
    assert_eq!(
        second["input"],
        serde_json::to_value(&prompt_two.input[2..]).unwrap()
    );

    let handshake = server.single_handshake();
    let openai_beta_header = handshake
        .header(OPENAI_BETA_HEADER)
        .expect("missing OpenAI-Beta header");
    assert!(
        openai_beta_header
            .split(',')
            .map(str::trim)
            .any(|value| value == WS_V2_BETA_HEADER_VALUE)
    );
    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
#[traced_test]
async fn responses_websocket_emits_websocket_telemetry_events() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness = websocket_harness(&server).await;
    harness.session_telemetry.reset_runtime_metrics();
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);

    stream_until_complete(&mut client_session, &harness, &prompt).await;

    tokio::time::sleep(Duration::from_millis(10)).await;

    let summary = harness
        .session_telemetry
        .runtime_metrics_summary()
        .expect("runtime metrics summary");
    assert_eq!(summary.api_calls.count, 0);
    assert_eq!(summary.streaming_events.count, 0);
    assert_eq!(summary.websocket_calls.count, 1);
    assert_eq!(summary.websocket_events.count, 2);

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_includes_timing_metrics_header_when_runtime_metrics_enabled() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        serde_json::json!({
            "type": "responsesapi.websocket_timing",
            "timing_metrics": {
                "responses_duration_excl_engine_and_client_tool_time_ms": 120,
                "engine_service_total_ms": 450,
                "engine_iapi_ttft_total_ms": 310,
                "engine_service_ttft_total_ms": 340,
                "engine_iapi_tbt_across_engine_calls_ms": 220,
                "engine_service_tbt_across_engine_calls_ms": 260
            }
        }),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness =
        websocket_harness_with_runtime_metrics(&server, /*runtime_metrics_enabled*/ true).await;
    harness.session_telemetry.reset_runtime_metrics();
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);

    stream_until_complete(&mut client_session, &harness, &prompt).await;
    tokio::time::sleep(Duration::from_millis(10)).await;

    let handshake = server.single_handshake();
    assert_eq!(
        handshake.header(X_RESPONSESAPI_INCLUDE_TIMING_METRICS_HEADER),
        Some("true".to_string())
    );

    let summary = harness
        .session_telemetry
        .runtime_metrics_summary()
        .expect("runtime metrics summary");
    assert_eq!(summary.responses_api_overhead_ms, 120);
    assert_eq!(summary.responses_api_inference_time_ms, 450);
    assert_eq!(summary.responses_api_engine_iapi_ttft_ms, 310);
    assert_eq!(summary.responses_api_engine_service_ttft_ms, 340);
    assert_eq!(summary.responses_api_engine_iapi_tbt_ms, 220.0);
    assert_eq!(summary.responses_api_engine_service_tbt_ms, 260.0);

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_omits_timing_metrics_header_when_runtime_metrics_disabled() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness =
        websocket_harness_with_runtime_metrics(&server, /*runtime_metrics_enabled*/ false).await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);

    stream_until_complete(&mut client_session, &harness, &prompt).await;

    let handshake = server.single_handshake();
    assert_eq!(
        handshake.header(X_RESPONSESAPI_INCLUDE_TIMING_METRICS_HEADER),
        None
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_emits_reasoning_included_event() {
    skip_if_no_network!();

    let server = start_websocket_server_with_headers(vec![WebSocketConnectionConfig {
        requests: vec![vec![ev_response_created("resp-1"), ev_completed("resp-1")]],
        response_headers: vec![("X-Reasoning-Included".to_string(), "true".to_string())],
        accept_delay: None,
        close_after_requests: true,
    }])
    .await;

    let harness = websocket_harness(&server).await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);
    let responses_metadata = turn_metadata(&harness, /*turn_id*/ None);

    let mut stream = client_session
        .stream(
            &prompt,
            &harness.model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            /*service_tier*/ None,
            &responses_metadata,
            &codex_rollout_trace::InferenceTraceContext::disabled(),
        )
        .await
        .expect("websocket stream failed");

    let mut saw_reasoning_included = false;
    while let Some(event) = stream.next().await {
        match event.expect("event") {
            ResponseEvent::ServerReasoningIncluded(true) => {
                saw_reasoning_included = true;
            }
            ResponseEvent::Completed { .. } => break,
            _ => {}
        }
    }

    assert!(saw_reasoning_included);
    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_emits_rate_limit_events() {
    skip_if_no_network!();

    let rate_limit_event = json!({
        "type": "codex.rate_limits",
        "plan_type": "plus",
        "rate_limits": {
            "allowed": true,
            "limit_reached": false,
            "primary": {
                "used_percent": 42,
                "window_minutes": 60,
                "reset_at": 1700000000
            },
            "secondary": null
        },
        "code_review_rate_limits": null,
        "credits": {
            "has_credits": true,
            "unlimited": false,
            "balance": "123"
        },
        "promo": null
    });

    let server = start_websocket_server_with_headers(vec![WebSocketConnectionConfig {
        requests: vec![vec![
            json!({
                "type": "codex.response.metadata",
                "headers": {"x-models-etag": "etag-123"},
            }),
            rate_limit_event,
            ev_response_created("resp-1"),
            ev_completed("resp-1"),
        ]],
        response_headers: vec![("X-Reasoning-Included".to_string(), "true".to_string())],
        accept_delay: None,
        close_after_requests: true,
    }])
    .await;

    let harness = websocket_harness(&server).await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);
    let responses_metadata = turn_metadata(&harness, /*turn_id*/ None);

    let mut stream = client_session
        .stream(
            &prompt,
            &harness.model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            /*service_tier*/ None,
            &responses_metadata,
            &codex_rollout_trace::InferenceTraceContext::disabled(),
        )
        .await
        .expect("websocket stream failed");

    let mut saw_rate_limits = None;
    let mut saw_models_etag = None;
    let mut saw_reasoning_included = false;

    while let Some(event) = stream.next().await {
        match event.expect("event") {
            ResponseEvent::RateLimits(snapshot) => {
                saw_rate_limits = Some(snapshot);
            }
            ResponseEvent::ModelsEtag(etag) => {
                saw_models_etag = Some(etag);
            }
            ResponseEvent::ServerReasoningIncluded(true) => {
                saw_reasoning_included = true;
            }
            ResponseEvent::Completed { .. } => break,
            _ => {}
        }
    }

    let rate_limits = saw_rate_limits.expect("missing rate limits");
    let primary = rate_limits.primary.expect("missing primary window");
    assert_eq!(primary.used_percent, 42.0);
    assert_eq!(primary.window_minutes, Some(60));
    assert_eq!(primary.resets_at, Some(1_700_000_000));
    assert_eq!(rate_limits.plan_type, Some(PlanType::Plus));
    let credits = rate_limits.credits.expect("missing credits");
    assert!(credits.has_credits);
    assert!(!credits.unlimited);
    assert_eq!(credits.balance.as_deref(), Some("123"));
    assert_eq!(saw_models_etag.as_deref(), Some("etag-123"));
    assert!(saw_reasoning_included);

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_usage_limit_error_emits_rate_limit_event() {
    skip_if_no_network!();

    let usage_limit_error = json!({
        "type": "error",
        "status": 429,
        "error": {
            "type": "usage_limit_reached",
            "message": "The usage limit has been reached",
            "plan_type": "pro",
            "resets_at": 1704067242,
            "resets_in_seconds": 1234
        },
        "headers": {
            "x-codex-primary-used-percent": "100.0",
            "x-codex-secondary-used-percent": "87.5",
            "x-codex-primary-over-secondary-limit-percent": "95.0",
            "x-codex-primary-window-minutes": "15",
            "x-codex-secondary-window-minutes": "60",
            "x-codex-credits-has-credits": "true",
            "x-codex-credits-unlimited": "false",
            "x-codex-credits-balance": "",
            "x-codex-rate-limit-reached-type": "workspace_member_usage_limit_reached"
        }
    });

    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-prewarm"),
            ev_completed("resp-prewarm"),
        ],
        vec![usage_limit_error],
    ]])
    .await;
    let mut builder = test_codex().with_config(|config| {
        config.model_provider.request_max_retries = Some(0);
        config.model_provider.stream_max_retries = Some(0);
    });
    let test = builder
        .build_with_websocket_server(&server)
        .await
        .expect("build websocket codex");

    let submission = test
        .codex
        .start_or_steer_turn(TurnInputRequest::user_input(vec![UserInput::Text {
            text: "hello".into(),
            text_elements: Vec::new(),
        }]))
        .await
        .expect("submission should succeed while emitting usage limit error events");

    let token_event =
        wait_for_event(&test.codex, |msg| matches!(msg, EventMsg::TokenCount(_))).await;
    let EventMsg::TokenCount(event) = token_event else {
        unreachable!();
    };

    let event_json = serde_json::to_value(&event).expect("serialize token count event");
    pretty_assertions::assert_eq!(
        event_json,
        json!({
            "info": null,
            "rate_limits": {
                "limit_id": "codex",
                "limit_name": null,
                "primary": {
                    "used_percent": 100.0,
                    "window_minutes": 15,
                    "resets_at": null
                },
                "secondary": {
                    "used_percent": 87.5,
                    "window_minutes": 60,
                    "resets_at": null
                },
                "credits": {
                    "has_credits": true,
                    "unlimited": false,
                    "balance": null
                },
                "individual_limit": null,
                "spend_control_reached": null,
                "plan_type": null,
                "rate_limit_reached_type": "workspace_member_usage_limit_reached"
            }
        })
    );

    let error_event = wait_for_event(&test.codex, |msg| matches!(msg, EventMsg::Error(_))).await;
    let EventMsg::Error(error_event) = error_event else {
        unreachable!();
    };
    assert!(
        error_event.message.contains("spend cap set by the owner"),
        "unexpected error message for submission {submission:?}: {}",
        error_event.message
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_invalid_request_error_with_status_is_forwarded() {
    skip_if_no_network!();

    let invalid_request_error = json!({
        "type": "error",
        "status": 400,
        "error": {
            "type": "invalid_request_error",
            "message": "Model 'castor-raikou-0205-ev3' does not support image inputs."
        }
    });

    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-prewarm"),
            ev_completed("resp-prewarm"),
        ],
        vec![invalid_request_error],
    ]])
    .await;
    let mut builder = test_codex().with_config(|config| {
        config.model_provider.request_max_retries = Some(0);
        config.model_provider.stream_max_retries = Some(0);
    });
    let test = builder
        .build_with_websocket_server(&server)
        .await
        .expect("build websocket codex");

    let submission = test
        .codex
        .start_or_steer_turn(TurnInputRequest::user_input(vec![UserInput::Text {
            text: "hello".into(),
            text_elements: Vec::new(),
        }]))
        .await
        .expect("submission should succeed while emitting invalid request events");

    let error_event = wait_for_event(&test.codex, |msg| matches!(msg, EventMsg::Error(_))).await;
    let EventMsg::Error(error_event) = error_event else {
        unreachable!();
    };
    assert!(
        error_event
            .message
            .to_lowercase()
            .contains("does not support image inputs"),
        "unexpected error message for submission {submission:?}: {}",
        error_event.message
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_connection_limit_error_reconnects_and_completes() {
    skip_if_no_network!();

    let websocket_connection_limit_error = json!({
        "type": "error",
        "status": 400,
        "error": {
            "type": "invalid_request_error",
            "code": "websocket_connection_limit_reached",
            "message": "Responses websocket connection limit reached (60 minutes). Create a new websocket connection to continue."
        }
    });

    let server = start_websocket_server(vec![
        vec![vec![websocket_connection_limit_error]],
        vec![vec![ev_response_created("resp-1"), ev_completed("resp-1")]],
    ])
    .await;
    let mut builder = test_codex().with_config(|config| {
        config.model_provider.request_max_retries = Some(0);
        config.model_provider.stream_max_retries = Some(1);
    });
    let test = builder
        .build_with_websocket_server(&server)
        .await
        .expect("build websocket codex");

    test.submit_turn("hello")
        .await
        .expect("submission should reconnect after websocket connection limit error");

    let total_websocket_requests: usize = server.connections().iter().map(Vec::len).sum();
    assert_eq!(total_websocket_requests, 2);
    let handshake_user_agents: Vec<_> = server
        .handshakes()
        .iter()
        .map(|handshake| handshake.header(USER_AGENT_HEADER))
        .collect();
    assert_eq!(
        handshake_user_agents,
        vec![
            Some(codex_login::default_client::get_codex_user_agent()),
            Some(codex_login::default_client::get_codex_user_agent()),
        ]
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_uses_incremental_create_on_prefix() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-1"),
            ev_assistant_message("msg_1", "assistant output"),
            ev_completed("resp-1"),
        ],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness(&server).await;
    let mut client_session = harness.client.new_session();
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let prompt_two = prompt_with_input(vec![
        message_item("hello"),
        assistant_message_item("1", "assistant output"),
        message_item("second"),
    ]);

    stream_until_complete(&mut client_session, &harness, &prompt_one).await;
    stream_until_complete(&mut client_session, &harness, &prompt_two).await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let first = connection.first().expect("missing request").body_json();
    let second = connection.get(1).expect("missing request").body_json();

    assert_eq!(first["type"].as_str(), Some("response.create"));
    assert_eq!(first["model"].as_str(), Some(MODEL));
    assert_eq!(first["stream"], serde_json::Value::Bool(true));
    assert_eq!(first["input"].as_array().map(Vec::len), Some(1));
    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second["previous_response_id"].as_str(), Some("resp-1"));
    assert_eq!(
        second["input"],
        serde_json::to_value(&prompt_two.input[2..]).expect("serialize incremental items")
    );

    assert_continuation_metrics(
        &harness.session_telemetry,
        &[
            (["full", "no_previous_request", "generation"], 1),
            (["incremental", "incremental", "generation"], 1),
        ],
    );
    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_lite_websocket_uses_incremental_create_on_prefix() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-1"),
            ev_assistant_message("msg_1", "assistant output"),
            ev_completed("resp-1"),
        ],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let mut harness = websocket_harness(&server).await;
    harness.model_info.use_responses_lite = true;
    let mut client_session = harness.client.new_session();
    let mut initial_item = message_item("hello");
    initial_item.set_id(Some(ResponseItemId::with_suffix("msg", "supplied")));
    let prompt_one = prompt_with_input_and_instructions(vec![initial_item.clone()], "base");
    let prompt_two = prompt_with_input_and_instructions(
        vec![
            initial_item,
            assistant_message_item("1", "assistant output"),
            message_item("second"),
        ],
        "base",
    );

    stream_until_complete(&mut client_session, &harness, &prompt_one).await;
    stream_until_complete(&mut client_session, &harness, &prompt_two).await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let first = connection.first().expect("missing request").body_json();
    let second = connection.get(1).expect("missing request").body_json();
    let first_input = first["input"].as_array().expect("request input");

    assert_eq!(first_input.len(), 2);
    assert!(
        first_input[0]["id"]
            .as_str()
            .is_some_and(|id| id.starts_with("msg_"))
    );
    assert_eq!(first_input[1]["id"], "msg_supplied");
    assert_eq!(second["previous_response_id"].as_str(), Some("resp-1"));
    assert_eq!(
        second["input"],
        serde_json::to_value(&prompt_two.input[2..]).expect("serialize incremental items")
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_forwards_turn_metadata_on_initial_and_incremental_create() {
    skip_if_no_network!();

    let mut first_output_item = ev_assistant_message("msg_1", "assistant output");
    first_output_item["item"]["internal_chat_message_metadata_passthrough"] =
        json!({"turn_id": "turn-123"});
    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-1"),
            first_output_item,
            ev_completed("resp-1"),
        ],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness(&server).await;
    let mut client_session = harness.client.new_session();
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let mut prior_assistant_output = assistant_message_item("1", "assistant output");
    prior_assistant_output.set_turn_id_if_missing("turn-123");
    let prompt_two = prompt_with_input(vec![
        message_item("hello"),
        prior_assistant_output,
        message_item("second"),
    ]);
    let first_responses_metadata =
        with_parent_turn(turn_metadata(&harness, Some("turn-123")), "parent-turn-123");
    let second_responses_metadata =
        with_parent_turn(turn_metadata(&harness, Some("turn-456")), "parent-turn-456");

    stream_until_complete_with_metadata(
        &mut client_session,
        &harness,
        &prompt_one,
        /*service_tier*/ None,
        &first_responses_metadata,
    )
    .await;
    stream_until_complete_with_metadata(
        &mut client_session,
        &harness,
        &prompt_two,
        /*service_tier*/ None,
        &second_responses_metadata,
    )
    .await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let first = connection.first().expect("missing request").body_json();
    let second = connection.get(1).expect("missing request").body_json();

    assert_eq!(first["type"].as_str(), Some("response.create"));
    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second["previous_response_id"].as_str(), Some("resp-1"));
    let first_metadata: serde_json::Value = serde_json::from_str(
        first["client_metadata"]["x-codex-turn-metadata"]
            .as_str()
            .expect("first turn metadata"),
    )
    .expect("first metadata should be valid json");
    let second_metadata: serde_json::Value = serde_json::from_str(
        second["client_metadata"]["x-codex-turn-metadata"]
            .as_str()
            .expect("second turn metadata"),
    )
    .expect("second metadata should be valid json");

    assert_eq!(first_metadata["turn_id"].as_str(), Some("turn-123"));
    assert_eq!(second_metadata["turn_id"].as_str(), Some("turn-456"));
    core_test_support::responses::assert_parent_turn(&first, Some("parent-turn-123"))
        .expect("initial WebSocket parent-turn metadata");
    core_test_support::responses::assert_parent_turn(&second, Some("parent-turn-456"))
        .expect("incremental WebSocket parent-turn metadata");
    assert_eq!(
        first["client_metadata"]["turn_id"].as_str(),
        first_metadata["turn_id"].as_str()
    );
    assert_eq!(
        second["client_metadata"]["turn_id"].as_str(),
        second_metadata["turn_id"].as_str()
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_sends_canonical_turn_metadata() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness = websocket_harness(&server).await;
    let mut client_session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);
    let responses_metadata = turn_metadata(&harness, Some("turn-123"));

    stream_until_complete_with_metadata(
        &mut client_session,
        &harness,
        &prompt,
        /*service_tier*/ None,
        &responses_metadata,
    )
    .await;

    let body = server
        .single_connection()
        .first()
        .expect("missing request")
        .body_json();

    assert_eq!(body["type"].as_str(), Some("response.create"));
    let turn_metadata: serde_json::Value = serde_json::from_str(
        body["client_metadata"]["x-codex-turn-metadata"]
            .as_str()
            .expect("turn metadata"),
    )
    .expect("valid turn metadata");
    assert_eq!(turn_metadata["turn_id"].as_str(), Some("turn-123"));
    assert_eq!(
        body["client_metadata"]["turn_id"].as_str(),
        turn_metadata["turn_id"].as_str()
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_uses_previous_response_id_when_prefix_after_completed() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-1"),
            ev_assistant_message("msg_1", "assistant output"),
            ev_completed("resp-1"),
        ],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness(&server).await;
    let mut client_session = harness.client.new_session();
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let prompt_two = prompt_with_input(vec![
        message_item("hello"),
        assistant_message_item("1", "assistant output"),
        message_item("second"),
    ]);

    stream_until_complete(&mut client_session, &harness, &prompt_one).await;
    stream_until_complete(&mut client_session, &harness, &prompt_two).await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let second = connection.get(1).expect("missing request").body_json();

    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second["previous_response_id"].as_str(), Some("resp-1"));
    assert_eq!(
        second["input"],
        serde_json::to_value(&prompt_two.input[2..]).expect("serialize incremental input")
    );

    server.shutdown().await;
}

#[test_case::test_case(vec![message_item("different")], "input_mismatch", "message"; "rewritten")]
#[test_case::test_case(
    vec![ResponseItem::Compaction {
        id: None,
        encrypted_content: "compacted".to_string(),
        internal_chat_message_metadata_passthrough: None,
    }],
    "input_mismatch",
    "compaction"; "replaced"
)]
#[test_case::test_case(
    vec![message_item("hello"), assistant_message_item("1", "changed output")],
    "input_mismatch",
    "message"; "rewritten server output"
)]
#[test_case::test_case(vec![message_item("hello")], "input_shortened", "none"; "missing server output")]
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_creates_on_non_prefix(
    input: Vec<ResponseItem>,
    reason: &str,
    current_item_type: &str,
) {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-1"),
            ev_assistant_message("msg_1", "assistant output"),
            ev_completed("resp-1"),
        ],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness(&server).await;
    let mut client_session = harness.client.new_session();
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let prompt_two = prompt_with_input(input);

    stream_until_complete(&mut client_session, &harness, &prompt_one).await;
    stream_until_complete(&mut client_session, &harness, &prompt_two).await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let second = connection.get(1).expect("missing request").body_json();

    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second["model"].as_str(), Some(MODEL));
    assert_eq!(second["stream"], serde_json::Value::Bool(true));
    assert_eq!(second.get("previous_response_id"), None);
    assert_eq!(
        second["input"],
        serde_json::to_value(&prompt_two.input).unwrap()
    );

    assert_continuation_metrics_with_item_types(
        &harness.session_telemetry,
        &[
            (
                ["full", reason, "generation", "message", current_item_type],
                1,
            ),
            (
                ["full", "no_previous_request", "generation", "none", "none"],
                1,
            ),
        ],
    );
    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_creates_when_instructions_change() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![ev_response_created("resp-1"), ev_completed("resp-1")],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness(&server).await;
    let mut client_session = harness.client.new_session();
    let prompt_one =
        prompt_with_input_and_instructions(vec![message_item("hello")], "base instructions one");
    let prompt_two = prompt_with_input_and_instructions(
        vec![message_item("hello"), message_item("second")],
        "base instructions two",
    );

    stream_until_complete(&mut client_session, &harness, &prompt_one).await;
    stream_until_complete(&mut client_session, &harness, &prompt_two).await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let second = connection.get(1).expect("missing request").body_json();

    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second.get("previous_response_id"), None);
    assert_eq!(second["input"][0]["role"], "developer");
    assert_eq!(
        second["input"][0]["content"][0]["text"],
        "base instructions two"
    );
    assert_eq!(
        &second["input"].as_array().unwrap()[1..],
        serde_json::to_value(&prompt_two.input)
            .unwrap()
            .as_array()
            .unwrap()
    );

    assert_continuation_metrics_with_item_types(
        &harness.session_telemetry,
        &[
            (
                ["full", "no_previous_request", "generation", "none", "none"],
                1,
            ),
            (
                ["full", "input_mismatch", "generation", "message", "message"],
                1,
            ),
        ],
    );
    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_v2_creates_with_previous_response_id_on_prefix() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![
            ev_response_created("resp-1"),
            ev_assistant_message("msg_1", "assistant output"),
            ev_completed("resp-1"),
        ],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness_with_v2(&server, /*runtime_metrics_enabled*/ true).await;
    let mut session = harness.client.new_session();
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let prompt_two = prompt_with_input(vec![
        message_item("hello"),
        assistant_message_item("1", "assistant output"),
        message_item("second"),
    ]);

    stream_until_complete(&mut session, &harness, &prompt_one).await;
    stream_until_complete(&mut session, &harness, &prompt_two).await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let first = connection.first().expect("missing request").body_json();
    let second = connection.get(1).expect("missing request").body_json();

    assert_eq!(first["type"].as_str(), Some("response.create"));
    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second["previous_response_id"].as_str(), Some("resp-1"));
    assert_eq!(
        second["input"],
        serde_json::to_value(&prompt_two.input[2..]).unwrap()
    );

    server.shutdown().await;
}

#[test_case::test_case(None, Some(json!({"type": "object", "properties": {}})), "text_changed"; "output schema")]
#[test_case::test_case(Some("changed-model"), None, "model_changed"; "model")]
#[test_case::test_case(Some("changed-model"), Some(json!({"type": "object"})), "model_changed"; "first changed property")]
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_v2_creates_without_previous_response_id_when_non_input_fields_change(
    model: Option<&str>,
    output_schema: Option<serde_json::Value>,
    reason: &str,
) {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![
        vec![ev_response_created("resp-1"), ev_completed("resp-1")],
        vec![ev_response_created("resp-2"), ev_completed("resp-2")],
    ]])
    .await;

    let harness = websocket_harness_with_v2(&server, /*runtime_metrics_enabled*/ true).await;
    let mut session = harness.client.new_session();
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let mut prompt_two = prompt_with_input(vec![message_item("hello"), message_item("second")]);
    prompt_two.output_schema = output_schema;
    let mut model_info = harness.model_info.clone();
    if let Some(model) = model {
        model_info.slug = model.to_string();
    }

    stream_until_complete(&mut session, &harness, &prompt_one).await;
    stream_until_complete_with_model_info(
        &mut session,
        &harness,
        &prompt_two,
        &model_info,
        "resp-2",
    )
    .await;

    let connection = server.single_connection();
    assert_eq!(connection.len(), 2);
    let second = connection.get(1).expect("missing request").body_json();

    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second.get("previous_response_id"), None);
    assert_eq!(
        second["input"],
        serde_json::to_value(&prompt_two.input).expect("serialize full input")
    );

    assert_continuation_metrics(
        &harness.session_telemetry,
        &[
            (["full", "no_previous_request", "generation"], 1),
            (["full", reason, "generation"], 1),
        ],
    );
    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_v2_after_error_uses_full_create_without_previous_response_id() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![
        vec![
            vec![ev_response_created("resp-1"), ev_completed("resp-1")],
            vec![json!({
                "type": "response.failed",
                "response": {
                    "error": {
                        "code": "invalid_prompt",
                        "message": "synthetic websocket failure"
                    }
                }
            })],
        ],
        vec![vec![ev_response_created("resp-3"), ev_completed("resp-3")]],
    ])
    .await;

    let harness = websocket_harness_with_v2(&server, /*runtime_metrics_enabled*/ true).await;
    let mut session = harness.client.new_session();
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let prompt_two = prompt_with_input(vec![message_item("hello"), message_item("second")]);
    let prompt_three = prompt_with_input(vec![
        message_item("hello"),
        message_item("second"),
        message_item("third"),
    ]);

    stream_until_complete(&mut session, &harness, &prompt_one).await;

    let responses_metadata = turn_metadata(&harness, /*turn_id*/ None);
    let mut second_stream = session
        .stream(
            &prompt_two,
            &harness.model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            /*service_tier*/ None,
            &responses_metadata,
            &codex_rollout_trace::InferenceTraceContext::disabled(),
        )
        .await
        .expect("websocket stream failed");
    let mut saw_error = false;
    while let Some(event) = second_stream.next().await {
        if event.is_err() {
            saw_error = true;
            break;
        }
    }
    assert!(saw_error, "expected second websocket stream to error");

    stream_until_complete(&mut session, &harness, &prompt_three).await;

    assert_eq!(server.handshakes().len(), 2);

    let connections = server.connections();
    assert_eq!(connections.len(), 2);
    let first_connection = connections.first().expect("missing first connection");
    assert_eq!(first_connection.len(), 2);

    let first = first_connection
        .first()
        .expect("missing first request")
        .body_json();
    let second = first_connection
        .get(1)
        .expect("missing second request")
        .body_json();
    let third = connections
        .get(1)
        .and_then(|connection| connection.first())
        .expect("missing third request")
        .body_json();

    assert_eq!(first["type"].as_str(), Some("response.create"));
    assert_eq!(second["type"].as_str(), Some("response.create"));
    assert_eq!(second["previous_response_id"].as_str(), Some("resp-1"));
    assert_eq!(third["type"].as_str(), Some("response.create"));
    assert_eq!(third.get("previous_response_id"), None);
    assert_eq!(
        third["input"],
        serde_json::to_value(&prompt_three.input).unwrap()
    );

    assert_continuation_metrics(
        &harness.session_telemetry,
        &[
            (["full", "connection_closed", "generation"], 1),
            (["full", "no_previous_request", "generation"], 1),
            (["incremental", "incremental", "generation"], 1),
        ],
    );

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_v2_surfaces_terminal_error_without_close_handshake() {
    skip_if_no_network!();

    let server = start_websocket_server_with_headers(vec![WebSocketConnectionConfig {
        requests: vec![
            vec![ev_response_created("resp-1"), ev_completed("resp-1")],
            vec![json!({
                "type": "response.failed",
                "response": {
                    "error": {
                        "code": "invalid_prompt",
                        "message": "synthetic websocket failure"
                    }
                }
            })],
        ],
        response_headers: Vec::new(),
        accept_delay: None,
        close_after_requests: false,
    }])
    .await;

    let harness = websocket_harness_with_v2(&server, /*runtime_metrics_enabled*/ true).await;
    let mut session = harness.client.new_session();
    let prompt_one = prompt_with_input(vec![message_item("hello")]);
    let prompt_two = prompt_with_input(vec![message_item("hello"), message_item("second")]);

    stream_until_complete(&mut session, &harness, &prompt_one).await;

    let responses_metadata = turn_metadata(&harness, /*turn_id*/ None);
    let mut second_stream = session
        .stream(
            &prompt_two,
            &harness.model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            /*service_tier*/ None,
            &responses_metadata,
            &codex_rollout_trace::InferenceTraceContext::disabled(),
        )
        .await
        .expect("websocket stream failed");

    let saw_error = tokio::time::timeout(Duration::from_secs(2), async {
        while let Some(event) = second_stream.next().await {
            if event.is_err() {
                return true;
            }
        }
        false
    })
    .await
    .expect("timed out waiting for terminal websocket error");

    assert!(saw_error, "expected second websocket stream to error");

    server.shutdown().await;
}

#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_v2_sets_openai_beta_header() {
    skip_if_no_network!();

    let server = start_websocket_server(vec![vec![vec![
        ev_response_created("resp-1"),
        ev_completed("resp-1"),
    ]]])
    .await;

    let harness = websocket_harness_with_v2(&server, /*runtime_metrics_enabled*/ true).await;
    let mut session = harness.client.new_session();
    let prompt = prompt_with_input(vec![message_item("hello")]);

    stream_until_complete(&mut session, &harness, &prompt).await;

    let handshake = server.single_handshake();
    let openai_beta_header = handshake
        .header(OPENAI_BETA_HEADER)
        .expect("missing OpenAI-Beta header");
    assert!(
        openai_beta_header
            .split(',')
            .map(str::trim)
            .any(|value| value == WS_V2_BETA_HEADER_VALUE)
    );
    server.shutdown().await;
}

fn message_item(text: &str) -> ResponseItem {
    ResponseItem::Message {
        status: None,
        encrypted_content: None,
        id: None,
        role: "user".into(),
        content: vec![ContentItem::InputText { text: text.into() }],
        phase: None,
        internal_chat_message_metadata_passthrough: None,
    }
}

fn assistant_message_item(id: &str, text: &str) -> ResponseItem {
    ResponseItem::Message {
        status: None,
        encrypted_content: None,
        id: Some(ResponseItemId::with_suffix("msg", id)),
        role: "assistant".into(),
        content: vec![ContentItem::OutputText {
            annotations: None,
            logprobs: None,
            text: text.into(),
        }],
        phase: None,
        internal_chat_message_metadata_passthrough: None,
    }
}

fn prompt_with_input(input: Vec<ResponseItem>) -> Prompt {
    let mut prompt = Prompt::default();
    prompt.input = input;
    prompt.base_instructions.text.clear();
    prompt
}

fn prompt_with_input_and_instructions(input: Vec<ResponseItem>, instructions: &str) -> Prompt {
    let mut prompt = prompt_with_input(input);
    prompt.base_instructions = BaseInstructions {
        text: instructions.to_string(),
        provenance: None,
    };
    prompt
}

fn websocket_provider(server: &WebSocketTestServer) -> ModelProviderInfo {
    websocket_provider_with_connect_timeout(server, /*websocket_connect_timeout_ms*/ None)
}

fn websocket_provider_with_connect_timeout(
    server: &WebSocketTestServer,
    websocket_connect_timeout_ms: Option<u64>,
) -> ModelProviderInfo {
    ModelProviderInfo {
        name: "mock-ws".into(),
        base_url: Some(format!("{}/v1", server.uri())),
        model_catalog_url: None,
        env_key: None,
        env_key_instructions: None,
        experimental_bearer_token: None,
        auth: None,
        gateway_oauth: None,
        aws: None,
        wire_api: WireApi::Responses,
        query_params: None,
        http_headers: None,
        env_http_headers: None,
        request_max_retries: Some(0),
        stream_max_retries: Some(0),
        stream_idle_timeout_ms: Some(5_000),
        websocket_connect_timeout_ms,
        requires_openai_auth: false,
        supports_websockets: true,
        supports_standalone_web_search: false,
        capabilities: None,
        include_internal_metadata: false,
    }
}

async fn websocket_harness(server: &WebSocketTestServer) -> WebsocketTestHarness {
    websocket_harness_with_runtime_metrics(server, /*runtime_metrics_enabled*/ false).await
}

async fn websocket_harness_for_codex_backend(server: &WebSocketTestServer) -> WebsocketTestHarness {
    let provider = ModelProviderInfo::create_openai_provider(Some(format!("{}/v1", server.uri())));
    websocket_harness_with_provider_options_and_auth(
        provider,
        /*runtime_metrics_enabled*/ false,
        /*concurrent_reasoning_summaries_enabled*/ false,
        /*enabled_features*/ &[],
        Some(CodexAuth::create_dummy_chatgpt_auth_for_testing()),
    )
    .await
}

async fn websocket_harness_with_runtime_metrics(
    server: &WebSocketTestServer,
    runtime_metrics_enabled: bool,
) -> WebsocketTestHarness {
    websocket_harness_with_options(server, runtime_metrics_enabled).await
}

async fn websocket_harness_with_v2(
    server: &WebSocketTestServer,
    runtime_metrics_enabled: bool,
) -> WebsocketTestHarness {
    websocket_harness_with_options(server, runtime_metrics_enabled).await
}

async fn websocket_harness_with_options(
    server: &WebSocketTestServer,
    runtime_metrics_enabled: bool,
) -> WebsocketTestHarness {
    websocket_harness_with_provider_options(
        websocket_provider(server),
        runtime_metrics_enabled,
        /*concurrent_reasoning_summaries_enabled*/ false,
        /*enabled_features*/ &[],
    )
    .await
}

async fn websocket_harness_with_provider_options(
    provider: ModelProviderInfo,
    runtime_metrics_enabled: bool,
    concurrent_reasoning_summaries_enabled: bool,
    enabled_features: &[Feature],
) -> WebsocketTestHarness {
    websocket_harness_with_provider_options_and_auth(
        provider,
        runtime_metrics_enabled,
        concurrent_reasoning_summaries_enabled,
        enabled_features,
        /*auth*/ None,
    )
    .await
}

async fn websocket_harness_with_provider_options_and_auth(
    provider: ModelProviderInfo,
    runtime_metrics_enabled: bool,
    concurrent_reasoning_summaries_enabled: bool,
    enabled_features: &[Feature],
    auth: Option<CodexAuth>,
) -> WebsocketTestHarness {
    let codex_home = TempDir::new().unwrap();
    let mut config = load_default_config_for_test(&codex_home).await;
    config.model = Some(MODEL.to_string());
    if runtime_metrics_enabled {
        config
            .features
            .enable(Feature::RuntimeMetrics)
            .expect("test config should allow feature update");
    }
    if concurrent_reasoning_summaries_enabled {
        config
            .features
            .enable(Feature::ConcurrentReasoningSummaries)
            .expect("test config should allow feature update");
    }
    for feature in enabled_features {
        config
            .features
            .enable(*feature)
            .expect("test config should allow feature update");
    }
    config.respect_system_proxy = config.features.enabled(Feature::RespectSystemProxy);
    let http_client_factory = config.http_client_factory();
    let outbound_proxy_policy = http_client_factory.outbound_proxy_policy();
    let config = Arc::new(config);
    let model_info = codex_core::test_support::construct_model_info_offline(MODEL, &config);
    let thread_id = ThreadId::new();
    let session_id = SessionId::new();
    let client_auth_manager = auth.map(|auth| {
        codex_core::test_support::auth_manager_from_auth_with_home(
            auth,
            codex_home.path().to_path_buf(),
        )
    });
    let auth_manager = client_auth_manager.clone().unwrap_or_else(|| {
        codex_core::test_support::auth_manager_from_auth(CodexAuth::from_api_key("Test API Key"))
    });
    let exporter = InMemoryMetricExporter::default();
    let metrics = MetricsClient::new(
        MetricsConfig::in_memory("test", "codex-core", env!("CARGO_PKG_VERSION"), exporter)
            .with_runtime_reader(),
    )
    .expect("in-memory metrics client");
    let session_telemetry = SessionTelemetry::new(
        thread_id,
        MODEL,
        model_info.slug.as_str(),
        /*account_id*/ None,
        Some("[email protected]".to_string()),
        auth_manager.auth_mode().map(TelemetryAuthMode::from),
        "test_originator".to_string(),
        /*log_user_prompts*/ false,
        "test".to_string(),
        SessionSource::Exec,
    )
    .with_metrics(metrics);
    let effort = None;
    let summary = ReasoningSummary::Auto;
    let client = ModelClient::new(
        client_auth_manager,
        AgentIdentityAuthPolicy::JwtOnly,
        thread_id,
        provider.clone(),
        SessionSource::Exec,
        "test_originator".to_string(),
        config.model_verbosity,
        config.features.enabled(Feature::ContentItemKinds),
        config.features.enabled(Feature::ReasoningEffortOverride),
        /*enable_request_compression*/ false,
        runtime_metrics_enabled,
        /*beta_features_header*/ None,
        /*concurrent_reasoning_summaries_enabled*/
        config
            .features
            .enabled(Feature::ConcurrentReasoningSummaries),
        config.features.enabled(Feature::OutputTokenReplay),
        /*attestation_provider*/ None,
        http_client_factory,
        config.workspace_routing_context(),
        Vec::new(),
    );

    WebsocketTestHarness {
        codex_home,
        auth_manager,
        client,
        outbound_proxy_policy,
        session_id,
        thread_id,
        model_info,
        effort,
        summary,
        session_telemetry,
    }
}

#[test_case::test_case(false; "resume")]
#[test_case::test_case(true; "fork")]
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn responses_websocket_restored_history_metric(fork: bool) -> anyhow::Result<()> {
    skip_if_no_network!(Ok(()));

    let exporter = InMemoryMetricExporter::default();
    let config =
        MetricsConfig::in_memory("test", "codex-core", "test", exporter).with_runtime_reader();
    let metrics = MetricsClient::new(config)?;
    let warmup = vec![ev_response_created("warmup"), ev_completed("warmup")];
    let turn = vec![ev_response_created("turn"), ev_completed("turn")];
    let server = start_websocket_server(vec![vec![warmup, turn]; 2]).await;
    let mut initial = test_codex().build_with_websocket_server(&server).await?;
    initial.submit_text_turn("original history marker").await?;
    let rollout_path = initial.codex.rollout_path().expect("persisted history");
    initial.codex.shutdown_and_wait().await?;

    let manager = &initial.thread_manager;
    let mut options = codex_core::StartThreadOptions::new(initial.config.clone());
    options.thread_extension_init.insert(metrics);
    let restored = if fork {
        manager
            .fork_legacy_thread(codex_core::ForkSnapshot::Interrupted, options, rollout_path)
            .await?
    } else {
        options.initial_history =
            codex_rollout::RolloutRecorder::get_rollout_history(&rollout_path).await?;
        manager.start_thread(options).await?
    };
    initial.codex = restored.thread;
    initial.submit_text_turn("continue").await?;
    assert_continuation_metrics(
        &initial.codex.session_telemetry(),
        &[
            (["full", "restored_history", "warmup"], 1),
            (["incremental", "incremental", "generation"], 1),
        ],
    );
    assert!(server.connections()[1].iter().any(|request| {
        request.body_json()["input"]
            .to_string()
            .contains("original history marker")
    }));
    initial.codex.shutdown_and_wait().await?;
    server.shutdown().await;
    Ok(())
}

fn assert_continuation_metrics(telemetry: &SessionTelemetry, expected: &[([&str; 3], u64)]) {
    let expected = expected
        .iter()
        .map(|([mode, reason, phase], count)| ([*mode, *reason, *phase, "none", "none"], *count))
        .collect::<Vec<_>>();
    assert_continuation_metrics_with_item_types(telemetry, &expected);
}

fn assert_continuation_metrics_with_item_types(
    telemetry: &SessionTelemetry,
    expected: &[([&str; 5], u64)],
) {
    use opentelemetry_sdk::metrics::data::AggregatedMetrics;
    use opentelemetry_sdk::metrics::data::MetricData;

    let snapshot = telemetry.snapshot_metrics().expect("metrics snapshot");
    let metric = snapshot
        .scope_metrics()
        .flat_map(opentelemetry_sdk::metrics::data::ScopeMetrics::metrics)
        .find(|metric| metric.name() == codex_otel::WEBSOCKET_CONTINUATION_COUNT_METRIC)
        .expect("continuation counter");
    let AggregatedMetrics::U64(MetricData::Sum(sum)) = metric.data() else {
        panic!("expected counter");
    };
    let mut expected: Vec<_> = expected
        .iter()
        .map(|(tags, count)| (tags.map(str::to_owned), *count))
        .collect();
    let mut actual: Vec<_> = sum
        .data_points()
        .map(|point| {
            let tags = [
                "mode",
                "reason",
                "phase",
                "previous_item_type",
                "current_item_type",
            ]
            .map(|key| {
                point
                    .attributes()
                    .find(|attr| attr.key.as_str() == key)
                    .unwrap()
                    .value
                    .to_string()
            });
            (tags, point.value())
        })
        .collect();
    actual.sort();
    expected.sort();
    assert_eq!(actual, expected);
}

async fn stream_until_complete(
    client_session: &mut ModelClientSession,
    harness: &WebsocketTestHarness,
    prompt: &Prompt,
) {
    stream_until_complete_with_service_tier(
        client_session,
        harness,
        prompt,
        /*service_tier*/ None,
    )
    .await;
}

async fn stream_until_complete_with_model_info(
    client_session: &mut ModelClientSession,
    harness: &WebsocketTestHarness,
    prompt: &Prompt,
    model_info: &ModelInfo,
    expected_response_id: &str,
) {
    let responses_metadata = turn_metadata(harness, /*turn_id*/ None);
    let mut stream = client_session
        .stream(
            prompt,
            model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            /*service_tier*/ None,
            &responses_metadata,
            &codex_rollout_trace::InferenceTraceContext::disabled(),
        )
        .await
        .expect("websocket stream failed");

    while let Some(event) = stream.next().await {
        if let ResponseEvent::Completed { response_id, .. } =
            event.expect("websocket stream failed")
        {
            assert_eq!(response_id, expected_response_id);
            return;
        }
    }
    panic!("websocket stream ended before completion");
}

async fn stream_until_complete_with_service_tier(
    client_session: &mut ModelClientSession,
    harness: &WebsocketTestHarness,
    prompt: &Prompt,
    service_tier: Option<ServiceTier>,
) {
    let responses_metadata = turn_metadata(harness, /*turn_id*/ None);
    stream_until_complete_with_metadata(
        client_session,
        harness,
        prompt,
        service_tier,
        &responses_metadata,
    )
    .await;
}

async fn stream_until_complete_with_metadata(
    client_session: &mut ModelClientSession,
    harness: &WebsocketTestHarness,
    prompt: &Prompt,
    service_tier: Option<ServiceTier>,
    responses_metadata: &CodexResponsesMetadata,
) {
    let mut stream = client_session
        .stream(
            prompt,
            &harness.model_info,
            &harness.session_telemetry,
            harness.effort.clone(),
            harness.summary,
            service_tier.map(|service_tier| service_tier.request_value().to_string()),
            responses_metadata,
            &codex_rollout_trace::InferenceTraceContext::disabled(),
        )
        .await
        .expect("websocket stream failed");

    while let Some(event) = stream.next().await {
        if matches!(event, Ok(ResponseEvent::Completed { .. })) {
            break;
        }
    }
}