codex/codex-rs/app-server/tests/suite/v2/turn_steer.rs

#![cfg(unix)]

use anyhow::Context;
use anyhow::Result;
use app_test_support::TestAppServer;
use app_test_support::create_command_execution_sse_response;
use app_test_support::create_mock_responses_server_sequence;
use app_test_support::create_mock_responses_server_sequence_unchecked;
use app_test_support::write_mock_responses_config_toml_with_chatgpt_base_url;
use codex_app_server::INPUT_TOO_LARGE_ERROR_CODE;
use codex_app_server::INVALID_PARAMS_ERROR_CODE;
use codex_app_server_protocol::AdditionalContextEntry;
use codex_app_server_protocol::AdditionalContextKind;
use codex_app_server_protocol::ClientRequest;
use codex_app_server_protocol::ItemStartedNotification;
use codex_app_server_protocol::JSONRPCError;
use codex_app_server_protocol::JSONRPCNotification;
use codex_app_server_protocol::RequestId;
use codex_app_server_protocol::ThreadItemPayload;
use codex_app_server_protocol::ThreadStartParams;
use codex_app_server_protocol::ThreadStartResponse;
use codex_app_server_protocol::TurnStartParams;
use codex_app_server_protocol::TurnStartResponse;
use codex_app_server_protocol::TurnSteerParams;
use codex_app_server_protocol::TurnSteerResponse;
use codex_app_server_protocol::UserInput as V2UserInput;
use codex_protocol::user_input::MAX_USER_INPUT_TEXT_CHARS;
use core_test_support::skip_if_remote;
use pretty_assertions::assert_eq;
use serde_json::Value;
use std::collections::HashMap;
use tempfile::TempDir;
use tokio::time::timeout;

use super::analytics::mount_analytics_capture;
use super::analytics::wait_for_analytics_event;

const DEFAULT_READ_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(10);

#[tokio::test]
async fn turn_steer_requires_active_turn() -> Result<()> {
    let tmp = TempDir::new()?;
    let codex_home = tmp.path().join("codex_home");
    std::fs::create_dir(&codex_home)?;

    let server = create_mock_responses_server_sequence(vec![]).await;
    write_mock_responses_config_toml_with_chatgpt_base_url(
        &codex_home,
        &server.uri(),
        &server.uri(),
    )?;
    mount_analytics_capture(&server, &codex_home).await?;

    let mut mcp = TestAppServer::builder()
        .with_codex_home(&codex_home)
        .without_managed_config()
        .build_initialized()
        .await?;

    let ThreadStartResponse { thread, .. } = mcp
        .start_thread(ThreadStartParams {
            model: Some("mock-model".to_string()),
            ..Default::default()
        })
        .await?;

    let steer_req = mcp
        .send_turn_steer_request(TurnSteerParams {
            thread_id: thread.id.clone(),
            client_user_message_id: Some("client-steer-message-1".to_string()),
            input: vec![V2UserInput::Text {
                text: "steer".to_string(),
                text_elements: Vec::new(),
            }],
            responsesapi_client_metadata: None,
            responsesapi_client_metadata_patch: None,
            additional_context: None,
            expected_turn_id: "turn-does-not-exist".to_string(),
        })
        .await?;
    let steer_err: JSONRPCError = timeout(
        DEFAULT_READ_TIMEOUT,
        mcp.read_stream_until_error_message(RequestId::Integer(steer_req)),
    )
    .await??;
    assert_eq!(steer_err.error.code, -32600);

    let event =
        wait_for_analytics_event(&server, DEFAULT_READ_TIMEOUT, "codex_turn_steer_event").await?;
    assert_eq!(event["event_params"]["thread_id"], thread.id);
    assert_eq!(event["event_params"]["result"], "rejected");
    assert_eq!(event["event_params"]["num_input_images"], 0);
    assert_eq!(
        event["event_params"]["expected_turn_id"],
        "turn-does-not-exist"
    );
    assert_eq!(
        event["event_params"]["accepted_turn_id"],
        serde_json::Value::Null
    );
    assert_eq!(event["event_params"]["rejection_reason"], "no_active_turn");

    let malformed_request = mcp
        .send_turn_steer_request(TurnSteerParams {
            thread_id: thread.id,
            client_user_message_id: None,
            input: vec![V2UserInput::Text {
                text: "rejected input".to_owned(),
                text_elements: Vec::new(),
            }],
            responsesapi_client_metadata: None,
            responsesapi_client_metadata_patch: None,
            additional_context: Some(HashMap::from([(
                "browser_info".to_owned(),
                AdditionalContextEntry {
                    value: "rejected context".to_owned(),
                    kind: AdditionalContextKind::Untrusted,
                    name: Some("agent.browser".to_owned()),
                },
            )])),
            expected_turn_id: "turn-does-not-exist".to_owned(),
        })
        .await?;
    let error = timeout(
        DEFAULT_READ_TIMEOUT,
        mcp.read_stream_until_error_message(RequestId::Integer(malformed_request)),
    )
    .await??;
    assert_eq!(
        error.error,
        codex_app_server_protocol::JSONRPCErrorError {
            code: -32600,
            message: "additional context name must match [a-z][a-z0-9_]{0,63}".to_owned(),
            data: None,
        }
    );

    Ok(())
}

#[tokio::test]
async fn turn_steer_rejects_oversized_text_input() -> Result<()> {
    // TODO(anp): Remove after the active-turn fixture can run in the selected remote environment.
    skip_if_remote!(
        Ok(()),
        "uses a host-local command and cwd fixture unavailable to remote executors"
    );

    #[cfg(target_os = "windows")]
    let shell_command = vec![
        "powershell".to_string(),
        "-Command".to_string(),
        "Start-Sleep -Seconds 10".to_string(),
    ];
    #[cfg(not(target_os = "windows"))]
    let shell_command = vec!["sleep".to_string(), "10".to_string()];

    let tmp = TempDir::new()?;
    let codex_home = tmp.path().join("codex_home");
    std::fs::create_dir(&codex_home)?;
    let working_directory = tmp.path().join("workdir");
    std::fs::create_dir(&working_directory)?;

    let server = create_mock_responses_server_sequence_unchecked(vec![
        create_command_execution_sse_response(
            shell_command.clone(),
            Some(&working_directory),
            Some(10_000),
            "call_sleep",
        )?,
    ])
    .await;
    write_mock_responses_config_toml_with_chatgpt_base_url(
        &codex_home,
        &server.uri(),
        &server.uri(),
    )?;
    mount_analytics_capture(&server, &codex_home).await?;

    let mut mcp = TestAppServer::builder()
        .with_codex_home(&codex_home)
        .without_managed_config()
        .build_initialized()
        .await?;

    let ThreadStartResponse { thread, .. } = mcp
        .start_thread(ThreadStartParams {
            model: Some("mock-model".to_string()),
            ..Default::default()
        })
        .await?;

    let TurnStartResponse { turn } = mcp
        .request(|request_id| ClientRequest::TurnStart {
            request_id,
            params: TurnStartParams {
                thread_id: thread.id.clone(),
                client_user_message_id: None,
                input: vec![V2UserInput::Text {
                    text: "run sleep".to_string(),
                    text_elements: Vec::new(),
                }],
                cwd: Some(working_directory.clone()),
                ..Default::default()
            },
        })
        .await?;

    let _task_started: JSONRPCNotification = timeout(
        DEFAULT_READ_TIMEOUT,
        mcp.read_stream_until_notification_message("turn/started"),
    )
    .await??;

    let oversized_input = "x".repeat(MAX_USER_INPUT_TEXT_CHARS + 1);
    let steer_req = mcp
        .send_turn_steer_request(TurnSteerParams {
            thread_id: thread.id.clone(),
            client_user_message_id: None,
            input: vec![V2UserInput::Text {
                text: oversized_input.clone(),
                text_elements: Vec::new(),
            }],
            responsesapi_client_metadata: None,
            responsesapi_client_metadata_patch: None,
            additional_context: None,
            expected_turn_id: turn.id.clone(),
        })
        .await?;
    let steer_err: JSONRPCError = timeout(
        DEFAULT_READ_TIMEOUT,
        mcp.read_stream_until_error_message(RequestId::Integer(steer_req)),
    )
    .await??;

    assert_eq!(steer_err.error.code, INVALID_PARAMS_ERROR_CODE);
    assert_eq!(
        steer_err.error.message,
        format!("Input exceeds the maximum length of {MAX_USER_INPUT_TEXT_CHARS} characters.")
    );
    let data = steer_err
        .error
        .data
        .expect("expected structured error data");
    assert_eq!(data["input_error_code"], INPUT_TOO_LARGE_ERROR_CODE);
    assert_eq!(data["max_chars"], MAX_USER_INPUT_TEXT_CHARS);
    assert_eq!(data["actual_chars"], oversized_input.chars().count());

    mcp.interrupt_turn_and_wait_for_aborted(thread.id, turn.id, DEFAULT_READ_TIMEOUT)
        .await?;

    Ok(())
}

#[test_case::test_case(false; "legacy_context")]
#[test_case::test_case(true; "named_context")]
#[tokio::test]
async fn turn_steer_returns_active_turn_id(tool_outputs: bool) -> Result<()> {
    // TODO(anp): Remove after the active-turn fixture can run in the selected remote environment.
    skip_if_remote!(
        Ok(()),
        "uses a host-local command and cwd fixture unavailable to remote executors"
    );

    #[cfg(target_os = "windows")]
    let shell_command = vec![
        "powershell".to_string(),
        "-Command".to_string(),
        "Start-Sleep -Seconds 2".to_string(),
    ];
    #[cfg(not(target_os = "windows"))]
    let shell_command = vec!["sleep".to_string(), "2".to_string()];

    let tmp = TempDir::new()?;
    let codex_home = tmp.path().join("codex_home");
    std::fs::create_dir(&codex_home)?;
    let working_directory = tmp.path().join("workdir");
    std::fs::create_dir(&working_directory)?;

    let server = create_mock_responses_server_sequence_unchecked(vec![
        create_command_execution_sse_response(
            shell_command.clone(),
            Some(&working_directory),
            Some(10_000),
            "call_sleep",
        )?,
        app_test_support::create_final_assistant_message_sse_response("Done")?,
    ])
    .await;
    write_mock_responses_config_toml_with_chatgpt_base_url(
        &codex_home,
        &server.uri(),
        &server.uri(),
    )?;
    mount_analytics_capture(&server, &codex_home).await?;

    let mut mcp = TestAppServer::builder()
        .with_codex_home(&codex_home)
        .without_managed_config()
        .build_initialized()
        .await?;

    let ThreadStartResponse { thread, .. } = mcp
        .start_thread(ThreadStartParams {
            model: Some("mock-model".to_string()),
            config: Some(HashMap::from([(
                "features.untrusted_as_tool_output".to_owned(),
                serde_json::json!(tool_outputs),
            )])),
            ..Default::default()
        })
        .await?;

    let TurnStartResponse { turn } = mcp
        .request(|request_id| ClientRequest::TurnStart {
            request_id,
            params: TurnStartParams {
                thread_id: thread.id.clone(),
                client_user_message_id: None,
                input: vec![V2UserInput::Text {
                    text: "run sleep".to_string(),
                    text_elements: Vec::new(),
                }],
                cwd: Some(working_directory.clone()),
                ..Default::default()
            },
        })
        .await?;

    let _task_started: JSONRPCNotification = timeout(
        DEFAULT_READ_TIMEOUT,
        mcp.read_stream_until_notification_message("turn/started"),
    )
    .await??;

    let steer: TurnSteerResponse = mcp
        .request(|request_id| ClientRequest::TurnSteer {
            request_id,
            params: TurnSteerParams {
                thread_id: thread.id.clone(),
                client_user_message_id: Some("client-steer-message-1".to_string()),
                input: vec![V2UserInput::Text {
                    text: "steer".to_string(),
                    text_elements: Vec::new(),
                }],
                responsesapi_client_metadata: None,
                responsesapi_client_metadata_patch: None,
                additional_context: Some(HashMap::from([(
                    "browser_info".to_owned(),
                    AdditionalContextEntry {
                        value: "tab one".to_owned(),
                        kind: AdditionalContextKind::Untrusted,
                        name: Some("browser".to_owned()),
                    },
                )])),
                expected_turn_id: turn.id.clone(),
            },
        })
        .await?;
    assert_eq!(steer.turn_id, turn.id);

    timeout(DEFAULT_READ_TIMEOUT, async {
        loop {
            let notification = mcp
                .read_stream_until_notification_message("item/started")
                .await?;
            let params = notification.params.expect("item/started params");
            let item_started: ItemStartedNotification =
                serde_json::from_value(params).expect("deserialize item/started notification");
            let ThreadItemPayload::UserMessage {
                client_id, content, ..
            } = item_started.item.payload
            else {
                continue;
            };
            if client_id == Some("client-steer-message-1".to_string()) {
                assert_eq!(
                    content,
                    vec![V2UserInput::Text {
                        text: "steer".to_string(),
                        text_elements: Vec::new(),
                    }]
                );
                return Ok::<(), anyhow::Error>(());
            }
        }
    })
    .await??;

    let event =
        wait_for_analytics_event(&server, DEFAULT_READ_TIMEOUT, "codex_turn_steer_event").await?;
    assert_eq!(event["event_params"]["thread_id"], thread.id);
    assert_eq!(event["event_params"]["session_id"], thread.session_id);
    assert_eq!(event["event_params"]["result"], "accepted");
    assert_eq!(event["event_params"]["num_input_images"], 0);
    assert_eq!(event["event_params"]["expected_turn_id"], turn.id);
    assert_eq!(event["event_params"]["accepted_turn_id"], turn.id);
    assert_eq!(
        event["event_params"]["rejection_reason"],
        serde_json::Value::Null
    );

    timeout(
        DEFAULT_READ_TIMEOUT,
        mcp.read_stream_until_notification_message("turn/completed"),
    )
    .await??;

    let requests = server
        .received_requests()
        .await
        .context("recorded requests")?;
    let body = requests
        .iter()
        .rfind(|request| request.url.path().ends_with("/responses"))
        .context("model request after steering")?
        .body_json::<Value>()?;
    let input = body["input"].as_array().context("input array")?;
    let context_index = if tool_outputs {
        input.iter().position(|item| {
            item["type"] == "function_call_output"
                && item["namespace"] == "agent"
                && item["name"] == "browser"
                && item["output"] == "tab one"
        })
    } else {
        input.iter().position(|item| {
            item["content"].as_array().is_some_and(|content| {
                content.iter().any(|part| {
                    part["text"] == "<external_browser_info>tab one</external_browser_info>"
                })
            })
        })
    }
    .context("steered context")?;
    let user_index = input
        .iter()
        .position(|item| {
            item["role"] == "user"
                && item["content"]
                    .as_array()
                    .is_some_and(|content| content.iter().any(|part| part["text"] == "steer"))
        })
        .context("steered user input")?;
    assert!(context_index < user_index);

    Ok(())
}

#[tokio::test]
async fn turn_steer_rejects_context_only_input_without_merging_context() -> Result<()> {
    // TODO(anp): Remove after the active-turn fixture can run in the selected remote environment.
    skip_if_remote!(
        Ok(()),
        "uses a host-local command and cwd fixture unavailable to remote executors"
    );

    let tmp = TempDir::new()?;
    let codex_home = tmp.path().join("codex_home");
    std::fs::create_dir(&codex_home)?;
    let working_directory = tmp.path().join("workdir");
    std::fs::create_dir(&working_directory)?;

    let server = create_mock_responses_server_sequence_unchecked(vec![
        create_command_execution_sse_response(
            vec!["sleep".to_string(), "1".to_string()],
            Some(&working_directory),
            Some(10_000),
            "call_sleep",
        )?,
        app_test_support::create_final_assistant_message_sse_response("Done")?,
    ])
    .await;
    write_mock_responses_config_toml_with_chatgpt_base_url(
        &codex_home,
        &server.uri(),
        &server.uri(),
    )?;
    mount_analytics_capture(&server, &codex_home).await?;

    let mut mcp = TestAppServer::builder()
        .with_codex_home(&codex_home)
        .without_managed_config()
        .build_initialized()
        .await?;

    let ThreadStartResponse { thread, .. } = mcp
        .start_thread(ThreadStartParams {
            model: Some("mock-model".to_string()),
            ..Default::default()
        })
        .await?;

    let TurnStartResponse { turn } = mcp
        .request(|request_id| ClientRequest::TurnStart {
            request_id,
            params: TurnStartParams {
                thread_id: thread.id.clone(),
                client_user_message_id: None,
                input: vec![V2UserInput::Text {
                    text: "run sleep".to_string(),
                    text_elements: Vec::new(),
                }],
                cwd: Some(working_directory),
                ..Default::default()
            },
        })
        .await?;
    timeout(
        DEFAULT_READ_TIMEOUT,
        mcp.read_stream_until_notification_message("turn/started"),
    )
    .await??;

    let additional_context = Some(HashMap::from([(
        "browser_info".to_string(),
        AdditionalContextEntry {
            value: "tab one".to_string(),
            kind: AdditionalContextKind::Untrusted,
            name: None,
        },
    )]));
    let steer_req = mcp
        .send_turn_steer_request(TurnSteerParams {
            thread_id: thread.id.clone(),
            client_user_message_id: None,
            input: Vec::new(),
            responsesapi_client_metadata: None,
            responsesapi_client_metadata_patch: None,
            additional_context,
            expected_turn_id: turn.id,
        })
        .await?;
    let steer_error: JSONRPCError = timeout(
        DEFAULT_READ_TIMEOUT,
        mcp.read_stream_until_error_message(RequestId::Integer(steer_req)),
    )
    .await??;
    assert_eq!(steer_error.error.code, -32600);
    assert_eq!(steer_error.error.message, "input must not be empty");

    timeout(
        DEFAULT_READ_TIMEOUT,
        mcp.read_stream_until_notification_message("turn/completed"),
    )
    .await??;

    let requests = server
        .received_requests()
        .await
        .context("failed to fetch received requests")?;
    let response_requests = requests
        .iter()
        .filter(|request| request.url.path().ends_with("/responses"))
        .collect::<Vec<_>>();
    assert_eq!(response_requests.len(), 2);
    let body = response_requests[1]
        .body_json::<Value>()
        .context("request body should be JSON")?;
    assert!(!body.to_string().contains("tab one"));

    Ok(())
}