codex/sdk/python/docs/api-reference.md

OpenAI Codex SDK — API Reference

Public surface of openai_codex for app-server v2.

This SDK surface is experimental. Turn streams are routed by turn ID so one client can consume multiple active turns concurrently. Thread starts default to ApprovalMode.auto_review; turn starts accept an optional approval_mode override.

Package Entry

from openai_codex import (
    Codex,
    AsyncCodex,
    ApprovalMode,
    ChatgptLoginHandle,
    DeviceCodeLoginHandle,
    AsyncChatgptLoginHandle,
    AsyncDeviceCodeLoginHandle,
    Thread,
    AsyncThread,
    TurnHandle,
    AsyncTurnHandle,
    TurnResult,
    Input,
    InputItem,
    RunInput,
    TextInput,
    ImageInput,
    LocalImageInput,
    SkillInput,
    MentionInput,
)
from openai_codex.types import (
    Account,
    AccountLoginCompletedNotification,
    CancelLoginAccountResponse,
    CancelLoginAccountStatus,
    GetAccountResponse,
    InitializeResponse,
    ThreadItem,
    ThreadTokenUsage,
    TurnError,
    TurnStatus,
)

• Version: openai_codex.__version__
• Requires Python >= 3.10
• Public app-server value and event types live in openai_codex.types

Codex (sync)

Codex(config: AppServerConfig | None = None)

Properties/methods:

• metadata -> InitializeResponse
• close() -> None
• login_api_key(api_key: str) -> None
• login_chatgpt() -> ChatgptLoginHandle
• login_chatgpt_device_code() -> DeviceCodeLoginHandle
• account(*, refresh_token: bool = False) -> GetAccountResponse
• logout() -> None
• thread_start(*, approval_mode=ApprovalMode.auto_review, base_instructions=None, config=None, cwd=None, developer_instructions=None, ephemeral=None, model=None, model_provider=None, personality=None, sandbox=None) -> Thread
• thread_list(*, archived=None, cursor=None, cwd=None, limit=None, model_providers=None, sort_key=None, source_kinds=None) -> ThreadListResponse
• thread_resume(thread_id: str, *, approval_mode=ApprovalMode.auto_review, base_instructions=None, config=None, cwd=None, developer_instructions=None, model=None, model_provider=None, personality=None, sandbox=None) -> Thread
• thread_fork(thread_id: str, *, approval_mode=ApprovalMode.auto_review, base_instructions=None, config=None, cwd=None, developer_instructions=None, model=None, model_provider=None, sandbox=None) -> Thread
• thread_archive(thread_id: str) -> ThreadArchiveResponse
• thread_unarchive(thread_id: str) -> Thread
• models(*, include_hidden: bool = False) -> ModelListResponse

Context manager:

with Codex() as codex:
    ...

AsyncCodex (async parity)

AsyncCodex(config: AppServerConfig | None = None)

Preferred usage:

async with AsyncCodex() as codex:
    ...

AsyncCodex initializes lazily. Context entry is the standard path because it ensures startup and shutdown are paired explicitly.

Properties/methods:

• metadata -> InitializeResponse
• close() -> Awaitable[None]
• login_api_key(api_key: str) -> Awaitable[None]
• login_chatgpt() -> Awaitable[AsyncChatgptLoginHandle]
• login_chatgpt_device_code() -> Awaitable[AsyncDeviceCodeLoginHandle]
• account(*, refresh_token: bool = False) -> Awaitable[GetAccountResponse]
• logout() -> Awaitable[None]
• thread_start(*, approval_mode=ApprovalMode.auto_review, base_instructions=None, config=None, cwd=None, developer_instructions=None, ephemeral=None, model=None, model_provider=None, personality=None, sandbox=None) -> Awaitable[AsyncThread]
• thread_list(*, archived=None, cursor=None, cwd=None, limit=None, model_providers=None, sort_key=None, source_kinds=None) -> Awaitable[ThreadListResponse]
• thread_resume(thread_id: str, *, approval_mode=ApprovalMode.auto_review, base_instructions=None, config=None, cwd=None, developer_instructions=None, model=None, model_provider=None, personality=None, sandbox=None) -> Awaitable[AsyncThread]
• thread_fork(thread_id: str, *, approval_mode=ApprovalMode.auto_review, base_instructions=None, config=None, cwd=None, developer_instructions=None, ephemeral=None, model=None, model_provider=None, sandbox=None) -> Awaitable[AsyncThread]
• thread_archive(thread_id: str) -> Awaitable[ThreadArchiveResponse]
• thread_unarchive(thread_id: str) -> Awaitable[AsyncThread]
• models(*, include_hidden: bool = False) -> Awaitable[ModelListResponse]

Async context manager:

async with AsyncCodex() as codex:
    ...

Login handles

ChatgptLoginHandle / AsyncChatgptLoginHandle

• login_id: str
• auth_url: str
• wait() -> AccountLoginCompletedNotification
• cancel() -> CancelLoginAccountResponse

Async handle methods return awaitables.

DeviceCodeLoginHandle / AsyncDeviceCodeLoginHandle

• login_id: str
• verification_url: str
• user_code: str
• wait() -> AccountLoginCompletedNotification
• cancel() -> CancelLoginAccountResponse

Async handle methods return awaitables.

wait() consumes only the completion notification for its matching login attempt. API-key login completes synchronously and does not return a handle.

Thread / AsyncThread

Thread and AsyncThread share the same shape and intent.

Thread

• run(input: str | Input, *, approval_mode=None, cwd=None, effort=None, model=None, output_schema=None, personality=None, sandbox_policy=None, service_tier=None, summary=None) -> TurnResult
• turn(input: str | Input, *, approval_mode=None, cwd=None, effort=None, model=None, output_schema=None, personality=None, sandbox_policy=None, service_tier=None, summary=None) -> TurnHandle
• read(*, include_turns: bool = False) -> ThreadReadResponse
• set_name(name: str) -> ThreadSetNameResponse
• compact() -> ThreadCompactStartResponse

AsyncThread

• run(input: str | Input, *, approval_mode=None, cwd=None, effort=None, model=None, output_schema=None, personality=None, sandbox_policy=None, service_tier=None, summary=None) -> Awaitable[TurnResult]
• turn(input: str | Input, *, approval_mode=None, cwd=None, effort=None, model=None, output_schema=None, personality=None, sandbox_policy=None, service_tier=None, summary=None) -> Awaitable[AsyncTurnHandle]
• read(*, include_turns: bool = False) -> Awaitable[ThreadReadResponse]
• set_name(name: str) -> Awaitable[ThreadSetNameResponse]
• compact() -> Awaitable[ThreadCompactStartResponse]

run(...) is the common-case convenience path. It accepts plain strings, starts the turn, consumes notifications until completion, and returns a small result object with:

• id: str
• status: TurnStatus
• error: TurnError | None
• started_at: int | None
• completed_at: int | None
• duration_ms: int | None
• final_response: str | None
• items: list[ThreadItem]
• usage: ThreadTokenUsage | None

final_response is None when the turn finishes without a final-answer or phase-less assistant message item.

Use turn(...) when you need low-level turn control (stream(), steer(), interrupt()) before collecting the turn result.

TurnHandle / AsyncTurnHandle

TurnHandle

• steer(input: str | Input) -> TurnSteerResponse
• interrupt() -> TurnInterruptResponse
• stream() -> Iterator[Notification]
• run() -> TurnResult

Behavior notes:

• stream() and run() consume only notifications for their own turn ID
• one Codex instance can stream multiple active turns concurrently

AsyncTurnHandle

• steer(input: str | Input) -> Awaitable[TurnSteerResponse]
• interrupt() -> Awaitable[TurnInterruptResponse]
• stream() -> AsyncIterator[Notification]
• run() -> Awaitable[TurnResult]

Behavior notes:

• stream() and run() consume only notifications for their own turn ID
• one AsyncCodex instance can stream multiple active turns concurrently

Inputs

@dataclass class TextInput: text: str
@dataclass class ImageInput: url: str
@dataclass class LocalImageInput: path: str
@dataclass class SkillInput: name: str; path: str
@dataclass class MentionInput: name: str; path: str

InputItem = TextInput | ImageInput | LocalImageInput | SkillInput | MentionInput
Input = list[InputItem] | InputItem
RunInput = Input | str

Use a plain str as shorthand for TextInput(...) anywhere a turn input is accepted: thread.run("..."), thread.turn("..."), and turn.steer("...").

Public Types

The SDK wrappers return and accept public app-server models wherever possible:

from openai_codex.types import (
    Account,
    AccountLoginCompletedNotification,
    CancelLoginAccountResponse,
    CancelLoginAccountStatus,
    GetAccountResponse,
    ThreadReadResponse,
    Turn,
    TurnStatus,
)

Retry + errors

from openai_codex import (
    retry_on_overload,
    JsonRpcError,
    MethodNotFoundError,
    InvalidParamsError,
    ServerBusyError,
    is_retryable_error,
)

• retry_on_overload(...) retries transient overload errors with exponential backoff + jitter.
• is_retryable_error(exc) checks if an exception is transient/overload-like.

Example

from openai_codex import Codex

with Codex() as codex:
    thread = codex.thread_start(model="gpt-5.4", config={"model_reasoning_effort": "high"})
    result = thread.run("Say hello in one sentence.")
    print(result.final_response)