openai-codex is a public beta. Install it with pip install openai-codex; public APIs may change before 1.0. While beta releases are the only published SDK releases, pip selects the latest beta. After a stable release exists, pass --pre to opt into newer prereleases.

The SDK and runtime packages are versioned independently. Each SDK release pins and installs one compatible runtime dependency automatically.

• A Thread is conversation state.
• A Turn is one model execution inside that thread.
• Multi-turn chat means multiple turns on the same Thread.

• Thread.run(...) starts a turn and returns TurnResult.
• TurnHandle.run() / AsyncTurnHandle.run() consumes events for an existing turn handle and returns the same TurnResult shape.
• TurnHandle.stream() / AsyncTurnHandle.stream() yields raw notifications (Notification) so you can react event-by-event.

Choose run() for most apps. Choose stream() for progress UIs, custom timeout logic, or custom parsing.

• Codex is the sync public API.
• AsyncCodex is an async replica of the same public API shape.
• Prefer async with AsyncCodex() for async code. It is the standard path for explicit startup/shutdown, and AsyncCodex initializes lazily on context entry or first awaited API use.

If your app is not already async, stay with Codex.

• login_api_key(...) authenticates immediately with an API key.
• login_chatgpt() starts browser login and returns a handle with auth_url.
• login_chatgpt_device_code() starts device-code login and returns a handle with verification_url and user_code.
• Interactive handles expose wait() for the matching account/login/completed notification and cancel() to stop that attempt.
• account() reads the current account state, and logout() clears it.

Public API keyword names are snake_case. The SDK still maps them to wire camelCase under the hood.

If you are migrating older code, update these names:

• approvalPolicy -> approval_policy
• baseInstructions -> base_instructions
• developerInstructions -> developer_instructions
• modelProvider -> model_provider
• modelProviders -> model_providers
• sortKey -> sort_key
• sourceKinds -> source_kinds
• outputSchema -> output_schema

Use the same sandbox= keyword for threads and turns:

from openai_codex import Sandbox

thread = codex.thread_start(sandbox=Sandbox.workspace_write)
result = thread.run("Review only.", sandbox=Sandbox.read_only)

The presets are:

• Sandbox.read_only: read files without allowing writes.
• Sandbox.workspace_write: the normal default for projects with a recorded trust decision; read files and write inside the workspace and configured writable roots.
• Sandbox.full_access: run without filesystem access restrictions.

When sandbox= is omitted, Codex uses its configured default. A turn sandbox override applies to that turn and subsequent turns.

The public API keeps only explicit lifecycle calls:

• thread_start(...) to create new threads
• thread_resume(thread_id, ...) to continue existing threads

This avoids duplicate ways to do the same operation and keeps behavior explicit.

Codex() is eager: it starts transport and calls initialize in __init__.

Common causes:

• installation is incomplete and the pinned openai-codex-cli-bin dependency is missing
• local codex_bin override points to a missing file
• a custom local Codex executable does not support the SDK operation being used

A turn is complete only when turn/completed arrives for that turn ID.

• run() waits for this automatically.
• With stream(), keep consuming notifications until completion.

Use retry_on_overload(...) for transient overload failures (ServerBusyError).

Do not blindly retry all errors. For InvalidParamsError or MethodNotFoundError, fix the input or use the runtime pinned by the SDK.

• Starting a new thread for every prompt when you wanted continuity.
• Forgetting to close() (or not using context managers).
• Reading Turn.items from live start/completed payloads instead of using TurnResult.items.
• Mixing SDK input classes with raw dicts incorrectly.