$ the-wire · showcase
Codex hardens WebSocket auth and network policy as SDKs add GCP storage
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
Codex's authentication and network-policy plumbing got extracted and tightened while openai-node and openai-python both shipped GCP external storage support and a new research model behind GPT-Rosalind.
Enforce network policy throughout HTTP and WebSocket requests openai/codex
Managed HTTP clients now run through a shared RequestBuilder that checks each redirect destination before following it, and policy denials survive error handling so callers can't retry them or record a revoked operation as a success. The stated purpose is keeping destination restrictions effective during redirects, response body reads, and established WebSocket traffic.
Retry transient OpenAI file blob upload failures openai/codex
A single failed blob upload used to abort the whole file upload even on a temporary 503 or an interrupted stream. Uploads now retry 503s plus timeout, connection, body, and request errors, at most five attempts inside a shared five-minute deadline, honoring x-ms-retry-after-ms and Retry-After and reopening contents for each attempt.
Extract WebSocket authentication into `codex-websocket-auth` (#47447) openai/codex
WebSocket auth arguments, settings, credential loading, and upgrade authorization moved into a shared codex-websocket-auth crate built on http types, consumed directly by the app server, transport, and CLI, with the AppServer prefix dropped from shared types. Token digests and JWT verification secrets are no longer printed in authentication policy Debug output.
v7.22.0 openai/openai-node
openai-node 7.22.0 adds GPT-6 Sol and Luna model identifiers to the API's known model set. Same-day releases also landed GCP as a supported external storage provider for organizations and gpt-rosalind-research in the Responses model union.
fix(helpers): use asyncio.get_running_loop() inside async methods (#3289) openai/openai-python
LocalAudioPlayer.play, LocalAudioPlayer.play_stream, and Microphone.record called asyncio.get_event_loop() from inside a running coroutine, deprecated since Python 3.10; they now use asyncio.get_running_loop(). The same release, openai-python 3.19.0, carries the GCP external storage and GPT-Rosalind additions.
