$ the-wire · showcase
WebSocket binary frames survive reconnect in openai-python 3.40.0
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
A reconnect bug that silently turned binary frames into text frames (or crashed them out as UnicodeDecodeError) is fixed, and the deprecated model shutdown dates are now written down where your IDE can show them.
feat(websocket): recover queued binary frames after reconnect (#4191) openai/openai-python
During reconnect, send_raw(bytes) used to decode queued bytes as UTF-8, so a valid UTF-8 payload changed from a binary frame to a text frame and anything else raised UnicodeDecodeError before the SDK could send it. Each frame now keeps its type and its place in the queue, and when recovery stops, unsent_frames returns the original text and bytes in order while unsent_messages retains its text-o...
0.165.0 openai/codex
by github-actions[bot]
Codex adds a built-in Google Cloud Vertex AI provider via model_provider = "gcp", GPT-6.1 Terra in the model catalogs, and reasoning effort changes mid-conversation enabled by default for supported models. The session picker also gains a repository filter and directory grouping across linked worktrees.
chore(api): document upcoming model shutdowns openai/openai-node
openai-node now documents announced shutdown dates for deprecated model options while leaving every existing model value available, including the March 2025 TTS snapshot in a January 6, 2027 notice. Audio translations are marked deprecated with migration to Realtime translations before February 26, 2027.
Preserve output boundaries through token truncation openai/codex
Truncation no longer loses the mapping between items: OutputBoundary and the token-budget formatting APIs remap item indices and UTF-8 text offsets to their post-truncation positions, and keep end boundaries ahead of generated omission notices. Code mode output formatting is unified, budgeting mixed text and audio in input order while preserving images and encrypted content.
chore: enforce generated-file ownership openai/openai-node
The generated-file customization budgets, reporters, workflows and zero-only policy are gone; ownership now rests on AGENTS.md and ordinary code review, with handwritten edits forbidden in generated files in favor of fixing the inputs or compiler and regenerating.
Action items
- → Upgrade openai-python to 3.40.0 if you send binary frames over Realtime and hit reconnects openai/openai-python [plan]