$ the-wire · showcase
Realtime reconnect opt-in in openai-node 7.39.0, strict-mode batch fix in openai-python
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
openai-node 7.39.0 adds opt-in bounded Realtime connection recovery that leaves default behavior untouched, while openai-python stops rejecting valid vector-store file-batch responses under strict validation.
feat(realtime): add opt-in bounded connection recovery openai/openai-node
The SDK now takes an opt-in option to reconnect itself on a temporary Realtime connection failure, reusing the shared WebSocket controller for both the ws and native WebSocket clients. Each attempt refreshes credentials and keeps the same connection object, but the replacement socket starts a new server session: state is not restored and attempted commands are not resent, so recovery at the tra...
v7.39.0 openai/openai-node
The release carrying the Realtime recovery option, dated 2026-10-18, lands as v7.39.0 off v7.38.0. Nothing in the release note beyond the one feature and a chores entry on neutral moderation examples, so the upgrade is additive unless you opt in.
fix(api): accept canonical batch responses in strict mode openai/openai-python
openai-python now accepts valid vector-store file-batch responses when strict response validation is enabled, preserving the returned object discriminator and continuing to validate the other response fields. If you turned on strict mode and were seeing batch endpoints fail on well-formed payloads, this is the fix.
Add observation-only model request lifecycle hooks openai/codex
Codex adds PreModelRequest and PostModelRequest hooks around every model request attempt, retries included, passing model, session, turn, and optional subagent context and reporting completion, failure, or cancellation with errors and token usage when available. The key constraint is in the design: hook output is diagnostic only and cannot veto a request or alter model context, with synchronous...
Prune unused standalone releases after successful installation openai/codex
Standalone installs now prune older releases after verifying the installed command, holding the install lock while doing it, and retaining the current and previously selected releases plus any used by running Codex processes or bundled helpers. Per Eric Traut's version of the change, cleanup only touches validated, installer-owned release directories and is skipped when filesystem or process ch...
