$ the-wire · showcase
Prewarmed hosted environments land in both SDKs, voice creation contract corrected
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
![openai-sdks[bot]](https://avatars.githubusercontent.com/in/3705508?v=4)
The day belongs to Agents infrastructure: prewarmed hosted environments shipped in tandem across openai-node and openai-python, alongside fixes that stop structured-output parsing from misfiring on commentary and stop token accounting from silently dropping detail counts.
feat(api): add prewarmed hosted environments openai/openai-node
Both SDKs now let you prepare an OpenAI-hosted environment before creating a beta Agents session, adding environment creation and paginated listing, session attachment by environment ID, a standalone ready status, and ready/failed webhook parsing. Prewarming access and the OpenAI-Beta: agents=v1 header remain required, so this is additive rather than a change to existing session flows.
feat(api): add prewarmed hosted environments openai/openai-python
The Python client gets the same hosted-environment surface as the Node one, so Agents prewarming is available regardless of which SDK your service is written in. Environment pagination retains response metadata, and existing hosted-environment configuration interfaces remain extendable.
fix(responses): parse structured text only in final message phases openai/openai-node
The Responses structured-output helpers used to parse every assistant message, so prose commentary raised a JSON parsing error and schema-valid commentary could land in output_parsed ahead of the final answer. The shared text-parsing boundary now checks message phase, parsing final_answer, absent, and null phases and leaving other explicit phases unparsed, with raw text, metadata, refusal handl...
fix(runners): preserve detailed token usage in totalUsage openai/openai-node
Runner totalUsage() and the totalUsage event rebuilt usage from only the three top-level totals, dropping cached-token and reasoning-token counts that individual completions already carried. The fix aggregates both detail objects across the run while preserving omitted fields and explicit zeros, and does not mutate source completions.
fix(api): correct custom voice creation parameters openai/openai-python
Custom voice creation now uses the supported consent-based audio upload contract, requiring a consent recording and an audio sample and keeping audio_sample as the documented response type. This shipped in openai-python 3.26.1, a release whose only change is this API parameter correction.
Action items
- → Upgrade openai to 3.26.1 if you create custom voices, since the consent-based parameter contract is the only change i... openai/openai-python [plan]
