$ the-wire · showcase
Codex binds backend auth at request time, SDKs add 12-hour prompt caching
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
![openai-sdks[bot]](https://avatars.githubusercontent.com/in/3705508?v=4)
Both SDKs shipped 12-hour Responses prompt-cache TTLs and typed cache-write usage, while Codex moved backend client credentials to per-request runtime authentication.
Bind reusable backend clients to runtime authentication openai/codex
Backend clients can now refresh tokens per request within the credential source and transport policy chosen at construction, but replacing credentials, destination, or account/organization selectors is rejected. Long-lived clients no longer need rebuilding around rotation, at the cost of failing loudly on any selector change.
feat(api): add 12-hour Responses caching and cache-write usage openai/openai-python
Python 3.45.0 accepts "12h" alongside "30m" in Responses request and response TTL types and exposes optional cache-write details split into 30-minute and 12-hour token counts. The PR notes that callers typing a returned TTL as exactly Literal["30m"] should allow "12h" too, and the usage breakdown now decodes as a typed model. The same change landed in openai-node's v7.48.0 line.
feat(api): expand agent turn errors and image and audio options openai/openai-node
Adds max_output_tokens as a recognized agent session turn failure code, image context compression on Responses requests, and noise-reduction strength on Realtime input audio. In Python, responses.parse() and responses.stream() now forward image through the request like responses.create(), closing a gap that made compression configurable only on one path.
feat(api): add workload identity providers and federation rules openai/openai-node
openai-node gained CRUD and cursor-paginated listing for OIDC workload identity providers plus federation rules on existing project service accounts, including attribute matchers and permissions, managed with an Admin API key. The companion project service tier call lets you read and update default and enabled tiers independently, including clearing the enabled optional list.
Expand Windows deny-read globs without invoking ripgrep (#59250) openai/codex
Codex now expands Windows deny-read globs with the ignore crate in-process, replacing the external rg scan and fallback walker, and drops scanner discovery from MXC sandbox dependency checks. Case-insensitive matching and scan depth limits are unchanged; directory links below the scan root are skipped while explicitly linked roots are followed. Three further Codex changes are routine: automatic...
