$ the-wire · showcase
Prompt cache TTLs and agent scheduling land in both SDKs
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
The generated spec for agent scheduling APIs, Live delegation controls, and prompt cache lifetimes reached openai-node and openai-python together, alongside a namespace-validation fix in the Python Responses runner.
feat(api): add delegation controls and prompt cache lifetimes openai/openai-node
Both openai-node and openai-python gained optional 30m and 12h TTLs on prompt cache breakpoints, including Responses message inputs, plus acknowledgement_delay_ms on Live delegation parameters. The same drop adds beta agent scheduling APIs for creating, listing, retrieving, updating, deleting, pausing, and resuming scheduled tasks, and an automations agent tool variant.
fix(responses): validate tool namespaces before runner dispatch openai/openai-python
Following the namespace-aware parsing added previously, a function call whose registered name arrived under an unrelated namespace could leave arguments unparsed while the Responses tool runner still dispatched the callback. The runner now rejects namespace mismatches before dispatch and requires parsed arguments for typed callbacks before caching them, leaving ordinary un-namespaced and deferr...
perf(agents): reuse action checks between stream events openai/openai-node
finalResult() was rechecking the same required-action list on every streamed event; it now caches successful checks per snapshot and handler predicate, resetting when the snapshot changes or the collection is released. Failure, cancellation, and final completion predicates still run on every call, and the count-based regression test fails without the change.
Prevent memory imports from racing with consolidation openai/codex
Concurrent memory imports and consolidation could modify the same workspace, leaving a partial import across memory versions or disturbing consolidation snapshots and baselines. A cross-process workspace lock now covers both paths: imports acquire every destination lock before touching a workspace and return a retryable busy error if one is held.
Move Guardian approval reason caps into context composition (#56477) openai/codex
The 512-token approval reason cap now applies inside guardian-context when composing synchronous full and delta reviews, before request budget enforcement, instead of truncating at the requesting host, and the exact action JSON is preserved. Elsewhere across codex, message boards got an explicitly selected append tool with no subscription or notification side effects, and a new experimental thr...
