$ the-wire · showcase
Codex closes an input race, API key expiry lands in both SDKs
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
Codex hardened turn finalization against inputs that arrived mid-drain, while openai-node and openai-python both shipped the same pair of API changes on the same day.
Close input admission while a turn finishes recording (#53452) openai/codex
Codex now marks active turns as finishing before draining pending input, so injections into a finishing turn are rejected and late client responses and notifications wait for the final batch to be recorded. Inputs arriving during finalization could previously be queued after the final batch drained, or recorded ahead of it.
Isolate arg0 helper directories by operating system (#53412) openai/codex
Helpers now live under CODEX_HOME/tmp/arg0-<os> and stale-directory cleanup is limited to the current operating system's root. That closes a Windows-and-WSL hazard: the two can share CODEX_HOME without sharing a lock namespace, so cleanup could delete another platform's live helpers, and older clients could still clean the shared tmp/arg0 root.
Allow turns when the agent message board is unavailable (#53369) openai/codex
With multi_agent_v2.disable_direct_message enabled, an unavailable agent message board could reject a turn when the tool router was constructed. The router no longer has to expose the board's post tool, and unavailable board calls now return an unsupported-call error to the model so the turn continues.
feat(api): support updating project API key expiry openai/openai-node
client.admin.organization.projects.apiKeys.update sets an API key's expiry relative to the request time, requiring a project ID, key ID, and expires_in_seconds, and returns the updated key resource; the Python client shipped the matching client.admin.organization.projects.api_keys.update. The same day also brought agent tool-list limits into the open: a 3 MiB compact JSON ceiling that is now do...
Remove selected protocol, apply-patch, MCP, and TUI tests (#53358) openai/codex
Also in codex: final polling of remote processes with still-open output streams got a regression test (exec_command session, write_stdin completes after exit), and eight protocol, apply-patch, MCP, and TUI tests were removed with runtime behavior unchanged.
