$ the-wire · showcase
Codex stops prompting for runtime-only grants, keeps its screen during editor handoff
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
2 people shipped this
Codex spent the day fixing the ways its own approval and rendering machinery gets in the way of the agent loop, while the Node and Python SDKs received documentation-only error coverage across nine pull requests.
Avoid stdin approval for runtime-only permission grants openai/codex
Retained runtime filesystem grants, including the ones used for plugin metrics, are now folded into the baseline permission comparison, so reachable-only grants no longer trigger an approval prompt for write_stdin or swallow input when sandbox approval is disabled. Review is still required for additional agent permissions, sandbox bypasses, and policy changes, and the approval description now r...
Keep Codex visible during external editor handoff openai/codex
Opening a separate-window editor from the fullscreen TUI used to leave the alternate screen and hide both your draft and the instruction to save and close. Codex now preserves and repaints the last frame while handing the terminal over, releases keyboard and mouse input modes, and tracks whether the alternate screen's keyboard mode is active so cleanup does not pop it twice if the editor switch...
Add a keep-and-next action to the warnings viewer openai/codex
Pressing k keeps the current warning and advances to the next one, closing the viewer only after the last warning, and kept warnings survive queued dismissal updates so a later keep decision wins for the same warning details. The companion change dismisses only the warning pages actually rendered when you close with Esc, F2, or Ctrl+C, with the footer hint now reading "dismiss & close"; unviewe...
Deduplicate retained instructions in async Guardian context openai/codex
Async Guardian requests were shipping retained user instructions even when the same request's transcript already carried complete copies. Async user transcript entries now include retained source metadata and order labels, and retained instructions are omitted when the current transcript holds the same complete source revision, while ordering guidance still rides along in every async sample.
chore(api): clarify documented API error responses openai/openai-python
The Python and TypeScript SDKs both landed the same set of documentation-only changes, describing applicable errors across organization, inference, media, and resource APIs plus fine-tuning, model, Batch, and stored chat completion responses. Methods, return types, exception types, and runtime error behavior are unchanged in every case.
