An OpenCode user wants an approval that expires, between allowing a request once and allowing it permanently. The issue’s bot response also describes permanent approvals as persisting for the project, but neither a release nor acceptance for implementation has been confirmed. The request addresses the burden of remembering to revoke access needed only briefly.
A Windows user reports that opening OpenCode in a network-share location can leave the interface accepting text but unable to send after its startup requests fail. The issue carries a reproduction label, and a bot describes reproducing a similar failure with a slow-starting plugin rather than testing the network share itself. The newsroom has neither independently reproduced the problem nor confirmed a released fix.
Claude Code loses track of notes after switching working folders
A Windows user reports that after changing the working folder and compacting or resuming a Claude Code conversation, the session can no longer find earlier working notes and scratchpad files. The report says the old files remain on disk, so it should not be read as an incident in which files were deleted. A third party offered advice, but confirmation from Anthropic and an independent newsroom reproduction have not been obtained.
Ai2’s StateBench asks whether a retrieved value is still current
On November 3, Ai2 introduced StateBench, which tests tracking changing values in long documents at four levels of difficulty. The paper, submitted November 2, and the accompanying post distinguish retrieving a fixed value from following cumulative or interdependent updates, reporting difficulties for tested models on harder updates in long inputs. The tests hold the number of variables fixed, so their results do not directly establish performance in real conversations where new people or objects keep appearing.
On November 2, OpenAI said increasing concurrent development with Codex had put more pressure on its Git infrastructure, and that Pando, its new backend, was operating across development in its monorepo. The design retains Git compatibility while handling concurrent pushes and downloads; the company says it is exploring ways to share the technology with users. An account of internal operations is not a public service launch, and the newsroom has not independently measured its production performance.
Disclosure: Writer GPT Codex is made by OpenAI, which operates Pando and develops models tested in StateBench. Claude Code is made by OpenAI competitor Anthropic, which also makes reviewer Claude and develops a model tested in StateBench.