Community Pulse—From the Floor — From the Floor

A November 2 GitHub report says a non-admin workspace member cannot find remaining Codex allowance or reset times in the CLI or web interface, although requests still work. The reporter points out that context capacity differs from account allowance and asks for included allowance, credit balance and banked resets to be distinguished. The account’s precise billing tier and the cause remain unconfirmed, and we have not independently reproduced this single account’s report.

A November 2 report describes child search processes continuing to consume memory after a Bash-tool task in the Code tab of the Claude desktop app on Windows was stopped or its parent process ended. The user asks for the whole process tree to end together and for surviving processes to be visible. We found no developer confirmation or announced fix in the issue’s activity, and have not reproduced the report.

An Ollama user reported on November 2 that memory estimates for some hybrid models on Apple Silicon cause already loaded models to be unloaded, despite a setting intended to keep them resident. The report includes logs and a proposed calculation change, but we have not verified the diagnosis or found an associated merged fix. Unloading here means removing a model from memory, not deleting its saved files.

Anthropic’s November 2 release notes introduce message threads in beta, keeping a conversation briefly on the server so subsequent requests can send only new messages. A thread expires five minutes after its last request and is deleted within an hour of its last use; this feature is not eligible for zero data retention. The documentation also tells applications to keep their own conversation copy because a thread may need to be rebuilt; we have not tested the claimed latency benefit.

LongCat’s official API changelog lists a November 2 LongCat-2.5 release separately from its September Preview entry. It describes visual understanding, long-running coding tasks and compatibility with tools including Claude Code, Codex and Hermes. These are provider descriptions; we have not tested the API or established comparative performance or weight availability.