So my long running agents are really not doing what is supposed to be. Noticed this a few times. It will complete the first goal fine, but if it went to the wrong path and I need to steer it towards a new path, it will start failing to be a long running agent.
Defeats the purpose of accessing it from my mobile, giving a prompt and catching it when its ready.
Tbh just using codex /goal is being more reliable than this, and it uses $$ from my sub and not API.
Right now I am stuck on planning next moves, loading conversation.
Will have to send a half done PR and catch things up from another session, which will have to load the context from the beginning.
Thanks for the Request IDs. What I can see in the logs matches some known error patterns we’re tracking, and I’ve added those IDs to those tickets to help our investigation.
Opus 4.7 has been a bit tricky with long-running tasks. If you’re hitting this regularly, I’d recommend trying 4.6 instead, or switching to a normal (not grind mode) cloud agent. Long-running agents are great for specific workflows, but for most tasks, a regular cloud agent will probably give you a smoother experience.
