Long sessions tend to end the same way: the primary model starts returning rate-limit errors, or the provider goes quiet mid-task. The work is half done, the context is fully loaded, and the fix is either to wait or to switch models by hand and hope continuity survives. If you run more than one...
Source: [Dev.to](https://dev.to/sertdisk/opencode-model-router-per-agent-model-fallback-chains-with-a-local-web-ui-4n23)