Issue with Kilo Code Orchestrator Mode in DeepSeek 3.1 – Fails to Split Tasks Properly

[kenken1985]

I've been noticing some issues with the Kilo Code Orchestrator mode in DeepSeek v3.1. It appears that the orchestrator is no longer effectively breaking down tasks into smaller subtasks as expected.

Instead, it either:

• Pushes the entire task into a single large code subtask, or
• Bypasses the Architect mode entirely and switches directly into Code mode.

This behavior often cause the task to be to big and cause the overthrow the context limit. It worked more effectively in earlier versions, so this seems like a potential regression.

Has anyone else experienced similar behavior? Any suggestions or workarounds?