Vision for text-only DeepSeek Harness agents: subagent-delegated image reading + Codex-style paste bridge on a configurable MiniMax/Kimi route