@zhang8019

Permission matrix for DeepSeek Harness: 3 sandbox modes x 4 approval strategies = 9 presets, with global and LLM-robot defaults, three-tier risk policies, audit log and git checkpoint.