dsh-superpowers

AI coding agents write fast, but speed is not quality. Without a methodology guardrail, models routinely skip brainstorming, ship features without tests, guess at bug fixes, and claim completion with zero verification evidence. The code looks like it works, but it will not survive a real code review.

dsh-superpowers ports the full obra/superpowers agentic skill framework into DeepSeek Harness, injecting a fixed set of engineering disciplines into every session: clarify intent, write a plan, dispatch subagents, enforce red-green-refactor, perform root-cause analysis before touching a fix, run code review in both directions, gather verification evidence, and only then merge. Fourteen skills cover the entire pipeline from requirement to integration. On session start a bootstrap paragraph is auto-injected into the system prompt, so the model checks for a matching skill before every reply, turning the agent into a self-disciplined engineering teammate.

It runs natively in DSH with zero external dependencies, pure Node built-in modules, and hot-reload of skill text on next load. Skills trigger automatically in natural language, can be invoked explicitly via slash commands, or skipped on demand. It is built for developers and teams doing serious work in DeepSeek Harness: engineers who want AI collaboration without giving up engineering rigor, tech leads who need subagent output that is reviewable, and anyone who wants the superpowers methodology inside their existing workflow.