dsh-generative-ideas

Drawing a roadmap is rarely about whether you can produce one; it is about whether you can produce several genuinely different ones and select the right one with evidence. Most teams quietly converge on a single narrative, skip a systematic scan of competitors and the open-source landscape, and end up deciding on gut feel. dsh-generative-ideas plugs into DeepSeek Harness to close that gap: it spawns a background headless agent that generates four distinct roadmap options in one pass, each carrying a name, thesis, phased acceptance items, risks, effort estimate, and a differentiator, shifting the decision from vibes to a structured comparison.

What sets the plugin apart is its deep-research doctrine. With the toggle on by default, the generation agent is mandated to study at least twelve competitors and record their concrete methods, read GitHub open-source repositories including READMEs, design docs, and issue trackers, and incorporate curated references from a .refs/ directory before it is even allowed to draft options. Every thesis names what the best-in-class competitor does differently, every phase draws on real patterns, and every risk is grounded in a failure another project actually encountered. A generation run takes roughly thirty to ninety seconds; the sidebar icon pulses while the agent works, and results persist even if you close and reopen the panel. On the compare step, three pre-flight actions are available: Reroll for a fresh set of options under the same focus, Push to force deep research back on and regenerate, and Discuss to drop into a free-form chat pass. Once an option is chosen, a single click exports it as goal.md into the target workspace root, ready to be picked up by a downstream agent.

It is built for developers, engineering leads, and small teams working inside DeepSeek Harness who want a structured, research-backed ideation loop rather than another round of unstructured brainstorming.