dsh-showme-html

An agent finishes a big task and dumps two hundred lines of Markdown. You read it all, yet you never actually saw the thing. When a dozen AI-generated variants sit side by side, you cannot name them — you point and say the third one, the one with the gradient, and whether you and the agent mean the same third one is a coin flip. Every step on the read-to-feedback path loses meaning.

dsh-showme-html gives the agent a show_html tool that renders a finished HTML page as an inline card inside the conversation. Every element on the page carries a stable, visible name (hero-banner-v3, f2, shot-07); the user clicks twice, the page drafts a one-line intent on his behalf, and that line lands in the input box or the feedback inbox with one press. Four preset styles (Soft, Swiss, Brutal, Blueprint) share the same semantic markup so switching skins requires zero tag changes. Three template skeletons (per-item review, candidate pick, illustrated report) turn page writing into data filling — a few hundred tokens instead of a 10 KB rewrite that might get the feedback channel wrong every time. Zero build step, zero dependencies; the host is the web server DSH is already running.

Who it is for: anyone working in a DSH or DeepSeek Harness workflow who regularly needs to visually compare AI outputs, pick among multiple variants, and give structured, element-level feedback. No extra services, no build pipeline — install and go.