dsh-ui-three-body

The most common frustration with AI agent workflows is rarely about model quality — it is about the absence of process. A one-line requirement jumps straight into code, the proposed design is invisible until too late, output formats drift, the agent declares victory with no verifiable acceptance, and tokens burn away with every extra round of clarification. dsh-ui-three-body turns DeepSeek Harness into a self-governed local delivery pipeline: requirements are decomposed into five verifiable facts and asked in a single pass; designs are pinned before implementation via clickable HTML prototypes; contracts are emitted as parseable objects to eliminate format drift; and acceptance is defined as executable assertions, not a blanket "done." The entire kernel adapts to task scale through S/M/L routing: small changes follow the laziest viable path, large tasks unlock orchestrated state machines and benchmark-driven optimization, with on-the-fly tier upgrades and downgrades. If the quality gate is not passed, nothing ships. Built-in self-healing (automatic backoff retries, provider failover, checkpoint resume) and an AI-mode token kill-switch (zero kernel tokens per turn when off) keep costs predictable.

The floating Zhizi on your screen — 11 dynamic skins, mouse-tracking pupils, a progress bar with a short goal title, and a random Three-Body quote every few seconds — doubles as both a live progress dashboard and a playful interaction surface.

Built for product managers, indie developers, and small teams who are tired of trusting an AI's word on delivery quality. You do not need a smarter model; you need a structured, auditable, locally owned pipeline where requirements are clarified once, designs are locked before code, contracts are machine-readable, and nothing crosses the finish line without passing the gate.