沐阳 H3 · Agent 预览

Every prompt-rewriting pipeline has the same debugging moment: the output looks wrong, and you have no idea whether it's the LLM's fault, the media references, or your own script. MiniMaxH3Viewer is the pack's answer - a unified preview node that renders the Media Agent's rewritten prompt, the storyboard summary, and the raw prompt into one formatted panel you can actually read. It's a display node, not a processing node, and it's deliberately simple.

It has one required input, text (a multiline string pre-filled with a demo template showing the format - a "final rewritten prompt" section and a storyboard/summary section), plus three optional inputs that you'll normally wire rather than type into: agent_prompt, summary_json, and myang_prompt, straight from the Media Agent's outputs. If you connect agent_prompt and summary_json, it renders them in the same readable layout; if summary_json isn't valid JSON it shrugs and falls back gracefully instead of erroring. The single output is preview_text, the formatted string, which you can route anywhere or just admire.

The detail that makes this node behave differently from most: it overrides IS_CHANGED to always return a change. That's deliberate. ComfyUI caches nodes between runs, and a cached preview node means the panel keeps showing the last run's text even after the agent produced new output - which is exactly the bug you'd spend an hour chasing. Forcing a re-run costs almost nothing (there's no model work here) and guarantees the preview always matches the current graph. Good design.

There's a subtlety about what the viewer is for, though. It's great for checking the agent's work before you commit tokens to a render - glance at the storyboard summary, confirm the media references look sane, then queue. It's not a substitute for reading myang_prompt when something's actually broken: the formatted view is for humans, and the underlying myang_prompt string is the ground truth you'd feed downstream. The viewer's text field doubles as a manual template, so you can also use it as a scratchpad for hand-writing prompts in the same style.

As a thin, well-scoped utility node, install is the one-liner shared across the pack: ComfyUI Manager → ComfyUI-MiniMaxH3-Myang → install → restart (hard-refresh the browser if the UI looks stale), or git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang.git into custom_nodes. No extra Python dependencies, no model files of its own - it needs only the agent outputs to be interesting.

Reach for this node whenever you're iterating on Media Agent output and can't tell what the LLM did. It's the "show your work" node, and in a pack that automates a lot of decisions for you, that's quietly the most useful thing in it.