MiniMax H3 Chain Preflight

Preflight is the cheapest node in the Contex Loop pack, which is exactly why it's the most useful. MiniMax H3 renders are minutes long and a 33B omni-modal model doesn't load in a blink; a Plan that's silently wrong - a scene duration that rounds to an invalid frame count, a reference timestamp off the canvas, a source track that won't line up, a resume whose predecessor dependencies changed - will eat all of that before you ever see a frame. This node catches it before the model loads. It is deliberately model-free: no weights, no VRAM, no waiting.

The README calls this "preflight checks timing, media, references, compatibility, and resume state before H3 loads," and that's the whole job in one sentence. Wire it between your Plan and Loop Start, and it validates Plan timing, the Source Timeline, reference registries, runtime compatibility, and any resume artifacts. Nothing about generation is touched - it's a report, not a render.

What it checks

• Plan timing - durations, context lengths, the AV clock math. H3 can't generate every frame count, so lengths round up to valid 17k+5 raw values; Preflight knows this and flags scenes that came out wrong.
• Media - the optional source_timeline and (legacy) source_audio routes: do the streams exist, do they decode, do their 24-fps extents match what the Plan expects.
• References - the optional tagged_references and reference_schedule registries, so a dangling @tag or an anchor off the canvas shows up before you queue.
• Resume state - start_clip above 1 means "resume here," and Preflight compares every saved predecessor dependency against the active Plan. verify_resume_history (on by default) is the strictness switch; only turn it off for an intentional advanced recovery from known artifacts.

There's also scene_range, a contiguous scene selection string if you only want to preflight part of a production.

Outputs

The socket set is the same as Plan Studio's preflight half: preflight (structured report), ready (boolean - true only when no blocking error remains), status (a concise error/warning count for a Text node), and report_json. The plan output passes the connected Plan through unchanged, so you can insert Preflight inline between Plan and Loop Start without re-routing anything.

A useful trick: put a Text node on status and glance at it before every queue, and wire ready into your own sanity switch if you want the graph to refuse to proceed on a red flag.

Install

This is a node from the ComfyUI-MiniMaxH3-Contex-Loop pack. Install once via ComfyUI Manager (search "ComfyUI-MiniMaxH3-Contex-Loop") or:

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop.git

Restart. No extra pip packages - the pack uses ComfyUI's built-ins. It does expect a current ComfyUI with native Add Guide for MiniMax H3 (PR #15439) for the generation half of the graph, and remember the H3 weights themselves are not bundled and are geofenced out of the US, EU, UK and South Korea by the MiniMax H3 Community License.

The honest caveat

Preflight is thorough, but it validates what the Plan declares against what's on disk - it can't predict the model deciding to melt a face. Treat a green ready as "your production is internally consistent," not "this render will be good." The point of the node is to kill the boring failures for free so your GPU time is spent on taste, not typos.