H16-3 · Window Relay Actual Calls (T8 EXP)

Timed prompts are one of those features that are easy to wire and hard to verify. You add a Prompt Relay plan, you project it onto an H16 window with MiniMaxH3H16RelayProjectEXPT8, you sample, and the result looks… fine? Different? You genuinely cannot tell whether the route re-weighted attention at frame 60 or whether the whole thing ran as dense with your timed events swept into the bit bucket.

This node closes that loop for one projected Relay window. It reads routed attention calls and completed forwards out of the finished window and shows them.

Why the pack treats this as a separate job

Because a wrapper can quietly take the same hook. The H3 stack in this pack has a lot of patch owners competing for the same entry point - attention backends, Turbo LoRAs, memory nodes, effect wrappers - and the author's policy is explicitly "preserve or delegate unknown patches, don't reject them". Which is friendly, and also means a Relay you wired might execute, might partially execute, or might land on a model another patch already owns.

The audit's job is to distinguish those cases with evidence rather than vibes. It reads the immutable record attached to that specific completed window result - not a mutable counter on the Apply/Project node, which tells you what was requested. That distinction is the entire reason the node exists, and the author is blunt about it: reading the old counters can't prove sampling succeeded.

Inputs and outputs

Three required inputs, all typed: window_result (from the H16 PASS2 window you're auditing), segment_spec (the window identity, so the audit can confirm plan SHA and source identity line up), and runtime (the T8_H16_RELAY_RUNTIME handle produced by MiniMaxH3H16RelayProjectEXPT8 for that window).

Out go cumulative_av_latent and report_json, and since it's marked as an output node, it works as a terminal - the report text also shows up in the node UI. Say it with me: the latent output is the H16-selected audio/video, and this node is a read-only observer. It doesn't alter your window.

Relay alone, or Relay plus EAV

Note the split the author made: this is the standalone Relay audit. If you've composed an external Relay and an EAV effect on the same window, use the EAV joint audit instead, because that one accounts for both call classes. Running the Relay-only audit on a Relay+EAV window will under-report, and you'll conclude your effect was a dud when it was the audit that was short-sighted.

Also note the EAV audit and this one take differently-typed runtimes. They're not interchangeable even though the schemas look similar at a glance.

Install

Same pack, same drill. Manager → MiniMax H3 Audio T8, or clone:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Full restart after install or update, and keep ComfyUI core, frontend and Manager current - this family leans on recent native H3 APIs, and a stale core is the usual reason an entire category turns red. The pack deliberately declares no pip dependencies; anything optional checks its own requirements only when you actually use it.

Where people get burned

The most common mistake is a data-flow pairing error: taking runtime from one window's projection and window_result/segment_spec from another. The pack is strict about identity, so you'll usually get a clean refusal instead of a plausible-looking wrong number - but if you're looping windows with a shared node and a widget you forgot to update, "usually" is doing heavy lifting.

The second is treating a green audit as a quality signal. It isn't. report_json can show routed attention and completed forwards and the clip can still look like a different take than the one in your head - the Relay effect is a conditioning change, and whether it reads as better is your eyeballs' call.

The third is cost. Every window you audit is another terminal node and another report to actually read. If you're debugging, audit window 0 and one mid-window. If everything works and you're shipping, you can live without the audit - just don't fool yourself that an un-audited Relay necessarily ran.