H3 Hermes Prompt Director

The two other H3 prompt nodes in this pack get a VLM to write a prompt. This one goes a step further: it uses Hermes Agent to research, draft, critique, and select a MiniMax H3 prompt, while TrentNodes keeps timing and format validation local. In plain terms - an agentic loop where the prompt is drafted, then reviewed against the H3 spec, then revised, instead of written once and repaired.

The division of labor is the interesting part. The agent gets to do the open-ended thinking (what should this shot actually be, what current H3 guidance says), and the node owns the things that must not be left to judgment: timing, format, validation. The README calls that "secure Ref2VA video/frames with optional audio" plus all Base modes, with untouched reference pass-throughs.

How it works

Required inputs set the job: h3_mode (ref, or the four base modes base_T2VA/base_I2VA/base_FL2VA/base_L2VA), subjects (the typed escape hatch), target_description (what the target video should show), quality_mode (fast/balanced/hero - a bounded Hermes candidate/review budget), and research_policy (never/when_uncertain/always - when Hermes may look up current H3 guidance).

It talks to a Hermes API, which is a real infra requirement: hermes_base_url defaults to http://127.0.0.1:8642 (a root-only plain-HTTP loopback URL), with hermes_provider and hermes_model for route selection. timeout_seconds (default 900) is a hard wall-clock limit, and poll_interval_seconds controls polling. This is not a node you run without a Hermes server standing by.

The subject-row system matches the Ultimate Cowboy node - kinds, names, descriptions, with subject_N_image slots that are <Picture N>. video_role (subject_source / structure_reference / edit_source / continuation_source) decides what the wired video is for, which is what determines the task type - wiring alone can't imply it. audio_role does the same for audio (none / reuse / reference). And cut_times accepts any Cut Detective output as ground truth for the [Shot N] times.

The validation story

Dialogue must reach a <d> block verbatim or the run retries once. constraint_notes get folded in as positive assertions because H3 has no negative field and the official format ends at non_diegetic_music. strict_duration rejects durations outside H3's official 4-15 second range. cleanup_policy handles the staged images the agent needs (delete_on_success / retain_24h / retain).

Outputs mirror the Ultimate node's honest wiring: h3_prompt, duration_seconds, fps, analysis_json, h3_checkpoint_hint, the six ref_image_N pass-throughs, ref_video/ref_video_audio/ref_audio, width/height/length on H3's grid, and label_map. Same "nothing is wired twice" guarantee - the assets you plug in come back out for the sampler.

Gotchas

• This is the heaviest of the three H3 nodes: it needs a running Hermes agent endpoint, not just an API key. Budget setup time accordingly, and mind timeout_seconds - agentic loops can run long.
• snap_duration_to_h3_grid (on by default) means the length output is what H3 actually produces, not what you asked for.
• quality_mode hero is a bounded but big budget - fine for one important prompt, slow for iteration.

Install

One of ~69 nodes in TrentNodes:

# ComfyUI Manager: search "Trent Nodes"

# or:
cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes.git
cd TrentNodes && pip install -r requirements.txt

The node ships with the pack; the Hermes server is the external dependency you'll need running on loopback. Choose between this and the two prompt nodes by how much agentic review your prompt actually needs - for most shots, the Auto Generator's validate-and-repair loop is the lighter-weight answer, and this is the one for prompts worth deliberating over.