🦊 Kumiho Search Items

Kumiho Search Items is the query layer for your Kumiho Cloud assets. Where Kumiho Load Asset needs a specific item, Search Items returns sets: every matching kref and the local path each one resolves to. If you've ever wanted "give me all the hero* textures in the characters space" as a batch of wires into a processing graph, this is the node. It's also the natural partner for Load Asset in a search → load → process → save loop.

How it works

Under the hood it calls the SDK's item_search() with the filters you set, then does the same BYO-storage resolution as Load Asset: for every match it pulls the latest revision, reads the artifact's location field, and hands you the real local path. Two behaviors are worth knowing before you rely on them:

• Items without a resolvable artifact are skipped. The source is explicit about it - a match whose latest revision has no file path gets dropped so downstream nodes never receive empty values. That's a thoughtful default, but it means the results are "items I can actually use", not "everything that matches."
• It always re-runs. The node's IS_CHANGED returns float("NaN"), the standard "never cache me" trick, so search results stay fresh even if the cloud catalog changed mid-session. The flip side (same footgun the plumbing docs warn about): an always-dirty node disables caching for everything downstream of it, so don't leave a search hanging off a hot loop you care about.

Inputs that matter

• name_filter - substring/wildcard match on item name, e.g. flux* or *portrait*.
• kind_filter - a dropdown of ~24 asset kinds: image, video, audio, checkpoint, lora, vae, controlnet, embedding, diffusion_model, clip, clip_vision, text_encoder, upscaler, ipadapter, mask, and more. Blank = no kind filter.
• context_filter - scopes by project/space path, like */characters/* or a full project prefix. Leave blank to default to the selected project.
• project - [Auto] or an explicit project from the dropdown (only your configured project if you're not authenticated).
• limit - caps results, default 50, max 500. Applied after resolution, so a high-limit query with lots of dead matches can do more work than you'd think.

Outputs

• krefs - list of matching kref URIs (iterative output).
• file_paths - the resolved local paths, same order and length as krefs. Pair the two lists, or wire file_paths into a path-reading loader for a batch pipeline.
• results_json - the full detail dump: every item's name, kind, kref, path, project and space, plus the filters that produced it and a status field (success, error, or pending).
• count - an INT, and here's the trap: it's the number of results with valid file paths, not raw search matches. If the node skipped half your hits for missing artifacts, count tells you the usable half.

Install & gotchas

Pack install, same as the other Kumiho nodes: ComfyUI Manager → search "Kumiho", or git clone https://github.com/kumihoclouds/kumiho-comfyui into custom_nodes and pip install -r requirements.txt. No models. Auth remains the mandatory step (pip install kumiho-cli, then kumiho-cli login).

Without the SDK or without auth, search isn't faking success: results_json gets "status": "error", count is 0, and the console logs SDK not available for search. That's actually a relief compared to some packs that would rather return garbage. If you're searching and getting an empty list, check auth first, then check whether your context_filter is more specific than the assets you expect - a filter like characters/* won't match items living directly under the project root. And keep in mind the lists are only as good as the latest revisions' artifacts, so "found nothing" often means "the newest version of that item has no file attached."