UI→API conversion emits frontend-only custom nodes (rgthree Bookmark) → unknown_class_type; consider pairing widgets by name via widgets_values_named

[sickbraintwo]

convert_ui_to_api (comfy_cli/workflow_to_api.py) skips only the frontend's own virtual nodes — _UI_ONLY_NODE_TYPES = {Note, MarkdownNote, PrimitiveNode, GetNode, SetNode, Reroute}. Popular packs ship more of them: rgthree's Bookmark (rgthree), Label (rgthree), Fast Groups Muter (rgthree)… are isVirtualNode in the frontend, have no Python class and never reach the executor. The converter still emits them, so comfy workflow validate / comfy workflow run (and comfy-mcp's validate_workflow / run_workflow, which call the same code) fail with unknown_class_type on a workflow that runs fine from the UI.

Repro: any UI-format workflow with one Bookmark (rgthree) node (rgthree-comfy installed):

comfy workflow validate --workflow my_ui_workflow.json
→ valid: false, errors: [{"type": "unknown_class_type", "node_id": "160", "class_type": "Bookmark (rgthree)"}]

Measured on comfy-cli 1.18.0 and on main (2026-09-15) with a 149-node workflow carrying 5 bookmarks: 5 × unknown_class_type, nothing else wrong with the graph (the same graph exported to API from the UI validates clean).

Two suggestions

1. Skip a node whose type is absent from object_info when it declares neither inputs nor outputs (decorative by construction), and/or extend _UI_ONLY_NODE_TYPES (and workflow_ops.UI_ONLY_NODE_TYPES) with rgthree's virtual types. We run exactly that as a 20-line wrapper loaded by a .pth in our venv: https://github.com/sickbraintwo/WextraUI/blob/main/tools/comfy_cli_patch.py — with it the same workflow converts to 60 clean nodes.

2. The frontend (≥ 1.5x) now saves widgets_values_named next to widgets_values (a {input name: value} map, present on every node of files saved by recent frontends). When it is there, pairing widget values by name instead of by position would make the converter — and comfy workflow slots — immune to the positional drift that produces pairing_suspect slots whenever a custom node draws a widget its schema does not list, or changed its input order between versions. We hit this on our own pack (40 suspect slots on one workflow) and fixed it on our side, but the name map is already in the file and would make the converter robust for every pack.

Happy to open a PR for either if you tell me which shape you prefer.

[sickbraintwo]

Two more gaps in the same converter, found by actually running a UI-format workflow through comfy workflow run / comfy-mcp run_workflow (the first report only validated it). Both measured on comfy-cli 1.18.0 against the frontend's own "Save (API)" export of the same graph.

1. A link that leaves a SetNode's pass-through output is lost. KJNodes' SetNode has an output that forwards its input. _Tracers.trace_get_set hops GetNode → SetNode → source, but when a consumer is wired directly to the SetNode's output, the SetNode itself is the source and nothing hops it: the consumer ends up with no input at all (here a StringConcatenate failed with a missing required input). The frontend's export resolves it to the node feeding the SetNode.

2. A bypassed node with no input of a compatible type forwards the wrong cable. trace_bypassed falls back to "the first linked input regardless of type". With a bypassed node whose output is STRING and whose first linked input is MODEL, the API graph wires MODEL into a STRING input; the executor rejects it (received_type mismatch), drops every output depending on it and the run ends success in 0.2 s with zero images — silently. The frontend leaves that output unplugged and the consumer keeps its widget default. validate shows these as edge_type_mismatch warnings, which is easy to read as cosmetic.

Both are fixed in the same local wrapper as before, by wrapping the two tracers (type check through _is_valid_connection, loop guard on SetNode chains): https://github.com/sickbraintwo/WextraUI/blob/main/tools/comfy_cli_patch.py (second try block). With it the UI file converts to the same graph as the frontend export and runs. Happy to turn either into a PR.