Current node's "Insert LoRA before / after" refused the most common workflows: a
checkpoint whose CLIP feeds both prompts (one link per output only) and UNet workflows
with their own CLIP loader (MODEL and CLIP had to come from one node).
- graph_splice.js: planLoraInsertion decides the links (after a node: every link of its
MODEL output; before one: only its own MODEL input) and, for LoRAs that train the text
encoder, the CLIP line (same node, a UNet workflow's one CLIP source, or before a
sampler the source its prompts share); spliceLora wires it, all or nothing. The old
analyzeModelChainInsertion / spliceModelChainNode / spliceAfterOutputs are gone.
- api/lora_info.py: GET /anomalous/lora_info reads the LoRA's safetensors header; the
picker asks it on confirm and inserts a model-only loader when there are no
text-encoder weights. The MCP add_lora uses the same helper.
- Insertion messages speak of the MODEL line; the fan-out ones are removed.
- tests/graph_splice.mjs (local) covers checkpoint, UNet + CLIP loader, before a
sampler, and the rollback.
Verified on 8199 through the Current node picker: after a checkpoint feeding two prompts
(full LoRA, both prompts moved) and before a sampler (model-only Flux LoRA).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A model-only LoRA loader was used for every UNet workflow, so a LoRA with text-encoder
weights lost them; and inserting after a checkpoint whose CLIP feeds both prompts was
refused (one link per output only).
- api/mcp_actions.py reads the LoRA's safetensors header: lora_te / te1 / te2 /
text_encoder keys mean it trains the text encoder.
- Such a LoRA becomes a LoraLoader taking MODEL from the model line's end and CLIP from
the same node (checkpoint, full LoRA loader) or a UNet workflow's one CLIP source;
others stay model-only.
- graph_splice.js: spliceAfterOutputs puts a node after outputs on different nodes and
moves every link each output had. The picker's spliceModelChainNode is back to its
committed behavior (the `types` option is gone).
Verified on 8199: UNet + CLIP loader with a text-encoder LoRA, a model-only one after it
and another text-encoder one; a checkpoint feeding two prompts; Ctrl+Z after each.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The second batch of MCP tools: they change things, through the same code the UI uses.
- Canvas, done by the open ComfyUI page: describe_canvas, set_prompt, set_model,
add_lora, place_combo, check_workflow_models, fix_workflow_models, run_workflow,
open_in_anomalous. api/mcp_bridge.py sends the action over ComfyUI's websocket and
waits; web/modules/mcp_bridge.js claims it (visible tab first), does it with
Anomalous's canvas code and answers. Each change is one Ctrl+Z step, shows a toast and
is logged as the AI's (activity entries carry via: "mcp", with their undo).
- This server: scan_models (new / look up again / picked models, via /anomalous/scan_all),
list_voices and speak (Anomalous TTS through /prompt and /history).
- graph_splice.js takes the chain's slot types, so a LoRA can go into a model-only (UNet,
Flux) chain as LoraLoaderModelOnly; the default (MODEL and CLIP) is unchanged.
- Nothing deletes or renames files; acting tools are not marked read-only, so AI apps
ask before calling them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
