"Pick models to scan" sits beside the scan button; the picks are named
under it and the button scans them, "clear" goes back to everything, so
the all/picked switch is gone. Under the button the page says what this
scan will do, each line with how many of its models are components.
The settings card (ui_scan_settings.js, holding the scan options and the
Civitai key prompt) unfolds in place instead of opening its own page.
The summary counts and lists models in component folders (VAE, text
encoders, CLIP, CLIP Vision, upscalers) apart. The counts at the top and
the not found / not scanned lists leave them out; a card and a list of
their own say that many are not on Civitai and that this is no fault.
The not found list looks up again just the models it shows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
"All models" now lists, from the scan summary, how many models a scan
looks up, keeps, leaves for "look up again" and leaves out, following
the offline and fetch-again switches. "Only ones I pick" names the picks
with their size and says when found ones would only be read again.
The picker (ui_scan_picker.js, GET /anomalous/scan_pickable) lists the
.safetensors files a scan reads with the scan page's states, folder,
base model and size, as a list or covers, narrowed by folder, state,
search and picked-only; "pick these" picks what is shown, shift-click a
run. Picks go to the scan as targets. The old inline-styled selector and
its batch_select route are gone; setWidgetValuePath keeps its own file.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The scan summary marks new models whose safetensors header reads but names no drawing model
kind (a voice model such as F5-TTS) as not_drawing; the scan page asks: don't scan them,
scan them, or see each. The choice lives in <model>.anomalous.json ("scan": "skip" / "scan",
model_identity.scan_choice), saved by POST /anomalous/scan_choice (api/scan_choice.py, this
computer only, at most 2000 models, only readable .safetensors layers).
- "skip": a scan of a whole folder leaves the model out (picked by hand it is still scanned),
the summary counts it apart (excluded) with its own list to change it back, the card and the
details say "不扫描"; the details have the same button.
- The unmatched count warns only for damaged files or lookups still to do; not-found rows say
the fingerprint is kept for repairing workflows.
- model_kind.inspect_file now gives format None when the header cannot be read, as its text said.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A model Civitai has no record of now has its safetensors header checked (model_identity.
safetensors_damage: too short, header longer than the file, not JSON, or a tensor ending past
the file's end; only the header is read). Such files get the unmatched reason "damaged": a red
badge on the card, a line in the details, a warning on the scan page (the summary counts them)
and first place in the unmatched list, with advice to download the model again.
Tests: header cases, a scan telling a damaged file from a whole one; the scan report fixtures
are whole safetensors files now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The picker, combo-save and activity fixes now ship in v1.58.0, so their entries
move from v1.59.0's Fixes to v1.58.0's (the code here already had them; the
picker keeps this branch's theme-aware border colour). Outdated docs/plans and
docs/audits are removed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/plans held task briefs written for other agents, most of them done or
abandoned (the toolbox shortcuts, the Material Library node capture), and
docs/audits a 2026-09-19 report on the old Model Doctor and Material Library.
Read publicly they looked like current plans. GEMINI.md now points front-end work
to AGENTS.md section 4 and docs/architecture/frontend.md; the audits row and the
audits ignore rules are gone. History keeps every file.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A search box above the clips (ui_tts_clip_search.js) finds clips by part of their
line or file name and lists the first 30 with their lines, to hear, use as the main
voice or add as an emotion (the new row gets the clip and its line). The lines come
once per dialog from Anomalous_TTS's GET /anomalous_tts/reference_lines (interface
14, by character name only); with an older node the request fails quietly and the
search covers file names. Picking a clip now fills its line from the same lines,
so the per-clip request is gone. Lines reach the page as text only (textContent).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With Anomalous_TTS interface 14 (GET /anomalous_tts/reference_text), picking a clip
that is not yet in use fills in the line the node finds for it (.txt / .lab /
annotation list); an older node gives no answer and the box stays blank as before.
The answer is used only while the box still holds that clip and is still blank, and
is dropped with the dialog. A blank line box shows the line the node uses for its
clip as greyed placeholder, so the settings are not changed by opening the dialog.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CHANGELOG heading and README badge drop "Beta". The recipes architecture note and
AD-013 said recipes stay beta until a stability decision; that decision is this
release.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PUT /anomalous/import/upload and POST /anomalous/placement/move wrote
into the model folders for any caller, while place and identify already
answered only the computer ComfyUI runs on. With ComfyUI started with
--listen, another device could upload model files or move models
between folders. Both now check mcp_server._is_local first and answer
403 not_local, which the import and Tidy screens already show.
Tests drive both routes through a proxy header and check nothing was
written or moved; they fail without the check. Full suite: structure
check, every tests/*.mjs and 234 Python tests pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018P2T2R7TFbJsc3C9maCD8B
Brings stage 1 into stage 2: the gallery drag fix and the change that
follows Anomalous_TTS interface 13 (no server path picker, no storage
change or move, folders shown with the node's settings file).
Conflicts: the picker CSS, the storage-change activity label and its
locale keys are dropped; stage 2's model download and backup entries stay.
tests/tts_setup_api.mjs now checks the pretrained download route instead
of the removed storage and library routes, and that no export sends a
path on the computer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018P2T2R7TFbJsc3C9maCD8B
Anomalous_TTS no longer takes a folder or file path on this computer from
a request (ComfyUI-Manager review of the node): the storage place, other
character folders and pretrained sources are set only in its settings
file, and imported files are always uploaded.
- GPT-SoVITS settings shows the storage place, other places and package
sources without change, move or remove buttons, and names the node's
settings file (status.settings_file) with a copy button and how to edit it.
- The import window drops "from this computer": no server path picker
(ui_tts_path_picker.js is gone), no local folder scan or local clip
preview; a very large upload is confirmed instead.
- tts_setup_api.js loses the removed routes; formatSize moves there.
- The storage-change activity entry, unused locale keys and picker CSS go.
Verified: tools/check_structure.mjs; tts_setup_api in Node with fetch
stubbed; the settings card, batch hero, add menu, file sources and player
rendered with jsdom and stubbed ComfyUI scripts. Not run in ComfyUI.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018P2T2R7TFbJsc3C9maCD8B
The models page's Tidy chip (with a count) reads every model's header
(model_kind) and lists files whose type does not match their folder - such
as an Anima/Flux/Wan diffusion model in checkpoints, which Civitai labels
"Checkpoint" - with why it matters, where it belongs (same subfolder, same
drive) and Move there, taking its cover, info and notes along; never over a
file. A whole checkpoint in diffusion models is marked as loading anyway.
Identical files are grouped by SHA-256 (scan hashes; same-size unscanned
files on Compare); each copy can go to the Recycle Bin. Moves are logged
(model_move). Also: opening the Tidy or Sources view stops cards of an
earlier load from being drawn into it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Models page: "+ Import" chip and model files dropped on the grid open an import
window. model_kind.py tells a file's type (LoRA, checkpoint, VAE, text encoder,
diffusion model, ControlNet, embedding, CLIP Vision, upscaler) and base model from
its safetensors / GGUF header only; .ckpt/.pt are never opened. api/model_import.py
finds the dropped file in Downloads/Desktop (name, size, time) and moves it, or
takes an upload; checks it against Civitai and the scanned models; never replaces
a file (" (2)"). New files are scanned after.
- Model cards show the Civitai version name (metadata version_name).
- Settings page in tabs (Look, Models, Workflows, Prompts, Backup, Help); the
download folder is a choice of ready-made folders or typed by hand; default is now
by base model ({base}).
- The scanner's offline base model uses model_kind (LoRAs, Qwen, Wan, Chroma...).
- AD-019: no Civitai browser inside the plugin.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Save as combo now keeps all picked nodes but subgraphs, with the links
between them. Model drop-downs and text boxes stay the slots; nodes
without slots are listed in the editor under "Nodes put back as saved"
with their values and go back exactly as saved. Inputs fed by links from
unpicked nodes are stored in structure.open, and the save dialog, the
editor and the card say which ones to wire after putting it down. Cards
of combos without models name their nodes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Settings -> Appearance -> Theme (and the ComfyUI setting
Anomalous.ModelBrowser.Theme, which replaces the Abyssal Scarlet switch)
picks Follow ComfyUI (default), Dark, Light or Abyssal Scarlet. Following
ComfyUI watches the dark-theme class ComfyUI sets on <html> for a dark
palette and repaints at once; an older Abyssal switch reads as Abyssal.
The light theme is html.amb-theme-light with its own --amb-* values. Hard
coded colors were turned into tokens whose dark values are the colors they
replace (checked: every stylesheet resolves to the same dark values):
rgba(var(--amb-ink-rgb), a) for tints, status shades, tooltip, float,
dialog, sunken, scrim and note tokens. Older rules and inline styles that
wrote a dark color now write var(--amb-lt-<role>, <that color>); the
--amb-lt-* tokens exist only in the light theme. Inside the plugin's
windows ComfyUI's palette variables (--bg-color, --fg-color ...) follow the
plugin's theme when the two differ. Surfaces over images stay dark, and the
image viewer gets the dark tokens back. The ready banner on a recipe reads
green instead of pink. Stylesheet cache versions are bumped.
Checked on a test ComfyUI in light and dark: home, models, gallery and its
viewer, workflows and a recipe's tabs, combos and a combo, prompts and the
docked drawer, voices, audio library, settings, scan, Model Check, current
node and its model picker, a model's detail with notes, share, save
recipe, update guide, versions, GPT-SoVITS settings and the prompt dialog;
light with a dark ComfyUI, and Abyssal Scarlet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The rail's Prompts entry used to fold the browser away and open Prompt
Studio in a drawer by the canvas, the one page that left the window. It is
now the browser's Prompts page (owner.promptPanel): the rail, the 7 key,
opening the browser with a text node selected, Current node's "Edit in
Prompt Studio" and a kept prompt all go there, and reopening the browser
returns to it. "Beside the canvas" on the page moves the studio into the
drawer as before (side and width remembered, browser folded away); the
drawer has "Back to the window" and Close.
Only one studio is shown at a time. The page's studio stops when another
page is shown, a workspace opens or the browser closes (its 400 ms poll
with it) and is made again when the page is shown; closing a workspace
opened from the Prompts page returns to it.
Checked live on a test ComfyUI: a CLIP Text Encode opens the page aimed at
it; docking, Back to the window and Close; 3 then 7; Workflows opened and
closed from the page; reopening the browser.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The "Open Prompt Studio shortcut" (Ctrl+Shift+L by default) goes: its row
in ComfyUI's settings, its command and Extensions menu entry, its default
keybinding and the keydown fallback that backed it. The command goes too,
so a binding a user recorded for it no longer opens anything. Prompt
Studio opens from the rail, with 7 while the browser is open, or by
opening the browser with a text node selected. The browser's own shortcut
(Ctrl+Shift+M) stays. Unused key-matching helpers are removed with it.
Checked on a test ComfyUI: the settings page lists only the browser's
shortcut, the command list has Open and Reset position, Ctrl+Shift+L does
nothing, and no console errors.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Opening the browser (the trigger, the action-bar button, the menu or its
shortcut) now follows the node selected on the canvas: a node with prompt
boxes and no model drop-down opens Prompt Studio, which targets that node;
any other node (a KSampler, a model loader) opens Current node; nothing
selected reopens the page used last. Settings -> Opening and window can turn
it off. Prompt Studio closing, the doctor banner and MCP actions still open
the browser as it was.
While the browser is open, the plain keys 1-9 go to the rail's pages from
the top. They are ignored in text fields, while an input method is
composing and while a dialog is open; ComfyUI binds no plain digits (W is
its Workflows key, so W/S were not used). The rail's tooltips name each
page with its number.
The rules are in shell_open_rules.js with tests/shell_open_rules.mjs.
Checked live on a test ComfyUI: a CLIP Text Encode opens Prompt Studio on
it, a KSampler opens Current node, 1/3/4/5/7/8 switch pages, digits typed
into the search box stay there, and the setting turns the rule off.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Settings -> Translation -> DeepL key saves the user's own DeepL API key;
translation then asks DeepL first, else Google's free endpoint, else
MyMemory. Saving a key translates one sentence and says whether it works or
why not (a refused key, a used-up month, busy). Only the key's tag changes,
so the page keeps its scroll place. The key stays on this computer: backups
leave it out like the Civitai key.
DeepL stopped taking the key as an auth_key parameter in 2025, so a key put
in config.json by hand no longer worked. It now goes in the Authorization
header with a JSON body; targets are ZH-HANS and EN-US. When DeepL fails,
the answer carries deepl_error beside the other provider's translation.
Checked live: a made-up key is refused by DeepL's server and the dialog
says so; removing it switches back to Google.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
By the user's decision, recipe packages are open again: the card's
export button and the detail's More menu make a .zip named after the
recipe (workflow, notes, tags, parameters, cover, model thumbnails;
versions only when ticked). Dropping one into the Workflows page's
share and import dialog inspects it and keeps it as a new recipe card,
optionally opened on the canvas.
An import never replaces a recipe any more: the "replace" path, which
erased the old recipe for good, is gone; a taken name gets " (2)".
Assets and versions move in before the recipe file, and a failure
removes only this import's new folders. The release gates and their
503 answers are removed. Tests: new card and renaming, nothing left
after a failed import, tampered and escaping packages refused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
share_code.js: a workflow's lean form (node id, type, widget values,
position to 10 px, only what differs from a fresh node; node-pack nodes
keep their ports), links by slot, groups and the anomalous_* extras,
deflated and spelled in Chinese characters (14 bits each) or base64url.
Encoding rebuilds and compares; subgraphs, native reroutes or any
difference share the whole workflow. AMB0/AMB1 still decode. About 4x
shorter than AMB0 on ComfyUI's default workflow (657 vs 2521), 13x on
an Easy-Use/rgthree graph.
ui_share.js replaces the share code dialogs that lived in main.js: the
code is made on opening with both copy buttons and the fingerprint
notes, and pasting a code or JSON, or dropping/choosing a workflow file,
an image or a backup .zip, imports it. tests/share_code.mjs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A backup now keeps each model's scan file (.info, small; on by default)
so names, trigger words and base models survive Civitai removing a
model, and on request Civitai's covers (stored once as the civitai_bak
copy; they can be hundreds of MB). Putting back writes a scan file only
where the model has none (replaced only with "replace") and Civitai's
cover only where it has no cover, with its civitai_bak copy so it stays
Civitai's and a later user cover still wins.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A backup also keeps a model list (each model's folder, path, size,
SHA-256 and link). Inspecting it names the models this computer lacks
(two copies of one file are one download); "Find downloads..." looks
them up like Model Check (Civitai by hash, the saved link, Hugging
Face), shows size, source and free space, and downloads the ticked ones
one after another into the path they had, each checked against its hash
and scanned after. A file found by name only starts unticked. Verified
on the isolated ComfyUI: an 18 MB Civitai LoRA came back to its folder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Export packs the stores (recipes, combos, materials, parameters, and
optionally ComfyUI's saved workflows), the plugin settings without the
Civitai key, and each model's user layer and own cover into one .zip to
keep on a cloud drive. Import lists what is new, the same or different
and the models it finds again (by SHA-256, else folder, path and size),
adds only what is missing by default, replaces differing files only on
request with this computer's going to the Recycle Bin first, merges a
model's name and notes field by field and puts the user's cover in
place of a Civitai one. Unsafe or foreign zips are refused. Both are
activity entries. tests/test_backup.py covers the rules; testing.md
names a single test file the way unittest finds it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The recurring traps (a :has() rule slowing every canvas frame, fixes
left on an unpushed local main, release checks, imports lost in a module
split) are now docs/architecture/lessons.md, so the links from
ARCHITECTURE.md, AGENTS.md and the docs indexes open for everyone.
AGENTS.md: tests live in tests/, .agents stays a local folder for
screenshots and logs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tests/ (Python backend tests, .mjs frontend rule tests) is no longer
ignored. tools/run_tests.mjs runs the structure check, every
tests/*.mjs and the Python tests with $PYTHON, the portable
python_embeded or python. docs/guides/testing.md says how to run and
write them; README links it. Developer notes no longer point at 18 test
files that were removed earlier (AGENTS.md table, backend, frontend and
update-guide notes).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Seven short "see it in action" steps in English and Chinese, each with a
still or a few-second animation (WebP, about 500 KB in all) from the
isolated demo ComfyUI: scan results, Model Check putting models back,
gallery search and drag to canvas, a combo dragged onto the canvas,
Prompt Studio tags, the activity log with Undo, character voices; the
model cover wall under the title. .gitignore keeps docs/images/*.webp.
The Activity row no longer says it only records: canvas changes have Undo.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each slot names its node (and node pack when not ComfyUI's own). A small
"译" button beside a prompt shows its translation in a read-only panel
(ui_translation_peek.js). On nodes with several models, saved values go
to the row with the same number (strength_02 under lora_02) and more
than four fold behind "+N"; empty optional rows (None) start unticked in
the save dialog so a ten-row LoRA stack does not make ten empty slots.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Models, LoRAs and prompts in their own groups, in the order they flow
(LoRA 1 next to the model, positive before negative); made-up names
renumbered along the chain. Covers looked up for every model, a whole
row opens the picker, names wrap to two lines, each model node's saved
values (strength, CLIP strength, type) shown read-only, prompt boxes
grow with their text. New captures keep nodes in flow order. The combo
editor's sticky top bar sat 20 px down with the page showing above it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The canvas's node menu (the frontend's getNodeMenuItems hook) offers
"Save as combo" for the picked nodes, or the right-clicked node when it
is not among them, when any of them is a model or text node; it opens
the same save dialog without opening the browser first, and the browser
shows the new combo afterwards.
Values other than the slots stay as they were saved (agreed): a combo
always puts down exactly what was saved; other values mean another combo.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A combo can now hold its own node structure: pick nodes on the canvas and
New -> From the nodes picked on the canvas keeps the model nodes (a
drop-down of model files) and the text nodes (a multiline text box by the
node's definition) with the links between them; other picked nodes and
links to them are left out, their ports stay open on the canvas.
Slots are per box, so a node with several model drop-downs has several; a
box fed by a link is no slot. Each model slot's folder is the one whose
files hold all its drop-down's model-file options (ComfyUI's /models
lists), never guessed from names. Every other box is stored by name and
put back untouched, in order (a count box adds its rows first).
The save dialog shows what is kept and dropped and lets slots be renamed
or unticked. The editor lists the slots: a picker of that folder's models
with covers, filtered by base model, and text boxes. Put on canvas makes
the nodes (missing node types and their packs are named before anything
is placed), fills the slots by name, links by port name and type, as one
Ctrl+Z step; a box or link that cannot be restored is reported, never
guessed. New -> Same structure as ... starts another combo from a saved
structure. The default combo is unchanged. MCP list_combos lists node
combos by slot and place_combo places them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Downloads: with the China mirror on, a Civitai match whose very file (same
SHA-256) is on Hugging Face too, by the workflow's link or ComfyUI-Manager's
list, is fetched from Hugging Face; Civitai's name, base model and page stay.
- A file downloaded from Hugging Face gets its model card's first example
image as cover and its base model, trigger words and example prompts as
notes, with the repository as link (api/hf_card.py), never over what the
user or a scan set.
- Text encoders, CLIP vision and upscalers are shown (and scanned) by
default, so shared workflows carry their fingerprints; older configs get
them turned on once (folder_defaults_version 2).
- The models page's search sits in the header (ui_model_search.js) instead
of inside the folder list; it searches the whole type by name, file name,
folder, trigger words, base model and notes, every word; Esc clears it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Models not on Civitai (Flux's VAE, T5 / CLIP text encoders, upscalers,
ControlNets...) are looked up by file name in the model-list.json of the
ComfyUI-Manager installed here (api/manager_catalog.py), keeping entries
that fit the node's folder type. With the workflow's fingerprint only the
listed file with that SHA-256 counts (same-name entries are each asked);
without one, a single entry is offered as "Download by name" with a
warning, never in Download all or the MCP's download_missing_models.
A site that stops answering is skipped for the rest of a lookup while the
others are still asked, so Civitai failing in mainland China no longer
stops Hugging Face through its mirror. GitHub raw file links count as
sources; a Hugging Face file behind a login offers its repository page.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Three MCP tools: download_missing_models (Model Check's Download all for
the open workflow, through the page so the placement rules and the put-in
afterwards are the UI's own), download_model (one model from a Civitai
version link, a Hugging Face / GitHub file link or a Civitai hash, into a
folder of its type) and download_status. The default models folder of a
type is shared by the dialog and the tools.
A Civitai version link without a matching file name now takes the
version's primary model file (people rename their files); a hash, when
there is one, still decides.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Settings > Model downloads can send Hugging Face downloads through
hf-mirror.com, on by default for the Chinese interface. The lookup and the
download use the chosen host whichever one the link names; when the mirror
sends a network outside China back to huggingface.co, the lookup follows
once for the size and SHA-256. A failed download from Hugging Face itself
suggests the mirror.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Model Check offers Download (and Download all) for a missing model whose
source names one exact file: Civitai by the workflow's fingerprint, or a
Civitai version / Hugging Face / GitHub file link the workflow carries.
A dialog shows each file's size, source and full destination, with the
models folder (when there are several) and subfolder changeable; Settings
> Model downloads sets the default: where the workflow says (the node
needs no change) or a set folder, {base} sorting by base model and
matching an existing folder in any spelling.
Downloads run one at a time into <file>.part, continue after an
interruption, are checked against the SHA-256 before they get their name,
never replace a file, and send the Civitai key to Civitai only. Finished
files are put into their nodes (one Ctrl+Z step each), scanned for info
and covers, and recorded in the activity log.
A change of fingerprints alone (after any scan) no longer takes a Ctrl+Z
step that seemed to do nothing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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>
An MCP endpoint (Streamable HTTP) inside ComfyUI's server, so Claude, Cursor, Cherry
Studio and other MCP apps can ask about the user's library without anything extra
installed and without memory or VRAM while nobody is connected.
- api/mcp_server.py: one JSON-RPC message per POST, answered with JSON. Serves the
2026-07-28 stateless protocol (headers must match the body) and the earlier
initialize-based versions; no sessions. This computer only: loopback peer, no proxy
headers, loopback Host and Origin (DNS rebinding).
- api/mcp_tools.py: 11 read-only tools: library overview, model search and details
(optional cover), scan status, output-image search and generation settings
(optional thumbnail), combos, workflow recipes, saved prompts, generated audio,
activity log. Ids resolve inside their roots; tool failures come back as isError.
- docs/guides/mcp.md: setup for each app, what can be asked, privacy note.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The bar starts with "Model Check:" so it is clear it comes from the browser.
- "Don't show again" turns the check on opening off and says where to turn it
back on; Settings → Workflows has the same switch (on by default), and turning
it on checks the open workflow right away. While off, the bar never shows.
- The tour's Model Check step mentions the button.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- While a combo group follows the pointer, Esc removes it again; the activity log
holds back until it is put down, so a cancelled group leaves no entry.
- The positive and negative prompt nodes are always laid out (positive empty when
the combo has no prompt); the negative takes Prompt Studio's common negative card
(starter cards moved to prompt_composition.js as PROMPT_PRESETS).
- The UNet "connect clip yourself" message shows once the group is down; it used to
pop up while the group followed the pointer, and closing it put the group down.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
"Use" tried to put a combo into the open workflow (main loader, LoRA
chain rebuilt, positive prompt) with a list of what changed and what was
skipped and why, which read as guesswork. A combo is a shortcut: its one
button, Put on canvas (放到画布), and dragging its card both lay out the
main model loader, LoRAs and prompt nodes as a new wired group, and never
change nodes already on the canvas. An existing node's values are changed
in Current node.
notebook_apply.js, the Use confirmation and its strings are removed; the
editor's More menu keeps only Delete.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
