backend:
- add 429 retry handling with backoff in CivArchive API requests
- add page-scraping fallback for CivArchive model payloads
- normalize CivArchive mirrors with dead/gated/paid/sha256 metadata
- include file metadata and mirror data in normalized archive files
- propagate additional dead/missing status field aliases
- include metadata in normalized CivitAI files
frontend:
- render mirror lists with source grouping, dead-link indicators, and skull icons
- add file badges (primary, type, format, size, fp, mirror count, id)
- improve secondary info in resolve/download methods to include base model
- enhance search panel secondary info with base model and filename
- on selecting a version with empty files, fetch fresh version details
- propagate base model context to CivitAI search state and display labels
Adds a fullscreen image preview dialog to the model info panel with
navigation, copy-prompt support, generation metadata display, and
referenced resources listing.
Extends CivitAI backend extraction to handle varied metadata key names
(negativePrompt, Clip skip, Model, etc.) and surfaces clip_skip, width,
height, tags, resources, and raw metadata through the new preview UI.
Renames z-index variable to reserve a dedicated layer for the preview
backdrop.
Adds a new 'Auto-fill model filter' setting under Search Defaults that
controls whether new searches default to 'Auto' (detecting the workflow
base model) or 'none' (Any model). The preference is persisted in
localStorage as ModelResolver.autoFillBaseModel and is respected across
search panel initialization, base model dropdown defaults, workflow state
cloning, and dirty-state checks.
- Add resolveBaseModelAlias with exact-then-fuzzy matching
- Use weighted scoring in getDominantWorkflowBaseModel to prefer
checkpoints/diffusion/unet (weight 8) over loras (weight 2)
- Skip URN-prefixed values during model reference extraction
- Refresh base model select labels after download panel re-render
- Update HuggingFace author index metadata and repo list
Replace the heavy autoResolveAfterDownload path (which re-analyzed the
entire workflow) with a more efficient refreshLocalMatchesForDownloadedMissing
call that queries /model_resolver/local-matches directly for the
specific downloaded file.
This removes the unnecessary re-analysis step, adds an explicit
force_rescan flag (coerced to bool on the backend for safe JSON
payload handling), propagates the missing model's category through
the activeDownloads tracking objects, and updates related UI text
to reflect the new narrower behavior.
Store references to queue title, toggle, and clear buttons as instance
properties in the constructor, replacing repeated DOM queries inside
updateQueueUI with direct reference lookups.
Raise CivArchiveSearchError in civarchive backend, catch and store it in
source_results in the workflow analyzer, surface it in resolve_download_methods
as an 'error' status, and display error messages in search_panel progress
items and tooltips.
Adds a checkbox in the resolver footer search menu that bypasses all
search result caches for the duration of a batch search run.
Frontend (queue_methods.js, resolve_download_methods.js,
resolver_dialog.js, search_panel.js, resolver-main.css):
- Introduce checkbox-type footer menu items (type: 'checkbox').
- Add 'Force search (ignore cache)' checkbox to the search menu,
wired to a new forceBatchSearch flag on ResolverManagerDialog.
- Pass forceSearch through searchMissingBatch and searchOnline to
the backend API call.
- When the manual search button is clicked and results already exist
for that model, forceSearch defaults to true.
- mergeSearchResults respects forceRefresh: if a forced search
returns no results for a source, that source is cleared from the
cached results.
Backend (__init__.py, core/sources/huggingface.py):
- force_search is read from the incoming API request and, when true,
clears all active source search caches (local, huggingface,
civitai, civarchive, lora_manager_archive) before searching.
- force_search is forwarded as force_refresh to
search_huggingface_for_file, which folds it into the cache key so
the cache is not hit, and propagates it to _get_author_index.
- Search result timestamps are no longer inherited from prior
cached signatures when force_search is active; a fresh timestamp
is written unconditionally.
Updated metadata/huggingface-author-index.json reflects the latest
cached author index after rebuild.
Add searched_at timestamps to search results so cached entries are accurately tracked and preserved across merges. Expand the clear-search-cache command to also invalidate popular databases, model list, and the new result timestamp store. Frontend gains new cache-clearing methods and a full resolver cache reset for maintenance. The search panel and missing browser now show source result dates and toggle the search button between 'Search' and 'Search Again' when results already exist.
Group Queue Exact, Apply Selected, and Clear Selected into a single
'Link' dropdown menu in the missing browser footer. Move Auto-Link 100%
into the same menu and show/hide it when exact local matches exist.
Add disabledWhen support to createFooterMenu items so buttons can be
visually and functionally disabled. Include disabled state styles for
.mr-footer-menu-item.
Simplify footer visibility logic and update references throughout.
- Add isLikelyWorkflowTabClickTarget to distinguish workflow tab clicks
from generic document clicks, preventing false 'workflow-change'
triggers and unnecessary refresh cycles.
- Add getWorkflowSignatureData to produce a stable, normalized signature
for workflow data (sorted nodes/links, top-level signature included).
- Update document-click handler to use a new 'workflow-tab-click' reason
via the click-target heuristic.
- Simplify refreshForActiveWorkflowChange retry logic by removing the
pending-graph-switch special case, since click origin is now explicit.
- Patch app.loadGraphData to dispatch custom event after loading,
ensuring workflow handlers fire when switching active workflow
- Add document-click handler to catch workflow changes initiated by clicks
- Pass event options through removeEventListener for proper cleanup
- Detect possible pending graph switch on document-click and reduce
max refresh attempts from 8 to 3 to avoid stale graph timers
The previous commit introduced server-side persistence for API keys and
preferences via a model_resolver_settings.json file and POST/GET routes.
This change reverts that behavior, returning to localStorage-only storage
for settings, which is simpler and avoids persisting sensitive credentials
to disk. Also removes the now-unnecessary .gitignore entry.
Replace all link dialog terminology with resolver terminology
throughout the codebase, including module names, CSS classes,
IDs, JS variables, SVG gradient IDs, and log module names.
Remove the old web/linker/ tree and introduce the new web/resolver/
tree with parallel structure. Also replace core/linker.py with
core/resolver.py.
Adds ARIA labels, tooltips, and aria-disabled attributes to all footer
buttons (Apply, Auto-link, Download All). Enforces a visible disabled
state via the .ml-btn-is-disabled class instead of relying on the
native disabled attribute, which is compatible with the custom dialog.
Add backend endpoints (/model_linker/model-list/status, /model_linker/model-list/update)
that let users check whether their bundled ComfyUI-Manager model-list.json is current
and download the latest version from GitHub, with SHA tracking and cache invalidation.
Add corresponding UI in the Options dialog under a new 'Local Database' section
with status display, remote check, and update controls.
Dariusz L