Introduce an agent skill framework for LLM-driven metadata enrichment:
- AgentCLI (py/agent_cli/): in-process wrappers around internal services
using standard relative imports, eliminating the need for sys.path hacks
- LLMService: centralized BYOK (bring-your-own-key) LLM client supporting
OpenAI, Ollama, and custom OpenAI-compatible endpoints
- PostProcessor: deterministic engine that applies LLM output via AgentCLI
(replaces old handler.py + _BASE_MODEL_ALIASES approach)
- SkillRegistry: filesystem-based skill discovery (skill.yaml + prompt.md)
- AgentService: orchestrates skill execution with WebSocket progress
- Frontend AgentManager: WebSocket listeners, skill execution, config UI
- Context menu entries (single + bulk) for "Enrich Metadata (Agent)"
- Settings UI for AI Provider configuration (BYOK)
- Full i18n support across 9 locales
Bug fixes found during review:
- aiohttp.web.json_response: status_code= -> status=
- settings_modal cancelEditApiKey: wrong argument position
- AgentManager.isLlmConfigured: allow Ollama without API key
- PostProcessor._merge_tags: lowercase all tags to match TagUpdateService
Replace native <select> with a searchable dropdown that:
- Filters options as the user types
- Shows filename-inferred suggestions at the top in a "Suggested" section
- Supports keyboard navigation (ArrowUp/Down/Enter/Escape)
- Allows typing custom values not in the list
- Removes dead .base-model-selector CSS
Adds 3 new i18n keys (baseModelSearchPlaceholder, baseModelSuggested,
baseModelNoMatch) with translations for all 9 locales.
Security hardening:
- Validate repo format with strict regex (reject .. traversal)
- Validate filename rejects path separators and ..
- Validate relative_path rejects absolute paths and ..
- Verify model_root is within configured scanner roots using
realpath + os.sep guard to prevent prefix-match bypass
- Add realpath-based escape detection for final dest_path
Bug fixes:
- Fix WebSocket leak in _downloadHfSingle: wrap ws.close() in
try/finally so it closes even if downloadHfModel() throws
- Same fix for batch HF download per-file WebSocket loop
Frontend hardening:
- Tighten HF repo regex: require huggingface.co for full URLs,
reject bare .. patterns
- Add 12 unit tests for detectUrlType() covering HF resolve,
HF repo, CivitAI, CivArchive, direct HTTP, edge cases
- Unify single-URL and multi-URL HF repo flows to use the same batch
preview interface (remove separate repoFileStep)
- Remove unnecessary cloud icon from HF batch preview items
- Use formatFileSize() instead of hardcoded MB text
- Change default selection to unchecked (no preselected files)
- Add select all / deselect all checkbox with dynamic Next button
- Clean up dead CSS, HTML template, and JS methods from removed
repoFileStep
- Add selectAll i18n key with translations for all 10 locales
- Fix batch progress bar name fallback for HF items
A model not being found on CivArchive by hash is a routine case (the
model simply isn't published there), not an error. The callers already
log the outcome at WARNING (bulk_metadata_refresh) or DEBUG
(metadata_sync_service) with full context, making this ERROR-level log
both misleading and redundant.
Cache corruption (NULL model_name/file_name from legacy DB rows or partial
writes) caused format_response to raise KeyError/AttributeError, failing the
entire /loras/list request and showing no models in the UI.
Fix across three layers:
- format_response (lora/checkpoint/embedding): replace direct dict[] access
with .get() fallbacks; return None for entries missing file_path
- handlers: filter None entries from list/excluded/fetch/duplicate/conflict
endpoints instead of letting them crash or appear as null in responses
- model_scanner: always use validate_batch repaired copies (previously
discarded when no invalid entries, leaving None values in raw_data)
- persistent_model_cache: add or-empty-string guards on read and write for
nullable TEXT columns (model_name, file_name, folder, base_model, etc.)
git clean -fd in _perform_git_update deleted untracked, non-ignored
directories (wildcards, stats, backups, civitai, caches, logs) during
portable-mode updates, since released tags do not list them in .gitignore.
Add -e excludes for all user-managed paths to both nightly and stable
update branches. Add regression tests for both paths.
Move NodeRegistry from a single global _nodes dict to a per-client
(_tab_nodes) structure so that multiple ComfyUI browser tabs no
longer overwrite each other's workflow node data during a
lora_registry_refresh cycle. The merged result is a union of all
known tabs' target nodes, eliminating the non-deterministic failure
where send-to-workflow could randomly target a tab lacking valid
targets.
- NodeRegistry.register_nodes(sid, nodes) replaces per-tab data
without affecting other tabs.
- NodeRegistry.get_merged_registry() returns the union across all
connected clients, together with tab_count / per-tab metadata.
- prepare_for_refresh() snapshots the current active sockets; caller
re-reads before merging so that newly-connected tabs are not pruned.
- workflow_registry.js sends api.clientId in the POST body so the
backend can identify which tab is registering.
- Add &withMeta=true to image info URL so API returns full generation
metadata (resources with hash/type) instead of null meta
- Fix checkpoint assignment guard: check modelId instead of id so non-
checkpoint types (upscaler) are not wrongly set as recipe checkpoint
- Skip modelVersionIds loop when resources/civitaiResources already
provided LoRAs, preventing hash-resolved duplicates
- Fix int/str type comparison in CivArchive get_model_version so
version ID matching works correctly
When CivitAI image API returns meta=null and modelVersionIds at root
level, the import flow now:
- Injects modelVersionIds + browsingLevel into a minimal metadata dict
so the parser can discover LoRAs and checkpoints (both import-from-url
and analyze-image paths)
- Adds checkpoint dedup + fallback in the parser's modelVersionIds
handler to avoid duplicate API calls
- Runs EXIF extraction unconditionally in analyze-image path, then
merges with API metadata (fixes gen params loss)
- Propagates preview_nsfw_level through all three import paths:
import-from-url, analyze-image (UI Import), and batch-import,
plus the frontend save flow
- Prefer file type (UNet/Diffusion Model) over baseModel name when
deciding whether a checkpoint routes to the unet folder
- Add UNet to backend primary file type whitelist
- Add Krea 2 to DIFFUSION_MODEL_BASE_MODELS
- Include UNet/Diffusion Model files in frontend file selection UI
- Use actual file type from CivitAI in download params instead of
hardcoded 'Model'
- Convert marquee selection from viewport to document coordinates so
scrolling during a drag no longer deselects off-screen cards.
- Add RAF-based auto-scroll when dragging near viewport edges.
- Compute off-screen card positions from VirtualScroller layout
parameters instead of relying on DOM queries.
The document-level click handler in SortDropdown.js called trigger.focus()
unconditionally on every click outside the sort group. When a model card
was clicked to open the modal, focus() triggered scrollIntoView on the
.sort-trigger button, perturbing .page-content.scrollTop and causing the
card grid to jump up a few pixels.
The same interference also broke the back-to-top smooth-scroll animation:
frame-by-frame focus/scroll perturbations caused VirtualScroller to
schedule repeated re-renders, interrupting the compositor-thread scroll.
Fix: only return focus to the trigger when the dropdown was actually open,
so ordinary page clicks (e.g. clicking a model card) never force focus.
When refreshing updates with a folder filter, versions already present in
other folders were excluded from the is_in_library check, making them
appear as available updates. When the user tried to download, the global
check found the file already exists and returned 'model already exists'.
Fix by also collecting the cross-folder version set when folder_path is
provided, and using the union (folder-filtered + cross-folder) for
is_in_library in both _build_record_from_remote and
_merge_with_local_versions.
The back-to-top button used scrollTo({top:0, behavior:'smooth'}) which
conflicts with VirtualScroller's DOM manipulations during the smooth
scroll animation. Each animation frame triggered handleScroll() ->
scheduleRender() -> renderItems(), causing the browser to interrupt
the smooth scroll animation mid-way, resulting in only ~1 page of
upward scroll instead of reaching the top.
Root cause: commit 311e89e9 fixed VirtualScroller to listen on the
correct scroll container (.page-content), but this meant every scroll
event during smooth animation now triggers expensive DOM operations
that abort the browser's compositor-thread smooth scroll animation.
Fix: use instant scroll (scrollTop = 0) so the position is set
immediately without triggering frame-by-frame VirtualScroller
interference.
_drain_stderr and _wait_until_ready both read from the same stderr pipe.
Starting the drain task before _wait_until_ready creates a race where the
drain task consumes aria2's early-exit error message before the startup
waiter can read it, resulting in an empty error message in the logs.
Also confirmed that --fsync does not exist as an aria2 option (exit code
28 = Invalid argument).
Exit code 28 (Invalid argument) indicates this user's aria2c does not
support the --fsync option. Remove it unconditionally; the stderr drain,
relaxed RPC timeouts, and increased retry coverage remain in place.
aria2 default --fsync=true calls fsync() after each write, which blocks
the entire single-threaded process on large files under Docker overlay.
Add --fsync=false to eliminate this blocking source.
Relax aiohttp session timeout: total=30 → sock_connect=10, sock_read=60
so that transient I/O delays don't cut off legitimate tellStatus RPCs.
Increase retry params (4 attempts, 3s delay) to give aria2 more recovery
time when blocked on synchronous I/O.
Root cause: aria2c subprocess stderr pipe (64 KB buffer) was never
drained. When enough error/warning output accumulated, aria2's write()
blocked, freezing the entire process including its RPC handler. The
tellStatus call then timed out after 30s with asyncio.TimeoutError(),
producing the empty error message in 'Failed to query aria2 download
status: '.
Fixes:
- Drain stderr in a background task so pipe never fills up
- Retry get_status() RPC calls up to 3 times on transient failure
- In the failure path, preserve .safetensors when .aria2 is absent
(the download was likely complete on disk)
In Vue render mode, ComfyUI's TransformPane uses a capture-phase wheel
handler (@wheel.capture) that fires before the tag element's bubble-phase
strength-adjustment listener. It checks wheelCapturedByFocusedElement(),
which requires data-capture-wheel on a focused element. The tag divs had
data-capture-wheel but were not focusable, so the check failed, causing
the capture handler to forward the event to the canvas (triggering zoom)
and stopPropagation() which prevented the strength handler from running.
Fix: move data-capture-wheel from individual tags to the container, make
it focusable (tabIndex=-1), and add a window-level capture-phase wheel
listener that focuses the container before TransformPane checks it.
- Add .grid-loading-overlay CSS: position:absolute inside card grid,
semi-transparent dark background, z-index 100, pointer-events:none
- Add showGridLoading() / hideGridLoading() to VirtualScroller:
creates/removes the scoped overlay inside the card grid only
- Modify loadMoreWithVirtualScroll(): replace full-page
state.loadingManager overlay with grid-scoped loading, defer
hide via requestAnimationFrame to eliminate blank-frame gap
- Clean up gridLoadingOverlay in dispose() to prevent DOM leak
- Replace page-specific header.search.placeholders.* keys with a single
header.search.placeholder key (value: "Search", no ellipsis)
- Keep header.search.notAvailable for the statistics page
- Remove unused placeholder/placeholders/notAvailable entries from all
10 locale files; preserve options and searchIn keys
- Update Jinja template and JS header to use the new unified key
- Fix Vue mode: text widgets (CLIPTextEncode, Prompt LM) had no
[data-testid=widget-layout-field-label], so findRowEl never matched.
Added fallback strategies: bare <label> text match and widget index match.
- Fix Vue mode: flash background pulse was never applied — @keyframes was
defined but no rule bound it to .lm-flash. Replaced with CSS transition
on .lm-flash-host class for value text color fade-in/fade-out.
- Fix Vue mode: -webkit-text-fill-color set by ComfyUI overrode
even with !important. Added -webkit-text-fill-color override to .lm-flash.
- Fix canvas mode: highlight rect was double-offset because onDrawForeground
ctx is pre-translated to node.pos. Removed background rect entirely per
design decision; kept text_color + inline color only.
- Add fade-in (250ms) / fade-out (400ms) for text color in both modes.
Canvas-drawn widgets use rAF color interpolation; DOM widgets use CSS
transition. Fixed hexToRgb to handle 3-digit hex shorthand (#DDD).
- Add hover dismissal to canvas mode via app.canvas.getWidgetAtCursor().
Vue mode already had it via mouseover listener.
- Replace 60fps rAF poll with 100ms setInterval for hover detection.
- Fix batch cleanup closure bug: isDomWidget evaluated per-widget instead
of per-call; fade rAF cancellers tracked per-widget in _lmFadeCancels map.
- Unify flash color from #66B3FF to LM brand accent #4299E0.
- Fix Vue fade-out: keep .lm-flash-host 300ms after removing .lm-flash so
CSS transition persists. Canvas DOM widgets: keep inline transition 300ms
after clearing color.
Sort by Most/Fewest versions first now works when Group by model is off.
- Backend: group items by modelId (respecting version_grouping setting),
count versions per group, sort groups by count, expand groups with
versions sorted by version id descending
- CSS: remove rule that hid the sort option in non-grouped mode
- Tests: add 3 tests covering desc, asc, and same_base variants
When viewing all versions of a model (VLM mode via 'x versions' button):
- Backend always sorts by version ID descending, ignoring current sort_by
- A temporary 'Newest version first' option is injected into the sort
dropdown (removed on exit, not a permanent option)
- The sort dropdown is disabled (greyed out) while VLM is active
- On clearing VLM, the previous sort preference is restored and the
dropdown re-enabled
- Handles stale VLM state (e.g. after page reload with leftover session)
- Covers all three model page types: loras, checkpoints, embeddings
Also fixes review nits:
- Correct i18n call pattern (defaultValue in options object)
- Shared _restoreSortAfterVlm() helper to avoid triple duplication
- group_by_model dedup now counts versions per group and attaches
version_count; respects update_flag_strategy (same_base) by
sub-grouping on base_model
- Card footer shows clickable 'x versions' link instead of version
name when grouped (hides HIGH/LOW badges); clicking triggers
View Local Versions without page reload
- Added 'Local Versions' sort option (versions_count), auto-hidden
when group_by_model is off
- Sort preference is saved/restored separately for normal and
grouped modes
- VLM flow (triggerVlmView, clearCustomFilter) uses resetAndReload()
via API instead of window.location.reload()
- Fixed cache mutation bug: version_count is now set on a shallow
copy, not the cached dict, preventing stale version_count leaking
into VLM responses
- i18n: all 9 locale files translated
