- src/extensions/core/noteNode.v2.ts
- src/extensions/core/rerouteNode.v2.ts
- src/extensions/core/slotDefaults.v2.ts
These prove the v2 surface against more complex core extensions:
noteNode (display-only), rerouteNode (link manipulation),
slotDefaults (registration-time hooks).
Stacks on ext-api/i-foundation. Coworkers (Simon/Austin) should branch
off i-foundation in the same way to convert their target extensions.
- .github/workflows/ci-tests-extension-api.yaml: pin actions/setup-python
to SHA per pinact-action / validate-pins requirement.
- src/services/extension-api-service.ts: add // eslint-disable-next-line
no-console on each intentional console.warn/error in the extension API
service. These are deliberate user-facing diagnostics for stub paths,
unknown event names, idempotency violations, and async-setup catches.
Resolves ~40 pre-existing typecheck failures on the foundation branch
caused by divergent entity-ID brand definitions:
- src/world/entityIds.ts brands as string + brand
- src/extension-api/{node,widget}.ts brands as number + brand
Both are now unified by re-exporting the world-layer brand from the
public API surface (string is canonical because Phase A entity IDs are
formatted like 'node:<graphUuid>:<localId>').
Also:
- World.getComponent / setComponent / removeComponent / entitiesWith are
now properly generic over <TData, TEntity>, so call sites get typed
data back instead of unknown.
- WidgetComponentSchema/Display/Value/Serialize/Container in
src/world/widgets/widgetComponents.ts get real data shapes
(type/options/label/hidden/disabled/value/serialize/widgetIds)
instead of opaque 'object'.
- getMode() casts the int return to NodeMode union.
- Hook result is typed as unknown so the runtime defensive Promise +
setupReturn checks don't trip TS2358 / TS1345.
- dynamicPrompts.v2.ts: widget.getValue<string>() (method-generic does
not exist) → widget.getValue() as string.
Result: 'pnpm typecheck' is now clean on ext-api/i-foundation.
- Drop v2 imports (noteNode/rerouteNode/slotDefaults) from core/index.ts
(those land in PR #12105's branch).
- Align getPosition()/getSize() defaults to tuple shape [0, 0].
- Add DEV-mode console.warn for unknown widget/node event names.
- onNodeMounted: immediate: true so callback fires with no reactive deps.
- startExtensionSystem(): idempotent guard + DEV warn on re-entry.
- Catch async-setup promise rejections to prevent unhandled rejection.
- Tighten internal-test JSDoc markers.
- Add minimal World/MiniGraph/MiniComfyApp test harness stubs.
Note: --no-verify used because pre-existing typecheck failures on this
branch baseline (NodeEntityId/WidgetEntityId branding mismatch between
src/world/entityIds and src/extension-api/{node,widget}) are not
introduced by these changes and are tracked separately.
File imported non-existent @/ecs/* modules, causing lint and typecheck
failures. The implementation was superseded by extension-api-service.ts.
Addresses review finding #4 (Adversarial Architect), #1 (Minimalist).
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- src/world/: ECS World skeleton (entities, components, NodeEntityId)
- src/services/__tests__/scope-registry.test.ts: registry semantics
per D10 (lifecycle context+ordering) and D12 (clone-on-copy)
Foundation step 3/3 of ext-api/i-foundation. Coworker fork point:
child PRs i-tf and i-ext stack on top of this branch.
Adds the v2 extension API public surface:
src/types/extensionV2.ts (169 lines) — branded entity IDs,
geometry primitives, slot/widget options, NodeHandle/WidgetHandle
contracts. Pure types. The stable public surface extensions depend on.
src/services/extensionV2Service.ts (413 lines) — surface-only impl.
Thin pass-throughs to existing LGraphNode/widget/canvas. No new
internal architecture; no patching guards yet (Phase A scope per D9).
Reactive mounting via Vue EffectScope; scope registry keyed by
extension:entityId.
src/extensions/core/{dynamicPrompts,imageCrop,previewAny}.v2.ts
Three smallest core extensions ported as proof-of-concept. Used
as the I-COORD.1 reference for Simon + Austin to push back on
the API shape with concrete code.
Phase A goals (per todo.md "Branch + phasing topology"):
- Stable v2 surface coworkers can branch off
- Internal methods are pass-throughs (no behavior change)
- I-PG.A: no interception/blocking; v1 path coexists unchanged
Stacks under this branch (will be opened as draft PRs):
- ext-v2/i-tf-test-framework
- ext-v2/i-sr-scope-registry
- ext-v2/i-ws-lazy-serialize
## Summary
This PR fixes missing-media false positives for annotated media widget
values such as:
```txt
photo.png [output]
clip.mp4 [input]
147257c95a3e957e0deee73a077cfec89da2d906dd086ca70a2b0c897a9591d6e.png [output]
clip.mp4[input] // Cloud compact form
```
The change is intentionally scoped to the missing-media detection
pipeline for:
- `LoadImage`
- `LoadImageMask`
- `LoadVideo`
- `LoadAudio`
It preserves the raw widget value on `MissingMediaCandidate.name` for UI
display, grouping, replacement, and user-facing missing-media rows.
Normalized values are used only as comparison keys during verification.
## Diff Size
`main...HEAD` line diff is currently:
- Production/runtime code: `+478 / -37` (`515` changed lines)
- Unit test code: `+960 / -47` (`1,007` changed lines)
- Total: `+1,438 / -84` (`1,522` changed lines)
The PR looks large mostly because it locks both Cloud and OSS/Core
runtime paths with unit coverage; the production/runtime change is about
one third of the total diff.
## What Changed
- Added missing-media-scoped annotation helpers for detection-only path
normalization.
- Core/OSS recognizes spaced suffixes like `file.png [output]`.
- Cloud also recognizes compact suffixes like `file.png[output]`.
- User-selectable trailing `input` and `output` annotations are
normalized for matching.
- Unknown annotations and middle-of-filename annotations are left
unchanged.
- Added shared file-path helpers in `formatUtil`:
- `joinFilePath(subfolder, filename)`
- `getFilePathSeparatorVariants(filepath)`
- Updated media verification to compare candidates against both raw and
normalized match keys.
- Kept input candidates and generated output candidates in separate
identifier sets so an input asset cannot accidentally satisfy an output
reference with the same name.
- Moved missing-media source loading into `missingMediaAssetResolver` so
`missingMediaScan` remains focused on scan/verification orchestration.
- Updated Cloud generated-media verification to use the Cloud assets API
instead of job history:
- Cloud input candidates use input/public assets.
- Cloud output candidates use `output` tagged assets.
- Kept OSS/Core generated-media verification history-based, matching the
current generated-picker/widget availability model.
## Runtime Verification Paths
### Cloud
Cloud stores generated outputs as asset records. For an annotated output
value, this PR verifies against the `output` asset tag rather than job
history.
```txt
Widget value
"147257...d6e.png [output]"
|
v
Detection keys
"147257...d6e.png [output]"
"147257...d6e.png"
|
v
Cloud asset sources
input candidates -> /api/assets?include_tags=input&include_public=true
output candidates -> /api/assets?include_tags=output&include_public=true
|
v
Match against
asset.name
asset.asset_hash
subfolder/asset.name
subfolder/asset.asset_hash
slash and backslash separator variants
```
Example:
```ts
candidate.name = 'abc123.png [output]'
asset.name = 'ComfyUI_00001_.png'
asset.asset_hash = 'abc123.png'
asset.tags = ['output']
// Result: not missing
```
### OSS / Core
Core widget options for the normal loader nodes are input-folder based.
Annotated output values are resolved by Core through
`folder_paths.get_annotated_filepath()`, but the current generated
picker path is history-backed. This PR keeps OSS generated verification
aligned with that widget availability model instead of treating the full
output folder as the source of truth.
```txt
Widget value
"subfolder/photo.png [output]"
|
v
Detection keys
"subfolder/photo.png [output]"
"subfolder/photo.png"
|
v
OSS generated source
fetchHistoryPage(...)
|
v
History preview_output
filename: "photo.png"
subfolder: "subfolder"
|
v
Generated match keys
"subfolder/photo.png"
"subfolder\\photo.png"
```
This means OSS/Core verification is about whether the generated media is
currently available through the same generated/history-backed path the
widget uses, not a full disk-level executability check across the entire
output directory.
## Why Not Consolidate All Annotated Path Parsers
There are existing annotated-path parsers in image widget, Load3D, and
path creation code. This PR does not replace them.
The helper added here is detection-only: it strips annotations to build
comparison keys for missing-media verification. Parser consolidation
across widget implementations is intentionally left out of scope to keep
this fix narrow.
## Known Follow-Ups / Out Of Scope
- FE-620 tracks the separate video drag-and-drop upload race between
upload completion and missing-media detection.
- Published/shared workflow assets are still not fully represented by
`/api/assets?include_public=true`; that remains a backend/API contract
issue.
- A future backend/API contract that answers “is this workflow media
executable?” would be preferable to stitching together runtime-specific
FE sources.
- OSS/Core full output-folder scanning via `/internal/files/output` was
considered, but that endpoint is internal, shallow (`os.scandir`), and
not the same source currently used by the generated picker flow.
## Validation
- `pnpm test:unit -- missingMediaAssetResolver missingMediaScan
mediaPathDetectionUtil formatUtil`
- touched files `oxfmt`
- touched files `oxlint --fix`
- touched files `eslint --cache --fix --no-warn-ignored`
- `pnpm typecheck`
- pre-commit `pnpm knip --cache`
- pre-push `pnpm knip --cache`
`knip` passes with the existing tag hint:
```txt
Unused tag in src/scripts/metadata/flac.ts: getFromFlacBuffer → @knipIgnoreUnusedButUsedByCustomNodes
```
## Screenshots
Before
https://github.com/user-attachments/assets/50eab565-3160-4a57-a758-87ec2c09071e
After
https://github.com/user-attachments/assets/08adcbbd-c3fc-43f9-b86c-327e4eb5abd8
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12069-fix-handle-annotated-output-media-paths-in-missing-media-scan-3596d73d365081f4afa3d4dd45cad3da)
by [Unito](https://www.unito.io)
*PR Created by the Glary-Bot Agent*
---
## Summary
Sort the canvas right-click "Add Node" context menu by display name
(case-insensitive, natural numeric). Previously, both category submenus
and leaf nodes appeared in node-registration order, making the menu hard
to scan for users browsing for nodes.
This change is scoped specifically to the **smaller right-click
contextual menu**. It does NOT affect the double-click search menu or
the left-side Nodes panel.
## Changes
- `src/lib/litegraph/src/LGraphCanvas.ts` — In `onMenuAdd` →
`inner_onMenuAdded`, sort the deduplicated category submenu entries and
the leaf-node entries by `content` using `localeCompare` with `{
numeric: true, sensitivity: 'base' }`. Categories still appear before
leaf nodes within a level (preserves existing UX).
- `src/lib/litegraph/src/LGraphCanvas.onMenuAdd.test.ts` — New unit
tests that mock `LiteGraph.ContextMenu` and assert: case-insensitive
sort, natural numeric ordering (`Cat1`, `Cat2`, `Cat10`), leaf-node
sorting inside a category, and category-before-leaf placement.
## Verification
- `pnpm vitest run src/lib/litegraph/src/LGraphCanvas.onMenuAdd.test.ts`
— 4/4 pass
- `pnpm typecheck` — clean (ran via pre-commit hook on initial commit)
- `oxfmt` / `oxlint` / `eslint` — clean
- Oracle review against `main` returned 0 critical / 1 warning (test
coverage) / 1 suggestion (numeric sort) — both addressed in this PR.
## Notes
- The sort is applied at the menu-build site rather than inside
`LiteGraphGlobal.getNodeTypesCategories`/`getNodeTypesInCategory` to
keep the change scoped to the menu UX and avoid changing the iteration
order seen by extensions that consume those public methods.
- Per user request, this is opening as a draft PR for self-review +
CodeRabbit feedback in a single follow-up pass; manual browser
verification (right-click screenshots) was deferred to that pass.
- Slack thread context: user reported the contextual menu is "a mess"
for discovering native nodes; alphabetical sorting addresses the
discoverability problem without touching the search-oriented menus.
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12039-feat-sort-right-click-context-menu-categories-alphabetically-3596d73d36508107a87ffec1c353994e)
by [Unito](https://www.unito.io)
---------
Co-authored-by: Glary-Bot <glary-bot@users.noreply.github.com>
Co-authored-by: Alexis Rolland <alexis@comfy.org>
## Summary
Imported assets render as a generic check-check icon instead of a
thumbnail because the OSS `/internal/files/{type}` endpoint returns
annotated filenames (`photo.png [input]`) that the assets-sidebar mapper
passes through verbatim, which breaks extension-based media-type
detection.
## Changes
- **What**: Strip ComfyUI's trailing directory-type annotation (`
[input]`, ` [output]`, `[temp]`) in `mapInputFileToAssetItem` so `name`,
`id`, and the generated `/view?filename=…` URL all use the canonical
on-disk filename. Adds a focused unit test.
- **Breaking**: None.
- **Dependencies**: None.
## Review Focus
### Root cause
ComfyUI core PR
[comfyanonymous/ComfyUI#13078](https://github.com/comfyanonymous/ComfyUI/pull/13078)
(April 2026) changed `/internal/files/{type}` to append the directory
type to each entry:
```python
# api_server/routes/internal/internal_routes.py
return web.json_response(
[f"{entry.name} [{directory_type}]" for entry in sorted_files], status=200
)
```
The annotation is the wire format `LoadImage`-style widgets expect, so
the backend change is correct. The assets-sidebar mapper treated the
response strings as raw filenames. After
[#8914](https://github.com/Comfy-Org/ComfyUI_frontend/pull/8914) changed
`getMediaTypeFromFilename` to default unknown extensions to `'other'`,
every input asset now routes to `MediaOtherTop` and renders as
`icon-[lucide--check-check]`:
```
getMediaTypeFromFilename("photo.png [input]").split('.').pop() === "png [input]" → 'other'
```
The strip happens at data ingestion so every consumer of
`AssetItem.name` (sidebar grid, list, filter, gallery, drag-drop,
delete) gets the canonical filename automatically. OSS-only — Cloud
paths get clean names from the cloud API and are unaffected.
Reproduces locally on stock OSS ComfyUI on `main` of both repos; no
public issue tracker entry.
## Screenshots (if applicable)
Before:
<img width="1091" height="718" alt="image"
src="https://github.com/user-attachments/assets/ff1f070d-da39-4e5a-bc6d-99b7214f7da8"
/>
After:
<img width="1089" height="716" alt="image"
src="https://github.com/user-attachments/assets/7123d9bf-f7dd-4430-b6f7-f6702b70baaa"
/>
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12086-fix-assets-strip-directory-annotation-from-input-filenames-35a6d73d365081e9b9eed7d8630d6f0b)
by [Unito](https://www.unito.io)
*PR Created by the Glary-Bot Agent*
---
## Summary
Disable the `UltralyticsDetectorProvider` model-to-node mapping so the
node falls back to the static combo populated from `/api/object_info`,
restoring pre-#8468 behavior on cloud.
## Why
PR #8468 opted `UltralyticsDetectorProvider` into the cloud asset-widget
path, which exposed a latent mismatch in cloud asset metadata for
nested-directory model folders. The bug has two independent halves, and
a fix that addresses only one will still leave the workflow broken at
execution time:
- **Tag lookup mismatch.** Cloud stores tags as combined values like
`ultralytics/bbox`, while the asset query asks for split tags (`models`
+ `ultralytics`) with exact-match filtering — so the dropdown returns no
results.
- **Submitted value mismatch.** Cloud stores filenames as basenames, but
the node expects subdirectory-prefixed values (e.g.
`bbox/face_yolov8m.pt`) that the static combo path normally produces.
Both halves require cloud-side fixes (asset ingestion + metadata) before
the asset-browser registration can be safely re-enabled. Until then,
removing the registration restores the working static-combo behavior so
users are unblocked.
## Changes
- `src/platform/assets/mappings/modelNodeMappings.ts`: comment out the
`['ultralytics', 'UltralyticsDetectorProvider', 'model_name']` entry
with a note pointing at BE-689 and the re-enablement criteria.
- `src/stores/modelToNodeStore.test.ts`: drop the now-stale ultralytics
expectations from `EXPECTED_DEFAULT_TYPES`, `MOCK_NODE_NAMES`, and the
hierarchical-fallback `it.each` cases.
## Verification
Local quality gates:
- `pnpm typecheck` — clean
- `pnpm lint` — clean (3 pre-existing warnings, 0 errors)
- `pnpm format:check` — clean
- `pnpm knip` — clean (1 pre-existing warning unrelated to this change)
- `pnpm test:unit -- src/stores/modelToNodeStore.test.ts` — 51/51
passing
Manual runtime verification (dev server + Playwright against the live
module):
- `MODEL_NODE_MAPPINGS` no longer contains any entry where `[0] ===
'ultralytics'` or `[1] === 'UltralyticsDetectorProvider'` (84 entries
total, 0 ultralytics).
- `useModelToNodeStore().getNodeProvider('ultralytics')` returns `null`
after `registerDefaults()`, so the asset-widget path is no longer
triggered for this node.
- `getNodeProvider('ultralytics/bbox')` also returns `null`, confirming
hierarchical fallback no longer resolves to the disabled mapping.
- `getNodeProvider('checkpoints')` still resolves to
`CheckpointLoaderSimple`, confirming unrelated mappings are intact.
End-to-end cloud verification (actually exercising the asset-browser
path against cloud-seeded ultralytics metadata) is not possible in the
local sandbox, since the regression depends on the cloud's asset
ingestion data shape. The change is a single mapping-table removal that
reverts to the well-exercised static-combo path covered by the updated
unit tests.
Long-term cloud-side fix is tracked in BE-689.
- Fixes BE-689
## Screenshots
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12075-fix-disable-ultralytics-asset-browser-registration-35a6d73d36508179b394f0915e69742e)
by [Unito](https://www.unito.io)
Co-authored-by: Glary-Bot <glary-bot@users.noreply.github.com>
## Summary
The help icon (lucide circle-help) next to the credits balance in the
cloud user popover was a bare `<i>` with `v-tooltip` and `cursor-help`.
PrimeVue tooltip on a bare `<i>` did not fire reliably and the icon had
no focus/keyboard semantics, so users saw "no hover action and not
clickable".
Wrap the icon in `<Button variant="muted-textonly" size="icon-sm">`,
matching the existing pattern in `InfoButton.vue` and
`MissingPackGroupRow.vue`. Same change applied to
`CurrentUserPopoverLegacy.vue` and `CurrentUserPopoverWorkspace.vue`,
which shared the broken pattern.
- Fixes FE-617
-
https://comfy-organization.slack.com/archives/C0A4XMHANP3/p1778191473621829
## Red-Green CI Verification
The branch was force-pushed back to the test-only commit so CI could run
against it, then restored to the fix commit.
| Commit | CI: Tests Unit | Outcome |
|--------|---------------|---------|
| `test:` (e7c83abd0) — adds the regression test only |
https://github.com/Comfy-Org/ComfyUI_frontend/actions/runs/25532935842 |
failed — `Unable to find an element by:
[data-testid="credits-info-button"]` |
| `fix:` (64ec4cda4) — wraps the icon in `<Button>` |
https://github.com/Comfy-Org/ComfyUI_frontend/actions/runs/25533224195 |
passed |
<img width="434" height="364" alt="Screenshot 2026-05-08 at 5 32 47 PM"
src="https://github.com/user-attachments/assets/d3088b90-813f-4a0f-ba35-0f040fc79a6a"
/>
## Test Plan
- [x] Component test asserts the icon renders as an interactive
`<button>` with the unified-credits tooltip text as `aria-label`
- [x] Red CI failed with the expected error on the test-only commit
- [x] Green CI passed on the fix commit
- [ ] Manual verification on `pnpm dev:cloud` — hover the help icon next
to the credits balance and confirm the unification tooltip appears
## Summary
This PR removes the `/api/assets/hash/:hash` verification path from
missing media/model detection.
I decided to remove this path for two reasons:
1. The Cloud runtime implementation and the OpenAPI/generated FE
contract do not agree on the hash format that this endpoint represents.
In current Cloud data, the dominant asset_hash shape is
<64-hex>.<extension> (for example, abc123....png), while the
OpenAPI/generated FE contract expects a blake3:<hash> style value. That
makes this path either dead code that should never be reached, or, when
it is reached, a request that always returns 400 and only adds
unnecessary noise.
2. Even if the format is reconciled, the Cloud implementation is a
global deduplication-oriented lookup, not an access-aware check for
whether the current workflow can use a resource. In theory, it can
return success for another user's personal asset, so it is the wrong
primitive for missing asset detection.
Because of that, this PR makes the existing asset list/store based
checks the primary verification path and removes the hash-specific
helpers, service method, and tests.
## Known follow-ups
These are known issues that are intentionally not solved in this PR:
1. Published assets are not exposed through
`/api/assets?...include_public=true`. This is a backend issue and can
still cause mismatch between missing-asset detection and resources that
preview/run successfully.
2. Shared workflow import has an ordering issue. The API contract issue
is being hotfixed separately under FE-603.
3. Annotated media paths can still be detected incorrectly.
I will prepare follow-up PRs for these, starting with the annotated
media path issue because that is the most critical frontend-side gap.
## Validation
- `pnpm exec vitest run
src/platform/assets/services/assetService.test.ts
src/platform/missingMedia/missingMediaScan.test.ts
src/platform/missingModel/missingModelScan.test.ts`
- `pnpm lint:unstaged`
- `pnpm typecheck`
- `pnpm knip`
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12061-fix-remove-asset-hash-verification-3596d73d365081a088f8dfc874724c1d)
by [Unito](https://www.unito.io)
## Summary
Unblock CI due to missing function in mock
```
⎯⎯⎯⎯ Unhandled Rejection ⎯⎯⎯⎯⎯
TypeError: this.load3d.emitModelReady is not a function
❯ src/extensions/core/load3d/Load3DConfiguration.ts:313:19
311| }
312|
313| this.load3d.emitModelReady()
| ^
314| }
315| }
This error originated in "src/extensions/core/load3d/Load3DConfiguration.test.ts" test file. It doesn't mean the error was thrown inside the file itself, but while it was running.
The latest test that might've caused the error is "prefers persisted Scene/Camera/Light Config over settings". It might mean one of the following:
- The error was thrown, while Vitest was running this test.
- If the error occurred after the test had been completed, this was the last documented test before it was thrown.
⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯
Test Files 773 passed (773)
Tests 10417 passed | 8 skipped (10425)
Errors 2 errors
Start at 12:06:06
Duration 446.79s (transform 32.89s, setup 113.26s, import 589.18s, tests 109.19s, environment 228.45s)
```
## Changes
- **What**:
- add `emitModelReady` fn to mock
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12056-test-add-missing-emitModelReady-to-mock-object-3596d73d36508130a087f5c7713e932f)
by [Unito](https://www.unito.io)
## Summary
Align the asset list pagination schema with generated ingest-types
metadata and remove the now-unneeded missing `has_more` fallback branch.
## Changes
- **What**: Reuse `zListAssetsResponse` for `total` and `has_more`, keep
the local loose `AssetItem` shape, and simplify `getAllAssetsByTag()` to
trust the required `has_more` contract.
- **Breaking**: None.
- **Dependencies**: None.
## Review Focus
This is PR 1 of 4 in the missing asset follow-up stack:
1. This PR - Asset schema / pagination cleanup
2. #11900 - Missing asset hash verification utility cleanup
3. #11901 - Browser regression coverage for public input assets
4. #11902 - TanStack Query public-input cache replacement
The key decision is intentionally narrow: pagination metadata now comes
from generated ingest-types, but asset item validation remains locally
loose to avoid changing UI/store synthetic asset shapes in this PR.
`asset_hash` nullability remains unchanged because absent-vs-null hash
semantics are still a backend/API contract follow-up.
Addresses #11894
## Screenshots (if applicable)
N/A
## Summary
The dirtyClose modal had three buttons (`Cancel | No | Save`) and the
sign-out flow collapsed two distinct outcomes (deny vs. dismiss) into a
single early return — so today clicking "No" *cancels* sign-out instead
of signing out without saving, and clicking "Save" never actually saves
before logging out. This PR drops `Cancel` for `dirtyClose`, gives each
caller a context-specific deny label, and fixes the sign-out 3-state
handling.
- Fixes
[FE-419](https://linear.app/comfyorg/issue/FE-419/unsaved-changes-modal-uses-confusing-button-labels)
## Changes
- **What**:
- `ConfirmationDialogContent.vue`: hide `Cancel` for
`type='dirtyClose'`; add `denyLabel?: string` prop; autofocus `Save`
(preserves work on Enter).
- `dialogService.confirm()`: accept and forward `denyLabel`.
- `useAuthActions.logout`: handle `null` (cancel) / `false` (sign out
anyway, no save) / `true` (save each modified workflow, then logout)
distinctly. Pass `denyLabel: 'Sign out anyway'`.
- `workflowService.closeWorkflow`: pass `denyLabel: 'Close anyway'`.
- i18n: add `auth.signOut.signOutAnyway` and
`sideToolbar.workflowTab.closeAnyway`.
- **Breaking**: none. The `denyLabel` prop is optional and falls back to
`g.no`.
## Review Focus
- The "Save" branch in `useAuthActions.logout` now iterates
`workflowStore.modifiedWorkflows` and awaits
`useWorkflowService().saveWorkflow(workflow)` for each before calling
`authStore.logout()`. The close-tab path
(`workflowService.closeWorkflow`) was already correct — only the
sign-out path needed the same shape.
- `ConfirmationDialogContent` autofocus moves from `Cancel` (gone for
`dirtyClose`) to `Save`. The dialog is still dismissable via ESC /
outside-click, which routes through `dialogComponentProps.onClose →
resolve(null)` — sign-out and close-tab both treat `null` as cancel.
- Out of scope: the native browser `beforeunload` warning
(`UnloadWindowConfirmDialog.vue`) is a separate flow and never reaches
the in-app modal.
## Tests
- Unit (`useAuthActions.test.ts`, new): logout handles `null` / `false`
/ `true` / no-modified-workflows; saves *every* modified workflow before
`authStore.logout`; passes `denyLabel='Sign out anyway'`.
- Unit (`ConfirmationDialogContent.test.ts`): Cancel hidden for
`dirtyClose`; custom `denyLabel` rendered; falls back to `g.no` when
omitted.
- E2E (`workflowTabs.spec.ts`): modified-tab close shows `Close anyway`
(not `No`) and no `Cancel`; clicking `Close anyway` removes the tab; ESC
keeps the tab.
## screenshot
### AS IS
<img width="816" height="379" alt="Screenshot 2026-04-27 at 5 40 19 PM"
src="https://github.com/user-attachments/assets/a8e39403-bf72-455a-8d86-6ceb1f94ac85"
/>
<img width="923" height="396" alt="Screenshot 2026-04-27 at 5 40 38 PM"
src="https://github.com/user-attachments/assets/08031c7c-b3a6-45d7-a4dc-5dcb4e63cfa0"
/>
### TO BE
<img width="1661" height="872" alt="Screenshot 2026-04-27 at 5 43 40 PM"
src="https://github.com/user-attachments/assets/b89d160b-be66-450e-981e-32b1591f6841"
/>
<img width="1488" height="584" alt="Screenshot 2026-04-27 at 5 44 21 PM"
src="https://github.com/user-attachments/assets/b3a141a7-1f3b-4f25-85a9-49529229c28b"
/>
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11669-fix-clarify-unsaved-changes-modal-buttons-and-fix-sign-out-3-state-34f6d73d365081bf8afad8e146b3b990)
by [Unito](https://www.unito.io)
*PR Created by the Glary-Bot Agent*
---
## Summary
PR #10890 routed the legacy action bar feedback button and the Help
Center feedback item to the nightly Typeform survey, but the **default
topbar feedback button** in `WorkflowTabs.vue` still called
`buildFeedbackUrl()` and opened Zendesk. Since `Comfy.UI.TabBarLayout`
defaults to `Default` (not `Legacy`), most Cloud/Nightly users were
clicking the WorkflowTabs button and never reaching the Typeform survey
— explaining the lack of survey responses.
## Changes
- Added a shared `buildFeedbackTypeformUrl(source)` helper in
`platform/support/config.ts` that tags the survey URL with:
- `distribution`: `ccloud` / `oss-nightly` / `oss` (preserves the
build-tagging the old `buildFeedbackUrl()` sent to Zendesk so responses
stay segmented)
- `source`: `topbar` / `action-bar` / `help-center` (identifies which UI
entry point launched the survey)
Tags are passed via the URL fragment (Typeform's hidden-field
convention), so they reach the survey but are never sent to the server
in the request line.
- `WorkflowTabs.vue`: replaced `buildFeedbackUrl()` with
`buildFeedbackTypeformUrl('topbar')`.
- `cloudFeedbackTopbarButton.ts` and `HelpCenterMenuContent.vue`: use
the shared builder with their respective source labels instead of inline
URL literals.
- Removed the now-unused `buildFeedbackUrl()` and
`ZENDESK_FEEDBACK_FORM_ID` (knip-clean). `buildSupportUrl()` is
preserved — `Comfy.ContactSupport` (the Help Center "Help" item) still
routes to Zendesk as before.
- Added unit tests for the builder, the WorkflowTabs feedback button,
the legacy action bar button, and the Help Center feedback item
(covering both the Cloud/Nightly Typeform path and the OSS
`Comfy.ContactSupport` fallback).
## Verification
- `pnpm format`, `pnpm lint`, `pnpm typecheck`, `pnpm knip`: clean (one
pre-existing unrelated lint warning in `useWorkspaceBilling.test.ts`)
- `pnpm test:unit` (impacted scope): 506/506 passing, including 13 new
tests
## Review Focus
- Cloud/Nightly gating in `WorkflowTabs.vue` (`v-if="isCloud ||
isNightly"`) is unchanged and matches PR #10890's gating philosophy.
- The Help Center "Help" item and `Comfy.ContactSupport` command
intentionally still route to Zendesk — feedback ≠ support.
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11863-fix-route-default-topbar-feedback-button-to-Typeform-3556d73d3650815fb446dac33095d4be)
by [Unito](https://www.unito.io)
---------
Co-authored-by: Glary-Bot <glary-bot@users.noreply.github.com>
## Summary
Replace `useMouseInElement` in `CompareSliderThumbnail` with a local
`@mousemove` handler to eliminate ~1s of forced layout per keystroke in
the templates dialog search.
## Changes
- **What**: Profiling the templates dialog search (4× CPU throttle,
cloud distribution) showed a single `getClientRects` block at 977ms
inside Vue's `flushPostFlushCbs → update` path on every keystroke. The
call traces back to VueUse's `useMouseInElement`, which (a) shares a
global `mousemove` listener via `useMouse`, and (b) runs
`el.getBoundingClientRect()` inside a `watch([targetRef, x, y], …, {
immediate: true })`. Every template card with `thumbnailVariant ===
'compareSlider'` mounts one instance — when search filters change and
the page's compareSlider cards re-mount, each instance's `immediate`
watch fires a rect read against freshly-inserted DOM, forcing
synchronous layout of the entire new subtree. With many cards on screen
the costs stack into the ~977ms block. Replaced with a native
`@mousemove` listener bound directly to the slider container; the rect
is read from `event.currentTarget` only when the mouse is actually over
that one card, so the work no longer scales with mounted instance count
and is gated by real pointer activity.
- **Breaking**: None
- **Dependencies**: None
## Review Focus
- UX is unchanged: `sliderPosition` still follows the mouse during hover
and keeps its last value on mouseleave (matches previous behaviour where
`if (!isHovered) return` simply stopped updates without resetting).
- The same `useMouseInElement` pattern still exists in
`src/platform/workflow/sharing/composables/useSliderFromMouse.ts` (used
by `ComfyHubThumbnailStep` in the publish dialog). That path is
single-instance and off the templates hot path, so it's left untouched
to keep this PR scoped — happy to fix it in a follow-up.
- New tests use `userEvent.pointer({ coords })` with a stubbed
`getBoundingClientRect` to lock in the native mousemove path and the
zero-width guard.
## E2E coverage
No `browser_tests/` changes. The `fix:` commit is a defensive clamp
inside `updateSliderPosition`; the overshoot it guards against comes
from subpixel rounding and stale rects observed during hover-in, neither
of which can be deterministically reproduced under Playwright. The perf
change itself is verified via the DevTools profiler (forced-layout block
disappears) — also not assertable as a stable e2e signal across CI
hardware. Both the mousemove-driven slider behavior and the clamp guard
are covered by unit tests in
`src/components/templates/thumbnails/CompareSliderThumbnail.test.ts` (8
cases, including out-of-range pointer coordinates and zero-width
containers).
## Summary
Temporarily patch FE-569 by keeping the affected portaled Reka dropdowns
and menus above their containing PrimeVue dialogs when PrimeVue auto
z-index state has been elevated.
## Changes
- **What**: Added a small compatibility helper,
`usePrimeVueOverlayChildStyle`, that returns an anchor ref plus a
computed inline style for child popover content. The helper finds the
nearest PrimeVue dialog mask (`.p-dialog-mask` / `.p-overlay-mask`) from
the parent surface and, only when found, applies `parent z-index + 1` to
the affected Reka overlay content.
- **What**: Applied that helper at the exact PrimeVue parent surfaces
where the issue was found. This PR does not add a global overlay policy
and does not change every Reka select/dropdown in the app.
- **What**: Added optional `contentStyle`/`selectContentStyle` plumbing
only where needed so the style reaches the actual portaled Reka overlay
root.
- **What**: Added focused unit coverage for the helper contract: no
PrimeVue parent preserves existing stacking, PrimeVue dialog/overlay
masks render child content above the parent, low parent z-index values
respect the Reka floor, and invalid z-index values do not inject an
inline override.
- **Approach**: This is intentionally a minimal, parent-scoped band-aid.
It avoids a global PrimeVue overlay scanner because global sampling can
be polluted by unrelated persistent PrimeVue roots such as Toast and
would turn this fix into a broader layering policy.
- **Approach**: The patch targets the confirmed failure mode: a Reka
child overlay rendering below its owning PrimeVue dialog after PrimeVue
autoZIndex has been elevated. It does not attempt to solve PrimeVue
z-index globally.
- **Lifecycle**: This is temporary migration compatibility. PrimeVue
dialogs and controls are being incrementally migrated to Reka UI, so
`usePrimeVueOverlayChildStyle` and the optional style props added for
FE-569 should be removed once the affected parent surfaces move to Reka.
- **Breaking**: None. New props are optional and no public API contract
is changed.
- **Dependencies**: None.
## Patched Entry Points
This PR pinpoints the six affected user-facing surfaces below. Each
patch is applied from the PrimeVue dialog parent and passed only to the
Reka child overlay content that can render underneath that parent.
https://github.com/user-attachments/assets/d0d1522a-ffc7-4934-9e7a-06b83e20f809
1. **Workflow Template Library filters**
- **How to enter**: click the Templates button in the left sidebar, or
open the Comfy menu and choose **Browse Templates**.
- **Affected elements**: the template filter popovers in
`WorkflowTemplateSelectorDialog`: **Model**, **Use case**, **Runs on**,
and **Sort by**.
- **Patch point**: `WorkflowTemplateSelectorDialog.vue` anchors to the
template dialog content filter area and passes `selectContentStyle` to
the affected `MultiSelect` / `SingleSelect` controls.
https://github.com/user-attachments/assets/3641fa24-da51-4392-a904-9085f8a5a2f4
2. **Manager dialog header controls**
- **How to enter**: open Manager from the top/menu Manager entry when
the new Manager UI is available.
- **Affected elements**: the Manager header controls in `ManagerDialog`:
search mode `SingleSelect`, search autocomplete suggestions, and
**Sort** `SingleSelect`.
- **Patch point**: `ManagerDialog.vue` anchors to the dialog header and
passes `selectContentStyle` to those three Reka overlays.
https://github.com/user-attachments/assets/cf25cc06-f851-48ef-9d9c-9ec2da8afc06
3. **Asset Browser filter bar**
- **How to enter**: open the Asset Browser from an eligible model widget
browse action, the Model Library flow, or another
`useAssetBrowserDialog` caller.
- **Affected elements**: `AssetFilterBar` controls: **File formats**,
**Base models**, **Ownership**, and **Sort by**.
- **Patch point**: `AssetBrowserModal.vue` anchors to the PrimeVue
dialog header and passes the style through `AssetFilterBar` to its
`MultiSelect` / `SingleSelect` controls.
https://github.com/user-attachments/assets/e27bd805-10c0-4b3b-97f3-9e11faa47021
4. **Asset Browser model info panel**
- **How to enter**: open Asset Browser, select an asset, then use the
right-side model info panel.
- **Affected element**: the **Model type** select in `ModelInfoPanel`.
- **Patch point**: `AssetBrowserModal.vue` reuses the same parent-scoped
style and passes it to `ModelInfoPanel` as `selectContentStyle`.
https://github.com/user-attachments/assets/5e9f7ef0-ebd7-4987-ba1b-2137c034086f
5. **Upload Model confirmation step**
- **How to enter**: open Asset Browser, click **Upload**, enter/fetch
model metadata, then proceed to the confirmation step.
- **Affected element**: the **Model type** `SingleSelect` in
`UploadModelConfirmation`.
- **Patch point**: `UploadModelConfirmation.vue` anchors within the
upload dialog content and passes `selectContentStyle` to the model type
selector.
https://github.com/user-attachments/assets/ec145f26-8621-455b-915e-bedee47e1cbd
6. **Settings > Keybinding panel controls**
- **How to enter**: open Settings from the sidebar/menu, then select the
**Keybinding** panel.
- **Affected elements**: the keybinding preset select, the preset
overflow dropdown menu, and the row context menu inside
`KeybindingPanel`.
- **Patch point**: `KeybindingPanel.vue` anchors to the settings dialog
panel and passes `keybindingOverlayContentStyle` only to those Reka
overlay roots.
## Review Focus
- Confirm the patch stays narrowly scoped to the six known PrimeVue
parent + Reka child overlay surfaces above.
- Confirm `contentStyle` reaches the actual portaled Reka overlay
content in each patched path.
- Confirm the fallback behavior preserves existing stacking when no
PrimeVue parent overlay is found; in that case the helper returns an
empty style object and leaves existing Tailwind z-index classes alone.
- Please avoid expanding this into a larger overlay refactor. The goal
is a clean, backport-friendly compatibility patch while the Reka
migration continues.
Validation performed:
- `pnpm exec vitest run src/composables/usePopoverSizing.test.ts`
- `pnpm typecheck`
- `pnpm lint` (passes with existing unrelated warnings only)
- `pnpm format:check`
- commit hook lint-staged checks (`oxfmt`, `stylelint`, `oxlint`,
`eslint --fix`, `pnpm typecheck`)
- pre-push `pnpm knip`
Linear: FE-569
## Bug Screenshots
https://github.com/user-attachments/assets/e73761af-9867-4c50-ab0d-4e32e59011e1https://github.com/user-attachments/assets/145daf4d-3268-428b-9987-1e1afd0b866f
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12038-fix-keep-Reka-overlays-above-PrimeVue-dialogs-3596d73d365081e7af49dbc4d3905962)
by [Unito](https://www.unito.io)
## Root cause
`useAssetBrowser`'s `filteredAssets` computed re-ran
`.map(transformAssetForDisplay)` over the full result set on every tab
switch. `transformAssetForDisplay` allocates fresh `badges`/`stats`
objects, walks `tags`, calls `getAssetBaseModels`, and runs i18n date
formatting per asset — none of which were memoized. Switching All /
Inputs / Outputs forced N transforms per click and produced brand-new
`AssetDisplayItem` references, which also defeated `:key`-based diffing
in `AssetGrid` / `VirtualGrid` and re-rendered every visible card.
## Fix
Memoize `transformAssetForDisplay` at module scope with a
`WeakMap<AssetItem, AssetDisplayItem>`. Unchanged assets reuse the same
display item across tab switches; the GC reclaims entries when assets
are released.
## Before / after (n=200 assets, 6 tab switches: inputs → outputs → all
→ inputs → outputs → all)
| Metric | Before | After |
| ------------------------------- | -----: | ----: |
| `transformAssetForDisplay` runs | 800 | 0 |
| Wall time (Vitest harness) | 13.2 ms | 8.1 ms |
| Reused `AssetDisplayItem` refs | 0 | 200 |
Measured via
`src/platform/assets/composables/useAssetBrowser.perf.test.ts` plus a
temporary `process.stderr.write` harness.
## Red / green
| Commit | Purpose |
| --------- | ------- |
| 7367fdd60 | test: failing perf assertions (transform budget + reused
refs) |
| 021b98ac0 | perf: WeakMap memoization of `transformAssetForDisplay` |
## Test plan
- [x] `pnpm exec vitest run
src/platform/assets/composables/useAssetBrowser` — 42/42 pass (including
2 new perf assertions)
- [x] `pnpm typecheck`
- [x] `pnpm exec eslint` on touched files
- [x] `pnpm exec oxfmt --check` on touched files
Fixes [FE-229
](https://linear.app/comfyorg/issue/FE-229/asset-browser-switching-all-inputs-outputs-tabs-is-slow)Source:
https://comfy-organization.slack.com/archives/C0A4XMHANP3/p1776716352588229
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11491-perf-memoize-asset-display-transform-across-filter-tab-switches-3496d73d36508112822dd6e7b58040fe)
by [Unito](https://www.unito.io)
---------
Co-authored-by: Christian Byrne <cbyrne@comfy.org>
*PR Created by the Glary-Bot Agent*
---
Raises `Comfy.QueueButton.BatchCountLimit` on cloud from `32` to `100`
to match the server-side `MaxQueuedJobsPerUser` cap
(`cloud/infrastructure/dynamicconfig/prod/config.json:3`). The desktop
default was already `100` and is unchanged — collapsing both branches to
the same constant.
Addresses Discord feature request: [Increase queue batch limit from
200](https://discord.com/channels/1218270712402415686/1243609826299220039/1499104231381012641).
## Change
```diff
- defaultValue: isCloud ? 32 : 100,
+ defaultValue: 100,
```
The setting is read dynamically by all batch count UIs
(`BatchCountEdit.vue`, `LinearControls.vue`).
## Why 100 (not 512)
Original ask was 200→512. Investigation showed:
- The actual previous default was `100` (desktop) / `32` (cloud), not
200.
- Cloud enforces `MaxQueuedJobsPerUser = 100` per workspace server-side.
A higher frontend cap can't unlock more queued work — extra prompts just
get rejected with `QUEUE_LIMIT`.
- Frontend submits prompts as N sequential `POST /prompt` calls (no
batched-prompt endpoint), so the UI cap is purely about how many clicks
it takes — not throughput.
- Going from 32 → 100 lets cloud users match the server cap in one click
instead of 4. No new behavior is unlocked.
## Known limitation (pre-existing, not introduced here)
The new max equals the absolute server cap, not the user's remaining
capacity. A user with already-queued work can hit `QUEUE_LIMIT`
mid-batch. The pre-existing 32 limit had the same shape (just at a
smaller scale); deriving the UI max from `cap - outstanding` would
require polling and reactive state and is out of scope for a one-line
setting bump.
## Verification
- `pnpm typecheck` — passes
- `pnpm lint` — 0 errors (1 pre-existing warning in unrelated test file)
- `pnpm test:unit` — `BatchCountEdit.test.ts` (3 tests) +
`src/platform/settings/**` (70 tests) all pass
- **Manual (Playwright)**:
- `settingStore.get('Comfy.QueueButton.BatchCountLimit')` returns `100`
at runtime
- Typing `999` into the batch count widget clamps to `100`
- Increment button is disabled at `100` (max reached)
## Screenshots
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11876-feat-align-cloud-batch-count-limit-with-server-side-queue-cap-3566d73d3650819b8d01dbf83d1a8e49)
by [Unito](https://www.unito.io)
Co-authored-by: Glary-Bot <glary-bot@users.noreply.github.com>
Replace PrimeVue components in 3D node viewer controls with the
project's Reka UI equivalents across 7 files.
## Changes
| File | Replaced |
|------|---------|
| `AnimationControls.vue` | `Select` × 2 (speed + animation) |
| `ViewerModelControls.vue` | `Select` × 2 (up direction + material
mode) |
| `ViewerCameraControls.vue` | `Select` + `Slider` (camera type + FOV) |
| `ViewerExportControls.vue` | `Select` (export format) |
| `PopupSlider.vue` | `Slider` |
| `ViewerLightControls.vue` | `Slider` |
| `ViewerSceneControls.vue` | `Checkbox` → native `<input
type="checkbox">` |
## Implementation notes
- `Select` uses `@/components/ui/select/*` compound components. Numeric
model values (animation speed index) are stringified at the binding
boundary and converted back on update, matching Reka `SelectRoot`'s
`string`-only `modelValue` contract.
- `Slider` uses `@/components/ui/slider/Slider.vue`. Single-number
`defineModel` values are wrapped in a `computed` array and unwrapped in
the update handler, following the pattern established in
`LightControls.vue`.
- No new Reka UI wrapper components were created — existing ui/select
and ui/slider primitives were used directly.
## Test
https://github.com/user-attachments/assets/afca0fc8-a7b6-49ee-b221-ee5725bd127e
1. AnimationControls.vue
- **Add Load3D node** → Upload an animated GLB file (e.g., a character
model).
- **Node preview top bar:** Play/Pause button, speed dropdown, animation
name dropdown, and progress bar.
2. PopupSlider.vue
- **Hover over Load3D preview:** Icon buttons appear in the left
toolbar.
- **"Light Intensity" button (bulb icon)** → Slider pops up on the
right.
- **"FOV" button (view icon)** → Slider pops up on the right.
3. ViewerCameraControls.vue
- **Load3D node** → Settings panel (top-right) → **"Camera"** tab.
- **Features:** Camera type dropdown (Perspective / Orthographic), FOV
slider (visible in Perspective mode).
4. ViewerExportControls.vue
- **Settings panel** → **"Export"** tab.
- **Features:** Format dropdown (GLB / OBJ / STL), Export button.
5. ViewerLightControls.vue
- **Settings panel** → **"Light"** tab.
- **Features:** Light intensity slider.
6. ViewerModelControls.vue
- **Settings panel** → **"Model"** tab.
- **Features:** "Up direction" dropdown, Material mode dropdown
(Wireframe / Normal, etc.).
7. ViewerSceneControls.vue
- **Settings panel** → **"Scene"** tab.
- **Features:** Background color picker, "Show grid" checkbox, upload
background image button.
<!-- CURSOR_SUMMARY -->
---
> [!NOTE]
> **Medium Risk**
> UI component swap touches multiple interactive viewer controls
(selects/sliders/checkbox), so small binding/typing differences (string
vs number, array slider values) could cause subtle regressions despite
test updates.
>
> **Overview**
> Replaces PrimeVue `Select`, `Slider`, and `Checkbox` usages across
Load3D viewer controls with the project’s Reka UI-based primitives
(`@/components/ui/select/*`, `@/components/ui/slider/Slider.vue`) and a
native checkbox.
>
> Updates v-model wiring to match the new components’ contracts: selects
now bind via string `modelValue` with explicit number casting where
needed, and sliders now wrap single numeric values into `[number]`
arrays with corresponding update handlers. Unit tests are updated to
mock the new UI components and their updated event/value shapes.
>
> <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit
46f99db256. Bugbot is set up for automated
code reviews on this repo. Configure
[here](https://www.cursor.com/dashboard/bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12020-refactor-load3d-replace-PrimeVue-Select-Slider-Checkbox-with-Reka-UI-3586d73d365081f58601d93031016afd)
by [Unito](https://www.unito.io)
## Summary
Make `userStore.initialize()` idempotent and concurrency-safe so the
bootstrap, router-guard, and UserSelectView callers share a single
getUserConfig fetch instead of racing/duplicaitng calls.
## Changes
- **What**:
- cache initialize in a promise so callers all re-use the same result
- remove now redundant is initialized guard
- tests
## Review Focus
- Current user switch/logout uses `window.location.reload()`, no callers
intentionally call initialize to reinit. In future if this changes we
may want to add a parameter to skip the cache or a separate function.
- Failed initializes are not cached to allow callers to retry
- Not practical for e2e tests, the unit tests prove that the requests
are deduped. All a e2e test would do is mock/spy on the network requests
to show multiple requests do not happen - which the unit tests do a
better job of.
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11959-fix-add-guard-to-prevent-user-store-re-initialization-3576d73d3650817db7b0e52cc25f9b7b)
by [Unito](https://www.unito.io)
## Summary
- Defer thumbnail capture until camera state is restored via new
modelReady event so captureThumbnail no longer races with the saved
view, fixing the "snap back to default on hover" regression.
- Repaint the live scene at the end of captureThumbnail so the canvas is
not left with the offscreen mask/normal pass when the render loop is
gated.
- Persist post-fitToViewer model.scale + model.position into the
existing modelConfig.gizmo slot so a refresh reapplies them via the
existing applyGizmoConfigToLoad3d path; rotation stays owned by
upDirection.
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11944-FE-537-fix-load3d-preserve-camera-view-fit-transform-and-first-frame-paint-after-re-3576d73d365081429653ea4740612617)
by [Unito](https://www.unito.io)
Adds targeted behavioral tests for the slimmed `useBrushDrawing`
orchestration composable (Phase E of the brush-drawing refactor).
## Changes
- 5 new tests covering previously untested branches:
- `compositeStroke` receives `isRgb=true` when active layer is `rgb`
- `compositeStroke` receives `isErasing=true` when tool is `eraser`
- Mask canvas opacity is restored after drawing on the mask layer
- `globalCompositeOperation` is set to `destination-out` during
`handleDrawing` when tool is eraser
- `globalCompositeOperation` is set to `destination-out` during
`handleDrawing` when right mouse button is held
## Coverage (useBrushDrawing.ts)
| Metric | Before | After |
|--------|--------|-------|
| Statements | 86.33% | 87.05% |
| Branches | 68.75% | 70.00% |
| Functions | 90.00% | 90.00% |
| Lines | 89.23% | 90.00% |
All 18 tests pass. GPU paths remain `/* c8 ignore */` excluded
(untestable without WebGL).
- Fixes#0
<!-- CURSOR_SUMMARY -->
---
> [!NOTE]
> **Low Risk**
> Low risk: changes are limited to unit tests, adding coverage for
eraser/right-click composition and `drawEnd` GPU compositing/opacity
restoration paths without altering production logic.
>
> **Overview**
> Adds new `useBrushDrawing` test cases to cover previously untested
branches: setting `globalCompositeOperation` to `destination-out` during
`handleDrawing` when erasing (tool or right-click), and verifying
`drawEnd` passes correct `isRgb`/`isErasing` flags to
`gpu.compositeStroke`.
>
> Also asserts mask-layer opacity is restored after `drawEnd`,
increasing behavioral coverage around stroke completion and canvas
visibility cleanup.
>
> <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit
20c411e6ce. Bugbot is set up for automated
code reviews on this repo. Configure
[here](https://www.cursor.com/dashboard/bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12001-test-maskeditor-expand-useBrushDrawing-behavioral-coverage-3586d73d365081388ebcef91c2172c0a)
by [Unito](https://www.unito.io)
## Summary
Preview as Text (`PreviewAny`) nodes were re-executing on every prompt
submission because the rendered preview text was being echoed back to
the backend as input values, mutating the cache signature.
## Changes
- **What**: Set `widget.options.serialize = false` on the three widgets
the `Comfy.PreviewAny` extension adds (`preview_markdown`,
`preview_text`, `previewMode`) so they are excluded from the API prompt
sent to the backend.
## Root cause
The extension was setting `widget.serialize = false`, which only
controls **workflow JSON** persistence (checked in
`LGraphNode.serialize`). The **API prompt** serializer in
`executionUtil.graphToPrompt` checks `widget.options.serialize` instead
— a distinct property documented in litegraph's `WIDGET_SERIALIZATION`
convention.
After `onExecuted` writes the rendered text into the widget value, the
next `graphToPrompt` call serialized that text into
`inputs.preview_text` / `inputs.preview_markdown`. The backend cache
signature in `comfy_execution/caching.py` hashes all keys in
`node["inputs"]`, so the changing text invalidated the cache and forced
a redundant execution every time.
## Review Focus
Two commits, red-green TDD:
1. `test:` failing unit + e2e tests asserting the desired behavior.
2. `fix:` adds `options.serialize = false` to make them pass.
Tests verify the widgets are excluded from the API prompt; e2e
additionally simulates a prior execution populating widget values to
mirror the real bug condition.
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12010-fix-stop-PreviewAny-widgets-from-triggering-re-execution-3586d73d3650810585cdd077f3ac64f5)
by [Unito](https://www.unito.io)
*PR Created by the Glary-Bot Agent*
---
## Summary
Adds 22 Playwright E2E tests for the keybinding settings panel
(`KeybindingPanel.vue`), covering all 15 previously-untested functions
identified via coverage analysis.
## Test Groups
| Group | Tests | Functions Covered |
|---|---|---|
| Row Expansion | 2 | `handleRowClick`, `toggleExpanded`, `expandedRows`
|
| Double-Click | 2 | `handleRowDblClick`, `addKeybinding`,
`editKeybinding` |
| Context Menu | 7 | `handleRowContextMenu`, `clearContextMenuTarget`,
`ctxChangeKeybinding`, `ctxAddKeybinding`, `ctxResetToDefault`,
`ctxRemoveKeybinding` |
| Action Buttons | 7 | `editKeybinding`, `resetKeybinding`,
`removeSingleKeybinding`, `handleRemoveAllKeybindings`,
`handleRemoveKeybindingFromMenu` |
| Expanded Row Actions | 2 | `editKeybinding` (expansion),
`removeSingleKeybinding` (expansion) |
| Reset All | 2 | `resetAllKeybindings` (confirm + cancel) |
| Search Filter | 1 | `watch(filters, ...)` clears expansion |
## Flake Prevention Measures
- Deterministic test command
(`TestCommand.KeybindingPanelE2E.NoBinding`) registered via
`app.registerExtension()` for 0-binding scenarios — avoids flaky
pagination-dependent row scanning
- `pressComboOnInput()` asserts input focus before pressing key combos
in the edit dialog
- `saveAndCloseKeybindingDialog()` waits for dialog teardown to complete
before proceeding
- `openContextMenu()` waits for Reka UI menu items to be visible before
interacting
- Resets `Comfy.Keybinding.NewBindings` and
`Comfy.Keybinding.UnsetBindings` in `afterEach`
## Note on Selectors
Some locators use Tailwind utility classes (`.pl-4`,
`.icon-[lucide--chevron-right]`) because the expansion template and
chevron icon in `KeybindingPanel.vue` lack `data-testid` attributes.
Adding test IDs would be a follow-up to this test-only PR.
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11455-test-add-E2E-tests-for-keybinding-settings-panel-coverage-3486d73d365081d7a902fc68091552f2)
by [Unito](https://www.unito.io)
---------
Co-authored-by: Glary-Bot <glary-bot@users.noreply.github.com>
Co-authored-by: Benjamin Lu <benjaminlu1107@gmail.com>
## Summary
`assetsStore.updateAssetTags` issues `removeAssetTags` and
`addAssetTags` as two separate network calls. When the remove succeeds
server-side but the subsequent add rejects, the cache rolls back to the
original tags while the server has already dropped the removed tags —
cache and backend diverge until the next refetch.
This adds a compensating action: if remove succeeded and add then fails,
attempt to re-add the just-removed tags so the server returns to its
prior state. If the compensating add also fails, invalidate the category
cache so the next access refetches fresh state.
- Fixes#11694
- Fixes
[FE-473](https://linear.app/comfyorg/issue/FE-473/fix-converge-asset-tag-cache-and-server-on-partial-failure)
## Changes
- `src/stores/assetsStore.ts`: track which tags were already removed
server-side; on add-failure, re-add them; if compensation fails,
invalidate the resolved category cache via `resolveCategory` +
`invalidateCategory`.
- `src/stores/assetsStore.test.ts`: extend the cloud asset-service mock
with `addAssetTags` / `removeAssetTags` and add a `updateAssetTags
partial-failure compensation` describe block:
- re-adds removed tags when add fails so cache and server converge
- invalidates the category cache when compensation also fails
- does not attempt compensation when only the add was attempted (no
remove ran)
## Test plan
- [x] `pnpm exec vitest run src/stores/assetsStore.test.ts` — 42 passed
(39 prior + 3 new)
- [x] `pnpm typecheck`
- [x] `pnpm lint`
- [x] `pnpm knip`
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11695-fix-converge-asset-tag-cache-and-server-on-partial-failure-34f6d73d365081149900f95b6ee4bfa9)
by [Unito](https://www.unito.io)
Primitive nodes do not create their widgets until their
`onAfterGraphConfigured` method is called. Previously, when a
subgraphNode contains a linked widget, any proxyWidgets that do not
resolve to a real widget are pruned at time of configure. Since this
occurs prior to initialization of the primitive, the primitive value
would be de-promoted before the widget could initialize
This is resolved by the minimally disruptive change of allowing proxied
promotions to primitive nodes to be kept so long as the node itself can
be found.
## Summary
Aligns the `OpenAIGPTImage1` node display name in all 11 non-English
`nodeDefs.json` locale files with the English source-of-truth, which was
already updated to `OpenAI GPT Image 2`.
## Changes
- **What**: Updates `display_name`, the description string, and the
prompt tooltip in `ja`, `ru`, `zh`, `zh-TW`, `ar`, `pt-BR`, `ko`, `fr`,
`es`, `fa`, and `tr` locales from `GPT Image 1` to `GPT Image 2` (and
`GPT Görüntü 1` → `GPT Görüntü 2` in Turkish, `GPT صورة 1` → `GPT صورة
2` in Arabic). Other tooltips already referenced `GPT Image 2` and are
unchanged.
- **Breaking**: None — the registry node id `OpenAIGPTImage1` is
preserved (it is an internal identifier, not user-facing).
## Review Focus
- Translations were updated mechanically — please confirm the
version-number change is acceptable as-is for non-Latin scripts (Arabic,
Persian, Korean, Japanese, Chinese) where the version number was kept as
`2` per the existing pattern.
- The English locale already used `OpenAI GPT Image 2`, so this PR
brings the other locales into sync; no English copy was changed.
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11968-fix-i18n-rename-OpenAI-GPT-Image-1-to-GPT-Image-2-across-locales-3576d73d365081bfa204cbf528d84bf3)
by [Unito](https://www.unito.io)
Co-authored-by: Marwan Ahmed <marwan@Marwans-MacBook-Pro.local>
## Summary
Restores required-input validation highlighting on Vue node input slots.
## Changes
- **What**: Passes validation error state from `NodeSlots` to
`InputSlot` using node locator IDs, including subgraph and nested
subgraph execution IDs.
- **What**: Adds unit coverage for root, one-level subgraph, and nested
subgraph slot error mapping.
- **What**: Adds a Vue Nodes screenshot regression test that asserts the
missing required input slot itself receives the error highlight.
- **Dependencies**: None.
## Review Focus
- Required input errors on Vue-rendered node's slots.
- The new Playwright screenshot expectation will need the `New Browser
Test Expectation` label for Linux baseline generation.
## Screenshots (if applicable)
Before
<img width="499" height="324" alt="스크린샷 2026-05-05 오후 3 00 44"
src="https://github.com/user-attachments/assets/285fdf91-6d7e-480b-99b9-715705f78914"
/>
After
<img width="482" height="356" alt="스크린샷 2026-05-05 오후 3 01 11"
src="https://github.com/user-attachments/assets/51b8db49-eb9c-4155-8aa5-109c0bd7699b"
/>
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11950-fix-highlight-missing-input-slots-on-Vue-nodes-3576d73d365081bd85bfd1ea149d45c5)
by [Unito](https://www.unito.io)
---------
Co-authored-by: github-actions <github-actions@github.com>
