Draft releases don't create git tags, but publish-pypi workflow
waits for the tag to exist. This caused core/* releases to hang
indefinitely since they were always drafts.
Change: only use draft for prereleases (alpha/beta/rc), publish
all stable releases (main and core/*) immediately.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
## Summary
Move run-button context assembly out of telemetry providers so telemetry
can initialize without importing app-mode/workspace state.
## Changes
- **What**: Providers now accept completed `RunButtonProperties`;
run-button call sites use a workspace composable to build that payload.
- **Dependencies**: None.
## Summary
- Migrate the website's top nav from bespoke components (`SiteNav`,
`MobileMenu`, `NavDesktopLink`, `PillButton`, `MaskRevealButton`) to
shadcn-vue primitives (`NavigationMenu`, `Sheet`, `Button`), split into
`HeaderMain` → `HeaderMainDesktop` + `HeaderMainMobile`.
- Mobile nav becomes a `Sheet` with drill-down sub-navigation, scroll
lock, sticky CTAs, sr-only i18n title/description, and a back-to-home
logo.
- Desktop nav uses `NavigationMenu` with shared viewport, featured
cards, `NavColumn` extraction, and centralized nav data in
`data/mainNavigation.ts`.
- Adds i18n strings for nav labels, dropdown column headers, close/back
affordances, and the mobile menu description.
## Test plan
- [ ] Desktop: hover PRODUCTS / COMMUNITY / COMPANY — dropdowns open
with featured card + columns, NEW badges render, external links show the
arrow-up-right icon, viewport is shared between triggers.
- [ ] Mobile (<lg): open hamburger sheet — verify logo + close, body
scroll is locked while open, top-level nav scrolls if it overflows, CTAs
stay pinned at bottom.
- [ ] Tap COMMUNITY / PRODUCTS / COMPANY — sub-panel slides over root
nav, in-sheet BACK returns to root, links navigate correctly.
- [ ] Reload `?locale=zh-CN`: dropdown labels, sheet title/description,
BACK / close affordances all localized.
- [ ] No regressions on `/cloud`, `/cloud/pricing`, `/customers`, etc. —
pages that swap CTAs (`BrandButton` → shadcn `Button`) still render at
every breakpoint.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: GitHub Action <action@github.com>
Co-authored-by: github-actions <github-actions@github.com>
Co-authored-by: Alexander Brown <drjkl@comfy.org>
Non preview, `canvasOnly` widgets like `control_after_generate` could be
displayed in the subgraph editor even though promoting them would have
no visual or functional effect when in vue mode.
In vue mode, these entries are now hidden from the list of candidate
items for promotion to reduce confusion.
## Summary
Account preconditions (sign-in / subscription / credits) on running a
workflow now open their modal directly and stay out of the error panel +
error count — previously `subscription_required` fell through to a red
"1 ERROR — Subscription required to queue workflows" banner. This covers
**both** paths: the `execution_error` websocket event and the `POST
/prompt` 402 queue paywall (`{ type: "PAYMENT_REQUIRED", message:
"Subscription required to queue workflows" }`), which is the exact
payload reported in #12840.
## Changes
- **What**: `execution_error` is classified by a pure
`accountPreconditionRouting` resolver (precedence sign-in > subscription
> credits) and routed to the existing modal via
`useAccountPreconditionDialog`; `executionStore` returns early for
preconditions so they never populate `lastExecutionError` /
`lastPromptError` / `lastNodeErrors` → fully excluded from the panel and
`totalErrorCount`. Runtime credit error at a node → credits modal (out
of panel; can name the node).
- **Queue paywall**: the `queuePrompt` catch resolves the same
precondition from the `POST /prompt` 402 response and opens the modal,
short-circuiting before `lastPromptError`, so the queue paywall stays
out of the panel too. The runtime matcher learns the `"Subscription
required to queue workflows"` message.
- **Breaking**: none.
## Before / After
Free-tier queue paywall (`POST /prompt` → 402) on a cloud build:
**Before** — raw `Subscription required to queue workflows` surfaced in
the error panel (no actionable upgrade):
<img width="1600" height="873" alt="before-error-panel"
src="https://github.com/user-attachments/assets/1b76b742-16bf-47e3-9245-17e35f8f1e70"
/>
**After** — clean subscription modal opens; nothing in the error panel
or error count:
<img width="1600" height="873" alt="after-subscription-modal"
src="https://github.com/user-attachments/assets/13d238cb-20bf-4795-a530-5abcf9968dc7"
/>
## Review Focus
- **Routing-only — the run button is intentionally untouched.** The
original AC#3 ("no Subscribe-to-Run button") is superseded by the FE-978
run-lock decision (pre-emptive role-aware lock, Figma 3253-18671).
Complements #12786 (FE-978 run-lock); disjoint file sets.
- Tests: `accountPreconditionRouting` / `useAccountPreconditionDialog` /
`executionStore` — each precondition routes to its modal and is excluded
from the panel/count; precedence resolves on co-occurrence. Plus
Playwright `browser_tests/tests/subscriptionPaywallError.spec.ts` — the
queue paywall (402) stays out of the error panel, with a control
asserting ordinary queue errors still surface. typecheck / oxlint /
eslint / stylelint / oxfmt / knip clean.
Fixes FE-878
Fixes#12840
## Summary
Adds `@comfyorg/comfyui-desktop-bridge-types` as a workspace package in
the frontend monorepo and changes the frontend app dependency to
`workspace:*`.
Adds a dedicated `Publish Desktop Bridge Types` workflow for publishing
`packages/comfyui-desktop-bridge-types` by its own package version,
without coupling it to the generated `@comfyorg/comfyui-frontend-types`
release. The generated frontend types package still emits a concrete
`@comfyorg/comfyui-desktop-bridge-types@0.1.2` dependency instead of
leaking workspace/catalog protocol references.
The Desktop2 missing-model path uses `window.__comfyDesktop2.isRemote()`
when available, but falls back to the legacy
`window.__comfyDesktop2Remote` marker so frontend rollout stays
compatible with older Desktop builds.
Paired Desktop PR: https://github.com/Comfy-Org/Comfy-Desktop/pull/1112
## Summary
Remove unused `export` keyword from `ExportFormatOption` interface. The
interface is only used internally in `constants.ts` and is not imported
elsewhere.
Fixes knip "unused exported types" error.
Co-authored-by: Connor Byrne <c.byrne@comfy.org>
## Summary
Removes the deprecated Group Nodes feature and replaces it with a
load-time migration that auto-converts any group nodes in a loaded
workflow into Subgraphs (with accepted lossiness).
## Changes
- **What**:
- `groupNode.ts` is now a migration-only extension.
`beforeConfigureGraph` registers temporary node types from
`extra.groupNodes` so instances are created during `configure`; a new
`afterConfigureGraph` hook converts every group node in the root graph
to a subgraph (via `LGraph.convertToSubgraph`), re-scanning until none
remain, then deletes `extra.groupNodes`. A failed conversion removes the
offending node so loading never hangs or breaks.
- Kept the minimum needed: `GroupNodeConfig` (builds the
input/output/widget maps), a slimmed `GroupNodeHandler` exposing a
rewritten `convertToNodes()` that no longer depends on the execution
DTOs, the `globalDefs`/`addCustomNodeDefs` path, and the `nodeDefStore`
`Object.assign` shim the migration relies on to detect group nodes.
- Deleted: the Manage Group Nodes dialog (`groupNodeManage.ts`/`.css`),
execution DTOs (`executableGroupNodeDto.ts`,
`executableGroupNodeChildDTO.ts`), the create/builder flow, recreate,
commands, keybindings, menus, the `isGroupNode` branches in the
right-side panel / error grouping / focus composable, the group-node
branches in node templates, dead i18n keys, and the now-unused
`serialise` clipboard helper.
- Rewrote `browser_tests/tests/groupNode.spec.ts` to assert
auto-conversion; deleted the `ManageGroupNode` page object and
`manageGroupNode()` helper.
- Net: ~2,700 lines removed across 23 files (7 files deleted).
- **Breaking**: Group nodes can no longer be created, managed, or
executed. Existing workflows still load — their group nodes are
converted to subgraphs on open.
## Review Focus
- The load-time migration in `afterConfigureGraph` and the rewritten
`GroupNodeHandler.convertToNodes()` (no longer uses the execution
`getInnerNodes()` / DTOs; derives inner node type/index from
`groupData.nodeData.nodes` and relies on `deserialiseAndCreate` +
selection ordering).
- Kept `nodeDefStore`'s `Object.assign(this, obj)` shim: the migration
depends on it to propagate the group-node marker symbol onto the
registered node definition.
### Accepted lossiness
- Group nodes nested inside subgraphs (or inside other group nodes)
convert into the root graph rather than their original container —
essentially nonexistent in real legacy workflows since group nodes
predate subgraphs.
- Temporary `workflow>name` node types stay registered for the session;
instantiating one auto-converts it to a subgraph.
## Verification
`pnpm typecheck`, `typecheck:browser`, `knip`, `oxlint`, `eslint`, and
`oxfmt` are green (also enforced by pre-commit hooks). Unit tests for
the touched files could not be run locally due to a pre-existing
environment error (`file:///assets/images/*.svg` passed to a Node
filename API at import time, which also fails on unmodified test files);
the browser spec requires a live server.
---------
Co-authored-by: Amp <amp@ampcode.com>
Co-authored-by: github-actions <github-actions@github.com>
Co-authored-by: AustinMroz <austin@comfy.org>
Co-authored-by: GitHub Action <action@github.com>
## Summary
Cloud users get yanked to `/cloud/survey` mid-workflow with no user
action. The redirect is **downstream of auth**: when the Cloud token is
briefly stale (token rotation / auth-refresh / reconnect window), the
authenticated survey-status check 401s, and the gate turned that
transient 401 into "survey not completed" → redirect.
Surveys are currently disabled for everyone on cloud via dynamicconfig
as the live mitigation. This PR lets us re-enable them safely
**without** waiting on the auth rework.
## Root cause
`getSurveyCompletedStatus()` returned `false` ("not completed") on
**any** non-200 — including a transient 401/403/5xx or network error —
and consumers treat `false` as a redirect to the survey. So a
stale-token 401 (or the page force-reload a 401 triggers in
`GraphCanvas`) bounced a working, already-onboarded user to the survey.
The real root cause of the transient 401s is a separate, still-open
effort: **FE-963** (reactive 401 re-mint + single retry), **FE-950/951**
(unified Cloud JWT), **BE-1125**. This PR does **not** fix those; it
stops the survey from being their user-visible casualty.
## The fix
`getSurveyCompletedStatus` now distinguishes the responses instead of
failing closed on all of them:
- **404** → not completed (show survey). This is the genuine signal: the
cloud backend (`GetSettingById`) returns 404 for a survey key that was
never stored, and a 404 is only reachable after a successful
authenticated read (a stale token 401s, never 404s), so it can't be a
transient false signal.
- **transient 401/403/5xx/network** → treat as completed (fail-safe), so
a working user is never bounced.
- **200** → completed iff `value` is non-empty (unchanged).
**No router change.** The `/` onboarding guard is untouched (router.ts
matches main), so the existing UX is preserved — a not-completed user is
still gated to the survey on load; only the spurious transient-failure
bounce is removed.
## Why #12301 was reverted, and how this differs
#12301 shipped a **blanket** fail-safe (`!response.ok → true`, 404
included), which made the survey unreachable for genuinely-not-completed
users (404 → "completed") and was reverted in #12344. This PR
special-cases **404 as the real not-completed signal** and fails safe
only on transient/ambiguous responses, so onboarding still works.
## Tests
- **Unit** (`auth.test.ts`): 200 non-empty → true; 200 empty / `null` /
missing `value` key → false; **404 → false**; 401/403/500/network →
true.
- **E2E** (`browser_tests/tests/cloudSurveyGate.spec.ts`, `@cloud`): a
transient 401 on `/` does **not** bounce a working user; a genuine 404
on `/` **does** route to the survey.
Linear: FE-739. Root cause (separate): FE-963 / FE-950 / FE-951 /
BE-1125.
## Summary
- **Pricing copy** — replace "concurrent API jobs" wording with
"workflows via API" on Creator/Pro plans, add a third feature line to
Standard, and add a new "Run Workflows via API" entry to the "What's
Included" section (existing feature11 → feature12).
- **Tutorial captions** — wire per-locale VTT caption tracks into
`TutorialDetailDialog`, so the CC button in `VideoPlayer` is now
functional for the learning tutorials. `LearningTutorial.caption` is
typed as `readonly VideoTrack[]`; `VideoTrack` is now exported from
`VideoPlayer.vue`.
- Tailwind class ordering in `PriceSection.vue` shifted due to the
formatter on touched lines.
## Test plan
- [ ] `/pricing` page (EN + zh-CN): Standard plan shows the new "1
workflow via API" line; Creator/Pro show "3/5 workflows via API";
"What's Included" lists "Run Workflows via API" above "Parallel job
execution".
- [ ] Open a learning tutorial → CC button is visible in the player →
toggling CC shows English captions in sync with playback.
- [ ] Tutorials without a `caption` entry hide the CC button.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: github-actions <github-actions@github.com>
## Summary
Fixes FE-485.
This updates workspace auth refresh handling so stale in-flight refresh
responses cannot overwrite a newer workspace context, and exhausted
transient token exchange failures preserve the existing workspace
context while its token is still valid.
## Changes
- Add commit-time request-id guards before `switchWorkspace` writes
workspace state, workspace token, `error`, or `sessionStorage`.
- Track the current workspace token expiry in memory and use it to
distinguish safe transient refresh failures from failures that must
clear context.
- Convert the stale refresh race coverage from expected-failing to a
normal passing regression test.
- Update transient retry coverage to assert valid context and
`sessionStorage` preservation.
## Browser / E2E coverage
No Playwright test was added because this bug is in the Pinia store race
between mocked token-exchange promises, request IDs, token expiry, and
`sessionStorage` commits. The deterministic unit spec directly controls
the ordering that is not practical to force through the browser without
real auth/session infrastructure and artificial network timing hooks.
## Validation
- `pnpm format -- src/platform/workspace/stores/workspaceAuthStore.ts
src/platform/workspace/stores/useWorkspaceAuth.test.ts`
- `pnpm exec vitest run
src/platform/workspace/stores/useWorkspaceAuth.test.ts`
- `pnpm exec eslint src/platform/workspace/stores/workspaceAuthStore.ts
src/platform/workspace/stores/useWorkspaceAuth.test.ts`
- `pnpm exec oxlint src/platform/workspace/stores/workspaceAuthStore.ts
src/platform/workspace/stores/useWorkspaceAuth.test.ts --type-aware`
- `pnpm exec vue-tsc --noEmit --pretty false`
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11726-fix-guard-workspace-auth-refresh-races-3506d73d365081b99df3c1bf3d0e008a)
by [Unito](https://www.unito.io)
---------
Co-authored-by: bymyself <cbyrne@comfy.org>
When right clicking an output asset from the assets sidebar panel, the
'insert as node in workflow' action was twice bugged
- The default type, as used for determining filename annotation, was set
to the type of the file. This meant that annotations would never be
applied to the filename
- `temp` outputs would incorrectly be assigned the `output` type.
- My fix for this one gives me a slightly bad taste in my mouth. Parsing
URLs isn't great, but it's cleaner than needing to scan the (potentially
sparse) full outputs to try and find the corresponding output.
## Summary
Update the default workflow to use a more modern model than SD1.5. This
new workflow uses Z-Image Turbo and is the same workflow as the one in
the README for consistency.
## Changes
- **What**: `src/scripts/defaultGraph.ts`
## Screenshots (if applicable)
<img width="1920" height="1152"
alt="{2DD28B9F-A9E7-4DD7-8F07-AF7241F5702E}"
src="https://github.com/user-attachments/assets/6e6ee298-a786-4a8c-adf3-6452df08a995"
/>
---------
Co-authored-by: Connor Byrne <c.byrne@comfy.org>
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
While a file is uploading to any of the file picker nodes (ie "Load
Image"), the 'select folder' icon is replaced with a loading spinner.
I had previously implemented this with a full progress bar indicating
the rate of upload, but found it particularly uninformative when used on
cloud.
<img width="659" height="494" alt="image"
src="https://github.com/user-attachments/assets/6d3ca82b-360b-44bc-b123-b276aae2c4d6"
/>
## Summary
- Merges the separate Playwright and Storybook bot comments into the
existing `<!-- COMFYUI_FRONTEND_PR_REPORT -->` unified comment, using
named sections via `upsert-comment-section`
- Each workflow independently upserts its own section without clobbering
others
- `pr-report.yaml` cleans up legacy `<!-- PLAYWRIGHT_TEST_STATUS -->`
and `<!-- STORYBOOK_BUILD_STATUS -->` comments on first run
- Perf report: collapses detailed metric tables in `<details>` by
default — only headline FPS/TBT/heap summary and a brief regression
count are visible inline
## Before → After
**Before:** 4–5 separate bot comments per PR (Playwright, Storybook,
Bundle/Perf/Coverage, Chromatic URLs) + CodeRabbit + Codecov
**After:** 1 unified comment with collapsible sections:
```
<!-- COMFYUI_FRONTEND_PR_REPORT -->
<!-- section:playwright:start --> ... <!-- section:playwright:end -->
<!-- section:storybook:start --> ... <!-- section:storybook:end -->
<!-- section:ci-metrics:start --> ... <!-- section:ci-metrics:end -->
<!-- section:chromatic:start --> ... <!-- section:chromatic:end -->
```
(CodeRabbit and Codecov are external and can't be merged)
## How it works
`upsert-comment-section` already existed for the website CI comment.
This PR extends that pattern to the main PR report comment:
1. **Playwright** — shell script writes to `SUMMARY_FILE` when set, CI
workflow uses `upsert-comment-section` with `section-name: playwright`
2. **Storybook** — same pattern, `section-name: storybook`
3. **Bundle/Perf/Coverage** — `pr-report.yaml` now also uses
`upsert-comment-section` instead of replacing the whole comment,
`section-name: ci-metrics`
4. **Chromatic** — `section-name: chromatic`, no script change needed
Both fork and non-fork PR paths updated.
## Perf report noise fix
Previously the full per-metric regression table was shown inline. Now:
- **Visible:** headline summary (avg FPS, P5 FPS, TBT, heap per test) +
"⚠️ N regressions detected" count
- **Collapsed by default:** regression detail table, all-metrics table,
historical variance, trend data
## Test plan
- [ ] Open a PR with frontend changes — confirm single unified bot
comment appears
- [ ] Confirm Playwright section appears with ⏳ then updates to results
- [ ] Confirm Storybook section appears only when storybook-relevant
files change
- [ ] Confirm legacy standalone Playwright/Storybook comments are
deleted by `pr-report.yaml`
- [ ] Confirm perf report shows headline only, details collapsed
- [ ] Test with a fork PR — confirm fork path also works
🤖 Generated with [Claude Code](https://claude.ai/claude-code)
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-11935-ci-consolidate-PR-bot-comments-into-one-unified-comment-3566d73d36508173964cdf596cfae1f8)
by [Unito](https://www.unito.io)
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Connor Byrne <c.byrne@comfy.org>
Co-authored-by: Alexander Brown <drjkl@comfy.org>
## Summary
Removes LiteGraph's vestigial `use_uuids` node-id mode, which was never
enabled anywhere in the codebase.
## Changes
- **What**: Deletes the `LiteGraph.use_uuids` flag and the node-id
branches that read it. Newly created/added/cloned nodes now always
receive integer ids. The `asSerialisable` sort guard drops its
`@ts-expect-error` in favor of an explicit `Number(a.id) - Number(b.id)`
comparison.
- **Breaking**: `LiteGraph.use_uuids` is removed from the public
`LiteGraph` global. It defaulted to `false` and nothing ever set it to
`true`, so behavior is unchanged for all real usage — but the property
no longer exists on the global surface. Litegraph changes can affect
downstream custom-node repos.
`createUuidv4` / `LiteGraph.uuidv4` are intentionally kept — still used
for subgraph ids, slot ids, and clipboard remapping.
## Review Focus
- Confirm hard removal (vs. deprecate-then-remove no-op getter) is
acceptable for the public `LiteGraph` surface.
- `asSerialisable` sort now coerces ids via `Number()`; all serialised
node ids are numeric, so ordering is unchanged.
## Notes
- 2 litegraph test suites (`LGraphCanvas.clipboard.test.ts`,
`SubgraphWidgetPromotion.test.ts`) fail at import with a pre-existing
asset-URL error unrelated to this change (confirmed they fail with these
changes stashed). All 909 tests that ran pass.
Co-authored-by: Amp <amp@ampcode.com>
## Summary
Stop Vitest from crashing on Windows when a unit test imports a Vue
component whose template references an absolute public-asset URL (e.g.
`<img src="/assets/images/default-template.png">`).
## Changes
- **What**: Pass `template.transformAssetUrls.includeAbsolute: false` to
`@vitejs/plugin-vue` under Vitest only. With no dev server, the plugin
forces `includeAbsolute: true`, compiling absolute public-asset URLs
into module imports. Vite's module runner then derives a rootless
`import.meta.url` of `file:///assets/...`, which has no drive letter on
Windows, so Vitest's `createRequire(import.meta.url)` throws
`ERR_INVALID_ARG_VALUE` and aborts collection of any test transitively
importing such a component. Disabling the absolute transform under tests
keeps these as plain string literals (identical rendered output).
- **Breaking**: None — change is scoped to `process.env.VITEST`; dev and
production builds are unchanged.
## Review Focus
`vite.config.mts` — the `vuePluginOptions` const is `undefined` outside
Vitest, so only the test runner is affected.
Co-authored-by: Amp <amp@ampcode.com>
## Summary
Remove unused `export` keyword from `Load3dSerializedBase` type. The
type is only used internally as the return type of `snapshotLoad3dState`
and is not imported elsewhere.
Fixes knip "unused exported types" error.
Co-authored-by: Connor Byrne <c.byrne@comfy.org>
## Summary
On first launch, the templates modal flashed open for a split second
before a deeplinked template (`?template=`) loaded, which felt broken.
## Changes
- **What**: Gate the first-launch templates modal on template URL
intent, alongside the existing shared-workflow (`?share=`) check. When a
template is being opened directly from the URL, the template modal no
longer opens. Behavior is unchanged when no template is in the URL — the
template modal still shows for first-time users.
- Test util: Added browser_tests/fixtures/utils/flashDetector.ts —
installs a pre-navigation requestAnimationFrame sampler that flags if a
[data-testid] element ever renders, even for a single frame. This
catches a brief flash that toBeHidden() (final-state only) cannot.
## Review Focus
`hasTemplateUrlIntent()` mirrors the existing
`hasSharedWorkflowIntent()` (direct `route.query` check plus
preserved-query fallback for the `/user-select` redirect path). Two
regression tests cover both the URL-param and preserved-intent cases.
**Coverage:**
- Unit (useWorkflowPersistenceV2.test.ts): the modal is not opened when
a template param is in the URL, and when template intent is preserved
across the /user-select redirect.
- E2E (templates.spec.ts): templates dialog never flashes when
first-time user opens a template link — verified red-without-fix,
green-with-fix.
Screen Recording
https://github.com/user-attachments/assets/636094d4-0ef0-4e42-af32-d4e6c7ec5731closes#12836
## Summary
Adds a team-gated, label-triggered multi-model Cursor review as a **thin
caller** for the reusable workflow in `Comfy-Org/github-workflows` — the
single source of truth for the panel, judge, prompts, and scripts. This
repo carries only the ~50-line caller, so there's no review logic to
drift out of sync.
## Changes
- **What**: `.github/workflows/pr-cursor-review.yaml` triggers on the
`cursor-review` label and calls
`Comfy-Org/github-workflows/.github/workflows/cursor-review.yml`, pinned
to `047ca48` (github-workflows#9, current main). Inheriting the reusable
workflow brings severity badges, line-anchored inline comments,
diff-size caps, prompt-injection hardening, and optional Slack DMs.
- **Config**: `diff_excludes` restated (overriding replaces the default
wholesale) with this repo's heavy paths added (Playwright snapshots,
generated manager types). Judge and panel both default to Opus 4.8 via
the reusable workflow — no overrides needed.
## Review Focus
- **Access control (the point).** Two layers, no allowlist: (1) only
triage+ users can apply a label in a public repo; (2) the reusable
workflow's secret-bearing jobs don't run on fork PRs, so
`CURSOR_API_KEY` is reachable only on internal branches.
- **Replaces a standalone draft.** Earlier revisions of this branch
carried a self-contained workflow + review/judge scripts; that
duplicated the reusable workflow, so it's been swapped for the thin
caller.
## Prerequisites (already done)
- `CURSOR_API_KEY` secret set on this repo.
- `cursor-review` label created.
- `SLACK_BOT_TOKEN` already present (enables the DM feature).
---------
Co-authored-by: GitHub Action <action@github.com>
## Summary
Enable Codecov carryforward on the `unit` and `e2e` flags so a missing
or late E2E coverage upload no longer produces a false `codecov/patch`
failure.
## Changes
- **What**: Add a `flags` block to `codecov.yml` marking `unit` and
`e2e` as `carryforward: true`. When a flag's upload is absent for a
commit, Codecov reuses that flag's last known coverage for unchanged
files instead of treating those lines as patch misses.
## Review Focus
The `e2e` flag is uploaded by the separate `ci-tests-e2e-coverage`
`workflow_run` job, which runs after CI and currently fails or skips
intermittently (uploaded with `fail_ci_if_error: false`). When that
upload is missing, the head report holds only the `unit` session, so
E2E-only code paths (canvas, vue-nodes, minimap, glsl, etc.) report as
uncovered and the patch status fails against the full-coverage base.
Carryforward fixes that symptom; the flaky coverage workflow remains a
separate root-cause fix.
Validated with `curl --data-binary @codecov.yml
https://codecov.io/validate` → `Valid!`. Carryforward takes effect after
merge once a `main` build establishes a baseline for both flags.
Co-authored-by: Amp <amp@ampcode.com>
## Summary
Harden the cloud assets media-type filter spec against a VirtualGrid
virtualization flake that intermittently failed CI at the
`waitForAssets(4)` precondition.
## Changes
- **What**: Replace the 7 `waitForAssets(MIXED_JOBS.length)`
preconditions in `assets-filter.spec.ts` with `waitForAssets()` (first
card visible = data loaded), and document why.
## Review Focus
CI artifact (`playwright-report-cloud`) from the failing run showed only
3 of 4 cards in the DOM — the 3D card was missing and the audio card
rendered taller (inline player). `VirtualGrid.vue` sizes its render
window from a single uniform `itemHeight` measured off the first card,
so the taller audio card pushes the 3D card out of the initial window
and it is virtualized out of the DOM until a re-measure (same cause as
#11635).
Requiring all 4 cards to be mounted simultaneously fights
virtualization. `tab.open()` already waits for the first card (data
loaded), and filtering reads the full asset store regardless of what is
mounted, so the per-filter count assertions still provide the real
coverage. No behavioral change to the tests — only the
readiness/precondition strategy.
Verified `pnpm exec eslint` and `pnpm typecheck:browser` pass. Cloud E2E
could not be run locally (needs a running frontend + backend); relies on
the cloud CI job for confirmation.
Related to #11635
---------
Co-authored-by: Amp <amp@ampcode.com>
## Summary
The `/cloud/pricing` page's "Add more credits anytime" feature claims
unused top-up credits **roll over** to the next month. Per product
(Pablo, #cloud), top-up credits **do not** roll over. Fix the copy to
match policy and the docs.
## Changes
- **What**: `pricing.included.feature5.description` (en + zh-CN) in
`apps/website/src/i18n/translations.ts`.
- Before: "Purchase additional credits at any time. Unused top-ups roll
over to the next month automatically for up to 1 year."
- After: "Purchase additional credits at any time. Top-up credits are
valid for 1 year from the date of purchase and do not roll over with
your monthly plan."
- Keeps the accurate 1-year validity window (matches
`docs/interface/credits.mdx`).
- Scope: this is the **only** rollover claim on the page. The tier cards
(lines 1224/1255/1286) only say "top-ups available" — no change needed.
- **Breaking**: none (copy only).
## Deliberately out of scope
The Slack thread surfaced an **unresolved** question: Pablo said top-ups
don't roll over *"as opposed to subscription credits which do,"* but
`docs/interface/credits.mdx` says monthly credits don't roll over
either. This PR makes **no claim about subscription/monthly credit
rollover** (the new copy is neutral: "do not roll over *with your
monthly plan*"). Resolving that — and the optional FAQ entry / docs
alignment Glary-Bot proposed — needs a product decision first. Flagging
for follow-up.
## Summary
Redesign the run error overlay so its count and copy use the same
grouped error semantics as the redesigned Error tab, while keeping
single-error toast copy precise.
This is a stacked follow-up to #12828. The parent PR redesigns the Error
tab cards and centralizes their grouped/row-based presentation. This PR
applies the same mental model to the compact overlay shown from the Run
button path, so users no longer see one count in the Error tab and a
different count in the overlay.
## Changes
- **What**: Align the error overlay count with the Error tab's grouped
error count.
- The overlay now reads the same grouped error surface that powers the
Error tab hero instead of independently summing raw store error counts.
- Validation/runtime/prompt execution groups use their grouped `count`.
- Missing model/media/node/swap groups keep the row/group count
semantics introduced by the Error tab redesign.
- This avoids cases where the overlay headline says one number while the
panel summarizes another.
- **What**: Preserve the existing single-error toast behavior while
adding explicit multi-row handling.
- A single true leaf still uses the catalog/resolver toast title and
toast message.
- A single grouped execution error with multiple node/input items uses
that group's title/message instead of the generic aggregate copy.
- A single missing model/media group with multiple model/file rows uses
the generic aggregate copy because it represents multiple actionable
rows.
- A single missing model/media row referenced by multiple nodes uses the
group title/message, since there is still only one model/file to
resolve.
- Multiple top-level groups continue to use the aggregate "N errors
found" style copy.
- **What**: Restyle the overlay toast to match the new error-surface
direction.
- Adds a compact dark card with a destructive left accent and a visible
outline for better contrast against the workspace.
- Keeps the close button in the card header area.
- Keeps the primary "View details" action, but adjusts spacing, size,
and typography to better match the Figma direction.
- Removes the older footer-style dismiss action so the overlay behaves
like a focused status toast rather than a secondary dialog.
- **What**: Share error grouping/count helpers instead of duplicating
local logic.
- Extracts execution item-list detection for reuse between the Error tab
render path and count logic.
- Extracts missing-model grouping/count helpers so missing-model row
count semantics have one implementation.
- Removes `groupedErrorMessages`, which became unused after the overlay
copy decision moved to grouped error state.
- **Breaking**: None.
- **Dependencies**: None.
## Review Focus
- **Stack boundary**: Please review this against
`jaeone/fe-816-error-card-redesign`, not against `main` directly. The
parent PR is #12828 and contains the Error tab card redesign that this
overlay work builds on.
- **Count semantics**: The visible overlay count intentionally changes
from raw store counts to grouped Error tab counts. This is not just a
refactor; it is the intended product behavior so the compact overlay and
the panel hero agree.
- **Overlay message branches**: The overlay can now choose between three
message modes. The goal is to keep precise single-error copy where it is
useful, but avoid showing one node-specific toast message when the
overlay actually represents multiple actionable rows.
| Branch | When it applies | Overlay title source | Overlay message
source | Example output |
| --- | --- | --- | --- | --- |
| Aggregate summary | More than one top-level error group, or one
missing model/media group with multiple actionable model/file rows. |
Generic aggregate title using grouped count. | Generic aggregate
message. | Title: `2 errors found`<br>Message: `Resolve them before
running the workflow.`<br>Example case: one Missing Models group
containing `first.safetensors` and `second.safetensors`. |
| Group summary | Exactly one error group that is not a true single
leaf, but should still be described by the group. This includes one
execution catalog group with multiple items, or one missing model/media
row referenced by multiple nodes. | The group's `displayTitle`. | The
group's `displayMessage`. | Title: `Missing connection`<br>Message:
`Required input slots have no connection feeding them.`<br>Example case:
one validation group with `KSampler - model` and `KSampler - positive`
rows. |
| Single leaf toast | Exactly one group, one actionable row/card, and at
most one node reference. | Resolver/catalog `toastTitle`. |
Resolver/catalog `toastMessage`. | Title: `Model missing`<br>Message:
`CheckpointLoaderSimple is missing missing.safetensors.`<br>Example
case: one missing model file referenced by one node. |
- **Scope control**: This PR intentionally does not redesign the full
Error Overlay flow beyond the compact toast/card. It also does not
revisit the deeper Error tab card layouts already handled in #12828.
- **Accessibility**: The toast keeps `role="status"` for polite
announcement semantics. The duplicate `aria-live` attribute was removed
during cleanup because `role="status"` already implies polite
live-region behavior.
## Validation
- `pnpm format`
- `pnpm lint`
- `pnpm typecheck`
- `pnpm knip`
- `pnpm test:unit src/components/error/useErrorOverlayState.test.ts
src/platform/missingModel/missingModelGrouping.test.ts
src/components/error/ErrorOverlay.test.ts
src/components/rightSidePanel/errors/useErrorGroups.test.ts`
- Pre-commit staged checks passed.
- Pre-push `knip` passed.
## Screenshots (if applicable)
<img width="454" height="179" alt="스크린샷 2026-06-16 오후 6 00 10"
src="https://github.com/user-attachments/assets/a85376ba-2b22-4cf8-a6fa-79f83fb8b244"
/>
<img width="453" height="179" alt="스크린샷 2026-06-16 오후 6 00 31"
src="https://github.com/user-attachments/assets/d9a1d4bd-92ab-451a-bb79-e7cfbc3af7c6"
/>
<img width="486" height="148" alt="스크린샷 2026-06-16 오후 6 00 55"
src="https://github.com/user-attachments/assets/b66faf96-65c8-4a22-9ff9-e8ccd450986e"
/>
<img width="395" height="127" alt="스크린샷 2026-06-16 오후 6 01 22"
src="https://github.com/user-attachments/assets/c64443f9-0eba-4b2b-8049-c1887c788b1e"
/>
<img width="384" height="134" alt="스크린샷 2026-06-16 오후 6 01 30"
src="https://github.com/user-attachments/assets/42f4fcae-b003-4df9-8f3a-0fda85a90880"
/>
<img width="376" height="129" alt="스크린샷 2026-06-16 오후 6 01 53"
src="https://github.com/user-attachments/assets/ce9030d0-2a98-4b38-9e7d-7a9c3103960f"
/>
<img width="379" height="128" alt="스크린샷 2026-06-16 오후 6 02 01"
src="https://github.com/user-attachments/assets/3d4ce356-1c22-4e3b-a1d0-fece351d9fbb"
/>
<img width="463" height="133" alt="스크린샷 2026-06-16 오후 6 02 33"
src="https://github.com/user-attachments/assets/6ae13a44-02aa-4167-8878-4906db468ad6"
/>
### before
<img width="539" height="475" alt="Screenshot 2026-06-17 at 9 52 50 PM"
src="https://github.com/user-attachments/assets/dd562cb4-870c-43de-aca4-81e0118735e9"
/>
### after
<img width="529" height="592" alt="Screenshot 2026-06-17 at 9 53 40 PM"
src="https://github.com/user-attachments/assets/ba8ed01e-ac91-4654-bfaa-d8f73923f378"
/>
Design-review polish on the team-workspace billing UI (3 of the items
from the Figma 'Team Plan - Workspaces' review). All three components
are on main.
### 1. Remove dead 'Upgrade' badge in account popover
`CurrentUserPopoverWorkspace.vue` — the white badge beside 'Plans &
pricing' was gated on `canUpgrade`, which is hardcoded `false` (PRO is
the only tier), so it never rendered. Removed the badge markup and the
dead computed. (Figma node 2797-724189.)
### 2. Subscribe-to-Run button height
`SubscribeToRun.vue` — used `size="sm"`, which didn't match the sibling
run/queue button (an `h-8` button group it swaps with in the same slot
via `CloudRunButtonWrapper`). Switched to `size="unset"` + `h-8
rounded-lg gap-1.5 px-4` to match.
### 3. Duplicate border/radius on member 'subscription inactive' dialog
`useSubscriptionDialog.ts` — the layout dialog already zeroes the
*content* border (`border-none shadow-none`), but the *root* pt kept
only `bg-transparent`, so the dialog frame's default border+radius
doubled with the card's own `rounded-2xl border`. Added `border-none
rounded-none shadow-none` to root so only the card's single
border/radius shows. (Figma node 3253-19473.)
Surfaced during Billing V1 design review (team workspaces).
---------
Co-authored-by: GitHub Action <action@github.com>
The workspace name was hard-clipped by the switcher panel's
`overflow-hidden` instead of truncating with an ellipsis. The name span
had `min-w-0` but its parent row `<button>` (`flex flex-1`) did not, so
the button kept its content width and the long name overflowed.
Add `min-w-0` to the row button (and the name span) so the flex chain
can shrink and the name truncates.
Follow-up to FE-769 (#12763, merged) — the component shipped without
this.
## Before / After
<img width="1370" height="538" alt="fe769-before-after"
src="https://github.com/user-attachments/assets/b05d6faf-9dfe-4c93-9941-3c0a9bbfcc2d"
/>
Verified live in the team-workspaces mock (long-named "Acme Studio
Workspace").
### after
<img width="703" height="302" alt="Screenshot 2026-06-17 at 9 54 34 PM"
src="https://github.com/user-attachments/assets/725c3175-b65f-4224-aca9-3de777c95e85"
/>
## Summary
Cancelled / inactive team plans keep members but lock runs; the run
button and the subscription-required dialog are now role-aware — owners
are routed to the pricing/subscribe flow, members (who cannot subscribe)
see "contact your workspace owner to resubscribe".
## Changes
- **What**: `SubscribeToRun.vue` becomes a role-aware locked run button
(owner → "Subscribe to Run"; member → neutral locked "Run" +
contact-owner tooltip; both open the subscription dialog).
`SubscriptionRequiredDialogContentWorkspace.vue` branches on role
(member → read-only contact-owner panel, no pricing/subscribe
affordance; owner → existing pricing/preview; member view suppressed for
`out_of_credits` so the active-but-low-credits path is unchanged).
`subscription.inactive.*` i18n keys.
- **Breaking**: none.
## Review Focus
- Role source =
`useWorkspaceUI().permissions.value.canManageSubscription` (owner /
personal = true, member = false) — the same accessor
`SubscriptionPanelContentWorkspace.vue` uses.
- **No BE work**: the run-gate already exists server-side
(`InactiveSubscriptionError`; `is_active` checked before funds). The
lock is gated on `is_active`, the same field the orchestrator uses, so
FE/BE stay consistent; leftover-credits-while-inactive remains blocked
by design.
- Complements #12785 (FE-878 precondition→modal routing); disjoint file
sets. Design: DES-197, Figma 3253-18670 / 3253-18671 / 3246-13962.
- Tests: `SubscribeToRun` (4) / `CloudRunButtonWrapper` (3) /
`SubscriptionRequiredDialogContentWorkspace` role cases — member sees
contact-owner (no subscribe), owner sees pricing, run locks on
`!is_active` and unlocks when active (22 total); full `test:unit` green.
Fixes FE-978
## Summary
Make `WidgetId` (`graphId:nodeId:name`) the single canonical widget
identity and represent subgraph promoted host widgets as ordinary
store-backed widgets addressed by it. This deletes two whole indirection
layers — the `world/*` widget-entity-IO layer and the
`PromotedWidgetView` runtime — leaving one model: a widget's data lives
in `widgetValueStore` keyed by `WidgetId`, and a `SubgraphNode` input
references it via `input.widgetId`.
**Net +304 lines across 107 files** (5,044 added / 4,740 deleted):
production code is net **−798** (1,521 added / 2,319 deleted) while
tests are net **+1,102** (3,523 added / 2,421 deleted). 10 files deleted
outright, 14 added.
## What got deleted
The old design wrapped every promoted subgraph widget in a synthetic
`IBaseWidget` "view" object with live getters that followed the source
widget, plus a manager to keep view identities stable, plus an IO
indirection layer over the store. All of it is gone:
- `promotedWidgetView.ts` — the `PromotedWidgetView` class (draw /
pointer / DOM-sync / projection / deepest-source resolution getters)
- `PromotedWidgetViewManager.ts` — view reconciliation/caching
- `world/widgetValueIO.ts` — the IO wrapper over `widgetValueStore`
- `world/entityIds.ts` + `world/brand.ts` — the `WidgetEntityId`
branded-id layer and the `entityId` field
- `widgetNodeTypeGuard.ts` — only used by the deleted view
- the per-`SubgraphNode` view machinery (`_promotedViewManager`,
`_cacheVersion`, view-key generation, DOM position-override cleanup) and
every now-dead `isPromotedWidgetView` branch across the panel, menu,
store, and util consumers
- `domWidgetStore` position-override APIs (`setPositionOverride` /
`clearPositionOverride`), only used to render a promoted DOM widget on a
different host node
## Why it's simpler
- One source of truth. A promoted host widget is `WidgetState` in the
store, seeded from the source at promotion (`registerWidget` with a
deep-cloned snapshot) and independent thereafter. No synthetic widget
objects, no runtime source-following, no view cache to invalidate.
- Resolution is data-driven. `resolveConcretePromotedWidget` walks
`SubgraphNode` inputs (`input.widgetId` + `resolveSubgraphInputTarget`)
instead of chasing view objects through `node.widgets`. This also
**fixes two-layer nested promotion** — the previously-skipped parity
test now passes and resolves through to the deepest concrete widget.
- The right-panel Parameters tab renders a subgraph node's promoted
widgets through the **same** store-backed path as ordinary node widgets:
display reads `WidgetState` via `widget.widgetId`, and value writes go
through `widgetValueStore.setValue(widgetId)`.
## Changes
- **What**:
- `WidgetId` branded type + `widgetId()` / `parseWidgetId()` /
`isWidgetId()` and a `WidgetState` type; `widgetValueStore` is
`WidgetId`-native (`registerWidget` / `getWidget` / `setValue` /
`deleteWidget`).
- Promotion creates host `WidgetState` then an input projection
(`input.widgetId`); demotion clears it; serialization and legacy
`proxyWidgets` migration round-trip through `input.widgetId`.
- `promotedInputWidget.ts` projects a store-backed ordinary widget from
an input slot; `SubgraphNode.widgets` is now a projected getter over
inputs (kept Litegraph-shaped so the canvas renderer and extensions
still read `node.widgets`). `invalidatePromotedViews()` is retained as a
no-op for extension compatibility.
- `promotedWidgetControl.ts` applies `control_after_generate` (e.g. seed
increment) on the host node, since the interior control widget is
link-fed and its value is dead; `syncPromotedComboHostOptions` mirrors
interior combo options onto host state.
- `multilineTextarea.ts` extracts the reusable multiline DOM-widget
behavior out of `useStringWidget` and adds promoted multiline
materialization via a `createPromotedHostWidget` app-layer hook (keeping
Litegraph core free of Vue/Pinia/DOM).
- Late-bound `LiteGraph` singleton holder (`litegraphInstance.ts`) to
break a widget-init import cycle.
- **Breaking**:
- `IBaseWidget.entityId` removed — use `widgetId`.
- `SubgraphNode.widgets` no longer exposes the old `PromotedWidgetView`
objects; promoted state lives in `widgetValueStore` keyed by
`input.widgetId` and `widgets` is a projection of inputs. The
`widget-promoted` event now carries the concrete interior widget.
Extension code reading `entityId` or relying on `PromotedWidgetView` is
affected.
## Review Focus
- Promotion / demotion / serialization round-tripping through
`input.widgetId` + `widgetValueStore`, incl. the legacy `proxyWidgets`
migration.
- Snapshot-at-promotion semantics (host widget does not follow the
source after creation), and the combo-options exception via
`syncPromotedComboHostOptions`.
- Two-layer nested resolution in `resolveConcretePromotedWidget` +
`SubgraphNode` nested-source resolution.
- The unified Parameters tab (`TabSubgraphInputs` → `SectionWidgets`):
value edit / rename / favorite / hide / reorder for promoted inputs are
wired through the store but warrant a visual/e2e pass.
- Litegraph-compat seams worth a careful read: projected
`SubgraphNode.widgets`, the canvas-edit `callback` bridge back to the
store, host-level `control_after_generate`, and the late-bound
`LiteGraph` holder / `domWidget.ts` import ordering.
---------
Co-authored-by: Amp <amp@ampcode.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: AustinMroz <austin@comfy.org>
Co-authored-by: GitHub Action <action@github.com>
When control is held, an active drag operation should cease applying
movements to nodes contained by selected groups. This functionality was
broken in vue mode because of unnecessary reimplementation of the code
for calculating items contained by groups during drag operations
## Summary
L1 prerequisite cleanup, scoped to a single type-preserving refactor:
extract `getAssetStoredFilename(asset)` to collapse the duplicated
`isCloud && asset.asset_hash ? asset.asset_hash : asset.name` branch
from `useMediaAssetActions.ts` into one helper in
`assetMetadataUtils.ts`. No behavior change.
Once BE-933/934 emit `file_path` and the cloud spec sync brings the
field into generated types, only the helper internals change (collapse
to `asset.file_path ?? asset.name`).
## Scope change (per review)
The `mockFeatureFlags` test util and the exported `FeatureFlags` type
that this PR originally also added have been **split out**. They had no
live consumer in the open stack — FE-729~732 (#12322 / #12335 / #12375 /
#12417) don't use them, and FE-780 / FE-781 (#12485 / #12486) still
hand-roll inline `vi.hoisted` mocks — so shipping them here would add a
public surface with no caller. They will be reintroduced bundled with
the first PR that actually adopts the util, where `featureFlag`'s return
type and the "all flags off vs. production defaults" semantics can be
validated against a real consumer.
## Review fixes carried in this PR
- Mock `@/platform/distribution/types` via an `importOriginal` spread so
`isDesktop` / `isNightly` survive the wholesale replacement (only
`isCloud` was re-hoisted before).
- Trimmed the `getAssetStoredFilename` JSDoc; the BE-933/934
future-collapse is now a one-line `TODO` rather than a design-doc
paragraph.
## Review Focus
- The helper is intentionally named `getAssetStoredFilename` to
disambiguate from the existing `getAssetFilename` (which targets
`user_metadata.filename` / `metadata.filename` for serialized-identifier
contexts — missing-model matching, filename schema validation) and
`getAssetDisplayFilename` (UI labels). Folding the `isCloud &&
asset_hash` fallback into either of those would regress
display/identifier sites where the cloud hash is never meant to surface.
- Fixes FE-733
- Parent: FE-601 (L1 umbrella)
- RFC: [Asset Identity
Semantics](https://www.notion.so/comfy-org/RFC-Asset-Identity-Semantics-35a6d73d365080e59d59c98cebae779b)
- Survey: [Asset FE Divergence Survey M1
Scope](https://www.notion.so/comfy-org/Assets-FE-Divergence-Survey-M1-Scope-3616d73d365080d0a9cbf5f2394c12f8)
- Slack thread:
https://comfy-organization.slack.com/archives/C0AUUTS2RQV/p1778815571949519
## Screenshots (if applicable)
N/A — no UI change.
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-12287-refactor-assets-extract-getAssetStoredFilename-helper-add-mockFeatureFlags-test-util-3616d73d365081c9a1c6e1982728a38a)
by [Unito](https://www.unito.io)
---------
Co-authored-by: GitHub Action <action@github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Alexander Brown <drjkl@comfy.org>
Co-authored-by: Matt Miller <matt@miller-media.com>
## Summary
Fix issue where variables typed in the `filename_prefix` of the save
nodes were not interpreted. See issue raised by users here:
https://github.com/Comfy-Org/ComfyUI/pull/13850#issuecomment-4700771342
## Changes
- **What**: Added the following nodes to
`src/extensions/core/saveImageExtraOutput.ts`
- `SaveImageAdvanced`
- `SaveSVGNode`
- `SaveAudioMP3`
- `SaveAudioOpus`
- `SaveAudioAdvanced`
## Summary
Fixes FE-874 by preventing terminal jobs that do not have an
`execution_end_time` from being grouped under the `Undated` section in
the expanded job history / queue UI.
The production change is intentionally small: when grouping completed or
failed jobs by date, the UI now falls back from `executionEndTimestamp`
to `createTime`.
```ts
ts = task.executionEndTimestamp ?? task.createTime
```
This keeps the existing preference for execution completion time when it
is available, while still giving pre-execution terminal jobs a
meaningful date bucket based on when the job was created.
## Root Cause
Some terminal jobs, especially failures that happen before execution
actually starts, can legitimately arrive without `execution_end_time`.
This matches the backend semantics: if there is no execution start,
there may be no execution end timestamp either.
Before this change, the frontend grouping logic treated missing
`executionEndTimestamp` as if there were no usable date at all for
terminal jobs. That caused failed jobs with a valid `create_time` to
appear in the `Undated` group.
At the same time, the list sorting logic already used `createTime`, so
those jobs could appear near the top of the list while still being
labeled as `Undated`. That mismatch made recent failed jobs look like
they had no date, even though the job creation timestamp was present.
## What Changed
- Updated `useJobList` date grouping for terminal jobs:
- `completed` and `failed` jobs still use `executionEndTimestamp` when
available.
- If `executionEndTimestamp` is missing, they now fall back to
`createTime`.
- Added regression coverage for terminal jobs without an execution end
timestamp:
- A failed job without `executionEndTimestamp` is grouped by
`createTime`.
- A completed job without `executionEndTimestamp` is also covered
because the production fallback applies to both terminal states in the
same code path.
- Cleaned up the `useJobList` test harness by replacing the mocked
`vue-i18n` module with a real `createI18n` instance per mount.
- This follows the repo testing guidance to avoid mocking `vue-i18n`.
- Each composable mount now receives a fresh i18n instance, avoiding
shared mutable i18n state between tests.
## User Impact
Failed jobs that never reached execution will no longer show up under
`Undated` when they still have a valid creation timestamp. They will
instead appear under the correct date group, such as `Today`,
`Yesterday`, or a localized month/day label.
This should make the expanded job history easier to scan and avoid the
confusing case where recent failed runs appear at the top while also
being labeled as undated.
## E2E Regression Coverage Rationale
I did not add a Playwright regression test under `browser_tests/` for
this fix because the regression is isolated to the timestamp selection
used by `useJobList` when it builds date groups. There is no changed
user interaction, navigation flow, API request shape, route handling, or
browser-only behavior.
The existing browser coverage already verifies that the Job History
sidebar opens, renders active and terminal jobs, and filters
completed/failed jobs using the mocked jobs route fixture. Adding an E2E
for this specific case would require creating another mocked `/api/jobs`
response with a terminal job that has `create_time` but no
`execution_end_time`, opening the sidebar, and asserting the rendered
date header. That would mostly duplicate the composable-level assertion
through the DOM while adding extra moving parts around relative date
labels, locale/timezone formatting, and the virtualized job list.
The regression is therefore covered more directly and deterministically
at the unit level in `src/composables/queue/useJobList.test.ts`. The new
test drives the same grouping pipeline that the UI consumes and asserts
that terminal jobs without `executionEndTimestamp` are grouped by
`createTime` instead of falling into `Undated`. I also verified the test
fails against the pre-fix implementation with `['Undated']` and passes
with the fallback.
## Notes
This PR does not attempt to synthesize an execution end time. The
backend can validly omit `execution_end_time` for jobs that never
started execution. The frontend fix is limited to display grouping: if
there is no execution end timestamp, use the already-present creation
timestamp as the grouping date.
If product requirements later need the exact terminal failure timestamp
for pre-execution failures, that would require a separate backend/API
timestamp such as a terminal-state or update timestamp. This PR only
fixes the current display fallback.
## Validation
Local validation run before publishing:
```bash
pnpm test:unit src/composables/queue/useJobList.test.ts
pnpm exec eslint src/composables/queue/useJobList.ts src/composables/queue/useJobList.test.ts
pnpm exec oxlint src/composables/queue/useJobList.ts src/composables/queue/useJobList.test.ts --type-aware
git diff --check -- src/composables/queue/useJobList.ts src/composables/queue/useJobList.test.ts
```
The commit hook also ran successfully during the final amend and passed:
```bash
pnpm exec oxfmt --write ...
pnpm exec oxlint --type-aware --fix ...
pnpm exec eslint --cache --fix ...
pnpm typecheck
```
## Screenshots
Before
<img width="346" height="624" alt="Screenshot 2026-06-17 1:35:06 AM"
src="https://github.com/user-attachments/assets/02269f57-038a-4f06-9892-0758ad84d2c7"
/>
After
<img width="352" height="632" alt="Screenshot 2026-06-17 1:35:37 AM"
src="https://github.com/user-attachments/assets/251cd762-2c88-4af6-8218-4af1915727b6"
/>
## Summary
Fixes a Nodes 2.0 node replacement regression where widgets that only
exist on the replacement node were not registered with the widget value
store, causing their Vue-rendered controls to fall back to component
defaults such as `0` instead of the replacement node's real widget
default.
The root cause is that `replaceWithMapping()` replaces the placeholder
node in-place by writing directly to `graph._nodes` and
`graph._nodes_by_id`. That path intentionally preserves the old node id,
but it also bypasses the normal `LGraph.add()` flow that binds widgets
to their owning node id. As a result, newly introduced bindable widgets
on the replacement node could exist on the LiteGraph node object while
remaining absent from `useWidgetValueStore`, which is the state Vue
Nodes reads from when rendering widget controls.
## Changes
- **What**: Bind every bindable widget on the replacement node to the
reused node id inside `replaceWithMapping()` after the replacement node
is inserted into the graph maps and before widget values are
transferred.
- **What**: Preserve the existing widget value transfer behavior for
mapped widgets. Because widgets are now bound before `newWidget.value =
oldValue` runs, transferred values are written through the normal widget
store state instead of only mutating the unbound widget object.
- **What**: Add a focused unit regression check that verifies
replacement-only widgets are bound with the reused node id during node
replacement.
- **What**: Extend the existing node replacement Playwright coverage to
assert the Vue Nodes rendered input for `KSampler.denoise` keeps the
expected replacement value after the replacement flow.
- **Breaking**: None.
- **Dependencies**: None.
## Review Focus
Please focus on the placement of the widget binding in
`replaceWithMapping()`. The binding happens after the new node has been
assigned the reused id and inserted into the graph's node maps, but
before mapped widget values are copied over from the old node. This
mirrors the important part of the normal graph add flow for widgets
while keeping the in-place replacement behavior intact.
The tests intentionally avoid asserting replacement-node fixture
defaults in isolation. The unit test verifies the actual new side effect
that prevents the regression: `setNodeId()` is called for a bindable
widget that was not present on the old node. The Playwright assertion
then covers the user-visible Nodes 2.0 symptom: the replacement widget
is rendered from the widget store instead of falling back to the Vue
numeric default.
Linear: FE-1070
## Validation
- `pnpm vitest run
src/platform/nodeReplacement/useNodeReplacement.test.ts`
- `pnpm typecheck:browser`
- `PLAYWRIGHT_LOCAL=1 PLAYWRIGHT_TEST_URL=http://localhost:5174 pnpm
exec playwright test --project=chromium
browser_tests/tests/nodeReplacement.spec.ts -g "Widget values are
preserved after replacement"`
- `pnpm lint`
- `pnpm typecheck`
- Commit hook also reran staged formatting/linting and `pnpm typecheck`
during the final amend.
## Screenshots (if applicable)
Before
https://github.com/user-attachments/assets/dc4e8137-d8aa-4a70-9973-5559ed84b90e
After
https://github.com/user-attachments/assets/4c70b9e4-d971-4e94-8d2f-12b0f2b00a09
## Summary
Fix Ctrl+C copy for large subgraphs by encoding clipboard metadata in
bounded byte chunks instead of spreading the full serialized payload
into a single `String.fromCharCode(...)` call.
## Root Cause
<img width="648" height="33" alt="스크린샷 2026-06-15 오후 4 46 52"
src="https://github.com/user-attachments/assets/09aec159-fd10-4979-bfb2-51aec9b51a63"
/>
Ctrl+C uses the native `copy` event path in `useCopy.ts` so ComfyUI can
write serialized node metadata into the system clipboard as `text/html`.
That metadata supports the cross-app / cross-window copy-paste path.
For Unicode safety, the current code first converts the serialized node
JSON to UTF-8 bytes with `TextEncoder`, then converts those bytes into a
binary string for `btoa`. The bug was in this conversion step:
```ts
String.fromCharCode(...Array.from(new TextEncoder().encode(serializedData)))
```
When a selected subgraph is large enough, the UTF-8 byte array becomes
too large to spread as function arguments. The browser throws
`RangeError: Maximum call stack size exceeded` before clipboard metadata
is written, so Ctrl+C appears to fail for large subgraphs.
The right-click / menu copy path was not affected in the same way
because it uses LiteGraph's internal `copyToClipboard()` path directly
and does not go through this system clipboard metadata encoding step.
## Changes
- **What**: Convert UTF-8 bytes to a binary string in `0x8000` byte
chunks before passing the result to `btoa`.
- **Why**: This preserves the existing UTF-8 safe cross-app clipboard
metadata format while avoiding the JavaScript argument-count limit that
caused the stack overflow.
- **Fallback**: Wrap system clipboard metadata encoding/writing in
`try/catch` so the internal `canvas.copyToClipboard()` result is still
produced even if the metadata bridge fails unexpectedly.
- **Dependencies**: None
## Review Focus
- Chunking is only used while building the binary string for base64
encoding. The clipboard payload format remains unchanged.
- Multi-byte UTF-8 data remains safe because chunking happens at the
byte-string construction layer; paste still reassembles the full byte
stream before `TextDecoder` decodes it.
- The unit test exercises the actual `useCopy` copy handler with a large
serialized payload, Unicode metadata, and a partial final chunk.
## Test Plan
- `vitest run src/composables/useCopy.test.ts`
- pre-commit hook: `oxfmt`, `oxlint`, `eslint`, `typecheck`
- pre-push hook: `pnpm knip`
No E2E was added because this regression is isolated to deterministic
clipboard metadata encoding in `useCopy`. The unit test exercises the
actual `copy` event handler with a large serialized payload and Unicode
metadata, avoiding a large workflow fixture and slower browser coverage
for behavior that does not require canvas rendering or end-to-end UI
orchestration.
Linear:
[FE-858](https://linear.app/comfyorg/issue/FE-858/bug-ctrlc-copy-keyboard-shortcut-does-not-work-on-large-subgraphs)
## Summary
Redesigns the Errors tab cards to match the new Figma error-panel spec
(file `Czv0JcCfcUiizeEZURevpq`): every error group is now wrapped in a
single bordered card led by an error-count summary hero, with each
category rendered as a collapsible section whose count lives in a
circular badge rather than a parenthetical title suffix.
> Rebased onto `main` after #12793 was merged. This PR now contains only
the error-card redesign slice.
## Changes
- **What**:
- **New `ErrorCardSection.vue`** — the shared section shell used by
every error type. Renders a 32px header (circular count badge + neutral
title + `actions` slot + collapse chevron) and a `TransitionCollapse`
body. Replaces the per-group `PropertiesAccordionItem`, dropping the old
octagon-alert icon + red title + `(n)` suffix and the sticky-header
behavior.
- **`TabErrors.vue`** — wraps all groups in one `rounded-lg` card
bordered with `secondary-background`. Adds a summary **hero** (large
severity-colored total count, vertical divider, "N Errors detected /
Resolve before running the workflow"). Moves the per-group action
buttons (Install All / Replace All / missing-model Refresh) into the
section's `actions` slot. Adds `getGroupCount()` / `totalErrorCount` and
switches content background to `interface-panel-surface`. Most of the
line count here is re-indentation from the template restructure, not
behavior change.
- **`missingErrorResolver.ts`** — drops the `formatCountTitle` helper so
display titles are `"Missing Models"` instead of `"Missing Models (4)"`;
the badge now carries the count. Toast titles/messages are untouched.
- **`ErrorNodeCard.vue`** — restyles the runtime/validation error-log
box to the Figma spec: borderless `base-foreground/5` surface, `ERROR
LOG` header, 12px non-mono body at 50% opacity, inset footer divider
with Get Help / Find on GitHub links.
- **Row components** (`MissingModelRow`, `MissingPackGroupRow`,
`SwapNodeGroupRow`, `MissingMediaCard`, `MissingNodeCard`,
`MissingModelCard`) — align spacing, fonts, badges, and button sizes
with Figma: 12px row labels, `size="sm"` (24px) action buttons, 16px
count badges (`rounded-sm`, `secondary-background-hover`, 9px), 32px
reference-row heights, `px-3` card padding. Model-name wrapping is kept
independent of its count badge and link button so they never reflow into
the metadata sub-label.
- **i18n** — adds `errorsDetected` (pluralized), `resolveBeforeRun`,
`expand`, `collapse` to `en/main.json`.
- **Breaking**: None. No store, composable, action, or data-flow changes
— all handlers and emitted events are preserved. The only user-visible
copy change is the removal of the `(n)` count suffix from section
titles.
## Review Focus
- **Title copy change**: `"Missing Models (4)"` → `"Missing Models"`.
Search-filter matching against the old `(n)` string no longer applies,
but the count is shown by the badge and the hero total.
- **Sticky header removed**: section headers no longer pin to the top on
scroll (intentional per the new design).
- **Collapse click target**: the old single-button header (which nested
action buttons inside a `<button>` — invalid HTML) is split into a
separate title button and chevron button. Behavior is unchanged and
accessibility improves; the empty space beside an action button no
longer toggles collapse.
- All semantic colors map to existing design-system tokens (no `dark:`
variants, no hardcoded hex). Verified the artifact hex values match the
tokens (e.g. `#262729` = `secondary-background`, `#e04e48` =
`destructive-background-hover`, `#171718` = `interface-panel-surface`).
## Follow-up
This PR intentionally keeps the error-count ownership cleanup out of the
current diff so the card redesign remains reviewable. A follow-up PR
will centralize error counting around a single source of truth so the
Errors tab summary hero, section badges, and any overlay surfaces cannot
drift from one another.
That follow-up will also address the current count mismatch in the
ErrorOverlay and continue the ErrorOverlay redesign there, instead of
expanding this PR after review.
## Screenshots (if applicable)
After
<img width="603" height="703" alt="스크린샷 2026-06-13 오후 1 00 02"
src="https://github.com/user-attachments/assets/065d7c19-9748-4e99-9b43-675a31e92949"
/>
<img width="601" height="197" alt="스크린샷 2026-06-13 오후 1 01 07"
src="https://github.com/user-attachments/assets/0fa1fbda-9091-4a45-9eca-e99c43089c0e"
/>
<img width="617" height="612" alt="스크린샷 2026-06-13 오후 1 02 43"
src="https://github.com/user-attachments/assets/3d67a057-bf65-4e51-bcf5-70ecce851826"
/>
<img width="495" height="723" alt="스크린샷 2026-06-13 오후 1 03 28"
src="https://github.com/user-attachments/assets/6dcc4021-0fc3-4955-a68b-c0533c66a3cf"
/>
---------
Co-authored-by: GitHub Action <action@github.com>
## Summary
The redesigned Settings dialog (Figma DES `3253-16079`) is **1280px**
wide, but it rendered at **960px**.
Root cause — the width was capped at 960 in **two** layers:
1. `useSettingsDialog.ts` → `SETTINGS_CONTENT_CLASS` (`max-w-[960px]`)
sizes the Reka dialog shell.
2. `SettingDialog.vue` → `<BaseModalLayout size="sm">` (`SIZE_CLASSES.sm
= max-w-[960px]`) sizes the modal content.
Widening only the shell leaves the inner `BaseModalLayout` at 960 (empty
space on the right). This sets both to **1280px** and lets
`BaseModalLayout` fill the shell (`size="full"`).
The dialog size is **not** a workspace-specific concern, so it applies
to all Settings (OSS + cloud) — no feature-flag gate.
Found during FE-768 designer QA.
## Verification
- Live: dialog measures 1280px, content area 1006px (was 960 / 688).
- `useSettingsDialog.test.ts`: `contentClass` is 1280px (`size:
'full'`).
- `pnpm typecheck` / `lint` / `format` / unit tests green.
## Test Plan
- [x] Settings dialog renders at 1280px with the content filling the
dialog
- [x] Unit test asserts the 1280px sizing
## Screenshots
Settings ▸ Plan & Credits at **1280px** (content fills the dialog; was
960px shell / 688px content area):
**Personal — Pro:**
<img width="720" alt="Settings dialog at 1280px — personal Pro"
src="https://github.com/user-attachments/assets/adc2fd9f-d249-469f-b947-1ec8f674cbb0"
/>
**Team:**
<img width="720" alt="Settings dialog at 1280px — team"
src="https://github.com/user-attachments/assets/e7378067-11a2-411b-b37b-98c8aecb82b1"
/>
---------
Co-authored-by: github-actions <github-actions@github.com>
Node previews are stored by `locatorId`, but sent from the server by
`executionId`. Normally, this difference is reconciled when the event is
received, but this step is skipped when the workflow is backgrounded.
Upon reloading the workflow, these backlogged `executionId`s were
incorrectly mapped directly onto node outputs. Any outputs located
inside a subgraph would then fail to display because `executionId`s are
now `locatorId`s.
This is solved by resolving any `executionId`s at time of output
restoration. Because `executionId`s can only leak into the outputs of
backgrounded workflows, it is safe for resolved `executionId`s to
overwrite any pre-existing `locatorId`s.
It might wind up cleaner to instead properly enforce that the
nodeOutputs cached by change tracker resolve a `locatorId` at time of
receipt. This would follow naturally for properly branded id types, but
would then require resolving `locatorId` from suspended workflows which
is a good bit more involved.
## Summary
Splits the **heavy, hard-to-test surface** out of the Phase 6 dialog
cutover (#12593) into its own independently reviewable, independently
testable PR — per @jtydhr88's review feedback that #12593 bundled too
many concepts (3D, mask editor, and the renderer cutover) to test
thoroughly at once.
This PR migrates only the four style-string dialog callers that carry
**Playwright screenshot baselines** and **maximize behavior** — the mask
editor and the 3D viewers — plus the shared dialog infrastructure they
need. **#12593 is rebased on top of this PR** and now contains only the
renderer cutover.
Parent:
[FE-571](https://linear.app/comfyorg/issue/FE-571/dialog-system-migration-primevue-reka-ui-parent)
This phase:
[FE-578](https://linear.app/comfyorg/issue/FE-578/phase-6-remove-primevue-dialogconfirmdialog-imports-clean-up-css)
## Why this is safe to land alone
**The global renderer default stays `'primevue'`.** Every caller
migrated here sets `renderer: 'reka'` explicitly, and the infra
additions are purely additive. So no other dialog changes behavior and
there is no half-migrated state — the default flip and the remaining
caller migrations all live in the stacked cutover (#12593).
## Changes
**Heavy callers → `renderer: 'reka'` + `size`/`contentClass`:**
- Mask editor (`useMaskEditor.ts`) — `mask-editor-dialog` hook class
moves to `contentClass` so `browser_tests` selectors keep working
unchanged
- 3D viewers ×4 (`ViewerControls.vue`, `AssetsSidebarTab.vue`,
`JobHistorySidebarTab.vue`, `load3d.ts`)
**Infra to reach Reka parity (additive):**
- `dialogStore`: `headerClass`/`bodyClass`/`footerClass` (Reka-path
analogues of `pt.header`/`pt.content`/`pt.footer`)
- `GlobalDialog`: forward the section classes; merge `bodyClass` into
the body wrapper
- `DialogContent`: maximized re-asserts its dimension classes after the
caller's `contentClass` so maximize wins, mirroring
`.p-dialog-maximized` `!important`
- `tailwind-utils`: teach tailwind-merge the `max-h-none` class so
maximize can release the caller's `max-height`
- `rekaPrimeVueBridge`: keep a backgrounded reka dialog from dismissing
when a stacked dialog opens on top of it
- `maskeditor/useKeyboard`: capture keydown so undo/redo survive the
Reka focus trap
## Quality gates
- [x] `pnpm typecheck` — clean
- [x] `pnpm lint` / `pnpm format` — clean (lint-staged)
- [x] `GlobalDialog.test.ts` — 25 passing (incl. new section-class +
maximize-override + stacked-dismiss tests)
- [x] Changed-source unit tests (`useMaskEditor`, `useKeyboard`,
`ViewerControls`, `load3d`) — 77 passing
- [ ] CI Playwright — mask editor baselines refreshed for the Reka
chrome (`browser_tests/tests/maskEditor.spec.ts-snapshots/*`)
## Out of scope (stacked in #12593)
The renderer cutover: `showConfirmDialog` flip, remaining
`dialogService`/composable callers (signin, top-up, workspace,
subscription, publish, share, …), **the `createDialog` default flip to
`'reka'`**, e2e selector retargeting, and the `ConfirmationService`
removal. PrimeVue branch deletion remains Phase 6b.
## 📸 Screenshots — before (PrimeVue) → after (Reka)
Captured via Chrome DevTools against this branch in cloud mode
(`cloud.comfy.org` backend), with an input image / `cube.obj` loaded.
Only the dialog **chrome** migrates (PrimeVue `Dialog` → Reka
`DialogContent`); the editor/viewer content is unchanged.
### Mask editor (`useMaskEditor`)
| Before (PrimeVue) | After (Reka) |
|---|---|
| <img width="430" alt="mask editor before"
src="https://github.com/user-attachments/assets/267e63b5-0832-409e-9c41-edf5ff96561f"
/> | <img width="430" alt="mask editor after"
src="https://github.com/user-attachments/assets/073cd824-8b01-4c07-99e1-a3a054906c7a"
/> |
### 3D viewer (`load3d` / `ViewerControls`)
| Before (PrimeVue) | After (Reka) |
|---|---|
| <img width="430" alt="3D viewer before"
src="https://github.com/user-attachments/assets/17b2cd2f-18e4-4d9a-9e0e-80ef833db216"
/> | <img width="430" alt="3D viewer after"
src="https://github.com/user-attachments/assets/9e20a7a5-4d22-40e6-8fa2-ece58b6e4d20"
/> |
### 3D viewer — maximized (maximize-wins dimension re-assertion in
`DialogContent`)
| Before (PrimeVue) | After (Reka) |
|---|---|
| <img width="430" alt="3D viewer maximized before"
src="https://github.com/user-attachments/assets/b705a4d5-4657-41ad-b6f3-95e54494ac9b"
/> | <img width="430" alt="3D viewer maximized after"
src="https://github.com/user-attachments/assets/188de427-ab58-45a9-8666-967b2908c320"
/> |
## Summary
Aligns the workspace switcher and creation flow to DES-246 (FE-769): the
switcher popover now opens to the **left** of the profile menu instead
of on top of it, team workspace rows drop the tier badge, and the
create-workspace dialog matches the design's copy and surface.
## Changes
- **What**:
- `CurrentUserPopoverWorkspace.vue`: replace the nested PrimeVue
`Popover` (rendered on top of the menu) with an inline panel anchored
left of the selector row (`right-full`, top-aligned, outside-click close
via VueUse)
- `WorkspaceSwitcherPopover.vue`: tier badge only renders on the
personal workspace row ("Remove the tier badge for team workspaces,
since there'll only be one plan now")
- Copy (`en/main.json`): switcher create label "Create a team workspace"
("Explicitly say 'team'"); create dialog message "Workspaces keep your
projects and files organized. Subscribe to a Team plan to invite
members.", label "Workspace name", placeholder "Ex: Comfy Org"
- `CreateWorkspaceDialogContent.vue`: surface matched to the Create
Workspace / Default frame — 512px width, muted name label, filled 40px
TextInput (`bg-secondary-background`, `rounded-lg`, `px-4`)
- Invite-flow copy deltas from DES-246: none — #12759 (FE-768) already
matches the design verbatim
Was stacked on #12762 (FE-778); that PR merged, so this is now rebased
onto `main` with only the FE-769 commits.
- Fixes
[FE-769](https://linear.app/comfyorg/issue/FE-769/updates-to-misc-ux)
## Review Focus
- The switcher panel now lives inside the profile popover DOM (no
teleport): clicking a row keeps the menu open, outside-click closes only
the panel
- The CREATOR badge on the profile-menu workspace selector row (visible
in the Figma frame) is intentionally not included — it needs
`is_original_owner` from the BE role-change work and ships with FE-770
- `leave-last-workspace -> auto-create` flow is deferred (not V1),
intentionally untouched
## Screenshots (if applicable)
| Before (overlaps menu, tier badges, "Create new workspace") | After
(left of menu, no team badges, "Create a team workspace") |
| --- | --- |
| <img width="700" alt="before"
src="https://github.com/user-attachments/assets/5522fcca-91b5-49e6-beaa-df1b88bed018"
/> | <img width="1100" alt="after"
src="https://github.com/user-attachments/assets/ce74d42e-19bd-4fe6-9477-b22e5964736d"
/> |
Create-workspace dialog with DES-246 copy and surface (512px, filled
input):
<img width="900" alt="create dialog after"
src="https://github.com/user-attachments/assets/e78eff0a-1c0e-4bbb-ac70-6cc1da996682"
/>
## Motivation
- **Why B8 exists**: two divergent BillingOp pollers poll the same
op-status endpoint with different policies (hand-rolled 2× / 5s cap / 30
attempts vs the store's 1.5× / 8s / 120s). Once B1 (FE-966) routes
personal flows through the facade, a single cancel op would be polled by
**both** — duplicate requests, with two timeout policies racing on the
same state.
- **The latent bug — silent failure on a money path**: the bespoke
poller treated timeout as a silent return, so `cancelSubscription`
resolved and the dialog showed "cancelled successfully" while the
backend op could still be pending or fail later. The root cause is
structural: the poll outcome was fire-and-forget — no caller consumed
it, so there was no channel through which a failure could surface.
- **The fix — outcome as an awaited contract**: `startOperation` returns
the terminal outcome as a `Promise<BillingOperation>`;
`cancelSubscription` awaits it and throws on any non-`succeeded`
terminal. Every outcome must now flow through the caller, making silence
structurally impossible (timeout/failure → error toast).
- **The trade-off this creates**: "the terminal promise always settles"
becomes load-bearing — the dialog's loading state hangs on it, and a
never-settling path would be worse than the old silence (a permanently
locked dialog). The terminal-promise hardening below, and its regression
tests, enforce that guarantee.
## Summary
Two divergent BillingOp pollers (hand-rolled
`useWorkspaceBilling.pollCancelStatus` vs `billingOperationStore`) are
unified into one — cancel-status polling now runs through
`billingOperationStore`; `pollCancelStatus` and its bespoke
backoff/timers/state are removed.
## Changes
- **What**: `billingOperationStore` gains `'cancel'` in `OperationType`;
`startOperation` now returns `Promise<BillingOperation>` resolving on
the terminal outcome (existing subscription / topup callers unaffected —
fire-and-forget preserved). `useWorkspaceBilling.cancelSubscription`
awaits the shared poller and throws on any non-`succeeded` terminal. One
backoff config = store's 1.5× / 8s / 120s.
- **Terminal-promise hardening**: the terminal promise always settles —
a rejected post-success status/balance refresh no longer leaves
`cancelSubscription` hanging with the dialog locked open
(`Promise.allSettled`), and a duplicate `startOperation` for an
in-flight op joins the same terminal promise instead of resolving
instantly with a `pending` snapshot (which the cancel path would read as
failure).
- **Breaking**: none — the `cancelSubscription(): Promise<void>`
contract is unchanged for `CancelSubscriptionDialogContent` /
`useBillingContext`.
## Review Focus
- **Intentional behavior change**: a cancel **timeout** is now a
terminal outcome that **throws** (`billingOperation.cancelTimeout`), so
the dialog surfaces an error toast — instead of the old silent
success-ish return (which could show success while the op was still
pending). Success / failure semantics otherwise preserved (success →
status refresh + `isSubscribed: false`; failure → throw with message).
- FE-only refactor (B8); cleanest after FE-904 but independent of it.
Relates FE-932 (shared status-refresh path).
## Tests
90 unit tests green across the four affected suites (fake timers for all
polling):
- **`billingOperationStore.test.ts` (33)** — cancel terminal outcomes
(succeeded / failed / timeout) resolve the awaited promise with the
right status + i18n message; cancel suppresses the store's toasts and
settings-dialog side effects; success refreshes status/balance and sets
`isSubscribed: false`; backoff progression, 8s cap, 2-min timeout;
transient poll errors keep polling. Regression guards for the hardening:
post-success refresh failure still settles the terminal promise
(reproduced as a hang before the fix), and a duplicate `startOperation`
joins the in-flight terminal promise instead of resolving with a
`pending` snapshot.
- **`useWorkspaceBilling.test.ts`** — `cancelSubscription` drives the
shared poller; throws the op error on `failed`, throws on `timeout`,
falls back to a generic message when `errorMessage` is absent; a failing
cancel API propagates without starting the poller.
- **`CancelSubscriptionDialogContent.test.ts` (8)** — locks the dialog
half of the behavior change: a rejected `cancelSubscription` shows an
error toast and keeps the dialog open; success closes the dialog with a
success toast.
- **`useSubscriptionCheckout.test.ts`** — unchanged, confirms
fire-and-forget callers are unaffected.
Both hardening regressions were proven red→green locally: before the fix
the terminal-hang test timed out at 5s and the duplicate-start test
resolved `pending`. Gates: vue-tsc / oxlint type-aware / eslint / oxfmt
clean; full CI green including all Playwright shards and the cloud
project. The existing `cancelSubscriptionDialog.spec.ts` e2e (@ui,
open/close/escape flows) is unchanged and green; cloud-backend e2e for
billing flows is tracked separately in FE-991.
Fixes FE-970
## Summary
Simplifies the Missing Models error card as the fifth slice of the
catalog-driven error-tab redesign. This PR is intentionally larger than
the previous slices because Missing Models is the only remaining error
type where the card UI, OSS download flow, Cloud import flow, and shared
model-import dialog all have to move together to preserve the resolution
path.
The high-level goal is to make Missing Models behave like the other
simplified error cards: show the exact missing item, show the affected
nodes, keep locate actions predictable, and only expose actions that can
actually resolve the problem.
This follows the staged error-tab cleanup plan:
1. #12683 refined validation, runtime, and prompt error presentation.
2. #12705 simplified missing media error presentation.
3. #12735 simplified missing node pack error presentation.
4. #12768 simplified swap node error presentation.
5. This PR simplifies missing model presentation and the model-import
handoff.
## Why This PR Is Larger
Missing Models has more resolution paths than the previous error groups:
- OSS can refresh model state, download individual models, and download
all available models.
- Cloud cannot download directly from the panel; it resolves supported
rows through the model import dialog.
- Some Cloud rows cannot be resolved through import at all because the
node/widget cannot consume imported model assets.
- Importing from Cloud needs to know the originating missing-model row
so it can lock the expected model type and apply the imported model back
to the affected widgets.
- Already-imported files can still be unusable if they were imported
under a different model type than the missing node expects.
Because of those constraints, splitting the card layout from the dialog
handoff would leave either a misleading Import button or an import
dialog that does not know what it is resolving. This PR keeps that
behavior in one reviewable unit.
## User-Facing Behavior
### Shared Missing Models Card
- Replaces the older grouped presentation with compact model rows.
- Shows each missing model as the primary row label.
- Shows model metadata as a smaller sublabel instead of using large
section headers.
- Keeps locate-node controls visually consistent with the other
simplified error cards.
- Keeps rows expandable when multiple nodes reference the same missing
model.
- Shows the affected node rows under expanded models.
- Allows single-reference rows to locate the affected node without
rendering an extra duplicate child row.
- Keeps unknown rows visible, including their affected nodes, instead of
silently hiding them.
- Removes the old library-select UI from the Missing Models card.
### OSS Behavior
- Keeps the refresh action available from the Missing Models group
header.
- Keeps individual Download actions for downloadable models.
- Moves file size out of the Download button label and into the row
sublabel.
- Keeps Download all when multiple downloadable models are available.
- Places Download all at the bottom of the card rather than competing
with the group header.
- Leaves rows without a download URL as non-downloadable instead of
rendering a broken action.
### Cloud Behavior
- Shows Import only for missing models that can be resolved by importing
a model asset of the required type.
- Separates models that cannot be resolved through Cloud import into an
Import Not Supported section.
- Gives unsupported rows a direct explanation: nodes referencing those
models do not support imported models, so users need to open the node
and choose a supported built-in model or replace the node with a
supported loader.
- Treats unknown model type/directory as unsupported for Cloud import,
because the import dialog cannot lock a valid model type and the node
cannot safely consume the imported asset.
- Keeps affected nodes visible in the unsupported section so users still
have a path to locate and replace the node manually.
## Cloud Import Dialog Changes
The shared model import dialog now accepts missing-model context when
opened from the Missing Models card.
When that context is present:
- The dialog shows which missing model will be replaced.
- The dialog lists the affected node/widget references that will be
updated.
- The model type selector is locked to the required model
directory/type.
- The Back/import-another path is disabled when it would break the
targeted missing-model flow.
- Import progress can be associated with the originating missing-model
row.
- After import completion, matching missing-model references are applied
automatically where possible.
- If the selected file is already imported under an incompatible model
type, the dialog shows a targeted failure state explaining why this
import cannot resolve the missing model.
This keeps the generic import dialog reusable while adding only the
context-specific behavior needed for Missing Models.
## Implementation Notes
- `MissingModelCard.vue` owns the card-level grouping and OSS/Cloud
section decisions.
- `MissingModelRow.vue` owns per-model row rendering, expansion, locate
actions, import/download actions, and row-level progress states.
- `useMissingModelInteractions.ts` remains the interaction layer for
locating nodes and applying resolved model selections.
- `UploadModelDialog.vue`, `UploadModelConfirmation.vue`,
`UploadModelFooter.vue`, `UploadModelProgress.vue`, and
`useUploadModelWizard.ts` receive the missing-model context needed by
the Cloud import handoff.
- `MissingModelLibrarySelect.vue` is removed because the simplified card
no longer exposes that inline selection path.
- Locale and selector changes are limited to the new simplified
row/section states and removed unused Missing Models strings.
## Tests Added / Updated
- Unit coverage for Missing Models card grouping and row states.
- Unit coverage for importable vs unsupported Cloud rows.
- Unit coverage for model row expansion, locate actions, progress
display, and action availability.
- Unit coverage for upload confirmation/footer/progress behavior when a
missing-model context is present.
- Unit coverage for incompatible already-imported model handling.
- E2E coverage for OSS Missing Models presentation.
- E2E coverage for mode-aware Missing Models interactions.
- Cloud E2E coverage for importable rows vs Import Not Supported rows.
- Cloud E2E coverage for opening the import dialog with missing-model
replacement context.
## Review Focus
- Cloud import eligibility: unsupported or unknown model rows should not
expose Import as if the row can be resolved automatically.
- Missing-model context in the import dialog: the required model type
should be locked, and the affected node/widget references should be
clear.
- OSS parity: OSS should keep refresh, individual Download, and Download
all while visually matching the simplified Cloud card where possible.
- Narrow side panel behavior: row labels may wrap, but link, primary
action, and locate controls should not overlap.
- Scope boundaries: this PR intentionally does not redesign Missing Node
Pack / Swap Node / Missing Media again; visual parity issues shared
across those cards can be handled in a follow-up unification pass if
needed.
## Validation
- `pnpm format`
- `pnpm lint`
- `pnpm typecheck`
- Related unit tests: 8 files / 84 tests passed
- `pnpm build`
- OSS Missing Models E2E: `errorsTabMissingModels.spec.ts` passed, 8/8
- Mode-aware Missing Models E2E subset passed, 11/11, excluding
unrelated local paste clipboard cases
- `pnpm build:cloud`
- Cloud Missing Models E2E: `errorsTabCloudMissingModels.spec.ts`
passed, 3/3
- Final Claude review: no Blocker or Major findings
## Breaking / Dependencies
- Breaking: none.
- Dependencies: none.
## Screenshots
OSS
<img width="575" height="393" alt="스크린샷 2026-06-12 오전 12 25 27"
src="https://github.com/user-attachments/assets/f5c44f95-711a-4d3d-99bd-f39ac2bb2012"
/>
<img width="659" height="351" alt="스크린샷 2026-06-12 오전 12 24 37"
src="https://github.com/user-attachments/assets/4bb65a47-c1aa-408b-836b-a1998412f815"
/>
Cloud
<img width="688" height="357" alt="스크린샷 2026-06-12 오전 12 23 59"
src="https://github.com/user-attachments/assets/9330a7e7-9f22-420f-82b3-dde0fb2b3dd1"
/>
<img width="531" height="437" alt="스크린샷 2026-06-12 오전 12 21 13"
src="https://github.com/user-attachments/assets/734bd911-f6f7-4872-8868-bb927ddeedd8"
/>
New import model flow
https://github.com/user-attachments/assets/c094c670-62b9-47ce-bfe1-2d09f4f7359d
## Summary
Update the oxc toolchain to 1.69 to fix the Windows fixed-size allocator
OOM that made oxlint commit ~130 GB per process and intermittently panic
with "Insufficient memory to create fixed-size allocator pool".
## Changes
- **What**: Bump catalog deps — `oxlint` 1.59.0 → 1.69.0,
`oxlint-tsgolint` 0.20.0 → 0.23.0, `oxfmt` 0.44.0 → 0.54.0,
`eslint-plugin-oxlint` 1.59.0 → 1.69.0. oxlint 1.69 includes the Windows
`VirtualAlloc` reserve/commit fix (oxc-project/oxc#22124, shipped in oxc
0.131.0).
- **What**: Add `--no-error-on-unmatched-pattern` to the lint-staged
oxlint command. oxlint 1.69 now exits non-zero when no lintable files
match (e.g. a yaml/lock-only commit), which broke the pre-commit hook;
this mirrors the existing stylelint `--allow-empty-input`.
- **Dependencies**: oxc toolchain only; no runtime/app dependencies
changed.
## Review Focus
- `pnpm format:check` reports no reformatting (oxfmt 0.54 produces
identical output on all 3655 files).
- `pnpm lint` passes (0 errors). Note: oxlint 1.69 adds
`typescript/no-useless-default-assignment`, which surfaces ~840
non-blocking warnings repo-wide (not auto-fixable). Left enabled here;
happy to disable in `.oxlintrc.json` in a follow-up if preferred.
- Verified locally that the new oxlint LSP commits ~0 GB instead of ~130
GB.
---------
Co-authored-by: Amp <amp@ampcode.com>
Co-authored-by: GitHub Action <action@github.com>
## Summary
PLY, SPZ, SPLAT and KSPLAT have no THREE.js exporter, so the 3D node now
offers them as export options only when the loaded file already uses
that format, and exports the original file unchanged instead of
attempting a conversion. Mesh files keep the convertible GLB/OBJ/STL/FBX
set.
Splat models (spz/splat/ksplat and gaussian-splat ply) previously had
exportable: false, which hid the export UI entirely; enable it so these
formats can be downloaded as-is. Also fixes STL export dropping its
originalURL fast-path arg.
## Screenshots (if applicable)
https://github.com/user-attachments/assets/8417c697-eb25-4cfb-af44-3855b814fa5d
## Summary
Behind `unified_cloud_auth` (default OFF), extract the `/auth/token`
Cloud-JWT mint out of `switchWorkspace` and add a parallel mint/refresh
lifecycle that writes a dedicated, **dormant** `unifiedToken` slot —
read by no consumer until PR3 — so this PR alone cannot change which
token any request carries.
Stacked on #12702 (PR1, flag registration). Base will auto-retarget to
`main` once #12702 merges.
## Changes
- **What**:
- Extract `requestToken(workspaceId?)` (network + parse only) from
`switchWorkspace`. An id-less `{}` body mints the personal-workspace
token; a concrete `workspace_id` keeps the legacy body byte-identical.
The legacy `switchWorkspace`/`refreshToken` path is behaviorally
unchanged (still owns its own state writes, `isLoading`, request-id, and
`scheduleTokenRefresh`).
- `mintAtLogin()` mints the personal default into `unifiedToken`, gated
on `unifiedCloudAuthEnabled` **only** (decoupled from
`teamWorkspacesEnabled`). Silent — no `isLoading` flash.
- `refreshUnified()` — parallel buffer-based refresh off the parsed
`expires_at` (reuses `TOKEN_REFRESH_BUFFER_MS`, no hardcoded TTL). The
legacy `refreshToken` is left untouched so its team-workspaces gate is
preserved.
- `remintUnifiedOnce()` — guarded single re-mint primitive for PR3's 401
path; a shared `unifiedRefreshRequestId` stale-guard prevents concurrent
mints clobbering each other (last mint to start wins).
- `clearWorkspaceContext()` tears down the unified slot + timer on
logout.
- `authStore`: mint at login for cloud users; add
`notifyTokenRefreshed()` and gate the Firebase `onIdTokenChanged`
rotation bump off under the flag (the unified lifecycle becomes the sole
rotation driver — no double rotation).
- **Breaking**: None. Every flag-OFF path is byte-for-byte the current
cascade.
## Review Focus
- **Dormancy**: `unifiedToken` is written only by the flag-gated mint
lifecycle and read by no consumer in this PR (the
`getAuthHeader`/`getAuthToken` flip is PR3). Tests assert
`workspaceToken` is never touched by `mintAtLogin`.
- **Legacy parity**: the `requestToken` extraction is a pure refactor —
the full existing `switchWorkspace`/`refreshToken` suite is the
regression net and stays green; flag-OFF + team-workspaces-OFF fires
zero network from any timer.
- **Concurrency**: `unifiedRefreshRequestId` stale-guard covers a
401-driven re-mint racing the scheduled refresh.
- **Rotation**: `notifyTokenRefreshed` fires only on a refresh re-mint
(never the initial login mint or a switch), and the legacy L129 bump is
gated off under the flag.
Tests cover legacy parity, the dormant slot, buffer-based refresh
(re-mint fires off parsed expiry), rotation-trigger semantics, the
single re-mint primitive (no loop on persistent 401), the concurrency
stale-guard, logout teardown, and full flag-OFF dormancy.
Part of FE-950 (single Cloud-JWT provider at login, Phase 1). Follow-up:
PR3 flips consumers + adds the 401-retry interceptor.
In moving styling to `documentElement`, #9516 introduced a regression
preventing themes from styling nodes.
| Before | After |
| ------ | ----- |
| <img width="360" alt="before"
src="https://github.com/user-attachments/assets/0a27ea1b-ff15-4524-b491-a0de9fee6ed2"
/> | <img width="360" alt="after"
src="https://github.com/user-attachments/assets/36aee446-6b7d-4b05-96de-39b19989af0d"
/>|
Note: Some elements (like the app mode toggle) are themed using
different variables (`--secondary-background` instead of
`--component-node-background`). I think the sanest approach is to define
`--secondary-background` to be `--component-node-background` by default,
but that feels better handled as a followup PR
---------
Co-authored-by: Alexander Brown <drjkl@comfy.org>
Co-authored-by: github-actions <github-actions@github.com>
## Summary
Adds the frontend telemetry attribution needed to analyze settings,
app-mode, and sharing funnel usage for MAR-321: re-enables three funnel
events that were disabled by default, attaches
app-mode/view-mode/dock-state context to UI click, run, and share
events, and adds a per-session `shell_layout` snapshot plus
right-side-panel toggle tracking.
## Changes
- **What**:
- Removes `setting_changed`, `template_filter_changed`, and
`ui_button_click` from the code-default `DEFAULT_DISABLED_EVENTS` lists
in the Mixpanel and PostHog providers, so these events now send by
default (see deployment note).
- `ui_button_click` now requires an `element_group`; all call sites are
tagged (`sidebar`, `queue`, `actionbar`, `breadcrumb`, `error_dialog`,
`errors_panel`, `graph_menu`, `graph_node`, `selection_toolbox`,
`node_library`, `workflow_actions`, `cloud_notification`, `app_mode`,
`top_menu`, `right_side_panel`) and the GTM provider forwards the field.
- Run events (`run_button_clicked`, GTM `run_workflow`) now carry
required `view_mode`/`is_app_mode` plus a new `dock_state`
(`docked`/`floating`), read from the `Comfy.MenuPosition.Docked`
localStorage key by a new `getActionbarDockState()` util.
- Share funnel events (`share_flow`, `share_link_opened`,
`shared_workflow_run`) now carry required `view_mode`/`is_app_mode`. A
new `useShareFlowContext()` composable dedupes the source/view-mode
context across the share dialog, URL copy field, and `useShareDialog`.
GTM `share_flow` forwards the new fields and still omits `share_id`.
- `shared_workflow_run` attribution is snapshotted onto the queued job
at queue time, so switching app/graph mode while a job runs no longer
misattributes the completion event (falls back to live values when no
snapshot exists).
- New `shell_layout` event fired once per session at graph-ready (cloud
only): `view_mode`, `is_app_mode`, `dock_state`, `actionbar_position`,
`active_sidebar_tab`, `right_side_panel_open`, `bottom_panel_open`,
`open_workflow_tabs`. Forwarded by Mixpanel and PostHog; not sent to
GTM.
- The right side panel open button (top menu) and close button now fire
`ui_button_click` (`right_side_panel_opened`/`right_side_panel_closed`),
covering the panel open-rate gap.
- **Dependencies**: None.
## Review Focus
- `view_mode`/`is_app_mode` changed from optional to required (typed as
`AppMode`) on run/share metadata — check no call sites were missed.
- The queue-time snapshot in `executionStore` (`queuedJob.viewMode ??
mode.value`) and its regression test.
- Share IDs remain limited to the providers/events that already carry
share attribution (GTM still strips `share_id`).
- `shell_layout` cadence is once per session (graph-ready idle
callback), matching the gap analysis's "session snapshot" wording.
Linear: MAR-321
Validation:
- `pnpm test:unit
src/platform/telemetry/providers/cloud/PostHogTelemetryProvider.test.ts
src/platform/telemetry/providers/cloud/MixpanelTelemetryProvider.test.ts
src/platform/telemetry/providers/cloud/GtmTelemetryProvider.test.ts
src/platform/telemetry/utils/getShellLayoutSnapshot.test.ts
src/platform/workflow/sharing/components/ShareWorkflowDialogContent.test.ts
src/platform/workflow/sharing/composables/useSharedWorkflowUrlLoader.test.ts
src/stores/executionStore.test.ts src/components/TopMenuSection.test.ts
src/components/graph/selectionToolbox/InfoButton.test.ts
src/components/rightSidePanel/errors/useErrorActions.test.ts
src/views/GraphView.test.ts`
- `pnpm typecheck`
- `pnpm lint`
- `pnpm knip`
- `git diff --check`
## Deployment note
`telemetry_disabled_events` is currently unset in the prod/staging/test
dynamicconfig rows, so the code-default change here is what enables
these events. The remote value remains available as a kill switch, but
it **replaces** the code defaults rather than merging: if ops sets it to
re-disable an event, the list must include every event that should stay
disabled (`tab_count_tracking`, `node_search`,
`node_search_result_selected`, `help_center_*`, `workflow_created`), not
just the new ones.
## What
Implements **B2 — Complete the billing facade** from the FE
billing-divergence survey. Adds the members missing from the shared
`BillingContext` so components stop bypassing `useBillingContext` with
raw `workspaceApi` calls.
Part of the billing convergence plan — **FE-904** (parent **FE-903**).
## Why this PR — an *enabling* refactor (near-zero standalone user
value)
On its own B2 changes no endpoint and is user-invisible (see
**Behavioral impact**). Its entire purpose is to be the **prerequisite**
that unblocks the rest of the convergence — it gives the facade a single
entry point and the missing capability/state surface that the next
levers depend on:
- **Unblocks B3 — repoint direct-bypass consumers (the next PR; a live
bug fix).** `SubscribeButton` (`current_tier`) and
`PostHogTelemetryProvider` (the `subscription_tier` person property)
currently read the **legacy** `useSubscription` tier, so the value is
**stale/empty for team users today** (telemetry + analytics are wrong
right now). They can only be repointed to a correct, workspace-aware
tier by sourcing it from the facade — which requires the **`tier`** (and
`renewalDate`, for `FreeTierDialog`) fields **this PR adds**. Without B2
there is literally no facade `tier` to read.
- **Unblocks B6 — orientation banners.** The 6 billing-state banners
need `billingStatus` / `subscriptionStatus`, exposed here.
- **Unblocks B1 — dispatcher flip (personal → workspace path).** B1 can
only collapse the personal/team fork once (a) every billing operation
flows through the facade — no raw `workspaceApi` bypass left — and (b)
the facade actually supports `resubscribe`/`topup`. This PR removes the
last bypasses and completes the action surface so the unified personal
path will work. (B1 itself stays gated on the BE-DATA unification.)
## Changes
- **Contract** (`composables/billing/types.ts`): `BillingActions` gains
`resubscribe()` and `topup(amountCents)`; `BillingState` gains
`billingStatus`, `subscriptionStatus`, `tier`, `renewalDate`. Exported
`BillingStatus`, `BillingSubscriptionStatus`, `CreateTopupResponse` from
`workspaceApi`.
- **Workspace adapter** (`useWorkspaceBilling`): real wiring —
`workspaceApi.resubscribe()` / `createTopup()`, surfaces
`statusData.billing_status` / `subscription_status` /
`subscription_tier` / `renewal_date`.
- **Legacy adapter** (`useLegacyBilling`): equivalents — `resubscribe` =
fresh checkout via `useSubscription`; `topup` converts **cents →
dollars** through `purchaseCredits`; `billingStatus` = `null` (no legacy
concept); `subscriptionStatus` synthesized from active/cancelled flags.
- **Dispatcher** (`useBillingContext`): proxies the new members.
- **Orphaned callers migrated** off raw `workspaceApi`:
- `SubscriptionPanelContentWorkspace.vue` → `resubscribe()`
- `useSubscriptionCheckout.ts` → `resubscribe()`
- `TopUpCreditsDialogContentWorkspace.vue` → `topup(amountCents)`
## Notes
- **Unit divergence absorbed:** the facade standardizes `topup` on
**cents**; the legacy adapter converts to dollars for
`/customers/credit`.
- **FE-only, no backend dependency** — safe to merge/deploy standalone;
independent of the B1 dispatcher flip (which is gated on the BE-DATA
unification).
## Behavioral impact (verified — safe to merge/deploy standalone)
This is a structural refactor: **endpoints, request payloads, and fetch
counts are unchanged**, and there is **no user-visible change** on
OSS/Desktop or Cloud-personal.
- **OSS / Desktop** (`teamWorkspacesEnabled` off): no change. The only
B2 code that runs is the eager `useAuthActions()` in `useLegacyBilling`
setup — side-effect-free, and already instantiated transitively via
`useSubscription` today. New computeds are lazy with zero readers; new
legacy `resubscribe`/`topup` are never invoked (their callers are
team-only surfaces).
- **Cloud personal**: no change. The migrated handlers are structurally
unreachable on the legacy path (dialog/panel variant gating,
`isCancelled` gated to `!isInPersonalWorkspace`).
- **Cloud team**: same endpoints/payloads/refresh counts. **One
intentional behavioral nuance:** routing `resubscribe`/`topup` through
the facade now toggles the shared `useBillingContext().isLoading` flag
during the call (the previous raw `workspaceApi` calls did not). This is
deliberate — it aligns these two with every other facade action
(`subscribe`, `cancelSubscription`, …). Net effect is at most a brief
loading-indicator flicker in the subscription panel; no change to
network, ordering, or state correctness.
> Follow-up (pre-existing, out of scope): **FE-932** — a completed
top-up refreshes balance but not status, so `subscription.hasFunds` can
be briefly stale. Predates B2 (`main` did balance-only too); to be fixed
with B6.
> Import-cycle note: this closes `useBillingContext → useLegacyBilling →
useAuthActions → useBillingContext`. It is module-eval safe — every
cross-cycle call is at composable-runtime, none at module top level.
## Verification
- `vue-tsc --noEmit`: clean.
- `oxlint --type-aware` on touched files: 0 errors / 0 warnings.
- Runtime no-op confirmed by an adversarial code-path review across the
3 build targets (OSS / Cloud-personal / Cloud-team).
- eslint + unit tests: deferred to CI.
Survey: **FE Billing API Divergence — Personal vs Team Workspace**
(Notion) — notes D4, P6, T1, E7, E9.
> Draft: opened for early review of the facade shape and the
legacy-equivalent semantics (esp. legacy `resubscribe` = fresh checkout,
and the cents/dollars conversion).
---------
Co-authored-by: Alexander Brown <drjkl@comfy.org>
## ELI-5
After you click Continue on the sign-in consent page, the page sends
your browser back to the app that asked. Our safety check only knew
about web-style addresses (`http://...`), so when the iOS app — whose
return address looks like `org.comfy.ios://...` — finished sign-in, the
page refused to deliver and showed "OAuth request failed." The fix:
instead of the page keeping a list of address styles it trusts, it now
asks the backend "what return address did this app register?" and goes
exactly there or nowhere. Truly dangerous addresses (ones that run code
in the page) stay banned outright.
## Problem
The consent success handler hard-allowlists `http(s)` for the
post-consent redirect (`oauthApi.ts`). That covers the loopback
redirects `comfy-desktop`/`comfy-cli` register, but rejects RFC 8252
reverse-DNS custom schemes — the callback shape native-app OAuth clients
use.
Live failure (prod, 2026-06-11, first `comfy-ios` sign-in test): user
approves consent → backend persists consent, consumes the auth request,
and mints an authorization code for `org.comfy.ios://oauth-callback` →
frontend throws `'unsafe scheme'` → user sees the generic **"OAuth
request failed"** → the code expires unused 60s later. Verified
end-to-end in the prod DB.
## Fix (final design — binding, not scheme lists)
Bind the post-consent navigation to the **challenge's registered
`redirect_uri`** (scheme + authority + path equality; the server only
appends `code`/`state` query params to the registered URI). The backend
supplies that field per-request — Comfy-Org/cloud#4230 — so the frontend
carries **zero per-client knowledge**: registering a future native
client is a backend-only change.
Layers:
1. **Executable-scheme denylist**
(`javascript:`/`data:`/`blob:`/`vbscript:`/`about:`) — unconditional;
the actual XSS line.
2. **Registration binding** when `challenge.redirect_uri` is present —
also rejects wrong-client redirects, which no scheme policy could.
3. **http(s)-only fallback** when the challenge doesn't surface
`redirect_uri` (older backend) — preserves today's behavior; the two PRs
can land in either order, but iOS sign-in needs both.
Also per the earlier review pass: navigation uses the parsed URL
(parser/sink consistency), malformed URLs throw structured errors,
single-navigation asserted.
## History
This PR went through three designs: dotted-scheme heuristic → four-lab
adversarial review (68 findings, Opus-judge consolidated) flagged the
heuristic as bypassable → exact scheme allowlist → product feedback
(hardcoding per-client schemes in shared frontend code doesn't scale and
shouldn't exist for non-product test apps) → registration binding, which
the review panel had independently flagged as the strongest option.
41/41 oauth tests passing.
## Tests
- Navigates: bound custom-scheme redirect
(`org.comfy.ios://oauth-callback?code=…` vs registered
`org.comfy.ios://oauth-callback`), http loopback (legacy fallback)
- Rejects: unbound custom scheme (fallback), wrong-client redirect
(`com.evil.app://…` vs registered iOS URI), path mutation, executable
schemes even when 'registered', malformed URLs
Related: BE-1341, BE-1350, Comfy-Org/cloud#4230.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
When performing mode toggle operations (like bypass or mute) with a
group (the colored rectangles) selected, nodes contained within the
group will be considered selected and will have their state toggled.
<img width="1024" height="1024" alt="AnimateDiff_00002"
src="https://github.com/user-attachments/assets/c4e9db17-3fe8-4fd8-9012-0e9a0bc59707"
/>
## Summary
Cleanup of the website download button: deduplicate the label lookup in
`DownloadLocalButton.vue`, and export a `Platform` domain type from
`useDownloadUrl` so the button spec, icon map, and analytics all consume
it instead of loose `string`s.
## Changes
- **What**:
- Hoist `t('download.hero.downloadLocal', locale)` into a single `label`
computed, reused by the fallback aria-labels and the button text
(previously evaluated in two places)
- Export `Platform` (`'windows' | 'mac'`) from `useDownloadUrl`;
`ButtonSpec.key`, the `ICONS` map (`Record<Platform, string>`), and
`captureDownloadClick` now consume it — adding a platform without an
icon becomes a compile error, and typo'd platform names can't reach
PostHog
- Drop the redundant `aria-hidden="true"` on the icon (`alt=""` already
marks it decorative)
## Review Focus
`ppformula-text-center` on the icon is intentionally kept — it is
positional (`position: relative; top: 0.19em`), not typographic, and
keeps the icon aligned with the optically-shifted PP Formula text.
No behavior change.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
## Summary
- Rebrand "Comfy Local" → "Comfy Desktop" across all user-facing copy on
`apps/website`. The nav already showed "Comfy Desktop"; this pulls the
rest of the site in line.
- Updates the homepage product card, the `/cloud` page FAQ + pricing
copy, the `/download` page SEO title/keywords, and matching Storybook
stories. Both `en` and `zh-CN`.
- Internal identifiers left untouched (i18n key names, the
`src/components/product/local/` folder, the `Product = 'local' | …` type
discriminator, and the `media.comfy.org/website/local/*` CDN paths).
## Test plan
- [ ] `pnpm --filter @comfyorg/website dev` and visually verify:
- [ ] `/` — product card section reads "Comfy Desktop" / "SEE DESKTOP
FEATURES"
- [ ] `/cloud` (en + zh-CN) — every FAQ "Local" reference now reads
"Comfy Desktop"
- [ ] `/download` — page `<title>` and meta keywords show new values
- [ ] Nav + footer still read "Comfy Desktop" / "DOWNLOAD DESKTOP"
(unchanged)
- [ ] `pnpm --filter @comfyorg/website test:e2e -- navigation.spec.ts`
passes
- [ ] Storybook: `Website/Common/ProductCard` and
`Website/Common/FooterLinkColumn` render with new labels
- [ ] Regenerate any screenshot baselines that snapshot the homepage
product cards or the `/cloud` FAQ section
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: github-actions <github-actions@github.com>
## Summary
Simplifies the Swap Nodes error card as the fourth slice of the
catalog/error-tab presentation refactor, aligning it with the newer
compact error-row patterns while preserving the existing replace and
locate behavior.
This follows the staged rollout plan from the earlier error-tab PRs:
1. #12683 refined execution-style errors: validation, runtime, and
prompt errors.
2. #12705 simplified missing media errors into flat, locatable rows.
3. #12735 simplified missing node pack errors and aligned grouped-row
behavior.
4. This PR applies the same simplification pass to Swap Nodes errors.
5. A later PR is expected to handle Missing Models, which is larger and
intentionally kept separate.
After the Missing Models slice lands, a follow-up consistency PR will
normalize the shared row/disclosure pattern across Missing Node Packs,
Swap Nodes, and Missing Models together. That follow-up will cover
parameterized i18n labels for disclosure controls, shared text-button
styling, and consistent disclosure semantics/accessibility across those
grouped rows.
## Changes
- **What**: Reworks the Swap Nodes card rows so each replacement group
is presented as a compact row with the source node type, replacement
target, replace action, and locate action.
- **What**: For a single affected node, the visible row label can be
clicked to locate the node, matching the interaction model used by the
newer missing-media and missing-node rows.
- **What**: For multiple affected nodes with the same replacement
target, the group renders a count badge and a disclosure row. Expanding
the group shows the affected node rows, each with its own locate action.
- **What**: Removes the old node-id badge path from Swap Nodes rows.
Node-id badges remain available to the other error cards that still own
that behavior.
- **What**: Keeps replacement behavior unchanged: per-group replacement
and replace-all still call through the existing node replacement store
flow.
- **What**: Adds regression coverage for the new grouped-row UI,
including same-type grouping in both Vue Nodes and LiteGraph render
modes.
- **Breaking**: None.
- **Dependencies**: None.
## Review Focus
Please focus on the Swap Nodes presentation and interaction symmetry
with the previous error-tab PRs:
- Single-node groups should remain directly locatable via the row label
and the locate icon.
- Multi-node groups should expose the count and expand/collapse behavior
without adding duplicate focusable disclosure controls.
- The visible row labels intentionally keep their own accessible names,
while the separate locate icon uses the generic `Locate node on canvas`
accessible name. This mirrors the established pattern from the previous
slices.
- The newly added Playwright fixture covers two same-type replaceable
nodes so duplicate group keys and grouped disclosure behavior are
exercised end-to-end.
## Validation
- `pnpm format`
- `pnpm test:unit
src/platform/nodeReplacement/components/SwapNodeGroupRow.test.ts`
- `pnpm test:browser:local browser_tests/tests/nodeReplacement.spec.ts
--project=chromium`
- Pre-commit hook: lint-staged, stylelint, oxfmt, oxlint, eslint, `pnpm
typecheck`, `pnpm typecheck:browser`
- Pre-push hook: `pnpm knip --cache`
- Additional parallel code review pass completed locally; no blocker or
major issues remained.
## Screenshots (if applicable)
This PR
<img width="561" height="362" alt="스크린샷 2026-06-11 오전 3 46 06"
src="https://github.com/user-attachments/assets/65395467-6c2f-4aa1-84c5-3d9614c00c80"
/>
old (Main)
<img width="611" height="798" alt="스크린샷 2026-06-11 오전 3 46 32"
src="https://github.com/user-attachments/assets/3862d5df-f839-40c0-9488-ce64b051378e"
/>
## Summary
Every `VirtualGrid` consumer (assets sidebar, manager dialog, widget
select dropdown) is blind to discrete mouse-wheel scrolling:
`useScroll`'s `throttle: 64` never reports scroll position, so the
virtualization window stays frozen — users see blank space below the
first viewport of items and `approach-end` (infinite scroll) never
fires. Trackpad scrolling masks the bug by emitting events faster than
the throttle window.
## Changes
- **What**: drop the `throttle` option from `useScroll` in `VirtualGrid`
and remove the `scrollThrottle` prop (no consumer passes it). Scroll
events are frame-aligned and the handler is cheap, so the throttle
bought nothing even when it worked.
- **Root cause**: VueUse ≥14 `throttleFilter` with `leading=false` (what
`useScroll` uses) marks spaced-out events as executed without executing
them — each event re-arms an `isLeading` restore timer that makes the
next event skip its invoke, and the trailing branch is unreachable when
`elapsed > duration`. Regression of vueuse#2390; still present on vueuse
`main`.
## Review Focus
- Verified live against staging: with the throttle, sidebar scrolled to
`scrollTop` 1250 while `scrollY` stayed 0 and the render window stayed
at `[0..3)` of 27 (blank viewport); with this fix, `scrollY` tracks 1:1
and the window advances. Bare-vs-throttled `useScroll` compared
side-by-side on the same element to isolate the cause.
- Unblocks wheel-scroll for #12780's dropdown infinite scroll with no
changes there.
- Fixes FE-990
🤖 Generated with [Claude Code](https://claude.com/claude-code)
## Summary
The rebranded `favicon.ico` renders with **opaque white corners** in
browsers — the rounded mark shows white slivers around its corners on
any background. This is a decode bug, not a design issue: the ICO
contains **BMP-format frames** whose alpha channel Chrome (and other
consumers) mishandle. Verified by loading the raw `.ico` in headless
Chrome on a dark page: corners render white instead of transparent.
Every surface that consumes the `.ico` directly shows the artifact —
Google search results, connector icon scraping, raw image views — while
browser tabs look fine because they prefer the SVG favicon.
## Changes
- **What**: Re-encode `apps/website/public/favicon.ico` and
`public/assets/favicon.ico` with **PNG-format ICO frames** (16/32/48).
PNG frames carry unambiguous alpha and decode correctly in all modern
browsers.
- **No design change**: identical rounded artwork, same transparency,
same sizes and filenames. SVG, PNGs, apple-touch-icon, and manifest
icons are untouched.
- **Breaking**: none.
## Review Focus
- Headless-Chrome verified: the old ico renders white corners on a dark
page; the re-encoded one renders transparent corners. (Comparison in PR
comments.)
- PNG-in-ICO is supported by all modern browsers and Google's favicon
pipeline.
- After merge, please add `needs-backport` + `cloud/1.45` so
cloud.comfy.org's copy gets the same fix.
## Screenshots (if applicable)
---------
Co-authored-by: GitHub Action <action@github.com>
Adds `comfy_usage_source: 'comfyui-frontend'` to the prompt body's
`extra_data`. The backend forwards this to API nodes' upstream requests
via the `Comfy-Usage-Source` header, so partner node API usage can be
attributed to the frontend.
Used in https://github.com/Comfy-Org/ComfyUI/pull/14404
Reads `system.deploy_environment` from `/system_stats` (added in
Comfy-Org/ComfyUI#14402) and sends it as the `Comfy-Env` header when
fetching `/releases`, matching the header name the backend already uses
for outbound API node requests. The header is omitted when the backend
doesn't report the field, so older backends are unaffected.
Note: api.comfy.org must allow `Comfy-Env` in
`Access-Control-Allow-Headers` for the CORS preflight to pass.
## ELI-5
When the app needs *all* the assets for a tag (like every input image),
it asks the server for them one page at a time. Today it says "give me
page starting at item #500" (offset paging). If items get added or
removed while it's flipping through, pages shift and it can show the
same thing twice or skip something.
This switches to "give me the page *after this bookmark*" (cursor
paging). The server now hands back a `next_cursor` bookmark with each
page; we pass it to fetch the next one. Bookmarks don't slip when the
list changes underneath, so the walk is stable and drift-free.
## What
Migrates the full-walk asset pager (`getAllAssetsByTag`) from offset to
keyset (`after` / `next_cursor`) pagination, now that the list-assets
endpoint exposes a cursor contract in the generated types.
- `handleAssetRequest` accepts an `after` cursor and sends it instead of
`offset` when present (the server ignores `offset` alongside a cursor)
- `getAllAssetsByTag` resumes each page from the prior response's
`next_cursor`, and terminates when `has_more` is false or `next_cursor`
is omitted
- `next_cursor` is exposed on the asset response schema; `after` is
threaded through `getAssetsByTag` / `getAssetsPageByTag` for
cursor-aware callers
- offset remains supported for random-access callers; only the full-walk
path changes
## Why
Offset pagination double-counts or skips records when the underlying set
changes mid-walk. Keyset cursors are stable under concurrent
inserts/deletes and scale better than deep offsets.
## Stacking
Based on `update-ingest-types` because the `after`/`next_cursor` types
land there first; this targets that branch and will retarget to the
default branch once it merges. Changes here touch only the asset
service/schema, disjoint from the generated types.
## Follow-ups
The asset store's bespoke offset loops (model loader, flat-output
infinite scroll) and the missing-media resolver still walk by offset;
those migrate in separate PRs.
## Tests
`assetService.test.ts` updated to assert the cursor walk, that the first
page carries neither `after` nor `offset`, that subsequent pages resume
from `next_cursor`, and that the walk halts when `next_cursor` is absent
even if `has_more` is true. Full asset/service + missing-media + store
suites pass locally (193 tests).
---------
Co-authored-by: mattmillerai <7741082+mattmillerai@users.noreply.github.com>
Co-authored-by: GitHub Action <action@github.com>
Adds a `website:download_button_clicked` PostHog event (with `platform`
property) fired when a user clicks the desktop installer download button
on comfy.org. Previously we only had `/download` pageviews as a proxy —
autocapture is not active on the website project, so these clicks were
untracked. Includes unit tests for the new capture helper.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
## Summary
Simplify the Missing Node Packs error card so it follows the new
error-tab item-row direction, with clearer pack rows, predictable locate
behavior, and focused E2E coverage.
This is the third PR in the staged error-tab simplification plan:
1. Merged: execution/prompt/validation error presentation and catalog
grouping in #12683.
2. Merged: missing media presentation simplification in #12705.
3. This PR: missing node pack presentation simplification.
4. Planned next: swap-node presentation simplification.
5. Planned later: missing model presentation and action-flow
simplification.
## Changes
- **What**: Refactors Missing Node Packs rows so pack-level and
node-level actions are easier to scan and more consistent with the rest
of the refreshed Errors tab.
- **What**: Removes the node-id badge from missing node pack rows,
matching the simplified item-row direction.
- **What**: Makes a single-node known pack row directly locatable from
the pack label, rather than rendering an extra child row.
- **What**: Keeps multi-node packs collapsed by default, with both the
chevron and pack title toggling the child node list.
- **What**: Keeps unknown packs expanded by default, including the
single-node unknown-pack case, so users can still see the unresolved
node type immediately.
- **What**: Keeps per-node child rows clickable for locate-on-canvas
behavior when a pack contains multiple affected nodes.
- **What**: Replaces missing-node-pack action labels with shared
`g.install` and `g.search` copy and removes now-unused English locale
keys.
- **What**: Adds targeted Playwright coverage for the simplified
missing-node-pack card, including unknown-pack default rows, row-label
locate behavior, and chevron/title expansion behavior.
- **Breaking**: None.
- **Dependencies**: None.
## Review Focus
Please focus on the missing-node-pack row behavior:
- Single known pack with one affected node should stay compact and
locate the node from the pack label or locate icon.
- Known packs with multiple affected nodes should show a count, start
collapsed, and expand/collapse from either the chevron or title.
- Unknown packs should expose the affected node rows immediately,
including when there is only one affected node.
- Locate actions should remain attached to the affected node rows, not
to the parent pack when there are multiple nodes.
- The E2E fixture intentionally uses two missing nodes with the same
`cnr_id` and node sizes of `[400, 200]` to follow browser-test asset
guidance.
## Validation
- `pnpm format:check`
- `pnpm lint`
- `pnpm typecheck`
- `pnpm knip --cache` via pre-push hook
- `pnpm test:unit
src/components/rightSidePanel/errors/MissingPackGroupRow.test.ts
src/components/rightSidePanel/errors/MissingNodeCard.test.ts --run`
- `pnpm test:browser:local
browser_tests/tests/propertiesPanel/errorsTabMissingNodes.spec.ts
--project=chromium`
- Pre-commit hook: staged formatting, linting, `pnpm typecheck`, and
`pnpm typecheck:browser`
## Screenshots
This PR
<img width="531" height="598" alt="스크린샷 2026-06-10 오전 1 54 31"
src="https://github.com/user-attachments/assets/9c0addeb-92d2-4cef-a4f3-35a87bbad308"
/>
old (Main)
<img width="509" height="807" alt="스크린샷 2026-06-10 오전 1 53 51"
src="https://github.com/user-attachments/assets/b8488f73-d8ed-4356-bd4c-fc678ea205f7"
/>
## What
https://a46f3266.comfy-storybook.pages.dev/?path=/docs/components-creditslider--docs&globals=theme:dark
<img width="560" height="288" alt="cs12644_dark"
src="https://github.com/user-attachments/assets/c06b5244-d178-4fa5-8bb9-61fd8595fe9b"
/>
<img width="560" height="288" alt="cs12644_light"
src="https://github.com/user-attachments/assets/16626333-43ba-4541-bd11-faaa8513b1e8"
/>
Adds **CreditSlider** — the team-plan credit-subscription slider from
Figma **DES-197** — as a standalone presentational component. **B4
standalone slice (FE-935)**; parent **FE-934**.
## Why it can ship now (no backend dependency)
The slider's 5 stops are **locked in DES-197**, so this component is
built and reviewable independently of the (still-TBD) backend slider
contract. Wiring it into the pricing table is deferred to FE-934.
## How it works
- 5 fixed stops — **200 / 400 / 700 / 1,400 / 2,500 USD** ↔ 42,200 /
84,400 / 147,700 / 295,400 / 527,500 credits; default **$700**.
- Snap-to-stop is guaranteed by driving the shared
`src/components/ui/slider/Slider.vue` (reka-ui) in **index space**
(`:min="0" :max="4" :step="1"`) — the thumb can only land on the 5
stops, with free keyboard-arrow support + ARIA from reka-ui.
- `v-model` carries the selected **USD** value; a `change` event also
emits `{ index, usd, credits }` for the future pricing-table wiring.
- Thresholds live in a typed constant `teamPlanCreditStops.ts` (sibling
to `tierPricing.ts`), **hardcoded per DES-197** with a `TODO(FE-934)` to
source from `GET /api/billing/plans` once the BE contract lands. The
credit figures equal `usdToCredits(usd)` (rate 211); a test guards
against rate drift.
## Files
- `src/platform/cloud/subscription/components/CreditSlider.vue` (+
`CreditSlider.stories.ts`, `CreditSlider.test.ts`)
- `src/platform/cloud/subscription/constants/teamPlanCreditStops.ts`
## Verification
- `vue-tsc --noEmit`: clean.
- `oxlint --type-aware`: 0 errors / 0 warnings.
- `vitest run`: **11/11 pass** — default stop, ArrowRight/ArrowLeft snap
to the adjacent stop (never in between), `change` payload, disabled
state, BE-sourced stops override, empty `stops` renders nothing, all 5
labels render, and the credit-rate-drift guard.
## Not in scope
- Wiring into `PricingTableWorkspace` / the team-plan card (FE-934,
blocked on the BE slider contract).
- The marketing caption and card layout around the slider (parent
component's concern).
Design source: Figma **DES-197** (Team Plan / Workspaces).
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Summary
Bumping the ranking of native nodes to improve their discoverability
## Changes
- **What**: Updated ranking of native nodes in
`public/assets/sorted-custom-node-map.json`
## Summary
Fixes a Vue node subgraph case where the missing-model refresh flow
clears the missing-model error, but the promoted combo widget remains in
an invalid visual state because its hosted options snapshot is stale.
## Changes
- **What**: After `reloadNodeDefs()` refreshes combo option lists and
extension `refreshComboInNodes` hooks run, resync hosted options
snapshots for promoted combo widgets so Vue-rendered subgraph nodes see
the newly available model option.
- **What**: Adds a focused E2E regression for the missing-model refresh
path on a subgraph with a promoted `ckpt_name` widget.
- **What**: Hardens the E2E by cleaning up its `/object_info` route
override and asserting the widget's `aria-invalid` state rather than a
Tailwind implementation class.
- **Breaking**: None.
- **Dependencies**: None.
## Review Focus
This PR is intentionally a minimal patch for the missing-model refresh
path, not a broader subgraph architecture change.
Root cause: `reloadNodeDefs()` updates the live LiteGraph combo widget
options, including widgets inside subgraphs. However, Vue nodes render
promoted widgets from a hosted `WidgetState.options` snapshot. When a
missing model is downloaded and the missing-model refresh button reloads
node definitions, the source combo receives the updated model list, but
the promoted host snapshot can still contain the old option list. The
missing-model error and node-level state are cleared, while the Vue
combo still computes itself as invalid from stale options.
The fix keeps the existing host snapshot model intact. It simply resyncs
promoted combo host options after the normal combo refresh pipeline
finishes. This avoids changing promoted-widget ownership,
`useProcessedWidgets` merge precedence, or broader subgraph internals
while addressing the reported stale invalid state.
Why the helper is in `app.ts`: this sync is currently a single-call-site
post-step of `reloadNodeDefs()`, and the ordering is load-bearing. It
must run after the core combo refresh loop and after extension
`refreshComboInNodes` hooks so it captures both built-in and
extension-provided option changes. Keeping the small private helper next
to the refresh orchestration makes that sequence explicit and avoids
adding a new public subgraph helper or introducing a more visible
dependency cycle through `promotionUtils` for a narrow patch.
This is stacked on
`jaeone/fe-942-bug-error-indicators-persist-after-resolving-missing-model`,
so it is opened as a draft until the base PR lands.
## Test Plan
- `pnpm knip`
- `pnpm exec oxfmt --check src/scripts/app.ts
browser_tests/tests/propertiesPanel/errorsTabModeAware.spec.ts`
- `pnpm exec eslint --cache --no-warn-ignored src/scripts/app.ts
browser_tests/tests/propertiesPanel/errorsTabModeAware.spec.ts`
- `pnpm typecheck`
- `pnpm test:browser:local --grep "Refreshing a resolved promoted
missing model clears the combo invalid state"`
- First local run hit a `beforeEach` timeout while the dev server was
still warming custom-node/conflict-detection output.
- Re-running against the warmed dev server passed.
## Screenshots (if applicable)
N/A. The behavioral E2E covers the visible invalid-state regression.
## Summary
`comfy-logo-single.svg` is still the old **yellow-on-blue** (`#172DD7`)
ComfyUI mark, while the favicon and the rest of the brand have moved to
**dark `#211927` + yellow `#F2FF59`**. This refreshes it to match.
## Changes
- **What**: Pure color swap on
`public/assets/images/comfy-logo-single.svg` (`#172DD7`→`#211927`,
yellows normalized to `#F2FF59`). Geometry/mask unchanged.
- **Breaking**: none.
## Review Focus
- This logo is the **in-app header mark** (`ComfyOrgHeader.vue`) and the
**PWA/cloud manifest icon**, so it's a visible brand change across OSS +
cloud + desktop — flagging for **brand/design sign-off** before merge.
- `comfy-logo-mono.svg` (black/white) is intentionally left as-is.
## Summary
- Thread `share_id` through shared workflow open/import, link creation,
auth completion, and run success telemetry
- Persist share attribution on loaded workflows and queued jobs so
shared runs can be joined back to the source link
- Add provider support for `share_link_opened` and `shared_workflow_run`
events across telemetry backends
## Behavior notes
- `execution_success` is now keyed off the success event's own
`prompt_id` (looked up in `queuedJobs`) instead of `activeJobId`. This
fixes successes for non-active jobs being reported with the wrong job
id, but may slightly shift `execution_success` event volume: successes
for jobs this client never queued or saw start are no longer tracked.
- Share auth attribution (`share_auth` preserved query) is cleared if
the user cancels the shared workflow dialog, so only users who proceed
past the dialog have signups attributed to the share link.
## Testing
- Added and updated unit tests for shared workflow loading, link
creation, auth attribution, workflow service loading, and execution
success
- Unit tests, `pnpm test:unit`, and repository checks for formatting,
linting, and type coverage passed
## Summary
Set `include_previews: true` on bulk asset export so exported zips
include preview files, using the new option added to the cloud export
API.
## Changes
- **What**: Added `include_previews` to `AssetExportOptions` and pass
`true` from the bulk export path in `useMediaAssetActions`.
## Review Focus
Field name matches the updated cloud `AssetExportOptions` contract.
## Summary
- Adds the Comfy affiliate program landing page at `/affiliates`,
composed from 8 reusable presentational blocks (`HeroSplit01`,
`ChecklistSplit01`, `StepsGrid01`, `BenefitsGrid01`, `DetailsTable01`,
`BrandAssetsGrid01`, `FAQSplit01`, `CtaCenter01`) and thin section
wrappers under `templates/affiliate/*`. The page funnels through: Hero →
Audience → How It Works → Why ComfyUI → Program Details → Brand Assets →
FAQ → closing "Ready to start earning?" CTA.
- Removes the older parallel implementation under
`components/affiliates/*` (7 Vue sections, 3 data files, the i18n
contract test) and the `affiliate-landing.*` translation keys — the page
now has a single canonical i18n surface under `affiliate.*`.
- Rewrites `e2e/affiliates.spec.ts` to target the new sections via
semantic queries (no testids were added to production), covering hero
indexability, the closing CTA's apply + terms links, FAQPage JSON-LD
entry count, and FAQ toggle behavior.
- Apply CTA links to the application form
(`forms.gle/RS8L2ttcuGap4Q1v6`, `target=_blank`); terms link goes to
`/affiliates/terms` (English-only by design, per the legal-reviewed
terms document).
## Test plan
- [ ] `pnpm --filter @comfyorg/website typecheck` passes (0 errors)
- [ ] `pnpm --filter @comfyorg/website test:unit` passes (96 tests
across 13 files)
- [ ] `pnpm --filter @comfyorg/website knip --cache` reports no new
orphans introduced by this branch
- [ ] Visit `/affiliates` — all 8 sections render in order on desktop
and mobile
- [ ] APPLY NOW (in closing CTA section) opens
`forms.gle/RS8L2ttcuGap4Q1v6` in a new tab
- [ ] "Read the affiliate program terms" link navigates to
`/affiliates/terms` in the same tab
- [ ] FAQ items expand/collapse on click; `<script
type="application/ld+json">` FAQPage entry contains all 8 Q/A pairs from
`src/data/affiliateFaq.ts`
- [ ] Run e2e: `pnpm --filter @comfyorg/website playwright test
e2e/affiliates.spec.ts`
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: GitHub Action <action@github.com>
## Summary
Add a new `/models` landing page (EN + zh-CN) for the marketing site,
plus supporting tweaks: a generator-side mechanism for old→new
model-slug 301 redirects, spacing/CTA polish on shared sections, and a
per-item layout option on GalleryCard.
## Changes
- **What**:
- New `/models` and `/zh-CN/models` pages composed of a hero (autoplay
video, modelName + i18n CTA), a creations gallery, and the shared
`AIModelsSection`.
- New `ModelsHeroSection.vue` and `ModelCreationsSection.vue`
components; localised strings under `models.list.*` / `models.hero.*` in
`translations.ts`.
- Reused `AIModelsSection` on the models page (replaces the duplicated
showcase markup that previously lived under `models/`) so the same
component now powers both the cloud product page and the models page.
- `generate-models.ts`: renamed the `grok` provider to **Grok Imagine**
(`grok-imagine` slug) and added a `LEGACY_SLUG_REDIRECTS` constant that
emits stub entries so the existing `canonicalSlug` mechanism in
`[slug].astro` issues a 301 from old slugs (`grok-image` →
`grok-imagine`).
- `model-metadata.ts`: renamed the metadata key from `grok-image` to
`grok-imagine` (hub slug unchanged).
- `GalleryCard.vue`: added per-item `objectFit` / `objectPosition`
overrides so individual gallery entries can opt out of the default
`cover` crop.
- `ModelsHeroSection.vue`: dropped the empty-string default on
`videoAriaLabel` and omit `aria-label` (with `aria-hidden="true"`) when
no label is provided — addresses the CodeRabbit accessibility note.
- Minor vertical-spacing tightening on `ModelCreationsSection` and the
shared `AIModelsSection`.
- **Breaking**: None. Old `/p/supported-models/grok-image` URLs 301 to
the new slug.
- **Dependencies**: None.
## Review Focus
- The `LEGACY_SLUG_REDIRECTS` constant in
`apps/website/scripts/generate-models.ts` is the new source of truth for
renamed slugs — future slug renames should follow the same pattern so
the redirect survives regeneration of `generated-models.json`.
- The shared `AIModelsSection` is now used in two places; confirm the
spacing tweak (`py-24 → py-16`, `mt-24 → mt-16 lg:mt-24`) still looks
correct on the cloud product page.
- `generated-models.json` is fully regenerated by `pnpm generate:models`
— diff size is large but mechanical.
## Screenshots (if applicable)
<!-- Add screenshots or video recording to help explain your changes -->
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Alexander Brown <drjkl@comfy.org>
## ELI-5
When an image on the page fails to load — a broken thumbnail, an expired
share
link, a flaky in-app browser — the app already handles it and shows a
fallback.
But under the hood, the image helper was *also* shouting "uncaught
error!" to the
browser's global error channel every time. Our monitoring hears that
shout and
logs it as a crash. With enough broken images (some in-app browsers
retry in a
loop), it became the single loudest "error" in our telemetry — for
something
that isn't actually broken. This tells the helper to handle the failure
quietly
instead of shouting.
## What
`useImage()` (from `@vueuse/core`) exposes load failures via its `error`
ref,
which every call site here already uses to render a fallback. But
vueuse's
default `onError` forwards the error to `globalThis.reportError`, so
each failed
`<img>` load also surfaces as an **unhandled** error to global error
monitoring.
This makes failed image loads — 404'd thumbnails, expired share links,
in-app
webviews that re-fetch on a loop — the highest-volume unhandled frontend
error
in our production telemetry, despite being expected and already handled
in the UI.
## Fix
Pass an explicit `onError` (a documented no-op) as the `useAsyncState`
options
argument at all four `useImage()` call sites:
- `components/common/ComfyImage.vue`
- `platform/workflow/sharing/components/ShareAssetThumbnail.vue`
- `platform/assets/components/MediaImageTop.vue`
- `platform/assets/components/AssetCard.vue`
The `error` ref is still set by `useAsyncState` before `onError` runs,
so the
fallback-UI behaviour is unchanged — the only difference is we stop
re-reporting
handled failures to the global error handler.
## Test plan
- [x] No behavioural change to the `error` ref / fallback rendering
(verified
against vueuse `useImage`/`useAsyncState` semantics: `error.value` is
assigned
independently of `onError`).
- [ ] CI lint/format/type checks.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: GitHub Action <action@github.com>
## ELI-5
Some custom nodes have a `getCustomWidgets()` function that's *supposed*
to hand
us a list of widgets. A few of them hand us back nothing
(null/undefined)
instead. We were trying to read that "nothing" like a list, which
crashes with
*"Cannot convert undefined or null to object"* — and because it happens
while
the app is still starting up, it can break the whole page. This PR just
says
"if there's nothing to register, skip it."
## What
`registerCustomWidgets` called `Object.entries(newWidgets)` directly.
When an
extension's `getCustomWidgets()` resolves to `null`/`undefined` (it's
typed
non-null, but extensions are untrusted and routinely violate the type),
this
throws `TypeError: Cannot convert undefined or null to object`.
The call site in `extensionService.ts` runs this inside a bare async
IIFE,
*outside* the `wrapWithErrorHandling` wrappers used for
keybindings/settings, so
the throw is unhandled and surfaces during app initialization.
## Why it matters
In production this is one of the highest-volume unhandled frontend
errors —
~2.6k events across **~1,160 distinct sessions/day**, all funneling
through this
one `Object.entries` call. Guarding the choke point silences it for
every
caller.
## Fix
- Keep `registerCustomWidgets` typed `Record<string,
ComfyWidgetConstructor>`
(the correct internal contract) and early-return on nullish input. The
runtime
guard defends against untrusted extensions that violate the type at the
boundary, without weakening the signature for legitimate callers.
- Add a regression test asserting
`registerCustomWidgets(null!/undefined!)` does
not throw (the `!` casts simulate the boundary violation).
## Test plan
- [x] `npx vitest run src/stores/widgetStore.test.ts` — 8 passing,
including the
new null/undefined case.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
## Summary
Client-side execution events (`execution_start` / `execution_success` /
`execution_error`) are now emitted only by the GTM provider, removing
the redundant Mixpanel and PostHog emissions that duplicated the
server-side PostHog execution pipeline.
## Changes
- **Removed** `trackWorkflowExecution`, `trackExecutionError`, and
`trackExecutionSuccess` from `MixpanelTelemetryProvider` and
`PostHogTelemetryProvider`, along with the now-unused
`lastTriggerSource` field and related type imports.
- **Kept** these methods on `GtmTelemetryProvider`. The
`TelemetryProvider` interface declares them optional and
`TelemetryRegistry` dispatches via optional chaining, so callers are
unchanged and Mixpanel/PostHog simply receive nothing for these events.
- **Added** GTM unit tests for `execution_start` and `execution_success`
(alongside the existing `execution_error` test) to pin the remaining
client-side path.
## Review Focus
- Execution telemetry on the client now flows exclusively to GTM;
PostHog execution data is expected to come solely from the server side,
so there should be no double-counting.
- The server-side PostHog execution pipeline is out of scope for this
frontend change — this PR only stops the client from emitting duplicate
execution events.
Reference: MAR-282
Prior context: Comfy-Org PR #3423.
Co-authored-by: Steven Tran <steventran@Stevens-MacBook-Air.local>
## Summary
Fixes the Desktop2 missing-model download path so the frontend calls the
Desktop2 download bridge directly when it is available, instead of
relying on the browser `<a download>` fallback that Desktop2 currently
has to intercept indirectly.
This addresses Linear FE-956, where missing-model downloads on Windows
could open the OS Save As dialog. The issue was reproducible when the
frontend language was not English: switching the UI language back to
English made the download succeed again.
## Root Cause
Desktop2 currently has compatibility logic that watches/intercepts the
frontend missing-model download flow from outside the FE code. That
interception depends on FE-rendered DOM details, including localized
accessible labels such as the missing-model download button
`aria-label`.
In English, Desktop2 could find the expected download controls and cache
the missing-model metadata before the FE-created `<a>` download was
clicked. In non-English locales, the localized label no longer matched
Desktop2's selector, so the Desktop2 interception path missed the
download. The FE then continued down the browser download path, which
Electron surfaced as a native Save As dialog on Windows.
## Changes
- Adds a narrow Desktop2 runtime bridge check in
`missingModelDownload.ts`:
- if `window.__comfyDesktop2.downloadModel` exists
- and `window.__comfyDesktop2Remote` is not set
- then FE calls `window.__comfyDesktop2.downloadModel(model.url,
model.name, model.directory)` directly and returns early.
- Keeps remote Desktop2 sessions on the existing browser fallback path
by preserving the `__comfyDesktop2Remote` guard.
- Leaves the existing OSS browser fallback and legacy desktop
`isDesktop` download-store path intact.
- Logs Desktop2 bridge failures so rejected promises or synchronous
bridge throws do not become unhandled errors.
- Adds regression coverage for:
- Desktop2 bridge path taking priority over browser and legacy desktop
fallbacks.
- rejected Desktop2 bridge calls being logged without falling back to
browser download.
- synchronously thrown Desktop2 bridge failures being logged without
crashing or falling back to browser download.
- remote Desktop2 sessions continuing to use browser fallback.
## User Impact
Desktop2 users should no longer depend on localized FE DOM text for
missing-model downloads. In particular, non-English UI locales should
route missing-model downloads through Desktop2's managed downloader
instead of opening the OS Save As dialog.
## Validation
- Manually verified the issue is fixed in Desktop2 using a locally built
FE dist served through ComfyUI with `--front-end-root`.
- Verified Korean locale no longer triggers the Save As dialog and the
missing-model download succeeds through Desktop2.
- Verified the new regression test fails when the production bridge fix
is reverted.
- Covered the FE-side contract with unit tests because a true end-to-end
assertion of the Windows native Save As dialog is not currently
practical in the FE browser-test infrastructure. The FE tests can verify
that clicking missing-model download routes into
`window.__comfyDesktop2.downloadModel`; they cannot directly prove
Electron/Windows native dialog behavior. That full native-dialog
regression belongs in Desktop2/Electron integration coverage.
- Ran:
- `pnpm exec oxfmt --check
src/platform/missingModel/missingModelDownload.ts
src/platform/missingModel/missingModelDownload.test.ts`
- `pnpm lint:unstaged`
- `pnpm exec vitest run
src/platform/missingModel/missingModelDownload.test.ts`
- `pnpm typecheck`
- `pnpm build`
- Pre-commit hook passed: `oxfmt`, `oxlint`, `eslint`, `typecheck`.
- Pre-push hook passed: `knip --cache` completed with existing tag hints
only.
- Ran a 3-round local Claude review loop; final verdict was approve with
no Blocker/Major findings.
## Follow-up Work
- Define and document the FE/Desktop2 bridge contract explicitly,
including the expected semantics of `downloadModel` resolving `false`
versus rejecting.
- Add a shared or canonical TypeScript declaration for
`window.__comfyDesktop2` and `window.__comfyDesktop2Remote` if more FE
code starts depending on these globals.
- Remove Desktop2's DOM/aria/class-based missing-model download
interception after a sufficient FE compatibility window, so Desktop2 no
longer depends on FE DOM structure or localized labels.
- Add Desktop2 integration/e2e coverage for missing-model downloads in
non-English locales, ideally including Windows where the Save As dialog
was observed. This is the right layer for a true native Save As
regression test.
- Optionally add a lighter FE browser E2E that injects a fake
`window.__comfyDesktop2.downloadModel` and verifies the missing-model UI
calls that bridge. This would validate the FE contract, but it would
still not replace Desktop2/Electron coverage for native dialog behavior.
- Decide on user-facing failure UX for Desktop2 bridge download failures
once Desktop2 defines whether failures, cancellations, and
already-queued downloads are represented by rejection or by `false`.
## Notes
This intentionally does not fall back to browser download when the
Desktop2 bridge resolves `false`. Falling back there could reintroduce
the exact Save As dialog behavior this PR fixes, and the meaning of
`false` should be clarified in the Desktop2 bridge contract before FE
invents user-facing behavior for it.
A true E2E test for this bug would need to exercise Desktop2/Electron on
Windows and assert that the native Save As dialog is not opened. The
current FE browser-test infrastructure cannot observe that native
Desktop2 behavior directly, so this PR uses focused unit regression
coverage for the FE routing contract plus manual Desktop2 verification.
## Summary
This is the second PR in the planned Workflow Overview error-tab
modernization stack. It follows the merged execution-error slice in
#12683 and focuses only on Missing Media / Missing Inputs.
The intent is to simplify missing-media errors from
media-type/file-resolution cards into compact node-level guidance.
Missing media is now presented as a flat list of affected node inputs,
matching the new catalog-driven item-row direction introduced by the
first PR while intentionally avoiding the larger missing-node,
swap-node, and missing-model redesigns.
This PR also removes the upload/library selection workflow from the
error tab. Missing media rows now help the user locate the affected
node/input directly, rather than attempting to resolve the missing file
inside the error card.
## Changes
- **What**: Simplifies the Missing Inputs card into a flat list of
affected node/widget rows.
- **What**: Displays each missing media row with a catalog-resolved item
label in the form `{nodeName} - {inputName}`.
- **What**: Makes both the row label and right-aligned locate icon
navigate to the affected node on canvas.
- **What**: Adds item-specific accessible labels for locate icon buttons
and an explicit focus-visible state for clickable row labels.
- **What**: Uses the existing `list-scale` transition pattern so missing
media rows animate consistently when they enter or leave.
- **What**: Removes the old image/video/audio sub-sections from the
missing-media card.
- **What**: Removes missing-media
upload/dropzone/library-selection/status/confirm/cancel UI, because
resolution is no longer handled inside this card.
- **What**: Removes stale E2E selectors and tests tied to the deleted
upload/library controls.
- **What**: Removes stale English missing-media locale keys for the
deleted controls; translated locale files are intentionally left to the
localization sync process.
- **What**: Removes stale missing-media store interaction state
(`expandState`, `uploadState`, `pendingSelection`) and the now-unused
name-based removal API that only supported the deleted upload/library
flow.
- **What**: Keeps the remaining widget/node/prefix candidate-removal
APIs used by the missing-media lifecycle.
- **What**: Carries `nodeType` through missing-media referencing nodes
so fallback labels remain correct when the same missing filename is
referenced by multiple node types.
- **Breaking**: None expected. This changes the missing-media error-tab
UX, but does not change the workflow schema or backend APIs.
- **Dependencies**: None.
## Planned stack
This PR is part of the same staged redesign plan described in #12683.
The goal is to keep each error category reviewable and stable before
moving to the next slice.
1. Execution errors, merged in #12683: validation, runtime, and prompt
error grouping/presentation.
2. Missing media, this PR: simplify image/video/audio missing-media
cards into node-level catalog item rows with locate actions.
3. Missing node and swap node: align missing-pack rows, nested node
references, install/replace actions, and locate behavior.
4. Missing model: unify OSS and Cloud presentation, simplify
download/import actions, and improve import/download progress behavior.
This PR intentionally does not change missing node packs, swap nodes,
missing models, runtime errors, validation errors, or prompt errors.
## Review Focus
- Missing media rows should now be node-level rows, not file-level rows
and not image/video/audio sections.
- If the same missing media filename is referenced by multiple nodes,
each affected node/input should render as its own row.
- Clicking either the row label or locate icon should focus the
corresponding affected node.
- Removed upload/library controls are intentional and should not be
restored in this PR.
- The missing-media store cleanup should not remove any state or API
still used by the remaining missing-media lifecycle.
- Only `rightSidePanel.missingMedia.missingMediaTitle` should remain in
the English locale for the old missing-media UI namespace; new row copy
lives in the error catalog.
- This PR should stay scoped to Missing Media / Missing Inputs and
should not include the larger missing model/node/swap redesign work.
## Screenshots (if applicable)
This PR
<img width="483" height="339" alt="스크린샷 2026-06-08 오후 8 26 43"
src="https://github.com/user-attachments/assets/e029411e-b1db-41cd-aa6e-2bd5d86762f6"
/>
Before (Main)
<img width="484" height="802" alt="스크린샷 2026-06-08 오후 8 32 19"
src="https://github.com/user-attachments/assets/c183c18d-babe-4f4e-b4e0-706ce41298de"
/>
## Validation
- `pnpm format`
- `pnpm test:unit src/platform/missingMedia/missingMediaStore.test.ts
src/components/rightSidePanel/errors/TabErrors.test.ts
src/platform/errorCatalog/errorMessageResolver.test.ts
src/platform/missingMedia/missingMediaScan.test.ts
src/platform/missingMedia/missingMediaAssetResolver.test.ts`
- `pnpm typecheck`
- `pnpm knip --cache`
- `pnpm test:browser:local
browser_tests/tests/propertiesPanel/errorsTabMissingMedia.spec.ts
--project=chromium`
- Pre-push hook: `pnpm knip --cache`
## Review notes
Claude review was run after the main implementation and again after
cleanup. No blocker or major issues remained. The only final minor
finding was stale inline i18n fixture data in `TabErrors.test.ts`, which
was removed in the final commit.
## What
The /download hero now surfaces both Windows AND macOS download buttons
when the post-mount UA sniff comes back empty (Linux desktop, mobile,
privacy-stripped UAs). Today those users see only the GitHub-install
button — no actual download link.
## Why
Field complaint: \"I can't find a download link at all on this page.\"
Reproduced on Chrome with a Linux UA — \`useDownloadUrl\` returns
\`platform: null\`, and \`<DownloadLocalButton v-show=\"platform\">\`
collapses to \`display: none\`. The user only sees \"INSTALL FROM
GITHUB\" next to it, with no obvious way to grab the Windows or Mac dmg.
Showing both is the better fallback — they're the two artifacts we
actually ship today (\`downloadUrls.windows\`, \`downloadUrls.macArm\`),
the UA path already encodes them, and the buttons exist as fragment
siblings so they slot into the parent's flex container next to the
existing GitHub button without extra layout wrapping.
## How
- \`useDownloadUrl\` now exposes a \`detected\` ref that flips to
\`true\` post-mount regardless of whether UA matched. Lets the component
distinguish three states:
- **\`!detected\`** → pre-hydration, render nothing (avoids SSR flicker)
- **\`detected && platform\`** → single matched CTA, current behaviour
- **\`detected && !platform\`** → NEW: render both Windows AND Mac
- \`DownloadLocalButton\` adds the fallback branch as a \`<template
v-else-if>\` fragment so the two buttons slot directly into the parent's
existing \`flex-col lg:flex-row\` layout next to the GitHub button. Each
carries its OS icon and an \`aria-label\` combining the existing
\"DOWNLOAD DESKTOP\" copy with the OS name so screen readers get equal
context to the detected-OS path.
- \`downloadUrls\` is now re-exported from the composable so the
fallback path uses the same source of truth instead of duplicating the
URLs.
## Tests
- [x] New \`@smoke\` e2e test (\`HeroSection falls back to both Windows
+ Mac when UA is unrecognized\`) — uses Linux UA, asserts both fallback
buttons appear with the right hrefs / target=\"_blank\", and confirms
the auto-detected single CTA is NOT also present
- [x] Existing Windows-UA test (\`HeroSection has download and GitHub
buttons\`) still passes — single CTA branch unchanged
- [ ] Manual: load /download on Linux Chrome → both buttons visible
- [ ] Manual: load /download on Windows → still single \"DOWNLOAD
DESKTOP\" button (no regression)
- [ ] Manual: load /download on macOS → still single \"DOWNLOAD
DESKTOP\" button with Apple icon
- [ ] Manual: load /download on mobile (which detection treats as null)
→ both fallback buttons visible
## Automated Ingest API Type Update
This PR updates the Ingest API TypeScript types and Zod schemas from the
latest cloud OpenAPI specification.
- Cloud commit: ca12913
- Generated using @hey-api/openapi-ts with Zod plugin
These types cover cloud-only endpoints (workspaces, billing, secrets,
assets, tasks, etc.).
Overlapping endpoints shared with the local ComfyUI Python backend are
excluded.
---------
Co-authored-by: mattmillerai <7741082+mattmillerai@users.noreply.github.com>
Co-authored-by: GitHub Action <action@github.com>
## Summary
https://us.posthog.com/project/204330/feature_flags/706651
Register the `unified_cloud_auth` server feature flag (default **off**)
so the upcoming single Cloud-JWT auth provider (FE-950/FE-951) can be
gated behind it. Pure plumbing — no behavior change.
## Changes
- **What**: Adds `ServerFeatureFlag.UNIFIED_CLOUD_AUTH`, the
`unifiedCloudAuthEnabled` getter in `useFeatureFlags` (via
`resolveFlag`, precedence: dev override > `remoteConfig` > server
feature), and the `unified_cloud_auth?: boolean` field on the
`RemoteConfig` type. Unit tests cover the server-feature wiring and the
dev-override toggle.
- This is PR 1 of 3 for FE-950. Nothing reads the flag yet; it resolves
`false` by default, so this is inert in every environment until a later
PR wires consumers.
## Review Focus
- Mirrors the existing `resolveFlag`-based flags (e.g.
`userSecretsEnabled`, `workflowSharingEnabled`) — deliberately the plain
shape, **not** the cloud-gated `teamWorkspacesEnabled` shape, since the
flag is only read from paths already `isCloud`-gated.
- Local toggle for development:
`localStorage.setItem('ff:unified_cloud_auth', 'true')` (DEV-only dev
override).
- Production delivery of the flag value (and staged % rollout) is
backend work — Comfy-Org/cloud#4139 (BE-1258), already adds
`unified_cloud_auth` to `/api/features`. This FE PR is independent of
it; with no server value the flag stays `false`.
Linear: FE-950
## Screenshots (if applicable)
N/A — no user-facing change (flag registration only).
## Automated Ingest API Type Update
This PR updates the Ingest API TypeScript types and Zod schemas from the
latest cloud OpenAPI specification.
- Cloud commit: 1dc7e7e
- Generated using @hey-api/openapi-ts with Zod plugin
These types cover cloud-only endpoints (workspaces, billing, secrets,
assets, tasks, etc.).
Overlapping endpoints shared with the local ComfyUI Python backend are
excluded.
---------
Co-authored-by: mattmillerai <7741082+mattmillerai@users.noreply.github.com>
Co-authored-by: GitHub Action <action@github.com>
## Summary
- Progressive container-query truncation for the **Keybinding** column
in the Settings → Keybinding panel so the **actions column stops getting
clipped** at narrow widths.
- Tiers (all driven by `@container/keybindings` on the cell content — no
JS resize listeners):
- `>= 16rem`: `Ctrl S , Ctrl Shift S + 1 more` (2 combos + `N-2 more`)
- `>= 12rem`: `Ctrl S + 2 more` (1 combo + `N-1 more`)
- `>= 8rem`: `Ctrl S + 2` (1 combo + compact `+N`)
- smallest: first combo only
- Actions column pinned with `whitespace-nowrap` so it always stays
fully visible — the keybindings column compresses first, not the icons.
- New `KeybindingList.vue` extracted from `KeybindingPanel.vue`; the
prior inline rendering was inlined twice (column body + we'd need it
elsewhere if expansion logic changes).
- Adds `g.nMoreKeybindingsCompact` and `g.keybindingListAriaLabel` i18n
keys (en only; other locales fall back to en).
## Before / after
### before
https://github.com/user-attachments/assets/227cdd6a-fd5d-41ba-8af2-2a49e2d9cf15
### after
<img width="936" height="812" alt="Screenshot 2026-05-13 at 11 05 12 AM"
src="https://github.com/user-attachments/assets/f750efd8-7d14-4a46-a92d-fa67e57c7909"
/>
<!-- Drag the PNGs from ~/Desktop/fe-523-screenshots/ into this section.
Recommended order: wide → 16rem → 12rem → 8rem, before then after at
each tier.
Captured by temp/scripts/capture-fe523-screenshots.ts at 4 dialog
widths. -->
_screenshots pending — drag/drop from `~/Desktop/fe-523-screenshots/`_
The narrow tier shows the original FE-523 bug clearly: on `main` the
Source and Actions columns are clipped without a scrollbar; with this PR
the keybinding column compresses first, keeping at least the edit icon
visible. Some action icons still get clipped at the absolute narrowest
tier — full sticky-actions behavior is a follow-up.
## Why draft
Manual visual verification of the four container-query tiers on a real
backend still needs to happen — the snapshot script runs against a
stubbed-out backend, so the dataset is sparse (the multi-combo rows
aren't injected). Moving out of draft after that pass.
## References
- Fixes
[FE-523](https://linear.app/comfyorg/issue/FE-523/update-keybinding-panel-responsiveness)
- Figma proposal:
https://www.figma.com/design/ZNLRans6JLM3hvCz5LlCQA/Keyboard-Shortcut?node-id=625-17286
- Slack thread:
https://comfy-organization.slack.com/archives/C075ANWQ8KS/p1776818615078629?thread_ts=1776808927.654249&cid=C075ANWQ8KS
## Test plan
- [x] `pnpm test:unit` — new `KeybindingList.test.ts` covers all 4 tiers
(0/1/2/3+ keybindings, count math `N-2` vs `N-1`, container class
present, aria-label assembled).
- [x] `pnpm typecheck`
- [x] `pnpm lint`
- [x] Snapshot diff at 4 widths (see Before / after section above once
images are attached).
- [ ] Manual on a real backend: command with 1 / 2 / 3 / 5 keybindings
each renders the correct `+N more` / `+N` value at each tier.
- [ ] Manual: confirm whether residual actions-column clipping at the
absolute narrowest tier warrants a follow-up.
## Summary
Publishes the Comfy Desktop privacy policy (currently bundled in-app
only) as a public page at \`/privacy/desktop\`, so users can read what
they consented to without re-opening the app and so the document is
verifiable / link-shareable. Also adds a single-sentence link from the
existing \`/privacy-policy\` page so people who land on the org-wide
policy can find the surface-specific one.
## Changes
- **What**: New marketing-site page at \`/privacy/desktop\` (en) and
\`/zh-CN/privacy/desktop\` (zh), backed by ~50 new \`desktop_privacy.*\`
i18n keys mirroring the in-app document. One-sentence addition to
\`privacy.intro.block.3\` linking to the new page.
- **Breaking**: None.
- **Dependencies**: None.
## Why now
Recent Discord audit by a user (cited Civitai as precedent) identified
six legitimate gaps between our website privacy policy and what GDPR /
UK GDPR Art 13 requires (named processors, lawful basis mapping,
retention specifics, international transfer mechanism, etc.). The
**in-app Desktop policy already addresses all of them**, but it isn't
publicly visible. This PR makes it visible at a URL.
A separate, counsel-reviewed PR will be needed to bring the org-wide
\`/privacy-policy\` page up to the same standard — that's deliberately
not in scope here.
## Content the new page covers
Mirrors the Desktop in-app document (effective 2026-05-19):
- Named processors: PostHog, Datadog, ToDesktop, Comfy Org analytics
warehouse
- Article 6 mapping: 6(1)(a) consent for analytics + crash reports;
6(1)(f) legitimate interests for update server
- Retention: 24 months analytics / 15 days crash / 36 months aggregate /
90 days update logs / removed on uninstall
- International transfers: Standard Contractual Clauses where required
- Data subject rights + \`privacy@comfy.org\` as contact
- Right to lodge complaint with ICO / EU supervisory authority / CPPA
## Review Focus
1. **i18n zh-CN placeholders.** Following the existing convention in
this file (see e.g. \`affiliate-landing.footerCta\`), zh-CN values use
the English text as placeholders. A follow-up PR will land Chinese
translations. Flag if there's a different preferred convention.
2. **Source-of-truth coupling.** This page is a publish of the bundled
in-app policy at \`src/renderer/src/lib/legalDocs.ts\`
(\`PRIVACY_POLICY\` constant) in the \`Comfy-Desktop\` repo. There's no
automated sync — the two need to stay aligned manually when the in-app
policy is revised. Worth a quarterly review reminder.
3. **\`privacy@comfy.org\` mailbox.** The policy directs all data
subject rights requests there. Confirm this inbox exists and is
monitored before merge.
4. **No \`noindex\` on the new page.** Unlike the existing
\`/privacy-policy\` page, this one is intentionally indexable — the
point is that it should be discoverable.
## Test plan
- [x] \`pnpm typecheck\` clean (0 errors, 0 warnings)
- [x] \`pnpm build\` succeeds; both \`/privacy/desktop/index.html\` and
\`/zh-CN/privacy/desktop/index.html\` are emitted
- [x] All 385 site pages still build
- [ ] After preview deploy: verify the link from \`/privacy-policy\` to
\`/privacy/desktop\` renders correctly, and the page table-of-contents
nav works (\`ContentSection\` infers sections from i18n keys)
## Companion items (not in this PR)
- Update \`comfy.org/privacy-policy\` (the org-wide policy) with named
processors, Art 6 mapping, retention specifics, and SCC language —
requires privacy counsel review
- Confirm Comfy Org Inc ICO registration status
- One-page determination from privacy counsel on whether Art 37 requires
a formal DPO (pseudonymous product telemetry is likely below threshold,
but the question deserves a documented answer)
2026-06-08 23:22:43 +00:00
662 changed files with 34208 additions and 17528 deletions
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.