feat: use demowright showTitleCard API for early title overlay

- Import showTitleCard/hideTitleCard from demowright/video-script
- Replace page.evaluate() hack with official demowright API
- CI clones demowright feat/show-title-card-api branch
- demowright PR: https://github.com/snomiao/demowright/pull/11

.claude/skills/backport-management/SKILL.md

@@ -11,10 +11,11 @@ Cherry-pick backport management for Comfy-Org/ComfyUI_frontend stable release br
 
 1. **Discover** — Collect candidates from Slack bot + git log gap (`reference/discovery.md`)
 2. **Analyze** — Categorize MUST/SHOULD/SKIP, check deps (`reference/analysis.md`)
-3. **Plan** — Order by dependency (leaf fixes first), group into waves per branch
-4. **Execute** — Label-driven automation → worktree fallback for conflicts (`reference/execution.md`)
-5. **Verify** — After each wave, verify branch integrity before proceeding
-6. **Log & Report** — Generate session report with mermaid diagram (`reference/logging.md`)
+3. **Human Review** — Present candidates in batches for interactive approval (see Interactive Approval Flow)
+4. **Plan** — Order by dependency (leaf fixes first), group into waves per branch
+5. **Execute** — Label-driven automation → worktree fallback for conflicts (`reference/execution.md`)
+6. **Verify** — After each wave, verify branch integrity before proceeding
+7. **Log & Report** — Generate session report (`reference/logging.md`)
 
 ## System Context
 
@@ -37,16 +38,29 @@ Cherry-pick backport management for Comfy-Org/ComfyUI_frontend stable release br
 
 **Critical: Match PRs to the correct target branches.**
 
-| Branch prefix | Scope                          | Example                                   |
-| ------------- | ------------------------------ | ----------------------------------------- |
-| `cloud/*`     | Cloud-hosted ComfyUI only      | App mode, cloud auth, cloud-specific UI   |
-| `core/*`      | Local/self-hosted ComfyUI only | Core editor, local workflows, node system |
+| Branch prefix | Scope                          | Example                                           |
+| ------------- | ------------------------------ | ------------------------------------------------- |
+| `cloud/*`     | Cloud-hosted ComfyUI only      | Team workspaces, cloud queue, cloud-only login    |
+| `core/*`      | Local/self-hosted ComfyUI only | Core editor, local workflows, node system         |
+| Both          | Shared infrastructure          | App mode, Firebase auth (API nodes), payment URLs |
 
-**⚠️ NEVER backport cloud-only PRs to `core/*` branches.** Cloud-only changes (app mode, cloud auth, cloud billing UI, cloud-specific API calls) are irrelevant to local users and waste effort. Before backporting any PR to a `core/*` branch, check:
+### What Goes Where
 
-- Does the PR title/description mention "app mode", "cloud", or cloud-specific features?
-- Does the PR only touch files like `appModeStore.ts`, cloud auth, or cloud-specific components?
-- If yes → skip for `core/*` branches (may still apply to `cloud/*` branches)
+**Both core + cloud:**
+
+- **App mode** PRs — app mode is NOT cloud-only
+- **Firebase auth** PRs — Firebase auth is on core for API nodes
+- **Payment redirect** PRs — payment infrastructure shared
+- **Bug fixes** touching shared components
+
+**Cloud-only (skip for core):**
+
+- Team workspaces
+- Cloud queue virtualization
+- Hide API key login
+- Cloud-specific UI behind cloud feature flags
+
+**⚠️ NEVER backport cloud-only PRs to `core/*` branches.** But do NOT assume "app mode" or "Firebase" = cloud-only. Check the actual files changed.
 
 ## ⚠️ Gotchas (Learn from Past Sessions)
 
@@ -67,6 +81,32 @@ The `pr-backport.yaml` action reports more conflicts than reality. `git cherry-p
 
 12 or 27 conflicting files can be trivial (snapshots, new files). **Categorize conflicts first**, then decide. See Conflict Triage below.
 
+### Accept-Theirs Can Produce Broken Hybrids
+
+When a PR **rewrites a component** (e.g., PrimeVue → Reka UI), the accept-theirs regex produces a broken mix of old and new code. The template may reference new APIs while the script still has old imports, or vice versa.
+
+**Detection:** Content conflicts with 4+ conflict markers in a single `.vue` file, especially when imports change between component libraries.
+
+**Fix:** Instead of accept-theirs regex, use `git show MERGE_SHA:path/to/file > path/to/file` to get the complete correct version from the merge commit on main. This bypasses the conflict entirely.
+
+### Cherry-Picks Can Reference Missing Dependencies
+
+When PR A on main depends on code introduced by PR B (which was merged before A), cherry-picking A brings in code that references B's additions. The cherry-pick succeeds but the branch is broken.
+
+**Common pattern:** Composables, component files, or type definitions introduced by an earlier PR and used by the cherry-picked PR.
+
+**Detection:** `pnpm typecheck` fails with "Cannot find module" or "is not defined" errors after cherry-pick.
+
+**Fix:** Use `git show MERGE_SHA:path/to/missing/file > path/to/missing/file` to bring the missing files from main. Always verify with typecheck.
+
+### Use `--no-verify` for Worktree Pushes
+
+Husky hooks fail in worktrees (can't find lint-staged config). Always use `git push --no-verify` and `git commit --no-verify` when working in `/tmp/` worktrees.
+
+### Automation Success Varies Wildly by Branch
+
+In the 2026-04-06 session: core/1.42 got 18/26 auto-PRs, cloud/1.42 got only 1/25. The cloud branch has more divergence. **Always plan for manual fallback** — don't assume automation will handle most PRs.
+
 ## Conflict Triage
 
 **Always categorize before deciding to skip. High conflict count ≠ hard conflicts.**
@@ -77,6 +117,8 @@ The `pr-backport.yaml` action reports more conflicts than reality. `git cherry-p
 | **Modify/delete (new file)** | PR introduces files not on target    | `git add $FILE` — keep the new file                             |
 | **Modify/delete (removed)**  | Target removed files the PR modifies | `git rm $FILE` — file no longer relevant                        |
 | **Content conflicts**        | Marker-based (`<<<<<<<`)             | Accept theirs via python regex (see below)                      |
+| **Component rewrites**       | 4+ markers in `.vue`, library change | Use `git show SHA:path > path` — do NOT accept-theirs           |
+| **Import-only conflicts**    | Only import lines differ             | Keep both imports if both used; remove unused after             |
 | **Add/add**                  | Both sides added same file           | Accept theirs, verify no logic conflict                         |
 | **Locale/JSON files**        | i18n key additions                   | Accept theirs, validate JSON after                              |
 
@@ -103,7 +145,7 @@ Skip these without discussion:
 - **Test-only / lint rule changes** — Not user-facing
 - **Revert pairs** — If PR A reverted by PR B, skip both. If fixed version (PR C) exists, backport only C.
 - **Features not on target branch** — e.g., Painter, GLSLShader, appModeStore on core/1.40
-- **Cloud-only PRs on core/\* branches** — App mode, cloud auth, cloud billing. These only affect cloud-hosted ComfyUI.
+- **Cloud-only PRs on core/\* branches** — Team workspaces, cloud queue, cloud-only login. (Note: app mode and Firebase auth are NOT cloud-only — see Branch Scope Rules)
 
 ## Wave Verification
 
@@ -122,6 +164,18 @@ git worktree remove /tmp/verify-TARGET --force
 
 If typecheck or tests fail, stop and investigate before continuing. A broken branch after wave N means all subsequent waves will compound the problem.
 
+### Fix PRs Are Normal
+
+Expect to create 1 fix PR per branch after verification. Common issues:
+
+1. **Component rewrite hybrids** — accept-theirs produced broken `.vue` files. Fix: overwrite with correct version from merge commit via `git show SHA:path > path`
+2. **Missing dependency files** — cherry-pick brought in code referencing composables/components not on the branch. Fix: add missing files from merge commit
+3. **Missing type properties** — cherry-picked code uses interface properties not yet on the branch (e.g., `key` on `ConfirmDialogOptions`). Fix: add the property to the interface
+4. **Unused imports** — conflict resolution kept imports that the branch doesn't use. Fix: remove unused imports
+5. **Wrong types from conflict resolution** — e.g., `{ top: number; right: number }` vs `{ top: number; left: number }`. Fix: match the return type of the actual function
+
+Create a fix PR on a branch from the target, verify typecheck passes, then merge with `--squash --admin`.
+
 ### Never Admin-Merge Without CI
 
 In a previous bulk session, all 69 backport PRs were merged with `gh pr merge --squash --admin`, bypassing required CI checks. This shipped 3 test failures to a release branch. **Lesson: `--admin` skips all branch protection, including required status checks.** Only use `--admin` after confirming CI has passed (e.g., `gh pr checks $PR` shows all green), or rely on auto-merge (`--auto --squash`) which waits for CI by design.
@@ -135,6 +189,43 @@ Large backport sessions (50+ PRs) are expensive and error-prone. Prefer continuo
 - Reserve session-style bulk backporting for catching up after gaps
 - When a release branch is created, immediately start the continuous process
 
+## Interactive Approval Flow
+
+After analysis, present ALL candidates (MUST, SHOULD, and borderline) to the human for interactive review before execution. Do not write a static decisions.md — collect approvals in conversation.
+
+### Batch Presentation
+
+Present PRs in batches of 5-10, grouped by theme (visual bugs, interaction bugs, cloud/auth, data correctness, etc.). Use this table format:
+
+```
+ #  | PR     | Title                                    | Target        | Rec  | Context
+----+--------+------------------------------------------+---------------+------+--------
+ 1  | #12345 | fix: broken thing                        | core+cloud/42 | Y    | Description here. Why it matters. Agent reasoning.
+ 2  | #12346 | fix: another issue                       | core/42       | N    | Only affects removed feature. Not on target branch.
+```
+
+Each row includes:
+
+- PR number and title
+- Target branches
+- Agent recommendation: `Rec: Y` or `Rec: N` with brief reasoning
+- 2-3 sentence context: what the PR does, why it matters (or doesn't)
+
+### Human Response Format
+
+- `Y` — approve for backport
+- `N` — skip
+- `?` — investigate (agent shows PR description, files changed, detailed take, then re-asks)
+- Any freeform question or comment triggers discussion before moving on
+- Bulk responses accepted (e.g. `1 Y, 2 Y, 3 N, 4 ?`)
+
+### Rules
+
+- ALL candidates are reviewed, not just MUST items
+- When human responds `?`, show the PR description, files changed, and agent's detailed analysis, then re-ask for their decision
+- When human asks a question about a PR, answer with context and recommendation, then wait for their decision
+- Do not proceed to execution until all batches are reviewed and every candidate has a Y or N
+
 ## Quick Reference
 
 ### Label-Driven Automation (default path)
@@ -150,13 +241,96 @@ gh api repos/Comfy-Org/ComfyUI_frontend/issues/$PR/labels \
 ```bash
 git worktree add /tmp/backport-$BRANCH origin/$BRANCH
 cd /tmp/backport-$BRANCH
+
+# For each PR:
+git fetch origin $BRANCH
 git checkout -b backport-$PR-to-$BRANCH origin/$BRANCH
 git cherry-pick -m 1 $MERGE_SHA
-# Resolve conflicts, push, create PR, merge
+# Resolve conflicts (see Conflict Triage)
+git push origin backport-$PR-to-$BRANCH --no-verify
+gh pr create --base $BRANCH --head backport-$PR-to-$BRANCH \
+  --title "[backport $BRANCH] $TITLE (#$PR)" \
+  --body "Backport of #$PR. [conflict notes]"
+gh pr merge $NEW_PR --squash --admin
+sleep 25
 ```
 
+### Efficient Batch: Test-Then-Resolve Pattern
+
+When many PRs need manual cherry-pick (e.g., cloud branches), test all first:
+
+```bash
+cd /tmp/backport-$BRANCH
+for pr in "${ORDER[@]}"; do
+  git checkout -b test-$pr origin/$BRANCH
+  if git cherry-pick -m 1 $SHA 2>/dev/null; then
+    echo "CLEAN: $pr"
+  else
+    echo "CONFLICT: $pr"
+    git cherry-pick --abort
+  fi
+  git checkout --detach HEAD
+  git branch -D test-$pr
+done
+```
+
+Then process clean PRs in a batch loop, conflicts individually.
+
 ### PR Title Convention
 
 ```
 [backport TARGET_BRANCH] Original Title (#ORIGINAL_PR)
 ```
+
+## Final Deliverables (Slack-Compatible)
+
+After execution completes, generate two files in `~/temp/backport-session/`. Both must be **Slack-compatible plain text** — no emojis, no markdown tables, no headers (`#`), no bold (`**`), no inline code. Use plain dashes, indentation, and line breaks only.
+
+### 1. Author Accountability Report
+
+File: `backport-author-accountability.md`
+
+Lists all backported PRs grouped by original author (via `gh pr view $PR --json author`). Surfaces who should be self-labeling.
+
+```
+Backport Session YYYY-MM-DD -- PRs that should have been labeled by authors
+
+- author-login
+    - #1234 fix: short title
+    - #5678 fix: another title
+- other-author
+    - #9012 fix: some other fix
+```
+
+Authors sorted alphabetically, 4-space indent for nested items.
+
+### 2. Slack Status Update
+
+File: `slack-status-update.md`
+
+A shareable summary of the session. Structure:
+
+```
+Backport session complete -- YYYY-MM-DD
+
+[1-sentence summary: N PRs backported to which branches. All pass typecheck.]
+
+Branches updated:
+- core/X.XX: N PRs + N fix PRs (N auto, N manual)
+- cloud/X.XX: N PRs + N fix PRs (N auto, N manual)
+- ...
+
+N total PRs created and merged (N backports + N fix PRs).
+
+Notable fixes included:
+- [category]: [list of fixes]
+- ...
+
+Conflict patterns encountered:
+- [pattern and how it was resolved]
+- ...
+
+N authors had PRs backported. See author accountability list for details.
+```
+
+No emojis, no tables, no bold, no headers. Plain text that pastes cleanly into Slack.

.claude/skills/backport-management/reference/analysis.md

@@ -23,10 +23,10 @@ For SHOULD items with conflicts: if conflict resolution requires more than trivi
 
 **Before categorizing, filter by branch scope:**
 
-| Target branch | Skip if PR is...                                                    |
-| ------------- | ------------------------------------------------------------------- |
-| `core/*`      | Cloud-only (app mode, cloud auth, cloud billing, cloud-specific UI) |
-| `cloud/*`     | Local-only features not present on cloud branch                     |
+| Target branch | Skip if PR is...                                                                                                  |
+| ------------- | ----------------------------------------------------------------------------------------------------------------- |
+| `core/*`      | Cloud-only (team workspaces, cloud queue, cloud-only login). Note: app mode and Firebase auth are NOT cloud-only. |
+| `cloud/*`     | Local-only features not present on cloud branch                                                                   |
 
 Cloud-only PRs backported to `core/*` are wasted effort — `core/*` branches serve local/self-hosted users who never see cloud features. Check PR titles, descriptions, and files changed for cloud-specific indicators.
 
@@ -61,8 +61,6 @@ done
 
 ## Human Review Checkpoint
 
-Present decisions.md before execution. Include:
+Use the Interactive Approval Flow (see SKILL.md) to review all candidates interactively. Do not write a static decisions.md for the human to edit — instead, present batches of 5-10 PRs with context and recommendations, and collect Y/N/? responses in conversation.
 
-1. All MUST/SHOULD/SKIP categorizations with rationale
-2. Questions for human (feature existence, scope, deps)
-3. Estimated effort per branch
+All candidates must be reviewed (MUST, SHOULD, and borderline items), not just a subset.

.claude/skills/backport-management/reference/execution.md

@@ -73,14 +73,22 @@ for PR in ${CONFLICT_PRS[@]}; do
   git cherry-pick -m 1 $MERGE_SHA
 
   # If conflict — NEVER skip based on file count alone!
-  # Categorize conflicts first: binary PNGs, modify/delete, content, add/add
+  # Categorize conflicts first: binary PNGs, modify/delete, content, add/add, component rewrites
   # See SKILL.md Conflict Triage table for resolution per type.
 
+  # For component rewrites (4+ markers in a .vue file, library migration):
+  # DO NOT use accept-theirs regex — it produces broken hybrids.
+  # Instead, use the complete file from the merge commit:
+  # git show $MERGE_SHA:path/to/file > path/to/file
+
+  # For simple content conflicts, accept theirs:
+  # python3 -c "import re; ..."
+
   # Resolve all conflicts, then:
   git add .
   GIT_EDITOR=true git cherry-pick --continue
 
-  git push origin backport-$PR-to-TARGET
+  git push origin backport-$PR-to-TARGET --no-verify
   NEW_PR=$(gh pr create --base TARGET_BRANCH --head backport-$PR-to-TARGET \
     --title "[backport TARGET] TITLE (#$PR)" \
     --body "Backport of #$PR..." | grep -oP '\d+$')
@@ -114,7 +122,30 @@ source ~/.nvm/nvm.sh && nvm use 24 && pnpm install && pnpm typecheck && pnpm tes
 git worktree remove /tmp/verify-TARGET --force
 ```
 
-If verification fails, stop and fix before proceeding to the next wave. Do not compound problems across waves.
+If verification fails, **do not skip** — create a fix PR:
+
+```bash
+# Stay in the verify worktree
+git checkout -b fix-backport-TARGET origin/TARGET_BRANCH
+
+# Common fixes:
+# 1. Component rewrite hybrids: overwrite with merge commit version
+git show MERGE_SHA:path/to/Component.vue > path/to/Component.vue
+
+# 2. Missing dependency files
+git show MERGE_SHA:path/to/missing.ts > path/to/missing.ts
+
+# 3. Missing type properties: edit the interface
+# 4. Unused imports: delete the import lines
+
+git add -A
+git commit --no-verify -m "fix: resolve backport typecheck issues on TARGET"
+git push origin fix-backport-TARGET --no-verify
+gh pr create --base TARGET --head fix-backport-TARGET --title "fix: resolve backport typecheck issues on TARGET" --body "..."
+gh pr merge $PR --squash --admin
+```
+
+Do not proceed to the next branch until typecheck passes.
 
 ## Conflict Resolution Patterns
 
@@ -142,7 +173,35 @@ git rm $FILE
 git checkout --theirs $FILE && git add $FILE
 ```
 
-### 4. Locale Files
+### 4. Component Rewrites (DO NOT accept-theirs)
+
+When a PR completely rewrites a component (e.g., PrimeVue → Reka UI), accept-theirs produces
+a broken hybrid with mismatched template/script sections.
+
+```bash
+# Use the complete correct file from the merge commit instead:
+git show $MERGE_SHA:src/components/input/MultiSelect.vue > src/components/input/MultiSelect.vue
+git show $MERGE_SHA:src/components/input/SingleSelect.vue > src/components/input/SingleSelect.vue
+git add src/components/input/MultiSelect.vue src/components/input/SingleSelect.vue
+```
+
+**Detection:** 4+ conflict markers in a single `.vue` file, imports changing between component
+libraries (PrimeVue → Reka UI, etc.), template structure completely different on each side.
+
+### 5. Missing Dependencies After Cherry-Pick
+
+Cherry-picks can succeed but leave the branch broken because the PR's code on main
+references composables/components introduced by an earlier PR.
+
+```bash
+# Add the missing file from the merge commit:
+git show $MERGE_SHA:src/composables/queue/useJobDetailsHover.ts > src/composables/queue/useJobDetailsHover.ts
+git show $MERGE_SHA:src/components/builder/BuilderSaveDialogContent.vue > src/components/builder/BuilderSaveDialogContent.vue
+```
+
+**Detection:** `pnpm typecheck` fails with "Cannot find module" or "X is not defined" after cherry-pick succeeds cleanly.
+
+### 6. Locale Files
 
 Usually adding new i18n keys — accept theirs, validate JSON:
 
@@ -176,8 +235,14 @@ gh pr checks $PR --watch --fail-fast && gh pr merge $PR --squash --admin
 8. **Always validate JSON** after resolving locale file conflicts
 9. **Dep refresh PRs** — skip on stable branches. Risk of transitive dep regressions outweighs audit cleanup. Cherry-pick individual CVE fixes instead.
 10. **Verify after each wave** — run `pnpm typecheck && pnpm test:unit` on the target branch after merging a batch. Catching breakage early prevents compounding errors.
-11. **Cloud-only PRs don't belong on core/\* branches** — app mode, cloud auth, and cloud-specific UI changes are irrelevant to local users. Always check PR scope against branch scope before backporting.
+11. **App mode and Firebase auth are NOT cloud-only** — they go to both core and cloud branches. Only team workspaces, cloud queue, and cloud-specific login are cloud-only.
 12. **Never admin-merge without CI** — `--admin` bypasses all branch protections including required status checks. A bulk session of 69 admin-merges shipped 3 test failures. Always wait for CI to pass first, or use `--auto --squash` which waits by design.
+13. **Accept-theirs regex breaks component rewrites** — when a PR migrates between component libraries (PrimeVue → Reka UI), the regex produces a broken hybrid. Use `git show SHA:path > path` to get the complete correct version instead.
+14. **Cherry-picks can silently bring in missing-dependency code** — if PR A references a composable introduced by PR B, cherry-picking A succeeds but typecheck fails. Always run typecheck after each wave and add missing files from the merge commit.
+15. **Fix PRs are expected** — plan for 1 fix PR per branch to resolve typecheck issues from conflict resolutions. This is normal, not a failure.
+16. **Use `--no-verify` in worktrees** — husky hooks fail in `/tmp/` worktrees. Always push/commit with `--no-verify`.
+17. **Automation success varies by branch** — core/1.42 got 18/26 auto-PRs (69%), cloud/1.42 got 1/25 (4%). Cloud branches diverge more. Plan for manual fallback.
+18. **Test-then-resolve pattern** — for branches with low automation success, run a dry-run loop to classify clean vs conflict PRs before processing. This is much faster than resolving conflicts serially.
 
 ## CI Failure Triage
 

.claude/skills/backport-management/reference/logging.md

@@ -2,26 +2,25 @@
 
 ## During Execution
 
-Maintain `execution-log.md` with per-branch tables:
+Maintain `execution-log.md` with per-branch tables (this is internal, markdown tables are fine here):
 
 ```markdown
-| PR#   | Title | CI Status                      | Status                            | Backport PR | Notes   |
-| ----- | ----- | ------------------------------ | --------------------------------- | ----------- | ------- |
-| #XXXX | Title | ✅ Pass / ❌ Fail / ⏳ Pending | ✅ Merged / ⏭️ Skip / ⏸️ Deferred | #YYYY       | Details |
+| PR#   | Title | Status | Backport PR | Notes   |
+| ----- | ----- | ------ | ----------- | ------- |
+| #XXXX | Title | merged | #YYYY       | Details |
 ```
 
 ## Wave Verification Log
 
-Track verification results per wave:
+Track verification results per wave within execution-log.md:
 
 ```markdown
-## Wave N Verification — TARGET_BRANCH
+Wave N Verification -- TARGET_BRANCH
 
 - PRs merged: #A, #B, #C
-- Typecheck: ✅ Pass / ❌ Fail
-- Unit tests: ✅ Pass / ❌ Fail
+- Typecheck: pass / fail
+- Fix PR: #YYYY (if needed)
 - Issues found: (if any)
-- Human review needed: (list any non-trivial conflict resolutions)
 ```
 
 ## Session Report Template
@@ -63,40 +62,42 @@ Track verification results per wave:
 - Feature branches that need tracking for future sessions?
 ```
 
-## Final Deliverable: Visual Summary
+## Final Deliverables
 
-At session end, generate a **mermaid diagram** showing all backported PRs organized by target branch and category (MUST/SHOULD), plus a summary table. Present this to the user as the final output.
+After all branches are complete and verified, generate these files in `~/temp/backport-session/`:
 
-```mermaid
-graph TD
-    subgraph branch1["☁️ cloud/X.XX — N PRs"]
-        C1["#XXXX title"]
-        C2["#XXXX title"]
-    end
+### 1. execution-log.md (internal)
 
-    subgraph branch2must["🔴 core/X.XX MUST — N PRs"]
-        M1["#XXXX title"]
-    end
+Per-branch tables with PR#, title, status, backport PR#, notes. Markdown tables are fine — this is for internal tracking, not Slack.
 
-    subgraph branch2should["🟡 core/X.XX SHOULD — N PRs"]
-        S1["#XXXX-#XXXX N auto-merged"]
-        S2["#XXXX-#XXXX N manual picks"]
-    end
+### 2. backport-author-accountability.md (Slack-compatible)
 
-    classDef cloudStyle fill:#1a3a5c,stroke:#4da6ff,color:#e0f0ff
-    classDef coreStyle fill:#1a4a2e,stroke:#4dff88,color:#e0ffe8
-    classDef mustStyle fill:#5c1a1a,stroke:#ff4d4d,color:#ffe0e0
-    classDef shouldStyle fill:#4a3a1a,stroke:#ffcc4d,color:#fff5e0
-```
+See SKILL.md "Final Deliverables" section. Plain text, no emojis/tables/headers/bold. Authors sorted alphabetically with PRs nested under each.
 
-Use the `mermaid` tool to render this diagram and present it alongside the summary table as the session's final deliverable.
+### 3. slack-status-update.md (Slack-compatible)
+
+See SKILL.md "Final Deliverables" section. Plain text summary that pastes cleanly into Slack. Includes branch counts, notable fixes, conflict patterns, author count.
+
+## Slack Formatting Rules
+
+Both shareable files (author accountability + status update) must follow these rules:
+
+- No emojis (no checkmarks, no arrows, no icons)
+- No markdown tables (use plain lists with dashes)
+- No headers (no # or ##)
+- No bold (\*_) or italic (_)
+- No inline code backticks
+- Use -- instead of em dash
+- Use plain dashes (-) for lists with 4-space indent for nesting
+- Line breaks between sections for readability
+
+These files should paste directly into a Slack message and look clean.
 
 ## Files to Track
 
-- `candidate_list.md` — all candidates per branch
-- `decisions.md` — MUST/SHOULD/SKIP with rationale
-- `wave-plan.md` — execution order
-- `execution-log.md` — real-time status
-- `backport-session-report.md` — final summary
+All in `~/temp/backport-session/`:
 
-All in `~/temp/backport-session/`.
+- `execution-plan.md` -- approved PRs with merge SHAs (input)
+- `execution-log.md` -- real-time status with per-branch tables (internal)
+- `backport-author-accountability.md` -- PRs grouped by author (Slack-compatible)
+- `slack-status-update.md` -- session summary (Slack-compatible)

.claude/skills/comfy-qa/REPRODUCE.md

@@ -0,0 +1,278 @@
+---
+name: reproduce-issue
+description: 'Reproduce a GitHub issue by researching prerequisites, setting up the environment (custom nodes, workflows, settings), and interactively exploring ComfyUI via playwright-cli until the bug is confirmed. Then records a clean demo video.'
+---
+
+# Issue Reproduction Skill
+
+Reproduce a reported GitHub issue against a running ComfyUI instance. This skill uses an interactive, agent-driven approach — not a static script. You will research, explore, retry, and adapt until the bug is reproduced, then record a clean demo.
+
+## Architecture
+
+Two videos are produced:
+
+1. **Research video** — the full exploration session: installing deps, trying things, failing, retrying, figuring out the bug. Valuable for debugging context.
+2. **Reproduce video** — a clean, minimal recording of just the reproduction steps. This is the demo you'd attach to the issue.
+
+```
+Phase 1: Research       → Read issue, understand prerequisites
+Phase 2: Environment    → Install custom nodes, load workflows, configure settings
+Phase 3: Explore        → [VIDEO 1: research] Interactively try to reproduce (retries OK)
+Phase 4: Record         → [VIDEO 2: reproduce] Clean recording of just the minimal repro steps
+Phase 5: Report         → Generate a structured reproduction report
+```
+
+## Prerequisites
+
+- ComfyUI server running (ask user for URL, default: `http://127.0.0.1:8188`)
+- `playwright-cli` installed: `npm install -g @playwright/cli@latest`
+- `gh` CLI (authenticated, for reading issues)
+- ComfyUI backend with Python environment (for installing custom nodes)
+
+## Phase 1: Research the Issue
+
+1. Fetch the issue details:
+
+   ```bash
+   gh issue view <number> --repo Comfy-Org/ComfyUI_frontend --json title,body,comments
+   ```
+
+2. Extract from the issue body:
+   - **Reproduction steps** (the exact sequence)
+   - **Prerequisites**: specific workflows, custom nodes, settings, models
+   - **Environment**: OS, browser, ComfyUI version
+   - **Media**: screenshots or videos showing the bug
+
+3. Search the codebase for related code:
+   - Find the feature/component mentioned in the issue
+   - Understand how it works currently
+   - Identify what state the UI needs to be in
+
+## Phase 2: Environment Setup
+
+Set up everything the issue requires BEFORE attempting reproduction.
+
+### Custom Nodes
+
+If the issue mentions custom nodes:
+
+```bash
+# Find the custom node repo
+# Clone into ComfyUI's custom_nodes directory
+cd <comfyui_path>/custom_nodes
+git clone <custom_node_repo_url>
+
+# Install dependencies if needed
+cd <custom_node_name>
+pip install -r requirements.txt 2>/dev/null || true
+
+# Restart ComfyUI server to load the new nodes
+```
+
+### Workflows
+
+If the issue references a specific workflow:
+
+```bash
+# Download workflow JSON if a URL is provided
+curl -L "<workflow_url>" -o /tmp/test-workflow.json
+
+# Load it via the API
+curl -X POST http://127.0.0.1:8188/api/workflow \
+  -H "Content-Type: application/json" \
+  -d @/tmp/test-workflow.json
+```
+
+Or load via playwright-cli:
+
+```bash
+playwright-cli goto "http://127.0.0.1:8188"
+# Drag-and-drop or use File > Open to load the workflow
+```
+
+### Settings
+
+If the issue requires specific settings:
+
+```bash
+# Use playwright-cli to open settings and change them
+playwright-cli press "Control+,"
+playwright-cli snapshot
+# Find and modify the relevant setting
+```
+
+## Phase 3: Interactive Exploration — Research Video
+
+Start recording the **research video** (Video 1). This captures the full exploration — mistakes, retries, dead ends — all valuable context.
+
+```bash
+# Open browser and start video recording
+playwright-cli open "http://127.0.0.1:8188"
+playwright-cli video-start
+
+# Take a snapshot to see current state
+playwright-cli snapshot
+
+# Interact based on what you see
+playwright-cli click <ref>
+playwright-cli fill <ref> "text"
+playwright-cli press "Control+s"
+
+# Check results
+playwright-cli snapshot
+playwright-cli screenshot --filename=/tmp/qa/research-step-1.png
+```
+
+### Key Principles
+
+- **Observe before acting**: Always `snapshot` before interacting
+- **Retry and adapt**: If a step fails, try a different approach
+- **Document what works**: Keep notes on which steps trigger the bug
+- **Don't give up**: Try multiple approaches if the first doesn't work
+- **Establish prerequisites**: Many bugs require specific UI state:
+  - Save a workflow first (File > Save)
+  - Make changes to dirty the workflow
+  - Open multiple tabs
+  - Add specific node types
+  - Change settings
+  - Resize the window
+
+### Common ComfyUI Interactions via playwright-cli
+
+| Action              | Command                                                        |
+| ------------------- | -------------------------------------------------------------- |
+| Open hamburger menu | `playwright-cli click` on the C logo button                    |
+| Navigate menu       | `playwright-cli hover <ref>` then `playwright-cli click <ref>` |
+| Add node            | Double-click canvas → type node name → select from results     |
+| Connect nodes       | Drag from output slot to input slot                            |
+| Save workflow       | `playwright-cli press "Control+s"`                             |
+| Save As             | Menu > File > Save As                                          |
+| Select node         | Click on the node                                              |
+| Delete node         | Select → `playwright-cli press "Delete"`                       |
+| Right-click menu    | `playwright-cli click <ref> --button right`                    |
+| Keyboard shortcut   | `playwright-cli press "Control+z"`                             |
+
+## Phase 4: Record Clean Demo — Reproduce Video (max 5 minutes)
+
+Once the bug is confirmed, **stop the research video** and **close the research browser**:
+
+```bash
+playwright-cli video-stop
+playwright-cli close
+```
+
+Now start a **fresh browser session** for the clean reproduce video (Video 2).
+
+**IMPORTANT constraints:**
+
+- **Max 5 minutes** — the reproduce video must be short and focused
+- **No environment setup** — server, user, custom nodes are already set up from Phase 3. Just log in and go.
+- **No exploration** — you already know the exact steps. Execute them quickly and precisely.
+- **Start video recording immediately**, execute steps, stop. Don't leave the recording running while thinking.
+
+1. **Open browser and start recording**:
+
+   ```bash
+   playwright-cli open "http://127.0.0.1:8188"
+   playwright-cli video-start
+   ```
+
+2. **Execute only the minimal reproduction steps** — no exploration, no mistakes. Just the clean sequence that demonstrates the bug. You already know exactly what works from Phase 3.
+
+3. **Take key screenshots** at critical moments:
+
+   ```bash
+   playwright-cli screenshot --filename=/tmp/qa/before-bug.png
+   # ... trigger the bug ...
+   playwright-cli screenshot --filename=/tmp/qa/bug-visible.png
+   ```
+
+4. **Stop recording and close** immediately after the bug is demonstrated:
+   ```bash
+   playwright-cli video-stop
+   playwright-cli close
+   ```
+
+## Phase 5: Generate Report
+
+Create a reproduction report at `tmp/qa/reproduce-report.md`:
+
+```markdown
+# Issue Reproduction Report
+
+- **Issue**: <issue_url>
+- **Title**: <issue_title>
+- **Date**: <today>
+- **Status**: Reproduced / Not Reproduced / Partially Reproduced
+
+## Environment
+
+- ComfyUI Server: <url>
+- OS: <os>
+- Custom Nodes Installed: <list or "none">
+- Settings Changed: <list or "none">
+
+## Prerequisites
+
+List everything that had to be set up before the bug could be triggered:
+
+1. ...
+2. ...
+
+## Reproduction Steps
+
+Minimal steps to reproduce (the clean sequence):
+
+1. ...
+2. ...
+3. ...
+
+## Expected Behavior
+
+<from the issue>
+
+## Actual Behavior
+
+<what actually happened>
+
+## Evidence
+
+- Research video: `research-video/video.webm` (full exploration session)
+- Reproduce video: `reproduce-video/video.webm` (clean minimal repro)
+- Screenshots: `before-bug.png`, `bug-visible.png`
+
+## Root Cause Analysis (if identified)
+
+<code pointers, hypothesis about what's going wrong>
+
+## Notes
+
+<any additional observations, workarounds discovered, related issues>
+```
+
+## Handling Failures
+
+If the bug **cannot be reproduced**:
+
+1. Document what you tried and why it didn't work
+2. Check if the issue was already fixed (search git log for related commits)
+3. Check if it's environment-specific (OS, browser, specific version)
+4. Set report status to "Not Reproduced" with detailed notes
+5. The report is still valuable — it saves others from repeating the same investigation
+
+## CI Integration
+
+In CI, this skill runs as a Claude Code agent with:
+
+- `ANTHROPIC_API_KEY` for Claude
+- `GEMINI_API_KEY` for initial issue analysis (optional)
+- ComfyUI server pre-started in the container
+- `playwright-cli` pre-installed
+
+The CI workflow:
+
+1. Gemini generates a reproduce guide (markdown) from the issue
+2. Claude agent receives the guide and runs this skill
+3. Claude explores interactively, installs dependencies, retries
+4. Claude records a clean demo once reproduced
+5. Video and report are uploaded as artifacts

.claude/skills/comfy-qa/SKILL.md

@@ -0,0 +1,283 @@
+---
+name: comfy-qa
+description: 'Comprehensive QA of ComfyUI frontend. Navigates all routes, tests all interactive features using playwright-cli, generates a report, and submits a draft PR. Works in CI and local environments, cross-platform.'
+---
+
+# ComfyUI Frontend QA Skill
+
+Automated quality assurance for the ComfyUI frontend. The pipeline reproduces reported bugs using Playwright E2E tests, records video evidence, and deploys reports to Cloudflare Pages.
+
+## Architecture Overview
+
+The QA pipeline uses a **three-phase approach**:
+
+1. **RESEARCH** — Claude writes Playwright E2E tests to reproduce bugs (assertion-backed, no hallucination)
+2. **REPRODUCE** — Deterministic replay of the research test with video recording
+3. **REPORT** — Deploy results to Cloudflare Pages with badge, video, and verdict
+
+### Key Design Decision
+
+Earlier iterations used AI vision (Gemini) to drive a browser and judge results from video. This was abandoned after discovering **AI reviewers hallucinate** — Gemini reported "REPRODUCED" when videos showed idle screens. The current approach uses **Playwright assertions** as the source of truth: if the test passes, the bug is proven.
+
+## Prerequisites
+
+- Node.js 22+
+- `pnpm` package manager
+- `gh` CLI (authenticated)
+- Playwright browsers: `npx playwright install chromium`
+- Environment variables:
+  - `GEMINI_API_KEY` — for PR analysis and video review
+  - `ANTHROPIC_API_KEY` — for Claude Agent SDK (research phase)
+  - `CLOUDFLARE_API_TOKEN` + `CLOUDFLARE_ACCOUNT_ID` — for report deployment
+
+## Pipeline Scripts
+
+| Script                            | Role                                                    | Model                         |
+| --------------------------------- | ------------------------------------------------------- | ----------------------------- |
+| `scripts/qa-analyze-pr.ts`        | Deep PR/issue analysis → QA guide                       | gemini-3.1-pro-preview        |
+| `scripts/qa-agent.ts`             | Research phase: Claude writes E2E tests                 | claude-sonnet-4-6 (Agent SDK) |
+| `scripts/qa-record.ts`            | Before/after video recording with Gemini-driven actions | gemini-3.1-pro-preview        |
+| `scripts/qa-reproduce.ts`         | Deterministic replay with narration                     | gemini-3-flash-preview        |
+| `scripts/qa-video-review.ts`      | Video comparison review                                 | gemini-3-flash-preview        |
+| `scripts/qa-generate-test.ts`     | Regression test generation from QA report               | gemini-3-flash-preview        |
+| `scripts/qa-deploy-pages.sh`      | Deploy to Cloudflare Pages + badge                      | —                             |
+| `scripts/qa-batch.sh`             | Batch-trigger QA for multiple issues                    | —                             |
+| `scripts/qa-report-template.html` | Report site (light/dark, seekbar, copy badge)           | —                             |
+
+## Triggering QA
+
+### Via GitHub Labels
+
+- **`qa-changes`** — Focused QA on a PR (Linux-only, before/after comparison)
+- **`qa-full`** — Full QA (3-OS matrix, after-only)
+- **`qa-issue`** — Reproduce a bug from an issue
+
+### Via Batch Script
+
+```bash
+# Trigger QA for specific issue numbers
+./scripts/qa-batch.sh 10394 10238 9996
+
+# From a triage file (top 5 Tier 1 issues)
+./scripts/qa-batch.sh --from tmp/issues.md --top 5
+
+# Preview without pushing
+./scripts/qa-batch.sh --dry-run 10394
+
+# Clean up old trigger branches
+./scripts/qa-batch.sh --cleanup
+```
+
+### Via Workflow Dispatch
+
+Go to Actions → "PR: QA" → Run workflow → choose mode (focused/full).
+
+## CI Workflow (`.github/workflows/pr-qa.yaml`)
+
+```
+resolve-matrix → analyze-pr ──┐
+                               ├→ qa-before (main branch, worktree build)
+                               ├→ qa-after  (PR branch)
+                               └→ report (video review, deploy, comment)
+```
+
+Before/after jobs run **in parallel** on separate runners for clean isolation.
+
+### Issue Reproduce Mode
+
+For issues (not PRs), the pipeline:
+
+1. Fetches the issue body and comments
+2. Runs `qa-analyze-pr.ts --type issue` to generate a QA guide
+3. Runs the research phase (Claude writes E2E test to reproduce)
+4. Records video of the test execution
+5. Posts results as a comment on the issue
+
+## Running Locally
+
+### Step 1: Environment Setup
+
+```bash
+# Ensure ComfyUI server is running
+# Default: http://127.0.0.1:8188
+
+# Install Playwright browsers
+npx playwright install chromium
+```
+
+### Step 2: Analyze the Issue/PR
+
+```bash
+# For a PR
+pnpm exec tsx scripts/qa-analyze-pr.ts \
+  --pr-number 10394 \
+  --repo Comfy-Org/ComfyUI_frontend \
+  --output-dir qa-guides
+
+# For an issue
+pnpm exec tsx scripts/qa-analyze-pr.ts \
+  --pr-number 10394 \
+  --repo Comfy-Org/ComfyUI_frontend \
+  --output-dir qa-guides \
+  --type issue
+```
+
+### Step 3: Record Before/After
+
+```bash
+# Before (main branch)
+pnpm exec tsx scripts/qa-record.ts \
+  --mode before \
+  --diff /tmp/pr-diff.txt \
+  --output-dir /tmp/qa-before \
+  --qa-guide qa-guides/qa-guide-1.json
+
+# After (PR branch)
+pnpm exec tsx scripts/qa-record.ts \
+  --mode after \
+  --diff /tmp/pr-diff.txt \
+  --output-dir /tmp/qa-after \
+  --qa-guide qa-guides/qa-guide-1.json
+```
+
+### Step 4: Review Videos
+
+```bash
+pnpm exec tsx scripts/qa-video-review.ts \
+  --artifacts-dir /tmp/qa-artifacts \
+  --video-file qa-session.mp4 \
+  --before-video qa-before-session.mp4 \
+  --output-dir /tmp/video-reviews \
+  --pr-context /tmp/pr-context.txt
+```
+
+## Research Phase Details (`qa-agent.ts`)
+
+Claude receives:
+
+- The issue description and comments
+- A QA guide from `qa-analyze-pr.ts`
+- An accessibility tree snapshot of the current UI
+
+Claude's tools:
+
+- **`inspect(selector?)`** — Read a11y tree to discover element selectors
+- **`writeTest(code)`** — Write a Playwright `.spec.ts` file
+- **`runTest()`** — Execute the test and get pass/fail + errors
+- **`done(verdict, summary, evidence, testCode)`** — Finish with verdict
+
+The test uses the project's Playwright fixtures (`comfyPageFixture`), giving access to `comfyPage.page`, `comfyPage.menu`, `comfyPage.settings`, etc.
+
+### Verdict Logic
+
+- **REPRODUCED** — Test passes (asserting the bug exists) → bug is proven
+- **NOT_REPRODUCIBLE** — Claude exhausted attempts, test cannot pass
+- **INCONCLUSIVE** — Agent timed out or encountered infrastructure issues
+
+Auto-completion: if a test passed but `done()` was never called, the pipeline auto-completes with REPRODUCED.
+
+## Manual QA (Fallback)
+
+When the automated pipeline isn't suitable (e.g., visual-only bugs, complex multi-step interactions), use **playwright-cli** for manual browser interaction:
+
+```bash
+# Install
+npm install -g @playwright/cli@latest
+
+# Open browser and navigate
+playwright-cli open http://127.0.0.1:8188
+
+# Get element references
+playwright-cli snapshot
+
+# Interact
+playwright-cli click e1
+playwright-cli fill e2 "test text"
+playwright-cli press Escape
+playwright-cli screenshot --filename=f.png
+```
+
+Snapshots return element references (`e1`, `e2`, …). Always run `snapshot` after navigation to refresh refs.
+
+## Manual QA Test Plan
+
+When performing manual QA (either via playwright-cli or the automated pipeline), systematically test each area below.
+
+### Application Load & Routes
+
+| Test              | Steps                                                        |
+| ----------------- | ------------------------------------------------------------ |
+| Root route loads  | Navigate to `/` — GraphView should render with canvas        |
+| User select route | Navigate to `/user-select` — user selection UI should appear |
+| 404 handling      | Navigate to `/nonexistent` — should handle gracefully        |
+
+### Canvas & Graph View
+
+| Test                      | Steps                                                          |
+| ------------------------- | -------------------------------------------------------------- |
+| Canvas renders            | The LiteGraph canvas is visible and interactive                |
+| Pan canvas                | Click and drag on empty canvas area                            |
+| Zoom in/out               | Use scroll wheel or Alt+=/Alt+-                                |
+| Add node via double-click | Double-click canvas to open search, type "KSampler", select it |
+| Delete node               | Select a node, press Delete key                                |
+| Connect nodes             | Drag from output slot to input slot                            |
+| Copy/Paste                | Select nodes, Ctrl+C then Ctrl+V                               |
+| Undo/Redo                 | Make changes, Ctrl+Z to undo, Ctrl+Y to redo                   |
+| Context menus             | Right-click node vs empty canvas — different menus             |
+
+### Sidebar Tabs
+
+| Test              | Steps                                 |
+| ----------------- | ------------------------------------- |
+| Workflows tab     | Press W — workflows sidebar opens     |
+| Node Library tab  | Press N — node library opens          |
+| Model Library tab | Press M — model library opens         |
+| Tab toggle        | Press same key again — sidebar closes |
+| Search in sidebar | Type in search box — results filter   |
+
+### Settings Dialog
+
+| Test             | Steps                                                |
+| ---------------- | ---------------------------------------------------- |
+| Open settings    | Press Ctrl+, or click settings button                |
+| Change a setting | Toggle a boolean setting — it persists after closing |
+| Search settings  | Type in settings search box — results filter         |
+| Close settings   | Press Escape or click close button                   |
+
+### Execution & Queue
+
+| Test           | Steps                                                 |
+| -------------- | ----------------------------------------------------- |
+| Queue prompt   | Load default workflow, click Queue — execution starts |
+| Queue progress | Progress indicator shows during execution             |
+| Interrupt      | Press Ctrl+Alt+Enter during execution — interrupts    |
+
+## Report Site
+
+Deployed to Cloudflare Pages at `https://comfy-qa.pages.dev/<branch>/`.
+
+Features:
+
+- Light/dark theme
+- Seekable video player with preload
+- Copy badge button (markdown)
+- Date-stamped badges (e.g., `QA0327`)
+- Vertical box badge for issues and PRs
+
+## Known Issues & Troubleshooting
+
+See `docs/qa/TROUBLESHOOTING.md` for common failures:
+
+- `set -euo pipefail` + grep with no match → append `|| true`
+- `__name is not defined` in `page.evaluate` → use `addScriptTag`
+- Cursor not visible in videos → monkey-patch `page.mouse` methods
+- Agent not calling `done()` → auto-complete from passing test
+
+## Backlog
+
+See `docs/qa/backlog.md` for planned improvements:
+
+- **Type B comparison**: Different commits for regression detection
+- **Type C comparison**: Cross-browser testing
+- **Pre-seed assets**: Upload test images before recording
+- **Lazy a11y tree**: Reduce token usage with `inspect(selector)` vs full dump

.claude/skills/comfy-qa/scripts/qa-agent.ts

@@ -0,0 +1,673 @@
+#!/usr/bin/env tsx
+/**
+ * QA Research Phase — Claude writes & debugs E2E tests to reproduce bugs
+ *
+ * Instead of driving a browser interactively, Claude:
+ * 1. Reads the issue + a11y snapshot of the UI
+ * 2. Writes a Playwright E2E test (.spec.ts) that reproduces the bug
+ * 3. Runs the test → reads errors → rewrites → repeats until it works
+ * 4. Outputs the passing test + verdict
+ *
+ * Tools:
+ *   - inspect(selector) — read a11y tree to understand UI state
+ *   - writeTest(code) — write a Playwright test file
+ *   - runTest() — execute the test and get results
+ *   - done(verdict, summary, testCode) — finish with the working test
+ */
+
+import type { Page } from '@playwright/test'
+/* eslint-disable import-x/no-unresolved */
+// @ts-expect-error — claude-agent-sdk has no type declarations for vue-tsc
+import { query, tool, createSdkMcpServer } from '@anthropic-ai/claude-agent-sdk'
+/* eslint-enable import-x/no-unresolved */
+import { z } from 'zod'
+import { mkdirSync, readFileSync, writeFileSync } from 'fs'
+import { execSync } from 'child_process'
+
+// ── Types ──
+
+interface ResearchOptions {
+  page: Page
+  issueContext: string
+  qaGuide: string
+  outputDir: string
+  serverUrl: string
+  anthropicApiKey?: string
+  maxTurns?: number
+  timeBudgetMs?: number
+}
+
+export type ReproMethod = 'e2e_test' | 'video' | 'both' | 'none'
+
+export interface ResearchResult {
+  verdict: 'REPRODUCED' | 'NOT_REPRODUCIBLE' | 'INCONCLUSIVE'
+  reproducedBy: ReproMethod
+  summary: string
+  evidence: string
+  testCode: string
+  videoScript?: string
+  log: Array<{
+    turn: number
+    timestampMs: number
+    toolName: string
+    toolInput: unknown
+    toolResult: string
+  }>
+}
+
+// ── Main research function ──
+
+export async function runResearchPhase(
+  opts: ResearchOptions
+): Promise<ResearchResult> {
+  const { page, issueContext, qaGuide, outputDir, serverUrl, anthropicApiKey } =
+    opts
+  const maxTurns = opts.maxTurns ?? 50
+
+  let agentDone = false
+  let finalVerdict: ResearchResult['verdict'] = 'INCONCLUSIVE'
+  let finalReproducedBy: ReproMethod = 'none'
+  let finalSummary = 'Agent did not complete'
+  let finalEvidence = ''
+  let finalTestCode = ''
+  let finalVideoScript = ''
+  let turnCount = 0
+  let lastPassedTurn = -1
+  const startTime = Date.now()
+  const researchLog: ResearchResult['log'] = []
+
+  const testDir = `${outputDir}/research`
+  mkdirSync(testDir, { recursive: true })
+  const testPath = `${testDir}/reproduce.spec.ts`
+
+  // Get initial a11y snapshot for context
+  let initialA11y = ''
+  try {
+    initialA11y = await page.locator('body').ariaSnapshot({ timeout: 5000 })
+    initialA11y = initialA11y.slice(0, 3000)
+  } catch {
+    initialA11y = '(could not capture initial a11y snapshot)'
+  }
+
+  // ── Tool: inspect ──
+  const inspectTool = tool(
+    'inspect',
+    'Read the current accessibility tree to understand UI state. Use this to discover element names, roles, and selectors for your test.',
+    {
+      selector: z
+        .string()
+        .optional()
+        .describe(
+          'Optional filter — only show elements matching this name/role. Omit for full tree.'
+        )
+    },
+    async (args: { selector?: string }) => {
+      let resultText: string
+      try {
+        const ariaText = await page
+          .locator('body')
+          .ariaSnapshot({ timeout: 5000 })
+        if (args.selector) {
+          const lines = ariaText.split('\n')
+          const matches = lines.filter((l: string) =>
+            l.toLowerCase().includes(args.selector!.toLowerCase())
+          )
+          resultText =
+            matches.length > 0
+              ? `Found "${args.selector}":\n${matches.slice(0, 15).join('\n')}`
+              : `"${args.selector}" not found. Full tree:\n${ariaText.slice(0, 2000)}`
+        } else {
+          resultText = ariaText.slice(0, 3000)
+        }
+      } catch (e) {
+        resultText = `inspect failed: ${e instanceof Error ? e.message : e}`
+      }
+
+      researchLog.push({
+        turn: turnCount,
+        timestampMs: Date.now() - startTime,
+        toolName: 'inspect',
+        toolInput: args,
+        toolResult: resultText.slice(0, 500)
+      })
+
+      return { content: [{ type: 'text' as const, text: resultText }] }
+    }
+  )
+
+  // ── Tool: readFixture ──
+  const readFixtureTool = tool(
+    'readFixture',
+    'Read a fixture or helper file from browser_tests/fixtures/ to understand the API. Use this to discover available methods on comfyPage helpers before writing your test.',
+    {
+      path: z
+        .string()
+        .describe(
+          'Relative path within browser_tests/fixtures/, e.g. "helpers/CanvasHelper.ts" or "components/Topbar.ts" or "ComfyPage.ts"'
+        )
+    },
+    async (args: { path: string }) => {
+      let resultText: string
+      try {
+        const fullPath = `${projectRoot}/browser_tests/fixtures/${args.path}`
+        const content = readFileSync(fullPath, 'utf-8')
+        resultText = content.slice(0, 4000)
+        if (content.length > 4000) {
+          resultText += `\n\n... (truncated, ${content.length} total chars)`
+        }
+      } catch (e) {
+        resultText = `Could not read fixture: ${e instanceof Error ? e.message : e}`
+      }
+
+      researchLog.push({
+        turn: turnCount,
+        timestampMs: Date.now() - startTime,
+        toolName: 'readFixture',
+        toolInput: args,
+        toolResult: resultText.slice(0, 500)
+      })
+
+      return { content: [{ type: 'text' as const, text: resultText }] }
+    }
+  )
+
+  // ── Tool: readTest ──
+  const readTestTool = tool(
+    'readTest',
+    'Read an existing E2E test file from browser_tests/tests/ to learn patterns and conventions used in this project.',
+    {
+      path: z
+        .string()
+        .describe(
+          'Relative path within browser_tests/tests/, e.g. "workflow.spec.ts" or "subgraph.spec.ts"'
+        )
+    },
+    async (args: { path: string }) => {
+      let resultText: string
+      try {
+        const fullPath = `${projectRoot}/browser_tests/tests/${args.path}`
+        const content = readFileSync(fullPath, 'utf-8')
+        resultText = content.slice(0, 4000)
+        if (content.length > 4000) {
+          resultText += `\n\n... (truncated, ${content.length} total chars)`
+        }
+      } catch (e) {
+        // List available test files if the path doesn't exist
+        try {
+          const { readdirSync } = await import('fs')
+          const files = readdirSync(`${projectRoot}/browser_tests/tests/`)
+            .filter((f: string) => f.endsWith('.spec.ts'))
+            .slice(0, 30)
+          resultText = `File not found: ${args.path}\n\nAvailable test files:\n${files.join('\n')}`
+        } catch {
+          resultText = `Could not read test: ${e instanceof Error ? e.message : e}`
+        }
+      }
+
+      researchLog.push({
+        turn: turnCount,
+        timestampMs: Date.now() - startTime,
+        toolName: 'readTest',
+        toolInput: args,
+        toolResult: resultText.slice(0, 500)
+      })
+
+      return { content: [{ type: 'text' as const, text: resultText }] }
+    }
+  )
+
+  // ── Tool: writeTest ──
+  const writeTestTool = tool(
+    'writeTest',
+    'Write a Playwright E2E test file that reproduces the bug. The test should assert the broken behavior exists.',
+    {
+      code: z
+        .string()
+        .describe('Complete Playwright test file content (.spec.ts)')
+    },
+    async (args: { code: string }) => {
+      writeFileSync(testPath, args.code)
+
+      researchLog.push({
+        turn: turnCount,
+        timestampMs: Date.now() - startTime,
+        toolName: 'writeTest',
+        toolInput: { path: testPath, codeLength: args.code.length },
+        toolResult: `Test written to ${testPath} (${args.code.length} chars)`
+      })
+
+      return {
+        content: [
+          {
+            type: 'text' as const,
+            text: `Test written to ${testPath}. Use runTest() to execute it.`
+          }
+        ]
+      }
+    }
+  )
+
+  // ── Tool: runTest ──
+  // Place test in browser_tests/ so Playwright config finds fixtures
+  const projectRoot = process.cwd()
+  const browserTestPath = `${projectRoot}/browser_tests/tests/qa-reproduce.spec.ts`
+
+  const runTestTool = tool(
+    'runTest',
+    'Run the Playwright test and get results. Returns stdout/stderr including assertion errors.',
+    {},
+    async () => {
+      turnCount++
+      // Copy the test to browser_tests/tests/ where Playwright expects it
+      const { copyFileSync } = await import('fs')
+      try {
+        copyFileSync(testPath, browserTestPath)
+      } catch {
+        // directory may not exist
+        mkdirSync(`${projectRoot}/browser_tests/tests`, { recursive: true })
+        copyFileSync(testPath, browserTestPath)
+      }
+
+      let resultText: string
+      try {
+        const output = execSync(
+          `cd "${projectRoot}" && npx playwright test browser_tests/tests/qa-reproduce.spec.ts --reporter=list --timeout=30000 --retries=0 --workers=1 2>&1`,
+          {
+            timeout: 90000,
+            encoding: 'utf-8',
+            env: {
+              ...process.env,
+              COMFYUI_BASE_URL: serverUrl
+            }
+          }
+        )
+        resultText = `TEST PASSED:\n${output.slice(-1500)}`
+      } catch (e) {
+        const err = e as { stdout?: string; stderr?: string; message?: string }
+        const output = (err.stdout || '') + '\n' + (err.stderr || '')
+        resultText = `TEST FAILED:\n${output.slice(-2000)}`
+      }
+
+      researchLog.push({
+        turn: turnCount,
+        timestampMs: Date.now() - startTime,
+        toolName: 'runTest',
+        toolInput: { testPath },
+        toolResult: resultText.slice(0, 1000)
+      })
+
+      // Auto-save passing test code for fallback completion — but only if
+      // the test contains a bug-specific assertion (not just a discovery/debug test)
+      if (resultText.startsWith('TEST PASSED')) {
+        try {
+          const code = readFileSync(browserTestPath, 'utf-8')
+          const hasBugAssertion =
+            /expect\s*\(/.test(code) &&
+            !/^\s*expect\([^)]+\)\.toBeDefined\(\)/m.test(code) &&
+            !/^\s*expect\([^)]+\)\.toBeGreaterThan\(0\)/m.test(code) &&
+            !/Inspect|Find|Debug|discover/i.test(
+              code.match(/test\(['"`]([^'"`]+)/)?.[1] ?? ''
+            )
+          if (hasBugAssertion) {
+            finalTestCode = code
+            lastPassedTurn = turnCount
+          }
+        } catch {
+          // ignore
+        }
+        resultText +=
+          '\n\n⚠️ Test PASSED — call done() now with verdict REPRODUCED and the test code. Do NOT write more tests.'
+      }
+
+      return { content: [{ type: 'text' as const, text: resultText }] }
+    }
+  )
+
+  // ── Tool: done ──
+  const doneTool = tool(
+    'done',
+    'Finish research with verdict and the final test code.',
+    {
+      verdict: z.enum(['REPRODUCED', 'NOT_REPRODUCIBLE', 'INCONCLUSIVE']),
+      reproducedBy: z
+        .enum(['e2e_test', 'video', 'both', 'none'])
+        .describe(
+          'How the bug was proven: e2e_test = Playwright assertion passed, video = visual evidence only, both = both methods, none = not reproduced'
+        ),
+      summary: z.string().describe('What you found and why'),
+      evidence: z.string().describe('Test output that proves the verdict'),
+      testCode: z
+        .string()
+        .describe(
+          'Final Playwright test code. If REPRODUCED, this test asserts the bug exists and passes.'
+        ),
+      videoScript: z
+        .string()
+        .optional()
+        .describe(
+          'Demowright video script for Phase 2 demo recording. REQUIRED when verdict is REPRODUCED. A separate test file using createVideoScript with title, segments, and outro. Do NOT include demowright imports in testCode.'
+        )
+    },
+    async (args: {
+      verdict: ResearchResult['verdict']
+      reproducedBy: ReproMethod
+      summary: string
+      evidence: string
+      testCode: string
+      videoScript?: string
+    }) => {
+      agentDone = true
+      finalVerdict = args.verdict
+      finalReproducedBy = args.reproducedBy
+      finalSummary = args.summary
+      finalEvidence = args.evidence
+      finalTestCode = args.testCode
+      finalVideoScript = args.videoScript ?? ''
+      writeFileSync(testPath, args.testCode)
+      if (args.videoScript) {
+        writeFileSync(`${outputDir}/video-script.spec.ts`, args.videoScript)
+      }
+      return {
+        content: [
+          { type: 'text' as const, text: `Research complete: ${args.verdict}` }
+        ]
+      }
+    }
+  )
+
+  // ── MCP Server ──
+  const server = createSdkMcpServer({
+    name: 'qa-research',
+    version: '1.0.0',
+    tools: [
+      inspectTool,
+      readFixtureTool,
+      readTestTool,
+      writeTestTool,
+      runTestTool,
+      doneTool
+    ]
+  })
+
+  // ── System prompt ──
+  const systemPrompt = `You are a senior QA engineer who writes Playwright E2E tests to reproduce reported bugs.
+
+## Your tools
+- inspect(selector?) — Read the accessibility tree to understand the current UI. Use to discover selectors, element names, and UI state.
+- readFixture(path) — Read fixture source code from browser_tests/fixtures/. Use to discover available methods. E.g. "helpers/CanvasHelper.ts", "components/Topbar.ts", "ComfyPage.ts"
+- readTest(path) — Read an existing test from browser_tests/tests/ to learn patterns. E.g. "workflow.spec.ts". Pass any name to list available files.
+- writeTest(code) — Write a Playwright test file (.spec.ts)
+- runTest() — Execute the test and get results (pass/fail + errors)
+- done(verdict, summary, evidence, testCode) — Finish with the final test
+
+## Workflow
+1. Read the issue description carefully
+2. Use inspect() to understand the current UI state and discover element selectors
+3. If unsure about the fixture API, use readFixture() to read the relevant helper source code
+4. If unsure about test patterns, use readTest() to read an existing test for reference
+5. Write a Playwright test that:
+   - Performs the exact reproduction steps from the issue
+   - Asserts the BROKEN behavior (the bug) — so the test PASSES when the bug exists
+6. Run the test with runTest()
+7. If it fails: read the error, fix the test, run again (max 5 attempts)
+8. Call done() with the final verdict and test code
+
+## Test writing guidelines
+- Import the project fixture: \`import { comfyPageFixture as test } from '../fixtures/ComfyPage'\`
+- Import expect: \`import { expect } from '@playwright/test'\`
+- The fixture provides \`comfyPage\` which has all the helpers listed below
+- If the bug IS present, the test should PASS. If the bug is fixed, the test would FAIL.
+- Keep tests focused and minimal — test ONLY the reported bug
+- Write ONE test, not multiple. Focus on the single clearest reproduction.
+- The test file will be placed in browser_tests/tests/qa-reproduce.spec.ts
+- Use \`comfyPage.nextFrame()\` after interactions that trigger UI updates
+- NEVER use \`page.waitForTimeout()\` — use Locator actions and retrying assertions instead
+- ALWAYS call done() when finished, even if the test passed — do not keep iterating after a passing test
+- Use \`expect.poll()\` for async assertions: \`await expect.poll(() => comfyPage.nodeOps.getGraphNodesCount()).toBe(8)\`
+- CRITICAL: Your assertions must be SPECIFIC TO THE BUG. A test that asserts \`expect(count).toBeGreaterThan(0)\` proves nothing — it would pass even without the bug. Instead assert the exact broken state, e.g. \`expect(clonedWidgets).toHaveLength(0)\` (missing widgets) or \`expect(zIndex).toBeLessThan(parentZIndex)\` (wrong z-order). If a test passes trivially, it's a false positive.
+- NEVER write "debug", "discovery", or "inspect node types" tests. These waste turns and produce false REPRODUCED verdicts. If you need to discover node type names, use inspect() or readFixture() — not a passing test.
+- If you cannot write a bug-specific assertion, call done() with verdict NOT_REPRODUCIBLE and explain why.
+
+## ComfyPage Fixture API Reference
+
+### Core properties
+- \`comfyPage.page\` — raw Playwright Page
+- \`comfyPage.canvas\` — Locator for #graph-canvas
+- \`comfyPage.queueButton\` — "Queue Prompt" button
+- \`comfyPage.runButton\` — "Run" button (new UI)
+- \`comfyPage.confirmDialog\` — ConfirmDialog (has .confirm, .delete, .overwrite, .reject locators + .click(name) method)
+- \`comfyPage.nextFrame()\` — wait for next requestAnimationFrame
+- \`comfyPage.setup()\` — navigate + wait for app ready (called automatically by fixture)
+
+### Menu (comfyPage.menu)
+- \`comfyPage.menu.topbar\` — Topbar helper:
+  - \`.triggerTopbarCommand(['File', 'Save As'])\` — navigate menu hierarchy
+  - \`.openTopbarMenu()\` / \`.closeTopbarMenu()\` — open/close hamburger
+  - \`.openSubmenu('File')\` — hover to open submenu, returns submenu Locator
+  - \`.getTabNames()\` — get all open workflow tab names
+  - \`.getActiveTabName()\` — get active tab name
+  - \`.getWorkflowTab(name)\` — get tab Locator
+  - \`.closeWorkflowTab(name)\` — close a tab
+  - \`.saveWorkflow(name)\` / \`.saveWorkflowAs(name)\` / \`.exportWorkflow(name)\`
+  - \`.switchTheme('dark' | 'light')\`
+- \`comfyPage.menu.workflowsTab\` — WorkflowsSidebarTab:
+  - \`.open()\` / \`.close()\` — toggle workflows sidebar
+  - \`.getTopLevelSavedWorkflowNames()\` — list saved workflow names
+- \`comfyPage.menu.nodeLibraryTab\` — NodeLibrarySidebarTab
+- \`comfyPage.menu.assetsTab\` — AssetsSidebarTab
+
+### Canvas (comfyPage.canvasOps)
+- \`.click({x, y})\` — click at position on canvas
+- \`.rightClick(x, y)\` — right-click (opens context menu)
+- \`.doubleClick()\` — double-click canvas (opens node search)
+- \`.clickEmptySpace()\` — click known empty area
+- \`.dragAndDrop(source, target)\` — drag from source to target position
+- \`.pan(offset, safeSpot?)\` — pan canvas by offset
+- \`.zoom(deltaY, steps?)\` — zoom via scroll wheel
+- \`.resetView()\` — reset zoom/pan to default
+- \`.getScale()\` / \`.setScale(n)\` — get/set canvas zoom
+- \`.getNodeCenterByTitle(title)\` — get screen coords of node center
+- \`.disconnectEdge()\` / \`.connectEdge()\` — default graph edge operations
+
+### Node Operations (comfyPage.nodeOps)
+- \`.getGraphNodesCount()\` — count all nodes
+- \`.getSelectedGraphNodesCount()\` — count selected nodes
+- \`.getNodes()\` — get all nodes
+- \`.getFirstNodeRef()\` — get NodeReference for first node
+- \`.getNodeRefById(id)\` — get NodeReference by ID
+- \`.getNodeRefsByType(type)\` — get all nodes of a type
+- \`.waitForGraphNodes(count)\` — wait until node count matches
+
+### Settings (comfyPage.settings)
+- \`.setSetting(id, value)\` — change a ComfyUI setting
+- \`.getSetting(id)\` — read current setting value
+
+### Keyboard (comfyPage.keyboard)
+- \`.undo()\` / \`.redo()\` — Ctrl+Z / Ctrl+Y
+- \`.bypass()\` — Ctrl+B
+- \`.selectAll()\` — Ctrl+A
+- \`.ctrlSend(key)\` — send Ctrl+key
+
+### Workflow (comfyPage.workflow)
+- \`.loadWorkflow(name)\` — load from browser_tests/assets/{name}.json
+- \`.setupWorkflowsDirectory(structure)\` — setup test directory
+- \`.deleteWorkflow(name)\`
+- \`.isCurrentWorkflowModified()\` — check dirty state
+
+### Context Menu (comfyPage.contextMenu)
+- \`.openFor(locator)\` — right-click locator and wait for menu
+- \`.clickMenuItem(name)\` — click a menu item by name
+- \`.isVisible()\` — check if context menu is showing
+- \`.assertHasItems(items)\` — assert menu contains items
+
+### Other helpers
+- \`comfyPage.settingDialog\` — SettingDialog component
+- \`comfyPage.searchBox\` / \`comfyPage.searchBoxV2\` — node search
+- \`comfyPage.toast\` — ToastHelper (\`.visibleToasts\`)
+- \`comfyPage.subgraph\` — SubgraphHelper
+- \`comfyPage.vueNodes\` — VueNodeHelpers
+- \`comfyPage.bottomPanel\` — BottomPanel
+- \`comfyPage.clipboard\` — ClipboardHelper
+- \`comfyPage.dragDrop\` — DragDropHelper
+
+### Available fixture files (use readFixture to explore)
+- ComfyPage.ts — main fixture with all helpers
+- helpers/CanvasHelper.ts, NodeOperationsHelper.ts, WorkflowHelper.ts
+- helpers/KeyboardHelper.ts, SettingsHelper.ts, SubgraphHelper.ts
+- components/Topbar.ts, ContextMenu.ts, SettingDialog.ts, SidebarTab.ts
+
+## Video Script (IMPORTANT — provide via done() tool)
+
+When calling done(), provide a \`videoScript\` parameter with a SEPARATE test file that uses demowright's createVideoScript.
+Do NOT put demowright imports in testCode — they won't resolve in Phase 1.
+
+The videoScript is a complete, standalone Playwright test file for Phase 2 demo recording:
+
+\`\`\`typescript
+import { comfyPageFixture as test } from '../fixtures/ComfyPage'
+import { createVideoScript, showTitleCard, hideTitleCard } from 'demowright/video-script'
+
+test('Demo: Bug Title', async ({ comfyPage }) => {
+  // Show title card IMMEDIATELY — covers the screen while setup runs behind it
+  await showTitleCard(comfyPage.page, 'Bug Title Here', { subtitle: 'Issue #NNNN' })
+
+  // Setup runs while title card is visible
+  await comfyPage.settings.setSetting('Comfy.UseNewMenu', 'Top')
+  await comfyPage.workflow.setupWorkflowsDirectory({})
+
+  // Remove early title card before script starts (script will show its own)
+  await hideTitleCard(comfyPage.page)
+
+  const script = createVideoScript()
+    .title('Bug Title Here', { subtitle: 'Issue #NNNN', durationMs: 4000 })
+    .segment('Step 1: description of what we do', async (pace) => {
+      await pace()  // narration finishes FIRST
+      await comfyPage.menu.topbar.saveWorkflow('name')  // THEN action
+      await comfyPage.page.waitForTimeout(2000)  // pause for viewer
+    })
+    .segment('Bug evidence: what we see proves the bug', async (pace) => {
+      await pace()
+      await comfyPage.page.waitForTimeout(5000)  // hold on evidence
+    })
+    .outro({ text: 'Bug Reproduced', subtitle: 'Summary' })
+
+  await script.render(comfyPage.page)
+})
+\`\`\`
+
+Key API:
+- \`.title(text, {subtitle?, durationMs?})\` — title card (4s default)
+- \`.segment(narrationText, async (pace) => { await pace(); ...actions... })\` — TTS narrated step
+- \`.outro({text?, subtitle?, durationMs?})\` — ending card
+- \`pace()\` — wait for narration audio to finish
+
+CRITICAL TIMING: Call \`await pace()\` FIRST in each segment, BEFORE the Playwright actions.
+This makes the narration play and finish, THEN the actions execute — so viewers hear what's about
+to happen before it happens. Pattern:
+
+\`\`\`typescript
+.segment('Now we save the workflow as a new name', async (pace) => {
+  await pace()  // narration finishes first
+  await comfyPage.menu.topbar.saveWorkflowAs('new-name')  // then action happens
+  await comfyPage.page.waitForTimeout(2000)  // pause so viewer sees the result
+})
+\`\`\`
+
+IMPORTANT RULES for videoScript:
+1. You MUST provide videoScript when verdict is REPRODUCED — every reproduced bug needs a narrated demo
+2. Call showTitleCard() BEFORE setup, run setup behind it, call hideTitleCard() before createVideoScript() — see example
+3. Call \`await pace()\` FIRST in each segment callback, BEFORE actions
+4. Add \`waitForTimeout(2000)\` after each action so viewers can see the result
+5. Final evidence segment: hold for 5+ seconds
+6. Reproduce the same steps as testCode but slower with clear narration
+
+## Current UI state (accessibility tree)
+${initialA11y}
+
+${qaGuide ? `## QA Analysis Guide\n${qaGuide}\n` : ''}
+## Issue to Reproduce
+${issueContext}`
+
+  // ── Run the agent ──
+  console.warn('Starting research phase (Claude writes E2E tests)...')
+
+  try {
+    for await (const message of query({
+      prompt:
+        'Write a Playwright E2E test that reproduces the reported bug. Use inspect() to discover selectors, readFixture() or readTest() if you need to understand the fixture API or see existing test patterns, writeTest() to write the test, runTest() to execute it. Iterate until it works or you determine the bug cannot be reproduced.',
+      options: {
+        model: 'claude-sonnet-4-6',
+        systemPrompt,
+        ...(anthropicApiKey ? { apiKey: anthropicApiKey } : {}),
+        maxTurns,
+        mcpServers: { 'qa-research': server },
+        allowedTools: [
+          'mcp__qa-research__inspect',
+          'mcp__qa-research__readFixture',
+          'mcp__qa-research__readTest',
+          'mcp__qa-research__writeTest',
+          'mcp__qa-research__runTest',
+          'mcp__qa-research__done'
+        ]
+      }
+    })) {
+      if (message.type === 'assistant' && message.message?.content) {
+        for (const block of message.message.content) {
+          if ('text' in block && block.text) {
+            console.warn(`  Claude: ${block.text.slice(0, 200)}`)
+          }
+          if ('name' in block) {
+            console.warn(
+              `  Tool: ${block.name}(${JSON.stringify(block.input).slice(0, 100)})`
+            )
+          }
+        }
+      }
+      if (agentDone) break
+    }
+  } catch (e) {
+    const errMsg = e instanceof Error ? e.message : String(e)
+    console.warn(`Research error: ${errMsg}`)
+
+    // Detect billing/auth errors and surface them clearly
+    if (
+      errMsg.includes('Credit balance is too low') ||
+      errMsg.includes('insufficient_quota') ||
+      errMsg.includes('rate_limit')
+    ) {
+      finalSummary = `API error: ${errMsg.slice(0, 200)}`
+      finalEvidence = 'Agent could not start due to API billing/auth issue'
+      console.warn(
+        '::error::Anthropic API credits exhausted — cannot run research phase'
+      )
+    }
+  }
+
+  // Auto-complete: if a test passed but done() was never called, use the passing test
+  if (!agentDone && lastPassedTurn >= 0 && finalTestCode) {
+    console.warn(
+      `Auto-completing: test passed at turn ${lastPassedTurn} but done() was not called`
+    )
+    finalVerdict = 'REPRODUCED'
+    finalReproducedBy = 'e2e_test'
+    finalSummary = `Test passed at turn ${lastPassedTurn} (auto-completed — agent did not call done())`
+    finalEvidence = `Test passed with exit code 0`
+  }
+
+  const result: ResearchResult = {
+    verdict: finalVerdict,
+    reproducedBy: finalReproducedBy,
+    summary: finalSummary,
+    evidence: finalEvidence,
+    testCode: finalTestCode,
+    videoScript: finalVideoScript || undefined,
+    log: researchLog
+  }
+
+  writeFileSync(`${testDir}/research-log.json`, JSON.stringify(result, null, 2))
+  console.warn(
+    `Research complete: ${finalVerdict} (${researchLog.length} tool calls)`
+  )
+
+  return result
+}

.claude/skills/comfy-qa/scripts/qa-analyze-pr.test.ts

@@ -0,0 +1,84 @@
+import { describe, expect, it } from 'vitest'
+
+import { extractMediaUrls } from './qa-analyze-pr'
+
+describe('extractMediaUrls', () => {
+  it('extracts markdown image URLs', () => {
+    const text = '![screenshot](https://example.com/image.png)'
+    expect(extractMediaUrls(text)).toEqual(['https://example.com/image.png'])
+  })
+
+  it('extracts multiple markdown images', () => {
+    const text = [
+      '![before](https://example.com/before.png)',
+      'Some text',
+      '![after](https://example.com/after.jpg)'
+    ].join('\n')
+    expect(extractMediaUrls(text)).toEqual([
+      'https://example.com/before.png',
+      'https://example.com/after.jpg'
+    ])
+  })
+
+  it('extracts raw URLs with media extensions', () => {
+    const text = 'Check this: https://cdn.example.com/demo.mp4 for details'
+    expect(extractMediaUrls(text)).toEqual(['https://cdn.example.com/demo.mp4'])
+  })
+
+  it('extracts GitHub user-attachments URLs', () => {
+    const text =
+      'https://github.com/user-attachments/assets/abc12345-6789-0def-1234-567890abcdef'
+    expect(extractMediaUrls(text)).toEqual([
+      'https://github.com/user-attachments/assets/abc12345-6789-0def-1234-567890abcdef'
+    ])
+  })
+
+  it('extracts private-user-images URLs', () => {
+    const text =
+      'https://private-user-images.githubusercontent.com/12345/abcdef-1234?jwt=token123'
+    expect(extractMediaUrls(text)).toEqual([
+      'https://private-user-images.githubusercontent.com/12345/abcdef-1234?jwt=token123'
+    ])
+  })
+
+  it('extracts URLs with query parameters', () => {
+    const text = 'https://example.com/image.png?w=800&h=600'
+    expect(extractMediaUrls(text)).toEqual([
+      'https://example.com/image.png?w=800&h=600'
+    ])
+  })
+
+  it('deduplicates URLs', () => {
+    const text = [
+      '![img](https://example.com/same.png)',
+      '![img2](https://example.com/same.png)',
+      'Also https://example.com/same.png'
+    ].join('\n')
+    expect(extractMediaUrls(text)).toEqual(['https://example.com/same.png'])
+  })
+
+  it('returns empty array for empty input', () => {
+    expect(extractMediaUrls('')).toEqual([])
+  })
+
+  it('returns empty array for text with no media URLs', () => {
+    expect(extractMediaUrls('Just some text without any URLs')).toEqual([])
+  })
+
+  it('handles mixed media types', () => {
+    const text = [
+      '![screen](https://example.com/screenshot.png)',
+      'Video: https://example.com/demo.webm',
+      '![gif](https://example.com/animation.gif)'
+    ].join('\n')
+    const urls = extractMediaUrls(text)
+    expect(urls).toContain('https://example.com/screenshot.png')
+    expect(urls).toContain('https://example.com/demo.webm')
+    expect(urls).toContain('https://example.com/animation.gif')
+  })
+
+  it('ignores non-http URLs in markdown', () => {
+    const text = '![local](./local-image.png)'
+    expect(extractMediaUrls(text)).toEqual([])
+  })
+})

.claude/skills/comfy-qa/scripts/qa-analyze-pr.ts

@@ -0,0 +1,799 @@
+#!/usr/bin/env tsx
+/**
+ * QA PR Analysis Script
+ *
+ * Deeply analyzes a PR using Gemini Pro to generate targeted QA guides
+ * for before/after recording sessions. Fetches PR thread, extracts media,
+ * and produces structured test plans.
+ *
+ * Usage:
+ *   pnpm exec tsx scripts/qa-analyze-pr.ts \
+ *     --pr-number 10270 \
+ *     --repo owner/repo \
+ *     --output-dir qa-guides/ \
+ *     [--model gemini-3.1-pro-preview]
+ *
+ * Env: GEMINI_API_KEY (required)
+ */
+
+import { execSync } from 'node:child_process'
+import { mkdirSync, readFileSync, writeFileSync } from 'node:fs'
+import { resolve } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+import { GoogleGenerativeAI } from '@google/generative-ai'
+
+// ── Types ──
+
+interface QaGuideStep {
+  action: string
+  description: string
+  expected_before?: string
+  expected_after?: string
+}
+
+interface QaGuide {
+  summary: string
+  test_focus: string
+  prerequisites: string[]
+  steps: QaGuideStep[]
+  visual_checks: string[]
+}
+
+interface PrThread {
+  title: string
+  body: string
+  labels: string[]
+  issueComments: string[]
+  reviewComments: string[]
+  reviews: string[]
+  diff: string
+}
+
+type TargetType = 'pr' | 'issue'
+
+interface Options {
+  prNumber: string
+  repo: string
+  outputDir: string
+  model: string
+  apiKey: string
+  mediaBudgetBytes: number
+  maxVideoBytes: number
+  type: TargetType
+}
+
+// ── CLI parsing ──
+
+function parseArgs(): Options {
+  const args = process.argv.slice(2)
+  const opts: Partial<Options> = {
+    model: 'gemini-3.1-pro-preview',
+    apiKey: process.env.GEMINI_API_KEY || '',
+    mediaBudgetBytes: 20 * 1024 * 1024,
+    maxVideoBytes: 10 * 1024 * 1024,
+    type: 'pr'
+  }
+
+  for (let i = 0; i < args.length; i++) {
+    switch (args[i]) {
+      case '--pr-number':
+        opts.prNumber = args[++i]
+        break
+      case '--repo':
+        opts.repo = args[++i]
+        break
+      case '--output-dir':
+        opts.outputDir = args[++i]
+        break
+      case '--model':
+        opts.model = args[++i]
+        break
+      case '--type':
+        opts.type = args[++i] as TargetType
+        break
+      case '--help':
+        console.warn(
+          'Usage: qa-analyze-pr.ts --pr-number <num> --repo <owner/repo> --output-dir <path> [--model <model>] [--type pr|issue]'
+        )
+        process.exit(0)
+    }
+  }
+
+  if (!opts.prNumber || !opts.repo || !opts.outputDir) {
+    console.error(
+      'Required: --pr-number <num> --repo <owner/repo> --output-dir <path>'
+    )
+    process.exit(1)
+  }
+
+  if (!opts.apiKey) {
+    console.error('GEMINI_API_KEY environment variable is required')
+    process.exit(1)
+  }
+
+  return opts as Options
+}
+
+// ── PR thread fetching ──
+
+function ghExec(cmd: string): string {
+  try {
+    return execSync(cmd, {
+      encoding: 'utf-8',
+      timeout: 30_000,
+      stdio: ['pipe', 'pipe', 'pipe']
+    }).trim()
+  } catch (err) {
+    console.warn(`gh command failed: ${cmd}`)
+    console.warn((err as Error).message)
+    return ''
+  }
+}
+
+function fetchPrThread(prNumber: string, repo: string): PrThread {
+  console.warn('Fetching PR thread...')
+
+  const prView = ghExec(
+    `gh pr view ${prNumber} --repo ${repo} --json title,body,labels`
+  )
+  const prData = prView
+    ? JSON.parse(prView)
+    : { title: '', body: '', labels: [] }
+
+  const issueCommentsRaw = ghExec(
+    `gh api repos/${repo}/issues/${prNumber}/comments --paginate`
+  )
+  const issueComments: string[] = issueCommentsRaw
+    ? JSON.parse(issueCommentsRaw).map((c: { body: string }) => c.body)
+    : []
+
+  const reviewCommentsRaw = ghExec(
+    `gh api repos/${repo}/pulls/${prNumber}/comments --paginate`
+  )
+  const reviewComments: string[] = reviewCommentsRaw
+    ? JSON.parse(reviewCommentsRaw).map((c: { body: string }) => c.body)
+    : []
+
+  const reviewsRaw = ghExec(
+    `gh api repos/${repo}/pulls/${prNumber}/reviews --paginate`
+  )
+  const reviews: string[] = reviewsRaw
+    ? JSON.parse(reviewsRaw)
+        .filter((r: { body: string }) => r.body)
+        .map((r: { body: string }) => r.body)
+    : []
+
+  const diff = ghExec(`gh pr diff ${prNumber} --repo ${repo}`)
+
+  console.warn(
+    `PR #${prNumber}: "${prData.title}" | ` +
+      `${issueComments.length} issue comments, ` +
+      `${reviewComments.length} review comments, ` +
+      `${reviews.length} reviews, ` +
+      `diff: ${diff.length} chars`
+  )
+
+  return {
+    title: prData.title || '',
+    body: prData.body || '',
+    labels: (prData.labels || []).map((l: { name: string }) => l.name),
+    issueComments,
+    reviewComments,
+    reviews,
+    diff
+  }
+}
+
+interface IssueThread {
+  title: string
+  body: string
+  labels: string[]
+  comments: string[]
+}
+
+function fetchIssueThread(issueNumber: string, repo: string): IssueThread {
+  console.warn('Fetching issue thread...')
+
+  const issueView = ghExec(
+    `gh issue view ${issueNumber} --repo ${repo} --json title,body,labels`
+  )
+  const issueData = issueView
+    ? JSON.parse(issueView)
+    : { title: '', body: '', labels: [] }
+
+  const commentsRaw = ghExec(
+    `gh api repos/${repo}/issues/${issueNumber}/comments --paginate`
+  )
+  const comments: string[] = commentsRaw
+    ? JSON.parse(commentsRaw).map((c: { body: string }) => c.body)
+    : []
+
+  console.warn(
+    `Issue #${issueNumber}: "${issueData.title}" | ` +
+      `${comments.length} comments`
+  )
+
+  return {
+    title: issueData.title || '',
+    body: issueData.body || '',
+    labels: (issueData.labels || []).map((l: { name: string }) => l.name),
+    comments
+  }
+}
+
+// ── Media extraction ──
+
+const MEDIA_EXTENSIONS = /\.(png|jpg|jpeg|gif|webp|mp4|webm|mov)$/i
+
+const MEDIA_URL_PATTERNS = [
+  // Markdown images: ![alt](url)
+  /!\[[^\]]*\]\(([^)]+)\)/g,
+  // GitHub user-attachments
+  /https:\/\/github\.com\/user-attachments\/assets\/[a-f0-9-]+/g,
+  // Private user images
+  /https:\/\/private-user-images\.githubusercontent\.com\/[^\s)"]+/g,
+  // Raw URLs with media extensions (standalone or in text)
+  /(?<!="|=')https?:\/\/[^\s)<>"]+\.(?:png|jpg|jpeg|gif|webp|mp4|webm|mov)(?:\?[^\s)<>"]*)?/gi
+]
+
+export function extractMediaUrls(text: string): string[] {
+  if (!text) return []
+
+  const urls = new Set<string>()
+
+  for (const pattern of MEDIA_URL_PATTERNS) {
+    // Reset lastIndex for global patterns
+    pattern.lastIndex = 0
+    let match: RegExpExecArray | null
+    while ((match = pattern.exec(text)) !== null) {
+      // For markdown images, the URL is in capture group 1
+      const url = match[1] || match[0]
+      // Clean trailing markdown/html artifacts
+      const cleaned = url.replace(/[)>"'\s]+$/, '')
+      if (cleaned.startsWith('http')) {
+        urls.add(cleaned)
+      }
+    }
+  }
+
+  return [...urls]
+}
+
+// ── Media downloading ──
+
+const ALLOWED_MEDIA_DOMAINS = [
+  'github.com',
+  'raw.githubusercontent.com',
+  'user-images.githubusercontent.com',
+  'private-user-images.githubusercontent.com',
+  'objects.githubusercontent.com',
+  'github.githubassets.com'
+]
+
+function isAllowedMediaDomain(url: string): boolean {
+  try {
+    const hostname = new URL(url).hostname
+    return ALLOWED_MEDIA_DOMAINS.some(
+      (domain) => hostname === domain || hostname.endsWith(`.${domain}`)
+    )
+  } catch {
+    return false
+  }
+}
+
+async function downloadMedia(
+  urls: string[],
+  outputDir: string,
+  budgetBytes: number,
+  maxVideoBytes: number
+): Promise<Array<{ path: string; mimeType: string }>> {
+  const downloaded: Array<{ path: string; mimeType: string }> = []
+  let totalBytes = 0
+
+  const mediaDir = resolve(outputDir, 'media')
+  mkdirSync(mediaDir, { recursive: true })
+
+  for (const url of urls) {
+    if (totalBytes >= budgetBytes) {
+      console.warn(
+        `Media budget exhausted (${totalBytes} bytes), skipping rest`
+      )
+      break
+    }
+
+    if (!isAllowedMediaDomain(url)) {
+      console.warn(`Skipping non-GitHub URL: ${url.slice(0, 80)}`)
+      continue
+    }
+
+    try {
+      const response = await fetch(url, {
+        signal: AbortSignal.timeout(15_000),
+        headers: { Accept: 'image/*,video/*' },
+        redirect: 'follow'
+      })
+
+      if (!response.ok) {
+        console.warn(`Failed to download ${url}: ${response.status}`)
+        continue
+      }
+
+      const contentLength = response.headers.get('content-length')
+      if (contentLength) {
+        const declaredSize = Number.parseInt(contentLength, 10)
+        if (declaredSize > budgetBytes - totalBytes) {
+          console.warn(
+            `Content-Length ${declaredSize} would exceed budget, skipping ${url}`
+          )
+          continue
+        }
+      }
+
+      const contentType = response.headers.get('content-type') || ''
+      const buffer = Buffer.from(await response.arrayBuffer())
+
+      // Skip oversized videos
+      const isVideo =
+        contentType.startsWith('video/') || /\.(mp4|webm|mov)$/i.test(url)
+      if (isVideo && buffer.length > maxVideoBytes) {
+        console.warn(
+          `Skipping large video ${url} (${(buffer.length / 1024 / 1024).toFixed(1)}MB > ${(maxVideoBytes / 1024 / 1024).toFixed(0)}MB cap)`
+        )
+        continue
+      }
+
+      if (totalBytes + buffer.length > budgetBytes) {
+        console.warn(`Would exceed budget, skipping ${url}`)
+        continue
+      }
+
+      const ext = guessExtension(url, contentType)
+      const filename = `media-${downloaded.length}${ext}`
+      const filepath = resolve(mediaDir, filename)
+      writeFileSync(filepath, buffer)
+      totalBytes += buffer.length
+
+      const mimeType = contentType.split(';')[0].trim() || guessMimeType(ext)
+
+      downloaded.push({ path: filepath, mimeType })
+      console.warn(
+        `Downloaded: ${url.slice(0, 80)}... (${(buffer.length / 1024).toFixed(0)}KB)`
+      )
+    } catch (err) {
+      console.warn(`Failed to download ${url}: ${(err as Error).message}`)
+    }
+  }
+
+  console.warn(
+    `Downloaded ${downloaded.length}/${urls.length} media files ` +
+      `(${(totalBytes / 1024 / 1024).toFixed(1)}MB)`
+  )
+  return downloaded
+}
+
+function guessExtension(url: string, contentType: string): string {
+  const urlMatch = url.match(MEDIA_EXTENSIONS)
+  if (urlMatch) return urlMatch[0].toLowerCase()
+
+  const typeMap: Record<string, string> = {
+    'image/png': '.png',
+    'image/jpeg': '.jpg',
+    'image/gif': '.gif',
+    'image/webp': '.webp',
+    'video/mp4': '.mp4',
+    'video/webm': '.webm'
+  }
+  return typeMap[contentType.split(';')[0]] || '.bin'
+}
+
+function guessMimeType(ext: string): string {
+  const map: Record<string, string> = {
+    '.png': 'image/png',
+    '.jpg': 'image/jpeg',
+    '.jpeg': 'image/jpeg',
+    '.gif': 'image/gif',
+    '.webp': 'image/webp',
+    '.mp4': 'video/mp4',
+    '.webm': 'video/webm',
+    '.mov': 'video/quicktime'
+  }
+  return map[ext] || 'application/octet-stream'
+}
+
+// ── Gemini analysis ──
+
+function buildIssueAnalysisPrompt(issue: IssueThread): string {
+  const allText = [
+    `# Issue: ${issue.title}`,
+    '',
+    '## Description',
+    issue.body,
+    '',
+    issue.comments.length > 0
+      ? `## Comments\n${issue.comments.join('\n\n---\n\n')}`
+      : ''
+  ]
+    .filter(Boolean)
+    .join('\n')
+
+  return `You are a senior QA engineer analyzing a bug report for ComfyUI frontend — a node-based visual workflow editor for AI image generation (Vue 3 + TypeScript).
+
+The UI has:
+- A large canvas (1280x720 viewport) showing a node graph centered at ~(640, 400)
+- Nodes are boxes with input/output slots connected by wires
+- A hamburger menu (top-left C logo) with File, Edit, Help submenus
+- Sidebars (Workflows, Node Library, Models)
+- A topbar with workflow tabs and Queue button
+- The default workflow loads with these nodes (approximate center coordinates):
+  - Load Checkpoint (~150, 300), CLIP Text Encode x2 (~450, 250 and ~450, 450)
+  - Empty Latent Image (~450, 600), KSampler (~750, 350), VAE Decode (~1000, 350), Save Image (~1200, 350)
+- Right-clicking ON a node shows node actions (Clone, Bypass, Convert, etc.)
+- Right-clicking on EMPTY canvas shows Add Node menu — different from node context menu
+
+Your task: Generate a DETAILED reproduction guide (8-15 steps) to trigger this bug on main.
+
+${allText}
+
+## Available test actions
+Each step must use one of these actions:
+
+### Menu actions
+- "openMenu" — clicks the Comfy hamburger menu (top-left C logo)
+- "hoverMenuItem" — hovers a top-level menu item to open submenu (label required)
+- "clickMenuItem" — clicks an item in the visible submenu (label required)
+
+### Element actions (by visible text)
+- "click" — clicks an element by visible text (text required)
+- "rightClick" — right-clicks an element to open context menu (text required)
+- "doubleClick" — double-clicks an element or coordinates (text or x,y)
+- "fillDialog" — fills dialog input and presses Enter (text required)
+- "pressKey" — presses a keyboard key (key required: Escape, Tab, Delete, Enter, etc.)
+
+### Canvas actions (by coordinates — viewport is 1280x720)
+- "clickCanvas" — click at coordinates (x, y required)
+- "rightClickCanvas" — right-click at coordinates (x, y required)
+- "doubleClick" — double-click at coordinates to open node search (x, y)
+- "dragCanvas" — drag from one point to another (fromX, fromY, toX, toY)
+- "scrollCanvas" — scroll wheel for zoom (x, y, deltaY: negative=zoom in, positive=zoom out)
+
+### Utility
+- "wait" — waits briefly (ms required, max 3000)
+- "screenshot" — takes a screenshot (name required)
+
+## Common ComfyUI interactions
+- Right-click a node → context menu with Clone, Bypass, Remove, Colors, etc.
+- Double-click empty canvas → opens node search dialog
+- Ctrl+C / Ctrl+V → copy/paste selected nodes
+- Delete key → remove selected node
+- Ctrl+G → group selected nodes
+- Drag from output slot to input slot → create connection
+- Click a node to select it, Shift+click for multi-select
+
+## Output format
+Return a JSON object with exactly one key: "reproduce", containing:
+{
+  "summary": "One sentence: what bug this issue reports",
+  "test_focus": "Specific behavior to reproduce",
+  "prerequisites": ["e.g. Load default workflow"],
+  "steps": [
+    {
+      "action": "clickCanvas",
+      "description": "Click on first node to select it",
+      "expected_before": "What should happen if the bug is present"
+    }
+  ],
+  "visual_checks": ["Specific visual evidence of the bug to look for"]
+}
+
+## Rules
+- Generate 8-15 DETAILED steps that actually trigger the reported bug.
+- Follow the issue's reproduction steps PRECISELY — translate them into available actions.
+- Use canvas coordinates for node interactions (nodes are typically in the center area 300-900 x 200-500).
+- Take screenshots BEFORE and AFTER critical actions to capture the bug state.
+- Do NOT just open a menu and screenshot — actually perform the full reproduction sequence.
+- Do NOT include login steps.
+- Output ONLY valid JSON, no markdown fences or explanation.`
+}
+
+function buildAnalysisPrompt(thread: PrThread): string {
+  const allText = [
+    `# PR: ${thread.title}`,
+    '',
+    '## Description',
+    thread.body,
+    '',
+    thread.issueComments.length > 0
+      ? `## Issue Comments\n${thread.issueComments.join('\n\n---\n\n')}`
+      : '',
+    thread.reviewComments.length > 0
+      ? `## Review Comments\n${thread.reviewComments.join('\n\n---\n\n')}`
+      : '',
+    thread.reviews.length > 0
+      ? `## Reviews\n${thread.reviews.join('\n\n---\n\n')}`
+      : '',
+    '',
+    '## Diff (truncated)',
+    '```',
+    thread.diff.slice(0, 8000),
+    '```'
+  ]
+    .filter(Boolean)
+    .join('\n')
+
+  return `You are a senior QA engineer analyzing a pull request for ComfyUI frontend (a Vue 3 + TypeScript web application for AI image generation workflows).
+
+Your task: Generate TWO targeted QA test guides — one for BEFORE the PR (main branch) and one for AFTER (PR branch).
+
+${allText}
+
+## Available test actions
+Each step must use one of these actions:
+- "openMenu" — clicks the Comfy hamburger menu (top-left C logo)
+- "hoverMenuItem" — hovers a top-level menu item to open submenu (label required)
+- "clickMenuItem" — clicks an item in the visible submenu (label required)
+- "fillDialog" — fills dialog input and presses Enter (text required)
+- "pressKey" — presses a keyboard key (key required)
+- "click" — clicks an element by visible text (text required)
+- "wait" — waits briefly (ms required, max 3000)
+- "screenshot" — takes a screenshot (name required)
+
+## Output format
+Return a JSON object with exactly two keys: "before" and "after", each containing:
+{
+  "summary": "One sentence: what this PR changes",
+  "test_focus": "Specific behaviors to verify in this recording",
+  "prerequisites": ["e.g. Load default workflow"],
+  "steps": [
+    {
+      "action": "openMenu",
+      "description": "Open the main menu to check file options",
+      "expected_before": "Old behavior description (before key only)",
+      "expected_after": "New behavior description (after key only)"
+    }
+  ],
+  "visual_checks": ["Specific visual elements to look for"]
+}
+
+## Rules
+- BEFORE guide: 2-4 steps, under 15 seconds. Show OLD/missing behavior.
+- AFTER guide: 3-6 steps, under 30 seconds. Prove the fix/feature works.
+- Focus on the SPECIFIC behavior changed by this PR, not generic testing.
+- Use information from PR description, screenshots, and comments to understand intended behavior.
+- Include at least one screenshot step in each guide.
+- Do NOT include login steps.
+- Menu pattern: openMenu -> hoverMenuItem -> clickMenuItem or screenshot.
+- Output ONLY valid JSON, no markdown fences or explanation.`
+}
+
+async function analyzeWithGemini(
+  thread: PrThread,
+  media: Array<{ path: string; mimeType: string }>,
+  model: string,
+  apiKey: string
+): Promise<{ before: QaGuide; after: QaGuide }> {
+  const genAI = new GoogleGenerativeAI(apiKey)
+  const geminiModel = genAI.getGenerativeModel({ model })
+
+  const prompt = buildAnalysisPrompt(thread)
+
+  const parts: Array<
+    { text: string } | { inlineData: { mimeType: string; data: string } }
+  > = [{ text: prompt }]
+
+  // Add media as inline data
+  for (const item of media) {
+    try {
+      const buffer = readFileSync(item.path)
+      parts.push({
+        inlineData: {
+          mimeType: item.mimeType,
+          data: buffer.toString('base64')
+        }
+      })
+    } catch (err) {
+      console.warn(
+        `Failed to read media ${item.path}: ${(err as Error).message}`
+      )
+    }
+  }
+
+  console.warn(
+    `Sending to ${model}: ${prompt.length} chars text, ${media.length} media files`
+  )
+
+  const result = await geminiModel.generateContent({
+    contents: [{ role: 'user', parts }],
+    generationConfig: {
+      temperature: 0.2,
+      maxOutputTokens: 8192,
+      responseMimeType: 'application/json'
+    }
+  })
+
+  let text = result.response.text()
+  // Strip markdown fences if present
+  text = text
+    .replace(/^```(?:json)?\n?/gm, '')
+    .replace(/```$/gm, '')
+    .trim()
+
+  console.warn('Gemini response received')
+  console.warn('Raw response (first 500 chars):', text.slice(0, 500))
+  const parsed = JSON.parse(text)
+
+  // Handle different response shapes from Gemini
+  let before: QaGuide
+  let after: QaGuide
+
+  if (Array.isArray(parsed) && parsed.length >= 2) {
+    // Array format: [before, after]
+    before = parsed[0]
+    after = parsed[1]
+  } else if (parsed.before && parsed.after) {
+    // Object format: { before, after }
+    before = parsed.before
+    after = parsed.after
+  } else {
+    // Try nested wrapper keys
+    const inner = parsed.qa_guide ?? parsed.guides ?? parsed
+    if (inner.before && inner.after) {
+      before = inner.before
+      after = inner.after
+    } else {
+      console.warn(
+        'Full response:',
+        JSON.stringify(parsed, null, 2).slice(0, 2000)
+      )
+      throw new Error(
+        `Unexpected response shape. Got keys: ${Object.keys(parsed).join(', ')}`
+      )
+    }
+  }
+
+  return { before, after }
+}
+
+async function analyzeIssueWithGemini(
+  issue: IssueThread,
+  media: Array<{ path: string; mimeType: string }>,
+  model: string,
+  apiKey: string
+): Promise<QaGuide> {
+  const genAI = new GoogleGenerativeAI(apiKey)
+  const geminiModel = genAI.getGenerativeModel({ model })
+
+  const prompt = buildIssueAnalysisPrompt(issue)
+
+  const parts: Array<
+    { text: string } | { inlineData: { mimeType: string; data: string } }
+  > = [{ text: prompt }]
+
+  for (const item of media) {
+    try {
+      const buffer = readFileSync(item.path)
+      parts.push({
+        inlineData: {
+          mimeType: item.mimeType,
+          data: buffer.toString('base64')
+        }
+      })
+    } catch (err) {
+      console.warn(
+        `Failed to read media ${item.path}: ${(err as Error).message}`
+      )
+    }
+  }
+
+  console.warn(
+    `Sending to ${model}: ${prompt.length} chars text, ${media.length} media files`
+  )
+
+  const result = await geminiModel.generateContent({
+    contents: [{ role: 'user', parts }],
+    generationConfig: {
+      temperature: 0.2,
+      maxOutputTokens: 8192,
+      responseMimeType: 'application/json'
+    }
+  })
+
+  let text = result.response.text()
+  text = text
+    .replace(/^```(?:json)?\n?/gm, '')
+    .replace(/```$/gm, '')
+    .trim()
+
+  console.warn('Gemini response received')
+  console.warn('Raw response (first 500 chars):', text.slice(0, 500))
+  const parsed = JSON.parse(text)
+
+  const guide: QaGuide =
+    parsed.reproduce ?? parsed.qa_guide?.reproduce ?? parsed
+  return guide
+}
+
+// ── Main ──
+
+async function main() {
+  const opts = parseArgs()
+  mkdirSync(opts.outputDir, { recursive: true })
+
+  if (opts.type === 'issue') {
+    await analyzeIssue(opts)
+  } else {
+    await analyzePr(opts)
+  }
+}
+
+async function analyzeIssue(opts: Options) {
+  const issue = fetchIssueThread(opts.prNumber, opts.repo)
+
+  const allText = [issue.body, ...issue.comments].join('\n')
+  const mediaUrls = extractMediaUrls(allText)
+  console.warn(`Found ${mediaUrls.length} media URLs`)
+
+  const media = await downloadMedia(
+    mediaUrls,
+    opts.outputDir,
+    opts.mediaBudgetBytes,
+    opts.maxVideoBytes
+  )
+
+  const guide = await analyzeIssueWithGemini(
+    issue,
+    media,
+    opts.model,
+    opts.apiKey
+  )
+
+  const beforePath = resolve(opts.outputDir, 'qa-guide-before.json')
+  writeFileSync(beforePath, JSON.stringify(guide, null, 2))
+
+  console.warn(`Wrote QA guide:`)
+  console.warn(`  Reproduce: ${beforePath}`)
+}
+
+async function analyzePr(opts: Options) {
+  const thread = fetchPrThread(opts.prNumber, opts.repo)
+
+  const allText = [
+    thread.body,
+    ...thread.issueComments,
+    ...thread.reviewComments,
+    ...thread.reviews
+  ].join('\n')
+  const mediaUrls = extractMediaUrls(allText)
+  console.warn(`Found ${mediaUrls.length} media URLs`)
+
+  const media = await downloadMedia(
+    mediaUrls,
+    opts.outputDir,
+    opts.mediaBudgetBytes,
+    opts.maxVideoBytes
+  )
+
+  const guides = await analyzeWithGemini(thread, media, opts.model, opts.apiKey)
+
+  const beforePath = resolve(opts.outputDir, 'qa-guide-before.json')
+  const afterPath = resolve(opts.outputDir, 'qa-guide-after.json')
+  writeFileSync(beforePath, JSON.stringify(guides.before, null, 2))
+  writeFileSync(afterPath, JSON.stringify(guides.after, null, 2))
+
+  console.warn(`Wrote QA guides:`)
+  console.warn(`  Before: ${beforePath}`)
+  console.warn(`  After:  ${afterPath}`)
+}
+
+function isExecutedAsScript(metaUrl: string): boolean {
+  const modulePath = fileURLToPath(metaUrl)
+  const scriptPath = process.argv[1] ? resolve(process.argv[1]) : ''
+  return modulePath === scriptPath
+}
+
+if (isExecutedAsScript(import.meta.url)) {
+  main().catch((err) => {
+    console.error('PR analysis failed:', err)
+    process.exit(1)
+  })
+}

.claude/skills/comfy-qa/scripts/qa-deploy-pages.sh

@@ -0,0 +1,413 @@
+#!/usr/bin/env bash
+# Deploy QA report to Cloudflare Pages.
+# Expected env vars: CLOUDFLARE_API_TOKEN, CLOUDFLARE_ACCOUNT_ID, RAW_BRANCH,
+#   BEFORE_SHA, AFTER_SHA, TARGET_NUM, TARGET_TYPE, REPO, RUN_ID
+# Writes outputs to GITHUB_OUTPUT: badge_status, url
+set -euo pipefail
+
+npm install -g wrangler@4.74.0 >/dev/null 2>&1
+
+DEPLOY_DIR=$(mktemp -d)
+mkdir -p "$DEPLOY_DIR"
+
+for os in Linux macOS Windows; do
+  DIR="qa-artifacts/qa-report-${os}-${RUN_ID}"
+  for prefix in qa qa-before; do
+    VID="${DIR}/${prefix}-session.mp4"
+    if [ -f "$VID" ]; then
+      DEST="$DEPLOY_DIR/${prefix}-${os}.mp4"
+      cp "$VID" "$DEST"
+      echo "Found ${prefix} ${os} video ($(du -h "$VID" | cut -f1))"
+    fi
+  done
+  # Copy multi-pass session videos (qa-session-1, qa-session-2, etc.)
+  for numbered in "$DIR"/qa-session-[0-9].mp4; do
+    [ -f "$numbered" ] || continue
+    NUM=$(basename "$numbered" | sed 's/qa-session-\([0-9]\).mp4/\1/')
+    DEST="$DEPLOY_DIR/qa-${os}-pass${NUM}.mp4"
+    cp "$numbered" "$DEST"
+    echo "Found pass ${NUM} ${os} video ($(du -h "$numbered" | cut -f1))"
+  done
+  # Generate GIF thumbnail from after video (or first pass)
+  THUMB_SRC="$DEPLOY_DIR/qa-${os}.mp4"
+  [ ! -f "$THUMB_SRC" ] && THUMB_SRC="$DEPLOY_DIR/qa-${os}-pass1.mp4"
+  if [ -f "$THUMB_SRC" ]; then
+    ffmpeg -y -ss 10 -i "$THUMB_SRC" -t 8 \
+      -vf "fps=8,scale=480:-1:flags=lanczos,split[s0][s1];[s0]palettegen=max_colors=64[p];[s1][p]paletteuse=dither=bayer" \
+      -loop 0 "$DEPLOY_DIR/qa-${os}-thumb.gif" 2>/dev/null \
+    || echo "GIF generation failed for ${os} (non-fatal)"
+  fi
+done
+
+# Build video cards and report sections
+CARDS=""
+# shellcheck disable=SC2034 # accessed via eval
+ICONS_Linux="🐧" ICONS_macOS="🍎" ICONS_Windows="🪟"
+CARD_COUNT=0
+DL_ICON="<svg width=14 height=14 viewBox='0 0 24 24' fill=none stroke=currentColor stroke-width=2><path d='M21 15v4a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2v-4'/><polyline points='7 10 12 15 17 10'/><line x1=12 y1=15 x2=12 y2=3'/></svg>"
+
+for os in Linux macOS Windows; do
+  eval "ICON=\$ICONS_${os}"
+  OS_LOWER=$(echo "$os" | tr '[:upper:]' '[:lower:]')
+  HAS_BEFORE=$([ -f "$DEPLOY_DIR/qa-before-${os}.mp4" ] && echo 1 || echo 0)
+  HAS_AFTER=$( { [ -f "$DEPLOY_DIR/qa-${os}.mp4" ] || [ -f "$DEPLOY_DIR/qa-${os}-pass1.mp4" ]; } && echo 1 || echo 0)
+  [ "$HAS_AFTER" = "0" ] && continue
+
+  # Collect all reports for this platform (single + multi-pass)
+  REPORT_FILES=""
+  REPORT_LINK=""
+  REPORT_HTML=""
+  for rpt in "video-reviews/${OS_LOWER}-qa-video-report.md" "video-reviews/${OS_LOWER}-pass"*-qa-video-report.md; do
+    [ -f "$rpt" ] && REPORT_FILES="${REPORT_FILES} ${rpt}"
+  done
+
+  if [ -n "$REPORT_FILES" ]; then
+    # Concatenate all reports into one combined report file
+    COMBINED_MD=""
+    for rpt in $REPORT_FILES; do
+      cp "$rpt" "$DEPLOY_DIR/$(basename "$rpt")"
+      RPT_MD=$(sed 's/&/\&amp;/g; s/</\&lt;/g; s/>/\&gt;/g' "$rpt")
+      [ -n "$COMBINED_MD" ] && COMBINED_MD="${COMBINED_MD}&#10;&#10;---&#10;&#10;"
+      COMBINED_MD="${COMBINED_MD}${RPT_MD}"
+    done
+    FIRST_REPORT=$(echo "$REPORT_FILES" | awk '{print $1}')
+    FIRST_BASENAME=$(basename "$FIRST_REPORT")
+    REPORT_LINK="<a class=dl href=${FIRST_BASENAME}><svg width=14 height=14 viewBox='0 0 24 24' fill=none stroke=currentColor stroke-width=2><path d='M14 2H6a2 2 0 0 0-2 2v16a2 2 0 0 0 2 2h12a2 2 0 0 0 2-2V8z'/><polyline points='14 2 14 8 20 8'/><line x1=16 y1=13 x2=8 y2=13/><line x1=16 y1=17 x2=8 y2=17'/></svg>Report</a>"
+    REPORT_HTML="<details class=report open><summary><svg width=14 height=14 viewBox='0 0 24 24' fill=none stroke=currentColor stroke-width=2><circle cx=12 cy=12 r=10/><line x1=12 y1=16 x2=12 y2=12/><line x1=12 y1=8 x2=12.01 y2=8'/></svg> AI Comparative Review</summary><div class=report-body data-md>${COMBINED_MD}</div></details>"
+  fi
+
+  if [ "$HAS_BEFORE" = "1" ]; then
+    CARDS="${CARDS}<div class='card reveal' style='--i:${CARD_COUNT}'><div class=card-header><span class=platform><span class=icon>${ICON}</span>${os}</span><span class=links>${REPORT_LINK}</span></div><div class=comparison><div class=comp-panel><div class=comp-label>Before <span class=comp-tag>main</span></div><div class=video-wrap><video controls preload=auto><source src=qa-before-${os}.mp4 type=video/mp4></video></div><div class=comp-dl><a class=dl href=qa-before-${os}.mp4 download>${DL_ICON}Before</a></div></div><div class=comp-panel><div class=comp-label>After <span class=comp-tag>PR</span></div><div class=video-wrap><video controls preload=auto><source src=qa-${os}.mp4 type=video/mp4></video></div><div class=comp-dl><a class=dl href=qa-${os}.mp4 download>${DL_ICON}After</a></div></div></div>${REPORT_HTML}</div>"
+  elif [ -f "$DEPLOY_DIR/qa-${os}.mp4" ]; then
+    CARDS="${CARDS}<div class='card reveal' style='--i:${CARD_COUNT}'><div class=video-wrap><video controls preload=auto><source src=qa-${os}.mp4 type=video/mp4></video></div><div class=card-body><span class=platform><span class=icon>${ICON}</span>${os}</span><span class=links><a class=dl href=qa-${os}.mp4 download>${DL_ICON}Download</a>${REPORT_LINK}</span></div>${REPORT_HTML}</div>"
+  else
+    PASS_VIDEOS=""
+    for pass_vid in "$DEPLOY_DIR/qa-${os}-pass"[0-9].mp4; do
+      [ -f "$pass_vid" ] || continue
+      PASS_NUM=$(basename "$pass_vid" | sed "s/qa-${os}-pass\([0-9]\).mp4/\1/")
+      PASS_VIDEOS="${PASS_VIDEOS}<div class=comp-panel><div class=comp-label>Pass ${PASS_NUM}</div><div class=video-wrap><video controls preload=auto><source src=qa-${os}-pass${PASS_NUM}.mp4 type=video/mp4></video></div><div class=comp-dl><a class=dl href=qa-${os}-pass${PASS_NUM}.mp4 download>${DL_ICON}Pass ${PASS_NUM}</a></div></div>"
+    done
+    CARDS="${CARDS}<div class='card reveal' style='--i:${CARD_COUNT}'><div class=card-header><span class=platform><span class=icon>${ICON}</span>${os}</span><span class=links>${REPORT_LINK}</span></div><div class=comparison>${PASS_VIDEOS}</div>${REPORT_HTML}</div>"
+  fi
+  CARD_COUNT=$((CARD_COUNT + 1))
+done
+
+# Build commit info and target link for the report header
+COMMIT_HTML=""
+REPO_URL="https://github.com/${REPO}"
+if [ -n "${TARGET_NUM:-}" ]; then
+  if [ "$TARGET_TYPE" = "issue" ]; then
+    COMMIT_HTML="<a href=${REPO_URL}/issues/${TARGET_NUM} class=sha title='Issue'>Issue #${TARGET_NUM}</a>"
+  else
+    COMMIT_HTML="<a href=${REPO_URL}/pull/${TARGET_NUM} class=sha title='Pull Request'>PR #${TARGET_NUM}</a>"
+  fi
+fi
+if [ -n "${BEFORE_SHA:-}" ]; then
+  SHORT_BEFORE="${BEFORE_SHA:0:7}"
+  COMMIT_HTML="${COMMIT_HTML:+${COMMIT_HTML} &middot; }<a href=${REPO_URL}/commit/${BEFORE_SHA} class=sha title='main branch'>main @ ${SHORT_BEFORE}</a>"
+fi
+if [ -n "${AFTER_SHA:-}" ]; then
+  SHORT_AFTER="${AFTER_SHA:0:7}"
+  AFTER_LABEL="PR"
+  [ -n "${TARGET_NUM:-}" ] && AFTER_LABEL="#${TARGET_NUM}"
+  COMMIT_HTML="${COMMIT_HTML:+${COMMIT_HTML} &middot; }<a href=${REPO_URL}/commit/${AFTER_SHA} class=sha title='PR head commit'>${AFTER_LABEL} @ ${SHORT_AFTER}</a>"
+fi
+if [ -n "${PIPELINE_SHA:-}" ]; then
+  SHORT_PIPE="${PIPELINE_SHA:0:7}"
+  COMMIT_HTML="${COMMIT_HTML:+${COMMIT_HTML} &middot; }<a href=${REPO_URL}/commit/${PIPELINE_SHA} class=sha title='QA pipeline version'>QA @ ${SHORT_PIPE}</a>"
+fi
+[ -n "$COMMIT_HTML" ] && COMMIT_HTML=" &middot; ${COMMIT_HTML}"
+
+RUN_LINK=""
+if [ -n "${RUN_URL:-}" ]; then
+  RUN_LINK=" &middot; <a href=\"${RUN_URL}\" class=sha title=\"GitHub Actions run\">CI Job</a>"
+fi
+
+# Timing info
+DEPLOY_TIME=$(date -u '+%Y-%m-%d %H:%M UTC')
+TIMING_HTML=""
+if [ -n "${RUN_START_TIME:-}" ]; then
+  TIMING_HTML=" &middot; <span class=sha title='Pipeline timing'>${RUN_START_TIME} &rarr; ${DEPLOY_TIME}</span>"
+fi
+
+# Generate index.html from template
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+TEMPLATE="$SCRIPT_DIR/qa-report-template.html"
+
+# Write dynamic content to temp files for safe substitution
+# Cloudflare Pages _headers file — enable range requests for video seeking
+cat > "$DEPLOY_DIR/_headers" <<'HEADERSEOF'
+/*.mp4
+  Accept-Ranges: bytes
+  Cache-Control: public, max-age=86400
+HEADERSEOF
+
+# Build purpose description from pr-context.txt
+PURPOSE_HTML=""
+if [ -f pr-context.txt ]; then
+  # Extract title line and first paragraph of description
+  PR_TITLE=$(grep -m1 '^Title:' pr-context.txt 2>/dev/null | sed 's/^Title: //' || true)
+  if [ "$TARGET_TYPE" = "issue" ]; then
+    PURPOSE_LABEL="Issue #${TARGET_NUM}"
+    PURPOSE_VERB="reports"
+  else
+    PURPOSE_LABEL="PR #${TARGET_NUM}"
+    PURPOSE_VERB="aims to"
+  fi
+  # Get first ~300 chars of description body (after "Description:" line)
+  PR_DESC=$(sed -n '/^Description:/,/^###/p' pr-context.txt 2>/dev/null | grep -v '^Description:\|^###' | head -5 | sed 's/&/\&amp;/g; s/</\&lt;/g; s/>/\&gt;/g' | tr '\n' ' ' | head -c 400 || true)
+  [ -z "$PR_DESC" ] && PR_DESC=$(sed -n '3,8p' pr-context.txt 2>/dev/null | sed 's/&/\&amp;/g; s/</\&lt;/g; s/>/\&gt;/g' | tr '\n' ' ' | head -c 400 || true)
+  # Build requirements from QA guide JSON
+  REQS_HTML=""
+  QA_GUIDE=$(ls qa-guides/qa-guide-*.json 2>/dev/null | head -1 || true)
+  if [ -f "$QA_GUIDE" ]; then
+    PREREQS=$(python3 -c "
+import json, sys, html
+try:
+  g = json.load(open(sys.argv[1]))
+  prereqs = g.get('prerequisites', [])
+  steps = g.get('steps', [])
+  focus = g.get('test_focus', '')
+  parts = []
+  if focus:
+    parts.append('<strong>Test focus:</strong> ' + html.escape(focus))
+  if prereqs:
+    parts.append('<strong>Prerequisites:</strong> ' + ', '.join(html.escape(p) for p in prereqs))
+  if steps:
+    parts.append('<strong>Steps:</strong> ' + ' → '.join(html.escape(s.get('description', str(s))) for s in steps[:6]))
+    if len(steps) > 6:
+      parts[-1] += ' → ...'
+  print('<br>'.join(parts))
+except: pass
+" "$QA_GUIDE" 2>/dev/null)
+    [ -n "$PREREQS" ] && REQS_HTML="<div class=purpose-reqs>${PREREQS}</div>"
+  fi
+
+  PURPOSE_HTML="<div class=purpose><div class=purpose-label>${PURPOSE_LABEL} ${PURPOSE_VERB}</div><strong>${PR_TITLE}</strong><br>${PR_DESC}${REQS_HTML}</div>"
+fi
+
+echo -n "$COMMIT_HTML" > "$DEPLOY_DIR/.commit_html"
+echo -n "$CARDS" > "$DEPLOY_DIR/.cards_html"
+echo -n "$RUN_LINK" > "$DEPLOY_DIR/.run_link"
+# Badge HTML with copy button (placeholder URL filled after deploy)
+echo -n '<div class="badge-bar"><img src="badge.svg" alt="QA Badge" class="badge-img"/><button class="copy-badge" title="Copy badge markdown" onclick="copyBadge()"><svg width=14 height=14 viewBox="0 0 24 24" fill=none stroke=currentColor stroke-width=2><rect x=9 y=9 width=13 height=13 rx=2/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/></svg></button></div>' > "$DEPLOY_DIR/.badge_html"
+echo -n "${TIMING_HTML:-}" > "$DEPLOY_DIR/.timing_html"
+echo -n "$PURPOSE_HTML" > "$DEPLOY_DIR/.purpose_html"
+python3 -c "
+import sys, pathlib
+d = pathlib.Path(sys.argv[1])
+t = pathlib.Path(sys.argv[2]).read_text()
+t = t.replace('{{COMMIT_HTML}}', (d / '.commit_html').read_text())
+t = t.replace('{{CARDS}}', (d / '.cards_html').read_text())
+t = t.replace('{{RUN_LINK}}', (d / '.run_link').read_text())
+t = t.replace('{{BADGE_HTML}}', (d / '.badge_html').read_text())
+t = t.replace('{{TIMING_HTML}}', (d / '.timing_html').read_text())
+t = t.replace('{{PURPOSE_HTML}}', (d / '.purpose_html').read_text())
+sys.stdout.write(t)
+" "$DEPLOY_DIR" "$TEMPLATE" > "$DEPLOY_DIR/index.html"
+rm -f "$DEPLOY_DIR/.commit_html" "$DEPLOY_DIR/.cards_html" "$DEPLOY_DIR/.run_link" "$DEPLOY_DIR/.badge_html" "$DEPLOY_DIR/.timing_html" "$DEPLOY_DIR/.purpose_html"
+
+cat > "$DEPLOY_DIR/404.html" <<'ERROREOF'
+<!DOCTYPE html><html lang=en><head><meta charset=utf-8><title>404</title>
+<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;600&display=swap" rel=stylesheet>
+<style>:root{--bg:oklch(8% 0.02 265);--fg:oklch(45% 0.01 265);--err:oklch(62% 0.22 25)}*{margin:0;padding:0;box-sizing:border-box}body{background:var(--bg);color:var(--fg);font-family:'Inter',system-ui,sans-serif;display:flex;align-items:center;justify-content:center;min-height:100vh}div{text-align:center}h1{color:var(--err);font-size:clamp(3rem,8vw,5rem);font-weight:700;letter-spacing:-.04em;margin-bottom:.5rem}p{font-size:1rem;max-width:32ch;line-height:1.5}</style>
+</head><body><div><h1>404</h1><p>File not found. The QA recording may have failed or been cancelled.</p></div></body></html>
+ERROREOF
+
+# Copy research log to deploy dir if it exists
+for rlog in qa-artifacts/*/research/research-log.json qa-artifacts/*/*/research/research-log.json qa-artifacts/before/*/research/research-log.json; do
+  if [ -f "$rlog" ]; then
+    cp "$rlog" "$DEPLOY_DIR/research-log.json"
+    echo "Found research log: $rlog"
+    break
+  fi
+done
+
+# Copy generated test code to deploy dir
+for tfile in qa-artifacts/*/research/reproduce.spec.ts qa-artifacts/*/*/research/reproduce.spec.ts qa-artifacts/before/*/research/reproduce.spec.ts; do
+  if [ -f "$tfile" ]; then
+    cp "$tfile" "$DEPLOY_DIR/reproduce.spec.ts"
+    echo "Found test code: $tfile"
+    break
+  fi
+done
+
+# Copy video script if available
+for vsfile in qa-artifacts/*/video-script.spec.ts qa-artifacts/*/*/video-script.spec.ts qa-artifacts/before/*/video-script.spec.ts; do
+  if [ -f "$vsfile" ]; then
+    cp "$vsfile" "$DEPLOY_DIR/video-script.spec.ts"
+    echo "Found video script: $vsfile"
+    break
+  fi
+done
+
+# Generate badge SVGs into deploy dir
+# Priority: research-log.json verdict (a11y-verified) > video review verdict (AI interpretation)
+REPRO_COUNT=0 INCONC_COUNT=0 NOT_REPRO_COUNT=0 TOTAL_REPORTS=0
+
+# Try research log first (ground truth from a11y assertions)
+RESEARCH_VERDICT=""
+REPRO_METHOD=""
+if [ -f "$DEPLOY_DIR/research-log.json" ]; then
+  RESEARCH_VERDICT=$(python3 -c "import json,sys; d=json.load(open(sys.argv[1])); print(d.get('verdict',''))" "$DEPLOY_DIR/research-log.json" 2>/dev/null || true)
+  REPRO_METHOD=$(python3 -c "import json,sys; d=json.load(open(sys.argv[1])); print(d.get('reproducedBy','none'))" "$DEPLOY_DIR/research-log.json" 2>/dev/null || true)
+  echo "Research verdict (a11y-verified): ${RESEARCH_VERDICT:-none} (by: ${REPRO_METHOD:-none})"
+  if [ -n "$RESEARCH_VERDICT" ]; then
+    TOTAL_REPORTS=1
+    case "$RESEARCH_VERDICT" in
+      REPRODUCED) REPRO_COUNT=1 ;;
+      NOT_REPRODUCIBLE) NOT_REPRO_COUNT=1 ;;
+      INCONCLUSIVE) INCONC_COUNT=1 ;;
+    esac
+  fi
+fi
+
+# Check video review verdicts (always, not just as fallback)
+VIDEO_REPRODUCED=false
+if [ -d video-reviews ]; then
+  for rpt in video-reviews/*-qa-video-report.md; do
+    [ -f "$rpt" ] || continue
+    VERDICT_JSON=$(grep -oP '"verdict":\s*"[A-Z_]+' "$rpt" 2>/dev/null | tail -1 | grep -oP '[A-Z_]+$' || true)
+
+    if [ -n "$VERDICT_JSON" ]; then
+      echo "Video review verdict: $VERDICT_JSON ($(basename "$rpt"))"
+      [ "$VERDICT_JSON" = "REPRODUCED" ] && VIDEO_REPRODUCED=true
+      # Only count video as separate report if no research log
+      if [ -z "$RESEARCH_VERDICT" ]; then
+        TOTAL_REPORTS=$((TOTAL_REPORTS + 1))
+        case "$VERDICT_JSON" in
+          REPRODUCED) REPRO_COUNT=$((REPRO_COUNT + 1)) ;;
+          NOT_REPRODUCIBLE) NOT_REPRO_COUNT=$((NOT_REPRO_COUNT + 1)) ;;
+          INCONCLUSIVE) INCONC_COUNT=$((INCONC_COUNT + 1)) ;;
+        esac
+      fi
+    elif [ -z "$RESEARCH_VERDICT" ]; then
+      TOTAL_REPORTS=$((TOTAL_REPORTS + 1))
+      # Fallback: grep Summary section (for older reports without ## Verdict)
+      SUMM=$(sed -n '/^## Summary/,/^## /p' "$rpt" 2>/dev/null | head -15)
+      if echo "$SUMM" | grep -iq 'INCONCLUSIVE'; then
+        INCONC_COUNT=$((INCONC_COUNT + 1))
+      elif echo "$SUMM" | grep -iq 'not reproduced\|could not reproduce\|could not be confirmed\|unable to reproduce\|fails\? to reproduce\|fails\? to perform\|was NOT\|NOT visible\|not observed\|fail.* to demonstrate\|does not demonstrate\|steps were not performed\|never.*tested\|never.*accessed\|not.* confirmed'; then
+        NOT_REPRO_COUNT=$((NOT_REPRO_COUNT + 1))
+      elif echo "$SUMM" | grep -iq 'reproduc\|confirm'; then
+        REPRO_COUNT=$((REPRO_COUNT + 1))
+        VIDEO_REPRODUCED=true
+      fi
+    fi
+  done
+fi
+
+# Upgrade reproduction method to "both" when E2E and video agree
+if [ "$REPRO_METHOD" = "e2e_test" ] && [ "$VIDEO_REPRODUCED" = "true" ]; then
+  REPRO_METHOD="both"
+  echo "Upgraded reproducedBy to 'both' (E2E + video review agree)"
+elif [ -z "$RESEARCH_VERDICT" ] && [ "$VIDEO_REPRODUCED" = "true" ]; then
+  REPRO_METHOD="video"
+fi
+FAIL_COUNT=$((TOTAL_REPORTS - REPRO_COUNT - NOT_REPRO_COUNT))
+[ "$FAIL_COUNT" -lt 0 ] && FAIL_COUNT=0
+echo "DEBUG verdict: repro=${REPRO_COUNT} not_repro=${NOT_REPRO_COUNT} inconc=${INCONC_COUNT} fail=${FAIL_COUNT} total=${TOTAL_REPORTS}"
+
+# Warn on verdict mismatch between E2E and video review
+if [ -n "$RESEARCH_VERDICT" ]; then
+  VIDEO_VERDICT=$(grep -oP '"verdict":\s*"[A-Z_]+' video-reviews/*-qa-video-report.md 2>/dev/null | tail -1 | grep -oP '[A-Z_]+$' || true)
+  if [ -n "$VIDEO_VERDICT" ] && [ "$RESEARCH_VERDICT" != "$VIDEO_VERDICT" ]; then
+    echo "⚠ Verdict mismatch: E2E=$RESEARCH_VERDICT vs Video=$VIDEO_VERDICT (E2E takes priority)"
+  fi
+fi
+
+echo "Verdict: ${REPRO_COUNT}✓ ${NOT_REPRO_COUNT}✗ ${FAIL_COUNT}⚠ / ${TOTAL_REPORTS}"
+
+# Badge text:
+#   Single pass: "REPRODUCED" / "NOT REPRODUCIBLE" / "INCONCLUSIVE"
+#   Multi pass:  "2✓ 0✗ 1⚠ / 3" with color based on dominant result
+REPRO_RESULT="" REPRO_COLOR="#9f9f9f"
+if [ "$TOTAL_REPORTS" -le 1 ]; then
+  # Single report — simple label
+  if [ "$REPRO_COUNT" -gt 0 ]; then
+    REPRO_RESULT="REPRODUCED" REPRO_COLOR="#2196f3"
+  elif [ "$NOT_REPRO_COUNT" -gt 0 ]; then
+    REPRO_RESULT="NOT REPRODUCIBLE" REPRO_COLOR="#9f9f9f"
+  elif [ "$FAIL_COUNT" -gt 0 ]; then
+    REPRO_RESULT="INCONCLUSIVE" REPRO_COLOR="#9f9f9f"
+  fi
+else
+  # Multi pass — show breakdown: X✓ Y✗ Z⚠ / N
+  PARTS=""
+  [ "$REPRO_COUNT" -gt 0 ] && PARTS="${REPRO_COUNT}✓"
+  [ "$NOT_REPRO_COUNT" -gt 0 ] && PARTS="${PARTS:+${PARTS} }${NOT_REPRO_COUNT}✗"
+  [ "$FAIL_COUNT" -gt 0 ] && PARTS="${PARTS:+${PARTS} }${FAIL_COUNT}⚠"
+  REPRO_RESULT="${PARTS} / ${TOTAL_REPORTS}"
+  # Color based on best outcome
+  if [ "$REPRO_COUNT" -gt 0 ]; then
+    REPRO_COLOR="#2196f3"
+  elif [ "$NOT_REPRO_COUNT" -gt 0 ]; then
+    REPRO_COLOR="#9f9f9f"
+  fi
+fi
+
+# Badge label: #NUM QA0327 (with today's date)
+QA_DATE=$(date -u '+%m%d')
+BADGE_LABEL="QA${QA_DATE}"
+[ -n "${TARGET_NUM:-}" ] && BADGE_LABEL="#${TARGET_NUM} QA${QA_DATE}"
+
+# For PRs, also extract fix quality from Overall Risk section
+FIX_RESULT="" FIX_COLOR="#4c1"
+if [ "$TARGET_TYPE" != "issue" ]; then
+  # Try structured JSON risk first
+  ALL_RISKS=$(grep -ohP '"risk":\s*"[a-z]+' video-reviews/*.md 2>/dev/null | grep -oP '[a-z]+$' || true)
+  if [ -n "$ALL_RISKS" ]; then
+    # Use worst risk across all reports
+    if echo "$ALL_RISKS" | grep -q 'high'; then
+      FIX_RESULT="MAJOR ISSUES" FIX_COLOR="#e05d44"
+    elif echo "$ALL_RISKS" | grep -q 'medium'; then
+      FIX_RESULT="MINOR ISSUES" FIX_COLOR="#dfb317"
+    elif echo "$ALL_RISKS" | grep -q 'low'; then
+      FIX_RESULT="APPROVED" FIX_COLOR="#4c1"
+    fi
+  else
+    # Fallback: grep Overall Risk section
+    RISK_TEXT=""
+    if [ -d video-reviews ]; then
+      RISK_TEXT=$(sed -n '/^## Overall Risk/,/^## /p' video-reviews/*.md 2>/dev/null | sed 's/\*//g' | head -20 || true)
+    fi
+    RISK_FIRST=$(echo "$RISK_TEXT" | grep -oiP '^\s*(high|medium|moderate|low|minimal|critical)' | head -1 | tr '[:upper:]' '[:lower:]' || true)
+    if [ -n "$RISK_FIRST" ]; then
+      case "$RISK_FIRST" in
+        *low*|*minimal*) FIX_RESULT="APPROVED" FIX_COLOR="#4c1" ;;
+        *medium*|*moderate*) FIX_RESULT="MINOR ISSUES" FIX_COLOR="#dfb317" ;;
+        *high*|*critical*) FIX_RESULT="MAJOR ISSUES" FIX_COLOR="#e05d44" ;;
+      esac
+    elif echo "$RISK_TEXT" | grep -iq 'no.*risk\|approved\|looks good'; then
+      FIX_RESULT="APPROVED" FIX_COLOR="#4c1"
+    fi
+  fi
+fi
+
+# Always use vertical box badge
+/tmp/gen-badge-box.sh "$DEPLOY_DIR/badge.svg" "$BADGE_LABEL" \
+  "$REPRO_COUNT" "$NOT_REPRO_COUNT" "$FAIL_COUNT" "$TOTAL_REPORTS" \
+  "$FIX_RESULT" "$FIX_COLOR" "$REPRO_METHOD"
+BADGE_STATUS="${REPRO_RESULT:-UNKNOWN}${FIX_RESULT:+ | Fix: ${FIX_RESULT}}"
+echo "badge_status=${BADGE_STATUS:-FINISHED}" >> "$GITHUB_OUTPUT"
+
+# Remove files exceeding Cloudflare Pages 25MB limit to prevent silent deploy failures
+MAX_SIZE=$((25 * 1024 * 1024))
+find "$DEPLOY_DIR" -type f -size +${MAX_SIZE}c | while read -r big_file; do
+  SIZE_MB=$(( $(stat -c%s "$big_file") / 1024 / 1024 ))
+  echo "Removing oversized file: $(basename "$big_file") (${SIZE_MB}MB > 25MB limit)"
+  rm "$big_file"
+done
+
+BRANCH=$(echo "$RAW_BRANCH" | sed 's/[^a-zA-Z0-9-]/-/g' | sed 's/--*/-/g' | sed 's/^-//;s/-$//' | cut -c1-28)
+
+DEPLOY_OUTPUT=$(wrangler pages deploy "$DEPLOY_DIR" \
+  --project-name="comfy-qa" \
+  --branch="$BRANCH" 2>&1) || true
+echo "$DEPLOY_OUTPUT" | tail -5
+
+URL=$(echo "$DEPLOY_OUTPUT" | grep -oE 'https://[a-zA-Z0-9.-]+\.pages\.dev\S*' | head -1 || true)
+FALLBACK_URL="https://${BRANCH}.comfy-qa.pages.dev"
+
+echo "url=${URL:-$FALLBACK_URL}" >> "$GITHUB_OUTPUT"
+echo "Deployed to: ${URL:-$FALLBACK_URL}"

.claude/skills/comfy-qa/scripts/qa-generate-test.ts

@@ -0,0 +1,208 @@
+#!/usr/bin/env tsx
+/**
+ * Generates a Playwright regression test (.spec.ts) from a QA report + PR diff.
+ * Uses Gemini to produce a test that asserts UIUX behavior verified during QA.
+ *
+ * Usage:
+ *   pnpm exec tsx scripts/qa-generate-test.ts \
+ *     --qa-report <path>       QA video review report (markdown)
+ *     --pr-diff <path>         PR diff file
+ *     --output <path>          Output .spec.ts file path
+ *     --model <name>           Gemini model (default: gemini-3-flash-preview)
+ */
+import { readFile, writeFile } from 'node:fs/promises'
+import { basename, resolve } from 'node:path'
+
+import { GoogleGenerativeAI } from '@google/generative-ai'
+
+interface CliOptions {
+  qaReport: string
+  prDiff: string
+  output: string
+  model: string
+}
+
+const DEFAULTS: CliOptions = {
+  qaReport: '',
+  prDiff: '',
+  output: '',
+  model: 'gemini-3-flash-preview'
+}
+
+// ── Fixture API reference for the prompt ────────────────────────────
+const FIXTURE_API = `
+## ComfyUI Playwright Test Fixture API
+
+Import pattern:
+\`\`\`typescript
+import { expect } from '@playwright/test'
+import { comfyPageFixture as test } from '../fixtures/ComfyPage'
+\`\`\`
+
+### Available helpers on \`comfyPage\`:
+- \`comfyPage.page\` — raw Playwright Page
+- \`comfyPage.menu.topbar\` — Topbar helper:
+  - \`.getTabNames(): Promise<string[]>\` — get all open tab names
+  - \`.getActiveTabName(): Promise<string>\` — get active tab name
+  - \`.saveWorkflow(name)\` — Save via File > Save dialog
+  - \`.saveWorkflowAs(name)\` — Save via File > Save As dialog
+  - \`.exportWorkflow(name)\` — Export via File > Export dialog
+  - \`.triggerTopbarCommand(path: string[])\` — e.g. ['File', 'Save As']
+  - \`.getWorkflowTab(name)\` — get a tab locator by name
+  - \`.closeWorkflowTab(name)\` — close a tab
+  - \`.openTopbarMenu()\` — open the hamburger menu
+  - \`.openSubmenu(label)\` — hover to open a submenu
+- \`comfyPage.menu.workflowsTab\` — Workflows sidebar:
+  - \`.open()\` / \`.close()\` — toggle sidebar
+  - \`.getTopLevelSavedWorkflowNames()\` — list saved workflows
+  - \`.getPersistedItem(name)\` — get a workflow item locator
+- \`comfyPage.workflow\` — WorkflowHelper:
+  - \`.loadWorkflow(name)\` — load from browser_tests/assets/{name}.json
+  - \`.setupWorkflowsDirectory(structure)\` — setup test directory
+  - \`.deleteWorkflow(name)\` — delete a workflow
+  - \`.isCurrentWorkflowModified(): Promise<boolean>\` — check dirty state
+  - \`.getUndoQueueSize()\` / \`.getRedoQueueSize()\`
+- \`comfyPage.settings.setSetting(key, value)\` — change settings
+- \`comfyPage.keyboard\` — KeyboardHelper:
+  - \`.undo()\` / \`.redo()\` / \`.bypass()\`
+- \`comfyPage.nodeOps\` — NodeOperationsHelper
+- \`comfyPage.canvas\` — CanvasHelper
+- \`comfyPage.contextMenu\` — ContextMenu
+- \`comfyPage.toast\` — ToastHelper
+- \`comfyPage.confirmDialog\` — confirmation dialog
+- \`comfyPage.nextFrame()\` — wait for Vue re-render
+
+### Test patterns:
+- Use \`test.describe('Name', { tag: '@ui' }, () => { ... })\` for UI tests
+- Use \`test.beforeEach\` to set up common state (settings, workflow dir)
+- Use \`expect(locator).toHaveScreenshot('name.png')\` for visual assertions
+- Use \`expect(locator).toBeVisible()\` / \`.toHaveText()\` for behavioral assertions
+- Use \`comfyPage.workflow.setupWorkflowsDirectory({})\` to ensure clean state
+`
+
+// ── Prompt builder ──────────────────────────────────────────────────
+function buildPrompt(qaReport: string, prDiff: string): string {
+  return `You are a Playwright test generator for the ComfyUI frontend.
+
+Your task: Generate a single .spec.ts regression test file that asserts the UIUX behavior
+described in the QA report below. The test must:
+
+1. Use the ComfyUI Playwright fixture API (documented below)
+2. Test UIUX behavior ONLY — element visibility, tab names, dialog states, workflow states
+3. NOT test code implementation details
+4. Be concise — only test the behavior that the PR changed
+5. Follow existing test conventions (see API reference)
+
+${FIXTURE_API}
+
+## QA Video Review Report
+${qaReport}
+
+## PR Diff (for context on what changed)
+${prDiff.slice(0, 8000)}
+
+## Output Requirements
+- Output ONLY the .spec.ts file content — no markdown fences, no explanations
+- Start with imports, end with closing brace
+- Use descriptive test names that explain the expected behavior
+- Add screenshot assertions where visual verification matters
+- Keep it focused: 2-5 test cases covering the core behavioral change
+- Use \`test.beforeEach\` for common setup (settings, workflow directory)
+- Tag the describe block with \`{ tag: '@ui' }\` or \`{ tag: '@workflow' }\` as appropriate
+`
+}
+
+// ── Gemini call ─────────────────────────────────────────────────────
+async function generateTest(
+  qaReport: string,
+  prDiff: string,
+  model: string
+): Promise<string> {
+  const apiKey = process.env.GEMINI_API_KEY
+  if (!apiKey) throw new Error('GEMINI_API_KEY env var required')
+
+  const genAI = new GoogleGenerativeAI(apiKey)
+  const genModel = genAI.getGenerativeModel({ model })
+
+  const prompt = buildPrompt(qaReport, prDiff)
+  console.warn(`Sending prompt to ${model} (${prompt.length} chars)...`)
+
+  const result = await genModel.generateContent({
+    contents: [{ role: 'user', parts: [{ text: prompt }] }],
+    generationConfig: {
+      temperature: 0.2,
+      maxOutputTokens: 8192
+    }
+  })
+
+  const text = result.response.text()
+
+  // Strip markdown fences if model wraps output
+  return text
+    .replace(/^```(?:typescript|ts)?\n?/, '')
+    .replace(/\n?```$/, '')
+    .trim()
+}
+
+// ── CLI ─────────────────────────────────────────────────────────────
+function parseArgs(): CliOptions {
+  const args = process.argv.slice(2)
+  const opts = { ...DEFAULTS }
+
+  for (let i = 0; i < args.length; i++) {
+    switch (args[i]) {
+      case '--qa-report':
+        opts.qaReport = args[++i]
+        break
+      case '--pr-diff':
+        opts.prDiff = args[++i]
+        break
+      case '--output':
+        opts.output = args[++i]
+        break
+      case '--model':
+        opts.model = args[++i]
+        break
+      case '--help':
+        console.warn(`Usage:
+  pnpm exec tsx scripts/qa-generate-test.ts [options]
+
+Options:
+  --qa-report <path>   QA video review report (markdown) [required]
+  --pr-diff <path>     PR diff file [required]
+  --output <path>      Output .spec.ts path [required]
+  --model <name>       Gemini model (default: gemini-3-flash-preview)`)
+        process.exit(0)
+    }
+  }
+
+  if (!opts.qaReport || !opts.prDiff || !opts.output) {
+    console.error('Missing required args. Run with --help for usage.')
+    process.exit(1)
+  }
+
+  return opts
+}
+
+async function main() {
+  const opts = parseArgs()
+
+  const qaReport = await readFile(resolve(opts.qaReport), 'utf-8')
+  const prDiff = await readFile(resolve(opts.prDiff), 'utf-8')
+
+  console.warn(
+    `QA report: ${basename(opts.qaReport)} (${qaReport.length} chars)`
+  )
+  console.warn(`PR diff: ${basename(opts.prDiff)} (${prDiff.length} chars)`)
+
+  const testCode = await generateTest(qaReport, prDiff, opts.model)
+
+  const outputPath = resolve(opts.output)
+  await writeFile(outputPath, testCode + '\n')
+  console.warn(`Generated test: ${outputPath} (${testCode.length} chars)`)
+}
+
+main().catch((err) => {
+  console.error(err)
+  process.exit(1)
+})

.claude/skills/comfy-qa/scripts/qa-report-template.html

@@ -0,0 +1,164 @@
+<!DOCTYPE html><html lang=en><head><meta charset=utf-8><meta name=viewport content="width=device-width,initial-scale=1"><title>QA Session Recordings</title>
+<link rel=preconnect href=https://fonts.googleapis.com><link rel=preconnect href=https://fonts.gstatic.com crossorigin><link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500&display=swap" rel=stylesheet>
+<script src="https://cdn.jsdelivr.net/npm/marked/marked.min.js"></script>
+<style>
+:root{--bg:oklch(97% 0.01 265);--surface:oklch(100% 0 0);--surface-up:oklch(94% 0.01 265);--fg:oklch(15% 0.02 265);--fg-muted:oklch(40% 0.01 265);--fg-dim:oklch(55% 0.01 265);--primary:oklch(50% 0.21 265);--primary-up:oklch(45% 0.21 265);--primary-glow:oklch(55% 0.15 265);--ok:oklch(45% 0.18 155);--err:oklch(50% 0.22 25);--border:oklch(85% 0.01 265);--border-faint:oklch(90% 0.01 265);--r:0.75rem;--r-lg:1rem;--ease-out:cubic-bezier(0.22,1,0.36,1);--dur-base:250ms;--dur-slow:500ms;--font:'Inter',system-ui,sans-serif;--font-mono:'JetBrains Mono',monospace}
+@media(prefers-color-scheme:dark){:root{--bg:oklch(8% 0.02 265);--surface:oklch(12% 0.02 265);--surface-up:oklch(16% 0.02 265);--fg:oklch(96% 0.01 95);--fg-muted:oklch(65% 0.01 265);--fg-dim:oklch(45% 0.01 265);--primary:oklch(62% 0.21 265);--primary-up:oklch(68% 0.21 265);--primary-glow:oklch(62% 0.15 265);--ok:oklch(62% 0.18 155);--err:oklch(62% 0.22 25);--border:oklch(22% 0.02 265);--border-faint:oklch(15% 0.01 265)}}
+*{margin:0;padding:0;box-sizing:border-box}
+body{background:var(--bg);color:var(--fg);font-family:var(--font);min-height:100vh;padding:clamp(1.5rem,4vw,3rem) clamp(1rem,3vw,2rem);position:relative}
+@media(prefers-color-scheme:dark){body::after{content:'';position:fixed;inset:0;pointer-events:none;opacity:.03;background:url("data:image/svg+xml,%3Csvg viewBox='0 0 256 256' xmlns='http://www.w3.org/2000/svg'%3E%3Cfilter id='n'%3E%3CfeTurbulence type='fractalNoise' baseFrequency='.85' numOctaves='4' stitchTiles='stitch'/%3E%3C/filter%3E%3Crect width='100%25' height='100%25' filter='url(%23n)'/%3E%3C/svg%3E")}}
+.container{max-width:1200px;margin:0 auto}
+header{display:flex;align-items:center;gap:1rem;margin-bottom:clamp(1.5rem,4vw,3rem);padding-bottom:1.25rem;border-bottom:1px solid var(--border)}
+.header-icon{width:36px;height:36px;display:grid;place-items:center;background:linear-gradient(135deg,oklch(100% 0 0/.06),oklch(100% 0 0/.02));backdrop-filter:blur(12px);border:1px solid oklch(100% 0 0/.1);border-radius:var(--r);flex-shrink:0}
+.header-icon svg{color:var(--primary)}
+h1{font-size:clamp(1.25rem,2.5vw,1.625rem);font-weight:700;letter-spacing:-.03em;background:linear-gradient(135deg,var(--fg),var(--fg-muted));-webkit-background-clip:text;-webkit-text-fill-color:transparent;background-clip:text}
+.meta{color:var(--fg-dim);font-size:.8125rem;margin-top:.15rem;letter-spacing:.01em}
+.grid{display:grid;grid-template-columns:repeat(auto-fill,minmax(min(480px,100%),1fr));gap:1.5rem}
+.card{background:var(--surface);border:1px solid var(--border);border-radius:var(--r-lg);overflow:hidden;transition:border-color var(--dur-base) var(--ease-out),box-shadow var(--dur-base) var(--ease-out),transform var(--dur-base) var(--ease-out)}
+.card:hover{border-color:var(--primary);box-shadow:0 4px 16px oklch(0% 0 0/.1);transform:translateY(-2px)}
+.video-wrap{position:relative;background:var(--surface);border-bottom:1px solid var(--border-faint)}
+.video-wrap video{width:100%;display:block;aspect-ratio:16/9;object-fit:contain}
+.card-body{padding:.75rem 1rem;display:flex;align-items:center;justify-content:space-between}
+.platform{display:flex;align-items:center;gap:.5rem;font-weight:600;font-size:.9375rem;letter-spacing:-.01em}
+.icon{font-size:1.125rem}
+.links{display:flex;gap:.5rem}
+.dl{color:var(--fg-muted);text-decoration:none;font-size:.75rem;font-weight:500;display:inline-flex;align-items:center;gap:.3rem;padding:.25rem .6rem;border-radius:9999px;border:1px solid var(--border);background:oklch(100% 0 0/.03);transition:all var(--dur-base) var(--ease-out)}
+.dl:hover{color:var(--primary-up);border-color:var(--primary);background:oklch(62% 0.21 265/.08)}
+.badge{font-size:.6875rem;font-weight:600;padding:.2rem .625rem;border-radius:9999px;text-transform:uppercase;letter-spacing:.05em}
+.card-header{padding:.75rem 1rem;display:flex;align-items:center;justify-content:space-between;border-bottom:1px solid var(--border-faint)}
+.comparison{display:grid;grid-template-columns:1fr 1fr;gap:0}
+.comp-panel{border-right:1px solid var(--border-faint)}
+.comp-panel:last-child{border-right:none}
+.comp-label{padding:.4rem .75rem;font-size:.7rem;font-weight:600;text-transform:uppercase;letter-spacing:.05em;color:var(--fg-muted);background:var(--surface);display:flex;align-items:center;gap:.4rem}
+.comp-tag{font-size:.6rem;padding:.1rem .4rem;border-radius:9999px;font-weight:600}
+.comp-panel:first-child .comp-tag{background:oklch(65% 0.01 265/.15);color:var(--fg-muted);border:1px solid var(--border)}
+.comp-panel:last-child .comp-tag{background:oklch(62% 0.18 155/.15);color:var(--ok);border:1px solid oklch(62% 0.18 155/.25)}
+.comp-dl{padding:.4rem .75rem;display:flex;justify-content:center}
+.report{border-top:1px solid var(--border-faint);padding:.75rem 1rem;font-size:.8125rem}
+.report summary{cursor:pointer;color:var(--fg-muted);font-weight:500;display:flex;align-items:center;gap:.4rem;user-select:none;transition:color var(--dur-base) var(--ease-out)}
+.report summary:hover{color:var(--fg)}
+.report summary svg{flex-shrink:0;opacity:.5}
+.report[open] summary{margin-bottom:.75rem;padding-bottom:.5rem;border-bottom:1px solid var(--border-faint)}
+.report-body{line-height:1.7;color:oklch(80% 0.01 265);overflow-x:auto}
+.report-body h1,.report-body h2{margin:1.25rem 0 .5rem;color:var(--fg);font-size:1rem;font-weight:600;letter-spacing:-.02em;border-bottom:1px solid var(--border-faint);padding-bottom:.4rem}
+.report-body h3{margin:.75rem 0 .4rem;color:var(--fg);font-size:.875rem;font-weight:600}
+.report-body p{margin:.4rem 0}
+.report-body ul,.report-body ol{margin:.4rem 0 .4rem 1.5rem}
+.report-body li{margin:.25rem 0}
+.report-body code{background:var(--surface-up);padding:.125rem .375rem;border-radius:.25rem;font-size:.7rem;font-family:var(--font-mono);border:1px solid var(--border-faint)}
+.report-body h3+p>code:first-child{background:oklch(62% 0.22 25/.15);color:var(--err);border-color:oklch(62% 0.22 25/.25)}
+.report-body h3+p>code:nth-child(2){background:oklch(62% 0.21 265/.15);color:var(--primary-up);border-color:oklch(62% 0.21 265/.25)}
+.report-body h3+p>code:nth-child(3){background:oklch(65% 0.01 265/.15);color:var(--fg-muted);border-color:var(--border)}
+.report-body table{width:100%;border-collapse:collapse;margin:.75rem 0;font-size:.75rem;border:1px solid var(--border);border-radius:var(--r);overflow:hidden}
+.report-body th,.report-body td{border:1px solid var(--border-faint);padding:.5rem .75rem;text-align:left;vertical-align:top;word-wrap:break-word}
+.report-body th{background:var(--surface-up);color:var(--fg);font-weight:600;font-size:.6875rem;text-transform:uppercase;letter-spacing:.05em;position:sticky;top:0;white-space:nowrap}
+.report-body tr:nth-child(even){background:color-mix(in oklch,var(--surface) 50%,transparent)}
+.report-body tr:hover{background:color-mix(in oklch,var(--surface-up) 50%,transparent)}
+.report-body strong{color:var(--fg)}
+.report-body hr{border:none;border-top:1px solid var(--border-faint);margin:1rem 0}
+@keyframes fade-up{from{opacity:0;transform:translateY(16px)}to{opacity:1;transform:translateY(0)}}
+.reveal{animation:fade-up var(--dur-slow) var(--ease-out) both;animation-delay:calc(var(--i,0) * 120ms)}
+@media(prefers-reduced-motion:reduce){.reveal{animation:none}}
+@media(max-width:480px){.grid{grid-template-columns:1fr}.card-body{flex-wrap:wrap;gap:.5rem}}
+.sha{color:var(--primary);text-decoration:none;font-family:var(--font-mono);font-size:.75rem;font-weight:500;padding:.1rem .4rem;border-radius:.25rem;background:oklch(62% 0.21 265/.08);border:1px solid oklch(62% 0.21 265/.15);transition:all var(--dur-base) var(--ease-out)}
+.sha:hover{background:oklch(62% 0.21 265/.15);border-color:var(--primary)}
+.badge-bar{display:flex;align-items:center;gap:.5rem;margin-bottom:1rem}
+.badge-img{height:20px;display:block}
+.copy-badge{background:oklch(100% 0 0/.06);border:1px solid var(--border);color:var(--fg-muted);padding:.3rem .4rem;border-radius:var(--r);cursor:pointer;display:inline-flex;align-items:center;transition:all var(--dur-base) var(--ease-out)}
+.copy-badge:hover{color:var(--primary-up);border-color:var(--primary);background:oklch(62% 0.21 265/.1)}
+.copy-badge.copied{color:var(--ok);border-color:var(--ok)}
+.vseek{width:100%;padding:0 .75rem;background:var(--surface);border-top:1px solid var(--border-faint);position:relative;height:24px;display:flex;align-items:center}
+.vseek input[type=range]{-webkit-appearance:none;appearance:none;width:100%;height:4px;background:var(--border);border-radius:2px;outline:none;cursor:pointer;position:relative;z-index:2}
+.vseek input[type=range]::-webkit-slider-thumb{-webkit-appearance:none;width:12px;height:12px;border-radius:50%;background:var(--primary);cursor:pointer;border:2px solid var(--bg);box-shadow:0 0 4px oklch(0% 0 0/.3)}
+.vseek input[type=range]::-moz-range-thumb{width:12px;height:12px;border-radius:50%;background:var(--primary);cursor:pointer;border:2px solid var(--bg)}
+.vseek .vbuf{position:absolute;left:.75rem;right:.75rem;height:4px;border-radius:2px;pointer-events:none;top:50%;transform:translateY(-50%)}
+.vseek .vbuf-bar{height:100%;background:oklch(62% 0.21 265/.25);border-radius:2px;transition:width 200ms linear}
+.vctrl{display:flex;align-items:center;gap:.375rem;padding:.5rem .75rem;background:var(--surface);border-top:1px solid var(--border-faint);flex-wrap:wrap}
+.vctrl button{background:oklch(100% 0 0/.06);border:1px solid var(--border);color:var(--fg-muted);font-size:.6875rem;font-weight:600;font-family:var(--font-mono);padding:.25rem .5rem;border-radius:.25rem;cursor:pointer;transition:all var(--dur-base) var(--ease-out);white-space:nowrap}
+.vctrl button:hover{color:var(--primary-up);border-color:var(--primary);background:oklch(62% 0.21 265/.1)}
+.vctrl button.active{color:var(--primary);border-color:var(--primary);background:oklch(62% 0.21 265/.15)}
+.vctrl .vtime{font-family:var(--font-mono);font-size:.6875rem;color:var(--fg-dim);min-width:10ch;text-align:center}
+.vctrl .vsep{width:1px;height:1rem;background:var(--border);flex-shrink:0}
+.vctrl .vhint{font-size:.6rem;color:var(--fg-dim);margin-left:auto}
+.purpose{background:linear-gradient(135deg,oklch(100% 0 0/.04),oklch(100% 0 0/.02));border:1px solid oklch(100% 0 0/.08);border-radius:var(--r-lg);padding:1rem 1.25rem;margin-bottom:1.5rem;font-size:.85rem;line-height:1.7;color:oklch(80% 0.01 265)}
+.purpose strong{color:var(--fg);font-weight:600}
+.purpose .purpose-label{font-size:.7rem;font-weight:600;text-transform:uppercase;letter-spacing:.05em;color:var(--fg-muted);margin-bottom:.4rem}
+.purpose .purpose-reqs{margin-top:.75rem;padding-top:.75rem;border-top:1px solid oklch(100% 0 0/.06);font-size:.8rem;color:oklch(70% 0.01 265);line-height:1.8}
+</style></head><body><div class=container>
+<header><div class=header-icon><svg width=20 height=20 viewBox="0 0 24 24" fill=none stroke=currentColor stroke-width=2 stroke-linecap=round stroke-linejoin=round><polygon points="23 7 16 12 23 17 23 7"/><rect x=1 y=5 width=15 height=14 rx=2 ry=2/></svg></div><div><h1>QA Session Recordings</h1><div class=meta>ComfyUI Frontend &middot; Automated QA{{COMMIT_HTML}}{{RUN_LINK}}{{TIMING_HTML}}</div>{{BADGE_HTML}}</div></header>
+{{PURPOSE_HTML}}<div class=grid>{{CARDS}}</div>
+<div id=research-section style="margin-top:2rem"></div>
+</div><script>
+// Load research-log.json and reproduce.spec.ts if available
+(async()=>{
+  const sec=document.getElementById('research-section');
+  try{
+    const [logRes,testRes,vsRes]=await Promise.allSettled([fetch('research-log.json'),fetch('reproduce.spec.ts'),fetch('video-script.spec.ts')]);
+    let html='';
+    if(logRes.status==='fulfilled'&&logRes.value.ok){
+      const log=await logRes.value.json();
+      // Show verdict banner for non-reproduced results
+      if(log.verdict&&log.verdict!=='REPRODUCED'){
+        const colors={NOT_REPRODUCIBLE:{bg:'oklch(25% 0.08 25)',border:'oklch(40% 0.15 25)',icon:'✗'},INCONCLUSIVE:{bg:'oklch(25% 0.06 80)',border:'oklch(40% 0.12 80)',icon:'⚠'}};
+        const c=colors[log.verdict]||colors.INCONCLUSIVE;
+        html+=`<div style="margin-bottom:1.5rem;padding:1.25rem;background:${c.bg};border:1px solid ${c.border};border-radius:var(--r-lg)"><div style="font-size:1.25rem;font-weight:700;margin-bottom:.5rem">${c.icon} ${log.verdict.replace(/_/g,' ')}</div><div style="font-size:.9rem;line-height:1.6;opacity:.9">${(log.summary||'No details available.').replace(/</g,'&lt;')}</div>${log.evidence?`<div style="margin-top:.75rem;padding:.75rem;background:oklch(0% 0 0/.2);border-radius:var(--r);font-family:var(--font-mono);font-size:.8rem;white-space:pre-wrap;max-height:200px;overflow:auto">${log.evidence.replace(/</g,'&lt;')}</div>`:''}</div>`;
+      }
+      html+=`<details style="margin-bottom:1.5rem"><summary style="cursor:pointer;font-weight:600;font-size:1rem;padding:.75rem 1rem;background:var(--surface);border:1px solid var(--border);border-radius:var(--r-lg)">Research Log &mdash; ${log.verdict||'?'} (${(log.log||[]).length||'?'} tool calls, ${((log.elapsedMs||0)/1000).toFixed(1)}s)</summary><div style="padding:1rem;background:var(--surface);border:1px solid var(--border);border-top:0;border-radius:0 0 var(--r-lg) var(--r-lg);overflow:auto;max-height:600px"><pre style="font-family:var(--font-mono);font-size:.8rem;line-height:1.6;white-space:pre-wrap">${JSON.stringify(log,null,2)}</pre></div></details>`;
+    }
+    if(testRes.status==='fulfilled'&&testRes.value.ok){
+      const code=await testRes.value.text();
+      html+=`<details><summary style="cursor:pointer;font-weight:600;font-size:1rem;padding:.75rem 1rem;background:var(--surface);border:1px solid var(--border);border-radius:var(--r-lg)">E2E Test Code (reproduce.spec.ts)</summary><div style="padding:1rem;background:var(--surface);border:1px solid var(--border);border-top:0;border-radius:0 0 var(--r-lg) var(--r-lg);overflow:auto;max-height:600px"><pre style="font-family:var(--font-mono);font-size:.8rem;line-height:1.6;white-space:pre-wrap">${code.replace(/</g,'&lt;').replace(/>/g,'&gt;')}</pre></div></details>`;
+    }
+    if(vsRes.status==='fulfilled'&&vsRes.value.ok){
+      const vsCode=await vsRes.value.text();
+      html+=`<details style="margin-top:1rem"><summary style="cursor:pointer;font-weight:600;font-size:1rem;padding:.75rem 1rem;background:var(--surface);border:1px solid var(--border);border-radius:var(--r-lg)">Video Script (video-script.spec.ts)</summary><div style="padding:1rem;background:var(--surface);border:1px solid var(--border);border-top:0;border-radius:0 0 var(--r-lg) var(--r-lg);overflow:auto;max-height:600px"><pre style="font-family:var(--font-mono);font-size:.8rem;line-height:1.6;white-space:pre-wrap">${vsCode.replace(/</g,'&lt;').replace(/>/g,'&gt;')}</pre></div></details>`;
+    }
+    if(html)sec.innerHTML=html;
+  }catch(e){console.warn('research load failed',e)}
+})();
+</script><script>
+function copyBadge(){const u=location.href.replace(/\/[^/]*$/,'/');const b=u+'badge.svg';const md='[![QA Badge]('+b+')]('+u+')';navigator.clipboard.writeText(md).then(()=>{const btn=document.querySelector('.copy-badge');btn.classList.add('copied');btn.innerHTML='<svg width=14 height=14 viewBox="0 0 24 24" fill=none stroke=currentColor stroke-width=2><polyline points="20 6 9 17 4 12"/></svg>';setTimeout(()=>{btn.classList.remove('copied');btn.innerHTML='<svg width=14 height=14 viewBox="0 0 24 24" fill=none stroke=currentColor stroke-width=2><rect x=9 y=9 width=13 height=13 rx=2/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/></svg>'},2000)})}
+document.querySelectorAll('[data-md]').forEach(el=>{const t=el.textContent;el.removeAttribute('data-md');el.innerHTML=marked.parse(t)});
+const FPS=30,FT=1/FPS,SPEEDS=[0.1,0.25,0.5,1,1.5,2];
+document.querySelectorAll('.video-wrap video').forEach(v=>{
+  v.playbackRate=1;
+  const c=document.createElement('div');c.className='vctrl';
+  const btn=(label,fn)=>{const b=document.createElement('button');b.textContent=label;b.onclick=fn;c.appendChild(b);return b};
+  const sep=()=>{const s=document.createElement('div');s.className='vsep';c.appendChild(s)};
+  const time=document.createElement('span');time.className='vtime';time.textContent='0:00.000';
+  btn('\u23EE',()=>{v.currentTime=0});
+  btn('\u25C0\u25C0',()=>{v.currentTime=Math.max(0,v.currentTime-FT*10)});
+  btn('\u25C0',()=>{v.pause();v.currentTime=Math.max(0,v.currentTime-FT)});
+  const playBtn=btn('\u25B6',()=>{v.paused?v.play():v.pause()});
+  btn('\u25B6\u25B6',()=>{v.pause();v.currentTime+=FT});
+  btn('\u25B6\u25B6\u25B6',()=>{v.currentTime+=FT*10});
+  sep();
+  const spdBtns=SPEEDS.map(s=>{const b=btn(s+'x',()=>{v.playbackRate=s;spdBtns.forEach(x=>x.classList.remove('active'));b.classList.add('active')});if(s===1)b.classList.add('active');return b});
+  sep();c.appendChild(time);
+  const hint=document.createElement('span');hint.className='vhint';hint.textContent='\u2190\u2192 frame \u2022 space play';c.appendChild(hint);
+  // Custom seekbar — works even without server range request support
+  const seekWrap=document.createElement('div');seekWrap.className='vseek';
+  const seekBar=document.createElement('input');seekBar.type='range';seekBar.min=0;seekBar.max=1000;seekBar.value=0;seekBar.step=1;
+  const bufWrap=document.createElement('div');bufWrap.className='vbuf';
+  const bufBar=document.createElement('div');bufBar.className='vbuf-bar';bufBar.style.width='0%';
+  bufWrap.appendChild(bufBar);seekWrap.appendChild(bufWrap);seekWrap.appendChild(seekBar);
+  let seeking=false;
+  seekBar.oninput=()=>{seeking=true;if(v.duration){v.currentTime=v.duration*(seekBar.value/1000)}};
+  seekBar.onchange=()=>{seeking=false};
+  v.closest('.video-wrap').after(seekWrap);
+  seekWrap.after(c);
+  v.ontimeupdate=()=>{
+    const m=Math.floor(v.currentTime/60),s=Math.floor(v.currentTime%60),ms=Math.floor((v.currentTime%1)*1000);
+    time.textContent=m+':'+(s<10?'0':'')+s+'.'+String(ms).padStart(3,'0');
+    if(!seeking&&v.duration){seekBar.value=Math.round((v.currentTime/v.duration)*1000)}
+  };
+  v.onprogress=v.onloadeddata=()=>{if(v.buffered.length&&v.duration){bufBar.style.width=(v.buffered.end(v.buffered.length-1)/v.duration*100)+'%'}};
+  v.onplay=()=>{playBtn.textContent='\u23F8'};v.onpause=()=>{playBtn.textContent='\u25B6'};
+  v.parentElement.addEventListener('keydown',e=>{
+    if(e.key==='ArrowLeft'){e.preventDefault();v.pause();v.currentTime=Math.max(0,v.currentTime-FT)}
+    if(e.key==='ArrowRight'){e.preventDefault();v.pause();v.currentTime+=FT}
+    if(e.key===' '){e.preventDefault();v.paused?v.play():v.pause()}
+  });
+  v.parentElement.setAttribute('tabindex','0');
+});
+</script></body></html>

.claude/skills/comfy-qa/scripts/qa-reproduce.ts

@@ -0,0 +1,253 @@
+#!/usr/bin/env tsx
+/**
+ * QA Reproduce Phase — Deterministic replay of research plan with narration
+ *
+ * Takes a reproduction plan from the research phase and replays it:
+ * 1. Execute each action deterministically (no AI decisions)
+ * 2. Capture a11y snapshot before/after each action
+ * 3. Gemini describes what visually changed (narration for humans)
+ * 4. Output: narration-log.json with full evidence chain
+ */
+
+import type { Page } from '@playwright/test'
+import { GoogleGenerativeAI } from '@google/generative-ai'
+import { mkdirSync, writeFileSync } from 'fs'
+
+import type { ActionResult } from './qa-record.js'
+
+// ── Types ──
+
+interface ReproductionStep {
+  action: Record<string, unknown> & { action: string }
+  expectedAssertion: string
+}
+
+interface NarrationEntry {
+  step: number
+  action: string
+  params: Record<string, unknown>
+  result: ActionResult
+  a11yBefore: unknown
+  a11yAfter: unknown
+  assertionExpected: string
+  assertionPassed: boolean
+  assertionActual: string
+  geminiNarration: string
+  timestampMs: number
+}
+
+export interface NarrationLog {
+  entries: NarrationEntry[]
+  allAssertionsPassed: boolean
+}
+
+interface ReproduceOptions {
+  page: Page
+  plan: ReproductionStep[]
+  geminiApiKey: string
+  outputDir: string
+}
+
+// ── A11y helpers ──
+
+interface A11yNode {
+  role: string
+  name: string
+  value?: string
+  checked?: boolean
+  disabled?: boolean
+  expanded?: boolean
+  children?: A11yNode[]
+}
+
+function searchA11y(node: A11yNode | null, selector: string): A11yNode | null {
+  if (!node) return null
+  const sel = selector.toLowerCase()
+  if (
+    node.name?.toLowerCase().includes(sel) ||
+    node.role?.toLowerCase().includes(sel)
+  ) {
+    return node
+  }
+  if (node.children) {
+    for (const child of node.children) {
+      const found = searchA11y(child, selector)
+      if (found) return found
+    }
+  }
+  return null
+}
+
+function summarizeA11y(node: A11yNode | null): string {
+  if (!node) return 'null'
+  const parts = [`role=${node.role}`, `name="${node.name}"`]
+  if (node.value !== undefined) parts.push(`value="${node.value}"`)
+  if (node.checked !== undefined) parts.push(`checked=${node.checked}`)
+  if (node.disabled) parts.push('disabled')
+  if (node.expanded !== undefined) parts.push(`expanded=${node.expanded}`)
+  return `{${parts.join(', ')}}`
+}
+
+// ── Subtitle overlay ──
+
+async function showSubtitle(page: Page, text: string, step: number) {
+  const encoded = encodeURIComponent(
+    text.slice(0, 120).replace(/'/g, "\\'").replace(/\n/g, ' ')
+  )
+  await page.addScriptTag({
+    content: `(function(){
+      var id='qa-subtitle';
+      var el=document.getElementById(id);
+      if(!el){
+        el=document.createElement('div');
+        el.id=id;
+        Object.assign(el.style,{position:'fixed',bottom:'32px',left:'50%',transform:'translateX(-50%)',zIndex:'2147483646',maxWidth:'90%',padding:'6px 14px',borderRadius:'6px',background:'rgba(0,0,0,0.8)',color:'rgba(255,255,255,0.95)',fontSize:'12px',fontFamily:'system-ui,sans-serif',fontWeight:'400',lineHeight:'1.4',pointerEvents:'none',textAlign:'center',whiteSpace:'normal'});
+        document.body.appendChild(el);
+      }
+      el.textContent='['+${step}+'] '+decodeURIComponent('${encoded}');
+    })()`
+  })
+}
+
+// ── Gemini visual narration ──
+
+async function geminiDescribe(
+  page: Page,
+  geminiApiKey: string,
+  focus: string
+): Promise<string> {
+  try {
+    const screenshot = await page.screenshot({ type: 'jpeg', quality: 70 })
+    const genAI = new GoogleGenerativeAI(geminiApiKey)
+    const model = genAI.getGenerativeModel({ model: 'gemini-3-flash-preview' })
+
+    const result = await model.generateContent([
+      {
+        text: `Describe in 1-2 sentences what you see on this ComfyUI screen. Focus on: ${focus}. Be factual — only describe what is visible.`
+      },
+      {
+        inlineData: {
+          mimeType: 'image/jpeg',
+          data: screenshot.toString('base64')
+        }
+      }
+    ])
+    return result.response.text().trim()
+  } catch (e) {
+    return `(Gemini narration failed: ${e instanceof Error ? e.message.slice(0, 50) : e})`
+  }
+}
+
+// ── Main reproduce function ──
+
+export async function runReproducePhase(
+  opts: ReproduceOptions
+): Promise<NarrationLog> {
+  const { page, plan, geminiApiKey, outputDir } = opts
+  const { executeAction } = await import('./qa-record.js')
+
+  const narrationDir = `${outputDir}/narration`
+  mkdirSync(narrationDir, { recursive: true })
+
+  const entries: NarrationEntry[] = []
+  const startMs = Date.now()
+
+  console.warn(`Reproduce phase: replaying ${plan.length} steps...`)
+
+  for (let i = 0; i < plan.length; i++) {
+    const step = plan[i]
+    const actionObj = step.action
+    const elapsed = Date.now() - startMs
+
+    // Show subtitle
+    await showSubtitle(page, `Step ${i + 1}: ${actionObj.action}`, i + 1)
+    console.warn(`  [${i + 1}/${plan.length}] ${actionObj.action}`)
+
+    // Capture a11y BEFORE
+    const a11yBefore = await page
+      .locator('body')
+      .ariaSnapshot({ timeout: 3000 })
+      .catch(() => null)
+
+    // Execute action
+    const result = await executeAction(
+      page,
+      actionObj as Parameters<typeof executeAction>[1],
+      outputDir
+    )
+    await new Promise((r) => setTimeout(r, 500))
+
+    // Capture a11y AFTER
+    const a11yAfter = await page
+      .locator('body')
+      .ariaSnapshot({ timeout: 3000 })
+      .catch(() => null)
+
+    // Check assertion
+    let assertionPassed = false
+    let assertionActual = ''
+    if (step.expectedAssertion) {
+      // Parse the expected assertion — e.g. "Settings dialog: visible" or "tab count: 2"
+      const parts = step.expectedAssertion.split(':').map((s) => s.trim())
+      const selectorName = parts[0]
+      const expectedState = parts.slice(1).join(':').trim()
+
+      const found = searchA11y(a11yAfter as A11yNode | null, selectorName)
+      assertionActual = found ? summarizeA11y(found) : 'NOT FOUND'
+
+      if (expectedState === 'visible' || expectedState === 'exists') {
+        assertionPassed = found !== null
+      } else if (expectedState === 'hidden' || expectedState === 'gone') {
+        assertionPassed = found === null
+      } else {
+        // Generic: check if the actual state contains the expected text
+        assertionPassed = assertionActual
+          .toLowerCase()
+          .includes(expectedState.toLowerCase())
+      }
+
+      console.warn(
+        `    Assertion: "${step.expectedAssertion}" → ${assertionPassed ? '✓ PASS' : '✗ FAIL'} (actual: ${assertionActual})`
+      )
+    }
+
+    // Gemini narration (visual description for humans)
+    const geminiNarration = await geminiDescribe(
+      page,
+      geminiApiKey,
+      `What changed after ${actionObj.action}?`
+    )
+
+    entries.push({
+      step: i + 1,
+      action: actionObj.action,
+      params: actionObj,
+      result,
+      a11yBefore,
+      a11yAfter,
+      assertionExpected: step.expectedAssertion,
+      assertionPassed,
+      assertionActual,
+      geminiNarration,
+      timestampMs: elapsed
+    })
+  }
+
+  // Final screenshot
+  await page.screenshot({ path: `${outputDir}/reproduce-final.png` })
+
+  const log: NarrationLog = {
+    entries,
+    allAssertionsPassed: entries.every((e) => e.assertionPassed)
+  }
+
+  writeFileSync(
+    `${narrationDir}/narration-log.json`,
+    JSON.stringify(log, null, 2)
+  )
+  console.warn(
+    `Reproduce phase complete: ${entries.filter((e) => e.assertionPassed).length}/${entries.length} assertions passed`
+  )
+
+  return log
+}

.claude/skills/comfy-qa/scripts/qa-video-review.test.ts

@@ -0,0 +1,150 @@
+import { describe, expect, it } from 'vitest'
+
+import {
+  extractPlatformFromArtifactDirName,
+  pickLatestVideosByPlatform,
+  selectVideoCandidateByFile
+} from './qa-video-review'
+
+describe('extractPlatformFromArtifactDirName', () => {
+  it('extracts and normalizes known qa artifact directory names', () => {
+    expect(
+      extractPlatformFromArtifactDirName('qa-report-Windows-22818315023')
+    ).toBe('windows')
+    expect(
+      extractPlatformFromArtifactDirName('qa-report-macOS-22818315023')
+    ).toBe('macos')
+    expect(
+      extractPlatformFromArtifactDirName('qa-report-Linux-22818315023')
+    ).toBe('linux')
+  })
+
+  it('falls back to slugifying unknown directory names', () => {
+    expect(extractPlatformFromArtifactDirName('custom platform run')).toBe(
+      'custom-platform-run'
+    )
+  })
+})
+
+describe('pickLatestVideosByPlatform', () => {
+  it('keeps only the latest candidate per platform', () => {
+    const selected = pickLatestVideosByPlatform([
+      {
+        platformName: 'windows',
+        videoPath: '/tmp/windows-old.mp4',
+        mtimeMs: 100
+      },
+      {
+        platformName: 'windows',
+        videoPath: '/tmp/windows-new.mp4',
+        mtimeMs: 200
+      },
+      {
+        platformName: 'linux',
+        videoPath: '/tmp/linux.mp4',
+        mtimeMs: 150
+      }
+    ])
+
+    expect(selected).toEqual([
+      {
+        platformName: 'linux',
+        videoPath: '/tmp/linux.mp4',
+        mtimeMs: 150
+      },
+      {
+        platformName: 'windows',
+        videoPath: '/tmp/windows-new.mp4',
+        mtimeMs: 200
+      }
+    ])
+  })
+})
+
+describe('selectVideoCandidateByFile', () => {
+  it('selects a single candidate by artifacts-relative path', () => {
+    const selected = selectVideoCandidateByFile(
+      [
+        {
+          platformName: 'windows',
+          videoPath: '/tmp/qa-artifacts/qa-report-Windows-1/qa-session.mp4',
+          mtimeMs: 100
+        },
+        {
+          platformName: 'linux',
+          videoPath: '/tmp/qa-artifacts/qa-report-Linux-1/qa-session.mp4',
+          mtimeMs: 200
+        }
+      ],
+      {
+        artifactsDir: '/tmp/qa-artifacts',
+        videoFile: 'qa-report-Linux-1/qa-session.mp4'
+      }
+    )
+
+    expect(selected).toEqual({
+      platformName: 'linux',
+      videoPath: '/tmp/qa-artifacts/qa-report-Linux-1/qa-session.mp4',
+      mtimeMs: 200
+    })
+  })
+
+  it('throws when basename matches multiple videos', () => {
+    expect(() =>
+      selectVideoCandidateByFile(
+        [
+          {
+            platformName: 'windows',
+            videoPath: '/tmp/qa-artifacts/qa-report-Windows-1/qa-session.mp4',
+            mtimeMs: 100
+          },
+          {
+            platformName: 'linux',
+            videoPath: '/tmp/qa-artifacts/qa-report-Linux-1/qa-session.mp4',
+            mtimeMs: 200
+          }
+        ],
+        {
+          artifactsDir: '/tmp/qa-artifacts',
+          videoFile: 'qa-session.mp4'
+        }
+      )
+    ).toThrow('matched 2 videos')
+  })
+
+  it('throws when there is no matching video', () => {
+    expect(() =>
+      selectVideoCandidateByFile(
+        [
+          {
+            platformName: 'windows',
+            videoPath: '/tmp/qa-artifacts/qa-report-Windows-1/qa-session.mp4',
+            mtimeMs: 100
+          }
+        ],
+        {
+          artifactsDir: '/tmp/qa-artifacts',
+          videoFile: 'qa-report-macOS-1/qa-session.mp4'
+        }
+      )
+    ).toThrow('No video matched')
+  })
+
+  it('throws when video file is missing', () => {
+    expect(() =>
+      selectVideoCandidateByFile(
+        [
+          {
+            platformName: 'windows',
+            videoPath: '/tmp/qa-artifacts/qa-report-Windows-1/qa-session.mp4',
+            mtimeMs: 100
+          }
+        ],
+        {
+          artifactsDir: '/tmp/qa-artifacts',
+          videoFile: '   '
+        }
+      )
+    ).toThrow('--video-file is required')
+  })
+})

.claude/skills/comfy-qa/scripts/qa-video-review.ts

@@ -0,0 +1,771 @@
+#!/usr/bin/env tsx
+import { mkdir, readFile, stat, writeFile } from 'node:fs/promises'
+import { basename, dirname, extname, relative, resolve } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+import { GoogleGenerativeAI } from '@google/generative-ai'
+import { globSync } from 'glob'
+
+interface CliOptions {
+  artifactsDir: string
+  videoFile: string
+  beforeVideo: string
+  outputDir: string
+  model: string
+  requestTimeoutMs: number
+  dryRun: boolean
+  prContext: string
+  targetUrl: string
+  passLabel: string
+}
+
+interface VideoCandidate {
+  platformName: string
+  videoPath: string
+  mtimeMs: number
+}
+
+const DEFAULT_OPTIONS: CliOptions = {
+  artifactsDir: './tmp/qa-artifacts',
+  videoFile: '',
+  beforeVideo: '',
+  outputDir: './tmp',
+  model: 'gemini-3-flash-preview',
+  requestTimeoutMs: 300_000,
+  dryRun: false,
+  prContext: '',
+  targetUrl: '',
+  passLabel: ''
+}
+
+const USAGE = `Usage:
+  pnpm exec tsx scripts/qa-video-review.ts [options]
+
+Options:
+  --artifacts-dir <path>        Artifacts root directory
+                                 (default: ./tmp/qa-artifacts)
+  --video-file <name-or-path>   Video file to analyze (required)
+                                 (supports basename or relative/absolute path)
+  --before-video <path>         Before video (main branch) for comparison
+                                 When provided, sends both videos to Gemini
+                                 for comparative before/after analysis
+  --output-dir <path>           Output directory for markdown reports
+                                 (default: ./tmp)
+  --model <name>                Gemini model
+                                 (default: gemini-3-flash-preview)
+  --request-timeout-ms <n>      Request timeout in milliseconds
+                                 (default: 300000)
+  --pr-context <file>           File with PR context (title, body, diff)
+                                 for PR-aware review
+  --target-url <url>            Issue or PR URL to include in the report
+  --pass-label <label>          Label for multi-pass reports (e.g. pass1)
+                                 Output becomes {platform}-{label}-qa-video-report.md
+  --dry-run                     Discover videos and output targets only
+  --help                        Show this help text
+
+Environment:
+  GEMINI_API_KEY                Required unless --dry-run
+`
+
+function parsePositiveInteger(rawValue: string, flagName: string): number {
+  const parsedValue = Number.parseInt(rawValue, 10)
+  if (!Number.isInteger(parsedValue) || parsedValue <= 0) {
+    throw new Error(`Invalid value for ${flagName}: "${rawValue}"`)
+  }
+  return parsedValue
+}
+
+function parseCliOptions(args: string[]): CliOptions {
+  const options: CliOptions = { ...DEFAULT_OPTIONS }
+
+  for (let index = 0; index < args.length; index += 1) {
+    const argument = args[index]
+    const nextValue = args[index + 1]
+    const requireValue = (flagName: string): string => {
+      if (!nextValue || nextValue.startsWith('--')) {
+        throw new Error(`Missing value for ${flagName}`)
+      }
+      index += 1
+      return nextValue
+    }
+
+    if (argument === '--help') {
+      process.stdout.write(USAGE)
+      process.exit(0)
+    }
+
+    if (argument === '--artifacts-dir') {
+      options.artifactsDir = requireValue(argument)
+      continue
+    }
+
+    if (argument === '--video-file') {
+      options.videoFile = requireValue(argument)
+      continue
+    }
+
+    if (argument === '--output-dir') {
+      options.outputDir = requireValue(argument)
+      continue
+    }
+
+    if (argument === '--model') {
+      options.model = requireValue(argument)
+      continue
+    }
+
+    if (argument === '--request-timeout-ms') {
+      options.requestTimeoutMs = parsePositiveInteger(
+        requireValue(argument),
+        argument
+      )
+      continue
+    }
+
+    if (argument === '--before-video') {
+      options.beforeVideo = requireValue(argument)
+      continue
+    }
+
+    if (argument === '--pr-context') {
+      options.prContext = requireValue(argument)
+      continue
+    }
+
+    if (argument === '--target-url') {
+      options.targetUrl = requireValue(argument)
+      continue
+    }
+
+    if (argument === '--pass-label') {
+      options.passLabel = requireValue(argument)
+      continue
+    }
+
+    if (argument === '--dry-run') {
+      options.dryRun = true
+      continue
+    }
+
+    throw new Error(`Unknown argument: ${argument}`)
+  }
+
+  return options
+}
+
+function normalizePlatformName(value: string): string {
+  const slug = value
+    .trim()
+    .toLowerCase()
+    .replace(/[^a-z0-9]+/g, '-')
+    .replace(/^-+|-+$/g, '')
+
+  return slug.length > 0 ? slug : 'unknown-platform'
+}
+
+export function extractPlatformFromArtifactDirName(dirName: string): string {
+  const matchedValue = dirName.match(/^qa-report-(.+?)(?:-\d+)?$/i)?.[1]
+  return normalizePlatformName(matchedValue ?? dirName)
+}
+
+function extractPlatformFromVideoPath(videoPath: string): string {
+  const artifactDirName = basename(dirname(videoPath))
+  return extractPlatformFromArtifactDirName(artifactDirName)
+}
+
+export function pickLatestVideosByPlatform(
+  candidates: VideoCandidate[]
+): VideoCandidate[] {
+  const latestByPlatform = new Map<string, VideoCandidate>()
+
+  for (const candidate of candidates) {
+    const current = latestByPlatform.get(candidate.platformName)
+    if (!current || candidate.mtimeMs > current.mtimeMs) {
+      latestByPlatform.set(candidate.platformName, candidate)
+    }
+  }
+
+  return [...latestByPlatform.values()].sort((a, b) =>
+    a.platformName.localeCompare(b.platformName)
+  )
+}
+
+function toProjectRelativePath(targetPath: string): string {
+  const relativePath = relative(process.cwd(), targetPath)
+  if (relativePath.startsWith('.')) {
+    return relativePath
+  }
+  return `./${relativePath}`
+}
+
+function errorToString(error: unknown): string {
+  return error instanceof Error ? error.message : String(error)
+}
+
+function normalizePathForMatch(value: string): string {
+  return value.replaceAll('\\', '/').replace(/^\.\/+/, '')
+}
+
+export function selectVideoCandidateByFile(
+  candidates: VideoCandidate[],
+  options: { artifactsDir: string; videoFile: string }
+): VideoCandidate {
+  const requestedValue = options.videoFile.trim()
+  if (requestedValue.length === 0) {
+    throw new Error('--video-file is required')
+  }
+
+  const artifactsRoot = resolve(options.artifactsDir)
+  const requestedAbsolutePath = resolve(requestedValue)
+  const requestedPathKey = normalizePathForMatch(requestedValue)
+
+  const matches = candidates.filter((candidate) => {
+    const candidateAbsolutePath = resolve(candidate.videoPath)
+    if (candidateAbsolutePath === requestedAbsolutePath) {
+      return true
+    }
+
+    const candidateBaseName = basename(candidate.videoPath)
+    if (candidateBaseName === requestedValue) {
+      return true
+    }
+
+    const relativeToCwd = normalizePathForMatch(
+      relative(process.cwd(), candidateAbsolutePath)
+    )
+    if (relativeToCwd === requestedPathKey) {
+      return true
+    }
+
+    const relativeToArtifacts = normalizePathForMatch(
+      relative(artifactsRoot, candidateAbsolutePath)
+    )
+    return relativeToArtifacts === requestedPathKey
+  })
+
+  if (matches.length === 1) {
+    return matches[0]
+  }
+
+  if (matches.length === 0) {
+    const availableVideos = candidates.map((candidate) =>
+      toProjectRelativePath(candidate.videoPath)
+    )
+    throw new Error(
+      [
+        `No video matched --video-file "${options.videoFile}".`,
+        'Available videos:',
+        ...availableVideos.map((videoPath) => `- ${videoPath}`)
+      ].join('\n')
+    )
+  }
+
+  throw new Error(
+    [
+      `--video-file "${options.videoFile}" matched ${matches.length} videos.`,
+      'Please pass a more specific path.',
+      ...matches.map((match) => `- ${toProjectRelativePath(match.videoPath)}`)
+    ].join('\n')
+  )
+}
+
+async function collectVideoCandidates(
+  artifactsDir: string
+): Promise<VideoCandidate[]> {
+  const absoluteArtifactsDir = resolve(artifactsDir)
+  const videoPaths = globSync('**/qa-session{,-[0-9]}.mp4', {
+    cwd: absoluteArtifactsDir,
+    absolute: true,
+    nodir: true
+  }).sort()
+
+  const candidates = await Promise.all(
+    videoPaths.map(async (videoPath) => {
+      const videoStat = await stat(videoPath)
+      return {
+        platformName: extractPlatformFromVideoPath(videoPath),
+        videoPath,
+        mtimeMs: videoStat.mtimeMs
+      }
+    })
+  )
+
+  return candidates
+}
+
+function getMimeType(filePath: string): string {
+  const ext = extname(filePath).toLowerCase()
+  const mimeMap: Record<string, string> = {
+    '.mp4': 'video/mp4',
+    '.webm': 'video/webm',
+    '.mov': 'video/quicktime',
+    '.avi': 'video/x-msvideo',
+    '.mkv': 'video/x-matroska',
+    '.m4v': 'video/mp4'
+  }
+  return mimeMap[ext] || 'video/mp4'
+}
+
+function buildReviewPrompt(options: {
+  platformName: string
+  videoPath: string
+  prContext: string
+  isComparative: boolean
+}): string {
+  const { platformName, videoPath, prContext, isComparative } = options
+
+  if (isComparative) {
+    return buildComparativePrompt(platformName, videoPath, prContext)
+  }
+
+  return buildSingleVideoPrompt(platformName, videoPath, prContext)
+}
+
+function buildComparativePrompt(
+  platformName: string,
+  videoPath: string,
+  prContext: string
+): string {
+  const lines = [
+    'You are a senior QA engineer performing a BEFORE/AFTER comparison review.',
+    '',
+    'You are given TWO videos:',
+    '- **Video 1 (BEFORE)**: The main branch BEFORE the PR. This shows the OLD behavior.',
+    '- **Video 2 (AFTER)**: The PR branch AFTER the changes. This shows the NEW behavior.',
+    '',
+    'Both videos show the same test steps executed on different code versions.',
+    ''
+  ]
+
+  if (prContext) {
+    lines.push('## PR Context', prContext, '')
+  }
+
+  lines.push(
+    '## Your Task',
+    `Platform: "${platformName}". After video: ${toProjectRelativePath(videoPath)}.`,
+    '',
+    '1. **BEFORE video**: Does it demonstrate the old behavior or bug that the PR aims to fix?',
+    '   Describe what you observe — this establishes the baseline.',
+    '2. **AFTER video**: Does it prove the PR fix works? Is the intended new behavior visible?',
+    '3. **Comparison**: What specifically changed between before and after?',
+    '4. **Regressions**: Did the PR introduce any new problems visible in the AFTER video',
+    '   that were NOT present in the BEFORE video?',
+    '',
+    'Note: Brief black frames during page transitions are NORMAL.',
+    'Note: Small cyan/purple dashed labels prefixed with "QA:" are annotations placed by the automated test script — they are NOT part of the application UI. Do not treat them as bugs or evidence.',
+    'Report only concrete, visible differences. Avoid speculation.',
+    '',
+    'Return markdown with these sections exactly:',
+    '## Summary',
+    '(What the PR changes, whether BEFORE confirms the old behavior, whether AFTER proves the fix)',
+    '',
+    '## Behavior Changes',
+    'Summarize ALL behavioral differences as a markdown TABLE:',
+    '| Behavior | Before (main) | After (PR) | Verdict |',
+    '',
+    '- **Behavior**: short name for the behavior (e.g. "Save shortcut label", "Menu hover style")',
+    '- **Before (main)**: how it works/looks in the BEFORE video',
+    '- **After (PR)**: how it works/looks in the AFTER video',
+    '- **Verdict**: `Fixed`, `Improved`, `Changed`, `Regression`, or `No Change`',
+    '',
+    'One row per distinct behavior. Include both changed AND unchanged key behaviors',
+    'that were tested, so reviewers can confirm nothing was missed.',
+    '',
+    '## Timeline Comparison',
+    'Present a chronological frame-by-frame comparison as a markdown TABLE:',
+    '| Time | Type | Severity | Before (main) | After (PR) |',
+    '',
+    '- **Time**: timestamp or range from the videos (e.g. `0:05-0:08`)',
+    '- **Type**: category such as `Visual`, `Behavior`, `Layout`, `Text`, `Animation`, `Menu`, `State`',
+    '- **Severity**: `None` (neutral change), `Fixed` (bug resolved), `Regression`, `Minor`, `Major`',
+    '- **Before (main)**: what is observed in the BEFORE video at that time',
+    '- **After (PR)**: what is observed in the AFTER video at that time',
+    '',
+    'Include one row per distinct observable difference. If behavior is identical at a timestamp,',
+    'omit that row. Focus on meaningful differences, not narrating every frame.',
+    '',
+    '## Confirmed Issues',
+    'For each issue, use this exact format:',
+    '',
+    '### [Short issue title]',
+    '`SEVERITY` `TIMESTAMP` `Confidence: LEVEL`',
+    '',
+    '[Description — specify whether it appears in BEFORE, AFTER, or both]',
+    '',
+    '**Evidence:** [What you observed at the given timestamp in which video]',
+    '',
+    '**Suggested Fix:** [Actionable recommendation]',
+    '',
+    '---',
+    '',
+    '## Possible Issues (Needs Human Verification)',
+    '## Overall Risk',
+    '(Assess whether the PR achieves its goal based on the before/after comparison)',
+    '',
+    '## Verdict',
+    'End your report with this EXACT JSON block (no markdown fence):',
+    '{"verdict": "REPRODUCED" | "NOT_REPRODUCIBLE" | "INCONCLUSIVE", "risk": "low" | "medium" | "high", "confidence": "high" | "medium" | "low"}',
+    '- REPRODUCED: the before video confirms the old behavior and the after video shows the fix working',
+    '- NOT_REPRODUCIBLE: the before video does not show the reported bug',
+    '- INCONCLUSIVE: the videos do not adequately demonstrate the behavior change'
+  )
+
+  return lines.filter(Boolean).join('\n')
+}
+
+function buildSingleVideoPrompt(
+  platformName: string,
+  videoPath: string,
+  prContext: string
+): string {
+  const lines = [
+    'You are a senior QA engineer reviewing a UI test session recording.',
+    '',
+    '## ANTI-HALLUCINATION RULES (READ FIRST)',
+    '- Describe ONLY what you can directly observe in the video frames',
+    '- NEVER infer or assume what "must have happened" between frames',
+    '- If a step is not visible in the video, say "NOT SHOWN" — do not guess',
+    '- Your job is to be a CAMERA — report facts, not interpretations',
+    ''
+  ]
+
+  const isIssueContext =
+    prContext &&
+    /^### Issue #|^Title:.*\bbug\b|^This video attempts to reproduce/im.test(
+      prContext
+    )
+
+  if (prContext) {
+    lines.push(
+      '## Phase 1: Blind Observation (describe what you SEE and HEAR)',
+      'First, describe every UI interaction chronologically WITHOUT knowing the expected outcome:',
+      '- What elements does the user click/hover/type?',
+      '- What dialogs/menus open and close?',
+      '- What keyboard indicators appear? (look for subtitle overlays)',
+      '- What is the BEFORE state and AFTER state of each action?',
+      '- **Audio**: Does the video have a TTS narration audio track? If yes, transcribe what the voice says. This narration describes the bug being reproduced.',
+      '',
+      '## Phase 2: Compare against expected behavior',
+      'Now compare your observations against the context below.',
+      'Only claim a match if your Phase 1 observations EXPLICITLY support it.',
+      ''
+    )
+
+    if (isIssueContext) {
+      lines.push(
+        '## Issue Context',
+        prContext,
+        '',
+        '## Comparison Questions',
+        '1. Did the video perform the reproduction steps described in the issue?',
+        '2. Did your Phase 1 observations show the reported bug behavior?',
+        '3. If the steps were not performed or the bug was not visible, say INCONCLUSIVE.',
+        ''
+      )
+    } else {
+      lines.push(
+        '## PR Context',
+        prContext,
+        '',
+        '## Comparison Questions',
+        '1. Did the video test the specific behavior the PR changes?',
+        '2. Did your Phase 1 observations show the expected before/after difference?',
+        '3. If the test was incomplete or inconclusive, say so honestly.',
+        ''
+      )
+    }
+  }
+
+  lines.push(
+    `Review this QA session video for platform "${platformName}".`,
+    `Source video: ${toProjectRelativePath(videoPath)}.`,
+    'The video shows the full test session — analyze it chronologically.',
+    'Focus on UI regressions, broken states, visual glitches, unreadable text, missing labels/i18n, and clear workflow failures.',
+    'Note: Brief black frames during page transitions are NORMAL and should NOT be reported as issues.',
+    'Note: Small cyan/purple dashed labels prefixed with "QA:" are annotations placed by the automated test script — they are NOT part of the application UI. Do not treat them as bugs or evidence.',
+    'Report only concrete, visible problems and avoid speculation.',
+    'If confidence is low, mark it explicitly.',
+    '',
+    'Return markdown with these sections exactly:',
+    '## Summary',
+    isIssueContext
+      ? '(Explain what bug was reported and whether the video confirms it is reproducible)'
+      : prContext
+        ? '(Explain what the PR intended and whether the video confirms it works)'
+        : '',
+    '## Confirmed Issues',
+    'For each confirmed issue, use this exact format (one block per issue):',
+    '',
+    '### [Short issue title]',
+    '`HIGH` `01:03` `Confidence: High`',
+    '',
+    '[Description of the issue — what went wrong and what was expected]',
+    '',
+    '**Evidence:** [What you observed in the video at the given timestamp]',
+    '',
+    '**Suggested Fix:** [Actionable recommendation]',
+    '',
+    '---',
+    '',
+    'The first line after the heading MUST be exactly three backtick-wrapped labels:',
+    '`SEVERITY` `TIMESTAMP` `Confidence: LEVEL`',
+    'Do NOT use a table for issues — use the block format above.',
+    '## Possible Issues (Needs Human Verification)',
+    '## Overall Risk',
+    '',
+    '## Narration',
+    'If the video contains a TTS audio narration track, transcribe it here.',
+    'If there is no audio or the video is silent, write "No narration detected."',
+    '',
+    '## Verdict',
+    'End your report with this EXACT JSON block (no markdown fence):',
+    '{"verdict": "REPRODUCED" | "NOT_REPRODUCIBLE" | "INCONCLUSIVE", "risk": "low" | "medium" | "high" | null, "confidence": "high" | "medium" | "low", "narrationDetected": true | false}',
+    '- REPRODUCED: the bug/behavior is clearly visible in the video',
+    '- NOT_REPRODUCIBLE: the steps were performed correctly but the bug was not observed',
+    '- INCONCLUSIVE: the reproduction steps were not performed or the video is insufficient',
+    '- narrationDetected: true if you heard TTS voice narration in the video, false if silent'
+  )
+
+  return lines.filter(Boolean).join('\n')
+}
+
+const MAX_VIDEO_BYTES = 100 * 1024 * 1024
+
+async function readVideoFile(videoPath: string): Promise<Buffer> {
+  const fileStat = await stat(videoPath)
+  if (fileStat.size > MAX_VIDEO_BYTES) {
+    throw new Error(
+      `Video ${basename(videoPath)} is ${formatBytes(fileStat.size)}, exceeds ${formatBytes(MAX_VIDEO_BYTES)} limit`
+    )
+  }
+  return readFile(videoPath)
+}
+
+async function requestGeminiReview(options: {
+  apiKey: string
+  model: string
+  platformName: string
+  videoPath: string
+  beforeVideoPath: string
+  timeoutMs: number
+  prContext: string
+}): Promise<string> {
+  const genAI = new GoogleGenerativeAI(options.apiKey)
+  const model = genAI.getGenerativeModel({ model: options.model })
+
+  const isComparative = options.beforeVideoPath.length > 0
+  const prompt = buildReviewPrompt({
+    platformName: options.platformName,
+    videoPath: options.videoPath,
+    prContext: options.prContext,
+    isComparative
+  })
+
+  const parts: Array<
+    { text: string } | { inlineData: { mimeType: string; data: string } }
+  > = [{ text: prompt }]
+
+  if (isComparative) {
+    const beforeBuffer = await readVideoFile(options.beforeVideoPath)
+    parts.push(
+      { text: 'Video 1 — BEFORE (main branch):' },
+      {
+        inlineData: {
+          mimeType: getMimeType(options.beforeVideoPath),
+          data: beforeBuffer.toString('base64')
+        }
+      }
+    )
+  }
+
+  const afterBuffer = await readVideoFile(options.videoPath)
+  if (isComparative) {
+    parts.push({ text: 'Video 2 — AFTER (PR branch):' })
+  }
+  parts.push({
+    inlineData: {
+      mimeType: getMimeType(options.videoPath),
+      data: afterBuffer.toString('base64')
+    }
+  })
+
+  const result = await model.generateContent(parts, {
+    timeout: options.timeoutMs
+  })
+  const response = result.response
+  const text = response.text()
+
+  if (!text || text.trim().length === 0) {
+    throw new Error('Gemini API returned no output text')
+  }
+
+  return text.trim()
+}
+
+function formatBytes(bytes: number): string {
+  if (bytes < 1024) return `${bytes} B`
+  if (bytes < 1024 * 1024) return `${(bytes / 1024).toFixed(1)} KB`
+  return `${(bytes / (1024 * 1024)).toFixed(1)} MB`
+}
+
+function buildReportMarkdown(input: {
+  platformName: string
+  model: string
+  videoPath: string
+  videoSizeBytes: number
+  beforeVideoPath?: string
+  beforeVideoSizeBytes?: number
+  reviewText: string
+  targetUrl?: string
+}): string {
+  const headerLines = [
+    `# ${input.platformName} QA Video Report`,
+    '',
+    `- Generated at: ${new Date().toISOString()}`,
+    `- Model: \`${input.model}\``
+  ]
+
+  if (input.targetUrl) {
+    headerLines.push(`- Target: ${input.targetUrl}`)
+  }
+
+  if (input.beforeVideoPath) {
+    headerLines.push(
+      `- Before video: \`${toProjectRelativePath(input.beforeVideoPath)}\` (${formatBytes(input.beforeVideoSizeBytes ?? 0)})`,
+      `- After video: \`${toProjectRelativePath(input.videoPath)}\` (${formatBytes(input.videoSizeBytes)})`,
+      '- Mode: **Comparative (before/after)**'
+    )
+  } else {
+    headerLines.push(
+      `- Source video: \`${toProjectRelativePath(input.videoPath)}\``,
+      `- Video size: ${formatBytes(input.videoSizeBytes)}`
+    )
+  }
+
+  headerLines.push('', '## AI Review', '')
+  return `${headerLines.join('\n')}${input.reviewText.trim()}\n`
+}
+
+async function reviewVideo(
+  video: VideoCandidate,
+  options: CliOptions,
+  apiKey: string
+): Promise<void> {
+  let prContext = ''
+  if (options.prContext) {
+    try {
+      prContext = await readFile(options.prContext, 'utf-8')
+      process.stdout.write(
+        `[${video.platformName}] Loaded PR context from ${options.prContext}\n`
+      )
+    } catch {
+      process.stdout.write(
+        `[${video.platformName}] Warning: Could not read PR context file ${options.prContext}\n`
+      )
+    }
+  }
+
+  const beforeVideoPath = options.beforeVideo
+    ? resolve(options.beforeVideo)
+    : ''
+
+  if (beforeVideoPath) {
+    const beforeStat = await stat(beforeVideoPath)
+    process.stdout.write(
+      `[${video.platformName}] Before video: ${toProjectRelativePath(beforeVideoPath)} (${formatBytes(beforeStat.size)})\n`
+    )
+  }
+
+  process.stdout.write(
+    `[${video.platformName}] Sending ${beforeVideoPath ? '2 videos (comparative)' : 'video'} to ${options.model}\n`
+  )
+
+  const reviewText = await requestGeminiReview({
+    apiKey,
+    model: options.model,
+    platformName: video.platformName,
+    videoPath: video.videoPath,
+    beforeVideoPath,
+    timeoutMs: options.requestTimeoutMs,
+    prContext
+  })
+
+  const videoStat = await stat(video.videoPath)
+  const passSegment = options.passLabel ? `-${options.passLabel}` : ''
+  const outputPath = resolve(
+    options.outputDir,
+    `${video.platformName}${passSegment}-qa-video-report.md`
+  )
+
+  const reportInput: Parameters<typeof buildReportMarkdown>[0] = {
+    platformName: video.platformName,
+    model: options.model,
+    videoPath: video.videoPath,
+    videoSizeBytes: videoStat.size,
+    reviewText,
+    targetUrl: options.targetUrl || undefined
+  }
+
+  if (beforeVideoPath) {
+    const beforeStat = await stat(beforeVideoPath)
+    reportInput.beforeVideoPath = beforeVideoPath
+    reportInput.beforeVideoSizeBytes = beforeStat.size
+  }
+
+  const reportMarkdown = buildReportMarkdown(reportInput)
+
+  await mkdir(dirname(outputPath), { recursive: true })
+  await writeFile(outputPath, reportMarkdown, 'utf-8')
+
+  process.stdout.write(
+    `[${video.platformName}] Wrote ${toProjectRelativePath(outputPath)}\n`
+  )
+}
+
+function isExecutedAsScript(metaUrl: string): boolean {
+  const modulePath = fileURLToPath(metaUrl)
+  const scriptPath = process.argv[1] ? resolve(process.argv[1]) : ''
+  return modulePath === scriptPath
+}
+
+async function main(): Promise<void> {
+  const options = parseCliOptions(process.argv.slice(2))
+  const candidates = await collectVideoCandidates(options.artifactsDir)
+
+  if (candidates.length === 0) {
+    process.stdout.write(
+      `No qa-session.mp4 files found under ${toProjectRelativePath(resolve(options.artifactsDir))}\n`
+    )
+    return
+  }
+
+  const selectedVideo = selectVideoCandidateByFile(candidates, {
+    artifactsDir: options.artifactsDir,
+    videoFile: options.videoFile
+  })
+
+  process.stdout.write(
+    `Selected ${selectedVideo.platformName}: ${toProjectRelativePath(selectedVideo.videoPath)}\n`
+  )
+
+  if (options.dryRun) {
+    process.stdout.write('\nDry run mode enabled, no API calls were made.\n')
+    return
+  }
+
+  const apiKey = process.env.GEMINI_API_KEY
+  if (!apiKey) {
+    throw new Error('GEMINI_API_KEY is required unless --dry-run is set')
+  }
+
+  await reviewVideo(selectedVideo, options, apiKey)
+}
+
+if (isExecutedAsScript(import.meta.url)) {
+  void main().catch((error: unknown) => {
+    const message = errorToString(error)
+    process.stderr.write(`qa-video-review failed: ${message}\n`)
+    process.exit(1)
+  })
+}

.claude/skills/comfy-qa/scripts/qa.ts

@@ -0,0 +1,513 @@
+#!/usr/bin/env tsx
+/**
+ * QA CLI — simplified entry point for local & CI QA runs
+ *
+ * Usage:
+ *   pnpm qa 10253                         # auto-detects issue vs PR
+ *   pnpm qa https://github.com/.../pull/10270
+ *   pnpm qa 10270 -t base                 # test PR base (reproduce bug)
+ *   pnpm qa 10270 -t both                 # test base + head
+ *   pnpm qa --uncommitted                 # test local uncommitted changes
+ *
+ * Automatically loads .env.local / .env for GEMINI_API_KEY, ANTHROPIC_API_KEY.
+ * Results are written to .comfy-qa/<number>/ by default.
+ */
+
+import { parseArgs } from 'node:util'
+import { config } from 'dotenv'
+import { existsSync, mkdirSync, writeFileSync } from 'fs'
+import { dirname, resolve } from 'path'
+import { execSync, spawn, spawnSync } from 'child_process'
+import { fileURLToPath } from 'url'
+
+// ── Constants ──
+
+const SCRIPT_DIR = dirname(fileURLToPath(import.meta.url))
+const RECORD_SCRIPT = resolve(SCRIPT_DIR, 'qa-record.ts')
+const DEFAULT_REPO = 'Comfy-Org/ComfyUI_frontend'
+const VALID_TARGETS = ['head', 'base', 'both'] as const
+const CLOUD_FALLBACK_URL = 'https://testcloud.comfy.org/'
+type PrTarget = (typeof VALID_TARGETS)[number]
+type TargetType = 'issue' | 'pr'
+
+// ── Load .env.local / .env ──
+
+for (const f of ['.env.local', '.env']) {
+  if (existsSync(f)) {
+    config({ path: f })
+    break
+  }
+}
+
+// ── Parse CLI ──
+
+const { values, positionals } = tryParseArgs()
+
+if (values.help) {
+  printUsage()
+  process.exit(0)
+}
+
+const serverUrl =
+  values.url || process.env.DEV_SERVER_COMFYUI_URL || 'http://127.0.0.1:8188'
+
+const prTarget = values.target as PrTarget
+if (!VALID_TARGETS.includes(prTarget)) {
+  console.error(
+    `Invalid --target "${prTarget}". Must be one of: ${VALID_TARGETS.join(', ')}`
+  )
+  process.exit(1)
+}
+
+// ── Ensure server is reachable (may fall back to cloud) ──
+
+const resolvedServerUrl = await ensureServer(serverUrl)
+
+// ── Dispatch by mode ──
+
+if (values.uncommitted) {
+  runUncommitted()
+} else {
+  const input = positionals[0]
+  if (!input) {
+    printUsage()
+    process.exit(1)
+  }
+  runTarget(input)
+}
+
+// ── Mode: uncommitted changes ──
+
+function runUncommitted(): never {
+  const diff = shell('git diff && git diff --staged')
+  if (!diff.trim()) {
+    console.error('No uncommitted changes found')
+    process.exit(1)
+  }
+
+  const outputDir = resolveOutputDir('.comfy-qa/local')
+  const diffFile = writeTmpFile(outputDir, 'uncommitted.diff', diff)
+
+  logHeader({ label: 'uncommitted changes', outputDir })
+  const code = runQaRecord('after', diffFile, outputDir)
+  exit(code, outputDir)
+}
+
+// ── Mode: issue or PR by number/URL ──
+
+function runTarget(input: string): never {
+  const { targetType, number, repo } = resolveTarget(input)
+  const outputDir = resolveOutputDir(`.comfy-qa/${number}`)
+
+  logHeader({
+    label: `${targetType} #${number} (${repo})`,
+    outputDir,
+    extra: targetType === 'pr' ? `Target:    ${prTarget}` : undefined
+  })
+
+  const diffFile =
+    targetType === 'issue'
+      ? fetchIssue(number, repo, outputDir)
+      : fetchPR(number, repo, outputDir)
+
+  let exitCode: number
+  if (targetType === 'issue') {
+    exitCode = runQaRecord('reproduce', diffFile, outputDir)
+  } else if (prTarget === 'both') {
+    exitCode = runPrBoth(diffFile, outputDir)
+  } else if (prTarget === 'base') {
+    exitCode = runQaRecord('before', diffFile, outputDir)
+  } else {
+    exitCode = runQaRecord('after', diffFile, outputDir)
+  }
+
+  exit(exitCode, outputDir)
+}
+
+// ── PR both phases ──
+
+function runPrBoth(diffFile: string, outputDir: string): number {
+  console.warn('\n=== Phase 1: Reproduce bug on base ===')
+  const baseDir = resolve(outputDir, 'base')
+  mkdirSync(baseDir, { recursive: true })
+  const baseCode = runQaRecord('before', diffFile, baseDir)
+  if (baseCode !== 0) {
+    console.warn('Base phase failed, continuing to head...')
+  }
+
+  console.warn('\n=== Phase 2: Demonstrate fix on head ===')
+  const headDir = resolve(outputDir, 'head')
+  mkdirSync(headDir, { recursive: true })
+  return runQaRecord('after', diffFile, headDir)
+}
+
+// ── Target resolution ──
+
+function resolveTarget(input: string): {
+  targetType: TargetType
+  number: string
+  repo: string
+} {
+  const urlMatch = input.match(
+    /github\.com\/([^/]+\/[^/]+)\/(issues|pull)\/(\d+)/
+  )
+
+  if (urlMatch) {
+    return {
+      repo: urlMatch[1],
+      targetType: urlMatch[2] === 'pull' ? 'pr' : 'issue',
+      number: urlMatch[3]
+    }
+  }
+
+  if (/^\d+$/.test(input)) {
+    return {
+      repo: DEFAULT_REPO,
+      targetType: detectType(input, DEFAULT_REPO),
+      number: input
+    }
+  }
+
+  console.error(`Cannot parse target: ${input}`)
+  console.error('Expected a GitHub URL or issue/PR number')
+  printUsage()
+  process.exit(1)
+}
+
+function detectType(number: string, repo: string): TargetType {
+  try {
+    const result = execSync(
+      `gh api repos/${repo}/issues/${number} --jq 'has("pull_request")'`,
+      { encoding: 'utf-8', timeout: 15000, stdio: ['pipe', 'pipe', 'pipe'] }
+    )
+    return result.trim() === 'true' ? 'pr' : 'issue'
+  } catch {
+    return 'issue'
+  }
+}
+
+// ── Data fetching ──
+
+function fetchIssue(number: string, repo: string, outputDir: string): string {
+  console.warn(`Fetching issue #${number}...`)
+  const body = shell(
+    `gh issue view ${number} --repo ${repo} --json title,body,labels --jq '"Title: " + .title + "\\n\\nLabels: " + ([.labels[].name] | join(", ")) + "\\n\\n" + .body'`
+  )
+  return writeTmpFile(outputDir, `issue-${number}.txt`, body)
+}
+
+function fetchPR(number: string, repo: string, outputDir: string): string {
+  console.warn(`Fetching PR #${number}...`)
+  const prJson = shell(
+    `gh pr view ${number} --repo ${repo} --json title,body,baseRefName,headRefName,baseRefOid,headRefOid`
+  )
+  const pr = JSON.parse(prJson) as {
+    title: string
+    body: string
+    baseRefName: string
+    headRefName: string
+    baseRefOid: string
+    headRefOid: string
+  }
+
+  console.warn(`  Base: ${pr.baseRefName} (${pr.baseRefOid.slice(0, 8)})`)
+  console.warn(`  Head: ${pr.headRefName} (${pr.headRefOid.slice(0, 8)})`)
+
+  let diff = ''
+  try {
+    diff = shell(`gh pr diff ${number} --repo ${repo}`)
+  } catch {
+    console.warn('Could not fetch PR diff')
+  }
+
+  writeTmpFile(
+    outputDir,
+    'refs.json',
+    JSON.stringify(
+      {
+        base: { ref: pr.baseRefName, sha: pr.baseRefOid },
+        head: { ref: pr.headRefName, sha: pr.headRefOid }
+      },
+      null,
+      2
+    )
+  )
+
+  return writeTmpFile(
+    outputDir,
+    `pr-${number}.txt`,
+    `Title: ${pr.title}\n\n${pr.body}\n\n--- DIFF ---\n\n${diff}`
+  )
+}
+
+// ── QA record runner ──
+
+function runQaRecord(
+  mode: string,
+  diffFile: string,
+  outputDir: string
+): number {
+  console.warn(`\nStarting QA ${mode} mode...\n`)
+  const r = spawnSync(
+    'pnpm',
+    [
+      'exec',
+      'tsx',
+      RECORD_SCRIPT,
+      '--mode',
+      mode,
+      '--diff',
+      diffFile,
+      '--output-dir',
+      outputDir,
+      '--url',
+      resolvedServerUrl
+    ],
+    { stdio: 'inherit', env: process.env }
+  )
+  return r.status ?? 1
+}
+
+// ── Server management ──
+
+async function ensureServer(url: string): Promise<string> {
+  if (await isReachable(url)) {
+    console.warn(`Server OK: ${url}`)
+    return url
+  }
+
+  console.warn(`Server not reachable at ${url}, attempting auto-start...`)
+  const port = new URL(url).port || '8188'
+
+  // Strategy 1: comfy-cli (pip install comfy-cli)
+  try {
+    execSync('which comfy', { stdio: 'pipe' })
+    console.warn('Starting ComfyUI via comfy-cli...')
+    const proc = spawn(
+      'comfy',
+      ['launch', '--background', '--', '--cpu', '--port', port],
+      {
+        stdio: 'ignore',
+        detached: true
+      }
+    )
+    proc.unref()
+    await waitForServer(url, 120000)
+    return url
+  } catch {
+    // comfy-cli not available
+  }
+
+  // Strategy 2: python main.py from TEST_COMFYUI_DIR or .comfy-qa/ComfyUI
+  const comfyDir = findComfyUIDir()
+  if (comfyDir) {
+    console.warn(`Starting ComfyUI from ${comfyDir}...`)
+    const proc = spawn('python', ['main.py', '--cpu', '--port', port], {
+      cwd: comfyDir,
+      stdio: 'ignore',
+      detached: true
+    })
+    proc.unref()
+    await waitForServer(url, 120000)
+    return url
+  }
+
+  // Strategy 3: clone ComfyUI and start
+  const cloneDir = resolve('.comfy-qa/ComfyUI')
+  if (!existsSync(resolve(cloneDir, 'main.py'))) {
+    console.warn('No ComfyUI installation found, cloning...')
+    try {
+      execSync(
+        `git clone --depth 1 https://github.com/comfyanonymous/ComfyUI.git "${cloneDir}"`,
+        { stdio: 'inherit', timeout: 120000 }
+      )
+      console.warn('Installing ComfyUI dependencies...')
+      execSync('pip install -r requirements.txt', {
+        cwd: cloneDir,
+        stdio: 'inherit',
+        timeout: 300000
+      })
+    } catch (err) {
+      console.warn(
+        `Clone/install failed: ${err instanceof Error ? err.message : err}`
+      )
+    }
+  }
+  if (existsSync(resolve(cloneDir, 'main.py'))) {
+    console.warn(`Starting ComfyUI from ${cloneDir}...`)
+    const proc = spawn('python', ['main.py', '--cpu', '--port', port], {
+      cwd: cloneDir,
+      stdio: 'ignore',
+      detached: true
+    })
+    proc.unref()
+    await waitForServer(url, 120000)
+    return url
+  }
+
+  // Strategy 4: fallback to testcloud
+  console.warn(`Local server failed. Falling back to ${CLOUD_FALLBACK_URL}`)
+  if (await isReachable(CLOUD_FALLBACK_URL)) {
+    console.warn(`Cloud server OK: ${CLOUD_FALLBACK_URL}`)
+    return CLOUD_FALLBACK_URL
+  }
+
+  console.error(`
+No ComfyUI server available. Tried:
+  1. ${url} (not reachable)
+  2. comfy-cli (not installed)
+  3. Local ComfyUI installation (not found)
+  4. ${CLOUD_FALLBACK_URL} (not reachable)
+
+Install: pip install comfy-cli && comfy install && comfy launch --cpu
+`)
+  process.exit(1)
+}
+
+function findComfyUIDir(): string | undefined {
+  const candidates = [
+    process.env.TEST_COMFYUI_DIR,
+    resolve('.comfy-qa/ComfyUI'),
+    '/home/ComfyUI'
+  ]
+  return candidates.find((d) => d && existsSync(resolve(d, 'main.py')))
+}
+
+async function isReachable(url: string): Promise<boolean> {
+  try {
+    const controller = new AbortController()
+    const timeout = setTimeout(() => controller.abort(), 5000)
+    const res = await fetch(url, { signal: controller.signal })
+    clearTimeout(timeout)
+    return res.ok || res.status === 200 || res.status === 304
+  } catch {
+    return false
+  }
+}
+
+async function waitForServer(url: string, timeoutMs: number): Promise<void> {
+  const start = Date.now()
+  while (Date.now() - start < timeoutMs) {
+    if (await isReachable(url)) {
+      console.warn('Server is ready')
+      return
+    }
+    await new Promise((r) => setTimeout(r, 2000))
+  }
+  console.error(`Server did not start within ${timeoutMs / 1000}s`)
+  process.exit(1)
+}
+
+// ── Utilities ──
+
+function shell(cmd: string): string {
+  return execSync(cmd, { encoding: 'utf-8', timeout: 30000 })
+}
+
+function writeTmpFile(
+  outputDir: string,
+  filename: string,
+  content: string
+): string {
+  const tmpDir = resolve(outputDir, '.tmp')
+  mkdirSync(tmpDir, { recursive: true })
+  const filePath = resolve(tmpDir, filename)
+  writeFileSync(filePath, content)
+  return filePath
+}
+
+function resolveOutputDir(defaultPath: string): string {
+  const dir = values.output ? resolve(values.output) : resolve(defaultPath)
+  mkdirSync(dir, { recursive: true })
+  return dir
+}
+
+function logHeader(opts: { label: string; outputDir: string; extra?: string }) {
+  console.warn(`QA target: ${opts.label}`)
+  console.warn(`Output:    ${opts.outputDir}`)
+  console.warn(`Server:    ${resolvedServerUrl}`)
+  if (values.ref) console.warn(`Ref:       ${values.ref}`)
+  if (opts.extra) console.warn(opts.extra)
+}
+
+function exit(code: number, outputDir: string): never {
+  console.warn('\n=== QA Complete ===')
+  console.warn(`Results: ${outputDir}`)
+  try {
+    console.warn(shell(`ls -la "${outputDir}"`))
+  } catch {
+    // not critical
+  }
+  process.exit(code)
+}
+
+function tryParseArgs() {
+  try {
+    const parsed = parseArgs({
+      args: process.argv.slice(2),
+      options: {
+        target: { type: 'string', short: 't', default: 'head' },
+        uncommitted: { type: 'boolean', default: false },
+        url: { type: 'string', default: '' },
+        ref: { type: 'string', default: '' },
+        output: { type: 'string', short: 'o', default: '' },
+        help: { type: 'boolean', short: 'h', default: false }
+      },
+      allowPositionals: true,
+      strict: true
+    })
+    return {
+      values: parsed.values as {
+        target: string
+        uncommitted: boolean
+        url: string
+        ref: string
+        output: string
+        help: boolean
+      },
+      positionals: parsed.positionals
+    }
+  } catch (err) {
+    console.error(`Error: ${err instanceof Error ? err.message : err}\n`)
+    printUsage()
+    process.exit(1)
+  }
+}
+
+function printUsage() {
+  console.warn(`
+QA CLI — Reproduce issues & test PRs for ComfyUI frontend
+
+Usage:
+  pnpm qa <number|url> [options]
+  pnpm qa --uncommitted
+
+Targets:
+  10253              Number (auto-detects issue vs PR via gh CLI)
+  https://github.com/Comfy-Org/ComfyUI_frontend/issues/10253
+  https://github.com/Comfy-Org/ComfyUI_frontend/pull/10270
+
+Options:
+  -t, --target <head|base|both>
+                     For PRs: which ref to test (default: head)
+                       head  — test the fix (PR head)
+                       base  — reproduce the bug (PR base)
+                       both  — base then head
+  --uncommitted      Test local uncommitted changes
+  --url <url>        ComfyUI server URL (default: from .env or http://127.0.0.1:8188)
+  --ref <ref>        Git ref to test against
+  -o, --output <dir> Override output directory (default: .comfy-qa/<number>)
+  -h, --help         Show this help
+
+Environment (auto-loaded from .env.local or .env):
+  GEMINI_API_KEY     Required — used for PR analysis, video review, TTS
+  ANTHROPIC_API_KEY  Optional locally — Claude Agent SDK auto-detects Claude Code session
+
+Examples:
+  pnpm qa 10253                  # reproduce an issue
+  pnpm qa 10270                  # test PR head (the fix)
+  pnpm qa 10270 -t base          # reproduce bug on PR base
+  pnpm qa 10270 -t both          # test base + head
+  pnpm qa --uncommitted          # test local changes
+`)
+}

.claude/skills/hardening-flaky-e2e-tests/SKILL.md

@@ -0,0 +1,246 @@
+---
+name: hardening-flaky-e2e-tests
+description: 'Diagnoses and fixes flaky Playwright e2e tests by replacing race-prone patterns with retry-safe alternatives. Use when triaging CI flakes, hardening spec files, fixing timing races, or asked to stabilize browser tests. Triggers on: flaky, flake, harden, stabilize, race condition in e2e, intermittent failure.'
+---
+
+# Hardening Flaky E2E Tests
+
+Fix flaky Playwright specs by identifying race-prone patterns and replacing them with retry-safe alternatives. This skill covers diagnosis, pattern matching, and mechanical transforms — not writing new tests (see `writing-playwright-tests` for that).
+
+## Workflow
+
+### 1. Gather CI Evidence
+
+```bash
+gh run list --workflow=ci-test.yaml --limit=5
+gh run download <run-id> -n playwright-report
+```
+
+- Open `report.json` and search for `"status": "flaky"` entries.
+- Collect file paths, test titles, and error messages.
+- Do NOT trust green checks alone — flaky tests that passed on retry still need fixing.
+- Use `error-context.md`, traces, and page snapshots before editing code.
+- Pull the newest run after each push instead of assuming the flaky set is unchanged.
+
+### 2. Classify the Flake
+
+Read the failing assertion and match it against the pattern table. Most flakes fall into one of these categories:
+
+| #   | Pattern                               | Signature in Code                                         | Fix                                                              |
+| --- | ------------------------------------- | --------------------------------------------------------- | ---------------------------------------------------------------- |
+| 1   | **Snapshot-then-assert**              | `expect(await evaluate()).toBe(x)`                        | `await expect.poll(() => evaluate()).toBe(x)`                    |
+| 2   | **Immediate count**                   | `const n = await loc.count(); expect(n).toBe(3)`          | `await expect(loc).toHaveCount(3)`                               |
+| 3   | **nextFrame after menu click**        | `clickMenuItem(x); nextFrame()`                           | `clickMenuItem(x); contextMenu.waitForHidden()`                  |
+| 4   | **Tight poll timeout**                | `expect.poll(..., { timeout: 250 })`                      | ≥2000 ms; prefer default 5000 ms                                 |
+| 5   | **Immediate evaluate after mutation** | `setSetting(k, v); expect(await evaluate()).toBe(x)`      | `await expect.poll(() => evaluate()).toBe(x)`                    |
+| 6   | **Screenshot without readiness**      | `loadWorkflow(); nextFrame(); toHaveScreenshot()`         | `waitForNodes()` or poll state first                             |
+| 7   | **Non-deterministic node order**      | `getNodeRefsByType('X')[0]` with >1 match                 | `getNodeRefById(id)` or guard `toHaveLength(1)`                  |
+| 8   | **Fake readiness helper**             | Helper clicks but doesn't assert state                    | Remove; poll the actual value                                    |
+| 9   | **Immediate graph state after drop**  | `expect(await getLinkCount()).toBe(1)`                    | `await expect.poll(() => getLinkCount()).toBe(1)`                |
+| 10  | **Immediate boundingBox/layout read** | `const box = await loc.boundingBox(); expect(box!.width)` | `await expect.poll(() => loc.boundingBox().then(b => b?.width))` |
+
+### 3. Apply the Transform
+
+#### Rule: Choose the Smallest Correct Assertion
+
+- **Locator state** → use built-in retrying assertions: `toBeVisible()`, `toHaveText()`, `toHaveCount()`, `toHaveClass()`
+- **Single async value** → `expect.poll(() => asyncFn()).toBe(expected)`
+- **Multiple assertions that must settle together** → `expect(async () => { ... }).toPass()`
+- **Never** use `waitForTimeout()` to hide a race.
+
+```typescript
+// ✅ Single value — use expect.poll
+await expect
+  .poll(() => comfyPage.page.evaluate(() => window.app!.graph.links.length))
+  .toBe(3)
+
+// ✅ Locator count — use toHaveCount
+await expect(comfyPage.page.locator('.dom-widget')).toHaveCount(2)
+
+// ✅ Multiple conditions — use toPass
+await expect(async () => {
+  expect(await node1.getValue()).toBe('foo')
+  expect(await node2.getValue()).toBe('bar')
+}).toPass({ timeout: 5000 })
+```
+
+#### Rule: Wait for the Real Readiness Boundary
+
+Visible is not always ready. Prefer user-facing assertions when possible; poll internal state only when there is no UI surface to assert on.
+
+Common readiness boundaries:
+
+| After this action...                   | Wait for...                                                  |
+| -------------------------------------- | ------------------------------------------------------------ |
+| Canvas interaction (drag, click node)  | `await comfyPage.nextFrame()`                                |
+| Menu item click                        | `await contextMenu.waitForHidden()`                          |
+| Workflow load                          | `await comfyPage.workflow.loadWorkflow(...)` (built-in wait) |
+| Settings write                         | Poll the setting value with `expect.poll()`                  |
+| Node pin/bypass/collapse toggle        | `await expect.poll(() => nodeRef.isPinned()).toBe(true)`     |
+| Graph mutation (add/remove node, link) | Poll link/node count                                         |
+| Clipboard write                        | Poll pasted value                                            |
+| Screenshot                             | Ensure nodes are rendered: `waitForNodes()` or poll state    |
+
+#### Rule: Expose Locators for Retrying Assertions
+
+When a helper returns a count via `await loc.count()`, callers can't use `toHaveCount()`. Expose the underlying `Locator` as a getter so callers choose between:
+
+```typescript
+// Helper exposes locator
+get domWidgets(): Locator {
+  return this.page.locator('.dom-widget')
+}
+
+// Caller uses retrying assertion
+await expect(comfyPage.domWidgets).toHaveCount(2)
+```
+
+Replace count methods with locator getters so callers can use retrying assertions directly.
+
+#### Rule: Fix Check-then-Act Races in Helpers
+
+```typescript
+// ❌ Race: count can change between check and waitFor
+const count = await locator.count()
+if (count > 0) {
+  await locator.waitFor({ state: 'hidden' })
+}
+
+// ✅ Direct: waitFor handles both cases
+await locator.waitFor({ state: 'hidden' })
+```
+
+#### Rule: Remove force:true from Clicks
+
+`force: true` bypasses actionability checks, hiding real animation/visibility races. Remove it and fix the underlying timing issue.
+
+```typescript
+// ❌ Hides the race
+await closeButton.click({ force: true })
+
+// ✅ Surfaces the real issue — fix with proper wait
+await closeButton.click()
+await dialog.waitForHidden()
+```
+
+#### Rule: Handle Non-deterministic Element Order
+
+When `getNodeRefsByType` returns multiple nodes, the order is not guaranteed. Don't use index `[0]` blindly.
+
+```typescript
+// ❌ Assumes order
+const node = (await comfyPage.nodeOps.getNodeRefsByType('CLIPTextEncode'))[0]
+
+// ✅ Find by ID or proximity
+const nodes = await comfyPage.nodeOps.getNodeRefsByType('CLIPTextEncode')
+let target = nodes[0]
+for (const n of nodes) {
+  const pos = await n.getPosition()
+  if (Math.abs(pos.y - expectedY) < minDist) target = n
+}
+```
+
+Or guard the assumption:
+
+```typescript
+const nodes = await comfyPage.nodeOps.getNodeRefsByType('CLIPTextEncode')
+expect(nodes).toHaveLength(1)
+const node = nodes[0]
+```
+
+#### Rule: Use toPass for Timing-sensitive Dismiss Guards
+
+Some UI elements (e.g. LiteGraph's graphdialog) have built-in dismiss delays. Retry the entire dismiss action:
+
+```typescript
+// ✅ Retry click+assert together
+await expect(async () => {
+  await comfyPage.canvas.click({ position: { x: 10, y: 10 } })
+  await expect(dialog).toBeHidden({ timeout: 500 })
+}).toPass({ timeout: 5000 })
+```
+
+### 4. Keep Changes Narrow
+
+- Shared helpers should drive setup to a stable boundary.
+- Do not encode one-spec timing assumptions into generic helpers.
+- If a race only matters to one spec, prefer a local wait in that spec.
+- If a helper fails before the real test begins, remove or relax the brittle precondition and let downstream UI interaction prove readiness.
+
+### 5. Verify Narrowly
+
+```bash
+# Targeted rerun with repetition
+pnpm test:browser:local -- browser_tests/tests/myFile.spec.ts --repeat-each 10
+
+# Single test by line number (avoids grep quoting issues on Windows)
+pnpm test:browser:local -- browser_tests/tests/myFile.spec.ts:42
+```
+
+- Use `--repeat-each 10` for targeted flake verification (use 20 for single test cases).
+- Verify with the smallest command that exercises the flaky path.
+
+### 6. Watch CI E2E Runs
+
+After pushing, use `gh` to monitor the E2E workflow:
+
+```bash
+# Find the run for the current branch
+gh run list --workflow="CI: Tests E2E" --branch=$(git branch --show-current) --limit=1
+
+# Watch it live (blocks until complete, streams logs)
+gh run watch <run-id>
+
+# One-liner: find and watch the latest E2E run for the current branch
+gh run list --workflow="CI: Tests E2E" --branch=$(git branch --show-current) --limit=1 --json databaseId --jq ".[0].databaseId" | xargs gh run watch
+```
+
+On Windows (PowerShell):
+
+```powershell
+# One-liner equivalent
+gh run watch (gh run list --workflow="CI: Tests E2E" --branch=$(git branch --show-current) --limit=1 --json databaseId --jq ".[0].databaseId")
+```
+
+After the run completes:
+
+```bash
+# Download the Playwright report artifact
+gh run download <run-id> -n playwright-report
+
+# View the run summary in browser
+gh run view <run-id> --web
+```
+
+Also watch the unit test workflow in parallel if you changed helpers:
+
+```bash
+gh run list --workflow="CI: Tests Unit" --branch=$(git branch --show-current) --limit=1
+```
+
+### 7. Pre-merge Checklist
+
+Before merging a flaky-test fix, confirm:
+
+- [ ] The latest CI artifact was inspected directly
+- [ ] The root cause is stated as a race or readiness mismatch
+- [ ] The fix waits on the real readiness boundary
+- [ ] The assertion primitive matches the job (poll vs toHaveCount vs toPass)
+- [ ] The fix stays local unless a shared helper truly owns the race
+- [ ] Local verification uses a targeted rerun
+- [ ] No behavioral changes to the test — only timing/retry strategy updated
+
+## Local Noise — Do Not Fix
+
+These are local distractions, not CI root causes:
+
+- Missing local input fixture files required by the test path
+- Missing local models directory
+- Teardown `EPERM` while restoring the local browser-test user data directory
+- Local screenshot baseline differences on Windows
+
+Rules:
+
+- First confirm whether it blocks the exact flaky path under investigation.
+- Do not commit temporary local assets used only for verification.
+- Do not commit local screenshot baselines.

.env_example

@@ -46,3 +46,11 @@ ALGOLIA_API_KEY=684d998c36b67a9a9fce8fc2d8860579
 # SENTRY_ORG=comfy-org
 # SENTRY_PROJECT=cloud-frontend-staging
 # SENTRY_PROJECT_PROD= # prod project slug for sourcemap uploads
+
+# ── QA Skill (scripts/qa-record.ts) ──
+# Required for automated bug reproduction via `pnpm exec tsx scripts/qa-record.ts`
+# GEMINI_API_KEY is required — used for PR analysis, video review, and TTS narration
+GEMINI_API_KEY=
+# ANTHROPIC_API_KEY is optional locally — Claude Agent SDK auto-detects Claude Code session
+# Required in CI (set as GitHub Actions secret)
+# ANTHROPIC_API_KEY=

.github/workflows/api-update-registry-api-types.yaml

@@ -1,107 +0,0 @@
-# Description: When upstream comfy-api is updated, click dispatch to update the TypeScript type definitions in this repo
-name: 'Api: Update Registry API Types'
-
-on:
-  # Manual trigger
-  workflow_dispatch:
-
-  # Triggered from comfy-api repo
-  repository_dispatch:
-    types: [comfy-api-updated]
-
-jobs:
-  update-registry-types:
-    runs-on: ubuntu-latest
-    permissions:
-      contents: write
-      pull-requests: write
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v6
-
-      - name: Install pnpm
-        uses: pnpm/action-setup@fc06bc1257f339d1d5d8b3a19a8cae5388b55320 # v4.4.0
-
-      - name: Setup Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version-file: '.nvmrc'
-          cache: 'pnpm'
-
-      - name: Install dependencies
-        run: pnpm install --frozen-lockfile
-
-      - name: Checkout comfy-api repository
-        uses: actions/checkout@v6
-        with:
-          repository: Comfy-Org/comfy-api
-          path: comfy-api
-          token: ${{ secrets.COMFY_API_PAT }}
-          clean: true
-
-      - name: Get API commit information
-        id: api-info
-        run: |
-          cd comfy-api
-          API_COMMIT=$(git rev-parse --short HEAD)
-          echo "commit=${API_COMMIT}" >> $GITHUB_OUTPUT
-          cd ..
-
-      - name: Generate API types
-        run: |
-          echo "Generating TypeScript types from comfy-api@${{ steps.api-info.outputs.commit }}..."
-          mkdir -p ./packages/registry-types/src
-          pnpm dlx openapi-typescript ./comfy-api/openapi.yml --output ./packages/registry-types/src/comfyRegistryTypes.ts
-
-      - name: Validate generated types
-        run: |
-          if [ ! -f ./packages/registry-types/src/comfyRegistryTypes.ts ]; then
-            echo "Error: Types file was not generated."
-            exit 1
-          fi
-
-          # Check if file is not empty
-          if [ ! -s ./packages/registry-types/src/comfyRegistryTypes.ts ]; then
-            echo "Error: Generated types file is empty."
-            exit 1
-          fi
-
-      - name: Lint generated types
-        run: |
-          echo "Linting generated Comfy Registry API types..."
-          pnpm lint:fix:no-cache -- ./packages/registry-types/src/comfyRegistryTypes.ts
-
-      - name: Check for changes
-        id: check-changes
-        run: |
-          if [[ -z $(git status --porcelain ./packages/registry-types/src/comfyRegistryTypes.ts) ]]; then
-            echo "No changes to Comfy Registry API types detected."
-            echo "changed=false" >> $GITHUB_OUTPUT
-            exit 0
-          else
-            echo "Changes detected in Comfy Registry API types."
-            echo "changed=true" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Create Pull Request
-        if: steps.check-changes.outputs.changed == 'true'
-        uses: peter-evans/create-pull-request@c0f553fe549906ede9cf27b5156039d195d2ece0 # v8.1.0
-        with:
-          token: ${{ secrets.PR_GH_TOKEN }}
-          commit-message: '[chore] Update Comfy Registry API types from comfy-api@${{ steps.api-info.outputs.commit }}'
-          title: '[chore] Update Comfy Registry API types from comfy-api@${{ steps.api-info.outputs.commit }}'
-          body: |
-            ## Automated API Type Update
-
-            This PR updates the Comfy Registry API types from the latest comfy-api OpenAPI specification.
-
-            - API commit: ${{ steps.api-info.outputs.commit }}
-            - Generated on: ${{ github.event.repository.updated_at }}
-
-            These types are automatically generated using openapi-typescript.
-          branch: update-registry-types-${{ steps.api-info.outputs.commit }}
-          base: main
-          labels: CNR
-          delete-branch: true
-          add-paths: |
-            packages/registry-types/src/comfyRegistryTypes.ts

.github/workflows/ci-tests-unit.yaml

@@ -31,4 +31,5 @@ jobs:
         uses: codecov/codecov-action@1af58845a975a7985b0beb0cbe6fbbb71a41dbad # v5.5.3
         with:
           files: coverage/lcov.info
+          token: ${{ secrets.CODECOV_TOKEN }}
           fail_ci_if_error: false

.github/workflows/ci-website-build.yaml

@@ -0,0 +1,33 @@
+# Description: Build and validate the marketing website (apps/website)
+name: 'CI: Website Build'
+
+on:
+  push:
+    branches: [main, master, website/*]
+    paths:
+      - 'apps/website/**'
+      - 'packages/design-system/**'
+      - 'pnpm-lock.yaml'
+  pull_request:
+    branches-ignore: [wip/*, draft/*, temp/*]
+    paths:
+      - 'apps/website/**'
+      - 'packages/design-system/**'
+      - 'pnpm-lock.yaml'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Setup frontend
+        uses: ./.github/actions/setup-frontend
+
+      - name: Build website
+        run: pnpm --filter @comfyorg/website build

.github/workflows/hub-ci.yaml

@@ -1,182 +0,0 @@
-name: Hub CI
-
-on:
-  push:
-    branches: [main]
-    paths:
-      - 'apps/hub/**'
-      - '.github/workflows/hub-ci.yaml'
-  pull_request:
-    branches: [main]
-    paths:
-      - 'apps/hub/**'
-      - '.github/workflows/hub-ci.yaml'
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
-permissions:
-  contents: read
-  pull-requests: write
-
-jobs:
-  lint:
-    name: Lint & Check
-    runs-on: ubuntu-latest
-    defaults:
-      run:
-        working-directory: apps/hub
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-
-      - name: Setup frontend
-        uses: ./.github/actions/setup-frontend
-
-      - name: Astro Check
-        run: pnpm run check
-
-      - name: Unit Tests
-        run: pnpm test
-
-      - name: Validate Templates
-        run: pnpm run validate:templates
-        continue-on-error: true
-
-  build:
-    name: Build Hub
-    runs-on: ubuntu-latest
-    defaults:
-      run:
-        working-directory: apps/hub
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-
-      - name: Setup frontend
-        uses: ./.github/actions/setup-frontend
-
-      - name: Build site
-        run: pnpm run build
-        env:
-          HUB_SKIP_SYNC: 'true'
-          SKIP_AI_GENERATION: 'true'
-
-      - name: Upload build artifact
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
-        with:
-          name: hub-build
-          path: apps/hub/dist
-          retention-days: 1
-
-  seo-audit:
-    name: SEO Audit
-    needs: build
-    runs-on: ubuntu-latest
-    defaults:
-      run:
-        working-directory: apps/hub
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-
-      - name: Setup frontend
-        uses: ./.github/actions/setup-frontend
-
-      - name: Download build artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: hub-build
-          path: apps/hub/dist
-
-      - name: Validate sitemap
-        id: sitemap
-        continue-on-error: true
-        run: |
-          echo "## Sitemap Validation" >> $GITHUB_STEP_SUMMARY
-          if pnpm run validate:sitemap 2>&1 | tee sitemap-output.txt; then
-            echo "✅ Sitemap validation passed" >> $GITHUB_STEP_SUMMARY
-            echo "status=passed" >> $GITHUB_OUTPUT
-          else
-            echo "❌ Sitemap validation failed" >> $GITHUB_STEP_SUMMARY
-            echo "status=failed" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Run SEO audit
-        id: seo
-        continue-on-error: true
-        run: |
-          echo "## SEO Audit" >> $GITHUB_STEP_SUMMARY
-          if pnpm run audit:seo 2>&1 | tee seo-output.txt; then
-            echo "✅ SEO audit passed" >> $GITHUB_STEP_SUMMARY
-            echo "status=passed" >> $GITHUB_OUTPUT
-          else
-            echo "⚠️ SEO audit found issues" >> $GITHUB_STEP_SUMMARY
-            echo "status=issues" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Check internal links
-        id: links
-        continue-on-error: true
-        run: |
-          echo "## Link Check" >> $GITHUB_STEP_SUMMARY
-          DIST_DIR="dist"
-          if [ ! -d "$DIST_DIR" ]; then
-            echo "⚠️ No build output found at $DIST_DIR" >> $GITHUB_STEP_SUMMARY
-            echo "status=skipped" >> $GITHUB_OUTPUT
-            exit 0
-          fi
-
-          BROKEN_FILE="broken-links.txt"
-          : > "$BROKEN_FILE"
-          BROKEN_COUNT=0
-          TOTAL_COUNT=0
-
-          for htmlfile in $(find "$DIST_DIR" -name '*.html' \
-            -not -path "$DIST_DIR/ar/*" -not -path "$DIST_DIR/es/*" -not -path "$DIST_DIR/fr/*" \
-            -not -path "$DIST_DIR/ja/*" -not -path "$DIST_DIR/ko/*" -not -path "$DIST_DIR/pt-BR/*" \
-            -not -path "$DIST_DIR/ru/*" -not -path "$DIST_DIR/tr/*" -not -path "$DIST_DIR/zh/*" \
-            -not -path "$DIST_DIR/zh-TW/*" | head -500); do
-            hrefs=$(grep -oP 'href="(/[^"]*)"' "$htmlfile" | sed 's/href="//;s/"$//' || true)
-            for href in $hrefs; do
-              TOTAL_COUNT=$((TOTAL_COUNT + 1))
-              clean="${href%%#*}"
-              clean="${clean%%\?*}"
-              if [ -z "$clean" ] || [ "$clean" = "/" ]; then continue; fi
-              found=false
-              if [[ "$clean" =~ \.[a-zA-Z0-9]+$ ]]; then
-                [ -f "${DIST_DIR}${clean}" ] && found=true
-              else
-                base="${clean%/}"
-                [ -f "${DIST_DIR}${base}/index.html" ] && found=true
-                [ "$found" = false ] && [ -f "${DIST_DIR}${base}.html" ] && found=true
-                [ "$found" = false ] && [ -f "${DIST_DIR}${clean}" ] && found=true
-                [ "$found" = false ] && [ -d "${DIST_DIR}${base}" ] && found=true
-              fi
-              if [ "$found" = false ]; then
-                BROKEN_COUNT=$((BROKEN_COUNT + 1))
-                echo "- \`${href}\` (in ${htmlfile#${DIST_DIR}/})" >> "$BROKEN_FILE"
-              fi
-            done
-          done
-
-          if [ "$BROKEN_COUNT" -eq 0 ]; then
-            echo "✅ All internal links valid ($TOTAL_COUNT checked)" >> $GITHUB_STEP_SUMMARY
-            echo "status=passed" >> $GITHUB_OUTPUT
-          else
-            echo "❌ Found $BROKEN_COUNT broken internal links out of $TOTAL_COUNT" >> $GITHUB_STEP_SUMMARY
-            head -n 50 "$BROKEN_FILE" >> $GITHUB_STEP_SUMMARY
-            echo "status=failed" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Upload SEO reports
-        if: always()
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
-        with:
-          name: hub-seo-reports
-          path: |
-            apps/hub/seo-output.txt
-            apps/hub/seo-summary.json
-            apps/hub/broken-links.txt
-          if-no-files-found: ignore

.github/workflows/hub-cron-rebuild.yaml

@@ -1,68 +0,0 @@
-name: Hub Cron Rebuild
-
-on:
-  schedule:
-    # Every 15 minutes — rebuilds the site to pick up new UGC workflows
-    # for search index, sitemap, filter pages, and pre-rendered detail pages.
-    - cron: '*/15 * * * *'
-  workflow_dispatch:
-
-concurrency:
-  group: hub-deploy-prod
-  cancel-in-progress: false
-
-permissions:
-  contents: read
-
-jobs:
-  rebuild:
-    runs-on: ubuntu-latest
-    env:
-      SKIP_AI_GENERATION: 'true'
-      PUBLIC_POSTHOG_KEY: ${{ secrets.HUB_POSTHOG_KEY }}
-      PUBLIC_GA_MEASUREMENT_ID: ${{ secrets.HUB_GA_MEASUREMENT_ID }}
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-
-      - name: Setup frontend
-        uses: ./.github/actions/setup-frontend
-
-      - name: Checkout templates data
-        uses: actions/checkout@v6
-        with:
-          repository: Comfy-Org/workflow_templates
-          path: _workflow_templates
-          sparse-checkout: templates
-          token: ${{ secrets.GH_TOKEN }}
-
-      - name: Restore content cache
-        uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
-        with:
-          path: apps/hub/.content-cache
-          key: hub-content-cache-cron-prod-${{ hashFiles('_workflow_templates/templates/**', 'apps/hub/src/**') }}
-          restore-keys: |
-            hub-content-cache-cron-prod-
-
-      - name: Sync templates
-        run: pnpm run sync
-        working-directory: apps/hub
-        env:
-          HUB_TEMPLATES_DIR: ${{ github.workspace }}/_workflow_templates/templates
-
-      - name: Build Astro site
-        run: pnpm run build
-        working-directory: apps/hub
-        env:
-          PUBLIC_HUB_API_URL: ${{ secrets.HUB_API_URL_PRODUCTION }}
-          PUBLIC_COMFY_CLOUD_URL: ${{ secrets.COMFY_CLOUD_URL_PRODUCTION }}
-          PUBLIC_APPROVED_ONLY: 'true'
-
-      - name: Deploy to Vercel
-        uses: amondnet/vercel-action@16e87c0a08142b0d0d33b76aeaf20823c381b9b9 # v25.2.0
-        with:
-          vercel-token: ${{ secrets.VERCEL_TOKEN }}
-          vercel-org-id: ${{ secrets.VERCEL_ORG_ID }}
-          vercel-project-id: ${{ secrets.HUB_VERCEL_PROJECT_ID }}
-          working-directory: apps/hub
-          vercel-args: '--prebuilt --prod'

.github/workflows/hub-deploy.yaml

@@ -1,80 +0,0 @@
-name: Deploy Hub
-
-on:
-  workflow_dispatch:
-    inputs:
-      skip_ai:
-        description: 'Skip AI content generation'
-        type: boolean
-        default: false
-      force_regenerate:
-        description: 'Force regenerate all content (ignore cache)'
-        type: boolean
-        default: false
-      template_filter:
-        description: 'Regenerate specific template only (e.g. "flux_schnell")'
-        type: string
-        default: ''
-
-concurrency:
-  group: hub-deploy-prod
-  cancel-in-progress: false
-
-permissions:
-  contents: read
-
-jobs:
-  build-deploy:
-    runs-on: ubuntu-latest
-    env:
-      OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-      PUBLIC_POSTHOG_KEY: ${{ secrets.HUB_POSTHOG_KEY }}
-      PUBLIC_GA_MEASUREMENT_ID: ${{ secrets.HUB_GA_MEASUREMENT_ID }}
-      SKIP_AI_GENERATION: ${{ inputs.skip_ai && 'true' || '' }}
-      FORCE_AI_REGENERATE: ${{ inputs.force_regenerate && 'true' || '' }}
-      AI_TEMPLATE_FILTER: ${{ inputs.template_filter }}
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-
-      - name: Setup frontend
-        uses: ./.github/actions/setup-frontend
-
-      - name: Checkout templates data
-        uses: actions/checkout@v6
-        with:
-          repository: Comfy-Org/workflow_templates
-          path: _workflow_templates
-          sparse-checkout: templates
-          token: ${{ secrets.GH_TOKEN }}
-
-      - name: Restore content cache
-        uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
-        with:
-          path: apps/hub/.content-cache
-          key: hub-content-cache-${{ hashFiles('_workflow_templates/templates/**', 'apps/hub/src/**') }}
-          restore-keys: |
-            hub-content-cache-
-
-      - name: Sync templates
-        run: pnpm run sync
-        working-directory: apps/hub
-        env:
-          HUB_TEMPLATES_DIR: ${{ github.workspace }}/_workflow_templates/templates
-
-      - name: Build Astro site
-        run: pnpm run build
-        working-directory: apps/hub
-        env:
-          PUBLIC_HUB_API_URL: ${{ secrets.HUB_API_URL_PRODUCTION }}
-          PUBLIC_COMFY_CLOUD_URL: ${{ secrets.COMFY_CLOUD_URL_PRODUCTION }}
-          PUBLIC_APPROVED_ONLY: 'true'
-
-      - name: Deploy to Vercel
-        uses: amondnet/vercel-action@16e87c0a08142b0d0d33b76aeaf20823c381b9b9 # v25.2.0
-        with:
-          vercel-token: ${{ secrets.VERCEL_TOKEN }}
-          vercel-org-id: ${{ secrets.VERCEL_ORG_ID }}
-          vercel-project-id: ${{ secrets.HUB_VERCEL_PROJECT_ID }}
-          working-directory: apps/hub
-          vercel-args: '--prebuilt --prod'

.github/workflows/hub-preview-cron.yaml

@@ -1,134 +0,0 @@
-name: Hub Preview Cron
-
-on:
-  schedule:
-    - cron: '*/15 * * * *'
-  workflow_dispatch:
-
-permissions:
-  contents: read
-  pull-requests: write
-
-jobs:
-  discover:
-    runs-on: ubuntu-latest
-    outputs:
-      matrix: ${{ steps.targets.outputs.matrix }}
-    steps:
-      - uses: actions/checkout@v6
-
-      - name: Build rebuild targets
-        id: targets
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          targets='[]'
-
-          # Main with production API (all workflows, no approved filter)
-          targets=$(echo "$targets" | jq -c '. + [{"ref": "main", "is_main": true, "pr": 0, "api_env": "production"}]')
-
-          # Main with test API
-          targets=$(echo "$targets" | jq -c '. + [{"ref": "main", "is_main": true, "pr": 0, "api_env": "test"}]')
-
-          # Find open PRs with the "preview-cron" label
-          prs=$(gh pr list --label "preview-cron" --state open --json number,headRefName)
-          for row in $(echo "$prs" | jq -c '.[]'); do
-            ref=$(echo "$row" | jq -r '.headRefName')
-            num=$(echo "$row" | jq -r '.number')
-            targets=$(echo "$targets" | jq -c \
-              --arg ref "$ref" --argjson num "$num" \
-              '. + [{"ref": $ref, "is_main": false, "pr": $num, "api_env": "test"}]')
-          done
-
-          echo "matrix={\"include\":$targets}" >> "$GITHUB_OUTPUT"
-          echo "### Rebuild targets" >> "$GITHUB_STEP_SUMMARY"
-          echo "$targets" | jq '.' >> "$GITHUB_STEP_SUMMARY"
-
-  rebuild:
-    needs: discover
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix: ${{ fromJson(needs.discover.outputs.matrix) }}
-    concurrency:
-      group: hub-preview-cron-${{ matrix.ref }}-${{ matrix.api_env }}
-      cancel-in-progress: true
-    env:
-      SKIP_AI_GENERATION: 'true'
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-        with:
-          ref: ${{ matrix.ref }}
-
-      - name: Setup frontend
-        uses: ./.github/actions/setup-frontend
-
-      - name: Checkout templates data
-        uses: actions/checkout@v6
-        with:
-          repository: Comfy-Org/workflow_templates
-          path: _workflow_templates
-          sparse-checkout: templates
-          token: ${{ secrets.GH_TOKEN }}
-
-      - name: Restore content cache
-        uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
-        with:
-          path: apps/hub/.content-cache
-          key: hub-content-cache-cron-${{ matrix.ref }}-${{ matrix.api_env }}-${{ hashFiles('_workflow_templates/templates/**', 'apps/hub/src/**') }}
-          restore-keys: |
-            hub-content-cache-cron-${{ matrix.ref }}-${{ matrix.api_env }}-
-
-      - name: Sync templates
-        run: pnpm run sync:en-only
-        working-directory: apps/hub
-        env:
-          HUB_TEMPLATES_DIR: ${{ github.workspace }}/_workflow_templates/templates
-
-      - name: Build Astro site
-        run: pnpm run build
-        working-directory: apps/hub
-        env:
-          PUBLIC_HUB_API_URL: ${{ matrix.api_env == 'test' && secrets.HUB_API_URL_PREVIEW || secrets.HUB_API_URL_PRODUCTION }}
-          PUBLIC_COMFY_CLOUD_URL: ${{ matrix.api_env == 'test' && secrets.COMFY_CLOUD_URL_PREVIEW || secrets.COMFY_CLOUD_URL_PRODUCTION }}
-
-      - name: Deploy to Vercel
-        id: deploy
-        uses: amondnet/vercel-action@16e87c0a08142b0d0d33b76aeaf20823c381b9b9 # v25.2.0
-        with:
-          vercel-token: ${{ secrets.VERCEL_TOKEN }}
-          vercel-org-id: ${{ secrets.VERCEL_ORG_ID }}
-          vercel-project-id: ${{ secrets.HUB_VERCEL_PROJECT_ID }}
-          working-directory: apps/hub
-          vercel-args: '--prebuilt'
-
-      - name: Alias main preview (prod API)
-        if: matrix.is_main && matrix.api_env == 'production' && secrets.HUB_PREVIEW_ALIAS
-        env:
-          PREVIEW_URL: ${{ steps.deploy.outputs.preview-url }}
-          ALIAS: ${{ secrets.HUB_PREVIEW_ALIAS }}
-          VERCEL_TOKEN_VAL: ${{ secrets.VERCEL_TOKEN }}
-          VERCEL_SCOPE: ${{ secrets.VERCEL_ORG_ID }}
-        run: |
-          npx vercel alias "$PREVIEW_URL" "$ALIAS" --token="$VERCEL_TOKEN_VAL" --scope="$VERCEL_SCOPE"
-
-      - name: Alias main preview (test API)
-        if: matrix.is_main && matrix.api_env == 'test' && secrets.HUB_PREVIEW_TEST_ALIAS
-        env:
-          PREVIEW_URL: ${{ steps.deploy.outputs.preview-url }}
-          ALIAS: ${{ secrets.HUB_PREVIEW_TEST_ALIAS }}
-          VERCEL_TOKEN_VAL: ${{ secrets.VERCEL_TOKEN }}
-          VERCEL_SCOPE: ${{ secrets.VERCEL_ORG_ID }}
-        run: |
-          npx vercel alias "$PREVIEW_URL" "$ALIAS" --token="$VERCEL_TOKEN_VAL" --scope="$VERCEL_SCOPE"
-
-      - name: Comment preview URL on PR
-        if: matrix.pr > 0
-        uses: marocchino/sticky-pull-request-comment@773744901bac0e8cbb5a0dc842800d45e9b2b405 # v2.9.4
-        with:
-          number: ${{ matrix.pr }}
-          header: hub-preview-cron
-          message: |
-            🔄 **Hub preview cron rebuilt:** ${{ steps.deploy.outputs.preview-url }}
-            _Last rebuild: ${{ github.event.head_commit.timestamp || 'manual trigger' }}_

.github/workflows/hub-preview.yaml

@@ -1,74 +0,0 @@
-name: Hub Preview
-
-on:
-  pull_request:
-    paths:
-      - 'apps/hub/**'
-  workflow_dispatch:
-
-concurrency:
-  group: hub-preview-${{ github.head_ref || github.run_id }}
-  cancel-in-progress: true
-
-permissions:
-  contents: read
-  pull-requests: write
-
-jobs:
-  preview:
-    runs-on: ubuntu-latest
-    env:
-      SKIP_AI_GENERATION: 'true'
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-
-      - name: Setup frontend
-        uses: ./.github/actions/setup-frontend
-
-      - name: Checkout templates data
-        uses: actions/checkout@v6
-        with:
-          repository: Comfy-Org/workflow_templates
-          path: _workflow_templates
-          sparse-checkout: templates
-          token: ${{ secrets.GH_TOKEN }}
-
-      - name: Restore content cache
-        uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
-        with:
-          path: apps/hub/.content-cache
-          key: hub-content-cache-preview-${{ hashFiles('_workflow_templates/templates/**', 'apps/hub/src/**') }}
-          restore-keys: |
-            hub-content-cache-preview-
-
-      - name: Sync templates
-        run: pnpm run sync:en-only
-        working-directory: apps/hub
-        env:
-          HUB_TEMPLATES_DIR: ${{ github.workspace }}/_workflow_templates/templates
-
-      - name: Build Astro site
-        run: pnpm run build
-        working-directory: apps/hub
-        env:
-          PUBLIC_HUB_API_URL: ${{ secrets.HUB_API_URL_PREVIEW }}
-          PUBLIC_COMFY_CLOUD_URL: ${{ secrets.COMFY_CLOUD_URL_PREVIEW }}
-
-      - name: Deploy preview to Vercel
-        id: deploy
-        uses: amondnet/vercel-action@16e87c0a08142b0d0d33b76aeaf20823c381b9b9 # v25.2.0
-        with:
-          vercel-token: ${{ secrets.VERCEL_TOKEN }}
-          vercel-org-id: ${{ secrets.VERCEL_ORG_ID }}
-          vercel-project-id: ${{ secrets.HUB_VERCEL_PROJECT_ID }}
-          working-directory: apps/hub
-          vercel-args: '--prebuilt'
-
-      - name: Comment preview URL
-        if: github.event_name == 'pull_request'
-        uses: marocchino/sticky-pull-request-comment@773744901bac0e8cbb5a0dc842800d45e9b2b405 # v2.9.4
-        with:
-          header: hub-vercel-preview
-          message: |
-            🚀 **Hub preview deployed:** ${{ steps.deploy.outputs.preview-url }}

.github/workflows/release-version-bump.yaml

@@ -142,6 +142,20 @@ jobs:
           fi
           echo "✅ Branch '$BRANCH' exists"
 
+      - name: Ensure packageManager field exists
+        run: |
+          if ! grep -q '"packageManager"' package.json; then
+            # Old branches (e.g. core/1.42) predate the packageManager field.
+            # Inject it so pnpm/action-setup can resolve the version.
+            node -e "
+              const fs = require('fs');
+              const pkg = JSON.parse(fs.readFileSync('package.json','utf8'));
+              pkg.packageManager = 'pnpm@10.33.0';
+              fs.writeFileSync('package.json', JSON.stringify(pkg, null, 2) + '\n');
+            "
+            echo "Injected packageManager into package.json for legacy branch"
+          fi
+
       - name: Install pnpm
         uses: pnpm/action-setup@fc06bc1257f339d1d5d8b3a19a8cae5388b55320 # v4.4.0
 

.gitignore

@@ -67,6 +67,9 @@ dist.zip
 /temp/
 /tmp/
 
+# QA local output
+/.comfy-qa/
+
 # Generated JSON Schemas
 /schemas/
 

.oxlintrc.json

@@ -12,8 +12,6 @@
     "playwright-report/*",
     "src/extensions/core/*",
     "src/scripts/*",
-    "apps/hub/scripts/**/*",
-    "apps/hub/src/scripts/*",
     "src/types/generatedManagerTypes.ts",
     "src/types/vue-shim.d.ts",
     "test-results/*",
@@ -66,6 +64,7 @@
         ]
       }
     ],
+    "no-unsafe-optional-chaining": "error",
     "no-self-assign": "allow",
     "no-unused-expressions": "off",
     "no-unused-private-class-members": "off",
@@ -106,8 +105,7 @@
         "allowInterfaces": "always"
       }
     ],
-    "vue/no-import-compiler-macros": "error",
-    "vue/no-dupe-keys": "error"
+    "vue/no-import-compiler-macros": "error"
   },
   "overrides": [
     {

AGENTS.md

@@ -179,6 +179,12 @@ This project uses **pnpm**. Always prefer scripts defined in `package.json` (e.g
 24. Do not use function expressions if it's possible to use function declarations instead
 25. Watch out for [Code Smells](https://wiki.c2.com/?CodeSmell) and refactor to avoid them
 
+## Design Standards
+
+Before implementing any user-facing feature, consult the [Comfy Design Standards](https://www.figma.com/design/QreIv5htUaSICNuO2VBHw0/Comfy-Design-Standards) Figma file. Use the Figma MCP to fetch it live — the file is the single source of truth and may be updated by designers at any time.
+
+See `docs/guidance/design-standards.md` for Figma file keys, section node IDs, and component references.
+
 ## Testing Guidelines
 
 See @docs/testing/\*.md for detailed patterns.
@@ -226,6 +232,7 @@ See @docs/testing/\*.md for detailed patterns.
 - shadcn/vue: <https://www.shadcn-vue.com/>
 - Reka UI: <https://reka-ui.com/>
 - PrimeVue: <https://primevue.org>
+- Comfy Design Standards: <https://www.figma.com/design/QreIv5htUaSICNuO2VBHw0/Comfy-Design-Standards>
 - ComfyUI: <https://docs.comfy.org>
 - Electron: <https://www.electronjs.org/docs/latest/>
 - Wiki: <https://deepwiki.com/Comfy-Org/ComfyUI_frontend/1-overview>

CONTRIBUTING.md

@@ -62,6 +62,37 @@ python main.py --port 8188 --cpu
 - Run `pnpm dev:electron` to start the dev server with electron API mocked
 - Run `pnpm dev:cloud` to start the dev server against the cloud backend (instead of local ComfyUI server)
 
+#### Testing with Cloud & Staging Environments
+
+Some features — particularly **partner/API nodes** (e.g. BFL, OpenAI, Stability AI) — require a cloud backend for authentication and billing. Running these against a local ComfyUI instance will result in permission errors or logged-out states. There are two ways to connect to a cloud/staging backend:
+
+**Option 1: Frontend — `pnpm dev:cloud`**
+
+The simplest approach. This proxies all API requests to the test cloud environment:
+
+```bash
+pnpm dev:cloud
+```
+
+This sets `DEV_SERVER_COMFYUI_URL` to `https://testcloud.comfy.org/` automatically. You can also set this variable manually in your `.env` file to target a different environment:
+
+```bash
+# .env
+DEV_SERVER_COMFYUI_URL=https://stagingcloud.comfy.org/
+```
+
+Any `*.comfy.org` URL automatically enables cloud mode, which includes the GCS media proxy needed for viewing generated images and videos. See [.env_example](.env_example) for all available cloud URLs.
+
+**Option 2: Backend — `--comfy-api-base`**
+
+Alternatively, launch the ComfyUI backend pointed at the staging API:
+
+```bash
+python main.py --comfy-api-base https://stagingapi.comfy.org --verbose
+```
+
+Then run `pnpm dev` as usual. This keeps the frontend in local mode but routes backend API calls through staging.
+
 #### Access dev server on touch devices
 
 Enable remote access to the dev server by setting `VITE_REMOTE_DEV` in `.env` to `true`.

apps/hub/.gitignore

@@ -1,9 +0,0 @@
-dist/
-.astro/
-.content-cache/
-src/content/templates/
-public/workflows/thumbnails/
-public/workflows/avatars/
-public/previews/
-public/search-index.json
-knowledge/tutorials/

apps/hub/astro.config.mjs

@@ -1,254 +0,0 @@
-import { defineConfig } from 'astro/config'
-import sitemap from '@astrojs/sitemap'
-import vercel from '@astrojs/vercel'
-import tailwindcss from '@tailwindcss/vite'
-import fs from 'node:fs'
-import path from 'node:path'
-import os from 'node:os'
-
-import vue from '@astrojs/vue'
-
-// Build template date lookup at config time
-const templatesDir = path.join(process.cwd(), 'src/content/templates')
-const templateDates = new Map()
-
-if (fs.existsSync(templatesDir)) {
-  const files = fs.readdirSync(templatesDir).filter((f) => f.endsWith('.json'))
-  for (const file of files) {
-    try {
-      const content = JSON.parse(
-        fs.readFileSync(path.join(templatesDir, file), 'utf-8')
-      )
-      if (content.name && content.date) {
-        templateDates.set(content.name, content.date)
-      }
-    } catch {
-      // Skip invalid JSON files
-    }
-  }
-}
-
-// Build timestamp used as lastmod fallback for pages without a specific date
-const buildDate = new Date().toISOString()
-
-// Supported locales (matches src/i18n/config.ts)
-const locales = [
-  'en',
-  'zh',
-  'zh-TW',
-  'ja',
-  'ko',
-  'es',
-  'fr',
-  'ru',
-  'tr',
-  'ar',
-  'pt-BR'
-]
-const nonDefaultLocales = locales.filter((l) => l !== 'en')
-
-// Custom sitemap pages for ISR routes not discovered at build time
-const siteOrigin = (
-  process.env.PUBLIC_SITE_ORIGIN || 'https://www.comfy.org'
-).replace(/\/$/, '')
-
-// Creator profile pages — extract unique usernames from synced templates
-const creatorUsernames = new Set()
-if (fs.existsSync(templatesDir)) {
-  const files = fs.readdirSync(templatesDir).filter((f) => f.endsWith('.json'))
-  for (const file of files) {
-    try {
-      const content = JSON.parse(
-        fs.readFileSync(path.join(templatesDir, file), 'utf-8')
-      )
-      if (content.username) creatorUsernames.add(content.username)
-    } catch {
-      // Skip invalid JSON
-    }
-  }
-}
-
-const creatorPages = [...creatorUsernames].map(
-  (u) => `${siteOrigin}/workflows/${u}/`
-)
-const localeCustomPages = nonDefaultLocales.map(
-  (locale) => `${siteOrigin}/${locale}/workflows/`
-)
-const customPages = [...creatorPages, ...localeCustomPages]
-
-// https://astro.build/config
-export default defineConfig({
-  site: (process.env.PUBLIC_SITE_ORIGIN || 'https://www.comfy.org').replace(
-    /\/$/,
-    ''
-  ),
-  prefetch: {
-    prefetchAll: false,
-    defaultStrategy: 'hover'
-  },
-  i18n: {
-    defaultLocale: 'en',
-    locales: locales,
-    routing: {
-      prefixDefaultLocale: false // English at root, others prefixed (/zh/, /ja/, etc.)
-    }
-  },
-  integrations: [
-    sitemap({
-      // Use custom filename to avoid collision with Framer's /sitemap.xml
-      filenameBase: 'sitemap-workflows',
-      // Include Framer's marketing sitemap in the index
-      customSitemaps: ['https://www.comfy.org/sitemap.xml'],
-      // Include on-demand locale pages that aren't discovered at build time
-      customPages: customPages,
-      serialize(item) {
-        const url = new URL(item.url)
-        const pathname = url.pathname
-
-        // Template detail pages: /workflows/{slug}/ or /{locale}/workflows/{slug}/
-        const templateMatch = pathname.match(
-          /^(?:\/([a-z]{2}(?:-[A-Z]{2})?))?\/workflows\/([^/]+)\/?$/
-        )
-        if (templateMatch) {
-          const slug = templateMatch[2]
-          const date = templateDates.get(slug)
-          item.lastmod = date ? new Date(date).toISOString() : buildDate
-          // @ts-expect-error - sitemap types are stricter than actual API
-          item.changefreq = 'monthly'
-          item.priority = 0.8
-          return item
-        }
-
-        // Homepage
-        if (pathname === '/' || pathname === '') {
-          item.lastmod = buildDate
-          // @ts-expect-error - sitemap types are stricter than actual API
-          item.changefreq = 'daily'
-          item.priority = 1.0
-          return item
-        }
-
-        // Workflows index (including localized versions)
-        if (pathname.match(/^(?:\/[a-z]{2}(?:-[A-Z]{2})?)?\/workflows\/?$/)) {
-          item.lastmod = buildDate
-          // @ts-expect-error - sitemap types are stricter than actual API
-          item.changefreq = 'daily'
-          item.priority = 0.9
-          return item
-        }
-
-        // Category pages: /workflows/category/{type}/ or /{locale}/workflows/category/{type}/
-        if (
-          pathname.match(
-            /^(?:\/[a-z]{2}(?:-[A-Z]{2})?)?\/workflows\/category\//
-          )
-        ) {
-          // @ts-expect-error - sitemap types are stricter than actual API
-          item.changefreq = 'weekly'
-          item.priority = 0.7
-          return item
-        }
-
-        // Model pages: /workflows/model/{model}/ or /{locale}/workflows/model/{model}/
-        if (
-          pathname.match(/^(?:\/[a-z]{2}(?:-[A-Z]{2})?)?\/workflows\/model\//)
-        ) {
-          // @ts-expect-error - sitemap types are stricter than actual API
-          item.changefreq = 'weekly'
-          item.priority = 0.6
-          return item
-        }
-
-        // Tag pages: /workflows/tag/{tag}/ or /{locale}/workflows/tag/{tag}/
-        if (
-          pathname.match(/^(?:\/[a-z]{2}(?:-[A-Z]{2})?)?\/workflows\/tag\//)
-        ) {
-          // @ts-expect-error - sitemap types are stricter than actual API
-          item.changefreq = 'weekly'
-          item.priority = 0.6
-          return item
-        }
-
-        // Default for other pages
-        // @ts-expect-error - sitemap types are stricter than actual API
-        item.changefreq = 'weekly'
-        item.priority = 0.5
-        return item
-      },
-      // Exclude OG image routes and legacy redirect pages from sitemap.
-      // Legacy redirects are /workflows/{slug}/ without a 12-char hex share_id suffix.
-      // Canonical detail pages are /workflows/{slug}-{shareId}/ (shareId = 12 hex chars).
-      filter: (page) => {
-        if (
-          page.includes('/workflows/og/') ||
-          page.includes('/workflows/og.png')
-        )
-          return false
-        // Check if this is a workflow detail path (not category/tag/model/creators)
-        const match = page.match(/\/workflows\/([^/]+)\/$/)
-        if (match) {
-          const segment = match[1]
-          // Skip known sub-paths
-          if (
-            ['category', 'tag', 'model', 'creators'].some((p) =>
-              page.includes(`/workflows/${p}/`)
-            )
-          )
-            return true
-          // Include if it has a share_id suffix (12 hex chars after last hyphen)
-          const lastHyphen = segment.lastIndexOf('-')
-          if (lastHyphen === -1) return false // No hyphen = legacy redirect
-          const candidate = segment.slice(lastHyphen + 1)
-          if (candidate.length === 12 && /^[0-9a-f]+$/.test(candidate))
-            return true
-          return false // Has hyphen but not a valid share_id = legacy redirect
-        }
-        return true
-      }
-    }),
-    vue()
-  ],
-  output: 'static',
-  adapter: vercel({
-    webAnalytics: { enabled: true },
-    skewProtection: true
-  }),
-
-  // Build performance optimizations
-  build: {
-    // Increase concurrency for faster builds on multi-core systems
-    concurrency: Math.max(1, os.cpus().length),
-    // Inline small stylesheets automatically
-    inlineStylesheets: 'auto'
-  },
-
-  // HTML compression
-  compressHTML: true,
-
-  // Image optimization settings
-  image: {
-    service: {
-      entrypoint: 'astro/assets/services/sharp',
-      config: {
-        // Limit input pixels to prevent memory issues with large images
-        limitInputPixels: 268402689 // ~16384x16384
-      }
-    }
-  },
-
-  // Responsive images for automatic srcset generation (now stable in Astro 5)
-  // Note: responsiveImages was moved from experimental to stable in Astro 5.x
-
-  vite: {
-    plugins: [tailwindcss()],
-    build: {
-      chunkSizeWarningLimit: 1000
-    },
-    optimizeDeps: {
-      include: ['web-vitals']
-    },
-    css: {
-      devSourcemap: false
-    }
-  }
-})

apps/hub/knowledge/concepts/3d-generation.md

@@ -1,22 +0,0 @@
-# 3D Generation
-
-3D generation creates three-dimensional models — meshes, point clouds, or multi-view images — from text or image inputs. This enables rapid prototyping of 3D assets without manual modeling. In ComfyUI, several approaches exist: image-to-3D (lifting a single photo into a mesh), text-to-3D (generating a 3D object from a description), and multi-view generation (producing consistent views of an object that can be reconstructed into 3D).
-
-## How It Works in ComfyUI
-
-- Key nodes involved: Model-specific loaders (`TripoSR`, `InstantMesh`, `StableZero123`), `LoadImage`, `Save3D` / `Preview3D`, `CRM` nodes
-- Typical workflow pattern: Load image → Load 3D model → Run inference → Preview 3D result → Export mesh
-
-## Key Settings
-
-- **Inference steps**: Number of denoising/reconstruction steps. More steps generally improve quality but increase generation time.
-- **Elevation angle**: Camera elevation for multi-view generation, controlling the vertical viewing angle of the generated views.
-- **Guidance scale**: How closely the model follows the input image or text. Higher values increase fidelity to the input but may reduce diversity.
-- **Output format**: Export format for the 3D mesh — OBJ, GLB, and PLY are common options, each suited to different downstream tools.
-
-## Tips
-
-- Clean single-object images on white or simple backgrounds work best for image-to-3D conversion.
-- Multi-view approaches (like Zero123) often produce better geometry than single-view methods.
-- Post-process generated meshes in Blender for cleanup, retopology, or texturing before production use.
-- Start with TripoSR for quick results — it generates meshes in seconds and is a good baseline to compare against other methods.

apps/hub/knowledge/concepts/_template-index.json

@@ -1,374 +0,0 @@
-{
-  "text-to-image": [
-    "01_get_started_text_to_image",
-    "api_bfl_flux2_max_sofa_swap",
-    "api_bfl_flux_1_kontext_max_image",
-    "api_bfl_flux_1_kontext_multiple_images_input",
-    "api_bfl_flux_1_kontext_pro_image",
-    "api_bfl_flux_pro_t2i",
-    "api_bytedance_seedream4",
-    "api_flux2",
-    "api_from_photo_2_miniature",
-    "api_google_gemini_image",
-    "api_grok_text_to_image",
-    "api_ideogram_v3_t2i",
-    "api_kling_omni_image",
-    "api_luma_photon_i2i",
-    "api_luma_photon_style_ref",
-    "api_nano_banana_pro",
-    "api_openai_dall_e_2_inpaint",
-    "api_openai_dall_e_2_t2i",
-    "api_openai_dall_e_3_t2i",
-    "api_openai_fashion_billboard_generator",
-    "api_openai_image_1_i2i",
-    "api_openai_image_1_inpaint",
-    "api_openai_image_1_multi_inputs",
-    "api_openai_image_1_t2i",
-    "api_recraft_image_gen_with_color_control",
-    "api_recraft_image_gen_with_style_control",
-    "api_recraft_style_reference",
-    "api_recraft_vector_gen",
-    "api_runway_reference_to_image",
-    "api_runway_text_to_image",
-    "api_stability_ai_i2i",
-    "api_stability_ai_sd3.5_i2i",
-    "api_stability_ai_sd3.5_t2i",
-    "api_stability_ai_stable_image_ultra_t2i",
-    "api_wan_text_to_image",
-    "default",
-    "flux1_dev_uso_reference_image_gen",
-    "flux1_krea_dev",
-    "flux_canny_model_example",
-    "flux_depth_lora_example",
-    "flux_dev_checkpoint_example",
-    "flux_dev_full_text_to_image",
-    "flux_fill_inpaint_example",
-    "flux_fill_outpaint_example",
-    "flux_redux_model_example",
-    "flux_schnell",
-    "flux_schnell_full_text_to_image",
-    "hidream_e1_1",
-    "hidream_e1_full",
-    "hidream_i1_dev",
-    "hidream_i1_fast",
-    "hidream_i1_full",
-    "image-qwen_image_edit_2511_lora_inflation",
-    "image_anima_preview",
-    "image_chroma1_radiance_text_to_image",
-    "image_chroma_text_to_image",
-    "image_flux2",
-    "image_flux2_fp8",
-    "image_flux2_klein_image_edit_4b_base",
-    "image_flux2_klein_image_edit_4b_distilled",
-    "image_flux2_klein_image_edit_9b_base",
-    "image_flux2_klein_image_edit_9b_distilled",
-    "image_flux2_klein_text_to_image",
-    "image_flux2_text_to_image",
-    "image_flux2_text_to_image_9b",
-    "image_kandinsky5_t2i",
-    "image_lotus_depth_v1_1",
-    "image_netayume_lumina_t2i",
-    "image_newbieimage_exp0_1-t2i",
-    "image_omnigen2_image_edit",
-    "image_omnigen2_t2i",
-    "image_ovis_text_to_image",
-    "image_qwen_Image_2512",
-    "image_qwen_image",
-    "image_qwen_image_2512_with_2steps_lora",
-    "image_qwen_image_controlnet_patch",
-    "image_qwen_image_instantx_controlnet",
-    "image_qwen_image_instantx_inpainting_controlnet",
-    "image_qwen_image_union_control_lora",
-    "image_z_image",
-    "image_z_image_turbo",
-    "image_z_image_turbo_fun_union_controlnet",
-    "sd3.5_large_blur",
-    "sd3.5_large_canny_controlnet_example",
-    "sd3.5_large_depth",
-    "sd3.5_simple_example",
-    "sdxl_refiner_prompt_example",
-    "sdxl_revision_text_prompts",
-    "sdxl_simple_example",
-    "sdxlturbo_example",
-    "templates-9grid_social_media-v2.0"
-  ],
-  "img2img": [
-    "02_qwen_Image_edit_subgraphed",
-    "api_luma_photon_i2i",
-    "api_meshy_multi_image_to_model",
-    "api_openai_image_1_i2i",
-    "api_runway_reference_to_image",
-    "api_stability_ai_i2i",
-    "api_stability_ai_sd3.5_i2i",
-    "flux1_dev_uso_reference_image_gen",
-    "flux_canny_model_example",
-    "flux_depth_lora_example",
-    "flux_fill_inpaint_example",
-    "flux_fill_outpaint_example",
-    "flux_kontext_dev_basic",
-    "flux_redux_model_example",
-    "image_chrono_edit_14B",
-    "image_qwen_image_edit",
-    "image_qwen_image_edit_2509",
-    "image_qwen_image_instantx_controlnet",
-    "image_qwen_image_instantx_inpainting_controlnet",
-    "sd3.5_large_blur",
-    "sd3.5_large_canny_controlnet_example",
-    "sd3.5_large_depth"
-  ],
-  "inpainting": [
-    "api_openai_dall_e_2_inpaint",
-    "api_openai_image_1_inpaint",
-    "api_stability_ai_audio_inpaint",
-    "flux_fill_inpaint_example",
-    "flux_fill_outpaint_example",
-    "image_flux.1_fill_dev_OneReward",
-    "image_qwen_image_instantx_inpainting_controlnet",
-    "video_wan2_2_14B_fun_inpaint",
-    "video_wan2_2_5B_fun_inpaint",
-    "video_wan_vace_inpainting",
-    "wan2.1_fun_inp"
-  ],
-  "outpainting": [
-    "api_bria_image_outpainting",
-    "flux_fill_outpaint_example",
-    "image_flux.1_fill_dev_OneReward",
-    "video_wan_vace_outpainting"
-  ],
-  "controlnet": [
-    "02_qwen_Image_edit_subgraphed",
-    "flux_canny_model_example",
-    "flux_depth_lora_example",
-    "flux_redux_model_example",
-    "image_lotus_depth_v1_1",
-    "image_qwen_image_controlnet_patch",
-    "image_qwen_image_edit_2509",
-    "image_qwen_image_instantx_controlnet",
-    "image_qwen_image_instantx_inpainting_controlnet",
-    "image_qwen_image_union_control_lora",
-    "image_z_image_turbo_fun_union_controlnet",
-    "sd3.5_large_canny_controlnet_example",
-    "sd3.5_large_depth",
-    "utility-depthAnything-v2-relative-video",
-    "utility-frame_interpolation-film",
-    "utility-lineart-video",
-    "utility-normal_crafter-video",
-    "utility-openpose-video",
-    "video_ltx2_canny_to_video",
-    "video_ltx2_depth_to_video",
-    "video_ltx2_pose_to_video",
-    "wan2.1_fun_control"
-  ],
-  "upscaling": [
-    "api_topaz_image_enhance",
-    "api_topaz_video_enhance",
-    "api_wavespeed_flshvsr_video_upscale",
-    "api_wavespped_image_upscale",
-    "api_wavespped_seedvr2_ai_image_fix",
-    "ultility_hitpaw_general_image_enhance",
-    "ultility_hitpaw_video_enhance",
-    "utility-gan_upscaler",
-    "utility-topaz_landscape_upscaler",
-    "utility_interpolation_image_upscale",
-    "utility_nanobanana_pro_ai_image_fix",
-    "utility_nanobanana_pro_illustration_upscale",
-    "utility_nanobanana_pro_product_upscale",
-    "utility_recraft_creative_image_upscale",
-    "utility_recraft_crisp_image_upscale",
-    "utility_seedvr2_image_upscale",
-    "utility_seedvr2_video_upscale",
-    "utility_topaz_illustration_upscale",
-    "utility_video_upscale"
-  ],
-  "video-generation": [
-    "03_video_wan2_2_14B_i2v_subgraphed",
-    "api_bytedace_seedance1_5_flf2v",
-    "api_bytedace_seedance1_5_image_to_video",
-    "api_bytedace_seedance1_5_text_to_video",
-    "api_bytedance_flf2v",
-    "api_bytedance_image_to_video",
-    "api_bytedance_text_to_video",
-    "api_grok_video",
-    "api_grok_video_edit",
-    "api_hailuo_minimax_i2v",
-    "api_hailuo_minimax_t2v",
-    "api_hailuo_minimax_video",
-    "api_kling2_6_i2v",
-    "api_kling2_6_t2v",
-    "api_kling_effects",
-    "api_kling_flf",
-    "api_kling_i2v",
-    "api_kling_motion_control",
-    "api_kling_omni_edit_video",
-    "api_kling_omni_i2v",
-    "api_kling_omni_t2v",
-    "api_kling_omni_v2v",
-    "api_ltxv_image_to_video",
-    "api_ltxv_text_to_video",
-    "api_luma_i2v",
-    "api_luma_t2v",
-    "api_moonvalley_image_to_video",
-    "api_moonvalley_text_to_video",
-    "api_moonvalley_video_to_video_motion_transfer",
-    "api_moonvalley_video_to_video_pose_control",
-    "api_openai_sora_video",
-    "api_pixverse_i2v",
-    "api_pixverse_t2v",
-    "api_pixverse_template_i2v",
-    "api_runway_first_last_frame",
-    "api_runway_gen3a_turbo_image_to_video",
-    "api_runway_gen4_turo_image_to_video",
-    "api_topaz_video_enhance",
-    "api_veo2_i2v",
-    "api_veo3",
-    "api_vidu_image_to_video",
-    "api_vidu_q2_flf2v",
-    "api_vidu_q2_i2v",
-    "api_vidu_q2_r2v",
-    "api_vidu_q2_t2v",
-    "api_vidu_q3_image_to_video",
-    "api_vidu_q3_text_to_video",
-    "api_vidu_reference_to_video",
-    "api_vidu_start_end_to_video",
-    "api_vidu_text_to_video",
-    "api_vidu_video_extension",
-    "api_wan2_6_i2v",
-    "api_wan2_6_t2v",
-    "api_wan_image_to_video",
-    "api_wan_r2v",
-    "api_wan_text_to_video",
-    "api_wavespeed_flshvsr_video_upscale",
-    "gsc_starter_2",
-    "hunyuan_video_text_to_video",
-    "image_to_video_wan",
-    "ltxv_image_to_video",
-    "ltxv_text_to_video",
-    "template-Animation_Trajectory_Control_Wan_ATI",
-    "templates-3D_logo_texture_animation",
-    "templates-6-key-frames",
-    "templates-car_product",
-    "templates-photo_to_product_vid",
-    "templates-sprite_sheet",
-    "templates-stitched_vid_contact_sheet",
-    "templates-textured_logo_elements",
-    "templates-textured_logotype-v2.1",
-    "text_to_video_wan",
-    "txt_to_image_to_video",
-    "ultility_hitpaw_video_enhance",
-    "utility-depthAnything-v2-relative-video",
-    "utility-frame_interpolation-film",
-    "utility-gan_upscaler",
-    "utility-lineart-video",
-    "utility-normal_crafter-video",
-    "utility-openpose-video",
-    "utility_seedvr2_video_upscale",
-    "utility_video_upscale",
-    "video-wan21_scail",
-    "video_humo",
-    "video_hunyuan_video_1.5_720p_i2v",
-    "video_hunyuan_video_1.5_720p_t2v",
-    "video_kandinsky5_i2v",
-    "video_kandinsky5_t2v",
-    "video_ltx2_canny_to_video",
-    "video_ltx2_depth_to_video",
-    "video_ltx2_i2v",
-    "video_ltx2_i2v_distilled",
-    "video_ltx2_pose_to_video",
-    "video_ltx2_t2v",
-    "video_ltx2_t2v_distilled",
-    "video_wan2.1_alpha_t2v_14B",
-    "video_wan2.1_fun_camera_v1.1_1.3B",
-    "video_wan2.1_fun_camera_v1.1_14B",
-    "video_wan2_1_infinitetalk",
-    "video_wan2_2_14B_animate",
-    "video_wan2_2_14B_flf2v",
-    "video_wan2_2_14B_fun_camera",
-    "video_wan2_2_14B_fun_control",
-    "video_wan2_2_14B_fun_inpaint",
-    "video_wan2_2_14B_i2v",
-    "video_wan2_2_14B_s2v",
-    "video_wan2_2_14B_t2v",
-    "video_wan2_2_5B_fun_control",
-    "video_wan2_2_5B_fun_inpaint",
-    "video_wan2_2_5B_ti2v",
-    "video_wan_ati",
-    "video_wan_vace_14B_ref2v",
-    "video_wan_vace_14B_t2v",
-    "video_wan_vace_14B_v2v",
-    "video_wan_vace_flf2v",
-    "video_wan_vace_inpainting",
-    "video_wan_vace_outpainting",
-    "video_wanmove_480p",
-    "video_wanmove_480p_hallucination",
-    "wan2.1_flf2v_720_f16",
-    "wan2.1_fun_control",
-    "wan2.1_fun_inp"
-  ],
-  "audio-generation": [
-    "05_audio_ace_step_1_t2a_song_subgraphed",
-    "api_kling2_6_i2v",
-    "api_kling2_6_t2v",
-    "api_stability_ai_audio_inpaint",
-    "api_stability_ai_audio_to_audio",
-    "api_stability_ai_text_to_audio",
-    "api_vidu_q3_image_to_video",
-    "api_vidu_q3_text_to_video",
-    "audio-chatterbox_tts",
-    "audio-chatterbox_tts_dialog",
-    "audio-chatterbox_tts_multilingual",
-    "audio-chatterbox_vc",
-    "audio_ace_step_1_5_checkpoint",
-    "audio_ace_step_1_5_split",
-    "audio_ace_step_1_5_split_4b",
-    "audio_ace_step_1_m2m_editing",
-    "audio_ace_step_1_t2a_instrumentals",
-    "audio_ace_step_1_t2a_song",
-    "audio_stable_audio_example",
-    "utility-audioseparation",
-    "video_wan2_1_infinitetalk",
-    "video_wan2_2_14B_s2v"
-  ],
-  "3d-generation": [
-    "04_hunyuan_3d_2.1_subgraphed",
-    "3d_hunyuan3d-v2.1",
-    "3d_hunyuan3d_image_to_model",
-    "3d_hunyuan3d_multiview_to_model",
-    "3d_hunyuan3d_multiview_to_model_turbo",
-    "api_from_photo_2_miniature",
-    "api_hunyuan3d_image_to_model",
-    "api_hunyuan3d_text_to_model",
-    "api_meshy_image_to_model",
-    "api_meshy_multi_image_to_model",
-    "api_meshy_text_to_model",
-    "api_rodin_gen2",
-    "api_rodin_image_to_model",
-    "api_rodin_multiview_to_model",
-    "api_tripo3_0_image_to_model",
-    "api_tripo3_0_text_to_model",
-    "api_tripo_image_to_model",
-    "api_tripo_multiview_to_model",
-    "api_tripo_text_to_model",
-    "templates-3D_logo_texture_animation",
-    "templates-qwen_multiangle"
-  ],
-  "lora": [
-    "flux_depth_lora_example",
-    "image-qwen_image_edit_2511_lora_inflation",
-    "image_qwen_image_2512_with_2steps_lora",
-    "image_qwen_image_union_control_lora"
-  ],
-  "embeddings": [],
-  "ip-adapter": [
-    "api_kling_omni_i2v",
-    "api_kling_omni_image",
-    "api_kling_omni_v2v",
-    "api_magnific_image_style_transfer",
-    "api_recraft_style_reference",
-    "api_vidu_q2_r2v",
-    "api_wan_r2v",
-    "templates-product_ad-v2.0"
-  ],
-  "samplers": [],
-  "cfg": [],
-  "vae": []
-}

apps/hub/knowledge/concepts/audio-generation.md

@@ -1,22 +0,0 @@
-# Audio Generation
-
-Audio generation in ComfyUI covers creating speech (text-to-speech), music, and sound effects from text prompts or reference audio. Dedicated audio models run within ComfyUI's node graph, letting you integrate audio creation into larger multimedia workflows — for example, generating a video and its soundtrack in a single pipeline.
-
-## How It Works in ComfyUI
-
-- Key nodes involved: Model-specific nodes (`CosyVoice` nodes for TTS, `StableAudio` nodes for music/SFX), audio preview and save nodes, `AudioScheduler`
-- Typical workflow pattern: Load audio model → Provide text/reference input → Generate audio → Preview/save audio
-
-## Key Settings
-
-- **Sample rate**: Output audio quality, typically 24000–48000 Hz. Higher rates capture more detail but produce larger files.
-- **Duration**: Length of generated audio in seconds. Longer durations may reduce quality or coherence depending on the model.
-- **Voice reference**: For voice cloning, a short audio clip of the target voice (3–10 seconds of clean speech works best).
-- **Text input**: The text to be spoken (TTS) or the description of the desired audio (music/SFX generation).
-
-## Tips
-
-- CosyVoice and F5-TTS are popular choices for text-to-speech in ComfyUI, each with dedicated custom nodes.
-- Stable Audio Open handles music and sound effect generation from text descriptions.
-- Use clean, noise-free reference audio clips for voice cloning to get the best results.
-- Keep text inputs short and well-punctuated for the highest quality speech output — long paragraphs may degrade in naturalness.

apps/hub/knowledge/concepts/cfg.md

@@ -1,23 +0,0 @@
-# CFG / Guidance Scale
-
-Classifier-Free Guidance (CFG) controls how strongly the model follows your text prompt versus generating freely. Higher CFG values produce outputs that adhere more closely to the prompt but can cause oversaturation and artifacts, while lower values yield more natural-looking images at the cost of reduced prompt control. Finding the right balance is essential for every workflow.
-
-## How It Works in ComfyUI
-
-- Key nodes: `KSampler` (the `cfg` parameter), `ModelSamplingDiscrete` (for advanced noise schedule configurations)
-- During each sampling step, the model generates both a conditioned prediction (with your prompt) and an unconditioned prediction (without it). CFG scales the difference between the two — higher values push the output further toward the conditioned prediction, amplifying prompt influence.
-
-## Key Settings
-
-- **cfg** (1.0–30.0): The guidance scale value. Recommended ranges vary by model architecture:
-  - SD 1.5 / SDXL: 7–8 is the standard starting point
-  - Flux: 1.0–4.0 (Flux uses much lower guidance)
-  - Video models (e.g., Wan, HunyuanVideo): 3.5–5.0
-
-## Tips
-
-- Start at 7 for SD-based models and 3.5 for Flux, then adjust based on results
-- Values above ~12 for SD models typically cause color oversaturation, harsh contrast, and visible artifacts
-- Values below ~3 for SD models tend to produce blurry or incoherent results
-- Some models like Flux Schnell use a guidance embedding baked into the model rather than traditional CFG — for these, the `cfg` parameter may have little or no effect
-- When experimenting, change CFG in increments of 0.5–1.0 to see its impact clearly

apps/hub/knowledge/concepts/controlnet.md

@@ -1,28 +0,0 @@
-# ControlNet
-
-ControlNet guides image generation using structural conditions extracted from reference images — such as edge maps, depth information, or human poses. Instead of relying solely on text prompts for composition, ControlNet lets you specify the spatial layout precisely. This bridges the gap between text-to-image flexibility and the structural precision needed for professional workflows.
-
-## How It Works in ComfyUI
-
-- Key nodes involved: `ControlNetLoader`, `ControlNetApplyAdvanced`, preprocessor nodes (`CannyEdgePreprocessor`, `DepthAnythingPreprocessor`, `DWPosePreprocessor`, `LineartPreprocessor`)
-- Typical workflow pattern: Load reference image → preprocess to extract condition (edges/depth/pose) → load ControlNet model → apply condition to sampling → generate image with structural guidance
-
-## ControlNet Types
-
-- **Canny**: Detects edges to preserve outlines and shapes
-- **Depth**: Captures spatial depth for accurate foreground/background placement
-- **OpenPose**: Extracts human body and hand poses for character positioning
-- **Normal Map**: Encodes surface orientation for consistent lighting and geometry
-- **Lineart**: Follows line drawings and illustrations as generation guides
-- **Scribble**: Uses rough sketches as loose compositional guides
-
-## Key Settings
-
-- **Strength**: Controls how strongly the condition guides generation (0.0–1.0). Values of 0.5–1.0 are typical. Higher values enforce the structure more rigidly; lower values allow the model more creative freedom.
-- **start_percent / end_percent**: Controls when the ControlNet activates during the sampling process. Starting at 0.0 and ending at 1.0 applies guidance throughout. Ending earlier (e.g., 0.8) lets the model refine fine details freely in final steps.
-
-## Tips
-
-- Always preprocess your input image with the appropriate preprocessor node before feeding it to ControlNet. Raw images will not produce correct conditioning.
-- Combine multiple ControlNets for precise control — for example, Depth for spatial layout plus OpenPose for character positioning. Stack them by chaining `ControlNetApplyAdvanced` nodes.
-- If your generation looks distorted or overcooked, lower the ControlNet strength. Values above 0.8 can fight with the text prompt and produce artifacts.

apps/hub/knowledge/concepts/embeddings.md

@@ -1,19 +0,0 @@
-# Textual Embeddings
-
-Textual embeddings are learned text representations that encode specific concepts, styles, or objects into the CLIP text encoder's vocabulary. These tiny files (~10–100 KB) effectively add new "words" to your prompt vocabulary, letting you reference complex visual concepts — a particular art style, a specific character, or a set of undesirable artifacts — with a single token. Because they operate at the text-encoding level, embeddings integrate seamlessly with your existing prompts and require no changes to the model itself.
-
-## How It Works in ComfyUI
-
-- Key nodes: `CLIPTextEncode` — reference embeddings directly in your prompt text using the syntax `embedding:name_of_embedding`
-- Typical workflow pattern: Place embedding files in `ComfyUI/models/embeddings/` → type `embedding:name_of_embedding` inside your positive or negative prompt in a `CLIPTextEncode` node → connect to sampler as usual
-
-## Key Settings
-
-- **Prompt weighting**: Embeddings have no dedicated strength slider, but you can adjust their influence with prompt weighting syntax, e.g., `(embedding:name_of_embedding:1.2)` to increase strength or `(embedding:name_of_embedding:0.6)` to soften it
-- **Placement**: Add embeddings to the negative prompt to suppress unwanted features, or to the positive prompt to invoke a learned concept
-
-## Tips
-
-- Embeddings are commonly used in negative prompts (e.g., `embedding:EasyNegative`, `embedding:bad-hands-5`) to reduce common artifacts like malformed hands or distorted faces
-- Make sure the embedding matches your base model version — an SD 1.5 embedding will not work correctly with an SDXL checkpoint
-- You can combine multiple embeddings with regular text in the same prompt for fine-grained control

apps/hub/knowledge/concepts/img2img.md

@@ -1,20 +0,0 @@
-# Image-to-Image
-
-Image-to-image (img2img) transforms an existing image using a text prompt while preserving the original structure and composition. Instead of starting from pure noise, the source image is encoded into latent space and partially noised, then the sampler denoises it guided by your prompt. This lets you restyle photos, refine AI-generated images, or apply creative modifications while keeping the overall layout intact.
-
-## How It Works in ComfyUI
-
-- Key nodes involved: `LoadImage`, `VAEEncode`, `CLIPTextEncode` (positive + negative), `KSampler`, `VAEDecode`, `SaveImage`
-- Typical workflow pattern: Load source image → encode to latent with VAE → encode text prompts → sample with partial denoise → decode latent to image → save
-
-## Key Settings
-
-- **Denoise Strength**: The most important setting, ranging from 0.0 to 1.0. Lower values (0.2–0.4) preserve more of the original image with subtle changes. Higher values (0.6–0.8) allow more creative freedom but deviate further from the source. A value of 1.0 effectively ignores the input image entirely.
-- **Steps**: Number of sampling steps. 20–30 is typical. Fewer steps may be sufficient at low denoise values since less transformation is needed.
-- **CFG Scale**: Controls prompt adherence, same as text-to-image. 7–8 is a standard starting point.
-
-## Tips
-
-- Start with a denoise strength of 0.5 and adjust up or down based on how much change you want. This gives a balanced mix of original structure and new content.
-- Your input image resolution should match the model's training resolution. Resize or crop your source image to 512×512 (SD 1.5) or 1024×1024 (SDXL) before loading to avoid quality issues.
-- Use img2img iteratively: generate an initial text-to-image result, then refine it with img2img at low denoise to fix details without losing the overall composition.

apps/hub/knowledge/concepts/inpainting.md

@@ -1,21 +0,0 @@
-# Inpainting
-
-Inpainting selectively regenerates parts of an image using a mask while leaving the rest untouched. You paint a mask over the area you want to change, provide a text prompt describing the desired replacement, and the model fills in only the masked region. This is essential for fixing defects, replacing objects, or refining specific details in an otherwise finished image.
-
-## How It Works in ComfyUI
-
-- Key nodes involved: `LoadImage`, `VAEEncodeForInpainting`, `CLIPTextEncode` (positive + negative), `KSampler`, `VAEDecode`, `SaveImage`
-- Typical workflow pattern: Load image + mask → encode with inpainting-aware VAE node → encode text prompts → sample → decode → save
-- The mask can be created using ComfyUI's built-in mask editor or loaded from an external image
-
-## Key Settings
-
-- **grow_mask_by**: Expands the mask by a number of pixels, helping the regenerated area blend smoothly with the surrounding image. 6–8 pixels is typical. Too little causes visible seams; too much affects areas you wanted to keep.
-- **Denoise Strength**: For inpainting, higher values (0.7–1.0) generally work best since you want the masked region to be fully regenerated. Lower values may produce inconsistent blending.
-- **Checkpoint**: Dedicated inpainting models like `512-inpainting-ema` produce significantly better edge blending than standard checkpoints.
-
-## Tips
-
-- Always expand your mask slightly beyond the target area. Tight masks create hard edges that look unnatural against the surrounding image.
-- Describe what you want to appear in the masked region, not what you want to remove. For example, prompt "a clear blue sky" rather than "remove the bird."
-- Use inpainting-specific checkpoints whenever possible. Standard models can inpaint but often struggle with seamless blending at mask boundaries.

apps/hub/knowledge/concepts/ip-adapter.md

@@ -1,21 +0,0 @@
-# IP-Adapter
-
-IP-Adapter (Image Prompt Adapter) uses reference images to guide generation style, composition, or subject instead of — or alongside — text prompts. Rather than describing what you want in words, you show the model an image, enabling "image prompting." This is especially powerful for transferring artistic style, maintaining character consistency across generations, or conveying visual concepts that are difficult to express in text.
-
-## How It Works in ComfyUI
-
-- Key nodes: `IPAdapterModelLoader`, `IPAdapterApply` (or `IPAdapterAdvanced`), `CLIPVisionLoader`, `CLIPVisionEncode`, `PrepImageForClipVision`
-- Typical workflow pattern: Load IP-Adapter model + CLIP Vision model → prepare and encode reference image → apply adapter to the main model → connect to sampler → decode
-
-## Key Settings
-
-- **weight** (0.0–1.0): Controls the influence of the reference image on the output. A range of 0.5–0.8 is typical; higher values make the output closer to the reference
-- **weight_type**: Determines how the reference is interpreted — `standard` for general use, `style transfer` for artistic style without copying content, `composition` for layout guidance
-- **start_at / end_at** (0.0–1.0): Controls when the adapter is active during sampling. Limiting the range (e.g., 0.0–0.8) can improve prompt responsiveness while retaining reference influence
-
-## Tips
-
-- Use the `style_transfer` weight type when you want to borrow an artistic style without reproducing the reference image's content
-- Combine IP-Adapter with a text prompt for the best results — the text adds detail and specificity on top of the visual guidance
-- Face-specific IP-Adapter models (e.g., `ip-adapter-faceid`) exist for portrait consistency across multiple generations
-- Lower the weight if your output looks too similar to the reference image

apps/hub/knowledge/concepts/lora.md

@@ -1,20 +0,0 @@
-# LoRA
-
-LoRA (Low-Rank Adaptation) is a technique for fine-tuning a base model's behavior using a small add-on file rather than retraining the entire model. LoRAs adjust a model's style, teach it specific subjects, or introduce new concepts — all in a file typically just 10–200 MB, compared to multi-gigabyte full checkpoints. This makes them easy to share, swap, and combine. In ComfyUI, you load LoRAs on top of a checkpoint and control how strongly they influence the output.
-
-## How It Works in ComfyUI
-
-- Key nodes involved: `LoraLoader` (loads one LoRA and applies it to both MODEL and CLIP), `LoraLoaderModelOnly` (applies to MODEL only, skipping CLIP for faster loading)
-- Typical workflow pattern: Load checkpoint → LoraLoader (attach LoRA) → CLIP Text Encode → KSampler → VAE Decode. Chain multiple `LoraLoader` nodes to stack LoRAs.
-
-## Key Settings
-
-- **strength_model**: Controls how much the LoRA affects the diffusion model. Range 0.0–1.0; typical values are 0.6–1.0. Higher values apply the LoRA effect more strongly.
-- **strength_clip**: Controls how much the LoRA affects text encoding. Usually set to the same value as strength_model, but can be adjusted independently for fine control.
-
-## Tips
-
-- Start with strength 0.7 and adjust up or down based on results — too high can cause oversaturation or artifacts.
-- Stacking too many LoRAs simultaneously can cause visual artifacts or conflicting styles; two or three is usually a safe limit.
-- Ensure the LoRA matches your base model architecture — SD 1.5 LoRAs will not work with SDXL checkpoints, and vice versa.
-- Many LoRAs require specific trigger words in your prompt to activate; always check the LoRA's documentation or model card.

apps/hub/knowledge/concepts/outpainting.md

@@ -1,20 +0,0 @@
-# Outpainting
-
-Outpainting extends an image beyond its original borders, generating new content that seamlessly continues the existing scene. Unlike inpainting which replaces content within an image, outpainting adds content outside the frame — expanding the canvas in any direction. This is useful for changing aspect ratios, adding environmental context, or creating panoramic compositions from a single image.
-
-## How It Works in ComfyUI
-
-- Key nodes involved: `LoadImage`, `ImagePadForOutpaint`, `VAEEncodeForInpainting`, `CLIPTextEncode` (positive + negative), `KSampler`, `VAEDecode`, `SaveImage`
-- Typical workflow pattern: Load image → pad image with transparent/noised borders → encode with inpainting VAE node (padded area becomes the mask) → encode text prompts → sample → decode → save
-
-## Key Settings
-
-- **Padding Pixels**: The number of pixels to extend on each side, typically 64–256. Smaller increments produce more coherent results since the model has more context relative to the new area.
-- **Denoise Strength**: Use high values (0.8–1.0) for outpainted regions since the padded area is essentially blank and needs full generation.
-- **Feathering**: Controls the gradient blend between the original image and the new content. Higher feathering values create smoother transitions and reduce visible seams.
-
-## Tips
-
-- Outpaint in stages rather than all at once. Extending by 128 pixels at a time and iterating produces far more coherent results than trying to add 512 pixels in a single pass.
-- Use a lower CFG scale (5–6) for outpainting. This allows the model to generate more natural, context-aware extensions rather than forcing strict prompt adherence that may clash with the existing image.
-- Include scene context in your prompt that matches the original image. If the source shows an indoor room, describe the room's style and lighting so the extension feels continuous.

apps/hub/knowledge/concepts/samplers.md

@@ -1,21 +0,0 @@
-# Samplers & Schedulers
-
-Samplers are the algorithms that iteratively denoise a random latent into a coherent image, while schedulers control the noise schedule — how much noise is removed at each step. Together they determine the image's quality, speed, and visual character. Choosing the right combination is one of the most impactful decisions in any generation workflow.
-
-## How It Works in ComfyUI
-
-- Key nodes: `KSampler` (main sampling node), `KSamplerAdvanced` (provides control over start/end steps for multi-pass workflows)
-- Typical workflow pattern: Load model → connect conditioning → configure sampler/scheduler/steps → sample → decode
-
-## Key Settings
-
-- **sampler_name**: The denoising algorithm. Common choices include `euler` (fast, good baseline), `euler_ancestral` (more creative variation), `dpmpp_2m` (balanced quality and speed), `dpmpp_2m_sde` (high quality, slightly slower), `dpmpp_3m_sde` (very high quality), and `uni_pc` (fast convergence)
-- **scheduler**: Controls the noise reduction curve. `normal` is linear, `karras` front-loads noise reduction for better detail, `exponential` and `sgm_uniform` (recommended for SDXL) are also available
-- **steps** (1–100): Number of denoising iterations. 20–30 is typical; more steps give diminishing returns. Flux and LCM models need far fewer (4–8 steps)
-
-## Tips
-
-- `euler` + `normal` is the safest starting combination for any model
-- `dpmpp_2m` + `karras` is a popular choice when you want higher quality with minimal speed cost
-- Ancestral samplers (`euler_ancestral`, any `_sde` variant) produce different results each run even with the same seed — useful for exploration, but not for reproducibility
-- Flux and LCM models converge much faster; using 20+ steps with them wastes time without improving quality

apps/hub/knowledge/concepts/text-to-image.md

@@ -1,21 +0,0 @@
-# Text-to-Image Generation
-
-Text-to-image is the foundational workflow in ComfyUI: you provide a text description (prompt) and the system generates an image from scratch. This is the starting point for most generative AI art. A diffusion model iteratively denoises a random latent image, guided by your text prompt encoded through CLIP, to produce a coherent image matching your description.
-
-## How It Works in ComfyUI
-
-- Key nodes involved: `CheckpointLoaderSimple`, `CLIPTextEncode` (positive + negative), `EmptyLatentImage`, `KSampler`, `VAEDecode`, `SaveImage`
-- Typical workflow pattern: Load checkpoint → encode text prompts → create empty latent → sample → decode latent to image → save
-
-## Key Settings
-
-- **Resolution**: Must match the model's training resolution. Use 512×512 for SD 1.5, 1024×1024 for SDXL and Flux models. Mismatched resolutions produce artifacts like duplicated limbs or distorted compositions.
-- **Steps**: Number of denoising iterations. 20–30 steps is a good balance between quality and speed. More steps refine details but with diminishing returns beyond 30.
-- **CFG Scale**: Controls how strongly the sampler follows your prompt. 7–8 is the typical range. Higher values increase prompt adherence but can introduce oversaturation or artifacts.
-- **Seed**: Determines the initial random noise. A fixed seed produces reproducible results, which is useful for iterating on prompts while keeping composition consistent.
-
-## Tips
-
-- Start with simple, descriptive prompts before adding stylistic modifiers. Complex prompts can conflict and produce muddy results.
-- Use the negative prompt `CLIPTextEncode` to specify what you want to avoid (e.g., "blurry, low quality, deformed hands") — this significantly improves output quality.
-- Always match your `EmptyLatentImage` resolution to the model you loaded. A 768×768 image on an SD 1.5 checkpoint will produce noticeably worse results than 512×512.

apps/hub/knowledge/concepts/upscaling.md

@@ -1,21 +0,0 @@
-# Upscaling
-
-Upscaling increases image resolution while adding detail, turning a small generated image into a large, sharp result. In ComfyUI, there are two main approaches: model-based upscaling, which uses trained AI models (like RealESRGAN or 4x-UltraSharp) to intelligently enlarge an image in one pass, and latent-based upscaling, which works in latent space with a KSampler to add new detail during the enlargement process. Model-based is faster, while latent-based offers more creative control.
-
-## How It Works in ComfyUI
-
-- Key nodes involved: `UpscaleModelLoader`, `ImageUpscaleWithModel`, `ImageScaleBy`, `LatentUpscale`, `VAEDecodeTiled`
-- Typical workflow pattern: Generate image → Upscale model loader → ImageUpscaleWithModel → Save image (model-based), or Generate latent → LatentUpscale → KSampler (lower denoise) → VAEDecode → Save image (latent-based)
-
-## Key Settings
-
-- **Upscale model**: The AI model used for model-based upscaling. `RealESRGAN_x4plus` is a reliable general-purpose choice; `4x-UltraSharp` excels at photo-realistic detail.
-- **Scale factor**: How much to enlarge — 2x and 4x are typical. Higher factors increase VRAM usage significantly.
-- **tile_size**: For tiled decoding/encoding of very large images. Range 512–1024; smaller tiles use less VRAM but take longer.
-
-## Tips
-
-- Model-based upscaling is faster but less creative; latent upscaling paired with a KSampler adds genuinely new detail.
-- Use `VAEDecodeTiled` for very large images to avoid out-of-memory errors.
-- Chain two 2x upscales instead of one 4x for better overall quality.
-- When using latent upscaling, set KSampler denoise to 0.3–0.5 to add detail without changing the composition.

apps/hub/knowledge/concepts/vae.md

@@ -1,20 +0,0 @@
-# VAE (Variational Autoencoder)
-
-The VAE encodes pixel images into a compact latent representation and decodes latents back into pixel images. All diffusion in Stable Diffusion and Flux happens in latent space — the VAE is the bridge between the images you see and the mathematical space where the model actually works. Every generation workflow ends with a VAE decode step to produce a viewable image.
-
-## How It Works in ComfyUI
-
-- Key nodes: `VAEDecode` (latent → image), `VAEEncode` (image → latent), `VAEDecodeTiled` (for large images to avoid out-of-memory errors), `VAELoader` (load a standalone VAE file)
-- Typical workflow pattern: Most checkpoints include a built-in VAE, so the `VAEDecode` node can pull directly from the loaded checkpoint. To use a different VAE, add a `VAELoader` node and connect it to `VAEDecode` instead.
-
-## Key Settings
-
-- **tile_size** (for `VAEDecodeTiled`): Size of each tile when decoding in chunks. Default is 512; reduce if you still encounter memory issues
-- **VAE choice**: VAE files are model-specific. Use `sdxl_vae.safetensors` for SDXL, `ae.safetensors` for Flux. Place files in `ComfyUI/models/vae/`
-
-## Tips
-
-- If colors look washed out or slightly off, try loading an external VAE — the VAE baked into a checkpoint is not always optimal, especially for community fine-tunes
-- Use `VAEDecodeTiled` for images larger than ~2048 px on either side to prevent out-of-memory crashes
-- SDXL and Flux each have their own VAE architecture — using the wrong one will produce corrupted output
-- When doing img2img or inpainting, the `VAEEncode` node converts your input image into the latent space the model expects

apps/hub/knowledge/concepts/video-generation.md

@@ -1,22 +0,0 @@
-# Video Generation
-
-Video generation creates video content from text prompts (T2V), reference images (I2V), or existing video (V2V) using specialized video diffusion models. Unlike image generation, video models must maintain temporal coherence across frames, ensuring smooth motion and consistent subjects. ComfyUI supports several leading open-source video models including WAN 2.1 and HunyuanVideo, each with their own loader and latent nodes.
-
-## How It Works in ComfyUI
-
-- Key nodes involved: Model-specific loaders (e.g. `WAN` video nodes, `HunyuanVideo` nodes, `LTXVLoader`), `EmptyHunyuanLatentVideo` / `EmptyLTXVLatentVideo`, `KSampler`, `VHS_VideoCombine` (from Video Helper Suite)
-- Typical workflow pattern: Load video model → Create empty video latent → KSampler (with video-aware scheduling) → VAE decode → VHS_VideoCombine → Save video
-
-## Key Settings
-
-- **Frame count**: Number of frames to generate. Typically 16–81 frames depending on the model; more frames require more VRAM and time.
-- **Resolution**: Often 512×320 or 848×480 for T2V. Higher resolutions need significantly more resources.
-- **FPS**: Frames per second for output, typically 8–24. Higher FPS gives smoother motion but requires more frames for the same duration.
-- **Motion scale/strength**: Controls the amount of movement in the generated video. Lower values produce subtle motion; higher values produce more dynamic scenes.
-
-## Tips
-
-- Start with fewer frames and lower resolution to test your prompt and settings before committing to a full-quality render.
-- Image-to-video (I2V) typically gives better coherence than text-to-video (T2V) because the model has a visual anchor.
-- Video Helper Suite (VHS) nodes are essential for loading, previewing, and saving video — install this custom node pack first.
-- WAN 2.1 and HunyuanVideo are currently the leading open models for quality video generation in ComfyUI.

apps/hub/knowledge/models/_aliases.json

@@ -1,88 +0,0 @@
-{
-  "Wan": "wan",
-  "Wan2.1": "wan",
-  "Wan2.2": "wan",
-  "Wan2.5": "wan",
-  "Wan2.6": "wan",
-  "Wan-Move": "wan",
-  "Motion Control": "wan",
-  "Flux": "flux",
-  "Flux.2": "flux",
-  "Flux.2 Dev": "flux",
-  "Flux.2 Klein": "flux",
-  "Kontext": "flux",
-  "BFL": "flux",
-  "SDXL": "sdxl",
-  "SD1.5": "sdxl",
-  "Stability": "sdxl",
-  "Reimagine": "sdxl",
-  "SD3.5": "sd3-5",
-  "SVD": "svd",
-  "Stable Audio": "stable-audio",
-  "Google": "gemini",
-  "Google Gemini": "gemini",
-  "Google Gemini Image": "gemini",
-  "Gemini3 Pro Image Preview": "gemini",
-  "Gemini-2.5-Flash": "gemini",
-  "Veo": "veo",
-  "Nano Banana Pro": "nano-banana-pro",
-  "nano-banana": "nano-banana-pro",
-  "OpenAI": "gpt-image-1",
-  "GPT-Image-1": "gpt-image-1",
-  "GPT-Image-1.5": "gpt-image-1",
-  "Qwen": "qwen",
-  "Qwen-Image": "qwen",
-  "Qwen-Image-Edit": "qwen",
-  "Qwen-Image-Layered": "qwen",
-  "Qwen-Image 2512": "qwen",
-  "Hunyuan Video": "hunyuan",
-  "Hunyuan3D": "hunyuan",
-  "Tencent": "hunyuan",
-  "LTX-2": "ltx-video",
-  "LTXV": "ltx-video",
-  "Lightricks": "ltx-video",
-  "ByteDance": "seedance",
-  "Seedance": "seedance",
-  "Seedream": "seedream",
-  "Seedream 4.0": "seedream",
-  "SeedVR2": "seedvr2",
-  "Vidu": "vidu",
-  "Vidu Q2": "vidu",
-  "Vidu Q3": "vidu",
-  "Kling": "kling",
-  "Kling O1": "kling",
-  "Kling2.6": "kling",
-  "ACE-Step": "ace-step",
-  "Chatter Box": "chatterbox",
-  "Recraft": "recraft",
-  "Runway": "runway",
-  "Luma": "luma",
-  "HiDream": "hidream",
-  "Tripo": "tripo",
-  "MiniMax": "minimax",
-  "Z-Image-Turbo": "z-image",
-  "Z-Image": "z-image",
-  "Grok": "grok",
-  "Moonvalley": "moonvalley",
-  "Topaz": "topaz",
-  "Kandinsky": "kandinsky",
-  "OmniGen": "omnigen",
-  "Magnific": "magnific",
-  "PixVerse": "pixverse",
-  "Meshy": "meshy",
-  "Rodin": "rodin",
-  "WaveSpeed": "wavespeed",
-  "Chroma": "chroma",
-  "BRIA": "bria",
-  "HitPaw": "hitpaw",
-  "NewBie": "newbie",
-  "Ovis-Image": "ovis-image",
-  "Ideogram": "ideogram",
-  "Anima": "anima",
-  "ChronoEdit": "chronoedit",
-  "Nvidia": "chronoedit",
-  "HuMo": "humo",
-  "FlashVSR": "flashvsr",
-  "Real-ESRGAN": "real-esrgan",
-  "Depth Anything\u00a0v2": "depth-anything-v2"
-}

apps/hub/knowledge/models/ace-step.md

@@ -1,47 +0,0 @@
-# ACE-Step
-
-ACE-Step is a foundation model for music generation developed by ACE Studio and StepFun. It uses diffusion-based generation with a Deep Compression AutoEncoder (DCAE) and a lightweight linear transformer to achieve state-of-the-art speed and musical coherence.
-
-## Model Variants
-
-### ACE-Step (3.5B)
-
-- 3.5B parameter diffusion model
-- DCAE encoder with linear transformer conditioning
-- 27 or 60 inference steps recommended
-- Apache 2.0 license
-
-## Key Features
-
-- 15x faster than LLM-based baselines (20 seconds for a 4-minute song on A100)
-- Full-song generation with lyrics and structure
-- Duration control for variable-length output
-- Music remixing and style transfer
-- Lyrics editing and vocal synthesis
-- Supports 16+ languages including English, Chinese, Japanese, Korean, French, German, Spanish, and more
-- Text-to-music from natural language descriptions
-
-## Hardware Requirements
-
-- RTX 3090: 12.76x real-time factor at 27 steps
-- RTX 4090: 34.48x real-time factor at 27 steps
-- NVIDIA A100: 27.27x real-time factor at 27 steps
-- Apple M2 Max: 2.27x real-time factor at 27 steps
-- Higher step counts (60) reduce speed by roughly half
-
-## Common Use Cases
-
-- Original music generation from text descriptions
-- Song remixing and style transfer
-- Lyrics-based music creation
-- Multi-language vocal music generation
-- Rapid music prototyping for content creators
-- Background music and soundtrack generation
-
-## Key Parameters
-
-- **steps**: Inference steps (27 for speed, 60 for quality)
-- **duration**: Target audio length in seconds (up to ~5 minutes)
-- **lyrics**: Song lyrics text input for vocal generation
-- **prompt**: Natural language description of desired music style and mood
-- **seed**: Random seed for reproducible generation (results are seed-sensitive)

apps/hub/knowledge/models/anima.md

@@ -1,46 +0,0 @@
-# Anima
-
-Anima is an API-based AI video generation platform that creates animated video content from text prompts, supporting character consistency and storyboard-driven workflows.
-
-## Model Variants
-
-### Anima Video Generation
-
-- Cloud-based video generation service
-- Supports multiple underlying AI models (Runway, Kling, Minimax, Luma)
-- Integrated text, image, and audio generation pipeline
-
-## Key Features
-
-- AI character generation with persistent identity across scenes
-- Storyboard-based workflow: script to visual scenes with narration
-- Multi-model integration (GPT-4, Claude, Gemini for text; FLUX, MidJourney for images)
-- Voice generation via ElevenLabs integration
-- Music composition via Suno integration
-- Autopilot mode for fully automated video creation
-- Prompt enhancement for optimized output quality
-- Template library for rapid content creation
-- Scene-by-scene generation with character consistency
-
-## Hardware Requirements
-
-- No local hardware required (cloud-based service)
-- Runs entirely through web API
-- Browser-based interface for interactive use
-
-## Common Use Cases
-
-- Animated story series production
-- Movie trailer and concept video creation
-- Kids bedtime story animation
-- Lofi music video generation
-- Marketing and explainer video content
-- Storyboard visualization
-
-## Key Parameters
-
-- **prompt**: Text description of the scene or story
-- **character**: Selected or generated character for identity consistency
-- **style**: Visual style preset (animation, cinematic, etc.)
-- **duration**: Target video length
-- **resolution**: Output video resolution

apps/hub/knowledge/models/bria.md

@@ -1,48 +0,0 @@
-# BRIA AI
-
-BRIA AI is an enterprise-focused visual generative AI platform that trains its models exclusively on licensed, ethically sourced data, ensuring commercially safe outputs with full IP indemnification.
-
-## Model Variants
-
-### BRIA Fibo
-
-- Flagship hyper-controllable text-to-image model
-- JSON-based control framework with 100+ disentangled visual attributes
-- Supports lighting, depth, color, composition, and camera control
-- Ideal for agentic workflows and enterprise-scale creative automation
-
-### BRIA Text-to-Image Lite
-
-- Fully private, self-hosted deployment of the Fibo pipeline
-- Designed for regulated industries requiring total data control
-- Runs on-premises with no external data transfer
-
-## Key Features
-
-- Trained on 100% licensed data from 20+ partners including Getty Images
-- Full IP indemnification for commercial use
-- Tri-layer content moderation for brand-safe outputs
-- Patented attribution engine compensating data owners by usage
-- ControlNet support for canny, depth, recoloring, and IP Adapter
-- Multilingual prompt support
-- Fine-tuning API for brand-specific customization
-
-## Hardware Requirements
-
-- Cloud-hosted API available (no local GPU required)
-- Self-hosted Lite version supports deployment on AWS and Azure
-- Open-source weights available on Hugging Face for local inference
-
-## Common Use Cases
-
-- Enterprise marketing and advertising content
-- E-commerce product photography
-- Brand-consistent visual asset generation
-- Storyboarding and concept art for media production
-
-## Key Parameters
-
-- **prompt**: Text description of desired image
-- **style**: Photorealistic, illustrative, or custom styles
-- **guidance_methods**: ControlNet canny, depth, recoloring, IP Adapter
-- **resolution**: Multiple aspect ratios supported

apps/hub/knowledge/models/chatterbox.md

@@ -1,52 +0,0 @@
-# Chatterbox
-
-Chatterbox is a family of state-of-the-art open-source text-to-speech models developed by Resemble AI, featuring zero-shot voice cloning and emotion control.
-
-## Model Variants
-
-### Chatterbox Turbo
-
-- 350M parameters, single-step mel decoding for low latency
-- Paralinguistic tags for non-speech sounds ([laugh], [cough], [chuckle])
-- English only, optimized for voice agents and production use
-
-### Chatterbox (Original)
-
-- 500M parameter Llama backbone, English only
-- CFG and exaggeration control for emotion intensity
-
-### Chatterbox Multilingual
-
-- 500M parameters, 23 languages (Arabic, Chinese, French, German, Hindi, Japanese, Korean, Spanish, and more)
-- Zero-shot voice cloning across languages
-
-## Key Features
-
-- Zero-shot voice cloning from a few seconds of reference audio
-- Emotion exaggeration control (first open-source model with this feature)
-- Built-in PerTh neural watermarking for responsible AI
-- Sub-200ms latency for real-time applications
-- Trained on 500K hours of cleaned speech data
-- MIT license (free for commercial use)
-- Outperforms ElevenLabs in subjective evaluations
-
-## Hardware Requirements
-
-- Minimum: NVIDIA GPU with CUDA support
-- Turbo model requires less VRAM than original due to smaller architecture
-- Runs on consumer GPUs (RTX 3060 and above)
-- CPU inference possible but significantly slower
-
-## Common Use Cases
-
-- Voice cloning for content creation
-- AI voice agents and assistants
-- Audiobook narration
-- Game and media dialogue generation
-
-## Key Parameters
-
-- **exaggeration**: Emotion intensity control (0.0 to 1.0, default 0.5)
-- **cfg_weight**: Classifier-free guidance weight (0.0 to 1.0, default 0.5)
-- **audio_prompt_path**: Path to reference audio clip for voice cloning
-- **language_id**: Language code for multilingual model (e.g., "fr", "zh", "ja")

apps/hub/knowledge/models/chroma.md

@@ -1,50 +0,0 @@
-# Chroma
-
-Chroma is an open-source 8.9 billion parameter text-to-image model based on the FLUX.1-schnell architecture, developed by Lodestone Rock and the community. It is fully Apache 2.0 licensed.
-
-## Model Variants
-
-### Chroma
-
-- 8.9B parameter model based on FLUX.1-schnell
-- Trained on a curated 5M sample dataset (from 20M candidates)
-- Apache 2.0 license for unrestricted use
-- Supports both tag-based and natural language prompting
-
-### Chroma XL
-
-- Experimental merge and fine-tune based on NoobAI-XL (SDXL architecture)
-- Low CFG (2.5-3.0) and low step count (8-12 steps)
-- Optimized for fast generation on consumer hardware
-
-## Key Features
-
-- Fully open-source with Apache 2.0 licensing
-- Diverse training data spanning anime, artistic, and photographic styles
-- Community-driven development with public training logs
-- Compatible with FLUX ecosystem (VAE, T5 text encoder)
-- ComfyUI workflow support
-- LoRA and fine-tuning compatible
-- GGUF quantized versions available for lower VRAM
-
-## Hardware Requirements
-
-- Base model: 24GB VRAM recommended (BF16)
-- Q8_0 quantized: ~13GB VRAM
-- Q4_0 quantized: ~7GB VRAM
-- Requires FLUX.1 VAE and T5 text encoder
-
-## Common Use Cases
-
-- Open-source text-to-image generation
-- Artistic and stylized image creation
-- Community model fine-tuning and experimentation
-- LoRA training for custom styles and characters
-
-## Key Parameters
-
-- **prompt**: Text description or tag-based prompt
-- **steps**: Inference steps (15-30 recommended)
-- **cfg_scale**: Guidance scale (1-4, model uses low CFG)
-- **resolution**: Output resolution (1024x1024 default)
-- **guidance**: Flux-style guidance parameter (around 4)

apps/hub/knowledge/models/chronoedit.md

@@ -1,58 +0,0 @@
-# ChronoEdit
-
-ChronoEdit is an image editing framework by NVIDIA that reframes editing as a video generation task, using temporal reasoning to ensure physically plausible and consistent edits.
-
-## Model Variants
-
-### ChronoEdit-14B
-
-- Full 14 billion parameter model for maximum quality
-- Built on pretrained video diffusion model architecture
-- Requires ~34GB VRAM (38GB with temporal reasoning enabled)
-
-### ChronoEdit-2B
-
-- Compact 2 billion parameter variant for efficiency
-- Maintains core temporal reasoning capabilities
-- Lower VRAM requirements for broader hardware compatibility
-
-### ChronoEdit-14B 8-Step Distilled LoRA
-
-- Distilled variant requiring only 8 inference steps
-- Faster generation with minimal quality loss
-- Uses flow-shift 2.0 and guidance-scale 1.0
-
-## Key Features
-
-- Treats image editing as a video generation task for temporal consistency
-- Temporal reasoning tokens simulate intermediate editing trajectories
-- Ensures physically plausible edits (object interactions, lighting, shadows)
-- Two-stage pipeline: temporal reasoning stage followed by editing frame generation
-- Prompt enhancer integration for improved editing instructions
-- LoRA fine-tuning support via DiffSynth-Studio
-- Upscaler LoRA available for super-resolution editing
-- PaintBrush LoRA for sketch-to-object editing
-- Apache-2.0 license
-
-## Hardware Requirements
-
-- 14B model: 34GB VRAM minimum (38GB with temporal reasoning)
-- 2B model: 12GB+ VRAM estimated
-- Supports model offloading to reduce peak VRAM
-- Linux only (not supported on Windows/macOS)
-
-## Common Use Cases
-
-- Physically consistent image editing (add/remove/modify objects)
-- World simulation for autonomous driving and robotics
-- Visualizing editing trajectories and reasoning
-- Image super-resolution via upscaler LoRA
-- Sketch-to-object conversion via PaintBrush LoRA
-
-## Key Parameters
-
-- **prompt**: Text description of the desired edit
-- **num_inference_steps**: Denoising steps (default ~50, or 8 with distilled LoRA)
-- **guidance_scale**: Prompt adherence strength (default ~7.5, or 1.0 with distilled LoRA)
-- **flow_shift**: Flow matching shift parameter (2.0 for distilled LoRA)
-- **enable_temporal_reasoning**: Toggle temporal reasoning stage for better consistency

apps/hub/knowledge/models/depth-anything-v2.md

@@ -1,60 +0,0 @@
-# Depth Anything V2
-
-Depth Anything V2 is a monocular depth estimation model trained on 595K synthetic labeled images and 62M+ real unlabeled images, providing robust relative depth maps from single images.
-
-## Model Variants
-
-### Depth-Anything-V2-Small
-
-- Lightweight variant for fast inference
-- ViT-S (Small) encoder backbone
-- Suitable for real-time applications
-
-### Depth-Anything-V2-Base
-
-- Mid-range variant balancing speed and accuracy
-- ViT-B (Base) encoder backbone
-
-### Depth-Anything-V2-Large
-
-- High-accuracy variant for detailed depth maps
-- ViT-L (Large) encoder backbone with 256 output features
-- Recommended for most production use cases
-
-### Depth-Anything-V2-Giant
-
-- Maximum accuracy variant
-- ViT-G (Giant) encoder backbone
-- Highest computational requirements
-
-## Key Features
-
-- More fine-grained depth detail than Depth Anything V1
-- More robust than V1 and Stable Diffusion-based alternatives (Marigold, Geowizard)
-- 10× faster than SD-based depth estimation models
-- Trained on large-scale synthetic + real data mixture
-- Produces relative (not metric) depth maps by default
-- DPT (Dense Prediction Transformer) decoder architecture
-
-## Hardware Requirements
-
-- Small: 2GB VRAM minimum
-- Base: 4GB VRAM minimum
-- Large: 6GB VRAM recommended
-- Giant: 12GB+ VRAM recommended
-- CPU inference supported for smaller variants
-
-## Common Use Cases
-
-- Depth map generation for compositing and VFX
-- ControlNet depth conditioning for image generation
-- 3D scene understanding and reconstruction
-- Foreground/background separation
-- Augmented reality occlusion
-- Video depth estimation for parallax effects
-
-## Key Parameters
-
-- **encoder**: Model size variant (vits, vitb, vitl, vitg)
-- **input_size**: Processing resolution (higher = more detail, more VRAM)
-- **output_type**: Raw depth array or normalized visualization

apps/hub/knowledge/models/flashvsr.md

@@ -1,50 +0,0 @@
-# FlashVSR
-
-FlashVSR is a diffusion-based streaming video super-resolution framework that achieves near real-time 4× upscaling through one-step inference with locality-constrained sparse attention.
-
-## Model Variants
-
-### FlashVSR v1
-
-- Initial release of the one-step streaming VSR model
-- Built on Wan2.1 1.3B video diffusion backbone
-- 4× super-resolution optimized
-
-### FlashVSR v1.1
-
-- Enhanced stability and fidelity over v1
-- Improved artifact handling across different aspect ratios
-- Recommended for production use
-
-## Key Features
-
-- One-step diffusion inference (no multi-step denoising required)
-- Streaming architecture with KV cache for sequential frame processing
-- Locality-Constrained Sparse Attention (LCSA) prevents artifacts at high resolutions
-- Tiny Conditional Decoder (TC Decoder) achieves 7× faster decoding than standard WanVAE
-- Three-stage distillation pipeline from multi-step to single-step inference
-- Runs at ~17 FPS for 768×1408 videos on a single A100 GPU
-- Up to 12× speedup over prior one-step diffusion VSR models
-- Scales reliably to ultra-high resolutions
-
-## Hardware Requirements
-
-- Minimum: 24GB VRAM (A100 or similar recommended)
-- Optimized for NVIDIA A100 GPUs
-- Significant VRAM required for high-resolution video processing
-- Multi-GPU inference not required but beneficial for throughput
-
-## Common Use Cases
-
-- Real-world video upscaling to 4K
-- AI-generated video enhancement and artifact removal
-- Long video super-resolution with temporal consistency
-- Streaming video quality improvement
-- Restoring compressed or low-resolution video footage
-
-## Key Parameters
-
-- **scale**: Upscaling factor (4× recommended for best results)
-- **tile_size**: Spatial tiling for memory management (0 = auto)
-- **input_resolution**: Source video resolution (outputs 4× larger)
-- **model_version**: v1 or v1.1 checkpoint selection

apps/hub/knowledge/models/flux.md

@@ -1,98 +0,0 @@
-# Flux
-
-Flux is a family of state-of-the-art text-to-image and image editing models developed by Black Forest Labs (BFL).
-
-## Model Variants
-
-### Flux.1 Schnell
-
-- Ultra-fast inference (1-4 steps)
-- 12B parameter rectified flow transformer
-- Apache 2.0 license (open source)
-- Best for rapid prototyping and real-time applications
-
-### Flux.1 Dev
-
-- High-quality 12B parameter development model
-- 20-50 steps for best results
-- Non-commercial license for research
-- Guidance-distilled for efficient generation
-
-### Flux.1 Pro
-
-- Highest quality Flux.1 outputs via commercial API
-- Best prompt adherence and detail
-
-### Flux.2 Dev
-
-- 32B parameter rectified flow transformer
-- Unified text-to-image, single-reference editing, and multi-reference editing
-- No fine-tuning needed for character/object/style reference
-- Up to 4MP photorealistic output with improved autoencoder
-- Non-commercial license; quantized versions available for consumer GPUs
-
-### Flux.2 Klein
-
-- Fastest Flux model family — sub-second inference on modern hardware
-- **Klein 4B**: ~8GB VRAM, Apache 2.0 license, ideal for edge deployment
-- **Klein 9B**: Best quality-to-latency ratio, non-commercial license
-- Base (undistilled) variants available for fine-tuning and LoRA training
-- Supports text-to-image, single-reference editing, and multi-reference editing
-
-### Flux.1 Kontext
-
-- In-context image generation and editing via text instructions
-- Available as Kontext Max (premium), Pro (API), and Dev (open-weights, 12B)
-- Character consistency across multiple scenes without fine-tuning
-- Typography manipulation and local editing within images
-
-### Flux.1 Fill
-
-- Dedicated inpainting and outpainting model
-- Maintains consistency with surrounding image context
-- Available as Fill Pro (API) and Fill Dev (open-weights)
-
-### Flux Redux / Canny / Depth
-
-- **Redux**: Image variation generation from reference images
-- **Canny**: Edge-detection-based structural conditioning
-- **Depth**: Depth-map-based structural conditioning for pose/layout control
-
-## Key Features
-
-- Excellent text rendering in images
-- Strong prompt following and instruction adherence
-- High resolution output (up to 4MP with Flux.2)
-- Multi-reference editing: combine up to 6 reference images
-- Consistent style and quality across generations
-
-## Hardware Requirements
-
-- Flux.2 Klein 4B: ~8GB VRAM (consumer GPUs like RTX 4070)
-- Flux.2 Klein 9B: ~20GB VRAM
-- Flux.1 models: 12GB VRAM minimum (fp16), 24GB recommended
-- Flux.2 Dev: 64GB+ VRAM native, FP8 quantized ~40GB
-- Quantized and weight-streaming options available for lower VRAM cards
-
-## Common Use Cases
-
-- Text-to-image generation
-- Iterative image editing via text instructions
-- Character-consistent multi-scene generation
-- Inpainting and outpainting
-- Style transfer and image variation
-- Structural conditioning (canny, depth)
-
-## Key Parameters
-
-- **steps**: 1-4 (Schnell/Klein distilled), 20-50 (Dev/Base)
-- **guidance_scale**: 3.5-4.0 typical for Flux.2, 3.5 for Flux.1
-- **resolution**: Up to 2048x2048 (Flux.1), up to 4MP (Flux.2)
-- **seed**: For reproducible generation
-- **prompt_upsampling**: Optional LLM-based prompt enhancement (Flux.2)
-
-## Blog References
-
-- [FLUX.2 Day-0 Support in ComfyUI](../blog/flux2-day-0-support.md) — FLUX.2 with 4MP output, multi-reference consistency, professional text rendering
-- [FLUX.2 [klein] 4B & 9B](../blog/flux2-klein-4b.md) — Fastest Flux models, sub-second inference, unified generation and editing
-- [The Complete AI Upscaling Handbook](../blog/upscaling-handbook.md) — Benchmarks for upscaling workflows

apps/hub/knowledge/models/flux.summary.md

@@ -1 +0,0 @@
-Flux is Black Forest Labs' family of text-to-image and image editing models. The lineup includes Flux.1 Schnell (ultra-fast, 1-4 steps, Apache 2.0), Flux.1 Dev (high-quality, 20-50 steps, non-commercial), Flux.1 Pro (commercial API), and the newer Flux.2 Dev (32B parameters, up to 4MP output, multi-reference editing without fine-tuning). Flux.2 Klein offers sub-second inference in 4B (~8GB VRAM, Apache 2.0) and 9B variants. Specialized models include Kontext (in-context editing, character consistency), Fill (inpainting/outpainting), Redux (image variations), and Canny/Depth (structural conditioning). Flux excels at text rendering in images, strong prompt adherence, and consistent multi-scene generation. VRAM ranges from ~8GB (Klein 4B) to 64GB+ (Flux.2 Dev native), with quantized options available. Key parameters: guidance_scale 3.5-4.0, resolution up to 4MP for Flux.2. Primary uses include text-to-image, iterative editing, style transfer, and structural conditioning.

apps/hub/knowledge/models/gemini.md

@@ -1,75 +0,0 @@
-# Gemini
-
-Gemini is Google DeepMind's multimodal AI model family with native image generation, editing, and video generation capabilities, accessible in ComfyUI through API nodes.
-
-## Model Variants
-
-### Gemini 3 Pro Image Preview
-
-- Most capable Gemini image model with advanced reasoning
-- Complex multi-turn image generation and editing
-- Up to 14 input images, native 4K output
-- Also known as Nano Banana Pro
-- Model ID: `gemini-3-pro-image-preview`
-
-### Gemini 2.5 Flash Image
-
-- Cost-effective image generation optimized for speed and low latency
-- Character consistency, multi-image fusion, and prompt-based editing
-- $0.039 per image (1290 output tokens per image)
-- Model ID: `gemini-2.5-flash-image`
-
-### Google Gemini (General)
-
-- Multimodal model for text, image understanding, and generation
-- Interleaved text-and-image output in conversational context
-- Supports image input for analysis and editing tasks
-
-### Veo 2
-
-- Text-to-video and image-to-video generation
-- 8-second video clips at 720p resolution
-- Realistic physics simulation and cinematic styles
-- Supports 16:9 and 9:16 aspect ratios
-- Model ID: `veo-2.0-generate-001`
-
-### Veo 3 / 3.1
-
-- Latest video generation with native audio (dialogue, SFX, ambient)
-- Up to 1080p and 4K resolution (Veo 3.1)
-- Style reference images for aesthetic control
-- 4, 6, or 8-second video duration options
-
-## Key Features
-
-- Native multimodal generation: text, images, and video in one model family
-- World knowledge from Google Search for factually accurate image generation
-- SynthID invisible watermarking on all generated content
-- Multi-image fusion and character consistency across generations
-- Clean text rendering across multiple languages
-- Prompt-based image editing without masks or complex workflows
-
-## Hardware Requirements
-
-- No local GPU required — all models accessed via cloud API
-- Available through ComfyUI API nodes, Google AI Studio, and Vertex AI
-- Requires API key and network access
-
-## Common Use Cases
-
-- Text-to-image and image editing via API nodes
-- Multi-turn conversational image generation
-- Video generation from text prompts or reference images
-- Product animation and social media video content
-- Style-consistent character and brand asset generation
-- Text rendering and translation in images
-
-## Key Parameters
-
-- **prompt**: Text description for generation or editing
-- **aspect_ratio**: 1:1, 3:4, 4:3, 9:16, 16:9, 21:9 (images); 16:9, 9:16 (video)
-- **temperature**: 0.0-2.0 (default 1.0 for image models)
-- **durationSeconds**: 4-8 seconds for Veo models
-- **sampleCount**: 1-4 output videos per request
-- **seed**: Integer for reproducible generation
-- **personGeneration**: Safety control — `allow_adult`, `dont_allow`, or `allow_all`

apps/hub/knowledge/models/gpt-image-1.md

@@ -1,62 +0,0 @@
-# GPT-Image-1
-
-GPT-Image-1 is OpenAI's natively multimodal image generation model, capable of generating and editing images from text and image inputs. It is accessed in ComfyUI through API nodes.
-
-## Model Variants
-
-### GPT-Image-1.5
-
-- Latest and most advanced GPT Image model
-- Best overall quality with superior instruction following
-- High input fidelity for the first 5 input images
-- Supports generate vs. edit action control
-- Multi-turn editing via the Responses API
-
-### GPT-Image-1
-
-- Production-grade image generation and editing
-- High input fidelity for the first input image
-- Supports up to 16 input images for editing
-- Up to 10 images per generation request
-
-### GPT-Image-1-Mini
-
-- Cost-effective variant for lower quality requirements
-- Same API surface as GPT-Image-1
-- Suitable for rapid prototyping and high-volume workloads
-
-## Key Features
-
-- Superior text rendering in generated images
-- Real-world knowledge for accurate depictions
-- Transparent background support (PNG and WebP)
-- Mask-based inpainting with prompt guidance
-- Multi-image editing: combine up to 16 reference images
-- Streaming partial image output during generation
-- Content moderation with adjustable strictness
-
-## Hardware Requirements
-
-- No local GPU required — cloud API service via OpenAI
-- Accessed through ComfyUI API nodes
-- Requires OpenAI API key and organization verification
-
-## Common Use Cases
-
-- Text-to-image generation with detailed prompts
-- Image editing and compositing from multiple references
-- Product photography and mockup generation
-- Inpainting with mask-guided editing
-- Transparent asset generation (stickers, logos, icons)
-- Multi-turn iterative image refinement
-
-## Key Parameters
-
-- **prompt**: Text description up to 32,000 characters
-- **size**: `1024x1024`, `1536x1024` (landscape), `1024x1536` (portrait), or `auto`
-- **quality**: `low`, `medium`, `high`, or `auto` (affects cost and detail)
-- **n**: Number of images to generate (1-10)
-- **background**: `transparent`, `opaque`, or `auto`
-- **output_format**: `png`, `jpeg`, or `webp`
-- **moderation**: `auto` (default) or `low` (less restrictive)
-- **input_fidelity**: `low` (default) or `high` for preserving input image details

apps/hub/knowledge/models/grok.md

@@ -1,56 +0,0 @@
-# Grok (Aurora)
-
-Aurora is xAI's autoregressive image generation model integrated into Grok, excelling at photorealistic rendering and precise text instruction following.
-
-## Model Variants
-
-### grok-2-image-1212
-
-- API-accessible image generation model
-- Generates multiple images from text prompts
-- $0.07 per generated image
-- OpenAI and Anthropic SDK compatible
-
-### Aurora (Consumer)
-
-- Autoregressive mixture-of-experts network
-- Trained on billions of text and image examples
-- Available via Grok on X platform, web, iOS, and Android
-
-### Grok Imagine
-
-- Video and image generation model
-- State-of-the-art quality across cost and latency
-- API available since January 2026
-
-## Key Features
-
-- Photorealistic image generation from text prompts
-- Precise text rendering within images
-- Accurate rendering of real-world entities, logos, and text
-- Image editing via uploaded photos with text instructions
-- Multi-image generation per request
-- Native multimodal input support
-
-## Hardware Requirements
-
-- Cloud API-based (no local GPU required)
-- All generation runs on xAI infrastructure
-- API access via console.x.ai
-
-## Common Use Cases
-
-- Photorealistic image generation
-- Text and logo rendering in images
-- Image editing and style transfer
-- Meme and social media content creation
-- Product visualization
-- Character and portrait generation
-
-## Key Parameters
-
-- **prompt**: Text description of desired image
-- **model**: Model identifier (grok-2-image-1212)
-- **n**: Number of images to generate
-- **response_format**: Output format (url or b64_json)
-- **size**: Image dimensions

apps/hub/knowledge/models/hidream.md

@@ -1,55 +0,0 @@
-# HiDream-I1
-
-HiDream-I1 is a 17B parameter image generation foundation model by HiDream.ai that achieves state-of-the-art quality using a sparse diffusion transformer architecture.
-
-## Model Variants
-
-### HiDream-I1 Full
-
-- Full 17B parameter sparse diffusion transformer
-- Uses Llama-3.1-8B-Instruct and T5-XXL as text encoders
-- VAE from FLUX.1 Schnell, MIT license
-
-### HiDream-I1 Dev
-
-- Distilled variant, faster inference with minor quality tradeoff
-
-### HiDream-I1 Fast
-
-- Further distilled for maximum speed, best for rapid prototyping
-
-### HiDream-E1
-
-- Instruction-based image editing model
-
-## Key Features
-
-- State-of-the-art HPS v2.1 score (33.82), surpassing Flux.1-dev, DALL-E 3, and Midjourney V6
-- Best-in-class prompt following on GenEval (0.83) and DPG-Bench (85.89)
-- Multiple output styles: photorealistic, cartoon, artistic, and more
-- Dual text encoding with Llama-3.1-8B-Instruct and T5-XXL for strong prompt adherence
-- MIT license for commercial use
-- Requires Flash Attention for optimal performance
-
-## Hardware Requirements
-
-- Minimum: 24GB VRAM (Full model), Dev and Fast variants run on lower VRAM
-- Recommended: 40GB+ VRAM for Full model at high resolution
-- CUDA 12.4+ recommended for Flash Attention
-- Llama-3.1-8B-Instruct weights downloaded automatically
-
-## Common Use Cases
-
-- High-fidelity text-to-image generation
-- Photorealistic image creation
-- Artistic and stylized illustrations
-- Instruction-based image editing (E1 variant)
-- Commercial image generation
-
-## Key Parameters
-
-- **model_type**: Variant selection (full, dev, fast)
-- **steps**: Inference steps (varies by variant; fewer for fast/dev)
-- **cfg_scale**: Guidance scale for prompt adherence
-- **resolution**: Output image dimensions
-- **prompt**: Detailed text description of desired image

apps/hub/knowledge/models/hitpaw.md

@@ -1,51 +0,0 @@
-# HitPaw
-
-HitPaw is an AI-powered visual enhancement platform providing image and video upscaling, restoration, and denoising through dedicated API services and desktop applications.
-
-## Model Variants
-
-### HitPaw Image Enhancer
-
-- AI-powered photo enhancement with super-resolution up to 8x
-- Face Clear Model: dual-model portrait upscaling (2x and 4x)
-- Face Natural Model: texture-preserving portrait enhancement
-- General Enhance Model: super-resolution for scenes and objects
-- High Fidelity Model: premium upscaling for DSLR and AIGC images
-- Generative Portrait/Enhance Models: diffusion-based restoration for heavily compressed images
-
-### HitPaw Video Enhancer (VikPea)
-
-- Frame-aware video restoration and ultra HD upscaling
-- Face Soft Model: face-optimized noise and blur reduction
-- Portrait Restore Model: multi-frame fusion for facial detail
-- General Restore Model: GAN-based restoration for broad scenarios
-- Ultra HD Model: premium upscaling from HD to ultra HD
-- Generative Model: diffusion-driven repair for low-resolution video
-
-## Key Features
-
-- One-click portrait and scene enhancement
-- Dual-model face and background processing pipelines
-- Batch processing and API access for automated workflows
-- Support for 30+ video input formats and 5 export formats
-- Multi-frame face restoration for temporal consistency in video
-- Denoising models for mobile and camera images
-
-## Hardware Requirements
-
-- Cloud API available (no local GPU required)
-- Desktop apps for Windows, Mac, Android, and iOS
-- API integration via HTTP-based interface
-
-## Common Use Cases
-
-- Upscaling AI-generated images to publication quality
-- Restoring old or low-resolution photos and videos
-- Enhancing portrait and landscape photography
-- Video quality improvement for content creators
-
-## Key Parameters
-
-- **model**: Select enhancement model per content type
-- **scale**: 2x or 4x super-resolution options
-- **format**: Output format selection (mp4, mov, mkv, m4v, avi for video)

apps/hub/knowledge/models/humo.md

@@ -1,47 +0,0 @@
-# HuMo
-
-HuMo is a human-centric video generation model by ByteDance that produces videos from collaborative multi-modal conditioning using text, image, and audio inputs.
-
-## Model Variants
-
-### HuMo (Wan2.1-T2V-1.3B based)
-
-- Built on the Wan2.1-T2V-1.3B video foundation model
-- Supports Text+Image (TI), Text+Audio (TA), and Text+Image+Audio (TIA) modes
-- Two-stage training: subject preservation then audio-visual sync
-
-## Key Features
-
-- Multi-modal conditioning: text, reference images, and audio simultaneously
-- Subject identity preservation from reference images across frames
-- Audio-driven lip synchronization with facial expression alignment
-- Focus-by-predicting strategy for facial region attention during audio sync
-- Time-adaptive guidance dynamically adjusts input weights across denoising steps
-- Minimal-invasive image injection maintains base model prompt understanding
-- Progressive two-stage training separates identity learning from audio sync
-- Supports text-controlled appearance editing while preserving identity
-
-## Hardware Requirements
-
-- Minimum: 24GB VRAM (RTX 3090/4090 or similar)
-- Multi-GPU inference supported via FSDP and sequence parallelism
-- Whisper-large-v3 audio encoder required for audio modes
-- Optional audio separator for cleaner speech input
-
-## Common Use Cases
-
-- Digital avatar and virtual presenter creation
-- Audio-driven talking head generation
-- Character-consistent video clips from reference photos
-- Lip-synced dialogue video from audio tracks
-- Prompted reenactment with identity preservation
-- Text-controlled outfit and style changes on consistent subjects
-
-## Key Parameters
-
-- **mode**: Generation mode (TI, TA, or TIA)
-- **scale_t**: Text guidance strength (default: 7.5)
-- **scale_a**: Audio guidance strength (default: 2.0)
-- **frames**: Number of output frames (97 at 25 FPS = ~4 seconds)
-- **height/width**: Output resolution (480p or 720p supported)
-- **steps**: Denoising steps (30-50 recommended)

apps/hub/knowledge/models/hunyuan.md

@@ -1,75 +0,0 @@
-# Hunyuan
-
-Hunyuan is Tencent's family of open-source generative models spanning text-to-image, text-to-video, and 3D asset generation.
-
-## Model Variants
-
-### Hunyuan-DiT
-
-- Text-to-image diffusion transformer with native Chinese and English support
-- 1.5B parameter DiT architecture, native 1024x1024 resolution
-- Bilingual text encoder for strong CJK text rendering in images
-- v1.2 is the latest version with improved quality
-
-### HunyuanVideo
-
-- Large-scale text-to-video and image-to-video generation model
-- 13B+ parameters, the largest open-source video generation model
-- Dual-stream to single-stream transformer architecture with full attention
-- MLLM text encoder (decoder-only LLM) for better instruction following
-- Causal 3D VAE with 4x temporal, 8x spatial, 16x channel compression
-- Generates 720p video (1280x720) at up to 129 frames (~5s at 24fps)
-- FP8 quantized weights available to reduce memory by ~10GB
-- Outperforms Runway Gen-3, Luma 1.6 in professional evaluations
-- 3 workflow templates available
-
-### Hunyuan3D 2.0
-
-- Image-to-3D and text-to-3D asset generation system
-- Two-stage pipeline: Hunyuan3D-DiT (shape) + Hunyuan3D-Paint (texture)
-- Flow-based diffusion transformer for geometry generation
-- High-resolution texture synthesis with geometric and diffusion priors
-- Outputs textured meshes in GLB/OBJ format
-- Outperforms both open and closed-source 3D generation models
-- 7 workflow templates available
-
-## Key Features
-
-- Native bilingual support (Chinese and English) across the family
-- Strong text rendering in generated images (Hunyuan-DiT)
-- State-of-the-art video generation quality (HunyuanVideo)
-- End-to-end 3D asset creation with texturing (Hunyuan3D)
-- Multi-resolution generation across all model types
-- Prompt rewrite system for improved generation quality (HunyuanVideo)
-
-## Hardware Requirements
-
-- Hunyuan-DiT: 11GB VRAM minimum (fp16), 16GB recommended
-- HunyuanVideo 540p (544x960): 45GB VRAM minimum
-- HunyuanVideo 720p (720x1280): 60GB VRAM minimum, 80GB recommended
-- HunyuanVideo FP8: Saves ~10GB compared to fp16 weights
-- Hunyuan3D 2.0: 16-24GB VRAM for shape + texture pipeline
-
-## Common Use Cases
-
-- Bilingual content creation and marketing materials
-- Asian-style artwork and illustrations
-- Text-in-image generation (Chinese/English)
-- High-quality video generation from text or image prompts
-- 3D asset creation for games, design, and prototyping
-- Textured mesh generation from reference images
-
-## Key Parameters
-
-- **steps**: 25-50 for Hunyuan-DiT (default 40), 50 for HunyuanVideo
-- **cfg_scale**: 5-8 for DiT (6 typical), 6.0 embedded for HunyuanVideo
-- **flow_shift**: 7.0 for HunyuanVideo flow matching scheduler
-- **video_length**: 129 frames for HunyuanVideo (~5s at 24fps)
-- **resolution**: 1024x1024 for DiT, 720x1280 or 544x960 for video
-- **negative_prompt**: Recommended for Hunyuan-DiT quality control
-
-## Blog References
-
-- [HunyuanVideo Native Support](../blog/hunyuanvideo-native-support.md) — 13B parameter video model, dual-stream transformer, MLLM text encoder
-- [HunyuanVideo 1.5 Native Support](../blog/hunyuanvideo-15-native-support.md) — Lightweight 8.3B model, 720p output, runs on 24GB consumer GPUs
-- [Hunyuan3D 2.0 and MultiView Native Support](../blog/hunyuan3d-20-native-support.md) — 3D model generation with PBR materials, 1.1B parameter multi-view model

apps/hub/knowledge/models/hunyuan.summary.md

@@ -1 +0,0 @@
-Hunyuan is Tencent's open-source generative model family spanning text-to-image, text-to-video, and 3D generation. Hunyuan-DiT is a 1.5B parameter text-to-image model with native Chinese and English support and strong CJK text rendering at 1024x1024 (11-16GB VRAM). HunyuanVideo is the largest open-source video model at 13B+ parameters, generating 720p video up to 129 frames (~5s at 24fps) using a dual-stream transformer with MLLM text encoder; it requires 45-80GB VRAM depending on resolution (FP8 saves ~10GB). Hunyuan3D 2.0 handles image-to-3D and text-to-3D generation via a two-stage pipeline producing textured GLB/OBJ meshes (16-24GB VRAM). Key strengths: bilingual content creation, state-of-the-art video quality surpassing Runway Gen-3, and end-to-end 3D asset creation. Typical parameters: 25-50 steps for DiT, 50 steps for video, cfg_scale 5-8.

apps/hub/knowledge/models/ideogram.md

@@ -1,52 +0,0 @@
-# Ideogram
-
-Ideogram is an AI image generation platform founded by former Google Brain researchers, known for industry-leading text rendering accuracy in generated images. It achieves approximately 90% text rendering accuracy compared to roughly 30% for competing tools.
-
-## Model Variants
-
-### Ideogram 3.0
-
-- Latest generation released March 2025
-- Highest ELO rating in human evaluations across diverse prompts
-- Style References support with up to 3 reference images
-- Random style feature with 4.3 billion style presets
-- Batch generation for scaled content production
-
-### Ideogram 2.0
-
-- Previous generation model
-- Available as alternative option in the platform
-- Solid text rendering and general image quality
-
-## Key Features
-
-- Best-in-class text rendering with accurate typography and spelling
-- Handles complex, multi-line text compositions and curved surfaces
-- Style modes: Realistic, Anime, 3D, Watercolor, Typography
-- Magic Prompt for automatic prompt enhancement
-- Canvas editing for post-generation refinement
-- Upscaler up to 8K resolution in 2x increments
-- Color palette control for brand consistency
-- API available for programmatic integration
-
-## Hardware Requirements
-
-- Cloud API only (no local GPU required)
-- API pricing at approximately $0.06 per image
-- Web interface with credit-based subscription plans
-
-## Common Use Cases
-
-- Marketing materials with branded text and logos
-- Social media graphics with text overlays
-- Product packaging and label design
-- Event posters, flyers, and invitations
-- Book covers and editorial design
-
-## Key Parameters
-
-- **prompt**: Text description with quoted text for typography
-- **model**: Version selection (2.0 or 3.0)
-- **style**: Realistic, Anime, 3D, Watercolor, Typography
-- **aspect_ratio**: 16 aspect ratio options available
-- **magic_prompt**: Toggle for automatic prompt enhancement

apps/hub/knowledge/models/kandinsky.md

@@ -1,51 +0,0 @@
-# Kandinsky
-
-Kandinsky is a family of open-source diffusion models for video and image generation, developed by Kandinsky Lab (Sber AI, Russia). The models support both English and Russian text prompts.
-
-## Model Variants
-
-### Kandinsky 5.0 Video Pro (19B)
-
-- HD video at 1280x768, 24fps (5 or 10 seconds)
-- Controllable camera motion via LoRA
-- Top-1 open-source T2V model on LMArena
-
-### Kandinsky 5.0 Video Lite (2B)
-
-- Lightweight model, #1 among open-source in its class
-- CFG-distilled (2x faster) and diffusion-distilled (6x faster) variants
-- Best Russian concept understanding in open source
-
-### Kandinsky 5.0 Image Lite (6B)
-
-- HD image output (1280x768, 1024x1024)
-- Strong text rendering; image editing variant available
-
-## Key Features
-
-- Bilingual support (English and Russian prompts)
-- Flow Matching architecture with MIT license
-- Camera control via trained LoRAs
-- ComfyUI and Diffusers integration
-- MagCache acceleration for faster inference
-
-## Hardware Requirements
-
-- Video Lite: 12GB VRAM minimum with optimizations
-- Video Pro: 24GB+ VRAM recommended
-- NF4 quantization and FlashAttention 2/3 or SDPA supported
-
-## Common Use Cases
-
-- Open-source video generation research
-- Russian and English bilingual content creation
-- Camera-controlled video synthesis
-- Image generation with text rendering
-- Fine-tuning with custom LoRAs
-
-## Key Parameters
-
-- **prompt**: Text description in English or Russian
-- **num_frames**: Number of output frames (5s or 10s)
-- **resolution**: Output resolution (up to 1280x768)
-- **steps**: Inference steps (varies by distillation level)

apps/hub/knowledge/models/kling.md

@@ -1,64 +0,0 @@
-# Kling
-
-Kling is a video and image generation platform developed by Kuaishou Technology. It offers text-to-video, image-to-video, video editing, audio generation, and virtual try-on capabilities through both a creative studio and a developer API.
-
-## Model Variants
-
-### Kling O1
-
-- First unified multimodal video model combining generation and editing
-- Built on Multimodal Visual Language (MVL) framework
-- Accepts text, image, video, and subject inputs in a single prompt
-- Supports video inpainting, outpainting, style re-rendering, and shot extension
-- Character and scene consistency via Element Library with director-like memory
-- Generates 3-10 second videos at up to 2K resolution
-
-### Kling 2.6
-
-- Simultaneous audio-visual generation in a single pass
-- Produces video with speech, sound effects, and ambient sounds together
-- Supports Chinese and English voice generation
-- Video content up to 10 seconds with synchronized audio
-- Deep semantic alignment between audio and visual dynamics
-
-### Kling (Base Models)
-
-- Text-to-video and image-to-video with Standard and Professional modes
-- Multi-image-to-video with multiple reference inputs
-- Camera control with 6 basic movements and 4 master shots
-- Video extension, lip-sync, and avatar generation
-- Start and end frame generation for controlled transitions
-
-## Key Features
-
-- Unified generation and editing in a single model (O1)
-- Simultaneous audio-visual generation (2.6)
-- Multi-subject consistency across shots and angles
-- Conversational editing via natural language prompts
-- Video effects center for special effects and transformations
-- Virtual try-on and image recognition capabilities
-- DeepSeek integration for prompt optimization
-
-## Hardware Requirements
-
-- Cloud API only; no local hardware required
-- Accessed via klingai.com creative studio or API platform
-- Standard and Professional generation modes (speed vs. quality tradeoff)
-
-## Common Use Cases
-
-- Film and television pre-production and shot generation
-- Social media content creation with audio
-- E-commerce product videos and virtual try-on
-- Advertising with one-click ad generation
-- Video post-production editing via text prompts
-- Multi-character narrative video creation
-
-## Key Parameters
-
-- **prompt**: Text description with positive and negative prompts
-- **mode**: Standard (fast) or Professional (high quality)
-- **duration**: Video length (3-10 seconds for O1, up to 10s for 2.6)
-- **aspect_ratio**: Width-to-height ratio for output
-- **camera_control**: Predefined camera movements and master shots
-- **creativity_strength**: Balance between reference fidelity and creative variation

apps/hub/knowledge/models/ltx-video.md

@@ -1,68 +0,0 @@
-# LTX-Video
-
-LTX-Video is Lightricks' open-source DiT-based video generation model, the first capable of generating high-quality videos in real-time.
-
-## Model Variants
-
-### LTX-Video 2 (v0.9.7/v0.9.8)
-
-- Major quality upgrade over the original release
-- Available in 2B and 13B parameter sizes
-- 13B dev: highest quality, requires more VRAM
-- 13B distilled: faster inference, fewer steps needed, slight quality trade-off
-- 2B distilled: lightweight option for lower VRAM usage
-- FP8 quantized versions available for all sizes (13B-dev, 13B-distilled, 2B-distilled)
-- Multi-condition generation: condition on multiple images or video segments at specific frames
-- Spatial and temporal upscaler models for enhanced resolution and frame rate
-- ICLoRA adapters for depth, pose, and canny edge conditioning
-- 9 workflow templates available
-
-### LTX-Video 0.9.1/0.9.6
-
-- Original public releases with 2B parameter DiT architecture
-- Text-to-video and image-to-video modes
-- 768x512 native resolution at 24fps
-- 0.9.6 distilled variant: 15x faster, real-time capable, no CFG required
-- Foundation for community fine-tunes
-
-## Key Features
-
-- Real-time video generation on high-end GPUs (first DiT model to achieve this)
-- Generates 30 FPS video at 1216x704 resolution faster than playback speed
-- Multi-condition generation with per-frame image/video conditioning and strength control
-- Temporal VAE for smooth, consistent motion
-- Multi-scale rendering pipeline mixing dev and distilled models for speed-quality balance
-- Latent upsampling pipeline for progressive resolution enhancement
-
-## Hardware Requirements
-
-- 2B model: 12GB VRAM minimum, 16GB recommended
-- 2B distilled FP8: 8-10GB VRAM
-- 13B model: 24-32GB VRAM (fp16)
-- 13B FP8: 16-20GB VRAM
-- 13B distilled: less VRAM than 13B dev, ideal for rapid iterations
-- 32GB+ system RAM recommended for all variants
-
-## Common Use Cases
-
-- Short-form video content and social media clips
-- Image-to-video animation from reference frames
-- Video-to-video transformation and extension
-- Multi-condition video generation (start/end frame, keyframes)
-- Depth, pose, and edge-conditioned video generation via ICLoRA
-- Rapid video prototyping and creative experimentation
-
-## Key Parameters
-
-- **num_frames**: Output frame count (divisible by 8 + 1, e.g. 97, 161, 257)
-- **steps**: 30-50 for dev models, 8-15 for distilled variants
-- **cfg_scale**: 3-5 typical for dev, not required for distilled
-- **width/height**: Divisible by 32, best under 720x1280 for 13B
-- **denoise_strength**: 0.3-0.5 when using latent upsampler refinement pass
-- **conditioning_strength**: Per-condition strength for multi-condition generation (default 1.0)
-- **seed**: For reproducible generation
-
-## Blog References
-
-- [LTX-Video 0.9.5 Day-1 Support](../blog/ltx-video-095-support.md) — Commercial license (OpenRail-M), multi-frame control, improved quality
-- [LTX-2: Open Source Audio-Video AI](../blog/ltx-2-open-source-audio-video.md) — Synchronized audio-video generation, NVFP4 for 3x speed / 60% less VRAM

apps/hub/knowledge/models/ltx-video.summary.md

@@ -1 +0,0 @@
-LTX-Video is Lightricks' open-source DiT-based video generation model, the first to achieve real-time video generation. LTX-Video 2 (v0.9.7/0.9.8) is available in 2B and 13B parameter sizes, with dev, distilled, and FP8 quantized variants. It supports multi-condition generation with per-frame image/video conditioning, spatial and temporal upscalers, and ICLoRA adapters for depth, pose, and canny conditioning. The 2B model needs 12-16GB VRAM (8-10GB FP8), while the 13B model requires 24-32GB (16-20GB FP8). It generates 30fps video at 1216x704 faster than playback speed. Earlier versions (0.9.1/0.9.6) established the 2B foundation with a 15x faster distilled variant. Primary uses: short-form video, image-to-video animation, video extension, and multi-condition keyframe generation. Key parameters: 30-50 steps for dev, 8-15 for distilled, cfg_scale 3-5, frames divisible by 8+1.

apps/hub/knowledge/models/luma.md

@@ -1,50 +0,0 @@
-# Luma
-
-Luma AI develops video and image generation models through its Dream Machine platform, powered by the Ray model family and Photon image model.
-
-## Model Variants
-
-### Ray3 / Ray3.14
-
-- Native 1080p video with reasoning-driven generation
-- World's first native 16-bit HDR video generation
-- Character reference, Modify Video, and Draft Mode (5x faster)
-
-### Ray2
-
-- Production-ready text-to-video and image-to-video
-- 5-9 second output at 24fps with coherent motion
-
-### Photon
-
-- Image generation with strong prompt following
-- Character and visual reference support
-- 1080p output at $0.016 per image
-
-## Key Features
-
-- Reasoning capability for understanding creative intent
-- Visual annotation for precise layout and motion control
-- HDR generation with 16-bit EXR export for pro workflows
-- Keyframe control, video extension, looping, and camera control
-
-## Hardware Requirements
-
-- API-only access via Luma AI API
-- No local hardware requirements
-- Available through Dream Machine web and iOS app
-
-## Common Use Cases
-
-- Cinematic video production and storytelling
-- Commercial advertising and product videos
-- Visual effects with Modify Video workflows
-- HDR content for professional post-production
-
-## Key Parameters
-
-- **prompt**: Text description for video generation
-- **keyframes**: Start and/or end frame images
-- **aspect_ratio**: Output dimensions and ratio
-- **loop**: Enable seamless looping
-- **camera_control**: Camera movement via text instructions

apps/hub/knowledge/models/magnific.md

@@ -1,47 +0,0 @@
-# Magnific
-
-Magnific is an AI-powered image upscaler and enhancer that uses generative AI to hallucinate new details and textures during the upscaling process.
-
-## Model Variants
-
-### Magnific Creative Upscaler
-
-- Generative upscaling up to 16x (max 10,000px per dimension)
-- AI engines: Illusio (illustration), Sharpy (photography), Sparkle (balanced)
-- Adds hallucinated details guided by text prompts
-
-### Magnific Precision Upscaler
-
-- Faithful high-fidelity upscaling without creative reinterpretation
-- Clean enlargement that stays true to the source image
-
-### Mystic Image Generator
-
-- Photorealistic text-to-image/image-to-image with LoRA styles at up to 4K
-
-## Key Features
-
-- Creativity slider controls AI-hallucinated detail level
-- HDR control for micro-contrast and crispness
-- Resemblance slider to balance fidelity vs. creative enhancement
-- Optimized modes for portraits, illustrations, video games, and film
-- API hosted on Freepik with Skin Enhancer endpoint
-
-## Hardware Requirements
-
-- Cloud-only service with no local hardware requirements
-- API available through Freepik's developer platform
-- Subscription-based with credit system
-
-## Common Use Cases
-
-- Upscaling AI-generated images for print and production
-- Enhancing low-resolution concept art and illustrations
-- Restoring old or compressed photographs
-
-## Key Parameters
-
-- Creativity: level of new detail hallucination (0-10)
-- HDR: micro-contrast and sharpness (-10 to 10)
-- Resemblance: fidelity to source image (-10 to 10)
-- Scale Factor: 2x, 4x, 8x, or 16x magnification

apps/hub/knowledge/models/meshy.md

@@ -1,49 +0,0 @@
-# Meshy
-
-Meshy is a popular AI 3D model generator enabling text-to-3D and image-to-3D creation with PBR textures and production-ready exports.
-
-## Model Variants
-
-### Meshy-6
-
-- Latest generation with highest quality geometry
-- Supports symmetry and pose control (A-pose, T-pose)
-- Configurable polygon counts up to 300,000
-
-### Meshy-5
-
-- Previous generation with art style support
-- Realistic and sculpture style options
-
-## Key Features
-
-- Text-to-3D with two-stage workflow (preview mesh, then refine textures)
-- Image-to-3D from photos, sketches, or illustrations
-- Multi-image input for multi-view reconstruction
-- AI texturing with PBR maps (diffuse, roughness, metallic, normal)
-- Automatic rigging and 500+ animation motion library
-- Smart remesh with quad or triangle topology control
-- Export in FBX, GLB, OBJ, STL, 3MF, USDZ, BLEND formats
-
-## Hardware Requirements
-
-- Cloud API-based (no local GPU required)
-- All generation runs on Meshy servers
-- API available on Pro tier and above
-
-## Common Use Cases
-
-- Game development asset creation
-- 3D printing and prototyping
-- Film and VFX previsualization
-- VR/AR content development
-- Product design and e-commerce
-
-## Key Parameters
-
-- **prompt**: Text description up to 600 characters
-- **ai_model**: Model version (meshy-5, meshy-6, latest)
-- **topology**: Mesh type (quad or triangle)
-- **target_polycount**: 100 to 300,000 polygons
-- **enable_pbr**: Generate PBR material maps
-- **pose_mode**: Character pose (a-pose, t-pose, or none)

apps/hub/knowledge/models/minimax.md

@@ -1,58 +0,0 @@
-# MiniMax
-
-MiniMax is a multi-modal AI company known for the Hailuo video generation models and Image-01, offering API-based video and image creation.
-
-## Model Variants
-
-### Hailuo 2.3
-
-- Latest video model with improved body movement and facial expressions
-- Supports anime, illustration, ink-wash, and game-CG styles
-- 768p or 1080p resolution, 6 or 10 second clips
-- Available in Quality and Fast variants
-
-### Hailuo 2.0 (Hailuo 02)
-
-- Native 1080p with Noise-aware Compute Redistribution (NCR)
-- 2.5x efficiency improvement over predecessors
-- Last-frame conditioning support
-
-### Image-01
-
-- Text-to-image generation with multiple output sizes
-
-### T2V-01-Director
-
-- Enhanced camera control with natural language commands
-- Pan, zoom, tracking shot, and shake directives
-
-## Key Features
-
-- Text-to-video and image-to-video generation
-- Up to 1080p resolution at 25fps
-- Video clips up to 10 seconds
-- Camera control with natural language commands
-- Subject consistency with reference images
-- Text-to-image generation with Image-01
-
-## Hardware Requirements
-
-- Cloud API-based (no local GPU required)
-- All generation runs on MiniMax servers
-- API access via platform.minimax.io
-
-## Common Use Cases
-
-- Social media video content creation
-- Cinematic short film production
-- Product advertising and e-commerce videos
-- Anime and illustrated content
-- Character-driven narrative scenes
-
-## Key Parameters
-
-- **prompt**: Text description for generation
-- **model**: Model selection (hailuo-2.3, hailuo-02, image-01)
-- **resolution**: Output resolution (768p or 1080p)
-- **duration**: Clip length (6 or 10 seconds for video)
-- **first_frame_image**: Reference image for image-to-video

apps/hub/knowledge/models/model-triage.json

@@ -1,762 +0,0 @@
-{
-  "generated": "2026-02-07",
-  "totalModels": 87,
-  "categories": {
-    "specific_model": [
-      {
-        "name": "Wan",
-        "category": "specific_model",
-        "templateCount": 36,
-        "priority": 108,
-        "docFile": "wan",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Nano Banana Pro",
-        "category": "specific_model",
-        "templateCount": 29,
-        "priority": 87,
-        "docFile": "nano-banana-pro",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Flux",
-        "category": "specific_model",
-        "templateCount": 24,
-        "priority": 72,
-        "docFile": "flux",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "SDXL",
-        "category": "specific_model",
-        "templateCount": 4,
-        "priority": 12,
-        "docFile": "sdxl",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "ACE-Step",
-        "category": "specific_model",
-        "templateCount": 7,
-        "priority": 21,
-        "docFile": "ace-step",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Seedance",
-        "category": "specific_model",
-        "templateCount": 6,
-        "priority": 18,
-        "docFile": "seedance",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Seedream",
-        "category": "specific_model",
-        "templateCount": 5,
-        "priority": 15,
-        "docFile": "seedream",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "HiDream",
-        "category": "specific_model",
-        "templateCount": 5,
-        "priority": 15,
-        "docFile": "hidream",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Stable Audio",
-        "category": "specific_model",
-        "templateCount": 4,
-        "priority": 12,
-        "docFile": "stable-audio",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Chatter Box",
-        "category": "specific_model",
-        "templateCount": 4,
-        "priority": 12,
-        "docFile": "chatterbox",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Z-Image-Turbo",
-        "category": "specific_model",
-        "templateCount": 4,
-        "priority": 12,
-        "docFile": "z-image",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Kandinsky",
-        "category": "specific_model",
-        "templateCount": 3,
-        "priority": 9,
-        "docFile": "kandinsky",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "OmniGen",
-        "category": "specific_model",
-        "templateCount": 3,
-        "priority": 9,
-        "docFile": "omnigen",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "SeedVR2",
-        "category": "specific_model",
-        "templateCount": 3,
-        "priority": 9,
-        "docFile": "seedvr2",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Chroma",
-        "category": "specific_model",
-        "templateCount": 2,
-        "priority": 6,
-        "docFile": "chroma",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "ChronoEdit",
-        "category": "specific_model",
-        "templateCount": 1,
-        "priority": 3,
-        "docFile": "chronoedit",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "HuMo",
-        "category": "specific_model",
-        "templateCount": 1,
-        "priority": 3,
-        "docFile": "humo",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "NewBie",
-        "category": "specific_model",
-        "templateCount": 1,
-        "priority": 3,
-        "docFile": "newbie",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Ovis-Image",
-        "category": "specific_model",
-        "templateCount": 1,
-        "priority": 3,
-        "docFile": "ovis-image",
-        "hasExistingDoc": false
-      }
-    ],
-    "provider_name": [
-      {
-        "name": "Google",
-        "category": "provider_name",
-        "templateCount": 29,
-        "priority": 0,
-        "mapsTo": ["gemini", "veo", "nano-banana-pro"],
-        "hasExistingDoc": false
-      },
-      {
-        "name": "BFL",
-        "category": "provider_name",
-        "templateCount": 28,
-        "priority": 0,
-        "mapsTo": ["flux"],
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Stability",
-        "category": "provider_name",
-        "templateCount": 19,
-        "priority": 0,
-        "mapsTo": ["sdxl", "stable-audio", "reimagine"],
-        "hasExistingDoc": false
-      },
-      {
-        "name": "ByteDance",
-        "category": "provider_name",
-        "templateCount": 11,
-        "priority": 0,
-        "mapsTo": ["seedance", "seedvr2", "seedream"],
-        "hasExistingDoc": false
-      },
-      {
-        "name": "OpenAI",
-        "category": "provider_name",
-        "templateCount": 11,
-        "priority": 0,
-        "mapsTo": ["gpt-image-1"],
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Lightricks",
-        "category": "provider_name",
-        "templateCount": 9,
-        "priority": 0,
-        "mapsTo": ["ltx-video"],
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Tencent",
-        "category": "provider_name",
-        "templateCount": 5,
-        "priority": 0,
-        "mapsTo": ["hunyuan"],
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Qwen",
-        "category": "provider_name",
-        "templateCount": 2,
-        "priority": 0,
-        "mapsTo": ["qwen"],
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Nvidia",
-        "category": "provider_name",
-        "templateCount": 1,
-        "priority": 0,
-        "mapsTo": [],
-        "hasExistingDoc": false
-      }
-    ],
-    "api_only": [
-      {
-        "name": "Vidu",
-        "category": "api_only",
-        "templateCount": 10,
-        "priority": 20,
-        "docFile": "vidu",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Kling",
-        "category": "api_only",
-        "templateCount": 9,
-        "priority": 18,
-        "docFile": "kling",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Recraft",
-        "category": "api_only",
-        "templateCount": 6,
-        "priority": 12,
-        "docFile": "recraft",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Runway",
-        "category": "api_only",
-        "templateCount": 5,
-        "priority": 10,
-        "docFile": "runway",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Tripo",
-        "category": "api_only",
-        "templateCount": 5,
-        "priority": 10,
-        "docFile": "tripo",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "GPT-Image-1",
-        "category": "api_only",
-        "templateCount": 4,
-        "priority": 8,
-        "docFile": "gpt-image-1",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "MiniMax",
-        "category": "api_only",
-        "templateCount": 4,
-        "priority": 8,
-        "docFile": "minimax",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Grok",
-        "category": "api_only",
-        "templateCount": 4,
-        "priority": 8,
-        "docFile": "grok",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Luma",
-        "category": "api_only",
-        "templateCount": 4,
-        "priority": 8,
-        "docFile": "luma",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Moonvalley",
-        "category": "api_only",
-        "templateCount": 4,
-        "priority": 8,
-        "docFile": "moonvalley",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Topaz",
-        "category": "api_only",
-        "templateCount": 4,
-        "priority": 8,
-        "docFile": "topaz",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "PixVerse",
-        "category": "api_only",
-        "templateCount": 3,
-        "priority": 6,
-        "docFile": "pixverse",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Meshy",
-        "category": "api_only",
-        "templateCount": 3,
-        "priority": 6,
-        "docFile": "meshy",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Rodin",
-        "category": "api_only",
-        "templateCount": 3,
-        "priority": 6,
-        "docFile": "rodin",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Magnific",
-        "category": "api_only",
-        "templateCount": 3,
-        "priority": 6,
-        "docFile": "magnific",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "WaveSpeed",
-        "category": "api_only",
-        "templateCount": 3,
-        "priority": 6,
-        "docFile": "wavespeed",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "BRIA",
-        "category": "api_only",
-        "templateCount": 2,
-        "priority": 4,
-        "docFile": "bria",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Veo",
-        "category": "api_only",
-        "templateCount": 2,
-        "priority": 4,
-        "docFile": "veo",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "HitPaw",
-        "category": "api_only",
-        "templateCount": 2,
-        "priority": 4,
-        "docFile": "hitpaw",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Z-Image",
-        "category": "api_only",
-        "templateCount": 1,
-        "priority": 2,
-        "docFile": "z-image",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Anima",
-        "category": "api_only",
-        "templateCount": 1,
-        "priority": 2,
-        "docFile": "anima",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Reimagine",
-        "category": "api_only",
-        "templateCount": 1,
-        "priority": 2,
-        "docFile": "reimagine",
-        "mapsTo": ["stability"],
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Ideogram",
-        "category": "api_only",
-        "templateCount": 1,
-        "priority": 2,
-        "docFile": "ideogram",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Gemini3 Pro Image Preview",
-        "category": "api_only",
-        "templateCount": 16,
-        "priority": 32,
-        "docFile": "gemini",
-        "hasExistingDoc": false
-      }
-    ],
-    "utility_model": [
-      {
-        "name": "SVD",
-        "category": "utility_model",
-        "templateCount": 1,
-        "priority": 1,
-        "docFile": "svd",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Real-ESRGAN",
-        "category": "utility_model",
-        "templateCount": 1,
-        "priority": 1,
-        "docFile": "real-esrgan",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Depth Anything v2",
-        "category": "utility_model",
-        "templateCount": 1,
-        "priority": 1,
-        "docFile": "depth-anything-v2",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "FlashVSR",
-        "category": "utility_model",
-        "templateCount": 1,
-        "priority": 1,
-        "docFile": "flashvsr",
-        "hasExistingDoc": false
-      }
-    ],
-    "variant": [
-      {
-        "name": "Wan2.1",
-        "category": "variant",
-        "templateCount": 21,
-        "priority": 0,
-        "mapsTo": "wan",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Wan2.2",
-        "category": "variant",
-        "templateCount": 15,
-        "priority": 0,
-        "mapsTo": "wan",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Qwen-Image-Edit",
-        "category": "variant",
-        "templateCount": 11,
-        "priority": 0,
-        "mapsTo": "qwen",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "LTX-2",
-        "category": "variant",
-        "templateCount": 9,
-        "priority": 0,
-        "mapsTo": "ltx-video",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Qwen-Image",
-        "category": "variant",
-        "templateCount": 7,
-        "priority": 0,
-        "mapsTo": "qwen",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Hunyuan3D",
-        "category": "variant",
-        "templateCount": 7,
-        "priority": 0,
-        "mapsTo": "hunyuan",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Google Gemini Image",
-        "category": "variant",
-        "templateCount": 6,
-        "priority": 0,
-        "mapsTo": "gemini",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Flux.2 Klein",
-        "category": "variant",
-        "templateCount": 6,
-        "priority": 0,
-        "mapsTo": "flux",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Kling O1",
-        "category": "variant",
-        "templateCount": 5,
-        "priority": 0,
-        "mapsTo": "kling",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Vidu Q2",
-        "category": "variant",
-        "templateCount": 5,
-        "priority": 0,
-        "mapsTo": "vidu",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "SD3.5",
-        "category": "variant",
-        "templateCount": 4,
-        "priority": 0,
-        "mapsTo": "sdxl",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Google Gemini",
-        "category": "variant",
-        "templateCount": 3,
-        "priority": 0,
-        "mapsTo": "gemini",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Flux.2 Dev",
-        "category": "variant",
-        "templateCount": 3,
-        "priority": 0,
-        "mapsTo": "flux",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Flux.2",
-        "category": "variant",
-        "templateCount": 3,
-        "priority": 0,
-        "mapsTo": "flux",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Wan2.5",
-        "category": "variant",
-        "templateCount": 3,
-        "priority": 0,
-        "mapsTo": "wan",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Kontext",
-        "category": "variant",
-        "templateCount": 3,
-        "priority": 0,
-        "mapsTo": "flux",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Wan2.6",
-        "category": "variant",
-        "templateCount": 3,
-        "priority": 0,
-        "mapsTo": "wan",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Hunyuan Video",
-        "category": "variant",
-        "templateCount": 3,
-        "priority": 0,
-        "mapsTo": "hunyuan",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Vidu Q3",
-        "category": "variant",
-        "templateCount": 2,
-        "priority": 0,
-        "mapsTo": "vidu",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "LTXV",
-        "category": "variant",
-        "templateCount": 2,
-        "priority": 0,
-        "mapsTo": "ltx-video",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Qwen-Image-Layered",
-        "category": "variant",
-        "templateCount": 2,
-        "priority": 0,
-        "mapsTo": "qwen",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "SD1.5",
-        "category": "variant",
-        "templateCount": 2,
-        "priority": 0,
-        "mapsTo": "sdxl",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Gemini-2.5-Flash",
-        "category": "variant",
-        "templateCount": 1,
-        "priority": 0,
-        "mapsTo": "gemini",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Qwen-Image 2512",
-        "category": "variant",
-        "templateCount": 1,
-        "priority": 0,
-        "mapsTo": "qwen",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Seedream 4.0",
-        "category": "variant",
-        "templateCount": 1,
-        "priority": 0,
-        "mapsTo": "seedream",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "GPT-Image-1.5",
-        "category": "variant",
-        "templateCount": 1,
-        "priority": 0,
-        "mapsTo": "gpt-image-1",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Kling2.6",
-        "category": "variant",
-        "templateCount": 1,
-        "priority": 0,
-        "mapsTo": "kling",
-        "hasExistingDoc": false
-      },
-      {
-        "name": "Wan-Move",
-        "category": "variant",
-        "templateCount": 1,
-        "priority": 0,
-        "mapsTo": "wan",
-        "hasExistingDoc": true
-      },
-      {
-        "name": "Motion Control",
-        "category": "variant",
-        "templateCount": 1,
-        "priority": 0,
-        "mapsTo": "wan",
-        "hasExistingDoc": false
-      }
-    ],
-    "skip": [
-      {
-        "name": "None",
-        "category": "skip",
-        "templateCount": 1,
-        "priority": 0,
-        "hasExistingDoc": false
-      },
-      {
-        "name": "nano-banana",
-        "category": "skip",
-        "templateCount": 1,
-        "priority": 0,
-        "note": "Duplicate of Nano Banana Pro",
-        "hasExistingDoc": false
-      }
-    ]
-  },
-  "priorityOrder": [
-    "wan",
-    "nano-banana-pro",
-    "flux",
-    "gemini",
-    "ace-step",
-    "vidu",
-    "kling",
-    "seedance",
-    "seedream",
-    "hidream",
-    "sdxl",
-    "stable-audio",
-    "chatterbox",
-    "z-image",
-    "recraft",
-    "runway",
-    "tripo",
-    "kandinsky",
-    "omnigen",
-    "seedvr2",
-    "gpt-image-1",
-    "minimax",
-    "grok",
-    "luma",
-    "moonvalley",
-    "topaz",
-    "chroma",
-    "pixverse",
-    "meshy",
-    "rodin",
-    "magnific",
-    "wavespeed",
-    "bria",
-    "veo",
-    "hitpaw",
-    "newbie",
-    "ovis-image",
-    "chronoedit",
-    "humo",
-    "anima",
-    "reimagine",
-    "ideogram",
-    "svd",
-    "real-esrgan",
-    "depth-anything-v2",
-    "flashvsr"
-  ]
-}

apps/hub/knowledge/models/moonvalley.md

@@ -1,53 +0,0 @@
-# Moonvalley (Marey)
-
-Marey is Moonvalley's AI video generation model for professional filmmakers, delivering studio-grade quality and trained exclusively on licensed footage.
-
-## Model Variants
-
-### Marey Realism v1.5
-
-- Latest production model with cinematic detail
-- 1080p resolution at 24fps, up to 5-second clips
-- Available via ComfyUI native nodes and fal.ai
-
-### Marey Director Controls
-
-- 3D-aware camera control from single images
-- Motion transfer from reference videos
-- Trajectory control for object path definition
-- Pose transfer and keyframing with multi-image timeline
-
-## Key Features
-
-- Text-to-video and image-to-video generation
-- Camera control with 3D scene understanding
-- Motion transfer from reference video clips
-- Trajectory control via drawn paths
-- Pose transfer for expressive character animation
-- Shot extension for seamless duration increase
-- Commercially safe (trained on licensed data only)
-
-## Hardware Requirements
-
-- Cloud API-based (no local GPU required)
-- Available via Moonvalley platform, ComfyUI, and fal.ai
-- Subscription tiers starting at $14.99/month
-
-## Common Use Cases
-
-- Professional film and commercial production
-- Cinematic B-roll generation
-- Previsualization and storyboarding
-- Music video and social media content
-- Product advertising with dynamic camera
-- Animation and character-driven storytelling
-
-## Key Parameters
-
-- **prompt**: Text description of desired video scene
-- **image**: Reference image for image-to-video mode
-- **camera_control**: Camera movement specification
-- **motion_reference**: Video reference for motion transfer
-- **trajectory**: Drawn path for object movement
-- **duration**: Clip length (up to 5 seconds)
-- **resolution**: Output resolution (up to 1080p at 24fps)

apps/hub/knowledge/models/nano-banana-pro.md

@@ -1,53 +0,0 @@
-# Nano Banana Pro
-
-Nano Banana Pro is Google DeepMind's flagship image generation and editing model, accessed through ComfyUI's API nodes. Internally it is the Gemini 3 Pro Image model, designed for production-ready high-fidelity visuals.
-
-## Model Variants
-
-### Nano Banana Pro (Gemini 3 Pro Image)
-
-- State-of-the-art reasoning-powered image generation
-- Supports up to 14 reference image inputs
-- Native 4K output resolution (up to 4096x4096)
-- Complex multi-turn image generation and editing
-- Model ID: `gemini-3-pro-image-preview`
-
-### Gemini 2.5 Flash Image (Nano Banana)
-
-- Cost-effective alternative optimized for speed
-- Balanced price-to-performance for interactive workflows
-- Character consistency and prompt-based editing
-- Model ID: `gemini-2.5-flash-image`
-
-## Key Features
-
-- **World knowledge**: Generates accurate real-world images using Google Search's knowledge base
-- **Text rendering**: Clean text generation with detection and translation across 10 languages
-- **Multi-image fusion**: Blend up to 14 input images into a single coherent output
-- **Studio controls**: Adjust angles, focus, color grading in generated images
-- **Character consistency**: Maintain subject identity across multiple generations
-- **Prompt-based editing**: Targeted transformations via natural language instructions
-
-## Hardware Requirements
-
-- No local GPU required — runs as a cloud API service
-- Accessed via ComfyUI API nodes (requires ComfyUI login and network access)
-- Available on Comfy Cloud or local ComfyUI with API node support
-
-## Common Use Cases
-
-- High-fidelity text-to-image generation
-- Multi-reference style transfer and image blending
-- Product visualization and mockups
-- Sketch-to-image and blueprint-to-3D visualization
-- Text rendering and translation in images
-- Iterative prompt-based image editing
-
-## Key Parameters
-
-- **prompt**: Text description of desired image or edit
-- **aspect_ratio**: Supported ratios include 1:1, 3:2, 4:3, 9:16, 16:9, 21:9
-- **temperature**: 0.0-2.0 (default 1.0)
-- **topP**: 0.0-1.0 (default 0.95)
-- **max_output_tokens**: Up to 32,768 tokens per response
-- **input images**: Up to 14 reference images per prompt

apps/hub/knowledge/models/newbie.md

@@ -1,43 +0,0 @@
-# NewBie
-
-NewBie image Exp0.1 is a 3.5B parameter open-source text-to-image model built on the Next-DiT architecture, developed by the NewBie-AI community. It is specifically pretrained on high-quality anime data for detailed and visually striking anime-style image generation.
-
-## Model Variants
-
-### NewBie image Exp0.1
-
-- 3.5B parameter DiT model based on Next-DiT architecture
-- Uses Gemma3-4B-it as primary text encoder with Jina CLIP v2 for pooled features
-- FLUX.1-dev 16-channel VAE for rich color rendering and fine texture detail
-- Supports natural language, tags, and XML structured prompts
-- Non-commercial community license (Newbie-NC-1.0) for model weights
-
-## Key Features
-
-- Exceptional anime and ACG (Anime, Comics, Games) style generation
-- XML structured prompting for improved attribute binding and element disentanglement
-- Strong multi-character scene generation with accurate attribute assignment
-- ComfyUI integration via dedicated custom nodes
-- LoRA training support with community trainer
-- Built on research from the Lumina architecture family
-
-## Hardware Requirements
-
-- Minimum: 12GB VRAM (bfloat16 or float16)
-- Recommended: 24GB VRAM for comfortable generation
-- Requires Gemma3-4B-it and Jina CLIP v2 text encoders
-- Python 3.10, PyTorch 2.6.0+, Transformers 4.57.1+
-
-## Common Use Cases
-
-- Anime and illustration generation
-- Character design with precise attribute control
-- Multi-character scene composition
-- Fan art and creative anime artwork
-
-## Key Parameters
-
-- **num_inference_steps**: 28 recommended
-- **height/width**: 1024x1024 native resolution
-- **prompt_format**: Natural language, tags, or XML structured
-- **torch_dtype**: bfloat16 recommended (float16 fallback)

apps/hub/knowledge/models/omnigen.md

@@ -1,53 +0,0 @@
-# OmniGen2
-
-OmniGen2 is a multimodal generation model with dual decoding pathways for text and image, built on the Qwen-VL-2.5 foundation by VectorSpaceLab.
-
-## Model Variants
-
-### OmniGen2
-
-- 3B vision-language encoder (Qwen-VL-2.5) + 4B image decoder
-- Dual decoding with unshared parameters for text and image
-- Decoupled image tokenizer
-- Apache 2.0 license
-
-### OmniGen v1
-
-- Earlier single-pathway architecture
-- Fewer capabilities than OmniGen2
-- Superseded by OmniGen2
-
-## Key Features
-
-- Text-to-image generation with high fidelity and aesthetics
-- Instruction-guided image editing (state-of-the-art among open-source models)
-- In-context generation combining multiple reference inputs (humans, objects, scenes)
-- Visual understanding inherited from Qwen-VL-2.5
-- CPU offload support reduces VRAM usage by nearly 50%
-- Sequential CPU offload available for under 3GB VRAM (slower inference)
-- Supports negative prompts and configurable guidance scales
-
-## Hardware Requirements
-
-- Minimum: NVIDIA RTX 3090 or equivalent (~17GB VRAM)
-- With CPU offload: ~9GB VRAM
-- With sequential CPU offload: under 3GB VRAM (significantly slower)
-- Flash Attention optional but recommended for best performance
-- CUDA 12.4+ recommended
-- Default output resolution: 1024x1024
-
-## Common Use Cases
-
-- Text-to-image generation
-- Instruction-based photo editing
-- Subject-driven image generation from reference photos
-- Multi-image composition and in-context editing
-
-## Key Parameters
-
-- **text_guidance_scale**: Controls adherence to text prompt (CFG)
-- **image_guidance_scale**: Controls similarity to reference image (1.2-2.0 for editing, 2.5-3.0 for in-context)
-- **num_inference_step**: Diffusion steps (default 50)
-- **max_pixels**: Maximum total pixel count for input images (default 1024x1024)
-- **negative_prompt**: Text describing undesired qualities (e.g., "blurry, low quality, watermark")
-- **scheduler**: ODE solver choice (euler or dpmsolver++)

apps/hub/knowledge/models/ovis-image.md

@@ -1,43 +0,0 @@
-# Ovis-Image
-
-Ovis-Image is a 7B text-to-image model by AIDC-AI, built on Ovis-U1, optimized for high-quality text rendering in generated images. It achieves state-of-the-art results on the CVTG-2K text rendering benchmark while remaining compact enough for single-GPU deployment.
-
-## Model Variants
-
-### Ovis-Image-7B
-
-- 2B (Ovis2.5-2B) + 7B parameter architecture
-- State-of-the-art on CVTG-2K benchmark for text rendering accuracy
-- Competitive with 20B+ models (Qwen-Image) and GPT-4o on text-centric tasks
-- Uses FLUX-based autoencoder for latent encoding
-- Apache 2.0 license
-
-## Key Features
-
-- Excellent text rendering with correct spelling and consistent typography
-- High fidelity on text-heavy, layout-sensitive prompts
-- Handles posters, banners, logos, UI mockups, and infographics
-- Supports diverse fonts, sizes, and aspect ratios
-- Strong performance on both English and Chinese text generation
-- Available via Diffusers library with OvisImagePipeline
-
-## Hardware Requirements
-
-- Minimum: 16GB VRAM (bfloat16)
-- Recommended: 24GB VRAM for comfortable use
-- Fits on a single high-end GPU
-- Tested with Python 3.10, PyTorch 2.6.0, Transformers 4.57.1
-
-## Common Use Cases
-
-- Generating posters and banners with accurate text
-- Logo and brand asset creation
-- UI mockup and infographic generation
-- Marketing materials with embedded typography
-
-## Key Parameters
-
-- **num_inference_steps**: 50 recommended
-- **guidance_scale**: 5.0
-- **resolution**: 1024x1024 native
-- **negative_prompt**: Supported for quality control

apps/hub/knowledge/models/pixverse.md

@@ -1,46 +0,0 @@
-# PixVerse
-
-PixVerse is an AI video generation platform founded in 2023 and backed by Alibaba, offering text-to-video and image-to-video capabilities with over 100 million registered users.
-
-## Model Variants
-
-### PixVerse V5.5
-
-- Latest model with improved fidelity, text-to-video, image-to-video, and modification
-
-### PixVerse R1
-
-- Real-time AI video generation model
-- Interactive control where users direct character actions as video unfolds
-
-### PixVerse V4.5 / V5
-
-- Previous generation models with strong cinematic quality and trending effects
-
-## Key Features
-
-- Text-to-video generation from natural language prompts
-- Image-to-video animation with realistic physics simulation
-- Fusion mode combining up to 3 images into one video
-- Key frame control and video extension with AI continuity
-- AI Video Modify for text-prompt-based editing
-
-## Hardware Requirements
-
-- Cloud-based platform with no local GPU required
-- Web app at app.pixverse.ai and mobile apps (iOS/Android)
-- API at platform.pixverse.ai for developer integration
-
-## Common Use Cases
-
-- Social media content creation (AI Kiss, Hug, Dance effects)
-- Marketing and promotional video production
-- Old photo revival and animation
-- Cinematic narrative and stylistic art generation
-
-## Key Parameters
-
-- prompt: text description of the desired video content
-- duration: video length (typically 5s clips)
-- resolution: output quality (360p to 720p+)
-- aspect_ratio: 16:9, 9:16, 1:1, and other ratios

apps/hub/knowledge/models/qwen.md

@@ -1,77 +0,0 @@
-# Qwen
-
-Qwen is Alibaba's family of vision-language and image generation models, spanning visual understanding, image editing, and image generation.
-
-## Model Variants
-
-### Qwen2.5-VL
-
-- Multimodal vision-language model from the Qwen team
-- Available in 3B, 7B, and 72B parameter sizes
-- Image understanding, video comprehension (1+ hour videos), and visual localization
-- Visual agent capabilities: computer use, phone use, dynamic tool calling
-- Structured output generation for invoices, forms, and tables
-- Dynamic resolution and frame rate training for video understanding
-- Optimized ViT encoder with window attention, SwiGLU, and RMSNorm
-
-### Qwen-Image-Edit
-
-- Specialized image editing model with instruction-following
-- Supports inpainting, outpainting, style transfer, and content-aware edits
-- 11 workflow templates available
-
-### Qwen-Image
-
-- Text-to-image generation model from the Qwen family
-- 7 workflow templates available
-
-### Qwen-Image-Layered
-
-- Layered image generation for composable outputs
-- Generates images with separate foreground/background layers
-- 2 workflow templates available
-
-### Qwen-Image 2512
-
-- Specific variant optimized for particular generation tasks
-- 1 workflow template available
-
-## Key Features
-
-- Strong visual understanding with state-of-the-art benchmark results
-- Native multi-language support including Chinese and English
-- Visual agent capabilities for computer and phone interaction
-- Video event capture with temporal segment pinpointing
-- Bounding box and point-based visual localization
-- Structured JSON output for document and table extraction
-- Instruction-based image editing with precise control
-
-## Hardware Requirements
-
-- 3B model: 6-8GB VRAM
-- 7B model: 16GB VRAM, flash_attention_2 recommended for multi-image/video
-- 72B model: Multi-GPU setup required (80GB+ per GPU)
-- Context length: 32,768 tokens default, extendable to 64K+ with YaRN
-- Dynamic pixel budget: 256-1280 tokens per image (configurable min/max pixels)
-
-## Common Use Cases
-
-- Image editing based on text instructions
-- Visual question answering and image description
-- Long video comprehension and event extraction
-- Document OCR and structured data extraction
-- Visual agent tasks (screen interaction, UI navigation)
-- Layered image generation for design workflows
-- Text-to-image generation with strong prompt following
-
-## Key Parameters
-
-- **max_new_tokens**: Controls output length for VL model responses
-- **min_pixels / max_pixels**: Control image token budget (e.g. 256x28x28 to 1280x28x28)
-- **temperature**: Generation diversity for text outputs
-- **resized_height / resized_width**: Direct image dimension control (rounded to nearest 28)
-- **fps**: Frame rate for video input processing in Qwen2.5-VL
-
-## Blog References
-
-- [Qwen Image Edit 2511 & Qwen Image Layered](../blog/qwen-image-edit-2511.md) — Better character consistency, RGBA layer decomposition, built-in LoRA support

apps/hub/knowledge/models/qwen.summary.md

@@ -1 +0,0 @@
-Qwen is Alibaba's family of vision-language and image generation models. Qwen2.5-VL is a multimodal vision-language model available in 3B (6-8GB VRAM), 7B (16GB), and 72B (multi-GPU 80GB+) sizes, capable of image understanding, hour-long video comprehension, visual localization, visual agent tasks (computer/phone use), and structured JSON output for document extraction. Qwen-Image-Edit provides instruction-based image editing with inpainting, outpainting, and style transfer. Qwen-Image handles text-to-image generation, while Qwen-Image-Layered produces composable foreground/background layer outputs. The family features native Chinese/English support, strong prompt following, and state-of-the-art visual understanding benchmarks. Key parameters include dynamic pixel budgets (256-1280 tokens per image), configurable frame rates for video input, and temperature for text diversity. Primary uses: image editing, visual QA, video comprehension, document OCR, and layered image generation.

apps/hub/knowledge/models/real-esrgan.md

@@ -1,61 +0,0 @@
-# Real-ESRGAN
-
-Real-ESRGAN is a practical image and video super-resolution model that extends ESRGAN with improved training on pure synthetic data for real-world restoration.
-
-## Model Variants
-
-### RealESRGAN_x4plus
-
-- General-purpose 4× upscaling model for real-world images
-- RRDB (Residual-in-Residual Dense Block) architecture
-- Handles noise, blur, JPEG compression artifacts
-
-### RealESRGAN_x4plus_anime_6B
-
-- Optimized for anime and illustration images
-- Smaller 6-block model for faster inference
-- Better edge preservation for line art
-
-### RealESRGAN_x2plus
-
-- 2× upscaling variant for moderate enlargement
-- Lower risk of hallucinated details
-
-### realesr-animevideov3
-
-- Lightweight model designed for anime video frames
-- Temporal consistency for video processing
-
-## Key Features
-
-- Trained entirely on synthetic degradation data (no paired real-world data needed)
-- Second-order degradation modeling simulates real-world compression chains
-- GFPGAN integration for face enhancement during upscaling
-- Tiling support for processing large images with limited VRAM
-- FP16 (half precision) inference for faster processing
-- NCNN Vulkan portable executables for cross-platform GPU support (Intel/AMD/NVIDIA)
-- Supports 2×, 3×, and 4× upscaling with arbitrary output scale via LANCZOS4 resize
-
-## Hardware Requirements
-
-- Minimum: 2GB VRAM with tiling enabled
-- Recommended: 4GB+ VRAM for comfortable use
-- NCNN Vulkan build runs on any GPU with Vulkan support
-- CPU inference supported but significantly slower
-
-## Common Use Cases
-
-- Upscaling old or low-resolution photographs
-- Enhancing compressed web images
-- Anime and manga image upscaling
-- Video frame super-resolution
-- Restoring degraded historical images
-- Pre-processing for print from low-resolution sources
-
-## Key Parameters
-
-- **outscale**: Final upsampling scale factor (default: 4)
-- **tile**: Tile size for memory management (0 = no tiling)
-- **face_enhance**: Enable GFPGAN face enhancement (default: false)
-- **model_name**: Select model variant (RealESRGAN_x4plus, anime_6B, etc.)
-- **denoise_strength**: Balance noise removal vs detail preservation (realesr-general-x4v3)

apps/hub/knowledge/models/recraft.md

@@ -1,50 +0,0 @@
-# Recraft
-
-Recraft is an AI image generation platform known for its V3 model and unique ability to produce both raster and vector (SVG) images from text prompts.
-
-## Model Variants
-
-### Recraft V3
-
-- Top-ranked model on Artificial Analysis Text-to-Image Leaderboard
-- Supports raster image generation at $0.04 per image
-- Supports vector SVG generation at $0.08 per image
-- Accurate text rendering at any size in generated images
-
-### Recraft 20B
-
-- More cost-effective variant at $0.022 per raster image
-- Vector generation at $0.044 per image
-- Suitable for high-volume production workflows
-
-## Key Features
-
-- Native vector SVG image generation from text prompts
-- Accurate text rendering (headlines, labels, signs) in images
-- Custom brand style creation from reference images
-- Generation in exact brand colors for brand consistency
-- AI-powered image vectorization (PNG/JPG to SVG)
-- Background removal, creative upscaling, and crisp upscaling
-- Multiple style presets: photorealism, clay, retro-pop, hand-drawn, 80s
-
-## Hardware Requirements
-
-- API-only access via Recraft API
-- No local hardware requirements
-- Available through Recraft Studio web interface
-
-## Common Use Cases
-
-- Logo and icon design (SVG output)
-- Brand-consistent marketing asset generation
-- Poster and advertisement creation with text
-- Scalable vector illustrations for web and print
-- Product mockup generation
-- SEO blog imagery at scale
-
-## Key Parameters
-
-- **prompt**: Text description of the desired image
-- **style**: Visual style (realistic_image, digital_illustration, vector_illustration, icon)
-- **colors**: Brand color palette for consistent output
-- **format**: Output format (raster PNG/JPG or vector SVG)

apps/hub/knowledge/models/rodin.md

@@ -1,57 +0,0 @@
-# Rodin
-
-Rodin is a 3D generation API by Hyper3D (DeemosTech) that creates production-ready 3D models from text or images with PBR materials.
-
-## Model Variants
-
-### Rodin Gen-2
-
-- Most advanced model with 10 billion parameters
-- Built on the BANG architecture
-- 4x improved geometric mesh quality over Gen-1
-- Generation time approximately 90 seconds
-
-### Rodin Gen-1.5 Regular
-
-- Detailed 3D assets with customizable quality
-- Adjustable polygon counts and 2K textures
-- Generation time approximately 70 seconds
-
-### Rodin Sketch
-
-- Fast prototyping with basic geometry and 1K textures
-- GLB format only, generation in approximately 20 seconds
-
-## Key Features
-
-- Text-to-3D and image-to-3D generation
-- Multi-view image input (up to 5 images) with fuse and concat modes
-- PBR and Shaded material options
-- Quad and triangle mesh modes
-- HighPack add-on for 4K textures and high-poly models
-- Bounding box ControlNet for dimension constraints
-- T/A pose control for humanoid models
-
-## Hardware Requirements
-
-- Cloud API-based (no local GPU required)
-- All generation runs on Hyper3D servers
-- API key required via hyper3d.ai dashboard
-
-## Common Use Cases
-
-- Game asset production
-- VR/AR content creation
-- Product visualization
-- Character modeling with pose control
-- Rapid 3D prototyping
-
-## Key Parameters
-
-- **prompt**: Text description for text-to-3D mode
-- **images**: Up to 5 reference images for image-to-3D
-- **quality**: Detail level (high, medium, low, extra-low)
-- **mesh_mode**: Face type (Quad or Raw triangles)
-- **material**: Material type (PBR, Shaded, or All)
-- **geometry_file_format**: Output format (glb, fbx, obj, stl, usdz)
-- **seed**: Randomization seed (0-65535)

apps/hub/knowledge/models/runway.md

@@ -1,50 +0,0 @@
-# Runway
-
-Runway is a generative AI company producing state-of-the-art video generation models, accessible via API and web interface.
-
-## Model Variants
-
-### Gen-3 Alpha
-
-- Text-to-video and image-to-video at 1280x768, 24fps
-- 5 or 10 second output, extendable up to 40 seconds
-- Photorealistic human character generation
-
-### Gen-3 Alpha Turbo
-
-- Faster, lower-cost variant (5 credits/sec vs 10)
-- Requires input image; supports first, middle, and last keyframes
-- Video extension up to 34 seconds total
-
-### Gen-4 Turbo
-
-- Latest generation with improved motion and prompt adherence
-- Image reference support and text-to-image (gen4_image)
-
-## Key Features
-
-- Advanced camera controls (Motion Brush, Director Mode)
-- C2PA provenance metadata for content authenticity
-- Expressive human characters with gestures and emotions
-- Wide range of cinematic styles and terminology support
-
-## Hardware Requirements
-
-- API-only access via Runway developer portal
-- No local hardware requirements
-- Enterprise tier available for higher rate limits
-
-## Common Use Cases
-
-- Film pre-visualization and storyboarding
-- Commercial advertisement production
-- Social media video content
-- Visual effects and motion graphics
-- Music video and artistic video creation
-
-## Key Parameters
-
-- **prompt**: Text description guiding video generation
-- **duration**: Output length (5 or 10 seconds)
-- **ratio**: Aspect ratio (1280:768 or 768:1280)
-- **keyframes**: Start, middle, and/or end frame images

apps/hub/knowledge/models/sd3-5.md

@@ -1,63 +0,0 @@
-# Stable Diffusion 3.5
-
-Stable Diffusion 3.5 is Stability AI's text-to-image model family based on the Multimodal Diffusion Transformer (MMDiT) architecture with rectified flow matching.
-
-## Model Variants
-
-### Stable Diffusion 3.5 Large
-
-- 8.1 billion parameter MMDiT model
-- Highest quality and prompt adherence in the SD family
-- 1 megapixel native resolution (1024×1024)
-- 28-50 inference steps recommended
-
-### Stable Diffusion 3.5 Large Turbo
-
-- Distilled version of SD 3.5 Large
-- 4-step inference for fast generation
-- Guidance scale of 0 (classifier-free guidance disabled)
-- Comparable quality to full model at fraction of the time
-
-### Stable Diffusion 3.5 Medium
-
-- 2.5 billion parameter MMDiT-X architecture
-- Designed for consumer hardware (9.9GB VRAM for transformer)
-- Dual attention blocks in first 12 transformer layers
-- Multi-resolution generation from 0.25 to 2 megapixels
-- Skip Layer Guidance recommended for better coherency
-
-## Key Features
-
-- Three text encoders: CLIP ViT-L, OpenCLIP ViT-bigG (77 tokens each), T5-XXL (256 tokens)
-- QK-normalization for stable training and easier fine-tuning
-- Rectified flow matching replaces traditional DDPM/DDIM sampling
-- Strong text rendering and typography in generated images
-- Diverse output styles (photography, 3D, painting, line art)
-- Highly customizable base for fine-tuning and LoRA training
-- T5-XXL encoder optional (can be removed to save memory with minimal quality loss)
-- Supports negative prompts for excluding unwanted elements
-
-## Hardware Requirements
-
-- Large: 24GB+ VRAM recommended (fp16), quantizable to fit smaller GPUs
-- Large Turbo: 16GB+ VRAM recommended
-- Medium: 10GB VRAM minimum (excluding text encoders)
-- NF4 quantization available via bitsandbytes for low-VRAM GPUs
-- CPU offloading supported via diffusers pipeline
-
-## Common Use Cases
-
-- Photorealistic image generation
-- Artistic illustration and concept art
-- Typography and text-heavy designs
-- Product visualization
-- Fine-tuning and LoRA development
-- ControlNet-guided generation
-
-## Key Parameters
-
-- **steps**: 28-50 for Large, 4 for Large Turbo, 20-40 for Medium
-- **guidance_scale**: 4.5-7.5 for Large/Medium, 0 for Large Turbo
-- **max_sequence_length**: T5 token limit (77 or 256, higher = better prompt understanding)
-- **resolution**: 1024×1024 native, flexible aspect ratios around 1MP
-- **negative_prompt**: Text describing elements to exclude (not supported by Turbo)

apps/hub/knowledge/models/sdxl.md

@@ -1,75 +0,0 @@
-# Stable Diffusion
-
-Stable Diffusion is Stability AI's family of open-source image and video generation models, spanning multiple architectures from U-Net to diffusion transformers.
-
-## Model Variants
-
-### SDXL (Stable Diffusion XL)
-
-- Stability AI's flagship text-to-image model (6.6B parameter U-Net)
-- Native 1024x1024 resolution with flexible aspect ratios around 1MP
-- Two text encoders (CLIP ViT-L + OpenCLIP ViT-bigG)
-- Optional refiner model for second-stage detail enhancement
-- Turbo and Lightning distilled variants for 1-4 step generation
-- Largest ecosystem of LoRAs, fine-tunes, and community models
-
-### SD3.5 (Stable Diffusion 3.5)
-
-- Diffusion transformer (DiT) architecture, successor to SDXL
-- Three text encoders (CLIP ViT-L, OpenCLIP ViT-bigG, T5-XXL) for stronger prompt following
-- Available in Large (8B) and Medium (2B) parameter sizes
-- Improved text rendering and compositional accuracy over SDXL
-- 4 workflow templates available
-
-### SD1.5 (Stable Diffusion 1.5)
-
-- The classic 512x512 latent diffusion model
-- Single CLIP ViT-L text encoder, 860M parameter U-Net
-- Still widely used for its massive LoRA and checkpoint ecosystem
-- Lower VRAM requirements make it accessible on consumer hardware
-- 2 workflow templates available
-
-### SVD (Stable Video Diffusion)
-
-- Image-to-video generation model based on Stable Diffusion
-- Generates short video clips (14 or 25 frames) from a single image
-- Related model for motion generation from static inputs
-
-### Stability API Products
-
-- Reimagine: Stability's API-based image variation and transformation service
-
-## Key Features
-
-- Excellent composition, layout, and photorealism (SDXL/SD3.5)
-- Large open-source ecosystem with thousands of community fine-tunes
-- Flexible aspect ratios and multi-resolution support
-- Dual/triple CLIP text encoding for nuanced prompt interpretation
-- Strong text rendering in SD3.5 via T5-XXL encoder
-
-## Hardware Requirements
-
-- SD1.5: 4-6GB VRAM (fp16), runs on most consumer GPUs
-- SDXL Base: 8GB VRAM minimum (fp16), 12GB recommended
-- SDXL Base + Refiner: 16GB+ VRAM
-- SD3.5 Medium: 8-12GB VRAM
-- SD3.5 Large: 16-24GB VRAM (fp16), quantized versions for 12GB cards
-
-## Common Use Cases
-
-- Photorealistic image generation
-- Artistic illustrations and concept art
-- Product photography and design
-- Character and portrait generation
-- LoRA-based custom style and subject training
-- Image-to-video with SVD
-
-## Key Parameters
-
-- **steps**: 20-40 for SDXL base, 15-25 for refiner, 28+ for SD3.5
-- **cfg_scale**: 5-10 (7 default for SDXL), 3.5-7 for SD3.5
-- **sampler**: DPM++ 2M Karras and Euler are popular for SDXL; Euler for SD3.5
-- **resolution**: 1024x1024 native for SDXL/SD3.5, 512x512 for SD1.5
-- **clip_skip**: Often set to 1-2; important for SD1.5 LoRA compatibility
-- **denoise_strength**: 0.7-0.8 when using the SDXL refiner (img2img)
-- **negative_prompt**: Supported in SDXL/SD1.5; not used in SD3.5 by default

apps/hub/knowledge/models/sdxl.summary.md

@@ -1 +0,0 @@
-Stable Diffusion is Stability AI's open-source image and video generation family. SDXL is the flagship text-to-image model (6.6B U-Net, dual CLIP encoders) generating 1024x1024 images with the largest ecosystem of LoRAs and community fine-tunes; it requires 8-12GB VRAM with Turbo/Lightning variants for 1-4 step generation. SD3.5 is the DiT-based successor with triple text encoders (including T5-XXL) in Large (8B, 16-24GB) and Medium (2B, 8-12GB) sizes, offering improved text rendering and compositional accuracy. SD1.5 remains popular for its massive ecosystem at just 4-6GB VRAM (512x512). SVD handles image-to-video generation (14 or 25 frames). Key parameters: 20-40 steps for SDXL, cfg_scale 5-10 (7 default), DPM++ 2M Karras sampler. Primary uses: photorealistic generation, artistic illustration, product photography, character generation, and LoRA-based custom training.

apps/hub/knowledge/models/seedance.md

@@ -1,64 +0,0 @@
-# Seedance
-
-Seedance is ByteDance's video generation model family, designed for cinematic, high-fidelity video creation from text and images. The 1.0 series established a standard for fluid motion and multi-shot consistency, while the 1.5 series adds native joint audio-visual generation.
-
-## Model Variants
-
-### Seedance 1.5 Pro
-
-- Native audio-visual generation producing video and audio in a single pass
-- Multilingual lip-sync supporting English, Mandarin, Japanese, Korean, and Spanish
-- 1080p output with 5-12 second duration
-- Advanced directorial camera controls (dolly zoom, tracking shots, whip pans)
-- Captures micro-expressions, non-verbal cues, and emotional transitions
-
-### Seedance 1.0 Pro
-
-- Production-quality 1080p video generation
-- Text-to-video and image-to-video with first and last frame control
-- Native multi-shot storytelling with subject and style consistency across cuts
-- Cinematic camera grammar interpretation (35mm film, noir lighting, drone shots)
-- 2-12 second video duration at 24-30fps
-
-### Seedance 1.0 Pro Fast
-
-- Faster, more cost-effective version of 1.0 Pro
-- Same capabilities with reduced generation time
-
-### Seedance 1.0 Lite
-
-- Optimized for speed and iteration at 720p or 1080p
-- Lower cost per generation for rapid prototyping
-
-## Key Features
-
-- Smooth, stable motion with wide dynamic range for large-scale movements
-- Native multi-shot storytelling maintaining consistency across transitions
-- Diverse stylistic expression (photorealism, cyberpunk, illustration, pixel art)
-- Precise prompt following for complex actions, multi-agent interactions, and camera work
-- Joint audio-visual synthesis with environmental sounds and dialogue (1.5)
-- Supports multiple aspect ratios (16:9, 9:16, 1:1, 4:3, 21:9, and more)
-
-## Hardware Requirements
-
-- Cloud API only; no local weights publicly available
-- Accessed via seed.bytedance.com, Scenario, fal.ai, and other API providers
-- 1080p 5-second video costs approximately $0.62 via fal.ai (Pro)
-- Lite version available at lower cost ($0.18 per 720p 5-second video)
-
-## Common Use Cases
-
-- Cinematic shorts and scene previsualization
-- Music video concept development
-- Product demonstration and marketing videos
-- Character-focused animation sequences
-- Social media content with audio (1.5)
-- Moodboard and style exploration for creative teams
-
-## Key Parameters
-
-- **prompt**: Text description of desired scene, action, and camera work
-- **image_url**: Source image for image-to-video generation (first frame)
-- **duration**: Video length (2-12 seconds for 1.0, 5-12 seconds for 1.5)
-- **resolution**: 480p, 720p, or 1080p output
-- **aspect_ratio**: 16:9, 9:16, 1:1, 4:3, 21:9, 9:21

apps/hub/knowledge/models/seedream.md

@@ -1,50 +0,0 @@
-# Seedream
-
-Seedream is ByteDance's text-to-image generation model, capable of producing high-quality images with strong text rendering, bilingual support (English and Chinese), and native high-resolution output.
-
-## Model Variants
-
-### Seedream 3.0
-
-- Native 2K resolution output without post-processing
-- Bilingual image generation (English and Chinese)
-- 3-second end-to-end generation for 1K images
-- Improved text rendering for small fonts and long text layouts
-
-### Seedream 4.0
-
-- Unified architecture for text-to-image and image editing
-- Native output up to 4K resolution
-- Multi-image reference input (up to 6 source images)
-- 1.8-second inference for 2K images
-- Batch input and output for multiple generations
-- Natural language image editing capabilities
-
-## Key Features
-
-- Accurate text rendering in both English and Chinese
-- Knowledge-driven generation for educational illustrations and charts
-- Strong character consistency across multiple angles
-- Prompt-based image editing without separate tools
-- Versatile style support from photorealism to anime
-- Leading scores on Artificial Analysis Image Arena
-
-## Hardware Requirements
-
-- API-only access via ByteDance Volcano Engine
-- No local hardware requirements for end users
-- Third-party API providers available (e.g., EvoLink)
-
-## Common Use Cases
-
-- Poster and advertisement design with embedded text
-- E-commerce product photography
-- Character design with multi-angle consistency
-- Educational illustration and infographic generation
-- Brand-consistent marketing materials
-
-## Key Parameters
-
-- **prompt**: Text description of the desired image
-- **resolution**: Output resolution (up to 4K supported)
-- **aspect_ratio**: Supports 16:9, 4:3, 1:1, and custom ratios

apps/hub/knowledge/models/seedvr2.md

@@ -1,47 +0,0 @@
-# SeedVR2
-
-SeedVR2 is a one-step diffusion-based video restoration model developed by ByteDance Seed and NTU S-Lab, published at ICLR 2026.
-
-## Model Variants
-
-### SeedVR2-3B
-
-- 3B parameter DiT with one-step inference for video and image upscaling
-- Available in FP16, FP8, and GGUF quantized formats
-
-### SeedVR2-7B
-
-- 7B parameter model with Sharp variant for maximum detail
-- Multi-GPU inference; supports 1080p and 2K on 4x H100-80GB
-
-### SeedVR (Original)
-
-- Multi-step diffusion model (CVPR 2025 Highlight)
-- Arbitrary-resolution restoration without pretrained diffusion prior
-
-## Key Features
-
-- One-step inference achieving 10x speedup over multi-step methods
-- Adaptive window attention with dynamic sizing for high-resolution inputs
-- Adversarial post-training against real data for faithful detail recovery
-- ComfyUI integration via official SeedVR2 Video Upscaler nodes
-- Apache 2.0 open-source license
-
-## Hardware Requirements
-
-- Minimum: 8-12GB VRAM with GGUF quantization and tiled VAE
-- Recommended: 24GB+ VRAM (RTX 4090) for 3B model at 1080p
-- High-end: 4x H100-80GB for 7B model at 2K resolution
-
-## Common Use Cases
-
-- Upscaling AI-generated video to 1080p or 4K
-- Restoring degraded or compressed video footage
-- Image super-resolution and detail recovery
-
-## Key Parameters
-
-- resolution: target shortest-edge resolution (720, 1080, 2160)
-- batch_size: frames per batch, must follow 4n+1 formula (5, 9, 13, 17, 21)
-- seed: random seed for reproducible generation
-- color_fix_type: wavelet, adain, hsv, or none

apps/hub/knowledge/models/stable-audio.md

@@ -1,53 +0,0 @@
-# Stable Audio Open
-
-Stable Audio Open 1.0 is Stability AI's open-source text-to-audio model for generating sound effects, production elements, and short musical clips.
-
-## Model Variants
-
-### Stable Audio Open 1.0
-
-- 1.2B parameter latent diffusion model
-- Transformer-based diffusion (DiT) architecture
-- T5-base text encoder for conditioning
-- Variational autoencoder for audio compression
-- Stability AI Community License (non-commercial)
-
-### Stable Audio (Commercial)
-
-- Full-length music generation up to 3 minutes with audio-to-audio and inpainting
-- Available via Stability AI platform API, commercial license
-
-## Key Features
-
-- Generates up to 47 seconds of stereo audio at 44.1kHz
-- Text-prompted sound effects, drum beats, ambient sounds, and foley
-- Variable-length output with timing control
-- Fine-tunable on custom audio datasets
-- Trained exclusively on Creative Commons licensed audio (CC0, CC BY, CC Sampling+)
-- Strong performance for sound effects and field recordings
-- Compatible with both stable-audio-tools and diffusers libraries
-
-## Hardware Requirements
-
-- Minimum: 8GB VRAM (fp16)
-- Recommended: 12GB+ VRAM for comfortable inference
-- Half-precision (fp16) supported for reduced memory
-- Chunked decoding available for memory-constrained setups
-- Inference speed: 8-20 diffusion steps per second depending on GPU
-
-## Common Use Cases
-
-- Sound effect and foley generation
-- Drum beats and instrument riff creation
-- Ambient soundscapes and background audio
-- Music production elements and samples
-- Audio prototyping for film and game sound design
-
-## Key Parameters
-
-- **steps**: Number of inference steps (100-200 recommended)
-- **cfg_scale**: Classifier-free guidance scale (typically 7)
-- **seconds_total**: Target audio duration (up to 47 seconds)
-- **seconds_start**: Start time offset for timing control
-- **negative_prompt**: Text describing undesired audio qualities
-- **sampler_type**: Diffusion sampler (dpmpp-3m-sde recommended)

apps/hub/knowledge/models/svd.md

@@ -1,55 +0,0 @@
-# Stable Video Diffusion
-
-Stable Video Diffusion (SVD) is Stability AI's image-to-video diffusion model that generates short video clips from a single conditioning image. In user studies, SVD was preferred over GEN-2 and PikaLabs for video quality.
-
-## Model Variants
-
-### SVD-XT (25 frames)
-
-- Generates 25 frames at 576x1024 resolution
-- Finetuned from the 14-frame SVD base model
-- Includes temporally consistent f8-decoder
-- Standard frame-wise decoder also available
-
-### SVD (14 frames)
-
-- Original release generating 14 frames
-- Foundation for community fine-tunes and extensions
-- Same 576x1024 native resolution
-
-## Key Features
-
-- Image-to-video generation from a single still image
-- Temporally consistent video output with finetuned decoder
-- Preferred over GEN-2 and PikaLabs in human evaluation studies
-- SynthID-compatible watermarking enabled by default
-- Latent diffusion architecture for efficient generation
-
-## Hardware Requirements
-
-- Minimum: 16GB VRAM
-- Recommended: A100 80GB for full quality (tested configuration)
-- SVD generation ~100s, SVD-XT ~180s on A100 80GB
-- Optimizations available for lower VRAM cards with quality tradeoffs
-
-## Common Use Cases
-
-- Animating still images into short video clips
-- Product visualization and motion graphics
-- Creative video experiments and art
-- Research on generative video models
-
-## Key Parameters
-
-- **num_frames**: 14 (SVD) or 25 (SVD-XT)
-- **resolution**: 576x1024 native
-- **conditioning_frame**: Input image at same resolution
-- **duration**: Up to ~4 seconds (25 frames)
-
-## Limitations
-
-- Short videos only (4 seconds maximum)
-- No text-based control (image conditioning only)
-- Cannot render legible text in output
-- Faces and people may not generate properly
-- May produce videos without motion or with very slow camera pans