fix: Fix variable insertion to only remove @ trigger on current line

components.

.claude/settings.json

@@ -1,11 +1,4 @@
 {
-  "enabledPlugins": {
-    "feature-dev@claude-plugins-official": true,
-    "context7@claude-plugins-official": true,
-    "typescript-lsp@claude-plugins-official": true,
-    "pyright-lsp@claude-plugins-official": true,
-    "ralph-loop@claude-plugins-official": true
-  },
   "hooks": {
     "PreToolUse": [
       {
@@ -18,5 +11,10 @@
         ]
       }
     ]
+  },
+  "enabledPlugins": {
+    "feature-dev@claude-plugins-official": true,
+    "context7@claude-plugins-official": true,
+    "ralph-loop@claude-plugins-official": true
   }
 }

.claude/skills/frontend-testing/SKILL.md

@@ -83,6 +83,9 @@ vi.mock('next/navigation', () => ({
   usePathname: () => '/test',
 }))
 
+// ✅ Zustand stores: Use real stores (auto-mocked globally)
+// Set test state with: useAppStore.setState({ ... })
+
 // Shared state for mocks (if needed)
 let mockSharedState = false
 
@@ -296,7 +299,7 @@ For each test file generated, aim for:
 For more detailed information, refer to:
 
 - `references/workflow.md` - **Incremental testing workflow** (MUST READ for multi-file testing)
-- `references/mocking.md` - Mock patterns and best practices
+- `references/mocking.md` - Mock patterns, Zustand store testing, and best practices
 - `references/async-testing.md` - Async operations and API calls
 - `references/domain-components.md` - Workflow, Dataset, Configuration testing
 - `references/common-patterns.md` - Frequently used testing patterns

.claude/skills/frontend-testing/references/mocking.md

@@ -37,16 +37,36 @@ Only mock these categories:
 1. **Third-party libraries with side effects** - `next/navigation`, external SDKs
 1. **i18n** - Always mock to return keys
 
+### Zustand Stores - DO NOT Mock Manually
+
+**Zustand is globally mocked** in `web/vitest.setup.ts`. Use real stores with `setState()`:
+
+```typescript
+// ✅ CORRECT: Use real store, set test state
+import { useAppStore } from '@/app/components/app/store'
+
+useAppStore.setState({ appDetail: { id: 'test', name: 'Test' } })
+render(<MyComponent />)
+
+// ❌ WRONG: Don't mock the store module
+vi.mock('@/app/components/app/store', () => ({ ... }))
+```
+
+See [Zustand Store Testing](#zustand-store-testing) section for full details.
+
 ## Mock Placement
 
 | Location | Purpose |
 |----------|---------|
-| `web/vitest.setup.ts` | Global mocks shared by all tests (for example `react-i18next`, `next/image`) |
+| `web/vitest.setup.ts` | Global mocks shared by all tests (`react-i18next`, `next/image`, `zustand`) |
+| `web/__mocks__/zustand.ts` | Zustand mock implementation (auto-resets stores after each test) |
 | `web/__mocks__/` | Reusable mock factories shared across multiple test files |
 | Test file | Test-specific mocks, inline with `vi.mock()` |
 
 Modules are not mocked automatically. Use `vi.mock` in test files, or add global mocks in `web/vitest.setup.ts`.
 
+**Note**: Zustand is special - it's globally mocked but you should NOT mock store modules manually. See [Zustand Store Testing](#zustand-store-testing).
+
 ## Essential Mocks
 
 ### 1. i18n (Auto-loaded via Global Mock)
@@ -276,6 +296,7 @@ const renderWithQueryClient = (ui: React.ReactElement) => {
 
 1. **Use real base components** - Import from `@/app/components/base/` directly
 1. **Use real project components** - Prefer importing over mocking
+1. **Use real Zustand stores** - Set test state via `store.setState()`
 1. **Reset mocks in `beforeEach`**, not `afterEach`
 1. **Match actual component behavior** in mocks (when mocking is necessary)
 1. **Use factory functions** for complex mock data
@@ -285,6 +306,7 @@ const renderWithQueryClient = (ui: React.ReactElement) => {
 ### ❌ DON'T
 
 1. **Don't mock base components** (`Loading`, `Button`, `Tooltip`, etc.)
+1. **Don't mock Zustand store modules** - Use real stores with `setState()`
 1. Don't mock components you can import directly
 1. Don't create overly simplified mocks that miss conditional logic
 1. Don't forget to clean up nock after each test
@@ -308,10 +330,151 @@ Need to use a component in test?
 ├─ Is it a third-party lib with side effects?
 │  └─ YES → Mock it (next/navigation, external SDKs)
 │
+├─ Is it a Zustand store?
+│  └─ YES → DO NOT mock the module!
+│           Use real store + setState() to set test state
+│           (Global mock handles auto-reset)
+│
 └─ Is it i18n?
    └─ YES → Uses shared mock (auto-loaded). Override only for custom translations
 ```
 
+## Zustand Store Testing
+
+### Global Zustand Mock (Auto-loaded)
+
+Zustand is globally mocked in `web/vitest.setup.ts` following the [official Zustand testing guide](https://zustand.docs.pmnd.rs/guides/testing). The mock in `web/__mocks__/zustand.ts` provides:
+
+- Real store behavior with `getState()`, `setState()`, `subscribe()` methods
+- Automatic store reset after each test via `afterEach`
+- Proper test isolation between tests
+
+### ✅ Recommended: Use Real Stores (Official Best Practice)
+
+**DO NOT mock store modules manually.** Import and use the real store, then use `setState()` to set test state:
+
+```typescript
+// ✅ CORRECT: Use real store with setState
+import { useAppStore } from '@/app/components/app/store'
+
+describe('MyComponent', () => {
+  it('should render app details', () => {
+    // Arrange: Set test state via setState
+    useAppStore.setState({
+      appDetail: {
+        id: 'test-app',
+        name: 'Test App',
+        mode: 'chat',
+      },
+    })
+
+    // Act
+    render(<MyComponent />)
+
+    // Assert
+    expect(screen.getByText('Test App')).toBeInTheDocument()
+    // Can also verify store state directly
+    expect(useAppStore.getState().appDetail?.name).toBe('Test App')
+  })
+
+  // No cleanup needed - global mock auto-resets after each test
+})
+```
+
+### ❌ Avoid: Manual Store Module Mocking
+
+Manual mocking conflicts with the global Zustand mock and loses store functionality:
+
+```typescript
+// ❌ WRONG: Don't mock the store module
+vi.mock('@/app/components/app/store', () => ({
+  useStore: (selector) => mockSelector(selector),  // Missing getState, setState!
+}))
+
+// ❌ WRONG: This conflicts with global zustand mock
+vi.mock('@/app/components/workflow/store', () => ({
+  useWorkflowStore: vi.fn(() => mockState),
+}))
+```
+
+**Problems with manual mocking:**
+
+1. Loses `getState()`, `setState()`, `subscribe()` methods
+1. Conflicts with global Zustand mock behavior
+1. Requires manual maintenance of store API
+1. Tests don't reflect actual store behavior
+
+### When Manual Store Mocking is Necessary
+
+In rare cases where the store has complex initialization or side effects, you can mock it, but ensure you provide the full store API:
+
+```typescript
+// If you MUST mock (rare), include full store API
+const mockStore = {
+  appDetail: { id: 'test', name: 'Test' },
+  setAppDetail: vi.fn(),
+}
+
+vi.mock('@/app/components/app/store', () => ({
+  useStore: Object.assign(
+    (selector: (state: typeof mockStore) => unknown) => selector(mockStore),
+    {
+      getState: () => mockStore,
+      setState: vi.fn(),
+      subscribe: vi.fn(),
+    },
+  ),
+}))
+```
+
+### Store Testing Decision Tree
+
+```
+Need to test a component using Zustand store?
+│
+├─ Can you use the real store?
+│  └─ YES → Use real store + setState (RECOMMENDED)
+│           useAppStore.setState({ ... })
+│
+├─ Does the store have complex initialization/side effects?
+│  └─ YES → Consider mocking, but include full API
+│           (getState, setState, subscribe)
+│
+└─ Are you testing the store itself (not a component)?
+   └─ YES → Test store directly with getState/setState
+            const store = useMyStore
+            store.setState({ count: 0 })
+            store.getState().increment()
+            expect(store.getState().count).toBe(1)
+```
+
+### Example: Testing Store Actions
+
+```typescript
+import { useCounterStore } from '@/stores/counter'
+
+describe('Counter Store', () => {
+  it('should increment count', () => {
+    // Initial state (auto-reset by global mock)
+    expect(useCounterStore.getState().count).toBe(0)
+
+    // Call action
+    useCounterStore.getState().increment()
+
+    // Verify state change
+    expect(useCounterStore.getState().count).toBe(1)
+  })
+
+  it('should reset to initial state', () => {
+    // Set some state
+    useCounterStore.setState({ count: 100 })
+    expect(useCounterStore.getState().count).toBe(100)
+
+    // After this test, global mock will reset to initial state
+  })
+})
+```
+
 ## Factory Function Pattern
 
 ```typescript

.claude/skills/vercel-react-best-practices/SKILL.md

@@ -0,0 +1,125 @@
+---
+name: vercel-react-best-practices
+description: React and Next.js performance optimization guidelines from Vercel Engineering. This skill should be used when writing, reviewing, or refactoring React/Next.js code to ensure optimal performance patterns. Triggers on tasks involving React components, Next.js pages, data fetching, bundle optimization, or performance improvements.
+license: MIT
+metadata:
+  author: vercel
+  version: "1.0.0"
+---
+
+# Vercel React Best Practices
+
+Comprehensive performance optimization guide for React and Next.js applications, maintained by Vercel. Contains 45 rules across 8 categories, prioritized by impact to guide automated refactoring and code generation.
+
+## When to Apply
+
+Reference these guidelines when:
+- Writing new React components or Next.js pages
+- Implementing data fetching (client or server-side)
+- Reviewing code for performance issues
+- Refactoring existing React/Next.js code
+- Optimizing bundle size or load times
+
+## Rule Categories by Priority
+
+| Priority | Category | Impact | Prefix |
+|----------|----------|--------|--------|
+| 1 | Eliminating Waterfalls | CRITICAL | `async-` |
+| 2 | Bundle Size Optimization | CRITICAL | `bundle-` |
+| 3 | Server-Side Performance | HIGH | `server-` |
+| 4 | Client-Side Data Fetching | MEDIUM-HIGH | `client-` |
+| 5 | Re-render Optimization | MEDIUM | `rerender-` |
+| 6 | Rendering Performance | MEDIUM | `rendering-` |
+| 7 | JavaScript Performance | LOW-MEDIUM | `js-` |
+| 8 | Advanced Patterns | LOW | `advanced-` |
+
+## Quick Reference
+
+### 1. Eliminating Waterfalls (CRITICAL)
+
+- `async-defer-await` - Move await into branches where actually used
+- `async-parallel` - Use Promise.all() for independent operations
+- `async-dependencies` - Use better-all for partial dependencies
+- `async-api-routes` - Start promises early, await late in API routes
+- `async-suspense-boundaries` - Use Suspense to stream content
+
+### 2. Bundle Size Optimization (CRITICAL)
+
+- `bundle-barrel-imports` - Import directly, avoid barrel files
+- `bundle-dynamic-imports` - Use next/dynamic for heavy components
+- `bundle-defer-third-party` - Load analytics/logging after hydration
+- `bundle-conditional` - Load modules only when feature is activated
+- `bundle-preload` - Preload on hover/focus for perceived speed
+
+### 3. Server-Side Performance (HIGH)
+
+- `server-cache-react` - Use React.cache() for per-request deduplication
+- `server-cache-lru` - Use LRU cache for cross-request caching
+- `server-serialization` - Minimize data passed to client components
+- `server-parallel-fetching` - Restructure components to parallelize fetches
+- `server-after-nonblocking` - Use after() for non-blocking operations
+
+### 4. Client-Side Data Fetching (MEDIUM-HIGH)
+
+- `client-swr-dedup` - Use SWR for automatic request deduplication
+- `client-event-listeners` - Deduplicate global event listeners
+
+### 5. Re-render Optimization (MEDIUM)
+
+- `rerender-defer-reads` - Don't subscribe to state only used in callbacks
+- `rerender-memo` - Extract expensive work into memoized components
+- `rerender-dependencies` - Use primitive dependencies in effects
+- `rerender-derived-state` - Subscribe to derived booleans, not raw values
+- `rerender-functional-setstate` - Use functional setState for stable callbacks
+- `rerender-lazy-state-init` - Pass function to useState for expensive values
+- `rerender-transitions` - Use startTransition for non-urgent updates
+
+### 6. Rendering Performance (MEDIUM)
+
+- `rendering-animate-svg-wrapper` - Animate div wrapper, not SVG element
+- `rendering-content-visibility` - Use content-visibility for long lists
+- `rendering-hoist-jsx` - Extract static JSX outside components
+- `rendering-svg-precision` - Reduce SVG coordinate precision
+- `rendering-hydration-no-flicker` - Use inline script for client-only data
+- `rendering-activity` - Use Activity component for show/hide
+- `rendering-conditional-render` - Use ternary, not && for conditionals
+
+### 7. JavaScript Performance (LOW-MEDIUM)
+
+- `js-batch-dom-css` - Group CSS changes via classes or cssText
+- `js-index-maps` - Build Map for repeated lookups
+- `js-cache-property-access` - Cache object properties in loops
+- `js-cache-function-results` - Cache function results in module-level Map
+- `js-cache-storage` - Cache localStorage/sessionStorage reads
+- `js-combine-iterations` - Combine multiple filter/map into one loop
+- `js-length-check-first` - Check array length before expensive comparison
+- `js-early-exit` - Return early from functions
+- `js-hoist-regexp` - Hoist RegExp creation outside loops
+- `js-min-max-loop` - Use loop for min/max instead of sort
+- `js-set-map-lookups` - Use Set/Map for O(1) lookups
+- `js-tosorted-immutable` - Use toSorted() for immutability
+
+### 8. Advanced Patterns (LOW)
+
+- `advanced-event-handler-refs` - Store event handlers in refs
+- `advanced-use-latest` - useLatest for stable callback refs
+
+## How to Use
+
+Read individual rule files for detailed explanations and code examples:
+
+```
+rules/async-parallel.md
+rules/bundle-barrel-imports.md
+rules/_sections.md
+```
+
+Each rule file contains:
+- Brief explanation of why it matters
+- Incorrect code example with explanation
+- Correct code example with explanation
+- Additional context and references
+
+## Full Compiled Document
+
+For the complete guide with all rules expanded: `AGENTS.md`

.claude/skills/vercel-react-best-practices/rules/advanced-event-handler-refs.md

@@ -0,0 +1,55 @@
+---
+title: Store Event Handlers in Refs
+impact: LOW
+impactDescription: stable subscriptions
+tags: advanced, hooks, refs, event-handlers, optimization
+---
+
+## Store Event Handlers in Refs
+
+Store callbacks in refs when used in effects that shouldn't re-subscribe on callback changes.
+
+**Incorrect (re-subscribes on every render):**
+
+```tsx
+function useWindowEvent(event: string, handler: (e) => void) {
+  useEffect(() => {
+    window.addEventListener(event, handler)
+    return () => window.removeEventListener(event, handler)
+  }, [event, handler])
+}
+```
+
+**Correct (stable subscription):**
+
+```tsx
+function useWindowEvent(event: string, handler: (e) => void) {
+  const handlerRef = useRef(handler)
+  useEffect(() => {
+    handlerRef.current = handler
+  }, [handler])
+
+  useEffect(() => {
+    const listener = (e) => handlerRef.current(e)
+    window.addEventListener(event, listener)
+    return () => window.removeEventListener(event, listener)
+  }, [event])
+}
+```
+
+**Alternative: use `useEffectEvent` if you're on latest React:**
+
+```tsx
+import { useEffectEvent } from 'react'
+
+function useWindowEvent(event: string, handler: (e) => void) {
+  const onEvent = useEffectEvent(handler)
+
+  useEffect(() => {
+    window.addEventListener(event, onEvent)
+    return () => window.removeEventListener(event, onEvent)
+  }, [event])
+}
+```
+
+`useEffectEvent` provides a cleaner API for the same pattern: it creates a stable function reference that always calls the latest version of the handler.

.claude/skills/vercel-react-best-practices/rules/advanced-use-latest.md

@@ -0,0 +1,49 @@
+---
+title: useLatest for Stable Callback Refs
+impact: LOW
+impactDescription: prevents effect re-runs
+tags: advanced, hooks, useLatest, refs, optimization
+---
+
+## useLatest for Stable Callback Refs
+
+Access latest values in callbacks without adding them to dependency arrays. Prevents effect re-runs while avoiding stale closures.
+
+**Implementation:**
+
+```typescript
+function useLatest<T>(value: T) {
+  const ref = useRef(value)
+  useLayoutEffect(() => {
+    ref.current = value
+  }, [value])
+  return ref
+}
+```
+
+**Incorrect (effect re-runs on every callback change):**
+
+```tsx
+function SearchInput({ onSearch }: { onSearch: (q: string) => void }) {
+  const [query, setQuery] = useState('')
+
+  useEffect(() => {
+    const timeout = setTimeout(() => onSearch(query), 300)
+    return () => clearTimeout(timeout)
+  }, [query, onSearch])
+}
+```
+
+**Correct (stable effect, fresh callback):**
+
+```tsx
+function SearchInput({ onSearch }: { onSearch: (q: string) => void }) {
+  const [query, setQuery] = useState('')
+  const onSearchRef = useLatest(onSearch)
+
+  useEffect(() => {
+    const timeout = setTimeout(() => onSearchRef.current(query), 300)
+    return () => clearTimeout(timeout)
+  }, [query])
+}
+```

.claude/skills/vercel-react-best-practices/rules/async-api-routes.md

@@ -0,0 +1,38 @@
+---
+title: Prevent Waterfall Chains in API Routes
+impact: CRITICAL
+impactDescription: 2-10× improvement
+tags: api-routes, server-actions, waterfalls, parallelization
+---
+
+## Prevent Waterfall Chains in API Routes
+
+In API routes and Server Actions, start independent operations immediately, even if you don't await them yet.
+
+**Incorrect (config waits for auth, data waits for both):**
+
+```typescript
+export async function GET(request: Request) {
+  const session = await auth()
+  const config = await fetchConfig()
+  const data = await fetchData(session.user.id)
+  return Response.json({ data, config })
+}
+```
+
+**Correct (auth and config start immediately):**
+
+```typescript
+export async function GET(request: Request) {
+  const sessionPromise = auth()
+  const configPromise = fetchConfig()
+  const session = await sessionPromise
+  const [config, data] = await Promise.all([
+    configPromise,
+    fetchData(session.user.id)
+  ])
+  return Response.json({ data, config })
+}
+```
+
+For operations with more complex dependency chains, use `better-all` to automatically maximize parallelism (see Dependency-Based Parallelization).

.claude/skills/vercel-react-best-practices/rules/async-defer-await.md

@@ -0,0 +1,80 @@
+---
+title: Defer Await Until Needed
+impact: HIGH
+impactDescription: avoids blocking unused code paths
+tags: async, await, conditional, optimization
+---
+
+## Defer Await Until Needed
+
+Move `await` operations into the branches where they're actually used to avoid blocking code paths that don't need them.
+
+**Incorrect (blocks both branches):**
+
+```typescript
+async function handleRequest(userId: string, skipProcessing: boolean) {
+  const userData = await fetchUserData(userId)
+  
+  if (skipProcessing) {
+    // Returns immediately but still waited for userData
+    return { skipped: true }
+  }
+  
+  // Only this branch uses userData
+  return processUserData(userData)
+}
+```
+
+**Correct (only blocks when needed):**
+
+```typescript
+async function handleRequest(userId: string, skipProcessing: boolean) {
+  if (skipProcessing) {
+    // Returns immediately without waiting
+    return { skipped: true }
+  }
+  
+  // Fetch only when needed
+  const userData = await fetchUserData(userId)
+  return processUserData(userData)
+}
+```
+
+**Another example (early return optimization):**
+
+```typescript
+// Incorrect: always fetches permissions
+async function updateResource(resourceId: string, userId: string) {
+  const permissions = await fetchPermissions(userId)
+  const resource = await getResource(resourceId)
+  
+  if (!resource) {
+    return { error: 'Not found' }
+  }
+  
+  if (!permissions.canEdit) {
+    return { error: 'Forbidden' }
+  }
+  
+  return await updateResourceData(resource, permissions)
+}
+
+// Correct: fetches only when needed
+async function updateResource(resourceId: string, userId: string) {
+  const resource = await getResource(resourceId)
+  
+  if (!resource) {
+    return { error: 'Not found' }
+  }
+  
+  const permissions = await fetchPermissions(userId)
+  
+  if (!permissions.canEdit) {
+    return { error: 'Forbidden' }
+  }
+  
+  return await updateResourceData(resource, permissions)
+}
+```
+
+This optimization is especially valuable when the skipped branch is frequently taken, or when the deferred operation is expensive.

.claude/skills/vercel-react-best-practices/rules/async-dependencies.md

@@ -0,0 +1,36 @@
+---
+title: Dependency-Based Parallelization
+impact: CRITICAL
+impactDescription: 2-10× improvement
+tags: async, parallelization, dependencies, better-all
+---
+
+## Dependency-Based Parallelization
+
+For operations with partial dependencies, use `better-all` to maximize parallelism. It automatically starts each task at the earliest possible moment.
+
+**Incorrect (profile waits for config unnecessarily):**
+
+```typescript
+const [user, config] = await Promise.all([
+  fetchUser(),
+  fetchConfig()
+])
+const profile = await fetchProfile(user.id)
+```
+
+**Correct (config and profile run in parallel):**
+
+```typescript
+import { all } from 'better-all'
+
+const { user, config, profile } = await all({
+  async user() { return fetchUser() },
+  async config() { return fetchConfig() },
+  async profile() {
+    return fetchProfile((await this.$.user).id)
+  }
+})
+```
+
+Reference: [https://github.com/shuding/better-all](https://github.com/shuding/better-all)

.claude/skills/vercel-react-best-practices/rules/async-parallel.md

@@ -0,0 +1,28 @@
+---
+title: Promise.all() for Independent Operations
+impact: CRITICAL
+impactDescription: 2-10× improvement
+tags: async, parallelization, promises, waterfalls
+---
+
+## Promise.all() for Independent Operations
+
+When async operations have no interdependencies, execute them concurrently using `Promise.all()`.
+
+**Incorrect (sequential execution, 3 round trips):**
+
+```typescript
+const user = await fetchUser()
+const posts = await fetchPosts()
+const comments = await fetchComments()
+```
+
+**Correct (parallel execution, 1 round trip):**
+
+```typescript
+const [user, posts, comments] = await Promise.all([
+  fetchUser(),
+  fetchPosts(),
+  fetchComments()
+])
+```

.claude/skills/vercel-react-best-practices/rules/async-suspense-boundaries.md

@@ -0,0 +1,99 @@
+---
+title: Strategic Suspense Boundaries
+impact: HIGH
+impactDescription: faster initial paint
+tags: async, suspense, streaming, layout-shift
+---
+
+## Strategic Suspense Boundaries
+
+Instead of awaiting data in async components before returning JSX, use Suspense boundaries to show the wrapper UI faster while data loads.
+
+**Incorrect (wrapper blocked by data fetching):**
+
+```tsx
+async function Page() {
+  const data = await fetchData() // Blocks entire page
+  
+  return (
+    <div>
+      <div>Sidebar</div>
+      <div>Header</div>
+      <div>
+        <DataDisplay data={data} />
+      </div>
+      <div>Footer</div>
+    </div>
+  )
+}
+```
+
+The entire layout waits for data even though only the middle section needs it.
+
+**Correct (wrapper shows immediately, data streams in):**
+
+```tsx
+function Page() {
+  return (
+    <div>
+      <div>Sidebar</div>
+      <div>Header</div>
+      <div>
+        <Suspense fallback={<Skeleton />}>
+          <DataDisplay />
+        </Suspense>
+      </div>
+      <div>Footer</div>
+    </div>
+  )
+}
+
+async function DataDisplay() {
+  const data = await fetchData() // Only blocks this component
+  return <div>{data.content}</div>
+}
+```
+
+Sidebar, Header, and Footer render immediately. Only DataDisplay waits for data.
+
+**Alternative (share promise across components):**
+
+```tsx
+function Page() {
+  // Start fetch immediately, but don't await
+  const dataPromise = fetchData()
+  
+  return (
+    <div>
+      <div>Sidebar</div>
+      <div>Header</div>
+      <Suspense fallback={<Skeleton />}>
+        <DataDisplay dataPromise={dataPromise} />
+        <DataSummary dataPromise={dataPromise} />
+      </Suspense>
+      <div>Footer</div>
+    </div>
+  )
+}
+
+function DataDisplay({ dataPromise }: { dataPromise: Promise<Data> }) {
+  const data = use(dataPromise) // Unwraps the promise
+  return <div>{data.content}</div>
+}
+
+function DataSummary({ dataPromise }: { dataPromise: Promise<Data> }) {
+  const data = use(dataPromise) // Reuses the same promise
+  return <div>{data.summary}</div>
+}
+```
+
+Both components share the same promise, so only one fetch occurs. Layout renders immediately while both components wait together.
+
+**When NOT to use this pattern:**
+
+- Critical data needed for layout decisions (affects positioning)
+- SEO-critical content above the fold
+- Small, fast queries where suspense overhead isn't worth it
+- When you want to avoid layout shift (loading → content jump)
+
+**Trade-off:** Faster initial paint vs potential layout shift. Choose based on your UX priorities.

.claude/skills/vercel-react-best-practices/rules/bundle-barrel-imports.md

@@ -0,0 +1,59 @@
+---
+title: Avoid Barrel File Imports
+impact: CRITICAL
+impactDescription: 200-800ms import cost, slow builds
+tags: bundle, imports, tree-shaking, barrel-files, performance
+---
+
+## Avoid Barrel File Imports
+
+Import directly from source files instead of barrel files to avoid loading thousands of unused modules. **Barrel files** are entry points that re-export multiple modules (e.g., `index.js` that does `export * from './module'`).
+
+Popular icon and component libraries can have **up to 10,000 re-exports** in their entry file. For many React packages, **it takes 200-800ms just to import them**, affecting both development speed and production cold starts.
+
+**Why tree-shaking doesn't help:** When a library is marked as external (not bundled), the bundler can't optimize it. If you bundle it to enable tree-shaking, builds become substantially slower analyzing the entire module graph.
+
+**Incorrect (imports entire library):**
+
+```tsx
+import { Check, X, Menu } from 'lucide-react'
+// Loads 1,583 modules, takes ~2.8s extra in dev
+// Runtime cost: 200-800ms on every cold start
+
+import { Button, TextField } from '@mui/material'
+// Loads 2,225 modules, takes ~4.2s extra in dev
+```
+
+**Correct (imports only what you need):**
+
+```tsx
+import Check from 'lucide-react/dist/esm/icons/check'
+import X from 'lucide-react/dist/esm/icons/x'
+import Menu from 'lucide-react/dist/esm/icons/menu'
+// Loads only 3 modules (~2KB vs ~1MB)
+
+import Button from '@mui/material/Button'
+import TextField from '@mui/material/TextField'
+// Loads only what you use
+```
+
+**Alternative (Next.js 13.5+):**
+
+```js
+// next.config.js - use optimizePackageImports
+module.exports = {
+  experimental: {
+    optimizePackageImports: ['lucide-react', '@mui/material']
+  }
+}
+
+// Then you can keep the ergonomic barrel imports:
+import { Check, X, Menu } from 'lucide-react'
+// Automatically transformed to direct imports at build time
+```
+
+Direct imports provide 15-70% faster dev boot, 28% faster builds, 40% faster cold starts, and significantly faster HMR.
+
+Libraries commonly affected: `lucide-react`, `@mui/material`, `@mui/icons-material`, `@tabler/icons-react`, `react-icons`, `@headlessui/react`, `@radix-ui/react-*`, `lodash`, `ramda`, `date-fns`, `rxjs`, `react-use`.
+
+Reference: [How we optimized package imports in Next.js](https://vercel.com/blog/how-we-optimized-package-imports-in-next-js)

.claude/skills/vercel-react-best-practices/rules/bundle-conditional.md

@@ -0,0 +1,31 @@
+---
+title: Conditional Module Loading
+impact: HIGH
+impactDescription: loads large data only when needed
+tags: bundle, conditional-loading, lazy-loading
+---
+
+## Conditional Module Loading
+
+Load large data or modules only when a feature is activated.
+
+**Example (lazy-load animation frames):**
+
+```tsx
+function AnimationPlayer({ enabled, setEnabled }: { enabled: boolean; setEnabled: React.Dispatch<React.SetStateAction<boolean>> }) {
+  const [frames, setFrames] = useState<Frame[] | null>(null)
+
+  useEffect(() => {
+    if (enabled && !frames && typeof window !== 'undefined') {
+      import('./animation-frames.js')
+        .then(mod => setFrames(mod.frames))
+        .catch(() => setEnabled(false))
+    }
+  }, [enabled, frames, setEnabled])
+
+  if (!frames) return <Skeleton />
+  return <Canvas frames={frames} />
+}
+```
+
+The `typeof window !== 'undefined'` check prevents bundling this module for SSR, optimizing server bundle size and build speed.

.claude/skills/vercel-react-best-practices/rules/bundle-defer-third-party.md

@@ -0,0 +1,49 @@
+---
+title: Defer Non-Critical Third-Party Libraries
+impact: MEDIUM
+impactDescription: loads after hydration
+tags: bundle, third-party, analytics, defer
+---
+
+## Defer Non-Critical Third-Party Libraries
+
+Analytics, logging, and error tracking don't block user interaction. Load them after hydration.
+
+**Incorrect (blocks initial bundle):**
+
+```tsx
+import { Analytics } from '@vercel/analytics/react'
+
+export default function RootLayout({ children }) {
+  return (
+    <html>
+      <body>
+        {children}
+        <Analytics />
+      </body>
+    </html>
+  )
+}
+```
+
+**Correct (loads after hydration):**
+
+```tsx
+import dynamic from 'next/dynamic'
+
+const Analytics = dynamic(
+  () => import('@vercel/analytics/react').then(m => m.Analytics),
+  { ssr: false }
+)
+
+export default function RootLayout({ children }) {
+  return (
+    <html>
+      <body>
+        {children}
+        <Analytics />
+      </body>
+    </html>
+  )
+}
+```

.claude/skills/vercel-react-best-practices/rules/bundle-dynamic-imports.md

@@ -0,0 +1,35 @@
+---
+title: Dynamic Imports for Heavy Components
+impact: CRITICAL
+impactDescription: directly affects TTI and LCP
+tags: bundle, dynamic-import, code-splitting, next-dynamic
+---
+
+## Dynamic Imports for Heavy Components
+
+Use `next/dynamic` to lazy-load large components not needed on initial render.
+
+**Incorrect (Monaco bundles with main chunk ~300KB):**
+
+```tsx
+import { MonacoEditor } from './monaco-editor'
+
+function CodePanel({ code }: { code: string }) {
+  return <MonacoEditor value={code} />
+}
+```
+
+**Correct (Monaco loads on demand):**
+
+```tsx
+import dynamic from 'next/dynamic'
+
+const MonacoEditor = dynamic(
+  () => import('./monaco-editor').then(m => m.MonacoEditor),
+  { ssr: false }
+)
+
+function CodePanel({ code }: { code: string }) {
+  return <MonacoEditor value={code} />
+}
+```

.claude/skills/vercel-react-best-practices/rules/bundle-preload.md

@@ -0,0 +1,50 @@
+---
+title: Preload Based on User Intent
+impact: MEDIUM
+impactDescription: reduces perceived latency
+tags: bundle, preload, user-intent, hover
+---
+
+## Preload Based on User Intent
+
+Preload heavy bundles before they're needed to reduce perceived latency.
+
+**Example (preload on hover/focus):**
+
+```tsx
+function EditorButton({ onClick }: { onClick: () => void }) {
+  const preload = () => {
+    if (typeof window !== 'undefined') {
+      void import('./monaco-editor')
+    }
+  }
+
+  return (
+    <button
+      onMouseEnter={preload}
+      onFocus={preload}
+      onClick={onClick}
+    >
+      Open Editor
+    </button>
+  )
+}
+```
+
+**Example (preload when feature flag is enabled):**
+
+```tsx
+function FlagsProvider({ children, flags }: Props) {
+  useEffect(() => {
+    if (flags.editorEnabled && typeof window !== 'undefined') {
+      void import('./monaco-editor').then(mod => mod.init())
+    }
+  }, [flags.editorEnabled])
+
+  return <FlagsContext.Provider value={flags}>
+    {children}
+  </FlagsContext.Provider>
+}
+```
+
+The `typeof window !== 'undefined'` check prevents bundling preloaded modules for SSR, optimizing server bundle size and build speed.

.claude/skills/vercel-react-best-practices/rules/client-event-listeners.md

@@ -0,0 +1,74 @@
+---
+title: Deduplicate Global Event Listeners
+impact: LOW
+impactDescription: single listener for N components
+tags: client, swr, event-listeners, subscription
+---
+
+## Deduplicate Global Event Listeners
+
+Use `useSWRSubscription()` to share global event listeners across component instances.
+
+**Incorrect (N instances = N listeners):**
+
+```tsx
+function useKeyboardShortcut(key: string, callback: () => void) {
+  useEffect(() => {
+    const handler = (e: KeyboardEvent) => {
+      if (e.metaKey && e.key === key) {
+        callback()
+      }
+    }
+    window.addEventListener('keydown', handler)
+    return () => window.removeEventListener('keydown', handler)
+  }, [key, callback])
+}
+```
+
+When using the `useKeyboardShortcut` hook multiple times, each instance will register a new listener.
+
+**Correct (N instances = 1 listener):**
+
+```tsx
+import useSWRSubscription from 'swr/subscription'
+
+// Module-level Map to track callbacks per key
+const keyCallbacks = new Map<string, Set<() => void>>()
+
+function useKeyboardShortcut(key: string, callback: () => void) {
+  // Register this callback in the Map
+  useEffect(() => {
+    if (!keyCallbacks.has(key)) {
+      keyCallbacks.set(key, new Set())
+    }
+    keyCallbacks.get(key)!.add(callback)
+
+    return () => {
+      const set = keyCallbacks.get(key)
+      if (set) {
+        set.delete(callback)
+        if (set.size === 0) {
+          keyCallbacks.delete(key)
+        }
+      }
+    }
+  }, [key, callback])
+
+  useSWRSubscription('global-keydown', () => {
+    const handler = (e: KeyboardEvent) => {
+      if (e.metaKey && keyCallbacks.has(e.key)) {
+        keyCallbacks.get(e.key)!.forEach(cb => cb())
+      }
+    }
+    window.addEventListener('keydown', handler)
+    return () => window.removeEventListener('keydown', handler)
+  })
+}
+
+function Profile() {
+  // Multiple shortcuts will share the same listener
+  useKeyboardShortcut('p', () => { /* ... */ }) 
+  useKeyboardShortcut('k', () => { /* ... */ })
+  // ...
+}
+```

.claude/skills/vercel-react-best-practices/rules/client-localstorage-schema.md

@@ -0,0 +1,71 @@
+---
+title: Version and Minimize localStorage Data
+impact: MEDIUM
+impactDescription: prevents schema conflicts, reduces storage size
+tags: client, localStorage, storage, versioning, data-minimization
+---
+
+## Version and Minimize localStorage Data
+
+Add version prefix to keys and store only needed fields. Prevents schema conflicts and accidental storage of sensitive data.
+
+**Incorrect:**
+
+```typescript
+// No version, stores everything, no error handling
+localStorage.setItem('userConfig', JSON.stringify(fullUserObject))
+const data = localStorage.getItem('userConfig')
+```
+
+**Correct:**
+
+```typescript
+const VERSION = 'v2'
+
+function saveConfig(config: { theme: string; language: string }) {
+  try {
+    localStorage.setItem(`userConfig:${VERSION}`, JSON.stringify(config))
+  } catch {
+    // Throws in incognito/private browsing, quota exceeded, or disabled
+  }
+}
+
+function loadConfig() {
+  try {
+    const data = localStorage.getItem(`userConfig:${VERSION}`)
+    return data ? JSON.parse(data) : null
+  } catch {
+    return null
+  }
+}
+
+// Migration from v1 to v2
+function migrate() {
+  try {
+    const v1 = localStorage.getItem('userConfig:v1')
+    if (v1) {
+      const old = JSON.parse(v1)
+      saveConfig({ theme: old.darkMode ? 'dark' : 'light', language: old.lang })
+      localStorage.removeItem('userConfig:v1')
+    }
+  } catch {}
+}
+```
+
+**Store minimal fields from server responses:**
+
+```typescript
+// User object has 20+ fields, only store what UI needs
+function cachePrefs(user: FullUser) {
+  try {
+    localStorage.setItem('prefs:v1', JSON.stringify({
+      theme: user.preferences.theme,
+      notifications: user.preferences.notifications
+    }))
+  } catch {}
+}
+```
+
+**Always wrap in try-catch:** `getItem()` and `setItem()` throw in incognito/private browsing (Safari, Firefox), when quota exceeded, or when disabled.
+
+**Benefits:** Schema evolution via versioning, reduced storage size, prevents storing tokens/PII/internal flags.

.claude/skills/vercel-react-best-practices/rules/client-passive-event-listeners.md

@@ -0,0 +1,48 @@
+---
+title: Use Passive Event Listeners for Scrolling Performance
+impact: MEDIUM
+impactDescription: eliminates scroll delay caused by event listeners
+tags: client, event-listeners, scrolling, performance, touch, wheel
+---
+
+## Use Passive Event Listeners for Scrolling Performance
+
+Add `{ passive: true }` to touch and wheel event listeners to enable immediate scrolling. Browsers normally wait for listeners to finish to check if `preventDefault()` is called, causing scroll delay.
+
+**Incorrect:**
+
+```typescript
+useEffect(() => {
+  const handleTouch = (e: TouchEvent) => console.log(e.touches[0].clientX)
+  const handleWheel = (e: WheelEvent) => console.log(e.deltaY)
+  
+  document.addEventListener('touchstart', handleTouch)
+  document.addEventListener('wheel', handleWheel)
+  
+  return () => {
+    document.removeEventListener('touchstart', handleTouch)
+    document.removeEventListener('wheel', handleWheel)
+  }
+}, [])
+```
+
+**Correct:**
+
+```typescript
+useEffect(() => {
+  const handleTouch = (e: TouchEvent) => console.log(e.touches[0].clientX)
+  const handleWheel = (e: WheelEvent) => console.log(e.deltaY)
+  
+  document.addEventListener('touchstart', handleTouch, { passive: true })
+  document.addEventListener('wheel', handleWheel, { passive: true })
+  
+  return () => {
+    document.removeEventListener('touchstart', handleTouch)
+    document.removeEventListener('wheel', handleWheel)
+  }
+}, [])
+```
+
+**Use passive when:** tracking/analytics, logging, any listener that doesn't call `preventDefault()`.
+
+**Don't use passive when:** implementing custom swipe gestures, custom zoom controls, or any listener that needs `preventDefault()`.

.claude/skills/vercel-react-best-practices/rules/client-swr-dedup.md

@@ -0,0 +1,56 @@
+---
+title: Use SWR for Automatic Deduplication
+impact: MEDIUM-HIGH
+impactDescription: automatic deduplication
+tags: client, swr, deduplication, data-fetching
+---
+
+## Use SWR for Automatic Deduplication
+
+SWR enables request deduplication, caching, and revalidation across component instances.
+
+**Incorrect (no deduplication, each instance fetches):**
+
+```tsx
+function UserList() {
+  const [users, setUsers] = useState([])
+  useEffect(() => {
+    fetch('/api/users')
+      .then(r => r.json())
+      .then(setUsers)
+  }, [])
+}
+```
+
+**Correct (multiple instances share one request):**
+
+```tsx
+import useSWR from 'swr'
+
+function UserList() {
+  const { data: users } = useSWR('/api/users', fetcher)
+}
+```
+
+**For immutable data:**
+
+```tsx
+import { useImmutableSWR } from '@/lib/swr'
+
+function StaticContent() {
+  const { data } = useImmutableSWR('/api/config', fetcher)
+}
+```
+
+**For mutations:**
+
+```tsx
+import { useSWRMutation } from 'swr/mutation'
+
+function UpdateButton() {
+  const { trigger } = useSWRMutation('/api/user', updateUser)
+  return <button onClick={() => trigger()}>Update</button>
+}
+```
+
+Reference: [https://swr.vercel.app](https://swr.vercel.app)

.claude/skills/vercel-react-best-practices/rules/js-batch-dom-css.md

@@ -0,0 +1,57 @@
+---
+title: Batch DOM CSS Changes
+impact: MEDIUM
+impactDescription: reduces reflows/repaints
+tags: javascript, dom, css, performance, reflow
+---
+
+## Batch DOM CSS Changes
+
+Avoid interleaving style writes with layout reads. When you read a layout property (like `offsetWidth`, `getBoundingClientRect()`, or `getComputedStyle()`) between style changes, the browser is forced to trigger a synchronous reflow.
+
+**Incorrect (interleaved reads and writes force reflows):**
+
+```typescript
+function updateElementStyles(element: HTMLElement) {
+  element.style.width = '100px'
+  const width = element.offsetWidth  // Forces reflow
+  element.style.height = '200px'
+  const height = element.offsetHeight  // Forces another reflow
+}
+```
+
+**Correct (batch writes, then read once):**
+
+```typescript
+function updateElementStyles(element: HTMLElement) {
+  // Batch all writes together
+  element.style.width = '100px'
+  element.style.height = '200px'
+  element.style.backgroundColor = 'blue'
+  element.style.border = '1px solid black'
+  
+  // Read after all writes are done (single reflow)
+  const { width, height } = element.getBoundingClientRect()
+}
+```
+
+**Better: use CSS classes**
+
+```css
+.highlighted-box {
+  width: 100px;
+  height: 200px;
+  background-color: blue;
+  border: 1px solid black;
+}
+```
+
+```typescript
+function updateElementStyles(element: HTMLElement) {
+  element.classList.add('highlighted-box')
+
+  const { width, height } = element.getBoundingClientRect()
+}
+```
+
+Prefer CSS classes over inline styles when possible. CSS files are cached by the browser, and classes provide better separation of concerns and are easier to maintain.

.claude/skills/vercel-react-best-practices/rules/js-cache-function-results.md

@@ -0,0 +1,80 @@
+---
+title: Cache Repeated Function Calls
+impact: MEDIUM
+impactDescription: avoid redundant computation
+tags: javascript, cache, memoization, performance
+---
+
+## Cache Repeated Function Calls
+
+Use a module-level Map to cache function results when the same function is called repeatedly with the same inputs during render.
+
+**Incorrect (redundant computation):**
+
+```typescript
+function ProjectList({ projects }: { projects: Project[] }) {
+  return (
+    <div>
+      {projects.map(project => {
+        // slugify() called 100+ times for same project names
+        const slug = slugify(project.name)
+        
+        return <ProjectCard key={project.id} slug={slug} />
+      })}
+    </div>
+  )
+}
+```
+
+**Correct (cached results):**
+
+```typescript
+// Module-level cache
+const slugifyCache = new Map<string, string>()
+
+function cachedSlugify(text: string): string {
+  if (slugifyCache.has(text)) {
+    return slugifyCache.get(text)!
+  }
+  const result = slugify(text)
+  slugifyCache.set(text, result)
+  return result
+}
+
+function ProjectList({ projects }: { projects: Project[] }) {
+  return (
+    <div>
+      {projects.map(project => {
+        // Computed only once per unique project name
+        const slug = cachedSlugify(project.name)
+        
+        return <ProjectCard key={project.id} slug={slug} />
+      })}
+    </div>
+  )
+}
+```
+
+**Simpler pattern for single-value functions:**
+
+```typescript
+let isLoggedInCache: boolean | null = null
+
+function isLoggedIn(): boolean {
+  if (isLoggedInCache !== null) {
+    return isLoggedInCache
+  }
+  
+  isLoggedInCache = document.cookie.includes('auth=')
+  return isLoggedInCache
+}
+
+// Clear cache when auth changes
+function onAuthChange() {
+  isLoggedInCache = null
+}
+```
+
+Use a Map (not a hook) so it works everywhere: utilities, event handlers, not just React components.
+
+Reference: [How we made the Vercel Dashboard twice as fast](https://vercel.com/blog/how-we-made-the-vercel-dashboard-twice-as-fast)

.claude/skills/vercel-react-best-practices/rules/js-cache-property-access.md

@@ -0,0 +1,28 @@
+---
+title: Cache Property Access in Loops
+impact: LOW-MEDIUM
+impactDescription: reduces lookups
+tags: javascript, loops, optimization, caching
+---
+
+## Cache Property Access in Loops
+
+Cache object property lookups in hot paths.
+
+**Incorrect (3 lookups × N iterations):**
+
+```typescript
+for (let i = 0; i < arr.length; i++) {
+  process(obj.config.settings.value)
+}
+```
+
+**Correct (1 lookup total):**
+
+```typescript
+const value = obj.config.settings.value
+const len = arr.length
+for (let i = 0; i < len; i++) {
+  process(value)
+}
+```

.claude/skills/vercel-react-best-practices/rules/js-cache-storage.md

@@ -0,0 +1,70 @@
+---
+title: Cache Storage API Calls
+impact: LOW-MEDIUM
+impactDescription: reduces expensive I/O
+tags: javascript, localStorage, storage, caching, performance
+---
+
+## Cache Storage API Calls
+
+`localStorage`, `sessionStorage`, and `document.cookie` are synchronous and expensive. Cache reads in memory.
+
+**Incorrect (reads storage on every call):**
+
+```typescript
+function getTheme() {
+  return localStorage.getItem('theme') ?? 'light'
+}
+// Called 10 times = 10 storage reads
+```
+
+**Correct (Map cache):**
+
+```typescript
+const storageCache = new Map<string, string | null>()
+
+function getLocalStorage(key: string) {
+  if (!storageCache.has(key)) {
+    storageCache.set(key, localStorage.getItem(key))
+  }
+  return storageCache.get(key)
+}
+
+function setLocalStorage(key: string, value: string) {
+  localStorage.setItem(key, value)
+  storageCache.set(key, value)  // keep cache in sync
+}
+```
+
+Use a Map (not a hook) so it works everywhere: utilities, event handlers, not just React components.
+
+**Cookie caching:**
+
+```typescript
+let cookieCache: Record<string, string> | null = null
+
+function getCookie(name: string) {
+  if (!cookieCache) {
+    cookieCache = Object.fromEntries(
+      document.cookie.split('; ').map(c => c.split('='))
+    )
+  }
+  return cookieCache[name]
+}
+```
+
+**Important (invalidate on external changes):**
+
+If storage can change externally (another tab, server-set cookies), invalidate cache:
+
+```typescript
+window.addEventListener('storage', (e) => {
+  if (e.key) storageCache.delete(e.key)
+})
+
+document.addEventListener('visibilitychange', () => {
+  if (document.visibilityState === 'visible') {
+    storageCache.clear()
+  }
+})
+```

.claude/skills/vercel-react-best-practices/rules/js-combine-iterations.md

@@ -0,0 +1,32 @@
+---
+title: Combine Multiple Array Iterations
+impact: LOW-MEDIUM
+impactDescription: reduces iterations
+tags: javascript, arrays, loops, performance
+---
+
+## Combine Multiple Array Iterations
+
+Multiple `.filter()` or `.map()` calls iterate the array multiple times. Combine into one loop.
+
+**Incorrect (3 iterations):**
+
+```typescript
+const admins = users.filter(u => u.isAdmin)
+const testers = users.filter(u => u.isTester)
+const inactive = users.filter(u => !u.isActive)
+```
+
+**Correct (1 iteration):**
+
+```typescript
+const admins: User[] = []
+const testers: User[] = []
+const inactive: User[] = []
+
+for (const user of users) {
+  if (user.isAdmin) admins.push(user)
+  if (user.isTester) testers.push(user)
+  if (!user.isActive) inactive.push(user)
+}
+```

.claude/skills/vercel-react-best-practices/rules/js-early-exit.md

@@ -0,0 +1,50 @@
+---
+title: Early Return from Functions
+impact: LOW-MEDIUM
+impactDescription: avoids unnecessary computation
+tags: javascript, functions, optimization, early-return
+---
+
+## Early Return from Functions
+
+Return early when result is determined to skip unnecessary processing.
+
+**Incorrect (processes all items even after finding answer):**
+
+```typescript
+function validateUsers(users: User[]) {
+  let hasError = false
+  let errorMessage = ''
+  
+  for (const user of users) {
+    if (!user.email) {
+      hasError = true
+      errorMessage = 'Email required'
+    }
+    if (!user.name) {
+      hasError = true
+      errorMessage = 'Name required'
+    }
+    // Continues checking all users even after error found
+  }
+  
+  return hasError ? { valid: false, error: errorMessage } : { valid: true }
+}
+```
+
+**Correct (returns immediately on first error):**
+
+```typescript
+function validateUsers(users: User[]) {
+  for (const user of users) {
+    if (!user.email) {
+      return { valid: false, error: 'Email required' }
+    }
+    if (!user.name) {
+      return { valid: false, error: 'Name required' }
+    }
+  }
+
+  return { valid: true }
+}
+```

.claude/skills/vercel-react-best-practices/rules/js-hoist-regexp.md

@@ -0,0 +1,45 @@
+---
+title: Hoist RegExp Creation
+impact: LOW-MEDIUM
+impactDescription: avoids recreation
+tags: javascript, regexp, optimization, memoization
+---
+
+## Hoist RegExp Creation
+
+Don't create RegExp inside render. Hoist to module scope or memoize with `useMemo()`.
+
+**Incorrect (new RegExp every render):**
+
+```tsx
+function Highlighter({ text, query }: Props) {
+  const regex = new RegExp(`(${query})`, 'gi')
+  const parts = text.split(regex)
+  return <>{parts.map((part, i) => ...)}</>
+}
+```
+
+**Correct (memoize or hoist):**
+
+```tsx
+const EMAIL_REGEX = /^[^\s@]+@[^\s@]+\.[^\s@]+$/
+
+function Highlighter({ text, query }: Props) {
+  const regex = useMemo(
+    () => new RegExp(`(${escapeRegex(query)})`, 'gi'),
+    [query]
+  )
+  const parts = text.split(regex)
+  return <>{parts.map((part, i) => ...)}</>
+}
+```
+
+**Warning (global regex has mutable state):**
+
+Global regex (`/g`) has mutable `lastIndex` state:
+
+```typescript
+const regex = /foo/g
+regex.test('foo')  // true, lastIndex = 3
+regex.test('foo')  // false, lastIndex = 0
+```

.claude/skills/vercel-react-best-practices/rules/js-index-maps.md

@@ -0,0 +1,37 @@
+---
+title: Build Index Maps for Repeated Lookups
+impact: LOW-MEDIUM
+impactDescription: 1M ops to 2K ops
+tags: javascript, map, indexing, optimization, performance
+---
+
+## Build Index Maps for Repeated Lookups
+
+Multiple `.find()` calls by the same key should use a Map.
+
+**Incorrect (O(n) per lookup):**
+
+```typescript
+function processOrders(orders: Order[], users: User[]) {
+  return orders.map(order => ({
+    ...order,
+    user: users.find(u => u.id === order.userId)
+  }))
+}
+```
+
+**Correct (O(1) per lookup):**
+
+```typescript
+function processOrders(orders: Order[], users: User[]) {
+  const userById = new Map(users.map(u => [u.id, u]))
+
+  return orders.map(order => ({
+    ...order,
+    user: userById.get(order.userId)
+  }))
+}
+```
+
+Build map once (O(n)), then all lookups are O(1).
+For 1000 orders × 1000 users: 1M ops → 2K ops.

.claude/skills/vercel-react-best-practices/rules/js-length-check-first.md

@@ -0,0 +1,49 @@
+---
+title: Early Length Check for Array Comparisons
+impact: MEDIUM-HIGH
+impactDescription: avoids expensive operations when lengths differ
+tags: javascript, arrays, performance, optimization, comparison
+---
+
+## Early Length Check for Array Comparisons
+
+When comparing arrays with expensive operations (sorting, deep equality, serialization), check lengths first. If lengths differ, the arrays cannot be equal.
+
+In real-world applications, this optimization is especially valuable when the comparison runs in hot paths (event handlers, render loops).
+
+**Incorrect (always runs expensive comparison):**
+
+```typescript
+function hasChanges(current: string[], original: string[]) {
+  // Always sorts and joins, even when lengths differ
+  return current.sort().join() !== original.sort().join()
+}
+```
+
+Two O(n log n) sorts run even when `current.length` is 5 and `original.length` is 100. There is also overhead of joining the arrays and comparing the strings.
+
+**Correct (O(1) length check first):**
+
+```typescript
+function hasChanges(current: string[], original: string[]) {
+  // Early return if lengths differ
+  if (current.length !== original.length) {
+    return true
+  }
+  // Only sort when lengths match
+  const currentSorted = current.toSorted()
+  const originalSorted = original.toSorted()
+  for (let i = 0; i < currentSorted.length; i++) {
+    if (currentSorted[i] !== originalSorted[i]) {
+      return true
+    }
+  }
+  return false
+}
+```
+
+This new approach is more efficient because:
+- It avoids the overhead of sorting and joining the arrays when lengths differ
+- It avoids consuming memory for the joined strings (especially important for large arrays)
+- It avoids mutating the original arrays
+- It returns early when a difference is found

.claude/skills/vercel-react-best-practices/rules/js-min-max-loop.md

@@ -0,0 +1,82 @@
+---
+title: Use Loop for Min/Max Instead of Sort
+impact: LOW
+impactDescription: O(n) instead of O(n log n)
+tags: javascript, arrays, performance, sorting, algorithms
+---
+
+## Use Loop for Min/Max Instead of Sort
+
+Finding the smallest or largest element only requires a single pass through the array. Sorting is wasteful and slower.
+
+**Incorrect (O(n log n) - sort to find latest):**
+
+```typescript
+interface Project {
+  id: string
+  name: string
+  updatedAt: number
+}
+
+function getLatestProject(projects: Project[]) {
+  const sorted = [...projects].sort((a, b) => b.updatedAt - a.updatedAt)
+  return sorted[0]
+}
+```
+
+Sorts the entire array just to find the maximum value.
+
+**Incorrect (O(n log n) - sort for oldest and newest):**
+
+```typescript
+function getOldestAndNewest(projects: Project[]) {
+  const sorted = [...projects].sort((a, b) => a.updatedAt - b.updatedAt)
+  return { oldest: sorted[0], newest: sorted[sorted.length - 1] }
+}
+```
+
+Still sorts unnecessarily when only min/max are needed.
+
+**Correct (O(n) - single loop):**
+
+```typescript
+function getLatestProject(projects: Project[]) {
+  if (projects.length === 0) return null
+  
+  let latest = projects[0]
+  
+  for (let i = 1; i < projects.length; i++) {
+    if (projects[i].updatedAt > latest.updatedAt) {
+      latest = projects[i]
+    }
+  }
+  
+  return latest
+}
+
+function getOldestAndNewest(projects: Project[]) {
+  if (projects.length === 0) return { oldest: null, newest: null }
+  
+  let oldest = projects[0]
+  let newest = projects[0]
+  
+  for (let i = 1; i < projects.length; i++) {
+    if (projects[i].updatedAt < oldest.updatedAt) oldest = projects[i]
+    if (projects[i].updatedAt > newest.updatedAt) newest = projects[i]
+  }
+  
+  return { oldest, newest }
+}
+```
+
+Single pass through the array, no copying, no sorting.
+
+**Alternative (Math.min/Math.max for small arrays):**
+
+```typescript
+const numbers = [5, 2, 8, 1, 9]
+const min = Math.min(...numbers)
+const max = Math.max(...numbers)
+```
+
+This works for small arrays, but can be slower or just throw an error for very large arrays due to spread operator limitations. Maximal array length is approximately 124000 in Chrome 143 and 638000 in Safari 18; exact numbers may vary - see [the fiddle](https://jsfiddle.net/qw1jabsx/4/). Use the loop approach for reliability.

.claude/skills/vercel-react-best-practices/rules/js-set-map-lookups.md

@@ -0,0 +1,24 @@
+---
+title: Use Set/Map for O(1) Lookups
+impact: LOW-MEDIUM
+impactDescription: O(n) to O(1)
+tags: javascript, set, map, data-structures, performance
+---
+
+## Use Set/Map for O(1) Lookups
+
+Convert arrays to Set/Map for repeated membership checks.
+
+**Incorrect (O(n) per check):**
+
+```typescript
+const allowedIds = ['a', 'b', 'c', ...]
+items.filter(item => allowedIds.includes(item.id))
+```
+
+**Correct (O(1) per check):**
+
+```typescript
+const allowedIds = new Set(['a', 'b', 'c', ...])
+items.filter(item => allowedIds.has(item.id))
+```

.claude/skills/vercel-react-best-practices/rules/js-tosorted-immutable.md

@@ -0,0 +1,57 @@
+---
+title: Use toSorted() Instead of sort() for Immutability
+impact: MEDIUM-HIGH
+impactDescription: prevents mutation bugs in React state
+tags: javascript, arrays, immutability, react, state, mutation
+---
+
+## Use toSorted() Instead of sort() for Immutability
+
+`.sort()` mutates the array in place, which can cause bugs with React state and props. Use `.toSorted()` to create a new sorted array without mutation.
+
+**Incorrect (mutates original array):**
+
+```typescript
+function UserList({ users }: { users: User[] }) {
+  // Mutates the users prop array!
+  const sorted = useMemo(
+    () => users.sort((a, b) => a.name.localeCompare(b.name)),
+    [users]
+  )
+  return <div>{sorted.map(renderUser)}</div>
+}
+```
+
+**Correct (creates new array):**
+
+```typescript
+function UserList({ users }: { users: User[] }) {
+  // Creates new sorted array, original unchanged
+  const sorted = useMemo(
+    () => users.toSorted((a, b) => a.name.localeCompare(b.name)),
+    [users]
+  )
+  return <div>{sorted.map(renderUser)}</div>
+}
+```
+
+**Why this matters in React:**
+
+1. Props/state mutations break React's immutability model - React expects props and state to be treated as read-only
+2. Causes stale closure bugs - Mutating arrays inside closures (callbacks, effects) can lead to unexpected behavior
+
+**Browser support (fallback for older browsers):**
+
+`.toSorted()` is available in all modern browsers (Chrome 110+, Safari 16+, Firefox 115+, Node.js 20+). For older environments, use spread operator:
+
+```typescript
+// Fallback for older browsers
+const sorted = [...items].sort((a, b) => a.value - b.value)
+```
+
+**Other immutable array methods:**
+
+- `.toSorted()` - immutable sort
+- `.toReversed()` - immutable reverse
+- `.toSpliced()` - immutable splice
+- `.with()` - immutable element replacement

.claude/skills/vercel-react-best-practices/rules/rendering-activity.md

@@ -0,0 +1,26 @@
+---
+title: Use Activity Component for Show/Hide
+impact: MEDIUM
+impactDescription: preserves state/DOM
+tags: rendering, activity, visibility, state-preservation
+---
+
+## Use Activity Component for Show/Hide
+
+Use React's `<Activity>` to preserve state/DOM for expensive components that frequently toggle visibility.
+
+**Usage:**
+
+```tsx
+import { Activity } from 'react'
+
+function Dropdown({ isOpen }: Props) {
+  return (
+    <Activity mode={isOpen ? 'visible' : 'hidden'}>
+      <ExpensiveMenu />
+    </Activity>
+  )
+}
+```
+
+Avoids expensive re-renders and state loss.

.claude/skills/vercel-react-best-practices/rules/rendering-animate-svg-wrapper.md

@@ -0,0 +1,47 @@
+---
+title: Animate SVG Wrapper Instead of SVG Element
+impact: LOW
+impactDescription: enables hardware acceleration
+tags: rendering, svg, css, animation, performance
+---
+
+## Animate SVG Wrapper Instead of SVG Element
+
+Many browsers don't have hardware acceleration for CSS3 animations on SVG elements. Wrap SVG in a `<div>` and animate the wrapper instead.
+
+**Incorrect (animating SVG directly - no hardware acceleration):**
+
+```tsx
+function LoadingSpinner() {
+  return (
+    <svg 
+      className="animate-spin"
+      width="24" 
+      height="24" 
+      viewBox="0 0 24 24"
+    >
+      <circle cx="12" cy="12" r="10" stroke="currentColor" />
+    </svg>
+  )
+}
+```
+
+**Correct (animating wrapper div - hardware accelerated):**
+
+```tsx
+function LoadingSpinner() {
+  return (
+    <div className="animate-spin">
+      <svg 
+        width="24" 
+        height="24" 
+        viewBox="0 0 24 24"
+      >
+        <circle cx="12" cy="12" r="10" stroke="currentColor" />
+      </svg>
+    </div>
+  )
+}
+```
+
+This applies to all CSS transforms and transitions (`transform`, `opacity`, `translate`, `scale`, `rotate`). The wrapper div allows browsers to use GPU acceleration for smoother animations.

.claude/skills/vercel-react-best-practices/rules/rendering-conditional-render.md

@@ -0,0 +1,40 @@
+---
+title: Use Explicit Conditional Rendering
+impact: LOW
+impactDescription: prevents rendering 0 or NaN
+tags: rendering, conditional, jsx, falsy-values
+---
+
+## Use Explicit Conditional Rendering
+
+Use explicit ternary operators (`? :`) instead of `&&` for conditional rendering when the condition can be `0`, `NaN`, or other falsy values that render.
+
+**Incorrect (renders "0" when count is 0):**
+
+```tsx
+function Badge({ count }: { count: number }) {
+  return (
+    <div>
+      {count && <span className="badge">{count}</span>}
+    </div>
+  )
+}
+
+// When count = 0, renders: <div>0</div>
+// When count = 5, renders: <div><span class="badge">5</span></div>
+```
+
+**Correct (renders nothing when count is 0):**
+
+```tsx
+function Badge({ count }: { count: number }) {
+  return (
+    <div>
+      {count > 0 ? <span className="badge">{count}</span> : null}
+    </div>
+  )
+}
+
+// When count = 0, renders: <div></div>
+// When count = 5, renders: <div><span class="badge">5</span></div>
+```

.claude/skills/vercel-react-best-practices/rules/rendering-content-visibility.md

@@ -0,0 +1,38 @@
+---
+title: CSS content-visibility for Long Lists
+impact: HIGH
+impactDescription: faster initial render
+tags: rendering, css, content-visibility, long-lists
+---
+
+## CSS content-visibility for Long Lists
+
+Apply `content-visibility: auto` to defer off-screen rendering.
+
+**CSS:**
+
+```css
+.message-item {
+  content-visibility: auto;
+  contain-intrinsic-size: 0 80px;
+}
+```
+
+**Example:**
+
+```tsx
+function MessageList({ messages }: { messages: Message[] }) {
+  return (
+    <div className="overflow-y-auto h-screen">
+      {messages.map(msg => (
+        <div key={msg.id} className="message-item">
+          <Avatar user={msg.author} />
+          <div>{msg.content}</div>
+        </div>
+      ))}
+    </div>
+  )
+}
+```
+
+For 1000 messages, browser skips layout/paint for ~990 off-screen items (10× faster initial render).

.claude/skills/vercel-react-best-practices/rules/rendering-hoist-jsx.md

@@ -0,0 +1,46 @@
+---
+title: Hoist Static JSX Elements
+impact: LOW
+impactDescription: avoids re-creation
+tags: rendering, jsx, static, optimization
+---
+
+## Hoist Static JSX Elements
+
+Extract static JSX outside components to avoid re-creation.
+
+**Incorrect (recreates element every render):**
+
+```tsx
+function LoadingSkeleton() {
+  return <div className="animate-pulse h-20 bg-gray-200" />
+}
+
+function Container() {
+  return (
+    <div>
+      {loading && <LoadingSkeleton />}
+    </div>
+  )
+}
+```
+
+**Correct (reuses same element):**
+
+```tsx
+const loadingSkeleton = (
+  <div className="animate-pulse h-20 bg-gray-200" />
+)
+
+function Container() {
+  return (
+    <div>
+      {loading && loadingSkeleton}
+    </div>
+  )
+}
+```
+
+This is especially helpful for large and static SVG nodes, which can be expensive to recreate on every render.
+
+**Note:** If your project has [React Compiler](https://react.dev/learn/react-compiler) enabled, the compiler automatically hoists static JSX elements and optimizes component re-renders, making manual hoisting unnecessary.

.claude/skills/vercel-react-best-practices/rules/rendering-hydration-no-flicker.md

@@ -0,0 +1,82 @@
+---
+title: Prevent Hydration Mismatch Without Flickering
+impact: MEDIUM
+impactDescription: avoids visual flicker and hydration errors
+tags: rendering, ssr, hydration, localStorage, flicker
+---
+
+## Prevent Hydration Mismatch Without Flickering
+
+When rendering content that depends on client-side storage (localStorage, cookies), avoid both SSR breakage and post-hydration flickering by injecting a synchronous script that updates the DOM before React hydrates.
+
+**Incorrect (breaks SSR):**
+
+```tsx
+function ThemeWrapper({ children }: { children: ReactNode }) {
+  // localStorage is not available on server - throws error
+  const theme = localStorage.getItem('theme') || 'light'
+  
+  return (
+    <div className={theme}>
+      {children}
+    </div>
+  )
+}
+```
+
+Server-side rendering will fail because `localStorage` is undefined.
+
+**Incorrect (visual flickering):**
+
+```tsx
+function ThemeWrapper({ children }: { children: ReactNode }) {
+  const [theme, setTheme] = useState('light')
+  
+  useEffect(() => {
+    // Runs after hydration - causes visible flash
+    const stored = localStorage.getItem('theme')
+    if (stored) {
+      setTheme(stored)
+    }
+  }, [])
+  
+  return (
+    <div className={theme}>
+      {children}
+    </div>
+  )
+}
+```
+
+Component first renders with default value (`light`), then updates after hydration, causing a visible flash of incorrect content.
+
+**Correct (no flicker, no hydration mismatch):**
+
+```tsx
+function ThemeWrapper({ children }: { children: ReactNode }) {
+  return (
+    <>
+      <div id="theme-wrapper">
+        {children}
+      </div>
+      <script
+        dangerouslySetInnerHTML={{
+          __html: `
+            (function() {
+              try {
+                var theme = localStorage.getItem('theme') || 'light';
+                var el = document.getElementById('theme-wrapper');
+                if (el) el.className = theme;
+              } catch (e) {}
+            })();
+          `,
+        }}
+      />
+    </>
+  )
+}
+```
+
+The inline script executes synchronously before showing the element, ensuring the DOM already has the correct value. No flickering, no hydration mismatch.
+
+This pattern is especially useful for theme toggles, user preferences, authentication states, and any client-only data that should render immediately without flashing default values.

.claude/skills/vercel-react-best-practices/rules/rendering-svg-precision.md

@@ -0,0 +1,28 @@
+---
+title: Optimize SVG Precision
+impact: LOW
+impactDescription: reduces file size
+tags: rendering, svg, optimization, svgo
+---
+
+## Optimize SVG Precision
+
+Reduce SVG coordinate precision to decrease file size. The optimal precision depends on the viewBox size, but in general reducing precision should be considered.
+
+**Incorrect (excessive precision):**
+
+```svg
+<path d="M 10.293847 20.847362 L 30.938472 40.192837" />
+```
+
+**Correct (1 decimal place):**
+
+```svg
+<path d="M 10.3 20.8 L 30.9 40.2" />
+```
+
+**Automate with SVGO:**
+
+```bash
+npx svgo --precision=1 --multipass icon.svg
+```

.claude/skills/vercel-react-best-practices/rules/rerender-defer-reads.md

@@ -0,0 +1,39 @@
+---
+title: Defer State Reads to Usage Point
+impact: MEDIUM
+impactDescription: avoids unnecessary subscriptions
+tags: rerender, searchParams, localStorage, optimization
+---
+
+## Defer State Reads to Usage Point
+
+Don't subscribe to dynamic state (searchParams, localStorage) if you only read it inside callbacks.
+
+**Incorrect (subscribes to all searchParams changes):**
+
+```tsx
+function ShareButton({ chatId }: { chatId: string }) {
+  const searchParams = useSearchParams()
+
+  const handleShare = () => {
+    const ref = searchParams.get('ref')
+    shareChat(chatId, { ref })
+  }
+
+  return <button onClick={handleShare}>Share</button>
+}
+```
+
+**Correct (reads on demand, no subscription):**
+
+```tsx
+function ShareButton({ chatId }: { chatId: string }) {
+  const handleShare = () => {
+    const params = new URLSearchParams(window.location.search)
+    const ref = params.get('ref')
+    shareChat(chatId, { ref })
+  }
+
+  return <button onClick={handleShare}>Share</button>
+}
+```

.claude/skills/vercel-react-best-practices/rules/rerender-dependencies.md

@@ -0,0 +1,45 @@
+---
+title: Narrow Effect Dependencies
+impact: LOW
+impactDescription: minimizes effect re-runs
+tags: rerender, useEffect, dependencies, optimization
+---
+
+## Narrow Effect Dependencies
+
+Specify primitive dependencies instead of objects to minimize effect re-runs.
+
+**Incorrect (re-runs on any user field change):**
+
+```tsx
+useEffect(() => {
+  console.log(user.id)
+}, [user])
+```
+
+**Correct (re-runs only when id changes):**
+
+```tsx
+useEffect(() => {
+  console.log(user.id)
+}, [user.id])
+```
+
+**For derived state, compute outside effect:**
+
+```tsx
+// Incorrect: runs on width=767, 766, 765...
+useEffect(() => {
+  if (width < 768) {
+    enableMobileMode()
+  }
+}, [width])
+
+// Correct: runs only on boolean transition
+const isMobile = width < 768
+useEffect(() => {
+  if (isMobile) {
+    enableMobileMode()
+  }
+}, [isMobile])
+```

.claude/skills/vercel-react-best-practices/rules/rerender-derived-state.md

@@ -0,0 +1,29 @@
+---
+title: Subscribe to Derived State
+impact: MEDIUM
+impactDescription: reduces re-render frequency
+tags: rerender, derived-state, media-query, optimization
+---
+
+## Subscribe to Derived State
+
+Subscribe to derived boolean state instead of continuous values to reduce re-render frequency.
+
+**Incorrect (re-renders on every pixel change):**
+
+```tsx
+function Sidebar() {
+  const width = useWindowWidth()  // updates continuously
+  const isMobile = width < 768
+  return <nav className={isMobile ? 'mobile' : 'desktop'} />
+}
+```
+
+**Correct (re-renders only when boolean changes):**
+
+```tsx
+function Sidebar() {
+  const isMobile = useMediaQuery('(max-width: 767px)')
+  return <nav className={isMobile ? 'mobile' : 'desktop'} />
+}
+```

.claude/skills/vercel-react-best-practices/rules/rerender-functional-setstate.md

@@ -0,0 +1,74 @@
+---
+title: Use Functional setState Updates
+impact: MEDIUM
+impactDescription: prevents stale closures and unnecessary callback recreations
+tags: react, hooks, useState, useCallback, callbacks, closures
+---
+
+## Use Functional setState Updates
+
+When updating state based on the current state value, use the functional update form of setState instead of directly referencing the state variable. This prevents stale closures, eliminates unnecessary dependencies, and creates stable callback references.
+
+**Incorrect (requires state as dependency):**
+
+```tsx
+function TodoList() {
+  const [items, setItems] = useState(initialItems)
+  
+  // Callback must depend on items, recreated on every items change
+  const addItems = useCallback((newItems: Item[]) => {
+    setItems([...items, ...newItems])
+  }, [items])  // ❌ items dependency causes recreations
+  
+  // Risk of stale closure if dependency is forgotten
+  const removeItem = useCallback((id: string) => {
+    setItems(items.filter(item => item.id !== id))
+  }, [])  // ❌ Missing items dependency - will use stale items!
+  
+  return <ItemsEditor items={items} onAdd={addItems} onRemove={removeItem} />
+}
+```
+
+The first callback is recreated every time `items` changes, which can cause child components to re-render unnecessarily. The second callback has a stale closure bug—it will always reference the initial `items` value.
+
+**Correct (stable callbacks, no stale closures):**
+
+```tsx
+function TodoList() {
+  const [items, setItems] = useState(initialItems)
+  
+  // Stable callback, never recreated
+  const addItems = useCallback((newItems: Item[]) => {
+    setItems(curr => [...curr, ...newItems])
+  }, [])  // ✅ No dependencies needed
+  
+  // Always uses latest state, no stale closure risk
+  const removeItem = useCallback((id: string) => {
+    setItems(curr => curr.filter(item => item.id !== id))
+  }, [])  // ✅ Safe and stable
+  
+  return <ItemsEditor items={items} onAdd={addItems} onRemove={removeItem} />
+}
+```
+
+**Benefits:**
+
+1. **Stable callback references** - Callbacks don't need to be recreated when state changes
+2. **No stale closures** - Always operates on the latest state value
+3. **Fewer dependencies** - Simplifies dependency arrays and reduces memory leaks
+4. **Prevents bugs** - Eliminates the most common source of React closure bugs
+
+**When to use functional updates:**
+
+- Any setState that depends on the current state value
+- Inside useCallback/useMemo when state is needed
+- Event handlers that reference state
+- Async operations that update state
+
+**When direct updates are fine:**
+
+- Setting state to a static value: `setCount(0)`
+- Setting state from props/arguments only: `setName(newName)`
+- State doesn't depend on previous value
+
+**Note:** If your project has [React Compiler](https://react.dev/learn/react-compiler) enabled, the compiler can automatically optimize some cases, but functional updates are still recommended for correctness and to prevent stale closure bugs.

.claude/skills/vercel-react-best-practices/rules/rerender-lazy-state-init.md

@@ -0,0 +1,58 @@
+---
+title: Use Lazy State Initialization
+impact: MEDIUM
+impactDescription: wasted computation on every render
+tags: react, hooks, useState, performance, initialization
+---
+
+## Use Lazy State Initialization
+
+Pass a function to `useState` for expensive initial values. Without the function form, the initializer runs on every render even though the value is only used once.
+
+**Incorrect (runs on every render):**
+
+```tsx
+function FilteredList({ items }: { items: Item[] }) {
+  // buildSearchIndex() runs on EVERY render, even after initialization
+  const [searchIndex, setSearchIndex] = useState(buildSearchIndex(items))
+  const [query, setQuery] = useState('')
+  
+  // When query changes, buildSearchIndex runs again unnecessarily
+  return <SearchResults index={searchIndex} query={query} />
+}
+
+function UserProfile() {
+  // JSON.parse runs on every render
+  const [settings, setSettings] = useState(
+    JSON.parse(localStorage.getItem('settings') || '{}')
+  )
+  
+  return <SettingsForm settings={settings} onChange={setSettings} />
+}
+```
+
+**Correct (runs only once):**
+
+```tsx
+function FilteredList({ items }: { items: Item[] }) {
+  // buildSearchIndex() runs ONLY on initial render
+  const [searchIndex, setSearchIndex] = useState(() => buildSearchIndex(items))
+  const [query, setQuery] = useState('')
+  
+  return <SearchResults index={searchIndex} query={query} />
+}
+
+function UserProfile() {
+  // JSON.parse runs only on initial render
+  const [settings, setSettings] = useState(() => {
+    const stored = localStorage.getItem('settings')
+    return stored ? JSON.parse(stored) : {}
+  })
+  
+  return <SettingsForm settings={settings} onChange={setSettings} />
+}
+```
+
+Use lazy initialization when computing initial values from localStorage/sessionStorage, building data structures (indexes, maps), reading from the DOM, or performing heavy transformations.
+
+For simple primitives (`useState(0)`), direct references (`useState(props.value)`), or cheap literals (`useState({})`), the function form is unnecessary.

.claude/skills/vercel-react-best-practices/rules/rerender-memo.md

@@ -0,0 +1,44 @@
+---
+title: Extract to Memoized Components
+impact: MEDIUM
+impactDescription: enables early returns
+tags: rerender, memo, useMemo, optimization
+---
+
+## Extract to Memoized Components
+
+Extract expensive work into memoized components to enable early returns before computation.
+
+**Incorrect (computes avatar even when loading):**
+
+```tsx
+function Profile({ user, loading }: Props) {
+  const avatar = useMemo(() => {
+    const id = computeAvatarId(user)
+    return <Avatar id={id} />
+  }, [user])
+
+  if (loading) return <Skeleton />
+  return <div>{avatar}</div>
+}
+```
+
+**Correct (skips computation when loading):**
+
+```tsx
+const UserAvatar = memo(function UserAvatar({ user }: { user: User }) {
+  const id = useMemo(() => computeAvatarId(user), [user])
+  return <Avatar id={id} />
+})
+
+function Profile({ user, loading }: Props) {
+  if (loading) return <Skeleton />
+  return (
+    <div>
+      <UserAvatar user={user} />
+    </div>
+  )
+}
+```
+
+**Note:** If your project has [React Compiler](https://react.dev/learn/react-compiler) enabled, manual memoization with `memo()` and `useMemo()` is not necessary. The compiler automatically optimizes re-renders.

.claude/skills/vercel-react-best-practices/rules/rerender-transitions.md

@@ -0,0 +1,40 @@
+---
+title: Use Transitions for Non-Urgent Updates
+impact: MEDIUM
+impactDescription: maintains UI responsiveness
+tags: rerender, transitions, startTransition, performance
+---
+
+## Use Transitions for Non-Urgent Updates
+
+Mark frequent, non-urgent state updates as transitions to maintain UI responsiveness.
+
+**Incorrect (blocks UI on every scroll):**
+
+```tsx
+function ScrollTracker() {
+  const [scrollY, setScrollY] = useState(0)
+  useEffect(() => {
+    const handler = () => setScrollY(window.scrollY)
+    window.addEventListener('scroll', handler, { passive: true })
+    return () => window.removeEventListener('scroll', handler)
+  }, [])
+}
+```
+
+**Correct (non-blocking updates):**
+
+```tsx
+import { startTransition } from 'react'
+
+function ScrollTracker() {
+  const [scrollY, setScrollY] = useState(0)
+  useEffect(() => {
+    const handler = () => {
+      startTransition(() => setScrollY(window.scrollY))
+    }
+    window.addEventListener('scroll', handler, { passive: true })
+    return () => window.removeEventListener('scroll', handler)
+  }, [])
+}
+```

.claude/skills/vercel-react-best-practices/rules/server-after-nonblocking.md

@@ -0,0 +1,73 @@
+---
+title: Use after() for Non-Blocking Operations
+impact: MEDIUM
+impactDescription: faster response times
+tags: server, async, logging, analytics, side-effects
+---
+
+## Use after() for Non-Blocking Operations
+
+Use Next.js's `after()` to schedule work that should execute after a response is sent. This prevents logging, analytics, and other side effects from blocking the response.
+
+**Incorrect (blocks response):**
+
+```tsx
+import { logUserAction } from '@/app/utils'
+
+export async function POST(request: Request) {
+  // Perform mutation
+  await updateDatabase(request)
+  
+  // Logging blocks the response
+  const userAgent = request.headers.get('user-agent') || 'unknown'
+  await logUserAction({ userAgent })
+  
+  return new Response(JSON.stringify({ status: 'success' }), {
+    status: 200,
+    headers: { 'Content-Type': 'application/json' }
+  })
+}
+```
+
+**Correct (non-blocking):**
+
+```tsx
+import { after } from 'next/server'
+import { headers, cookies } from 'next/headers'
+import { logUserAction } from '@/app/utils'
+
+export async function POST(request: Request) {
+  // Perform mutation
+  await updateDatabase(request)
+  
+  // Log after response is sent
+  after(async () => {
+    const userAgent = (await headers()).get('user-agent') || 'unknown'
+    const sessionCookie = (await cookies()).get('session-id')?.value || 'anonymous'
+    
+    logUserAction({ sessionCookie, userAgent })
+  })
+  
+  return new Response(JSON.stringify({ status: 'success' }), {
+    status: 200,
+    headers: { 'Content-Type': 'application/json' }
+  })
+}
+```
+
+The response is sent immediately while logging happens in the background.
+
+**Common use cases:**
+
+- Analytics tracking
+- Audit logging
+- Sending notifications
+- Cache invalidation
+- Cleanup tasks
+
+**Important notes:**
+
+- `after()` runs even if the response fails or redirects
+- Works in Server Actions, Route Handlers, and Server Components
+
+Reference: [https://nextjs.org/docs/app/api-reference/functions/after](https://nextjs.org/docs/app/api-reference/functions/after)

.claude/skills/vercel-react-best-practices/rules/server-cache-lru.md

@@ -0,0 +1,41 @@
+---
+title: Cross-Request LRU Caching
+impact: HIGH
+impactDescription: caches across requests
+tags: server, cache, lru, cross-request
+---
+
+## Cross-Request LRU Caching
+
+`React.cache()` only works within one request. For data shared across sequential requests (user clicks button A then button B), use an LRU cache.
+
+**Implementation:**
+
+```typescript
+import { LRUCache } from 'lru-cache'
+
+const cache = new LRUCache<string, any>({
+  max: 1000,
+  ttl: 5 * 60 * 1000  // 5 minutes
+})
+
+export async function getUser(id: string) {
+  const cached = cache.get(id)
+  if (cached) return cached
+
+  const user = await db.user.findUnique({ where: { id } })
+  cache.set(id, user)
+  return user
+}
+
+// Request 1: DB query, result cached
+// Request 2: cache hit, no DB query
+```
+
+Use when sequential user actions hit multiple endpoints needing the same data within seconds.
+
+**With Vercel's [Fluid Compute](https://vercel.com/docs/fluid-compute):** LRU caching is especially effective because multiple concurrent requests can share the same function instance and cache. This means the cache persists across requests without needing external storage like Redis.
+
+**In traditional serverless:** Each invocation runs in isolation, so consider Redis for cross-process caching.
+
+Reference: [https://github.com/isaacs/node-lru-cache](https://github.com/isaacs/node-lru-cache)

.claude/skills/vercel-react-best-practices/rules/server-cache-react.md

@@ -0,0 +1,76 @@
+---
+title: Per-Request Deduplication with React.cache()
+impact: MEDIUM
+impactDescription: deduplicates within request
+tags: server, cache, react-cache, deduplication
+---
+
+## Per-Request Deduplication with React.cache()
+
+Use `React.cache()` for server-side request deduplication. Authentication and database queries benefit most.
+
+**Usage:**
+
+```typescript
+import { cache } from 'react'
+
+export const getCurrentUser = cache(async () => {
+  const session = await auth()
+  if (!session?.user?.id) return null
+  return await db.user.findUnique({
+    where: { id: session.user.id }
+  })
+})
+```
+
+Within a single request, multiple calls to `getCurrentUser()` execute the query only once.
+
+**Avoid inline objects as arguments:**
+
+`React.cache()` uses shallow equality (`Object.is`) to determine cache hits. Inline objects create new references each call, preventing cache hits.
+
+**Incorrect (always cache miss):**
+
+```typescript
+const getUser = cache(async (params: { uid: number }) => {
+  return await db.user.findUnique({ where: { id: params.uid } })
+})
+
+// Each call creates new object, never hits cache
+getUser({ uid: 1 })
+getUser({ uid: 1 })  // Cache miss, runs query again
+```
+
+**Correct (cache hit):**
+
+```typescript
+const getUser = cache(async (uid: number) => {
+  return await db.user.findUnique({ where: { id: uid } })
+})
+
+// Primitive args use value equality
+getUser(1)
+getUser(1)  // Cache hit, returns cached result
+```
+
+If you must pass objects, pass the same reference:
+
+```typescript
+const params = { uid: 1 }
+getUser(params)  // Query runs
+getUser(params)  // Cache hit (same reference)
+```
+
+**Next.js-Specific Note:**
+
+In Next.js, the `fetch` API is automatically extended with request memoization. Requests with the same URL and options are automatically deduplicated within a single request, so you don't need `React.cache()` for `fetch` calls. However, `React.cache()` is still essential for other async tasks:
+
+- Database queries (Prisma, Drizzle, etc.)
+- Heavy computations
+- Authentication checks
+- File system operations
+- Any non-fetch async work
+
+Use `React.cache()` to deduplicate these operations across your component tree.
+
+Reference: [React.cache documentation](https://react.dev/reference/react/cache)

.claude/skills/vercel-react-best-practices/rules/server-parallel-fetching.md

@@ -0,0 +1,83 @@
+---
+title: Parallel Data Fetching with Component Composition
+impact: CRITICAL
+impactDescription: eliminates server-side waterfalls
+tags: server, rsc, parallel-fetching, composition
+---
+
+## Parallel Data Fetching with Component Composition
+
+React Server Components execute sequentially within a tree. Restructure with composition to parallelize data fetching.
+
+**Incorrect (Sidebar waits for Page's fetch to complete):**
+
+```tsx
+export default async function Page() {
+  const header = await fetchHeader()
+  return (
+    <div>
+      <div>{header}</div>
+      <Sidebar />
+    </div>
+  )
+}
+
+async function Sidebar() {
+  const items = await fetchSidebarItems()
+  return <nav>{items.map(renderItem)}</nav>
+}
+```
+
+**Correct (both fetch simultaneously):**
+
+```tsx
+async function Header() {
+  const data = await fetchHeader()
+  return <div>{data}</div>
+}
+
+async function Sidebar() {
+  const items = await fetchSidebarItems()
+  return <nav>{items.map(renderItem)}</nav>
+}
+
+export default function Page() {
+  return (
+    <div>
+      <Header />
+      <Sidebar />
+    </div>
+  )
+}
+```
+
+**Alternative with children prop:**
+
+```tsx
+async function Header() {
+  const data = await fetchHeader()
+  return <div>{data}</div>
+}
+
+async function Sidebar() {
+  const items = await fetchSidebarItems()
+  return <nav>{items.map(renderItem)}</nav>
+}
+
+function Layout({ children }: { children: ReactNode }) {
+  return (
+    <div>
+      <Header />
+      {children}
+    </div>
+  )
+}
+
+export default function Page() {
+  return (
+    <Layout>
+      <Sidebar />
+    </Layout>
+  )
+}
+```

.claude/skills/vercel-react-best-practices/rules/server-serialization.md

@@ -0,0 +1,38 @@
+---
+title: Minimize Serialization at RSC Boundaries
+impact: HIGH
+impactDescription: reduces data transfer size
+tags: server, rsc, serialization, props
+---
+
+## Minimize Serialization at RSC Boundaries
+
+The React Server/Client boundary serializes all object properties into strings and embeds them in the HTML response and subsequent RSC requests. This serialized data directly impacts page weight and load time, so **size matters a lot**. Only pass fields that the client actually uses.
+
+**Incorrect (serializes all 50 fields):**
+
+```tsx
+async function Page() {
+  const user = await fetchUser()  // 50 fields
+  return <Profile user={user} />
+}
+
+'use client'
+function Profile({ user }: { user: User }) {
+  return <div>{user.name}</div>  // uses 1 field
+}
+```
+
+**Correct (serializes only 1 field):**
+
+```tsx
+async function Page() {
+  const user = await fetchUser()
+  return <Profile name={user.name} />
+}
+
+'use client'
+function Profile({ name }: { name: string }) {
+  return <div>{name}</div>
+}
+```

.github/workflows/autofix.yml

@@ -82,6 +82,6 @@ jobs:
       # mdformat breaks YAML front matter in markdown files. Add --exclude for directories containing YAML front matter.
       - name: mdformat
         run: |
-          uvx --python 3.13 mdformat . --exclude ".claude/skills/**/SKILL.md"
+          uvx --python 3.13 mdformat . --exclude ".claude/skills/**"
 
       - uses: autofix-ci/action@635ffb0c9798bd160680f18fd73371e355b85f27

.gitignore

@@ -209,6 +209,7 @@ api/.vscode
 .history
 
 .idea/
+web/migration/
 
 # pnpm
 /.pnpm-store

api/app_factory.py

@@ -71,6 +71,8 @@ def create_app() -> DifyApp:
 
 
 def initialize_extensions(app: DifyApp):
+    # Initialize Flask context capture for workflow execution
+    from context.flask_app_context import init_flask_context
     from extensions import (
         ext_app_metrics,
         ext_blueprints,
@@ -100,6 +102,8 @@ def initialize_extensions(app: DifyApp):
         ext_warnings,
     )
 
+    init_flask_context()
+
     extensions = [
         ext_timezone,
         ext_logging,

api/commands.py

@@ -862,8 +862,27 @@ def clear_free_plan_tenant_expired_logs(days: int, batch: int, tenant_ids: list[
 
 
 @click.command("clean-workflow-runs", help="Clean expired workflow runs and related data for free tenants.")
-@click.option("--days", default=30, show_default=True, help="Delete workflow runs created before N days ago.")
+@click.option(
+    "--before-days",
+    "--days",
+    default=30,
+    show_default=True,
+    type=click.IntRange(min=0),
+    help="Delete workflow runs created before N days ago.",
+)
 @click.option("--batch-size", default=200, show_default=True, help="Batch size for selecting workflow runs.")
+@click.option(
+    "--from-days-ago",
+    default=None,
+    type=click.IntRange(min=0),
+    help="Lower bound in days ago (older). Must be paired with --to-days-ago.",
+)
+@click.option(
+    "--to-days-ago",
+    default=None,
+    type=click.IntRange(min=0),
+    help="Upper bound in days ago (newer). Must be paired with --from-days-ago.",
+)
 @click.option(
     "--start-from",
     type=click.DateTime(formats=["%Y-%m-%d", "%Y-%m-%dT%H:%M:%S"]),
@@ -882,8 +901,10 @@ def clear_free_plan_tenant_expired_logs(days: int, batch: int, tenant_ids: list[
     help="Preview cleanup results without deleting any workflow run data.",
 )
 def clean_workflow_runs(
-    days: int,
+    before_days: int,
     batch_size: int,
+    from_days_ago: int | None,
+    to_days_ago: int | None,
     start_from: datetime.datetime | None,
     end_before: datetime.datetime | None,
     dry_run: bool,
@@ -894,11 +915,24 @@ def clean_workflow_runs(
     if (start_from is None) ^ (end_before is None):
         raise click.UsageError("--start-from and --end-before must be provided together.")
 
+    if (from_days_ago is None) ^ (to_days_ago is None):
+        raise click.UsageError("--from-days-ago and --to-days-ago must be provided together.")
+
+    if from_days_ago is not None and to_days_ago is not None:
+        if start_from or end_before:
+            raise click.UsageError("Choose either day offsets or explicit dates, not both.")
+        if from_days_ago <= to_days_ago:
+            raise click.UsageError("--from-days-ago must be greater than --to-days-ago.")
+        now = datetime.datetime.now()
+        start_from = now - datetime.timedelta(days=from_days_ago)
+        end_before = now - datetime.timedelta(days=to_days_ago)
+        before_days = 0
+
     start_time = datetime.datetime.now(datetime.UTC)
     click.echo(click.style(f"Starting workflow run cleanup at {start_time.isoformat()}.", fg="white"))
 
     WorkflowRunCleanup(
-        days=days,
+        days=before_days,
         batch_size=batch_size,
         start_from=start_from,
         end_before=end_before,

api/context/__init__.py

@@ -0,0 +1,74 @@
+"""
+Core Context - Framework-agnostic context management.
+
+This module provides context management that is independent of any specific
+web framework. Framework-specific implementations register their context
+capture functions at application initialization time.
+
+This ensures the workflow layer remains completely decoupled from Flask
+or any other web framework.
+"""
+
+import contextvars
+from collections.abc import Callable
+
+from core.workflow.context.execution_context import (
+    ExecutionContext,
+    IExecutionContext,
+    NullAppContext,
+)
+
+# Global capturer function - set by framework-specific modules
+_capturer: Callable[[], IExecutionContext] | None = None
+
+
+def register_context_capturer(capturer: Callable[[], IExecutionContext]) -> None:
+    """
+    Register a context capture function.
+
+    This should be called by framework-specific modules (e.g., Flask)
+    during application initialization.
+
+    Args:
+        capturer: Function that captures current context and returns IExecutionContext
+    """
+    global _capturer
+    _capturer = capturer
+
+
+def capture_current_context() -> IExecutionContext:
+    """
+    Capture current execution context.
+
+    This function uses the registered context capturer. If no capturer
+    is registered, it returns a minimal context with only contextvars
+    (suitable for non-framework environments like tests or standalone scripts).
+
+    Returns:
+        IExecutionContext with captured context
+    """
+    if _capturer is None:
+        # No framework registered - return minimal context
+        return ExecutionContext(
+            app_context=NullAppContext(),
+            context_vars=contextvars.copy_context(),
+        )
+
+    return _capturer()
+
+
+def reset_context_provider() -> None:
+    """
+    Reset the context capturer.
+
+    This is primarily useful for testing to ensure a clean state.
+    """
+    global _capturer
+    _capturer = None
+
+
+__all__ = [
+    "capture_current_context",
+    "register_context_capturer",
+    "reset_context_provider",
+]

api/context/flask_app_context.py

@@ -0,0 +1,198 @@
+"""
+Flask App Context - Flask implementation of AppContext interface.
+"""
+
+import contextvars
+from collections.abc import Generator
+from contextlib import contextmanager
+from typing import Any, final
+
+from flask import Flask, current_app, g
+
+from context import register_context_capturer
+from core.workflow.context.execution_context import (
+    AppContext,
+    IExecutionContext,
+)
+
+
+@final
+class FlaskAppContext(AppContext):
+    """
+    Flask implementation of AppContext.
+
+    This adapts Flask's app context to the AppContext interface.
+    """
+
+    def __init__(self, flask_app: Flask) -> None:
+        """
+        Initialize Flask app context.
+
+        Args:
+            flask_app: The Flask application instance
+        """
+        self._flask_app = flask_app
+
+    def get_config(self, key: str, default: Any = None) -> Any:
+        """Get configuration value from Flask app config."""
+        return self._flask_app.config.get(key, default)
+
+    def get_extension(self, name: str) -> Any:
+        """Get Flask extension by name."""
+        return self._flask_app.extensions.get(name)
+
+    @contextmanager
+    def enter(self) -> Generator[None, None, None]:
+        """Enter Flask app context."""
+        with self._flask_app.app_context():
+            yield
+
+    @property
+    def flask_app(self) -> Flask:
+        """Get the underlying Flask app instance."""
+        return self._flask_app
+
+
+def capture_flask_context(user: Any = None) -> IExecutionContext:
+    """
+    Capture current Flask execution context.
+
+    This function captures the Flask app context and contextvars from the
+    current environment. It should be called from within a Flask request or
+    app context.
+
+    Args:
+        user: Optional user object to include in context
+
+    Returns:
+        IExecutionContext with captured Flask context
+
+    Raises:
+        RuntimeError: If called outside Flask context
+    """
+    # Get Flask app instance
+    flask_app = current_app._get_current_object()  # type: ignore
+
+    # Save current user if available
+    saved_user = user
+    if saved_user is None:
+        # Check for user in g (flask-login)
+        if hasattr(g, "_login_user"):
+            saved_user = g._login_user
+
+    # Capture contextvars
+    context_vars = contextvars.copy_context()
+
+    return FlaskExecutionContext(
+        flask_app=flask_app,
+        context_vars=context_vars,
+        user=saved_user,
+    )
+
+
+@final
+class FlaskExecutionContext:
+    """
+    Flask-specific execution context.
+
+    This is a specialized version of ExecutionContext that includes Flask app
+    context. It provides the same interface as ExecutionContext but with
+    Flask-specific implementation.
+    """
+
+    def __init__(
+        self,
+        flask_app: Flask,
+        context_vars: contextvars.Context,
+        user: Any = None,
+    ) -> None:
+        """
+        Initialize Flask execution context.
+
+        Args:
+            flask_app: Flask application instance
+            context_vars: Python contextvars
+            user: Optional user object
+        """
+        self._app_context = FlaskAppContext(flask_app)
+        self._context_vars = context_vars
+        self._user = user
+        self._flask_app = flask_app
+
+    @property
+    def app_context(self) -> FlaskAppContext:
+        """Get Flask app context."""
+        return self._app_context
+
+    @property
+    def context_vars(self) -> contextvars.Context:
+        """Get context variables."""
+        return self._context_vars
+
+    @property
+    def user(self) -> Any:
+        """Get user object."""
+        return self._user
+
+    def __enter__(self) -> "FlaskExecutionContext":
+        """Enter the Flask execution context."""
+        # Restore context variables
+        for var, val in self._context_vars.items():
+            var.set(val)
+
+        # Save current user from g if available
+        saved_user = None
+        if hasattr(g, "_login_user"):
+            saved_user = g._login_user
+
+        # Enter Flask app context
+        self._cm = self._app_context.enter()
+        self._cm.__enter__()
+
+        # Restore user in new app context
+        if saved_user is not None:
+            g._login_user = saved_user
+
+        return self
+
+    def __exit__(self, *args: Any) -> None:
+        """Exit the Flask execution context."""
+        if hasattr(self, "_cm"):
+            self._cm.__exit__(*args)
+
+    @contextmanager
+    def enter(self) -> Generator[None, None, None]:
+        """Enter Flask execution context as context manager."""
+        # Restore context variables
+        for var, val in self._context_vars.items():
+            var.set(val)
+
+        # Save current user from g if available
+        saved_user = None
+        if hasattr(g, "_login_user"):
+            saved_user = g._login_user
+
+        # Enter Flask app context
+        with self._flask_app.app_context():
+            # Restore user in new app context
+            if saved_user is not None:
+                g._login_user = saved_user
+            yield
+
+
+def init_flask_context() -> None:
+    """
+    Initialize Flask context capture by registering the capturer.
+
+    This function should be called during Flask application initialization
+    to register the Flask-specific context capturer with the core context module.
+
+    Example:
+        app = Flask(__name__)
+        init_flask_context()  # Register Flask context capturer
+
+    Note:
+        This function does not need the app instance as it uses Flask's
+        `current_app` to get the app when capturing context.
+    """
+    register_context_capturer(capture_flask_context)

api/controllers/console/app/generator.py

@@ -55,6 +55,35 @@ class InstructionTemplatePayload(BaseModel):
     type: str = Field(..., description="Instruction template type")
 
 
+class ContextGeneratePayload(BaseModel):
+    """Payload for generating extractor code node."""
+
+    workflow_id: str = Field(..., description="Workflow ID")
+    node_id: str = Field(..., description="Current tool/llm node ID")
+    parameter_name: str = Field(..., description="Parameter name to generate code for")
+    language: str = Field(default="python3", description="Code language (python3/javascript)")
+    prompt_messages: list[dict[str, Any]] = Field(
+        ..., description="Multi-turn conversation history, last message is the current instruction"
+    )
+    model_config_data: dict[str, Any] = Field(..., alias="model_config", description="Model configuration")
+
+
+class SuggestedQuestionsPayload(BaseModel):
+    """Payload for generating suggested questions."""
+
+    workflow_id: str = Field(..., description="Workflow ID")
+    node_id: str = Field(..., description="Current tool/llm node ID")
+    parameter_name: str = Field(..., description="Parameter name")
+    language: str = Field(
+        default="English", description="Language for generated questions (e.g. English, Chinese, Japanese)"
+    )
+    model_config_data: dict[str, Any] | None = Field(
+        default=None,
+        alias="model_config",
+        description="Model configuration (optional, uses system default if not provided)",
+    )
+
+
 def reg(cls: type[BaseModel]):
     console_ns.schema_model(cls.__name__, cls.model_json_schema(ref_template=DEFAULT_REF_TEMPLATE_SWAGGER_2_0))
 
@@ -64,6 +93,8 @@ reg(RuleCodeGeneratePayload)
 reg(RuleStructuredOutputPayload)
 reg(InstructionGeneratePayload)
 reg(InstructionTemplatePayload)
+reg(ContextGeneratePayload)
+reg(SuggestedQuestionsPayload)
 
 
 @console_ns.route("/rule-generate")
@@ -278,3 +309,74 @@ class InstructionGenerationTemplateApi(Resource):
                 return {"data": INSTRUCTION_GENERATE_TEMPLATE_CODE}
             case _:
                 raise ValueError(f"Invalid type: {args.type}")
+
+
+@console_ns.route("/context-generate")
+class ContextGenerateApi(Resource):
+    @console_ns.doc("generate_with_context")
+    @console_ns.doc(description="Generate with multi-turn conversation context")
+    @console_ns.expect(console_ns.models[ContextGeneratePayload.__name__])
+    @console_ns.response(200, "Content generated successfully")
+    @console_ns.response(400, "Invalid request parameters or workflow not found")
+    @console_ns.response(402, "Provider quota exceeded")
+    @setup_required
+    @login_required
+    @account_initialization_required
+    def post(self):
+        from core.llm_generator.utils import deserialize_prompt_messages
+
+        args = ContextGeneratePayload.model_validate(console_ns.payload)
+        _, current_tenant_id = current_account_with_tenant()
+
+        prompt_messages = deserialize_prompt_messages(args.prompt_messages)
+
+        try:
+            return LLMGenerator.generate_with_context(
+                tenant_id=current_tenant_id,
+                workflow_id=args.workflow_id,
+                node_id=args.node_id,
+                parameter_name=args.parameter_name,
+                language=args.language,
+                prompt_messages=prompt_messages,
+                model_config=args.model_config_data,
+            )
+        except ProviderTokenNotInitError as ex:
+            raise ProviderNotInitializeError(ex.description)
+        except QuotaExceededError:
+            raise ProviderQuotaExceededError()
+        except ModelCurrentlyNotSupportError:
+            raise ProviderModelCurrentlyNotSupportError()
+        except InvokeError as e:
+            raise CompletionRequestError(e.description)
+
+
+@console_ns.route("/context-generate/suggested-questions")
+class SuggestedQuestionsApi(Resource):
+    @console_ns.doc("generate_suggested_questions")
+    @console_ns.doc(description="Generate suggested questions for context generation")
+    @console_ns.expect(console_ns.models[SuggestedQuestionsPayload.__name__])
+    @console_ns.response(200, "Questions generated successfully")
+    @setup_required
+    @login_required
+    @account_initialization_required
+    def post(self):
+        args = SuggestedQuestionsPayload.model_validate(console_ns.payload)
+        _, current_tenant_id = current_account_with_tenant()
+
+        try:
+            return LLMGenerator.generate_suggested_questions(
+                tenant_id=current_tenant_id,
+                workflow_id=args.workflow_id,
+                node_id=args.node_id,
+                parameter_name=args.parameter_name,
+                language=args.language,
+                model_config=args.model_config_data,
+            )
+        except ProviderTokenNotInitError as ex:
+            raise ProviderNotInitializeError(ex.description)
+        except QuotaExceededError:
+            raise ProviderQuotaExceededError()
+        except ModelCurrentlyNotSupportError:
+            raise ProviderModelCurrentlyNotSupportError()
+        except InvokeError as e:
+            raise CompletionRequestError(e.description)

api/controllers/console/app/workflow.py

@@ -46,6 +46,8 @@ from models.workflow import Workflow
 from services.app_generate_service import AppGenerateService
 from services.errors.app import WorkflowHashNotEqualError
 from services.errors.llm import InvokeRateLimitError
+from services.workflow.entities import MentionGraphRequest, MentionParameterSchema
+from services.workflow.mention_graph_service import MentionGraphService
 from services.workflow_service import DraftWorkflowDeletionError, WorkflowInUseError, WorkflowService
 
 logger = logging.getLogger(__name__)
@@ -188,6 +190,15 @@ class DraftWorkflowTriggerRunAllPayload(BaseModel):
     node_ids: list[str]
 
 
+class MentionGraphPayload(BaseModel):
+    """Request payload for generating mention graph."""
+
+    parent_node_id: str = Field(description="ID of the parent node that uses the extracted value")
+    parameter_key: str = Field(description="Key of the parameter being extracted")
+    context_source: list[str] = Field(description="Variable selector for the context source")
+    parameter_schema: dict[str, Any] = Field(description="Schema of the parameter to extract")
+
+
 def reg(cls: type[BaseModel]):
     console_ns.schema_model(cls.__name__, cls.model_json_schema(ref_template=DEFAULT_REF_TEMPLATE_SWAGGER_2_0))
 
@@ -205,6 +216,7 @@ reg(WorkflowListQuery)
 reg(WorkflowUpdatePayload)
 reg(DraftWorkflowTriggerRunPayload)
 reg(DraftWorkflowTriggerRunAllPayload)
+reg(MentionGraphPayload)
 
 
 # TODO(QuantumGhost): Refactor existing node run API to handle file parameter parsing
@@ -1166,3 +1178,54 @@ class DraftWorkflowTriggerRunAllApi(Resource):
                     "status": "error",
                 }
             ), 400
+
+
+@console_ns.route("/apps/<uuid:app_id>/workflows/draft/mention-graph")
+class MentionGraphApi(Resource):
+    """
+    API for generating Mention LLM node graph structures.
+
+    This endpoint creates a complete graph structure containing an LLM node
+    configured to extract values from list[PromptMessage] variables.
+    """
+
+    @console_ns.doc("generate_mention_graph")
+    @console_ns.doc(description="Generate a Mention LLM node graph structure")
+    @console_ns.doc(params={"app_id": "Application ID"})
+    @console_ns.expect(console_ns.models[MentionGraphPayload.__name__])
+    @console_ns.response(200, "Mention graph generated successfully")
+    @console_ns.response(400, "Invalid request parameters")
+    @console_ns.response(403, "Permission denied")
+    @setup_required
+    @login_required
+    @account_initialization_required
+    @get_app_model(mode=[AppMode.ADVANCED_CHAT, AppMode.WORKFLOW])
+    @edit_permission_required
+    def post(self, app_model: App):
+        """
+        Generate a Mention LLM node graph structure.
+
+        Returns a complete graph structure containing a single LLM node
+        configured for extracting values from list[PromptMessage] context.
+        """
+
+        payload = MentionGraphPayload.model_validate(console_ns.payload or {})
+
+        parameter_schema = MentionParameterSchema(
+            name=payload.parameter_schema.get("name", payload.parameter_key),
+            type=payload.parameter_schema.get("type", "string"),
+            description=payload.parameter_schema.get("description", ""),
+        )
+
+        request = MentionGraphRequest(
+            parent_node_id=payload.parent_node_id,
+            parameter_key=payload.parameter_key,
+            context_source=payload.context_source,
+            parameter_schema=parameter_schema,
+        )
+
+        with Session(db.engine) as session:
+            service = MentionGraphService(session)
+            response = service.generate_mention_graph(tenant_id=app_model.tenant_id, request=request)
+
+        return response.model_dump()

api/controllers/console/app/workflow_draft_variable.py

@@ -17,7 +17,7 @@ from controllers.console.wraps import account_initialization_required, edit_perm
 from controllers.web.error import InvalidArgumentError, NotFoundError
 from core.file import helpers as file_helpers
 from core.variables.segment_group import SegmentGroup
-from core.variables.segments import ArrayFileSegment, FileSegment, Segment
+from core.variables.segments import ArrayFileSegment, ArrayPromptMessageSegment, FileSegment, Segment
 from core.variables.types import SegmentType
 from core.workflow.constants import CONVERSATION_VARIABLE_NODE_ID, SYSTEM_VARIABLE_NODE_ID
 from extensions.ext_database import db
@@ -58,6 +58,8 @@ def _convert_values_to_json_serializable_object(value: Segment):
         return value.value.model_dump()
     elif isinstance(value, ArrayFileSegment):
         return [i.model_dump() for i in value.value]
+    elif isinstance(value, ArrayPromptMessageSegment):
+        return value.to_object()
     elif isinstance(value, SegmentGroup):
         return [_convert_values_to_json_serializable_object(i) for i in value.value]
     else:

api/controllers/console/auth/activate.py

@@ -69,6 +69,13 @@ class ActivateCheckApi(Resource):
         if invitation:
             data = invitation.get("data", {})
             tenant = invitation.get("tenant", None)
+
+            # Check workspace permission
+            if tenant:
+                from libs.workspace_permission import check_workspace_member_invite_permission
+
+                check_workspace_member_invite_permission(tenant.id)
+
             workspace_name = tenant.name if tenant else None
             workspace_id = tenant.id if tenant else None
             invitee_email = data.get("email") if data else None

api/controllers/console/workspace/members.py

@@ -107,6 +107,12 @@ class MemberInviteEmailApi(Resource):
         inviter = current_user
         if not inviter.current_tenant:
             raise ValueError("No current tenant")
+
+        # Check workspace permission for member invitations
+        from libs.workspace_permission import check_workspace_member_invite_permission
+
+        check_workspace_member_invite_permission(inviter.current_tenant.id)
+
         invitation_results = []
         console_web_url = dify_config.CONSOLE_WEB_URL
 

api/controllers/console/workspace/workspace.py

@@ -20,6 +20,7 @@ from controllers.console.error import AccountNotLinkTenantError
 from controllers.console.wraps import (
     account_initialization_required,
     cloud_edition_billing_resource_check,
+    only_edition_enterprise,
     setup_required,
 )
 from enums.cloud_plan import CloudPlan
@@ -28,6 +29,7 @@ from libs.helper import TimestampField
 from libs.login import current_account_with_tenant, login_required
 from models.account import Tenant, TenantStatus
 from services.account_service import TenantService
+from services.enterprise.enterprise_service import EnterpriseService
 from services.feature_service import FeatureService
 from services.file_service import FileService
 from services.workspace_service import WorkspaceService
@@ -288,3 +290,31 @@ class WorkspaceInfoApi(Resource):
         db.session.commit()
 
         return {"result": "success", "tenant": marshal(WorkspaceService.get_tenant_info(tenant), tenant_fields)}
+
+
+@console_ns.route("/workspaces/current/permission")
+class WorkspacePermissionApi(Resource):
+    """Get workspace permissions for the current workspace."""
+
+    @setup_required
+    @login_required
+    @account_initialization_required
+    @only_edition_enterprise
+    def get(self):
+        """
+        Get workspace permission settings.
+        Returns permission flags that control workspace features like member invitations and owner transfer.
+        """
+        _, current_tenant_id = current_account_with_tenant()
+
+        if not current_tenant_id:
+            raise ValueError("No current tenant")
+
+        # Get workspace permissions from enterprise service
+        permission = EnterpriseService.WorkspacePermissionService.get_permission(current_tenant_id)
+
+        return {
+            "workspace_id": permission.workspace_id,
+            "allow_member_invite": permission.allow_member_invite,
+            "allow_owner_transfer": permission.allow_owner_transfer,
+        }, 200

api/controllers/console/wraps.py

@@ -286,13 +286,12 @@ def enable_change_email(view: Callable[P, R]):
 def is_allow_transfer_owner(view: Callable[P, R]):
     @wraps(view)
     def decorated(*args: P.args, **kwargs: P.kwargs):
-        _, current_tenant_id = current_account_with_tenant()
-        features = FeatureService.get_features(current_tenant_id)
-        if features.is_allow_transfer_workspace:
-            return view(*args, **kwargs)
+        from libs.workspace_permission import check_workspace_owner_transfer_permission
 
-        # otherwise, return 403
-        abort(403)
+        _, current_tenant_id = current_account_with_tenant()
+        # Check both billing/plan level and workspace policy level permissions
+        check_workspace_owner_transfer_permission(current_tenant_id)
+        return view(*args, **kwargs)
 
     return decorated
 

api/core/app/apps/advanced_chat/generate_response_converter.py

@@ -82,7 +82,7 @@ class AdvancedChatAppGenerateResponseConverter(AppGenerateResponseConverter):
                 data = cls._error_to_stream_response(sub_stream_response.err)
                 response_chunk.update(data)
             else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
             yield response_chunk
 
     @classmethod
@@ -110,7 +110,7 @@ class AdvancedChatAppGenerateResponseConverter(AppGenerateResponseConverter):
             }
 
             if isinstance(sub_stream_response, MessageEndStreamResponse):
-                sub_stream_response_dict = sub_stream_response.model_dump(mode="json")
+                sub_stream_response_dict = sub_stream_response.model_dump(mode="json", exclude_none=True)
                 metadata = sub_stream_response_dict.get("metadata", {})
                 sub_stream_response_dict["metadata"] = cls._get_simple_metadata(metadata)
                 response_chunk.update(sub_stream_response_dict)
@@ -120,6 +120,6 @@ class AdvancedChatAppGenerateResponseConverter(AppGenerateResponseConverter):
             elif isinstance(sub_stream_response, NodeStartStreamResponse | NodeFinishStreamResponse):
                 response_chunk.update(sub_stream_response.to_ignore_detail_dict())
             else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
 
             yield response_chunk

api/core/app/apps/agent_chat/generate_response_converter.py

@@ -81,7 +81,7 @@ class AgentChatAppGenerateResponseConverter(AppGenerateResponseConverter):
                 data = cls._error_to_stream_response(sub_stream_response.err)
                 response_chunk.update(data)
             else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
             yield response_chunk
 
     @classmethod
@@ -109,7 +109,7 @@ class AgentChatAppGenerateResponseConverter(AppGenerateResponseConverter):
             }
 
             if isinstance(sub_stream_response, MessageEndStreamResponse):
-                sub_stream_response_dict = sub_stream_response.model_dump(mode="json")
+                sub_stream_response_dict = sub_stream_response.model_dump(mode="json", exclude_none=True)
                 metadata = sub_stream_response_dict.get("metadata", {})
                 sub_stream_response_dict["metadata"] = cls._get_simple_metadata(metadata)
                 response_chunk.update(sub_stream_response_dict)
@@ -117,6 +117,6 @@ class AgentChatAppGenerateResponseConverter(AppGenerateResponseConverter):
                 data = cls._error_to_stream_response(sub_stream_response.err)
                 response_chunk.update(data)
             else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
 
             yield response_chunk

api/core/app/apps/chat/generate_response_converter.py

@@ -81,7 +81,7 @@ class ChatAppGenerateResponseConverter(AppGenerateResponseConverter):
                 data = cls._error_to_stream_response(sub_stream_response.err)
                 response_chunk.update(data)
             else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
             yield response_chunk
 
     @classmethod
@@ -109,7 +109,7 @@ class ChatAppGenerateResponseConverter(AppGenerateResponseConverter):
             }
 
             if isinstance(sub_stream_response, MessageEndStreamResponse):
-                sub_stream_response_dict = sub_stream_response.model_dump(mode="json")
+                sub_stream_response_dict = sub_stream_response.model_dump(mode="json", exclude_none=True)
                 metadata = sub_stream_response_dict.get("metadata", {})
                 sub_stream_response_dict["metadata"] = cls._get_simple_metadata(metadata)
                 response_chunk.update(sub_stream_response_dict)
@@ -117,6 +117,6 @@ class ChatAppGenerateResponseConverter(AppGenerateResponseConverter):
                 data = cls._error_to_stream_response(sub_stream_response.err)
                 response_chunk.update(data)
             else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
 
             yield response_chunk

api/core/app/apps/common/workflow_response_converter.py

@@ -70,6 +70,8 @@ class _NodeSnapshot:
     """Empty string means the node is not executing inside an iteration."""
     loop_id: str = ""
     """Empty string means the node is not executing inside a loop."""
+    mention_parent_id: str = ""
+    """Empty string means the node is not an extractor node."""
 
 
 class WorkflowResponseConverter:
@@ -131,6 +133,7 @@ class WorkflowResponseConverter:
             start_at=event.start_at,
             iteration_id=event.in_iteration_id or "",
             loop_id=event.in_loop_id or "",
+            mention_parent_id=event.in_mention_parent_id or "",
         )
         node_execution_id = NodeExecutionId(event.node_execution_id)
         self._node_snapshots[node_execution_id] = snapshot
@@ -287,6 +290,7 @@ class WorkflowResponseConverter:
                 created_at=int(snapshot.start_at.timestamp()),
                 iteration_id=event.in_iteration_id,
                 loop_id=event.in_loop_id,
+                mention_parent_id=event.in_mention_parent_id,
                 agent_strategy=event.agent_strategy,
             ),
         )
@@ -373,6 +377,7 @@ class WorkflowResponseConverter:
                 files=self.fetch_files_from_node_outputs(event.outputs or {}),
                 iteration_id=event.in_iteration_id,
                 loop_id=event.in_loop_id,
+                mention_parent_id=event.in_mention_parent_id,
             ),
         )
 
@@ -422,6 +427,7 @@ class WorkflowResponseConverter:
                 files=self.fetch_files_from_node_outputs(event.outputs or {}),
                 iteration_id=event.in_iteration_id,
                 loop_id=event.in_loop_id,
+                mention_parent_id=event.in_mention_parent_id,
                 retry_index=event.retry_index,
             ),
         )

api/core/app/apps/completion/generate_response_converter.py

@@ -79,7 +79,7 @@ class CompletionAppGenerateResponseConverter(AppGenerateResponseConverter):
                 data = cls._error_to_stream_response(sub_stream_response.err)
                 response_chunk.update(data)
             else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
             yield response_chunk
 
     @classmethod
@@ -106,7 +106,7 @@ class CompletionAppGenerateResponseConverter(AppGenerateResponseConverter):
             }
 
             if isinstance(sub_stream_response, MessageEndStreamResponse):
-                sub_stream_response_dict = sub_stream_response.model_dump(mode="json")
+                sub_stream_response_dict = sub_stream_response.model_dump(mode="json", exclude_none=True)
                 metadata = sub_stream_response_dict.get("metadata", {})
                 if not isinstance(metadata, dict):
                     metadata = {}
@@ -116,6 +116,6 @@ class CompletionAppGenerateResponseConverter(AppGenerateResponseConverter):
                 data = cls._error_to_stream_response(sub_stream_response.err)
                 response_chunk.update(data)
             else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
 
             yield response_chunk

api/core/app/apps/pipeline/generate_response_converter.py

@@ -60,7 +60,7 @@ class WorkflowAppGenerateResponseConverter(AppGenerateResponseConverter):
                 data = cls._error_to_stream_response(sub_stream_response.err)
                 response_chunk.update(cast(dict, data))
             else:
-                response_chunk.update(sub_stream_response.model_dump())
+                response_chunk.update(sub_stream_response.model_dump(exclude_none=True))
             yield response_chunk
 
     @classmethod
@@ -91,5 +91,5 @@ class WorkflowAppGenerateResponseConverter(AppGenerateResponseConverter):
             elif isinstance(sub_stream_response, NodeStartStreamResponse | NodeFinishStreamResponse):
                 response_chunk.update(cast(dict, sub_stream_response.to_ignore_detail_dict()))
             else:
-                response_chunk.update(sub_stream_response.model_dump())
+                response_chunk.update(sub_stream_response.model_dump(exclude_none=True))
             yield response_chunk

api/core/app/apps/workflow/app_generator.py

@@ -8,7 +8,7 @@ from typing import Any, Literal, Union, overload
 from flask import Flask, current_app
 from pydantic import ValidationError
 from sqlalchemy import select
-from sqlalchemy.orm import Session, sessionmaker
+from sqlalchemy.orm import sessionmaker
 
 import contexts
 from configs import dify_config
@@ -23,6 +23,7 @@ from core.app.apps.workflow.generate_response_converter import WorkflowAppGenera
 from core.app.apps.workflow.generate_task_pipeline import WorkflowAppGenerateTaskPipeline
 from core.app.entities.app_invoke_entities import InvokeFrom, WorkflowAppGenerateEntity
 from core.app.entities.task_entities import WorkflowAppBlockingResponse, WorkflowAppStreamResponse
+from core.db.session_factory import session_factory
 from core.helper.trace_id_helper import extract_external_trace_id_from_args
 from core.model_runtime.errors.invoke import InvokeAuthorizationError
 from core.ops.ops_trace_manager import TraceQueueManager
@@ -476,7 +477,7 @@ class WorkflowAppGenerator(BaseAppGenerator):
         :return:
         """
         with preserve_flask_contexts(flask_app, context_vars=context):
-            with Session(db.engine, expire_on_commit=False) as session:
+            with session_factory.create_session() as session:
                 workflow = session.scalar(
                     select(Workflow).where(
                         Workflow.tenant_id == application_generate_entity.app_config.tenant_id,

api/core/app/apps/workflow/generate_response_converter.py

@@ -60,7 +60,7 @@ class WorkflowAppGenerateResponseConverter(AppGenerateResponseConverter):
                 data = cls._error_to_stream_response(sub_stream_response.err)
                 response_chunk.update(data)
             else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
             yield response_chunk
 
     @classmethod
@@ -91,5 +91,5 @@ class WorkflowAppGenerateResponseConverter(AppGenerateResponseConverter):
             elif isinstance(sub_stream_response, NodeStartStreamResponse | NodeFinishStreamResponse):
                 response_chunk.update(sub_stream_response.to_ignore_detail_dict())
             else:
-                response_chunk.update(sub_stream_response.model_dump(mode="json"))
+                response_chunk.update(sub_stream_response.model_dump(mode="json", exclude_none=True))
             yield response_chunk

api/core/app/apps/workflow_app_runner.py

@@ -385,6 +385,7 @@ class WorkflowBasedAppRunner:
                     start_at=event.start_at,
                     in_iteration_id=event.in_iteration_id,
                     in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                     inputs=inputs,
                     process_data=process_data,
                     outputs=outputs,
@@ -405,6 +406,7 @@ class WorkflowBasedAppRunner:
                     start_at=event.start_at,
                     in_iteration_id=event.in_iteration_id,
                     in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                     agent_strategy=event.agent_strategy,
                     provider_type=event.provider_type,
                     provider_id=event.provider_id,
@@ -428,6 +430,7 @@ class WorkflowBasedAppRunner:
                     execution_metadata=execution_metadata,
                     in_iteration_id=event.in_iteration_id,
                     in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                 )
             )
         elif isinstance(event, NodeRunFailedEvent):
@@ -444,6 +447,7 @@ class WorkflowBasedAppRunner:
                     execution_metadata=event.node_run_result.metadata,
                     in_iteration_id=event.in_iteration_id,
                     in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                 )
             )
         elif isinstance(event, NodeRunExceptionEvent):
@@ -460,6 +464,7 @@ class WorkflowBasedAppRunner:
                     execution_metadata=event.node_run_result.metadata,
                     in_iteration_id=event.in_iteration_id,
                     in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                 )
             )
         elif isinstance(event, NodeRunStreamChunkEvent):
@@ -469,6 +474,7 @@ class WorkflowBasedAppRunner:
                     from_variable_selector=list(event.selector),
                     in_iteration_id=event.in_iteration_id,
                     in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                 )
             )
         elif isinstance(event, NodeRunRetrieverResourceEvent):
@@ -477,6 +483,7 @@ class WorkflowBasedAppRunner:
                     retriever_resources=event.retriever_resources,
                     in_iteration_id=event.in_iteration_id,
                     in_loop_id=event.in_loop_id,
+                    in_mention_parent_id=event.in_mention_parent_id,
                 )
             )
         elif isinstance(event, NodeRunAgentLogEvent):

api/core/app/entities/queue_entities.py

@@ -190,6 +190,8 @@ class QueueTextChunkEvent(AppQueueEvent):
     """iteration id if node is in iteration"""
     in_loop_id: str | None = None
     """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""
 
 
 class QueueAgentMessageEvent(AppQueueEvent):
@@ -229,6 +231,8 @@ class QueueRetrieverResourcesEvent(AppQueueEvent):
     """iteration id if node is in iteration"""
     in_loop_id: str | None = None
     """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""
 
 
 class QueueAnnotationReplyEvent(AppQueueEvent):
@@ -306,6 +310,8 @@ class QueueNodeStartedEvent(AppQueueEvent):
     node_run_index: int = 1  # FIXME(-LAN-): may not used
     in_iteration_id: str | None = None
     in_loop_id: str | None = None
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""
     start_at: datetime
     agent_strategy: AgentNodeStrategyInit | None = None
 
@@ -328,6 +334,8 @@ class QueueNodeSucceededEvent(AppQueueEvent):
     """iteration id if node is in iteration"""
     in_loop_id: str | None = None
     """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""
     start_at: datetime
 
     inputs: Mapping[str, object] = Field(default_factory=dict)
@@ -383,6 +391,8 @@ class QueueNodeExceptionEvent(AppQueueEvent):
     """iteration id if node is in iteration"""
     in_loop_id: str | None = None
     """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""
     start_at: datetime
 
     inputs: Mapping[str, object] = Field(default_factory=dict)
@@ -407,6 +417,8 @@ class QueueNodeFailedEvent(AppQueueEvent):
     """iteration id if node is in iteration"""
     in_loop_id: str | None = None
     """loop id if node is in loop"""
+    in_mention_parent_id: str | None = None
+    """parent node id if this is an extractor node event"""
     start_at: datetime
 
     inputs: Mapping[str, object] = Field(default_factory=dict)

api/core/app/entities/task_entities.py

@@ -262,6 +262,7 @@ class NodeStartStreamResponse(StreamResponse):
         extras: dict[str, object] = Field(default_factory=dict)
         iteration_id: str | None = None
         loop_id: str | None = None
+        mention_parent_id: str | None = None
         agent_strategy: AgentNodeStrategyInit | None = None
 
     event: StreamEvent = StreamEvent.NODE_STARTED
@@ -285,6 +286,7 @@ class NodeStartStreamResponse(StreamResponse):
                 "extras": {},
                 "iteration_id": self.data.iteration_id,
                 "loop_id": self.data.loop_id,
+                "mention_parent_id": self.data.mention_parent_id,
             },
         }
 
@@ -320,6 +322,7 @@ class NodeFinishStreamResponse(StreamResponse):
         files: Sequence[Mapping[str, Any]] | None = []
         iteration_id: str | None = None
         loop_id: str | None = None
+        mention_parent_id: str | None = None
 
     event: StreamEvent = StreamEvent.NODE_FINISHED
     workflow_run_id: str
@@ -349,6 +352,7 @@ class NodeFinishStreamResponse(StreamResponse):
                 "files": [],
                 "iteration_id": self.data.iteration_id,
                 "loop_id": self.data.loop_id,
+                "mention_parent_id": self.data.mention_parent_id,
             },
         }
 
@@ -384,6 +388,7 @@ class NodeRetryStreamResponse(StreamResponse):
         files: Sequence[Mapping[str, Any]] | None = []
         iteration_id: str | None = None
         loop_id: str | None = None
+        mention_parent_id: str | None = None
         retry_index: int = 0
 
     event: StreamEvent = StreamEvent.NODE_RETRY
@@ -414,6 +419,7 @@ class NodeRetryStreamResponse(StreamResponse):
                 "files": [],
                 "iteration_id": self.data.iteration_id,
                 "loop_id": self.data.loop_id,
+                "mention_parent_id": self.data.mention_parent_id,
                 "retry_index": self.data.retry_index,
             },
         }

api/core/file/file_manager.py

@@ -1,4 +1,5 @@
 import base64
+import logging
 from collections.abc import Mapping
 
 from configs import dify_config
@@ -10,7 +11,10 @@ from core.model_runtime.entities import (
     TextPromptMessageContent,
     VideoPromptMessageContent,
 )
-from core.model_runtime.entities.message_entities import PromptMessageContentUnionTypes
+from core.model_runtime.entities.message_entities import (
+    MultiModalPromptMessageContent,
+    PromptMessageContentUnionTypes,
+)
 from core.tools.signature import sign_tool_file
 from extensions.ext_storage import storage
 
@@ -18,6 +22,8 @@ from . import helpers
 from .enums import FileAttribute
 from .models import File, FileTransferMethod, FileType
 
+logger = logging.getLogger(__name__)
+
 
 def get_attr(*, file: File, attr: FileAttribute):
     match attr:
@@ -89,6 +95,8 @@ def to_prompt_message_content(
         "format": f.extension.removeprefix("."),
         "mime_type": f.mime_type,
         "filename": f.filename or "",
+        # Encoded file reference for context restoration: "transfer_method:related_id" or "remote:url"
+        "file_ref": _encode_file_ref(f),
     }
     if f.type == FileType.IMAGE:
         params["detail"] = image_detail_config or ImagePromptMessageContent.DETAIL.LOW
@@ -96,6 +104,17 @@ def to_prompt_message_content(
     return prompt_class_map[f.type].model_validate(params)
 
 
+def _encode_file_ref(f: File) -> str | None:
+    """Encode file reference as 'transfer_method:id_or_url' string."""
+    if f.transfer_method == FileTransferMethod.REMOTE_URL:
+        return f"remote:{f.remote_url}" if f.remote_url else None
+    elif f.transfer_method == FileTransferMethod.LOCAL_FILE:
+        return f"local:{f.related_id}" if f.related_id else None
+    elif f.transfer_method == FileTransferMethod.TOOL_FILE:
+        return f"tool:{f.related_id}" if f.related_id else None
+    return None
+
+
 def download(f: File, /):
     if f.transfer_method in (
         FileTransferMethod.TOOL_FILE,
@@ -164,3 +183,128 @@ def _to_url(f: File, /):
         return sign_tool_file(tool_file_id=f.related_id, extension=f.extension)
     else:
         raise ValueError(f"Unsupported transfer method: {f.transfer_method}")
+
+
+def restore_multimodal_content(
+    content: MultiModalPromptMessageContent,
+) -> MultiModalPromptMessageContent:
+    """
+    Restore base64_data or url for multimodal content from file_ref.
+
+    file_ref format: "transfer_method:id_or_url" (e.g., "local:abc123", "remote:https://...")
+
+    Args:
+        content: MultiModalPromptMessageContent with file_ref field
+
+    Returns:
+        MultiModalPromptMessageContent with restored base64_data or url
+    """
+    # Skip if no file reference or content already has data
+    if not content.file_ref:
+        return content
+    if content.base64_data or content.url:
+        return content
+
+    try:
+        file = _build_file_from_ref(
+            file_ref=content.file_ref,
+            file_format=content.format,
+            mime_type=content.mime_type,
+            filename=content.filename,
+        )
+        if not file:
+            return content
+
+        # Restore content based on config
+        if dify_config.MULTIMODAL_SEND_FORMAT == "base64":
+            restored_base64 = _get_encoded_string(file)
+            return content.model_copy(update={"base64_data": restored_base64})
+        else:
+            restored_url = _to_url(file)
+            return content.model_copy(update={"url": restored_url})
+
+    except Exception as e:
+        logger.warning("Failed to restore multimodal content: %s", e)
+        return content
+
+
+def _build_file_from_ref(
+    file_ref: str,
+    file_format: str | None,
+    mime_type: str | None,
+    filename: str | None,
+) -> File | None:
+    """
+    Build a File object from encoded file_ref string.
+
+    Args:
+        file_ref: Encoded reference "transfer_method:id_or_url"
+        file_format: The file format/extension (without dot)
+        mime_type: The mime type
+        filename: The filename
+
+    Returns:
+        File object with storage_key loaded, or None if not found
+    """
+    from sqlalchemy import select
+    from sqlalchemy.orm import Session
+
+    from extensions.ext_database import db
+    from models.model import UploadFile
+    from models.tools import ToolFile
+
+    # Parse file_ref: "method:value"
+    if ":" not in file_ref:
+        logger.warning("Invalid file_ref format: %s", file_ref)
+        return None
+
+    method, value = file_ref.split(":", 1)
+    extension = f".{file_format}" if file_format else None
+
+    if method == "remote":
+        return File(
+            tenant_id="",
+            type=FileType.IMAGE,
+            transfer_method=FileTransferMethod.REMOTE_URL,
+            remote_url=value,
+            extension=extension,
+            mime_type=mime_type,
+            filename=filename,
+            storage_key="",
+        )
+
+    # Query database for storage_key
+    with Session(db.engine) as session:
+        if method == "local":
+            stmt = select(UploadFile).where(UploadFile.id == value)
+            upload_file = session.scalar(stmt)
+            if upload_file:
+                return File(
+                    tenant_id=upload_file.tenant_id,
+                    type=FileType(upload_file.extension)
+                    if hasattr(FileType, upload_file.extension.upper())
+                    else FileType.IMAGE,
+                    transfer_method=FileTransferMethod.LOCAL_FILE,
+                    related_id=value,
+                    extension=extension or ("." + upload_file.extension if upload_file.extension else None),
+                    mime_type=mime_type or upload_file.mime_type,
+                    filename=filename or upload_file.name,
+                    storage_key=upload_file.key,
+                )
+        elif method == "tool":
+            stmt = select(ToolFile).where(ToolFile.id == value)
+            tool_file = session.scalar(stmt)
+            if tool_file:
+                return File(
+                    tenant_id=tool_file.tenant_id,
+                    type=FileType.IMAGE,
+                    transfer_method=FileTransferMethod.TOOL_FILE,
+                    related_id=value,
+                    extension=extension,
+                    mime_type=mime_type or tool_file.mimetype,
+                    filename=filename or tool_file.name,
+                    storage_key=tool_file.file_key,
+                )
+
+    logger.warning("File not found for file_ref: %s", file_ref)
+    return None

api/core/llm_generator/llm_generator.py

@@ -1,8 +1,8 @@
 import json
 import logging
 import re
-from collections.abc import Sequence
-from typing import Protocol, cast
+from collections.abc import Mapping, Sequence
+from typing import Any, Protocol, cast
 
 import json_repair
 
@@ -393,6 +393,488 @@ class LLMGenerator:
             logger.exception("Failed to invoke LLM model, model: %s", model_config.get("name"))
             return {"output": "", "error": f"An unexpected error occurred: {str(e)}"}
 
+    @classmethod
+    def generate_with_context(
+        cls,
+        tenant_id: str,
+        workflow_id: str,
+        node_id: str,
+        parameter_name: str,
+        language: str,
+        prompt_messages: list[PromptMessage],
+        model_config: dict,
+    ) -> dict:
+        """
+        Generate extractor code node based on conversation context.
+
+        Args:
+            tenant_id: Tenant/workspace ID
+            workflow_id: Workflow ID
+            node_id: Current tool/llm node ID
+            parameter_name: Parameter name to generate code for
+            language: Code language (python3/javascript)
+            prompt_messages: Multi-turn conversation history (last message is instruction)
+            model_config: Model configuration (provider, name, completion_params)
+
+        Returns:
+            dict with CodeNodeData format:
+            - variables: Input variable selectors
+            - code_language: Code language
+            - code: Generated code
+            - outputs: Output definitions
+            - message: Explanation
+            - error: Error message if any
+        """
+        from sqlalchemy import select
+        from sqlalchemy.orm import Session
+
+        from services.workflow_service import WorkflowService
+
+        # Get workflow
+        with Session(db.engine) as session:
+            stmt = select(App).where(App.id == workflow_id)
+            app = session.scalar(stmt)
+            if not app:
+                return cls._error_response(f"App {workflow_id} not found")
+
+            workflow = WorkflowService().get_draft_workflow(app_model=app)
+            if not workflow:
+                return cls._error_response(f"Workflow for app {workflow_id} not found")
+
+        # Get upstream nodes via edge backtracking
+        upstream_nodes = cls._get_upstream_nodes(workflow.graph_dict, node_id)
+
+        # Get current node info
+        current_node = cls._get_node_by_id(workflow.graph_dict, node_id)
+        if not current_node:
+            return cls._error_response(f"Node {node_id} not found")
+
+        # Get parameter info
+        parameter_info = cls._get_parameter_info(
+            tenant_id=tenant_id,
+            node_data=current_node.get("data", {}),
+            parameter_name=parameter_name,
+        )
+
+        # Build system prompt
+        system_prompt = cls._build_extractor_system_prompt(
+            upstream_nodes=upstream_nodes,
+            current_node=current_node,
+            parameter_info=parameter_info,
+            language=language,
+        )
+
+        # Construct complete prompt_messages with system prompt
+        complete_messages: list[PromptMessage] = [
+            SystemPromptMessage(content=system_prompt),
+            *prompt_messages,
+        ]
+
+        from core.llm_generator.output_parser.structured_output import invoke_llm_with_structured_output
+
+        # Get model instance and schema
+        provider = model_config.get("provider", "")
+        model_name = model_config.get("name", "")
+        model_instance = ModelManager().get_model_instance(
+            tenant_id=tenant_id,
+            model_type=ModelType.LLM,
+            provider=provider,
+            model=model_name,
+        )
+
+        model_schema = model_instance.model_type_instance.get_model_schema(model_name, model_instance.credentials)
+        if not model_schema:
+            return cls._error_response(f"Model schema not found for {model_name}")
+
+        model_parameters = model_config.get("completion_params", {})
+        json_schema = cls._get_code_node_json_schema()
+
+        try:
+            response = invoke_llm_with_structured_output(
+                provider=provider,
+                model_schema=model_schema,
+                model_instance=model_instance,
+                prompt_messages=complete_messages,
+                json_schema=json_schema,
+                model_parameters=model_parameters,
+                stream=False,
+                tenant_id=tenant_id,
+            )
+
+            return cls._parse_code_node_output(
+                response.structured_output, language, parameter_info.get("type", "string")
+            )
+
+        except InvokeError as e:
+            return cls._error_response(str(e))
+        except Exception as e:
+            logger.exception("Failed to generate with context, model: %s", model_config.get("name"))
+            return cls._error_response(f"An unexpected error occurred: {str(e)}")
+
+    @classmethod
+    def _error_response(cls, error: str) -> dict:
+        """Return error response in CodeNodeData format."""
+        return {
+            "variables": [],
+            "code_language": "python3",
+            "code": "",
+            "outputs": {},
+            "message": "",
+            "error": error,
+        }
+
+    @classmethod
+    def generate_suggested_questions(
+        cls,
+        tenant_id: str,
+        workflow_id: str,
+        node_id: str,
+        parameter_name: str,
+        language: str,
+        model_config: dict | None = None,
+    ) -> dict:
+        """
+        Generate suggested questions for context generation.
+
+        Returns dict with questions array and error field.
+        """
+        from sqlalchemy import select
+        from sqlalchemy.orm import Session
+
+        from core.llm_generator.output_parser.structured_output import invoke_llm_with_structured_output
+        from services.workflow_service import WorkflowService
+
+        # Get workflow context (reuse existing logic)
+        with Session(db.engine) as session:
+            stmt = select(App).where(App.id == workflow_id)
+            app = session.scalar(stmt)
+            if not app:
+                return {"questions": [], "error": f"App {workflow_id} not found"}
+
+            workflow = WorkflowService().get_draft_workflow(app_model=app)
+            if not workflow:
+                return {"questions": [], "error": f"Workflow for app {workflow_id} not found"}
+
+        upstream_nodes = cls._get_upstream_nodes(workflow.graph_dict, node_id)
+        current_node = cls._get_node_by_id(workflow.graph_dict, node_id)
+        if not current_node:
+            return {"questions": [], "error": f"Node {node_id} not found"}
+
+        parameter_info = cls._get_parameter_info(
+            tenant_id=tenant_id,
+            node_data=current_node.get("data", {}),
+            parameter_name=parameter_name,
+        )
+
+        # Build prompt
+        system_prompt = cls._build_suggested_questions_prompt(
+            upstream_nodes=upstream_nodes,
+            current_node=current_node,
+            parameter_info=parameter_info,
+            language=language,
+        )
+
+        prompt_messages: list[PromptMessage] = [
+            SystemPromptMessage(content=system_prompt),
+        ]
+
+        # Get model instance - use default if model_config not provided
+        model_manager = ModelManager()
+        if model_config:
+            provider = model_config.get("provider", "")
+            model_name = model_config.get("name", "")
+            model_instance = model_manager.get_model_instance(
+                tenant_id=tenant_id,
+                model_type=ModelType.LLM,
+                provider=provider,
+                model=model_name,
+            )
+        else:
+            model_instance = model_manager.get_default_model_instance(
+                tenant_id=tenant_id,
+                model_type=ModelType.LLM,
+            )
+            model_name = model_instance.model
+
+        model_schema = model_instance.model_type_instance.get_model_schema(model_name, model_instance.credentials)
+        if not model_schema:
+            return {"questions": [], "error": f"Model schema not found for {model_name}"}
+
+        completion_params = model_config.get("completion_params", {}) if model_config else {}
+        model_parameters = {**completion_params, "max_tokens": 256}
+        json_schema = cls._get_suggested_questions_json_schema()
+
+        try:
+            response = invoke_llm_with_structured_output(
+                provider=model_instance.provider,
+                model_schema=model_schema,
+                model_instance=model_instance,
+                prompt_messages=prompt_messages,
+                json_schema=json_schema,
+                model_parameters=model_parameters,
+                stream=False,
+                tenant_id=tenant_id,
+            )
+
+            questions = response.structured_output.get("questions", []) if response.structured_output else []
+            return {"questions": questions, "error": ""}
+
+        except InvokeError as e:
+            return {"questions": [], "error": str(e)}
+        except Exception as e:
+            logger.exception("Failed to generate suggested questions, model: %s", model_name)
+            return {"questions": [], "error": f"An unexpected error occurred: {str(e)}"}
+
+    @classmethod
+    def _build_suggested_questions_prompt(
+        cls,
+        upstream_nodes: list[dict],
+        current_node: dict,
+        parameter_info: dict,
+        language: str = "English",
+    ) -> str:
+        """Build minimal prompt for suggested questions generation."""
+        # Simplify upstream nodes to reduce tokens
+        sources = [f"{n['title']}({','.join(n.get('outputs', {}).keys())})" for n in upstream_nodes[:5]]
+        param_type = parameter_info.get("type", "string")
+        param_desc = parameter_info.get("description", "")[:100]
+
+        return f"""Suggest 3 code generation questions for extracting data.
+Sources: {", ".join(sources)}
+Target: {parameter_info.get("name")}({param_type}) - {param_desc}
+Output 3 short, practical questions in {language}."""
+
+    @classmethod
+    def _get_suggested_questions_json_schema(cls) -> dict:
+        """Return JSON Schema for suggested questions."""
+        return {
+            "type": "object",
+            "properties": {
+                "questions": {
+                    "type": "array",
+                    "items": {"type": "string"},
+                    "minItems": 3,
+                    "maxItems": 3,
+                    "description": "3 suggested questions",
+                },
+            },
+            "required": ["questions"],
+        }
+
+    @classmethod
+    def _get_code_node_json_schema(cls) -> dict:
+        """Return JSON Schema for structured output."""
+        return {
+            "type": "object",
+            "properties": {
+                "variables": {
+                    "type": "array",
+                    "items": {
+                        "type": "object",
+                        "properties": {
+                            "variable": {"type": "string", "description": "Variable name in code"},
+                            "value_selector": {
+                                "type": "array",
+                                "items": {"type": "string"},
+                                "description": "Path like [node_id, output_name]",
+                            },
+                        },
+                        "required": ["variable", "value_selector"],
+                    },
+                },
+                "code": {"type": "string", "description": "Generated code with main function"},
+                "outputs": {
+                    "type": "object",
+                    "additionalProperties": {
+                        "type": "object",
+                        "properties": {"type": {"type": "string"}},
+                    },
+                    "description": "Output definitions, key is output name",
+                },
+                "explanation": {"type": "string", "description": "Brief explanation of the code"},
+            },
+            "required": ["variables", "code", "outputs", "explanation"],
+        }
+
+    @classmethod
+    def _get_upstream_nodes(cls, graph_dict: Mapping[str, Any], node_id: str) -> list[dict]:
+        """
+        Get all upstream nodes via edge backtracking.
+
+        Traverses the graph backwards from node_id to collect all reachable nodes.
+        """
+        from collections import defaultdict
+
+        nodes = {n["id"]: n for n in graph_dict.get("nodes", [])}
+        edges = graph_dict.get("edges", [])
+
+        # Build reverse adjacency list
+        reverse_adj: dict[str, list[str]] = defaultdict(list)
+        for edge in edges:
+            reverse_adj[edge["target"]].append(edge["source"])
+
+        # BFS to find all upstream nodes
+        visited: set[str] = set()
+        queue = [node_id]
+        upstream: list[dict] = []
+
+        while queue:
+            current = queue.pop(0)
+            for source in reverse_adj.get(current, []):
+                if source not in visited:
+                    visited.add(source)
+                    queue.append(source)
+                    if source in nodes:
+                        upstream.append(cls._extract_node_info(nodes[source]))
+
+        return upstream
+
+    @classmethod
+    def _get_node_by_id(cls, graph_dict: Mapping[str, Any], node_id: str) -> dict | None:
+        """Get node by ID from graph."""
+        for node in graph_dict.get("nodes", []):
+            if node["id"] == node_id:
+                return node
+        return None
+
+    @classmethod
+    def _extract_node_info(cls, node: dict) -> dict:
+        """Extract minimal node info with outputs based on node type."""
+        node_type = node["data"]["type"]
+        node_data = node.get("data", {})
+
+        # Build outputs based on node type (only type, no description to reduce tokens)
+        outputs: dict[str, str] = {}
+        match node_type:
+            case "start":
+                for var in node_data.get("variables", []):
+                    name = var.get("variable", var.get("name", ""))
+                    outputs[name] = var.get("type", "string")
+            case "llm":
+                outputs["text"] = "string"
+            case "code":
+                for name, output in node_data.get("outputs", {}).items():
+                    outputs[name] = output.get("type", "string")
+            case "http-request":
+                outputs = {"body": "string", "status_code": "number", "headers": "object"}
+            case "knowledge-retrieval":
+                outputs["result"] = "array[object]"
+            case "tool":
+                outputs = {"text": "string", "json": "object"}
+            case _:
+                outputs["output"] = "string"
+
+        info: dict = {
+            "id": node["id"],
+            "title": node_data.get("title", node["id"]),
+            "outputs": outputs,
+        }
+        # Only include description if not empty
+        desc = node_data.get("desc", "")
+        if desc:
+            info["desc"] = desc
+
+        return info
+
+    @classmethod
+    def _get_parameter_info(cls, tenant_id: str, node_data: dict, parameter_name: str) -> dict:
+        """Get parameter info from tool schema using ToolManager."""
+        default_info = {"name": parameter_name, "type": "string", "description": ""}
+
+        if node_data.get("type") != "tool":
+            return default_info
+
+        try:
+            from core.app.entities.app_invoke_entities import InvokeFrom
+            from core.tools.entities.tool_entities import ToolProviderType
+            from core.tools.tool_manager import ToolManager
+
+            provider_type_str = node_data.get("provider_type", "")
+            provider_type = ToolProviderType(provider_type_str) if provider_type_str else ToolProviderType.BUILT_IN
+
+            tool_runtime = ToolManager.get_tool_runtime(
+                provider_type=provider_type,
+                provider_id=node_data.get("provider_id", ""),
+                tool_name=node_data.get("tool_name", ""),
+                tenant_id=tenant_id,
+                invoke_from=InvokeFrom.DEBUGGER,
+            )
+
+            parameters = tool_runtime.get_merged_runtime_parameters()
+            for param in parameters:
+                if param.name == parameter_name:
+                    return {
+                        "name": param.name,
+                        "type": param.type.value if hasattr(param.type, "value") else str(param.type),
+                        "description": param.llm_description
+                        or (param.human_description.en_US if param.human_description else ""),
+                        "required": param.required,
+                    }
+        except Exception as e:
+            logger.debug("Failed to get parameter info from ToolManager: %s", e)
+
+        return default_info
+
+    @classmethod
+    def _build_extractor_system_prompt(
+        cls,
+        upstream_nodes: list[dict],
+        current_node: dict,
+        parameter_info: dict,
+        language: str,
+    ) -> str:
+        """Build system prompt for extractor code generation."""
+        upstream_json = json.dumps(upstream_nodes, indent=2, ensure_ascii=False)
+        param_type = parameter_info.get("type", "string")
+        return f"""You are a code generator for workflow automation.
+
+Generate {language} code to extract/transform upstream node outputs for the target parameter.
+
+## Upstream Nodes
+{upstream_json}
+
+## Target
+Node: {current_node["data"].get("title", current_node["id"])}
+Parameter: {parameter_info.get("name")} ({param_type}) - {parameter_info.get("description", "")}
+
+## Requirements
+- Write a main function that returns type: {param_type}
+- Use value_selector format: ["node_id", "output_name"]
+"""
+
+    @classmethod
+    def _parse_code_node_output(cls, content: Mapping[str, Any] | None, language: str, parameter_type: str) -> dict:
+        """
+        Parse structured output to CodeNodeData format.
+
+        Args:
+            content: Structured output dict from invoke_llm_with_structured_output
+            language: Code language
+            parameter_type: Expected parameter type
+
+        Returns dict with variables, code_language, code, outputs, message, error.
+        """
+        if content is None:
+            return cls._error_response("Empty or invalid response from LLM")
+
+        # Validate and normalize variables
+        variables = [
+            {"variable": v.get("variable", ""), "value_selector": v.get("value_selector", [])}
+            for v in content.get("variables", [])
+            if isinstance(v, dict)
+        ]
+
+        outputs = content.get("outputs", {"result": {"type": parameter_type}})
+
+        return {
+            "variables": variables,
+            "code_language": language,
+            "code": content.get("code", ""),
+            "outputs": outputs,
+            "message": content.get("explanation", ""),
+            "error": "",
+        }
+
     @staticmethod
     def instruction_modify_legacy(
         tenant_id: str, flow_id: str, current: str, instruction: str, model_config: dict, ideal_output: str | None

api/core/llm_generator/output_parser/file_ref.py

@@ -0,0 +1,188 @@
+"""
+File reference detection and conversion for structured output.
+
+This module provides utilities to:
+1. Detect file reference fields in JSON Schema (format: "dify-file-ref")
+2. Convert file ID strings to File objects after LLM returns
+"""
+
+import uuid
+from collections.abc import Mapping
+from typing import Any
+
+from core.file import File
+from core.variables.segments import ArrayFileSegment, FileSegment
+from factories.file_factory import build_from_mapping
+
+FILE_REF_FORMAT = "dify-file-ref"
+
+
+def is_file_ref_property(schema: dict) -> bool:
+    """Check if a schema property is a file reference."""
+    return schema.get("type") == "string" and schema.get("format") == FILE_REF_FORMAT
+
+
+def detect_file_ref_fields(schema: Mapping[str, Any], path: str = "") -> list[str]:
+    """
+    Recursively detect file reference fields in schema.
+
+    Args:
+        schema: JSON Schema to analyze
+        path: Current path in the schema (used for recursion)
+
+    Returns:
+        List of JSON paths containing file refs, e.g., ["image_id", "files[*]"]
+    """
+    file_ref_paths: list[str] = []
+    schema_type = schema.get("type")
+
+    if schema_type == "object":
+        for prop_name, prop_schema in schema.get("properties", {}).items():
+            current_path = f"{path}.{prop_name}" if path else prop_name
+
+            if is_file_ref_property(prop_schema):
+                file_ref_paths.append(current_path)
+            elif isinstance(prop_schema, dict):
+                file_ref_paths.extend(detect_file_ref_fields(prop_schema, current_path))
+
+    elif schema_type == "array":
+        items_schema = schema.get("items", {})
+        array_path = f"{path}[*]" if path else "[*]"
+
+        if is_file_ref_property(items_schema):
+            file_ref_paths.append(array_path)
+        elif isinstance(items_schema, dict):
+            file_ref_paths.extend(detect_file_ref_fields(items_schema, array_path))
+
+    return file_ref_paths
+
+
+def convert_file_refs_in_output(
+    output: Mapping[str, Any],
+    json_schema: Mapping[str, Any],
+    tenant_id: str,
+) -> dict[str, Any]:
+    """
+    Convert file ID strings to File objects based on schema.
+
+    Args:
+        output: The structured_output from LLM result
+        json_schema: The original JSON schema (to detect file ref fields)
+        tenant_id: Tenant ID for file lookup
+
+    Returns:
+        Output with file references converted to File objects
+    """
+    file_ref_paths = detect_file_ref_fields(json_schema)
+    if not file_ref_paths:
+        return dict(output)
+
+    result = _deep_copy_dict(output)
+
+    for path in file_ref_paths:
+        _convert_path_in_place(result, path.split("."), tenant_id)
+
+    return result
+
+
+def _deep_copy_dict(obj: Mapping[str, Any]) -> dict[str, Any]:
+    """Deep copy a mapping to a mutable dict."""
+    result: dict[str, Any] = {}
+    for key, value in obj.items():
+        if isinstance(value, Mapping):
+            result[key] = _deep_copy_dict(value)
+        elif isinstance(value, list):
+            result[key] = [_deep_copy_dict(item) if isinstance(item, Mapping) else item for item in value]
+        else:
+            result[key] = value
+    return result
+
+
+def _convert_path_in_place(obj: dict, path_parts: list[str], tenant_id: str) -> None:
+    """Convert file refs at the given path in place, wrapping in Segment types."""
+    if not path_parts:
+        return
+
+    current = path_parts[0]
+    remaining = path_parts[1:]
+
+    # Handle array notation like "files[*]"
+    if current.endswith("[*]"):
+        key = current[:-3] if current != "[*]" else None
+        target = obj.get(key) if key else obj
+
+        if isinstance(target, list):
+            if remaining:
+                # Nested array with remaining path - recurse into each item
+                for item in target:
+                    if isinstance(item, dict):
+                        _convert_path_in_place(item, remaining, tenant_id)
+            else:
+                # Array of file IDs - convert all and wrap in ArrayFileSegment
+                files: list[File] = []
+                for item in target:
+                    file = _convert_file_id(item, tenant_id)
+                    if file is not None:
+                        files.append(file)
+                # Replace the array with ArrayFileSegment
+                if key:
+                    obj[key] = ArrayFileSegment(value=files)
+        return
+
+    if not remaining:
+        # Leaf node - convert the value and wrap in FileSegment
+        if current in obj:
+            file = _convert_file_id(obj[current], tenant_id)
+            if file is not None:
+                obj[current] = FileSegment(value=file)
+            else:
+                obj[current] = None
+    else:
+        # Recurse into nested object
+        if current in obj and isinstance(obj[current], dict):
+            _convert_path_in_place(obj[current], remaining, tenant_id)
+
+
+def _convert_file_id(file_id: Any, tenant_id: str) -> File | None:
+    """
+    Convert a file ID string to a File object.
+
+    Tries multiple file sources in order:
+    1. ToolFile (files generated by tools/workflows)
+    2. UploadFile (files uploaded by users)
+    """
+    if not isinstance(file_id, str):
+        return None
+
+    # Validate UUID format
+    try:
+        uuid.UUID(file_id)
+    except ValueError:
+        return None
+
+    # Try ToolFile first (files generated by tools/workflows)
+    try:
+        return build_from_mapping(
+            mapping={
+                "transfer_method": "tool_file",
+                "tool_file_id": file_id,
+            },
+            tenant_id=tenant_id,
+        )
+    except ValueError:
+        pass
+
+    # Try UploadFile (files uploaded by users)
+    try:
+        return build_from_mapping(
+            mapping={
+                "transfer_method": "local_file",
+                "upload_file_id": file_id,
+            },
+            tenant_id=tenant_id,
+        )
+    except ValueError:
+        pass
+
+    # File not found in any source
+    return None

api/core/llm_generator/output_parser/structured_output.py

@@ -8,6 +8,7 @@ import json_repair
 from pydantic import TypeAdapter, ValidationError
 
 from core.llm_generator.output_parser.errors import OutputParserError
+from core.llm_generator.output_parser.file_ref import convert_file_refs_in_output
 from core.llm_generator.prompts import STRUCTURED_OUTPUT_PROMPT
 from core.model_manager import ModelInstance
 from core.model_runtime.callbacks.base_callback import Callback
@@ -57,6 +58,7 @@ def invoke_llm_with_structured_output(
     stream: Literal[True],
     user: str | None = None,
     callbacks: list[Callback] | None = None,
+    tenant_id: str | None = None,
 ) -> Generator[LLMResultChunkWithStructuredOutput, None, None]: ...
 @overload
 def invoke_llm_with_structured_output(
@@ -72,6 +74,7 @@ def invoke_llm_with_structured_output(
     stream: Literal[False],
     user: str | None = None,
     callbacks: list[Callback] | None = None,
+    tenant_id: str | None = None,
 ) -> LLMResultWithStructuredOutput: ...
 @overload
 def invoke_llm_with_structured_output(
@@ -87,6 +90,7 @@ def invoke_llm_with_structured_output(
     stream: bool = True,
     user: str | None = None,
     callbacks: list[Callback] | None = None,
+    tenant_id: str | None = None,
 ) -> LLMResultWithStructuredOutput | Generator[LLMResultChunkWithStructuredOutput, None, None]: ...
 def invoke_llm_with_structured_output(
     *,
@@ -101,20 +105,28 @@ def invoke_llm_with_structured_output(
     stream: bool = True,
     user: str | None = None,
     callbacks: list[Callback] | None = None,
+    tenant_id: str | None = None,
 ) -> LLMResultWithStructuredOutput | Generator[LLMResultChunkWithStructuredOutput, None, None]:
     """
-    Invoke large language model with structured output
-    1. This method invokes model_instance.invoke_llm with json_schema
-    2. Try to parse the result as structured output
+    Invoke large language model with structured output.
 
+    This method invokes model_instance.invoke_llm with json_schema and parses
+    the result as structured output.
+
+    :param provider: model provider name
+    :param model_schema: model schema entity
+    :param model_instance: model instance to invoke
     :param prompt_messages: prompt messages
-    :param json_schema: json schema
+    :param json_schema: json schema for structured output
     :param model_parameters: model parameters
     :param tools: tools for tool calling
     :param stop: stop words
     :param stream: is stream response
     :param user: unique user id
     :param callbacks: callbacks
+    :param tenant_id: tenant ID for file reference conversion. When provided and
+                      json_schema contains file reference fields (format: "dify-file-ref"),
+                      file IDs in the output will be automatically converted to File objects.
     :return: full response or stream response chunk generator result
     """
 
@@ -153,8 +165,18 @@ def invoke_llm_with_structured_output(
                 f"Failed to parse structured output, LLM result is not a string: {llm_result.message.content}"
             )
 
+        structured_output = _parse_structured_output(llm_result.message.content)
+
+        # Convert file references if tenant_id is provided
+        if tenant_id is not None:
+            structured_output = convert_file_refs_in_output(
+                output=structured_output,
+                json_schema=json_schema,
+                tenant_id=tenant_id,
+            )
+
         return LLMResultWithStructuredOutput(
-            structured_output=_parse_structured_output(llm_result.message.content),
+            structured_output=structured_output,
             model=llm_result.model,
             message=llm_result.message,
             usage=llm_result.usage,
@@ -186,8 +208,18 @@ def invoke_llm_with_structured_output(
                     delta=event.delta,
                 )
 
+            structured_output = _parse_structured_output(result_text)
+
+            # Convert file references if tenant_id is provided
+            if tenant_id is not None:
+                structured_output = convert_file_refs_in_output(
+                    output=structured_output,
+                    json_schema=json_schema,
+                    tenant_id=tenant_id,
+                )
+
             yield LLMResultChunkWithStructuredOutput(
-                structured_output=_parse_structured_output(result_text),
+                structured_output=structured_output,
                 model=model_schema.model,
                 prompt_messages=prompt_messages,
                 system_fingerprint=system_fingerprint,

api/core/llm_generator/utils.py

@@ -0,0 +1,45 @@
+"""Utility functions for LLM generator."""
+
+from core.model_runtime.entities.message_entities import (
+    AssistantPromptMessage,
+    PromptMessage,
+    PromptMessageRole,
+    SystemPromptMessage,
+    ToolPromptMessage,
+    UserPromptMessage,
+)
+
+
+def deserialize_prompt_messages(messages: list[dict]) -> list[PromptMessage]:
+    """
+    Deserialize list of dicts to list[PromptMessage].
+
+    Expected format:
+        [
+            {"role": "user", "content": "..."},
+            {"role": "assistant", "content": "..."},
+        ]
+    """
+    result: list[PromptMessage] = []
+    for msg in messages:
+        role = PromptMessageRole.value_of(msg["role"])
+        content = msg.get("content", "")
+
+        match role:
+            case PromptMessageRole.USER:
+                result.append(UserPromptMessage(content=content))
+            case PromptMessageRole.ASSISTANT:
+                result.append(AssistantPromptMessage(content=content))
+            case PromptMessageRole.SYSTEM:
+                result.append(SystemPromptMessage(content=content))
+            case PromptMessageRole.TOOL:
+                result.append(ToolPromptMessage(content=content, tool_call_id=msg.get("tool_call_id", "")))
+
+    return result
+
+
+def serialize_prompt_messages(messages: list[PromptMessage]) -> list[dict]:
+    """
+    Serialize list[PromptMessage] to list of dicts.
+    """
+    return [{"role": msg.role.value, "content": msg.content} for msg in messages]

api/core/memory/README.md

@@ -0,0 +1,267 @@
+# Memory Module
+
+This module provides memory management for LLM conversations, enabling context retention across dialogue turns.
+
+## Overview
+
+The memory module contains two types of memory implementations:
+
+1. **TokenBufferMemory** - Conversation-level memory (existing)
+2. **NodeTokenBufferMemory** - Node-level memory (**Chatflow only**)
+
+> **Note**: `NodeTokenBufferMemory` is only available in **Chatflow** (advanced-chat mode).
+> This is because it requires both `conversation_id` and `node_id`, which are only present in Chatflow.
+> Standard Workflow mode does not have `conversation_id` and therefore cannot use node-level memory.
+
+```
+┌─────────────────────────────────────────────────────────────────────────────┐
+│                           Memory Architecture                               │
+├─────────────────────────────────────────────────────────────────────────────┤
+│                                                                             │
+│  ┌─────────────────────────────────────────────────────────────────────-┐   │
+│  │                      TokenBufferMemory                               │   │
+│  │  Scope: Conversation                                                 │   │
+│  │  Storage: Database (Message table)                                   │   │
+│  │  Key: conversation_id                                                │   │
+│  └─────────────────────────────────────────────────────────────────────-┘   │
+│                                                                             │
+│  ┌─────────────────────────────────────────────────────────────────────-┐   │
+│  │                    NodeTokenBufferMemory                             │   │
+│  │  Scope: Node within Conversation                                     │   │
+│  │  Storage: WorkflowNodeExecutionModel.outputs["context"]              │   │
+│  │  Key: (conversation_id, node_id, workflow_run_id)                    │   │
+│  └─────────────────────────────────────────────────────────────────────-┘   │
+│                                                                             │
+└─────────────────────────────────────────────────────────────────────────────┘
+```
+
+---
+
+## TokenBufferMemory (Existing)
+
+### Purpose
+
+`TokenBufferMemory` retrieves conversation history from the `Message` table and converts it to `PromptMessage` objects for LLM context.
+
+### Key Features
+
+- **Conversation-scoped**: All messages within a conversation are candidates
+- **Thread-aware**: Uses `parent_message_id` to extract only the current thread (supports regeneration scenarios)
+- **Token-limited**: Truncates history to fit within `max_token_limit`
+- **File support**: Handles `MessageFile` attachments (images, documents, etc.)
+
+### Data Flow
+
+```
+Message Table                    TokenBufferMemory              LLM
+     │                                  │                        │
+     │  SELECT * FROM messages          │                        │
+     │  WHERE conversation_id = ?       │                        │
+     │  ORDER BY created_at DESC        │                        │
+     ├─────────────────────────────────▶│                        │
+     │                                  │                        │
+     │                    extract_thread_messages()              │
+     │                                  │                        │
+     │                    build_prompt_message_with_files()      │
+     │                                  │                        │
+     │                    truncate by max_token_limit            │
+     │                                  │                        │
+     │                                  │  Sequence[PromptMessage]
+     │                                  ├───────────────────────▶│
+     │                                  │                        │
+```
+
+### Thread Extraction
+
+When a user regenerates a response, a new thread is created:
+
+```
+Message A (user)
+    └── Message A' (assistant)
+            └── Message B (user)
+                    └── Message B' (assistant)
+            └── Message A'' (assistant, regenerated)  ← New thread
+                    └── Message C (user)
+                            └── Message C' (assistant)
+```
+
+`extract_thread_messages()` traces back from the latest message using `parent_message_id` to get only the current thread: `[A, A'', C, C']`
+
+### Usage
+
+```python
+from core.memory.token_buffer_memory import TokenBufferMemory
+
+memory = TokenBufferMemory(conversation=conversation, model_instance=model_instance)
+history = memory.get_history_prompt_messages(max_token_limit=2000, message_limit=100)
+```
+
+---
+
+## NodeTokenBufferMemory
+
+### Purpose
+
+`NodeTokenBufferMemory` provides **node-scoped memory** within a conversation. Each LLM node in a workflow can maintain its own independent conversation history.
+
+### Use Cases
+
+1. **Multi-LLM Workflows**: Different LLM nodes need separate context
+2. **Iterative Processing**: An LLM node in a loop needs to accumulate context across iterations
+3. **Specialized Agents**: Each agent node maintains its own dialogue history
+
+### Design: Zero Extra Storage
+
+**Key insight**: LLM node already saves complete context in `outputs["context"]`.
+
+Each LLM node execution outputs:
+```python
+outputs = {
+    "text": clean_text,
+    "context": self._build_context(prompt_messages, clean_text),  # Complete dialogue history!
+    ...
+}
+```
+
+This `outputs["context"]` contains:
+- All previous user/assistant messages (excluding system prompt)
+- The current assistant response
+
+**No separate storage needed** - we just read from the last execution's `outputs["context"]`.
+
+### Benefits
+
+| Aspect | Old Design (Object Storage) | New Design (outputs["context"]) |
+|--------|----------------------------|--------------------------------|
+| Storage | Separate JSON file | Already in WorkflowNodeExecutionModel |
+| Concurrency | Race condition risk | No issue (each execution is INSERT) |
+| Cleanup | Need separate cleanup task | Follows node execution lifecycle |
+| Migration | Required | None |
+| Complexity | High | Low |
+
+### Data Flow
+
+```
+WorkflowNodeExecutionModel        NodeTokenBufferMemory           LLM Node
+     │                                  │                           │
+     │                                  │◀── get_history_prompt_messages()
+     │                                  │                           │
+     │  SELECT outputs FROM             │                           │
+     │  workflow_node_executions        │                           │
+     │  WHERE workflow_run_id = ?       │                           │
+     │  AND node_id = ?                 │                           │
+     │◀─────────────────────────────────┤                           │
+     │                                  │                           │
+     │  outputs["context"]              │                           │
+     ├─────────────────────────────────▶│                           │
+     │                                  │                           │
+     │                    deserialize PromptMessages                │
+     │                                  │                           │
+     │                    truncate by max_token_limit               │
+     │                                  │                           │
+     │                                  │  Sequence[PromptMessage]  │
+     │                                  ├──────────────────────────▶│
+     │                                  │                           │
+```
+
+### Thread Tracking
+
+Thread extraction still uses `Message` table's `parent_message_id` structure:
+
+1. Query `Message` table for conversation → get thread's `workflow_run_ids`
+2. Get the last completed `workflow_run_id` in the thread
+3. Query `WorkflowNodeExecutionModel` for that execution's `outputs["context"]`
+
+### API
+
+```python
+class NodeTokenBufferMemory:
+    def __init__(
+        self,
+        app_id: str,
+        conversation_id: str,
+        node_id: str,
+        tenant_id: str,
+        model_instance: ModelInstance,
+    ):
+        """Initialize node-level memory."""
+        ...
+
+    def get_history_prompt_messages(
+        self,
+        *,
+        max_token_limit: int = 2000,
+        message_limit: int | None = None,
+    ) -> Sequence[PromptMessage]:
+        """
+        Retrieve history as PromptMessage sequence.
+        
+        Reads from last completed execution's outputs["context"].
+        """
+        ...
+
+    # Legacy methods (no-op, kept for compatibility)
+    def add_messages(self, *args, **kwargs) -> None: pass
+    def flush(self) -> None: pass
+    def clear(self) -> None: pass
+```
+
+### Configuration
+
+Add to `MemoryConfig` in `core/workflow/nodes/llm/entities.py`:
+
+```python
+class MemoryMode(StrEnum):
+    CONVERSATION = "conversation"  # Use TokenBufferMemory (default)
+    NODE = "node"                  # Use NodeTokenBufferMemory (Chatflow only)
+
+class MemoryConfig(BaseModel):
+    role_prefix: RolePrefix | None = None
+    window: MemoryWindowConfig | None = None
+    query_prompt_template: str | None = None
+    mode: MemoryMode = MemoryMode.CONVERSATION
+```
+
+**Mode Behavior:**
+
+| Mode           | Memory Class          | Scope                    | Availability  |
+| -------------- | --------------------- | ------------------------ | ------------- |
+| `conversation` | TokenBufferMemory     | Entire conversation      | All app modes |
+| `node`         | NodeTokenBufferMemory | Per-node in conversation | Chatflow only |
+
+> When `mode=node` is used in a non-Chatflow context (no conversation_id), it falls back to no memory.
+
+---
+
+## Comparison
+
+| Feature        | TokenBufferMemory        | NodeTokenBufferMemory              |
+| -------------- | ------------------------ | ---------------------------------- |
+| Scope          | Conversation             | Node within Conversation           |
+| Storage        | Database (Message table) | WorkflowNodeExecutionModel.outputs |
+| Thread Support | Yes                      | Yes                                |
+| File Support   | Yes (via MessageFile)    | Yes (via context serialization)    |
+| Token Limit    | Yes                      | Yes                                |
+| Use Case       | Standard chat apps       | Complex workflows                  |
+
+---
+
+## Extending to Other Nodes
+
+Currently only **LLM Node** outputs `context` in its outputs. To enable node memory for other nodes:
+
+1. Add `outputs["context"] = self._build_context(prompt_messages, response)` in the node
+2. The `NodeTokenBufferMemory` will automatically pick it up
+
+Nodes that could potentially support this:
+- `question_classifier`
+- `parameter_extractor`
+- `agent`
+
+---
+
+## Future Considerations
+
+1. **Cleanup**: Node memory lifecycle follows `WorkflowNodeExecutionModel`, which already has cleanup mechanisms
+2. **Compression**: For very long conversations, consider summarization strategies
+3. **Extension**: Other nodes may benefit from node-level memory

api/core/memory/__init__.py

@@ -0,0 +1,11 @@
+from core.memory.base import BaseMemory
+from core.memory.node_token_buffer_memory import (
+    NodeTokenBufferMemory,
+)
+from core.memory.token_buffer_memory import TokenBufferMemory
+
+__all__ = [
+    "BaseMemory",
+    "NodeTokenBufferMemory",
+    "TokenBufferMemory",
+]

api/core/memory/base.py

@@ -0,0 +1,83 @@
+"""
+Base memory interfaces and types.
+
+This module defines the common protocol for memory implementations.
+"""
+
+from abc import ABC, abstractmethod
+from collections.abc import Sequence
+
+from core.model_runtime.entities import ImagePromptMessageContent, PromptMessage
+
+
+class BaseMemory(ABC):
+    """
+    Abstract base class for memory implementations.
+
+    Provides a common interface for both conversation-level and node-level memory.
+    """
+
+    @abstractmethod
+    def get_history_prompt_messages(
+        self,
+        *,
+        max_token_limit: int = 2000,
+        message_limit: int | None = None,
+    ) -> Sequence[PromptMessage]:
+        """
+        Get history prompt messages.
+
+        :param max_token_limit: Maximum tokens for history
+        :param message_limit: Maximum number of messages
+        :return: Sequence of PromptMessage for LLM context
+        """
+        pass
+
+    def get_history_prompt_text(
+        self,
+        human_prefix: str = "Human",
+        ai_prefix: str = "Assistant",
+        max_token_limit: int = 2000,
+        message_limit: int | None = None,
+    ) -> str:
+        """
+        Get history prompt as formatted text.
+
+        :param human_prefix: Prefix for human messages
+        :param ai_prefix: Prefix for assistant messages
+        :param max_token_limit: Maximum tokens for history
+        :param message_limit: Maximum number of messages
+        :return: Formatted history text
+        """
+        from core.model_runtime.entities import (
+            PromptMessageRole,
+            TextPromptMessageContent,
+        )
+
+        prompt_messages = self.get_history_prompt_messages(
+            max_token_limit=max_token_limit,
+            message_limit=message_limit,
+        )
+
+        string_messages = []
+        for m in prompt_messages:
+            if m.role == PromptMessageRole.USER:
+                role = human_prefix
+            elif m.role == PromptMessageRole.ASSISTANT:
+                role = ai_prefix
+            else:
+                continue
+
+            if isinstance(m.content, list):
+                inner_msg = ""
+                for content in m.content:
+                    if isinstance(content, TextPromptMessageContent):
+                        inner_msg += f"{content.data}\n"
+                    elif isinstance(content, ImagePromptMessageContent):
+                        inner_msg += "[image]\n"
+                string_messages.append(f"{role}: {inner_msg.strip()}")
+            else:
+                message = f"{role}: {m.content}"
+                string_messages.append(message)
+
+        return "\n".join(string_messages)

api/core/memory/node_token_buffer_memory.py

@@ -0,0 +1,197 @@
+"""
+Node-level Token Buffer Memory for Chatflow.
+
+This module provides node-scoped memory within a conversation.
+Each LLM node in a workflow can maintain its own independent conversation history.
+
+Note: This is only available in Chatflow (advanced-chat mode) because it requires
+both conversation_id and node_id.
+
+Design:
+- History is read directly from WorkflowNodeExecutionModel.outputs["context"]
+- No separate storage needed - the context is already saved during node execution
+- Thread tracking leverages Message table's parent_message_id structure
+"""
+
+import logging
+from collections.abc import Sequence
+from typing import cast
+
+from sqlalchemy import select
+from sqlalchemy.orm import Session
+
+from core.file import file_manager
+from core.memory.base import BaseMemory
+from core.model_manager import ModelInstance
+from core.model_runtime.entities import (
+    AssistantPromptMessage,
+    MultiModalPromptMessageContent,
+    PromptMessage,
+    PromptMessageRole,
+    SystemPromptMessage,
+    ToolPromptMessage,
+    UserPromptMessage,
+)
+from core.model_runtime.entities.message_entities import PromptMessageContentUnionTypes
+from core.prompt.utils.extract_thread_messages import extract_thread_messages
+from extensions.ext_database import db
+from models.model import Message
+from models.workflow import WorkflowNodeExecutionModel
+
+logger = logging.getLogger(__name__)
+
+
+class NodeTokenBufferMemory(BaseMemory):
+    """
+    Node-level Token Buffer Memory.
+
+    Provides node-scoped memory within a conversation. Each LLM node can maintain
+    its own independent conversation history.
+
+    Key design: History is read directly from WorkflowNodeExecutionModel.outputs["context"],
+    which is already saved during node execution. No separate storage needed.
+    """
+
+    def __init__(
+        self,
+        app_id: str,
+        conversation_id: str,
+        node_id: str,
+        tenant_id: str,
+        model_instance: ModelInstance,
+    ):
+        self.app_id = app_id
+        self.conversation_id = conversation_id
+        self.node_id = node_id
+        self.tenant_id = tenant_id
+        self.model_instance = model_instance
+
+    def _get_thread_workflow_run_ids(self) -> list[str]:
+        """
+        Get workflow_run_ids for the current thread by querying Message table.
+        Returns workflow_run_ids in chronological order (oldest first).
+        """
+        with Session(db.engine, expire_on_commit=False) as session:
+            stmt = (
+                select(Message)
+                .where(Message.conversation_id == self.conversation_id)
+                .order_by(Message.created_at.desc())
+                .limit(500)
+            )
+            messages = list(session.scalars(stmt).all())
+
+        if not messages:
+            return []
+
+        # Extract thread messages using existing logic
+        thread_messages = extract_thread_messages(messages)
+
+        # For newly created message, its answer is temporarily empty, skip it
+        if thread_messages and not thread_messages[0].answer and thread_messages[0].answer_tokens == 0:
+            thread_messages.pop(0)
+
+        # Reverse to get chronological order, extract workflow_run_ids
+        return [msg.workflow_run_id for msg in reversed(thread_messages) if msg.workflow_run_id]
+
+    def _deserialize_prompt_message(self, msg_dict: dict) -> PromptMessage:
+        """Deserialize a dict to PromptMessage based on role."""
+        role = msg_dict.get("role")
+        if role in (PromptMessageRole.USER, "user"):
+            return UserPromptMessage.model_validate(msg_dict)
+        elif role in (PromptMessageRole.ASSISTANT, "assistant"):
+            return AssistantPromptMessage.model_validate(msg_dict)
+        elif role in (PromptMessageRole.SYSTEM, "system"):
+            return SystemPromptMessage.model_validate(msg_dict)
+        elif role in (PromptMessageRole.TOOL, "tool"):
+            return ToolPromptMessage.model_validate(msg_dict)
+        else:
+            return PromptMessage.model_validate(msg_dict)
+
+    def _deserialize_context(self, context_data: list[dict]) -> list[PromptMessage]:
+        """Deserialize context data from outputs to list of PromptMessage."""
+        messages = []
+        for msg_dict in context_data:
+            try:
+                msg = self._deserialize_prompt_message(msg_dict)
+                msg = self._restore_multimodal_content(msg)
+                messages.append(msg)
+            except Exception as e:
+                logger.warning("Failed to deserialize prompt message: %s", e)
+        return messages
+
+    def _restore_multimodal_content(self, message: PromptMessage) -> PromptMessage:
+        """
+        Restore multimodal content (base64 or url) from file_ref.
+
+        When context is saved, base64_data is cleared to save storage space.
+        This method restores the content by parsing file_ref (format: "method:id_or_url").
+        """
+        content = message.content
+        if content is None or isinstance(content, str):
+            return message
+
+        # Process list content, restoring multimodal data from file references
+        restored_content: list[PromptMessageContentUnionTypes] = []
+        for item in content:
+            if isinstance(item, MultiModalPromptMessageContent):
+                # restore_multimodal_content preserves the concrete subclass type
+                restored_item = file_manager.restore_multimodal_content(item)
+                restored_content.append(cast(PromptMessageContentUnionTypes, restored_item))
+            else:
+                restored_content.append(item)
+
+        return message.model_copy(update={"content": restored_content})
+
+    def get_history_prompt_messages(
+        self,
+        *,
+        max_token_limit: int = 2000,
+        message_limit: int | None = None,
+    ) -> Sequence[PromptMessage]:
+        """
+        Retrieve history as PromptMessage sequence.
+        History is read directly from the last completed node execution's outputs["context"].
+        """
+        _ = message_limit  # unused, kept for interface compatibility
+
+        thread_workflow_run_ids = self._get_thread_workflow_run_ids()
+        if not thread_workflow_run_ids:
+            return []
+
+        # Get the last completed workflow_run_id (contains accumulated context)
+        last_run_id = thread_workflow_run_ids[-1]
+
+        with Session(db.engine, expire_on_commit=False) as session:
+            stmt = select(WorkflowNodeExecutionModel).where(
+                WorkflowNodeExecutionModel.workflow_run_id == last_run_id,
+                WorkflowNodeExecutionModel.node_id == self.node_id,
+                WorkflowNodeExecutionModel.status == "succeeded",
+            )
+            execution = session.scalars(stmt).first()
+
+            if not execution:
+                return []
+
+            outputs = execution.outputs_dict
+            if not outputs:
+                return []
+
+            context_data = outputs.get("context")
+
+        if not context_data or not isinstance(context_data, list):
+            return []
+
+        prompt_messages = self._deserialize_context(context_data)
+        if not prompt_messages:
+            return []
+
+        # Truncate by token limit
+        try:
+            current_tokens = self.model_instance.get_llm_num_tokens(prompt_messages)
+            while current_tokens > max_token_limit and len(prompt_messages) > 1:
+                prompt_messages.pop(0)
+                current_tokens = self.model_instance.get_llm_num_tokens(prompt_messages)
+        except Exception as e:
+            logger.warning("Failed to count tokens for truncation: %s", e)
+
+        return prompt_messages

api/core/memory/token_buffer_memory.py

@@ -5,12 +5,12 @@ from sqlalchemy.orm import sessionmaker
 
 from core.app.app_config.features.file_upload.manager import FileUploadConfigManager
 from core.file import file_manager
+from core.memory.base import BaseMemory
 from core.model_manager import ModelInstance
 from core.model_runtime.entities import (
     AssistantPromptMessage,
     ImagePromptMessageContent,
     PromptMessage,
-    PromptMessageRole,
     TextPromptMessageContent,
     UserPromptMessage,
 )
@@ -24,7 +24,7 @@ from repositories.api_workflow_run_repository import APIWorkflowRunRepository
 from repositories.factory import DifyAPIRepositoryFactory
 
 
-class TokenBufferMemory:
+class TokenBufferMemory(BaseMemory):
     def __init__(
         self,
         conversation: Conversation,
@@ -115,10 +115,14 @@ class TokenBufferMemory:
                 return AssistantPromptMessage(content=prompt_message_contents)
 
     def get_history_prompt_messages(
-        self, max_token_limit: int = 2000, message_limit: int | None = None
+        self,
+        *,
+        max_token_limit: int = 2000,
+        message_limit: int | None = None,
     ) -> Sequence[PromptMessage]:
         """
         Get history prompt messages.
+
         :param max_token_limit: max token limit
         :param message_limit: message limit
         """
@@ -200,44 +204,3 @@ class TokenBufferMemory:
                 curr_message_tokens = self.model_instance.get_llm_num_tokens(prompt_messages)
 
         return prompt_messages
-
-    def get_history_prompt_text(
-        self,
-        human_prefix: str = "Human",
-        ai_prefix: str = "Assistant",
-        max_token_limit: int = 2000,
-        message_limit: int | None = None,
-    ) -> str:
-        """
-        Get history prompt text.
-        :param human_prefix: human prefix
-        :param ai_prefix: ai prefix
-        :param max_token_limit: max token limit
-        :param message_limit: message limit
-        :return:
-        """
-        prompt_messages = self.get_history_prompt_messages(max_token_limit=max_token_limit, message_limit=message_limit)
-
-        string_messages = []
-        for m in prompt_messages:
-            if m.role == PromptMessageRole.USER:
-                role = human_prefix
-            elif m.role == PromptMessageRole.ASSISTANT:
-                role = ai_prefix
-            else:
-                continue
-
-            if isinstance(m.content, list):
-                inner_msg = ""
-                for content in m.content:
-                    if isinstance(content, TextPromptMessageContent):
-                        inner_msg += f"{content.data}\n"
-                    elif isinstance(content, ImagePromptMessageContent):
-                        inner_msg += "[image]\n"
-
-                string_messages.append(f"{role}: {inner_msg.strip()}")
-            else:
-                message = f"{role}: {m.content}"
-                string_messages.append(message)
-
-        return "\n".join(string_messages)

api/core/model_runtime/entities/message_entities.py

@@ -91,6 +91,9 @@ class MultiModalPromptMessageContent(PromptMessageContent):
     mime_type: str = Field(default=..., description="the mime type of multi-modal file")
     filename: str = Field(default="", description="the filename of multi-modal file")
 
+    # File reference for context restoration, format: "transfer_method:related_id" or "remote:url"
+    file_ref: str | None = Field(default=None, description="Encoded file reference for restoration")
+
     @property
     def data(self):
         return self.url or f"data:{self.mime_type};base64,{self.base64_data}"
@@ -276,7 +279,5 @@ class ToolPromptMessage(PromptMessage):
 
         :return: True if prompt message is empty, False otherwise
         """
-        if not super().is_empty() and not self.tool_call_id:
-            return False
-
-        return True
+        # ToolPromptMessage is not empty if it has content OR has a tool_call_id
+        return super().is_empty() and not self.tool_call_id

api/core/prompt/advanced_prompt_transform.py

@@ -5,7 +5,7 @@ from core.app.entities.app_invoke_entities import ModelConfigWithCredentialsEnti
 from core.file import file_manager
 from core.file.models import File
 from core.helper.code_executor.jinja2.jinja2_formatter import Jinja2Formatter
-from core.memory.token_buffer_memory import TokenBufferMemory
+from core.memory.base import BaseMemory
 from core.model_runtime.entities import (
     AssistantPromptMessage,
     PromptMessage,
@@ -43,7 +43,7 @@ class AdvancedPromptTransform(PromptTransform):
         files: Sequence[File],
         context: str | None,
         memory_config: MemoryConfig | None,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
         model_config: ModelConfigWithCredentialsEntity,
         image_detail_config: ImagePromptMessageContent.DETAIL | None = None,
     ) -> list[PromptMessage]:
@@ -84,7 +84,7 @@ class AdvancedPromptTransform(PromptTransform):
         files: Sequence[File],
         context: str | None,
         memory_config: MemoryConfig | None,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
         model_config: ModelConfigWithCredentialsEntity,
         image_detail_config: ImagePromptMessageContent.DETAIL | None = None,
     ) -> list[PromptMessage]:
@@ -145,7 +145,7 @@ class AdvancedPromptTransform(PromptTransform):
         files: Sequence[File],
         context: str | None,
         memory_config: MemoryConfig | None,
-        memory: TokenBufferMemory | None,
+        memory: BaseMemory | None,
         model_config: ModelConfigWithCredentialsEntity,
         image_detail_config: ImagePromptMessageContent.DETAIL | None = None,
     ) -> list[PromptMessage]:
@@ -270,7 +270,7 @@ class AdvancedPromptTransform(PromptTransform):
 
     def _set_histories_variable(
         self,
-        memory: TokenBufferMemory,
+        memory: BaseMemory,
         memory_config: MemoryConfig,
         raw_prompt: str,
         role_prefix: MemoryConfig.RolePrefix,

api/core/prompt/entities/advanced_prompt_entities.py

@@ -1,3 +1,4 @@
+from enum import StrEnum
 from typing import Literal
 
 from pydantic import BaseModel
@@ -5,6 +6,13 @@ from pydantic import BaseModel
 from core.model_runtime.entities.message_entities import PromptMessageRole
 
 
+class MemoryMode(StrEnum):
+    """Memory mode for LLM nodes."""
+
+    CONVERSATION = "conversation"  # Use TokenBufferMemory (default, existing behavior)
+    NODE = "node"  # Use NodeTokenBufferMemory (Chatflow only)
+
+
 class ChatModelMessage(BaseModel):
     """
     Chat Message.
@@ -48,3 +56,4 @@ class MemoryConfig(BaseModel):
     role_prefix: RolePrefix | None = None
     window: WindowConfig
     query_prompt_template: str | None = None
+    mode: MemoryMode = MemoryMode.CONVERSATION

api/core/prompt/prompt_transform.py

@@ -1,7 +1,7 @@
 from typing import Any
 
 from core.app.entities.app_invoke_entities import ModelConfigWithCredentialsEntity
-from core.memory.token_buffer_memory import TokenBufferMemory
+from core.memory.base import BaseMemory
 from core.model_manager import ModelInstance
 from core.model_runtime.entities.message_entities import PromptMessage
 from core.model_runtime.entities.model_entities import ModelPropertyKey
@@ -11,7 +11,7 @@ from core.prompt.entities.advanced_prompt_entities import MemoryConfig
 class PromptTransform:
     def _append_chat_histories(
         self,
-        memory: TokenBufferMemory,
+        memory: BaseMemory,
         memory_config: MemoryConfig,
         prompt_messages: list[PromptMessage],
         model_config: ModelConfigWithCredentialsEntity,
@@ -52,7 +52,7 @@ class PromptTransform:
 
     def _get_history_messages_from_memory(
         self,
-        memory: TokenBufferMemory,
+        memory: BaseMemory,
         memory_config: MemoryConfig,
         max_token_limit: int,
         human_prefix: str | None = None,
@@ -73,7 +73,7 @@ class PromptTransform:
         return memory.get_history_prompt_text(**kwargs)
 
     def _get_history_messages_list_from_memory(
-        self, memory: TokenBufferMemory, memory_config: MemoryConfig, max_token_limit: int
+        self, memory: BaseMemory, memory_config: MemoryConfig, max_token_limit: int
     ) -> list[PromptMessage]:
         """Get memory messages."""
         return list(

api/core/tools/tool_manager.py

@@ -1047,6 +1047,8 @@ class ToolManager:
                         continue
                     tool_input = ToolNodeData.ToolInput.model_validate(tool_configurations.get(parameter.name, {}))
                     if tool_input.type == "variable":
+                        if not isinstance(tool_input.value, list):
+                            raise ToolParameterError(f"Invalid variable selector for {parameter.name}")
                         variable = variable_pool.get(tool_input.value)
                         if variable is None:
                             raise ToolParameterError(f"Variable {tool_input.value} does not exist")
@@ -1056,6 +1058,11 @@ class ToolManager:
                     elif tool_input.type == "mixed":
                         segment_group = variable_pool.convert_template(str(tool_input.value))
                         parameter_value = segment_group.text
+                    elif tool_input.type == "mention":
+                        # Mention type not supported in agent mode
+                        raise ToolParameterError(
+                            f"Mention type not supported in agent for parameter '{parameter.name}'"
+                        )
                     else:
                         raise ToolParameterError(f"Unknown tool input type '{tool_input.type}'")
                     runtime_parameters[parameter.name] = parameter_value

api/core/tools/workflow_as_tool/tool.py

@@ -5,7 +5,6 @@ import logging
 from collections.abc import Generator, Mapping, Sequence
 from typing import Any, cast
 
-from flask import has_request_context
 from sqlalchemy import select
 
 from core.db.session_factory import session_factory
@@ -29,6 +28,21 @@ from models.workflow import Workflow
 logger = logging.getLogger(__name__)
 
 
+def _try_resolve_user_from_request() -> Account | EndUser | None:
+    """
+    Try to resolve user from Flask request context.
+
+    Returns None if not in a request context or if user is not available.
+    """
+    # Note: `current_user` is a LocalProxy. Never compare it with None directly.
+    # Use _get_current_object() to dereference the proxy
+    user = getattr(current_user, "_get_current_object", lambda: current_user)()
+    # Check if we got a valid user object
+    if user is not None and hasattr(user, "id"):
+        return user
+    return None
+
+
 class WorkflowTool(Tool):
     """
     Workflow tool.
@@ -209,21 +223,13 @@ class WorkflowTool(Tool):
         Returns:
             Account | EndUser | None: The resolved user object, or None if resolution fails.
         """
-        if has_request_context():
-            return self._resolve_user_from_request()
-        else:
-            return self._resolve_user_from_database(user_id=user_id)
+        # Try to resolve user from request context first
+        user = _try_resolve_user_from_request()
+        if user is not None:
+            return user
 
-    def _resolve_user_from_request(self) -> Account | EndUser | None:
-        """
-        Resolve user from Flask request context.
-        """
-        try:
-            # Note: `current_user` is a LocalProxy. Never compare it with None directly.
-            return getattr(current_user, "_get_current_object", lambda: current_user)()
-        except Exception as e:
-            logger.warning("Failed to resolve user from request context: %s", e)
-            return None
+        # Fall back to database resolution
+        return self._resolve_user_from_database(user_id=user_id)
 
     def _resolve_user_from_database(self, user_id: str) -> Account | EndUser | None:
         """

api/core/variables/__init__.py

@@ -4,6 +4,7 @@ from .segments import (
     ArrayFileSegment,
     ArrayNumberSegment,
     ArrayObjectSegment,
+    ArrayPromptMessageSegment,
     ArraySegment,
     ArrayStringSegment,
     FileSegment,
@@ -20,6 +21,7 @@ from .variables import (
     ArrayFileVariable,
     ArrayNumberVariable,
     ArrayObjectVariable,
+    ArrayPromptMessageVariable,
     ArrayStringVariable,
     ArrayVariable,
     FileVariable,
@@ -42,6 +44,8 @@ __all__ = [
     "ArrayNumberVariable",
     "ArrayObjectSegment",
     "ArrayObjectVariable",
+    "ArrayPromptMessageSegment",
+    "ArrayPromptMessageVariable",
     "ArraySegment",
     "ArrayStringSegment",
     "ArrayStringVariable",

api/core/variables/segments.py

@@ -6,6 +6,7 @@ from typing import Annotated, Any, TypeAlias
 from pydantic import BaseModel, ConfigDict, Discriminator, Tag, field_validator
 
 from core.file import File
+from core.model_runtime.entities import PromptMessage
 
 from .types import SegmentType
 
@@ -208,6 +209,15 @@ class ArrayBooleanSegment(ArraySegment):
     value: Sequence[bool]
 
 
+class ArrayPromptMessageSegment(ArraySegment):
+    value_type: SegmentType = SegmentType.ARRAY_PROMPT_MESSAGE
+    value: Sequence[PromptMessage]
+
+    def to_object(self):
+        """Convert to JSON-serializable format for database storage and frontend."""
+        return [msg.model_dump() for msg in self.value]
+
+
 def get_segment_discriminator(v: Any) -> SegmentType | None:
     if isinstance(v, Segment):
         return v.value_type
@@ -248,6 +258,7 @@ SegmentUnion: TypeAlias = Annotated[
         | Annotated[ArrayObjectSegment, Tag(SegmentType.ARRAY_OBJECT)]
         | Annotated[ArrayFileSegment, Tag(SegmentType.ARRAY_FILE)]
         | Annotated[ArrayBooleanSegment, Tag(SegmentType.ARRAY_BOOLEAN)]
+        | Annotated[ArrayPromptMessageSegment, Tag(SegmentType.ARRAY_PROMPT_MESSAGE)]
     ),
     Discriminator(get_segment_discriminator),
 ]

api/core/variables/types.py

@@ -45,6 +45,7 @@ class SegmentType(StrEnum):
     ARRAY_OBJECT = "array[object]"
     ARRAY_FILE = "array[file]"
     ARRAY_BOOLEAN = "array[boolean]"
+    ARRAY_PROMPT_MESSAGE = "array[message]"
 
     NONE = "none"
 

api/core/variables/utils.py

@@ -3,8 +3,10 @@ from typing import Any
 
 import orjson
 
+from core.model_runtime.entities import PromptMessage
+
 from .segment_group import SegmentGroup
-from .segments import ArrayFileSegment, FileSegment, Segment
+from .segments import ArrayFileSegment, ArrayPromptMessageSegment, FileSegment, Segment
 
 
 def to_selector(node_id: str, name: str, paths: Iterable[str] = ()) -> Sequence[str]:
@@ -16,7 +18,7 @@ def to_selector(node_id: str, name: str, paths: Iterable[str] = ()) -> Sequence[
 
 def segment_orjson_default(o: Any):
     """Default function for orjson serialization of Segment types"""
-    if isinstance(o, ArrayFileSegment):
+    if isinstance(o, (ArrayFileSegment, ArrayPromptMessageSegment)):
         return [v.model_dump() for v in o.value]
     elif isinstance(o, FileSegment):
         return o.value.model_dump()
@@ -24,6 +26,8 @@ def segment_orjson_default(o: Any):
         return [segment_orjson_default(seg) for seg in o.value]
     elif isinstance(o, Segment):
         return o.value
+    elif isinstance(o, PromptMessage):
+        return o.model_dump()
     raise TypeError(f"Object of type {type(o).__name__} is not JSON serializable")
 
 

api/core/variables/variables.py

@@ -12,6 +12,7 @@ from .segments import (
     ArrayFileSegment,
     ArrayNumberSegment,
     ArrayObjectSegment,
+    ArrayPromptMessageSegment,
     ArraySegment,
     ArrayStringSegment,
     BooleanSegment,
@@ -110,6 +111,10 @@ class ArrayBooleanVariable(ArrayBooleanSegment, ArrayVariable):
     pass
 
 
+class ArrayPromptMessageVariable(ArrayPromptMessageSegment, ArrayVariable):
+    pass
+
+
 class RAGPipelineVariable(BaseModel):
     belong_to_node_id: str = Field(description="belong to which node id, shared means public")
     type: str = Field(description="variable type, text-input, paragraph, select, number,  file, file-list")
@@ -160,6 +165,7 @@ Variable: TypeAlias = Annotated[
         | Annotated[ArrayObjectVariable, Tag(SegmentType.ARRAY_OBJECT)]
         | Annotated[ArrayFileVariable, Tag(SegmentType.ARRAY_FILE)]
         | Annotated[ArrayBooleanVariable, Tag(SegmentType.ARRAY_BOOLEAN)]
+        | Annotated[ArrayPromptMessageVariable, Tag(SegmentType.ARRAY_PROMPT_MESSAGE)]
         | Annotated[SecretVariable, Tag(SegmentType.SECRET)]
     ),
     Discriminator(get_segment_discriminator),

api/core/workflow/context/__init__.py

@@ -0,0 +1,22 @@
+"""
+Execution Context - Context management for workflow execution.
+
+This package provides Flask-independent context management for workflow
+execution in multi-threaded environments.
+"""
+
+from core.workflow.context.execution_context import (
+    AppContext,
+    ExecutionContext,
+    IExecutionContext,
+    NullAppContext,
+    capture_current_context,
+)
+
+__all__ = [
+    "AppContext",
+    "ExecutionContext",
+    "IExecutionContext",
+    "NullAppContext",
+    "capture_current_context",
+]

api/core/workflow/context/execution_context.py

@@ -0,0 +1,216 @@
+"""
+Execution Context - Abstracted context management for workflow execution.
+"""
+
+import contextvars
+from abc import ABC, abstractmethod
+from collections.abc import Generator
+from contextlib import AbstractContextManager, contextmanager
+from typing import Any, Protocol, final, runtime_checkable
+
+
+class AppContext(ABC):
+    """
+    Abstract application context interface.
+
+    This abstraction allows workflow execution to work with or without Flask
+    by providing a common interface for application context management.
+    """
+
+    @abstractmethod
+    def get_config(self, key: str, default: Any = None) -> Any:
+        """Get configuration value by key."""
+        pass
+
+    @abstractmethod
+    def get_extension(self, name: str) -> Any:
+        """Get Flask extension by name (e.g., 'db', 'cache')."""
+        pass
+
+    @abstractmethod
+    def enter(self) -> AbstractContextManager[None]:
+        """Enter the application context."""
+        pass
+
+
+@runtime_checkable
+class IExecutionContext(Protocol):
+    """
+    Protocol for execution context.
+
+    This protocol defines the interface that all execution contexts must implement,
+    allowing both ExecutionContext and FlaskExecutionContext to be used interchangeably.
+    """
+
+    def __enter__(self) -> "IExecutionContext":
+        """Enter the execution context."""
+        ...
+
+    def __exit__(self, *args: Any) -> None:
+        """Exit the execution context."""
+        ...
+
+    @property
+    def user(self) -> Any:
+        """Get user object."""
+        ...
+
+
+@final
+class ExecutionContext:
+    """
+    Execution context for workflow execution in worker threads.
+
+    This class encapsulates all context needed for workflow execution:
+    - Application context (Flask app or standalone)
+    - Context variables for Python contextvars
+    - User information (optional)
+
+    It is designed to be serializable and passable to worker threads.
+    """
+
+    def __init__(
+        self,
+        app_context: AppContext | None = None,
+        context_vars: contextvars.Context | None = None,
+        user: Any = None,
+    ) -> None:
+        """
+        Initialize execution context.
+
+        Args:
+            app_context: Application context (Flask or standalone)
+            context_vars: Python contextvars to preserve
+            user: User object (optional)
+        """
+        self._app_context = app_context
+        self._context_vars = context_vars
+        self._user = user
+
+    @property
+    def app_context(self) -> AppContext | None:
+        """Get application context."""
+        return self._app_context
+
+    @property
+    def context_vars(self) -> contextvars.Context | None:
+        """Get context variables."""
+        return self._context_vars
+
+    @property
+    def user(self) -> Any:
+        """Get user object."""
+        return self._user
+
+    @contextmanager
+    def enter(self) -> Generator[None, None, None]:
+        """
+        Enter this execution context.
+
+        This is a convenience method that creates a context manager.
+        """
+        # Restore context variables if provided
+        if self._context_vars:
+            for var, val in self._context_vars.items():
+                var.set(val)
+
+        # Enter app context if available
+        if self._app_context is not None:
+            with self._app_context.enter():
+                yield
+        else:
+            yield
+
+    def __enter__(self) -> "ExecutionContext":
+        """Enter the execution context."""
+        self._cm = self.enter()
+        self._cm.__enter__()
+        return self
+
+    def __exit__(self, *args: Any) -> None:
+        """Exit the execution context."""
+        if hasattr(self, "_cm"):
+            self._cm.__exit__(*args)
+
+
+class NullAppContext(AppContext):
+    """
+    Null implementation of AppContext for non-Flask environments.
+
+    This is used when running without Flask (e.g., in tests or standalone mode).
+    """
+
+    def __init__(self, config: dict[str, Any] | None = None) -> None:
+        """
+        Initialize null app context.
+
+        Args:
+            config: Optional configuration dictionary
+        """
+        self._config = config or {}
+        self._extensions: dict[str, Any] = {}
+
+    def get_config(self, key: str, default: Any = None) -> Any:
+        """Get configuration value by key."""
+        return self._config.get(key, default)
+
+    def get_extension(self, name: str) -> Any:
+        """Get extension by name."""
+        return self._extensions.get(name)
+
+    def set_extension(self, name: str, extension: Any) -> None:
+        """Set extension by name."""
+        self._extensions[name] = extension
+
+    @contextmanager
+    def enter(self) -> Generator[None, None, None]:
+        """Enter null context (no-op)."""
+        yield
+
+
+class ExecutionContextBuilder:
+    """
+    Builder for creating ExecutionContext instances.
+
+    This provides a fluent API for building execution contexts.
+    """
+
+    def __init__(self) -> None:
+        self._app_context: AppContext | None = None
+        self._context_vars: contextvars.Context | None = None
+        self._user: Any = None
+
+    def with_app_context(self, app_context: AppContext) -> "ExecutionContextBuilder":
+        """Set application context."""
+        self._app_context = app_context
+        return self
+
+    def with_context_vars(self, context_vars: contextvars.Context) -> "ExecutionContextBuilder":
+        """Set context variables."""
+        self._context_vars = context_vars
+        return self
+
+    def with_user(self, user: Any) -> "ExecutionContextBuilder":
+        """Set user."""
+        self._user = user
+        return self
+
+    def build(self) -> ExecutionContext:
+        """Build the execution context."""
+        return ExecutionContext(
+            app_context=self._app_context,
+            context_vars=self._context_vars,
+            user=self._user,
+        )
+
+
+def capture_current_context() -> IExecutionContext:
+    """
+    Capture current execution context from the calling environment.
+
+    Returns:
+        IExecutionContext with captured context
+    """
+    from context import capture_current_context
+
+    return capture_current_context()