feat(platform): add endedAt field and fix execution analytics timestamps (#11759)

## Summary

This PR adds proper execution end time tracking and fixes timestamp
handling throughout the execution analytics system.

### Key Changes

1. **Added `endedAt` field to database schema** - Executions now have a
dedicated field for tracking when they finish
2. **Fixed timestamp nullable handling** - `started_at` and `ended_at`
are now properly nullable in types
3. **Fixed chart aggregation** - Reduced threshold from ≥3 to ≥1
executions per day
4. **Improved timestamp display** - Moved timestamps to expandable
details section in analytics table
5. **Fixed nullable timestamp bugs** - Updated all frontend code to
handle null timestamps correctly

## Problem Statement

### Issue 1: Missing Execution End Times
Previously, executions used `updatedAt` (last DB update) as a proxy for
"end time". This broke when adding correctness scores retroactively -
the end time would change to whenever the score was added, not when the
execution actually finished.

### Issue 2: Chart Shows Only One Data Point
The accuracy trends chart showed only one data point despite having
executions across multiple days. Root cause: aggregation required ≥3
executions per day.

### Issue 3: Incorrect Type Definitions
Manually maintained types defined `started_at` and `ended_at` as
non-nullable `Date`, contradicting reality where QUEUED executions
haven't started yet.

## Solution

### Database Schema (`schema.prisma`)
```prisma
model AgentGraphExecution {
  // ...
  startedAt DateTime?
  endedAt   DateTime?  // NEW FIELD
  // ...
}
```

### Execution Lifecycle
- **QUEUED**: `startedAt = null`, `endedAt = null` (not started)
- **RUNNING**: `startedAt = set`, `endedAt = null` (in progress)  
- **COMPLETED/FAILED/TERMINATED**: `startedAt = set`, `endedAt = set`
(finished)

### Migration Strategy
```sql
-- Add endedAt column
ALTER TABLE "AgentGraphExecution" ADD COLUMN "endedAt" TIMESTAMP(3);

-- Backfill ONLY terminal executions (prevents marking RUNNING executions as ended)
UPDATE "AgentGraphExecution"
SET "endedAt" = "updatedAt"
WHERE "endedAt" IS NULL
  AND "executionStatus" IN ('COMPLETED', 'FAILED', 'TERMINATED');
```

## Changes by Component

### Backend

**`schema.prisma`**
- Added `endedAt` field to `AgentGraphExecution`

**`execution.py`**
- Made `started_at` and `ended_at` optional with Field descriptions
- Updated `from_db()` to use `endedAt` instead of `updatedAt`
- `update_graph_execution_stats()` sets `endedAt` when status becomes
terminal

**`execution_analytics_routes.py`**
- Removed `created_at`/`updated_at` from `ExecutionAnalyticsResult` (DB
metadata, not execution data)
- Kept only `started_at`/`ended_at` (actual execution runtime)
- Made settings global (avoid recreation)
- Moved OpenAI key validation to `_process_batch` (only check when LLM
actually runs)

**`analytics.py`**
- Fixed aggregation: `COUNT(*) >= 1` (was 3) - include all days with ≥1
execution
- Uses `createdAt` for chart grouping (when execution was queued)

**`late_execution_monitor.py`**
- Handle optional `started_at` with fallback to `datetime.min` for
sorting
- Display "Not started" when `started_at` is null

### Frontend

**Type Definitions**
- Fixed manually maintained `types.ts`: `started_at: Date | null` (was
non-nullable)
- Generated types were already correct

**Analytics Components**
- `AnalyticsResultsTable.tsx`: Show only `started_at`/`ended_at` in
2-column expandable grid
- `ExecutionAnalyticsForm.tsx`: Added filter explanation UI

**Monitoring Components** - Fixed null handling bugs:
- `OldAgentLibraryView.tsx`: Handle null in reduce function
- `agent-runs-selector-list.tsx`: Safe sorting with `?.getTime() ?? 0`
- `AgentFlowList.tsx`: Filter/sort with null checks
- `FlowRunsStatus.tsx`: Filter null timestamps
- `FlowRunsTimeline.tsx`: Filter executions with null timestamps before
rendering
- `monitoring/page.tsx`: Safe sorting
- `ActivityItem.tsx`: Fallback to "recently" for null timestamps

## Benefits

✅ **Accurate End Times**: `endedAt` is frozen when execution finishes,
not updated later
✅ **Type Safety**: Nullable types match reality, exposing real bugs  
✅ **Better UX**: Chart shows all days with data (not just days with ≥3
executions)
✅ **Bug Fixes**: 7+ frontend components now handle null timestamps
correctly
✅ **Documentation**: Field descriptions explain when timestamps are null

## Testing

### Backend
```bash
cd autogpt_platform/backend
poetry run format  # ✅ All checks passed
poetry run lint    # ✅ All checks passed
```

### Frontend  
```bash
cd autogpt_platform/frontend
pnpm format        # ✅ All checks passed
pnpm lint          # ✅ All checks passed
pnpm types         # ✅ All type errors fixed
```

### Test Data Generation
Created script to generate 35 test executions across 7 days with
correctness scores:
```bash
poetry run python scripts/generate_test_analytics_data.py
```

## Migration Notes

⚠️ **Important**: The migration only backfills `endedAt` for executions
with terminal status (COMPLETED, FAILED, TERMINATED). Active executions
(QUEUED, RUNNING) correctly keep `endedAt = null`.

## Breaking Changes

None - this is backward compatible:
- `endedAt` is nullable, existing code that doesn't use it is unaffected
- Frontend already used generated types which were correct
- Migration safely backfills historical data

<!-- CURSOR_SUMMARY -->
---

> [!NOTE]
> Introduces explicit execution end-time tracking and normalizes
timestamp handling across backend and frontend.
> 
> - Adds `endedAt` to `AgentGraphExecution` (schema + migration);
backfills terminal executions; sets `endedAt` on terminal status updates
> - Makes `GraphExecutionMeta.started_at/ended_at` optional; updates
`from_db()` to use DB `endedAt`; exposes timestamps in
`ExecutionAnalyticsResult`
> - Moves OpenAI key validation into batch processing; instantiates
`Settings` once
> - Accuracy trends: reduce daily aggregation threshold to `>= 1`;
optional historical series
> - Monitoring/analytics UI: results table shows/export
`started_at`/`ended_at`; adds chart filter explainer
> - Frontend null-safety: update types (`Date | null`) and fix
sorting/filtering/rendering for nullable timestamps across monitoring
and library views
> - Late execution monitor: safe sorting/display when `started_at` is
null
> - OpenAPI specs updated for new/nullable fields
> 
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
1d987ca6e5. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->

---------

Co-authored-by: Nicholas Tindle <nicholas.tindle@agpt.co>

.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: () => void) {
+  useEffect(() => {
+    window.addEventListener(event, handler)
+    return () => window.removeEventListener(event, handler)
+  }, [event, handler])
+}
+```
+
+**Correct (stable subscription):**
+
+```tsx
+function useWindowEvent(event: string, handler: () => void) {
+  const handlerRef = useRef(handler)
+  useEffect(() => {
+    handlerRef.current = handler
+  }, [handler])
+
+  useEffect(() => {
+    const listener = () => handlerRef.current()
+    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: () => 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)
+  useEffect(() => {
+    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 }: { enabled: 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])
+
+  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-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 +1,82 @@
+---
+title: Batch DOM CSS Changes
+impact: MEDIUM
+impactDescription: reduces reflows/repaints
+tags: javascript, dom, css, performance, reflow
+---
+
+## Batch DOM CSS Changes
+
+Avoid changing styles one property at a time. Group multiple CSS changes together via classes or `cssText` to minimize browser reflows.
+
+**Incorrect (multiple reflows):**
+
+```typescript
+function updateElementStyles(element: HTMLElement) {
+  // Each line triggers a reflow
+  element.style.width = '100px'
+  element.style.height = '200px'
+  element.style.backgroundColor = 'blue'
+  element.style.border = '1px solid black'
+}
+```
+
+**Correct (add class - single reflow):**
+
+```typescript
+// CSS file
+.highlighted-box {
+  width: 100px;
+  height: 200px;
+  background-color: blue;
+  border: 1px solid black;
+}
+
+// JavaScript
+function updateElementStyles(element: HTMLElement) {
+  element.classList.add('highlighted-box')
+}
+```
+
+**Correct (change cssText - single reflow):**
+
+```typescript
+function updateElementStyles(element: HTMLElement) {
+  element.style.cssText = `
+    width: 100px;
+    height: 200px;
+    background-color: blue;
+    border: 1px solid black;
+  `
+}
+```
+
+**React example:**
+
+```tsx
+// Incorrect: changing styles one by one
+function Box({ isHighlighted }: { isHighlighted: boolean }) {
+  const ref = useRef<HTMLDivElement>(null)
+  
+  useEffect(() => {
+    if (ref.current && isHighlighted) {
+      ref.current.style.width = '100px'
+      ref.current.style.height = '200px'
+      ref.current.style.backgroundColor = 'blue'
+    }
+  }, [isHighlighted])
+  
+  return <div ref={ref}>Content</div>
+}
+
+// Correct: toggle class
+function Box({ isHighlighted }: { isHighlighted: boolean }) {
+  return (
+    <div className={isHighlighted ? 'highlighted-box' : ''}>
+      Content
+    </div>
+  )
+}
+```
+
+Prefer CSS classes over inline styles when possible. Classes are cached by the browser and provide better separation of concerns.

.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/join 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 for very large arrays due to spread operator limitations. 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 +1,26 @@
+---
+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.

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

@@ -0,0 +1,79 @@
+---
+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 Layout({ children }: { children: ReactNode }) {
+  const header = await fetchHeader()
+  return (
+    <div>
+      <div>{header}</div>
+      {children}
+    </div>
+  )
+}
+
+async function Sidebar() {
+  const items = await fetchSidebarItems()
+  return <nav>{items.map(renderItem)}</nav>
+}
+
+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>
+}
+```

autogpt_platform/backend/backend/api/conn_manager.py

@@ -122,24 +122,6 @@ class ConnectionManager:
 
         return len(connections)
 
-    async def broadcast_to_all(self, *, method: WSMethod, data: dict) -> int:
-        """Broadcast a message to all active websocket connections."""
-        message = WSMessage(
-            method=method,
-            data=data,
-        ).model_dump_json()
-
-        connections = tuple(self.active_connections)
-        if not connections:
-            return 0
-
-        await asyncio.gather(
-            *(connection.send_text(message) for connection in connections),
-            return_exceptions=True,
-        )
-
-        return len(connections)
-
     async def _subscribe(self, channel_key: str, websocket: WebSocket) -> str:
         if channel_key not in self.subscriptions:
             self.subscriptions[channel_key] = set()

autogpt_platform/backend/backend/api/features/admin/execution_analytics_routes.py

@@ -28,6 +28,7 @@ from backend.executor.manager import get_db_async_client
 from backend.util.settings import Settings
 
 logger = logging.getLogger(__name__)
+settings = Settings()
 
 
 class ExecutionAnalyticsRequest(BaseModel):
@@ -63,6 +64,8 @@ class ExecutionAnalyticsResult(BaseModel):
     score: Optional[float]
     status: str  # "success", "failed", "skipped"
     error_message: Optional[str] = None
+    started_at: Optional[datetime] = None
+    ended_at: Optional[datetime] = None
 
 
 class ExecutionAnalyticsResponse(BaseModel):
@@ -173,64 +176,30 @@ async def get_execution_analytics_config(
         # Return with provider prefix for clarity
         return f"{provider_name}: {model_name}"
 
-    # Get all models from the registry (dynamic, not hardcoded enum)
-    from backend.data import llm_registry
-    from backend.server.v2.llm import db as llm_db
-
-    # Get the recommended model from the database (configurable via admin UI)
-    recommended_model_slug = await llm_db.get_recommended_model_slug()
-
-    # Build the available models list
-    first_enabled_slug = None
-    for registry_model in llm_registry.iter_dynamic_models():
-        # Only include enabled models in the list
-        if not registry_model.is_enabled:
-            continue
-
-        # Track first enabled model as fallback
-        if first_enabled_slug is None:
-            first_enabled_slug = registry_model.slug
-
-        model_enum = LlmModel(registry_model.slug)  # Create enum instance from slug
-        label = generate_model_label(model_enum)
+    # Include all LlmModel values (no more filtering by hardcoded list)
+    recommended_model = LlmModel.GPT4O_MINI.value
+    for model in LlmModel:
+        label = generate_model_label(model)
         # Add "(Recommended)" suffix to the recommended model
-        if registry_model.slug == recommended_model_slug:
+        if model.value == recommended_model:
             label += " (Recommended)"
 
         available_models.append(
             ModelInfo(
-                value=registry_model.slug,
+                value=model.value,
                 label=label,
-                provider=registry_model.metadata.provider,
+                provider=model.provider,
             )
         )
 
     # Sort models by provider and name for better UX
     available_models.sort(key=lambda x: (x.provider, x.label))
 
-    # Handle case where no models are available
-    if not available_models:
-        logger.warning(
-            "No enabled LLM models found in registry. "
-            "Ensure models are configured and enabled in the LLM Registry."
-        )
-        # Provide a placeholder entry so admins see meaningful feedback
-        available_models.append(
-            ModelInfo(
-                value="",
-                label="No models available - configure in LLM Registry",
-                provider="none",
-            )
-        )
-
-    # Use the DB recommended model, or fallback to first enabled model
-    final_recommended = recommended_model_slug or first_enabled_slug or ""
-
     return ExecutionAnalyticsConfig(
         available_models=available_models,
         default_system_prompt=DEFAULT_SYSTEM_PROMPT,
         default_user_prompt=DEFAULT_USER_PROMPT,
-        recommended_model=final_recommended,
+        recommended_model=recommended_model,
     )
 
 
@@ -258,11 +227,6 @@ async def generate_execution_analytics(
     )
 
     try:
-        # Validate model configuration
-        settings = Settings()
-        if not settings.secrets.openai_internal_api_key:
-            raise HTTPException(status_code=500, detail="OpenAI API key not configured")
-
         # Get database client
         db_client = get_db_async_client()
 
@@ -354,6 +318,8 @@ async def generate_execution_analytics(
                     ),
                     status="skipped",
                     error_message=None,  # Not an error - just already processed
+                    started_at=execution.started_at,
+                    ended_at=execution.ended_at,
                 )
             )
 
@@ -383,6 +349,9 @@ async def _process_batch(
 ) -> list[ExecutionAnalyticsResult]:
     """Process a batch of executions concurrently."""
 
+    if not settings.secrets.openai_internal_api_key:
+        raise HTTPException(status_code=500, detail="OpenAI API key not configured")
+
     async def process_single_execution(execution) -> ExecutionAnalyticsResult:
         try:
             # Generate activity status and score using the specified model
@@ -421,6 +390,8 @@ async def _process_batch(
                     score=None,
                     status="skipped",
                     error_message="Activity generation returned None",
+                    started_at=execution.started_at,
+                    ended_at=execution.ended_at,
                 )
 
             # Update the execution stats
@@ -450,6 +421,8 @@ async def _process_batch(
                 summary_text=activity_response["activity_status"],
                 score=activity_response["correctness_score"],
                 status="success",
+                started_at=execution.started_at,
+                ended_at=execution.ended_at,
             )
 
         except Exception as e:
@@ -463,6 +436,8 @@ async def _process_batch(
                 score=None,
                 status="failed",
                 error_message=str(e),
+                started_at=execution.started_at,
+                ended_at=execution.ended_at,
             )
 
     # Process all executions in the batch concurrently

autogpt_platform/backend/backend/api/features/admin/llm_routes.py

@@ -1,557 +0,0 @@
-import logging
-
-import autogpt_libs.auth
-import fastapi
-
-from backend.data import llm_registry
-from backend.data.block_cost_config import refresh_llm_costs
-from backend.server.v2.llm import db as llm_db
-from backend.server.v2.llm import model as llm_model
-
-logger = logging.getLogger(__name__)
-
-router = fastapi.APIRouter(
-    tags=["llm", "admin"],
-    dependencies=[fastapi.Security(autogpt_libs.auth.requires_admin_user)],
-)
-
-
-async def _refresh_runtime_state() -> None:
-    """Refresh the LLM registry and clear all related caches to ensure real-time updates."""
-    logger.info("Refreshing LLM registry runtime state...")
-    try:
-        # Refresh registry from database
-        await llm_registry.refresh_llm_registry()
-        refresh_llm_costs()
-
-        # Clear block schema caches so they're regenerated with updated model options
-        from backend.data.block import BlockSchema
-
-        BlockSchema.clear_all_schema_caches()
-        logger.info("Cleared all block schema caches")
-
-        # Clear the /blocks endpoint cache so frontend gets updated schemas
-        try:
-            from backend.api.features.v1 import _get_cached_blocks
-
-            _get_cached_blocks.cache_clear()
-            logger.info("Cleared /blocks endpoint cache")
-        except Exception as e:
-            logger.warning("Failed to clear /blocks cache: %s", e)
-
-        # Clear the v2 builder providers cache (if it exists)
-        try:
-            from backend.api.features.builder import db as builder_db
-
-            if hasattr(builder_db, "_get_all_providers"):
-                builder_db._get_all_providers.cache_clear()
-                logger.info("Cleared v2 builder providers cache")
-        except Exception as e:
-            logger.debug("Could not clear v2 builder cache: %s", e)
-
-        # Notify all executor services to refresh their registry cache
-        from backend.data.llm_registry import publish_registry_refresh_notification
-
-        await publish_registry_refresh_notification()
-        logger.info("Published registry refresh notification")
-    except Exception as exc:
-        logger.exception(
-            "LLM runtime state refresh failed; caches may be stale: %s", exc
-        )
-
-
-@router.get(
-    "/providers",
-    summary="List LLM providers",
-    response_model=llm_model.LlmProvidersResponse,
-)
-async def list_llm_providers(include_models: bool = True):
-    providers = await llm_db.list_providers(include_models=include_models)
-    return llm_model.LlmProvidersResponse(providers=providers)
-
-
-@router.post(
-    "/providers",
-    summary="Create LLM provider",
-    response_model=llm_model.LlmProvider,
-)
-async def create_llm_provider(request: llm_model.UpsertLlmProviderRequest):
-    provider = await llm_db.upsert_provider(request=request)
-    await _refresh_runtime_state()
-    return provider
-
-
-@router.patch(
-    "/providers/{provider_id}",
-    summary="Update LLM provider",
-    response_model=llm_model.LlmProvider,
-)
-async def update_llm_provider(
-    provider_id: str,
-    request: llm_model.UpsertLlmProviderRequest,
-):
-    provider = await llm_db.upsert_provider(request=request, provider_id=provider_id)
-    await _refresh_runtime_state()
-    return provider
-
-
-@router.get(
-    "/models",
-    summary="List LLM models",
-    response_model=llm_model.LlmModelsResponse,
-)
-async def list_llm_models(provider_id: str | None = fastapi.Query(default=None)):
-    models = await llm_db.list_models(provider_id=provider_id)
-    return llm_model.LlmModelsResponse(models=models)
-
-
-@router.post(
-    "/models",
-    summary="Create LLM model",
-    response_model=llm_model.LlmModel,
-)
-async def create_llm_model(request: llm_model.CreateLlmModelRequest):
-    model = await llm_db.create_model(request=request)
-    await _refresh_runtime_state()
-    return model
-
-
-@router.patch(
-    "/models/{model_id}",
-    summary="Update LLM model",
-    response_model=llm_model.LlmModel,
-)
-async def update_llm_model(
-    model_id: str,
-    request: llm_model.UpdateLlmModelRequest,
-):
-    model = await llm_db.update_model(model_id=model_id, request=request)
-    await _refresh_runtime_state()
-    return model
-
-
-@router.patch(
-    "/models/{model_id}/toggle",
-    summary="Toggle LLM model availability",
-    response_model=llm_model.ToggleLlmModelResponse,
-)
-async def toggle_llm_model(
-    model_id: str,
-    request: llm_model.ToggleLlmModelRequest,
-):
-    """
-    Toggle a model's enabled status, optionally migrating workflows when disabling.
-
-    If disabling a model and `migrate_to_slug` is provided, all workflows using
-    this model will be migrated to the specified replacement model before disabling.
-    A migration record is created which can be reverted later using the revert endpoint.
-
-    Optional fields:
-    - `migration_reason`: Reason for the migration (e.g., "Provider outage")
-    - `custom_credit_cost`: Custom pricing override for billing during migration
-    """
-    try:
-        result = await llm_db.toggle_model(
-            model_id=model_id,
-            is_enabled=request.is_enabled,
-            migrate_to_slug=request.migrate_to_slug,
-            migration_reason=request.migration_reason,
-            custom_credit_cost=request.custom_credit_cost,
-        )
-        await _refresh_runtime_state()
-        if result.nodes_migrated > 0:
-            logger.info(
-                "Toggled model '%s' to %s and migrated %d nodes to '%s' (migration_id=%s)",
-                result.model.slug,
-                "enabled" if request.is_enabled else "disabled",
-                result.nodes_migrated,
-                result.migrated_to_slug,
-                result.migration_id,
-            )
-        return result
-    except ValueError as exc:
-        logger.warning("Model toggle validation failed: %s", exc)
-        raise fastapi.HTTPException(status_code=400, detail=str(exc)) from exc
-    except Exception as exc:
-        logger.exception("Failed to toggle LLM model %s: %s", model_id, exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to toggle model availability",
-        ) from exc
-
-
-@router.get(
-    "/models/{model_id}/usage",
-    summary="Get model usage count",
-    response_model=llm_model.LlmModelUsageResponse,
-)
-async def get_llm_model_usage(model_id: str):
-    """Get the number of workflow nodes using this model."""
-    try:
-        return await llm_db.get_model_usage(model_id=model_id)
-    except ValueError as exc:
-        raise fastapi.HTTPException(status_code=404, detail=str(exc)) from exc
-    except Exception as exc:
-        logger.exception("Failed to get model usage %s: %s", model_id, exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to get model usage",
-        ) from exc
-
-
-@router.delete(
-    "/models/{model_id}",
-    summary="Delete LLM model and migrate workflows",
-    response_model=llm_model.DeleteLlmModelResponse,
-)
-async def delete_llm_model(
-    model_id: str,
-    replacement_model_slug: str = fastapi.Query(
-        ..., description="Slug of the model to migrate existing workflows to"
-    ),
-):
-    """
-    Delete a model and automatically migrate all workflows using it to a replacement model.
-
-    This endpoint:
-    1. Validates the replacement model exists and is enabled
-    2. Counts how many workflow nodes use the model being deleted
-    3. Updates all AgentNode.constantInput->model fields to the replacement
-    4. Deletes the model record
-    5. Refreshes all caches and notifies executors
-
-    Example: DELETE /admin/llm/models/{id}?replacement_model_slug=gpt-4o
-    """
-    try:
-        result = await llm_db.delete_model(
-            model_id=model_id, replacement_model_slug=replacement_model_slug
-        )
-        await _refresh_runtime_state()
-        logger.info(
-            "Deleted model '%s' and migrated %d nodes to '%s'",
-            result.deleted_model_slug,
-            result.nodes_migrated,
-            result.replacement_model_slug,
-        )
-        return result
-    except ValueError as exc:
-        # Validation errors (model not found, replacement invalid, etc.)
-        logger.warning("Model deletion validation failed: %s", exc)
-        raise fastapi.HTTPException(status_code=400, detail=str(exc)) from exc
-    except Exception as exc:
-        logger.exception("Failed to delete LLM model %s: %s", model_id, exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to delete model and migrate workflows",
-        ) from exc
-
-
-# ============================================================================
-# Migration Management Endpoints
-# ============================================================================
-
-
-@router.get(
-    "/migrations",
-    summary="List model migrations",
-    response_model=llm_model.LlmMigrationsResponse,
-)
-async def list_llm_migrations(
-    include_reverted: bool = fastapi.Query(
-        default=False, description="Include reverted migrations in the list"
-    ),
-):
-    """
-    List all model migrations.
-
-    Migrations are created when disabling a model with the migrate_to_slug option.
-    They can be reverted to restore the original model configuration.
-    """
-    try:
-        migrations = await llm_db.list_migrations(include_reverted=include_reverted)
-        return llm_model.LlmMigrationsResponse(migrations=migrations)
-    except Exception as exc:
-        logger.exception("Failed to list migrations: %s", exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to list migrations",
-        ) from exc
-
-
-@router.get(
-    "/migrations/{migration_id}",
-    summary="Get migration details",
-    response_model=llm_model.LlmModelMigration,
-)
-async def get_llm_migration(migration_id: str):
-    """Get details of a specific migration."""
-    try:
-        migration = await llm_db.get_migration(migration_id)
-        if not migration:
-            raise fastapi.HTTPException(
-                status_code=404, detail=f"Migration '{migration_id}' not found"
-            )
-        return migration
-    except fastapi.HTTPException:
-        raise
-    except Exception as exc:
-        logger.exception("Failed to get migration %s: %s", migration_id, exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to get migration",
-        ) from exc
-
-
-@router.post(
-    "/migrations/{migration_id}/revert",
-    summary="Revert a model migration",
-    response_model=llm_model.RevertMigrationResponse,
-)
-async def revert_llm_migration(
-    migration_id: str,
-    request: llm_model.RevertMigrationRequest | None = None,
-):
-    """
-    Revert a model migration, restoring affected workflows to their original model.
-
-    This only reverts the specific nodes that were part of the migration.
-    The source model must exist for the revert to succeed.
-
-    Options:
-    - `re_enable_source_model`: Whether to re-enable the source model if disabled (default: True)
-
-    Response includes:
-    - `nodes_reverted`: Number of nodes successfully reverted
-    - `nodes_already_changed`: Number of nodes that were modified since migration (not reverted)
-    - `source_model_re_enabled`: Whether the source model was re-enabled
-
-    Requirements:
-    - Migration must not already be reverted
-    - Source model must exist
-    """
-    try:
-        re_enable = request.re_enable_source_model if request else True
-        result = await llm_db.revert_migration(
-            migration_id,
-            re_enable_source_model=re_enable,
-        )
-        await _refresh_runtime_state()
-        logger.info(
-            "Reverted migration '%s': %d nodes restored from '%s' to '%s' "
-            "(%d already changed, source re-enabled=%s)",
-            migration_id,
-            result.nodes_reverted,
-            result.target_model_slug,
-            result.source_model_slug,
-            result.nodes_already_changed,
-            result.source_model_re_enabled,
-        )
-        return result
-    except ValueError as exc:
-        logger.warning("Migration revert validation failed: %s", exc)
-        raise fastapi.HTTPException(status_code=400, detail=str(exc)) from exc
-    except Exception as exc:
-        logger.exception("Failed to revert migration %s: %s", migration_id, exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to revert migration",
-        ) from exc
-
-
-# ============================================================================
-# Creator Management Endpoints
-# ============================================================================
-
-
-@router.get(
-    "/creators",
-    summary="List model creators",
-    response_model=llm_model.LlmCreatorsResponse,
-)
-async def list_llm_creators():
-    """
-    List all model creators.
-
-    Creators are organizations that create/train models (e.g., OpenAI, Meta, Anthropic).
-    This is distinct from providers who host/serve the models (e.g., OpenRouter).
-    """
-    try:
-        creators = await llm_db.list_creators()
-        return llm_model.LlmCreatorsResponse(creators=creators)
-    except Exception as exc:
-        logger.exception("Failed to list creators: %s", exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to list creators",
-        ) from exc
-
-
-@router.get(
-    "/creators/{creator_id}",
-    summary="Get creator details",
-    operation_id="getV2GetLlmCreatorDetails",
-    response_model=llm_model.LlmModelCreator,
-)
-async def get_llm_creator(creator_id: str):
-    """Get details of a specific model creator."""
-    try:
-        creator = await llm_db.get_creator(creator_id)
-        if not creator:
-            raise fastapi.HTTPException(
-                status_code=404, detail=f"Creator '{creator_id}' not found"
-            )
-        return creator
-    except fastapi.HTTPException:
-        raise
-    except Exception as exc:
-        logger.exception("Failed to get creator %s: %s", creator_id, exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to get creator",
-        ) from exc
-
-
-@router.post(
-    "/creators",
-    summary="Create model creator",
-    response_model=llm_model.LlmModelCreator,
-)
-async def create_llm_creator(request: llm_model.UpsertLlmCreatorRequest):
-    """
-    Create a new model creator.
-
-    A creator represents an organization that creates/trains AI models,
-    such as OpenAI, Anthropic, Meta, or Google.
-    """
-    try:
-        creator = await llm_db.upsert_creator(request=request)
-        await _refresh_runtime_state()
-        logger.info("Created model creator '%s' (%s)", creator.display_name, creator.id)
-        return creator
-    except Exception as exc:
-        logger.exception("Failed to create creator: %s", exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to create creator",
-        ) from exc
-
-
-@router.patch(
-    "/creators/{creator_id}",
-    summary="Update model creator",
-    response_model=llm_model.LlmModelCreator,
-)
-async def update_llm_creator(
-    creator_id: str,
-    request: llm_model.UpsertLlmCreatorRequest,
-):
-    """Update an existing model creator."""
-    try:
-        creator = await llm_db.upsert_creator(request=request, creator_id=creator_id)
-        await _refresh_runtime_state()
-        logger.info("Updated model creator '%s' (%s)", creator.display_name, creator_id)
-        return creator
-    except Exception as exc:
-        logger.exception("Failed to update creator %s: %s", creator_id, exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to update creator",
-        ) from exc
-
-
-@router.delete(
-    "/creators/{creator_id}",
-    summary="Delete model creator",
-    response_model=dict,
-)
-async def delete_llm_creator(creator_id: str):
-    """
-    Delete a model creator.
-
-    This will remove the creator association from all models that reference it
-    (sets creatorId to NULL), but will not delete the models themselves.
-    """
-    try:
-        await llm_db.delete_creator(creator_id)
-        await _refresh_runtime_state()
-        logger.info("Deleted model creator '%s'", creator_id)
-        return {"success": True, "message": f"Creator '{creator_id}' deleted"}
-    except ValueError as exc:
-        logger.warning("Creator deletion validation failed: %s", exc)
-        raise fastapi.HTTPException(status_code=404, detail=str(exc)) from exc
-    except Exception as exc:
-        logger.exception("Failed to delete creator %s: %s", creator_id, exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to delete creator",
-        ) from exc
-
-
-# ============================================================================
-# Recommended Model Endpoints
-# ============================================================================
-
-
-@router.get(
-    "/recommended-model",
-    summary="Get recommended model",
-    response_model=llm_model.RecommendedModelResponse,
-)
-async def get_recommended_model():
-    """
-    Get the currently recommended LLM model.
-
-    The recommended model is shown to users as the default/suggested option
-    in model selection dropdowns.
-    """
-    try:
-        model = await llm_db.get_recommended_model()
-        return llm_model.RecommendedModelResponse(
-            model=model,
-            slug=model.slug if model else None,
-        )
-    except Exception as exc:
-        logger.exception("Failed to get recommended model: %s", exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to get recommended model",
-        ) from exc
-
-
-@router.post(
-    "/recommended-model",
-    summary="Set recommended model",
-    response_model=llm_model.SetRecommendedModelResponse,
-)
-async def set_recommended_model(request: llm_model.SetRecommendedModelRequest):
-    """
-    Set a model as the recommended model.
-
-    This clears the recommended flag from any other model and sets it on
-    the specified model. The model must be enabled to be set as recommended.
-
-    The recommended model is displayed to users as the default/suggested
-    option in model selection dropdowns throughout the platform.
-    """
-    try:
-        model, previous_slug = await llm_db.set_recommended_model(request.model_id)
-        await _refresh_runtime_state()
-        logger.info(
-            "Set recommended model to '%s' (previous: %s)",
-            model.slug,
-            previous_slug or "none",
-        )
-        return llm_model.SetRecommendedModelResponse(
-            model=model,
-            previous_recommended_slug=previous_slug,
-            message=f"Model '{model.display_name}' is now the recommended model",
-        )
-    except ValueError as exc:
-        logger.warning("Set recommended model validation failed: %s", exc)
-        raise fastapi.HTTPException(status_code=400, detail=str(exc)) from exc
-    except Exception as exc:
-        logger.exception("Failed to set recommended model: %s", exc)
-        raise fastapi.HTTPException(
-            status_code=500,
-            detail="Failed to set recommended model",
-        ) from exc

autogpt_platform/backend/backend/api/features/admin/llm_routes_test.py

@@ -1,405 +0,0 @@
-from unittest.mock import AsyncMock
-
-import fastapi
-import fastapi.testclient
-import pytest
-import pytest_mock
-from autogpt_libs.auth.jwt_utils import get_jwt_payload
-from pytest_snapshot.plugin import Snapshot
-
-import backend.api.features.admin.llm_routes as llm_routes
-
-app = fastapi.FastAPI()
-app.include_router(llm_routes.router)
-
-client = fastapi.testclient.TestClient(app)
-
-
-@pytest.fixture(autouse=True)
-def setup_app_admin_auth(mock_jwt_admin):
-    """Setup admin auth overrides for all tests in this module"""
-    app.dependency_overrides[get_jwt_payload] = mock_jwt_admin["get_jwt_payload"]
-    yield
-    app.dependency_overrides.clear()
-
-
-def test_list_llm_providers_success(
-    mocker: pytest_mock.MockFixture,
-    configured_snapshot: Snapshot,
-) -> None:
-    """Test successful listing of LLM providers"""
-    # Mock the database function
-    mock_providers = [
-        {
-            "id": "provider-1",
-            "name": "openai",
-            "display_name": "OpenAI",
-            "description": "OpenAI LLM provider",
-            "supports_tools": True,
-            "supports_json_output": True,
-            "supports_reasoning": False,
-            "supports_parallel_tool": True,
-            "metadata": {},
-            "models": [],
-        },
-        {
-            "id": "provider-2",
-            "name": "anthropic",
-            "display_name": "Anthropic",
-            "description": "Anthropic LLM provider",
-            "supports_tools": True,
-            "supports_json_output": True,
-            "supports_reasoning": False,
-            "supports_parallel_tool": True,
-            "metadata": {},
-            "models": [],
-        },
-    ]
-
-    mocker.patch(
-        "backend.api.features.admin.llm_routes.llm_db.list_providers",
-        new=AsyncMock(return_value=mock_providers),
-    )
-
-    response = client.get("/admin/llm/providers")
-
-    assert response.status_code == 200
-    response_data = response.json()
-    assert len(response_data["providers"]) == 2
-    assert response_data["providers"][0]["name"] == "openai"
-
-    # Snapshot test the response
-    configured_snapshot.assert_match(response_data, "list_llm_providers_success.json")
-
-
-def test_list_llm_models_success(
-    mocker: pytest_mock.MockFixture,
-    configured_snapshot: Snapshot,
-) -> None:
-    """Test successful listing of LLM models"""
-    # Mock the database function
-    mock_models = [
-        {
-            "id": "model-1",
-            "slug": "gpt-4o",
-            "display_name": "GPT-4o",
-            "description": "GPT-4 Optimized",
-            "provider_id": "provider-1",
-            "context_window": 128000,
-            "max_output_tokens": 16384,
-            "is_enabled": True,
-            "capabilities": {},
-            "metadata": {},
-            "costs": [
-                {
-                    "id": "cost-1",
-                    "credit_cost": 10,
-                    "credential_provider": "openai",
-                    "metadata": {},
-                }
-            ],
-        }
-    ]
-
-    mocker.patch(
-        "backend.api.features.admin.llm_routes.llm_db.list_models",
-        new=AsyncMock(return_value=mock_models),
-    )
-
-    response = client.get("/admin/llm/models")
-
-    assert response.status_code == 200
-    response_data = response.json()
-    assert len(response_data["models"]) == 1
-    assert response_data["models"][0]["slug"] == "gpt-4o"
-
-    # Snapshot test the response
-    configured_snapshot.assert_match(response_data, "list_llm_models_success.json")
-
-
-def test_create_llm_provider_success(
-    mocker: pytest_mock.MockFixture,
-    configured_snapshot: Snapshot,
-) -> None:
-    """Test successful creation of LLM provider"""
-    mock_provider = {
-        "id": "new-provider-id",
-        "name": "groq",
-        "display_name": "Groq",
-        "description": "Groq LLM provider",
-        "supports_tools": True,
-        "supports_json_output": True,
-        "supports_reasoning": False,
-        "supports_parallel_tool": False,
-        "metadata": {},
-    }
-
-    mocker.patch(
-        "backend.api.features.admin.llm_routes.llm_db.upsert_provider",
-        new=AsyncMock(return_value=mock_provider),
-    )
-
-    mock_refresh = mocker.patch(
-        "backend.api.features.admin.llm_routes._refresh_runtime_state",
-        new=AsyncMock(),
-    )
-
-    request_data = {
-        "name": "groq",
-        "display_name": "Groq",
-        "description": "Groq LLM provider",
-        "supports_tools": True,
-        "supports_json_output": True,
-        "supports_reasoning": False,
-        "supports_parallel_tool": False,
-        "metadata": {},
-    }
-
-    response = client.post("/admin/llm/providers", json=request_data)
-
-    assert response.status_code == 200
-    response_data = response.json()
-    assert response_data["name"] == "groq"
-    assert response_data["display_name"] == "Groq"
-
-    # Verify refresh was called
-    mock_refresh.assert_called_once()
-
-    # Snapshot test the response
-    configured_snapshot.assert_match(response_data, "create_llm_provider_success.json")
-
-
-def test_create_llm_model_success(
-    mocker: pytest_mock.MockFixture,
-    configured_snapshot: Snapshot,
-) -> None:
-    """Test successful creation of LLM model"""
-    mock_model = {
-        "id": "new-model-id",
-        "slug": "gpt-4.1-mini",
-        "display_name": "GPT-4.1 Mini",
-        "description": "Latest GPT-4.1 Mini model",
-        "provider_id": "provider-1",
-        "context_window": 128000,
-        "max_output_tokens": 16384,
-        "is_enabled": True,
-        "capabilities": {},
-        "metadata": {},
-        "costs": [
-            {
-                "id": "cost-id",
-                "credit_cost": 5,
-                "credential_provider": "openai",
-                "metadata": {},
-            }
-        ],
-    }
-
-    mocker.patch(
-        "backend.api.features.admin.llm_routes.llm_db.create_model",
-        new=AsyncMock(return_value=mock_model),
-    )
-
-    mock_refresh = mocker.patch(
-        "backend.api.features.admin.llm_routes._refresh_runtime_state",
-        new=AsyncMock(),
-    )
-
-    request_data = {
-        "slug": "gpt-4.1-mini",
-        "display_name": "GPT-4.1 Mini",
-        "description": "Latest GPT-4.1 Mini model",
-        "provider_id": "provider-1",
-        "context_window": 128000,
-        "max_output_tokens": 16384,
-        "is_enabled": True,
-        "capabilities": {},
-        "metadata": {},
-        "costs": [
-            {
-                "credit_cost": 5,
-                "credential_provider": "openai",
-                "metadata": {},
-            }
-        ],
-    }
-
-    response = client.post("/admin/llm/models", json=request_data)
-
-    assert response.status_code == 200
-    response_data = response.json()
-    assert response_data["slug"] == "gpt-4.1-mini"
-    assert response_data["is_enabled"] is True
-
-    # Verify refresh was called
-    mock_refresh.assert_called_once()
-
-    # Snapshot test the response
-    configured_snapshot.assert_match(response_data, "create_llm_model_success.json")
-
-
-def test_update_llm_model_success(
-    mocker: pytest_mock.MockFixture,
-    configured_snapshot: Snapshot,
-) -> None:
-    """Test successful update of LLM model"""
-    mock_model = {
-        "id": "model-1",
-        "slug": "gpt-4o",
-        "display_name": "GPT-4o Updated",
-        "description": "Updated description",
-        "provider_id": "provider-1",
-        "context_window": 256000,
-        "max_output_tokens": 32768,
-        "is_enabled": True,
-        "capabilities": {},
-        "metadata": {},
-        "costs": [
-            {
-                "id": "cost-1",
-                "credit_cost": 15,
-                "credential_provider": "openai",
-                "metadata": {},
-            }
-        ],
-    }
-
-    mocker.patch(
-        "backend.api.features.admin.llm_routes.llm_db.update_model",
-        new=AsyncMock(return_value=mock_model),
-    )
-
-    mock_refresh = mocker.patch(
-        "backend.api.features.admin.llm_routes._refresh_runtime_state",
-        new=AsyncMock(),
-    )
-
-    request_data = {
-        "display_name": "GPT-4o Updated",
-        "description": "Updated description",
-        "context_window": 256000,
-        "max_output_tokens": 32768,
-    }
-
-    response = client.patch("/admin/llm/models/model-1", json=request_data)
-
-    assert response.status_code == 200
-    response_data = response.json()
-    assert response_data["display_name"] == "GPT-4o Updated"
-    assert response_data["context_window"] == 256000
-
-    # Verify refresh was called
-    mock_refresh.assert_called_once()
-
-    # Snapshot test the response
-    configured_snapshot.assert_match(response_data, "update_llm_model_success.json")
-
-
-def test_toggle_llm_model_success(
-    mocker: pytest_mock.MockFixture,
-    configured_snapshot: Snapshot,
-) -> None:
-    """Test successful toggling of LLM model enabled status"""
-    mock_model = {
-        "id": "model-1",
-        "slug": "gpt-4o",
-        "display_name": "GPT-4o",
-        "description": "GPT-4 Optimized",
-        "provider_id": "provider-1",
-        "context_window": 128000,
-        "max_output_tokens": 16384,
-        "is_enabled": False,
-        "capabilities": {},
-        "metadata": {},
-        "costs": [],
-    }
-
-    mocker.patch(
-        "backend.api.features.admin.llm_routes.llm_db.toggle_model",
-        new=AsyncMock(return_value=mock_model),
-    )
-
-    mock_refresh = mocker.patch(
-        "backend.api.features.admin.llm_routes._refresh_runtime_state",
-        new=AsyncMock(),
-    )
-
-    request_data = {"is_enabled": False}
-
-    response = client.patch("/admin/llm/models/model-1/toggle", json=request_data)
-
-    assert response.status_code == 200
-    response_data = response.json()
-    assert response_data["is_enabled"] is False
-
-    # Verify refresh was called
-    mock_refresh.assert_called_once()
-
-    # Snapshot test the response
-    configured_snapshot.assert_match(response_data, "toggle_llm_model_success.json")
-
-
-def test_delete_llm_model_success(
-    mocker: pytest_mock.MockFixture,
-    configured_snapshot: Snapshot,
-) -> None:
-    """Test successful deletion of LLM model with migration"""
-    mock_response = {
-        "deleted_model_slug": "gpt-3.5-turbo",
-        "deleted_model_display_name": "GPT-3.5 Turbo",
-        "replacement_model_slug": "gpt-4o-mini",
-        "nodes_migrated": 42,
-        "message": "Successfully deleted model 'GPT-3.5 Turbo' (gpt-3.5-turbo) "
-        "and migrated 42 workflow node(s) to 'gpt-4o-mini'.",
-    }
-
-    mocker.patch(
-        "backend.api.features.admin.llm_routes.llm_db.delete_model",
-        new=AsyncMock(return_value=type("obj", (object,), mock_response)()),
-    )
-
-    mock_refresh = mocker.patch(
-        "backend.api.features.admin.llm_routes._refresh_runtime_state",
-        new=AsyncMock(),
-    )
-
-    response = client.delete(
-        "/admin/llm/models/model-1?replacement_model_slug=gpt-4o-mini"
-    )
-
-    assert response.status_code == 200
-    response_data = response.json()
-    assert response_data["deleted_model_slug"] == "gpt-3.5-turbo"
-    assert response_data["nodes_migrated"] == 42
-    assert response_data["replacement_model_slug"] == "gpt-4o-mini"
-
-    # Verify refresh was called
-    mock_refresh.assert_called_once()
-
-    # Snapshot test the response
-    configured_snapshot.assert_match(response_data, "delete_llm_model_success.json")
-
-
-def test_delete_llm_model_validation_error(
-    mocker: pytest_mock.MockFixture,
-) -> None:
-    """Test deletion fails with proper error when validation fails"""
-    mocker.patch(
-        "backend.api.features.admin.llm_routes.llm_db.delete_model",
-        new=AsyncMock(side_effect=ValueError("Replacement model 'invalid' not found")),
-    )
-
-    response = client.delete("/admin/llm/models/model-1?replacement_model_slug=invalid")
-
-    assert response.status_code == 400
-    assert "Replacement model 'invalid' not found" in response.json()["detail"]
-
-
-def test_delete_llm_model_missing_replacement(
-    mocker: pytest_mock.MockFixture,
-) -> None:
-    """Test deletion fails when replacement_model_slug is not provided"""
-    response = client.delete("/admin/llm/models/model-1")
-
-    # FastAPI will return 422 for missing required query params
-    assert response.status_code == 422

autogpt_platform/backend/backend/api/features/builder/db.py

@@ -15,7 +15,6 @@ from backend.blocks import load_all_blocks
 from backend.blocks.llm import LlmModel
 from backend.data.block import AnyBlockSchema, BlockCategory, BlockInfo, BlockSchema
 from backend.data.db import query_raw_with_schema
-from backend.data.llm_registry import get_all_model_slugs_for_validation
 from backend.integrations.providers import ProviderName
 from backend.util.cache import cached
 from backend.util.models import Pagination
@@ -32,14 +31,7 @@ from .model import (
 )
 
 logger = logging.getLogger(__name__)
-
-
-def _get_llm_models() -> list[str]:
-    """Get LLM model names for search matching from the registry."""
-    return [
-        slug.lower().replace("-", " ") for slug in get_all_model_slugs_for_validation()
-    ]
-
+llm_models = [name.name.lower().replace("_", " ") for name in LlmModel]
 
 MAX_LIBRARY_AGENT_RESULTS = 100
 MAX_MARKETPLACE_AGENT_RESULTS = 100
@@ -504,8 +496,8 @@ async def _get_static_counts():
 def _matches_llm_model(schema_cls: type[BlockSchema], query: str) -> bool:
     for field in schema_cls.model_fields.values():
         if field.annotation == LlmModel:
-            # Check if query matches any value in llm_models from registry
-            if any(query in name for name in _get_llm_models()):
+            # Check if query matches any value in llm_models
+            if any(query in name for name in llm_models):
                 return True
     return False
 

autogpt_platform/backend/backend/api/features/chat/service.py

@@ -299,9 +299,6 @@ async def stream_chat_completion(
             f"new message_count={len(session.messages)}"
         )
 
-    if len(session.messages) > config.max_context_messages:
-        raise ValueError(f"Max messages exceeded: {config.max_context_messages}")
-
     logger.info(
         f"Upserting session: {session.session_id} with user id {session.user_id}, "
         f"message_count={len(session.messages)}"

autogpt_platform/backend/backend/api/features/chat/tools/__init__.py

@@ -7,9 +7,15 @@ from backend.api.features.chat.model import ChatSession
 from .add_understanding import AddUnderstandingTool
 from .agent_output import AgentOutputTool
 from .base import BaseTool
+from .create_agent import CreateAgentTool
+from .edit_agent import EditAgentTool
 from .find_agent import FindAgentTool
+from .find_block import FindBlockTool
 from .find_library_agent import FindLibraryAgentTool
+from .get_doc_page import GetDocPageTool
 from .run_agent import RunAgentTool
+from .run_block import RunBlockTool
+from .search_docs import SearchDocsTool
 
 if TYPE_CHECKING:
     from backend.api.features.chat.response_model import StreamToolOutputAvailable
@@ -17,10 +23,16 @@ if TYPE_CHECKING:
 # Single source of truth for all tools
 TOOL_REGISTRY: dict[str, BaseTool] = {
     "add_understanding": AddUnderstandingTool(),
+    "create_agent": CreateAgentTool(),
+    "edit_agent": EditAgentTool(),
     "find_agent": FindAgentTool(),
+    "find_block": FindBlockTool(),
     "find_library_agent": FindLibraryAgentTool(),
     "run_agent": RunAgentTool(),
+    "run_block": RunBlockTool(),
     "agent_output": AgentOutputTool(),
+    "search_docs": SearchDocsTool(),
+    "get_doc_page": GetDocPageTool(),
 }
 
 # Export individual tool instances for backwards compatibility

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/__init__.py

@@ -0,0 +1,29 @@
+"""Agent generator package - Creates agents from natural language."""
+
+from .core import (
+    apply_agent_patch,
+    decompose_goal,
+    generate_agent,
+    generate_agent_patch,
+    get_agent_as_json,
+    save_agent_to_library,
+)
+from .fixer import apply_all_fixes
+from .utils import get_blocks_info
+from .validator import validate_agent
+
+__all__ = [
+    # Core functions
+    "decompose_goal",
+    "generate_agent",
+    "generate_agent_patch",
+    "apply_agent_patch",
+    "save_agent_to_library",
+    "get_agent_as_json",
+    # Fixer
+    "apply_all_fixes",
+    # Validator
+    "validate_agent",
+    # Utils
+    "get_blocks_info",
+]

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/client.py

@@ -0,0 +1,25 @@
+"""OpenRouter client configuration for agent generation."""
+
+import os
+
+from openai import AsyncOpenAI
+
+# Configuration - use OPEN_ROUTER_API_KEY for consistency with chat/config.py
+OPENROUTER_API_KEY = os.getenv("OPEN_ROUTER_API_KEY")
+AGENT_GENERATOR_MODEL = os.getenv("AGENT_GENERATOR_MODEL", "anthropic/claude-opus-4.5")
+
+# OpenRouter client (OpenAI-compatible API)
+_client: AsyncOpenAI | None = None
+
+
+def get_client() -> AsyncOpenAI:
+    """Get or create the OpenRouter client."""
+    global _client
+    if _client is None:
+        if not OPENROUTER_API_KEY:
+            raise ValueError("OPENROUTER_API_KEY environment variable is required")
+        _client = AsyncOpenAI(
+            base_url="https://openrouter.ai/api/v1",
+            api_key=OPENROUTER_API_KEY,
+        )
+    return _client

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/core.py

@@ -0,0 +1,390 @@
+"""Core agent generation functions."""
+
+import copy
+import json
+import logging
+import uuid
+from typing import Any
+
+from backend.api.features.library import db as library_db
+from backend.data.graph import Graph, Link, Node, create_graph
+
+from .client import AGENT_GENERATOR_MODEL, get_client
+from .prompts import DECOMPOSITION_PROMPT, GENERATION_PROMPT, PATCH_PROMPT
+from .utils import get_block_summaries, parse_json_from_llm
+
+logger = logging.getLogger(__name__)
+
+
+async def decompose_goal(description: str, context: str = "") -> dict[str, Any] | None:
+    """Break down a goal into steps or return clarifying questions.
+
+    Args:
+        description: Natural language goal description
+        context: Additional context (e.g., answers to previous questions)
+
+    Returns:
+        Dict with either:
+        - {"type": "clarifying_questions", "questions": [...]}
+        - {"type": "instructions", "steps": [...]}
+        Or None on error
+    """
+    client = get_client()
+    prompt = DECOMPOSITION_PROMPT.format(block_summaries=get_block_summaries())
+
+    full_description = description
+    if context:
+        full_description = f"{description}\n\nAdditional context:\n{context}"
+
+    try:
+        response = await client.chat.completions.create(
+            model=AGENT_GENERATOR_MODEL,
+            messages=[
+                {"role": "system", "content": prompt},
+                {"role": "user", "content": full_description},
+            ],
+            temperature=0,
+        )
+
+        content = response.choices[0].message.content
+        if content is None:
+            logger.error("LLM returned empty content for decomposition")
+            return None
+
+        result = parse_json_from_llm(content)
+
+        if result is None:
+            logger.error(f"Failed to parse decomposition response: {content[:200]}")
+            return None
+
+        return result
+
+    except Exception as e:
+        logger.error(f"Error decomposing goal: {e}")
+        return None
+
+
+async def generate_agent(instructions: dict[str, Any]) -> dict[str, Any] | None:
+    """Generate agent JSON from instructions.
+
+    Args:
+        instructions: Structured instructions from decompose_goal
+
+    Returns:
+        Agent JSON dict or None on error
+    """
+    client = get_client()
+    prompt = GENERATION_PROMPT.format(block_summaries=get_block_summaries())
+
+    try:
+        response = await client.chat.completions.create(
+            model=AGENT_GENERATOR_MODEL,
+            messages=[
+                {"role": "system", "content": prompt},
+                {"role": "user", "content": json.dumps(instructions, indent=2)},
+            ],
+            temperature=0,
+        )
+
+        content = response.choices[0].message.content
+        if content is None:
+            logger.error("LLM returned empty content for agent generation")
+            return None
+
+        result = parse_json_from_llm(content)
+
+        if result is None:
+            logger.error(f"Failed to parse agent JSON: {content[:200]}")
+            return None
+
+        # Ensure required fields
+        if "id" not in result:
+            result["id"] = str(uuid.uuid4())
+        if "version" not in result:
+            result["version"] = 1
+        if "is_active" not in result:
+            result["is_active"] = True
+
+        return result
+
+    except Exception as e:
+        logger.error(f"Error generating agent: {e}")
+        return None
+
+
+def json_to_graph(agent_json: dict[str, Any]) -> Graph:
+    """Convert agent JSON dict to Graph model.
+
+    Args:
+        agent_json: Agent JSON with nodes and links
+
+    Returns:
+        Graph ready for saving
+    """
+    nodes = []
+    for n in agent_json.get("nodes", []):
+        node = Node(
+            id=n.get("id", str(uuid.uuid4())),
+            block_id=n["block_id"],
+            input_default=n.get("input_default", {}),
+            metadata=n.get("metadata", {}),
+        )
+        nodes.append(node)
+
+    links = []
+    for link_data in agent_json.get("links", []):
+        link = Link(
+            id=link_data.get("id", str(uuid.uuid4())),
+            source_id=link_data["source_id"],
+            sink_id=link_data["sink_id"],
+            source_name=link_data["source_name"],
+            sink_name=link_data["sink_name"],
+            is_static=link_data.get("is_static", False),
+        )
+        links.append(link)
+
+    return Graph(
+        id=agent_json.get("id", str(uuid.uuid4())),
+        version=agent_json.get("version", 1),
+        is_active=agent_json.get("is_active", True),
+        name=agent_json.get("name", "Generated Agent"),
+        description=agent_json.get("description", ""),
+        nodes=nodes,
+        links=links,
+    )
+
+
+def _reassign_node_ids(graph: Graph) -> None:
+    """Reassign all node and link IDs to new UUIDs.
+
+    This is needed when creating a new version to avoid unique constraint violations.
+    """
+    # Create mapping from old node IDs to new UUIDs
+    id_map = {node.id: str(uuid.uuid4()) for node in graph.nodes}
+
+    # Reassign node IDs
+    for node in graph.nodes:
+        node.id = id_map[node.id]
+
+    # Update link references to use new node IDs
+    for link in graph.links:
+        link.id = str(uuid.uuid4())  # Also give links new IDs
+        if link.source_id in id_map:
+            link.source_id = id_map[link.source_id]
+        if link.sink_id in id_map:
+            link.sink_id = id_map[link.sink_id]
+
+
+async def save_agent_to_library(
+    agent_json: dict[str, Any], user_id: str, is_update: bool = False
+) -> tuple[Graph, Any]:
+    """Save agent to database and user's library.
+
+    Args:
+        agent_json: Agent JSON dict
+        user_id: User ID
+        is_update: Whether this is an update to an existing agent
+
+    Returns:
+        Tuple of (created Graph, LibraryAgent)
+    """
+    from backend.data.graph import get_graph_all_versions
+
+    graph = json_to_graph(agent_json)
+
+    if is_update:
+        # For updates, keep the same graph ID but increment version
+        # and reassign node/link IDs to avoid conflicts
+        if graph.id:
+            existing_versions = await get_graph_all_versions(graph.id, user_id)
+            if existing_versions:
+                latest_version = max(v.version for v in existing_versions)
+                graph.version = latest_version + 1
+                # Reassign node IDs (but keep graph ID the same)
+                _reassign_node_ids(graph)
+                logger.info(f"Updating agent {graph.id} to version {graph.version}")
+    else:
+        # For new agents, always generate a fresh UUID to avoid collisions
+        graph.id = str(uuid.uuid4())
+        graph.version = 1
+        # Reassign all node IDs as well
+        _reassign_node_ids(graph)
+        logger.info(f"Creating new agent with ID {graph.id}")
+
+    # Save to database
+    created_graph = await create_graph(graph, user_id)
+
+    # Add to user's library (or update existing library agent)
+    library_agents = await library_db.create_library_agent(
+        graph=created_graph,
+        user_id=user_id,
+        create_library_agents_for_sub_graphs=False,
+    )
+
+    return created_graph, library_agents[0]
+
+
+async def get_agent_as_json(
+    graph_id: str, user_id: str | None
+) -> dict[str, Any] | None:
+    """Fetch an agent and convert to JSON format for editing.
+
+    Args:
+        graph_id: Graph ID or library agent ID
+        user_id: User ID
+
+    Returns:
+        Agent as JSON dict or None if not found
+    """
+    from backend.data.graph import get_graph
+
+    # Try to get the graph (version=None gets the active version)
+    graph = await get_graph(graph_id, version=None, user_id=user_id)
+    if not graph:
+        return None
+
+    # Convert to JSON format
+    nodes = []
+    for node in graph.nodes:
+        nodes.append(
+            {
+                "id": node.id,
+                "block_id": node.block_id,
+                "input_default": node.input_default,
+                "metadata": node.metadata,
+            }
+        )
+
+    links = []
+    for node in graph.nodes:
+        for link in node.output_links:
+            links.append(
+                {
+                    "id": link.id,
+                    "source_id": link.source_id,
+                    "sink_id": link.sink_id,
+                    "source_name": link.source_name,
+                    "sink_name": link.sink_name,
+                    "is_static": link.is_static,
+                }
+            )
+
+    return {
+        "id": graph.id,
+        "name": graph.name,
+        "description": graph.description,
+        "version": graph.version,
+        "is_active": graph.is_active,
+        "nodes": nodes,
+        "links": links,
+    }
+
+
+async def generate_agent_patch(
+    update_request: str, current_agent: dict[str, Any]
+) -> dict[str, Any] | None:
+    """Generate a patch to update an existing agent.
+
+    Args:
+        update_request: Natural language description of changes
+        current_agent: Current agent JSON
+
+    Returns:
+        Patch dict or clarifying questions, or None on error
+    """
+    client = get_client()
+    prompt = PATCH_PROMPT.format(
+        current_agent=json.dumps(current_agent, indent=2),
+        block_summaries=get_block_summaries(),
+    )
+
+    try:
+        response = await client.chat.completions.create(
+            model=AGENT_GENERATOR_MODEL,
+            messages=[
+                {"role": "system", "content": prompt},
+                {"role": "user", "content": update_request},
+            ],
+            temperature=0,
+        )
+
+        content = response.choices[0].message.content
+        if content is None:
+            logger.error("LLM returned empty content for patch generation")
+            return None
+
+        return parse_json_from_llm(content)
+
+    except Exception as e:
+        logger.error(f"Error generating patch: {e}")
+        return None
+
+
+def apply_agent_patch(
+    current_agent: dict[str, Any], patch: dict[str, Any]
+) -> dict[str, Any]:
+    """Apply a patch to an existing agent.
+
+    Args:
+        current_agent: Current agent JSON
+        patch: Patch dict with operations
+
+    Returns:
+        Updated agent JSON
+    """
+    agent = copy.deepcopy(current_agent)
+    patches = patch.get("patches", [])
+
+    for p in patches:
+        patch_type = p.get("type")
+
+        if patch_type == "modify":
+            node_id = p.get("node_id")
+            changes = p.get("changes", {})
+
+            for node in agent.get("nodes", []):
+                if node["id"] == node_id:
+                    _deep_update(node, changes)
+                    logger.debug(f"Modified node {node_id}")
+                    break
+
+        elif patch_type == "add":
+            new_nodes = p.get("new_nodes", [])
+            new_links = p.get("new_links", [])
+
+            agent["nodes"] = agent.get("nodes", []) + new_nodes
+            agent["links"] = agent.get("links", []) + new_links
+            logger.debug(f"Added {len(new_nodes)} nodes, {len(new_links)} links")
+
+        elif patch_type == "remove":
+            node_ids_to_remove = set(p.get("node_ids", []))
+            link_ids_to_remove = set(p.get("link_ids", []))
+
+            # Remove nodes
+            agent["nodes"] = [
+                n for n in agent.get("nodes", []) if n["id"] not in node_ids_to_remove
+            ]
+
+            # Remove links (both explicit and those referencing removed nodes)
+            agent["links"] = [
+                link
+                for link in agent.get("links", [])
+                if link["id"] not in link_ids_to_remove
+                and link["source_id"] not in node_ids_to_remove
+                and link["sink_id"] not in node_ids_to_remove
+            ]
+
+            logger.debug(
+                f"Removed {len(node_ids_to_remove)} nodes, {len(link_ids_to_remove)} links"
+            )
+
+    return agent
+
+
+def _deep_update(target: dict, source: dict) -> None:
+    """Recursively update a dict with another dict."""
+    for key, value in source.items():
+        if key in target and isinstance(target[key], dict) and isinstance(value, dict):
+            _deep_update(target[key], value)
+        else:
+            target[key] = value

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/fixer.py

@@ -0,0 +1,606 @@
+"""Agent fixer - Fixes common LLM generation errors."""
+
+import logging
+import re
+import uuid
+from typing import Any
+
+from .utils import (
+    ADDTODICTIONARY_BLOCK_ID,
+    ADDTOLIST_BLOCK_ID,
+    CODE_EXECUTION_BLOCK_ID,
+    CONDITION_BLOCK_ID,
+    CREATEDICT_BLOCK_ID,
+    CREATELIST_BLOCK_ID,
+    DATA_SAMPLING_BLOCK_ID,
+    DOUBLE_CURLY_BRACES_BLOCK_IDS,
+    GET_CURRENT_DATE_BLOCK_ID,
+    STORE_VALUE_BLOCK_ID,
+    UNIVERSAL_TYPE_CONVERTER_BLOCK_ID,
+    get_blocks_info,
+    is_valid_uuid,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def fix_agent_ids(agent: dict[str, Any]) -> dict[str, Any]:
+    """Fix invalid UUIDs in agent and link IDs."""
+    # Fix agent ID
+    if not is_valid_uuid(agent.get("id", "")):
+        agent["id"] = str(uuid.uuid4())
+        logger.debug(f"Fixed agent ID: {agent['id']}")
+
+    # Fix node IDs
+    id_mapping = {}  # Old ID -> New ID
+    for node in agent.get("nodes", []):
+        if not is_valid_uuid(node.get("id", "")):
+            old_id = node.get("id", "")
+            new_id = str(uuid.uuid4())
+            id_mapping[old_id] = new_id
+            node["id"] = new_id
+            logger.debug(f"Fixed node ID: {old_id} -> {new_id}")
+
+    # Fix link IDs and update references
+    for link in agent.get("links", []):
+        if not is_valid_uuid(link.get("id", "")):
+            link["id"] = str(uuid.uuid4())
+            logger.debug(f"Fixed link ID: {link['id']}")
+
+        # Update source/sink IDs if they were remapped
+        if link.get("source_id") in id_mapping:
+            link["source_id"] = id_mapping[link["source_id"]]
+        if link.get("sink_id") in id_mapping:
+            link["sink_id"] = id_mapping[link["sink_id"]]
+
+    return agent
+
+
+def fix_double_curly_braces(agent: dict[str, Any]) -> dict[str, Any]:
+    """Fix single curly braces to double in template blocks."""
+    for node in agent.get("nodes", []):
+        if node.get("block_id") not in DOUBLE_CURLY_BRACES_BLOCK_IDS:
+            continue
+
+        input_data = node.get("input_default", {})
+        for key in ("prompt", "format"):
+            if key in input_data and isinstance(input_data[key], str):
+                original = input_data[key]
+                # Fix simple variable references: {var} -> {{var}}
+                fixed = re.sub(
+                    r"(?<!\{)\{([a-zA-Z_][a-zA-Z0-9_]*)\}(?!\})",
+                    r"{{\1}}",
+                    original,
+                )
+                if fixed != original:
+                    input_data[key] = fixed
+                    logger.debug(f"Fixed curly braces in {key}")
+
+    return agent
+
+
+def fix_storevalue_before_condition(agent: dict[str, Any]) -> dict[str, Any]:
+    """Add StoreValueBlock before ConditionBlock if needed for value2."""
+    nodes = agent.get("nodes", [])
+    links = agent.get("links", [])
+
+    # Find all ConditionBlock nodes
+    condition_node_ids = {
+        node["id"] for node in nodes if node.get("block_id") == CONDITION_BLOCK_ID
+    }
+
+    if not condition_node_ids:
+        return agent
+
+    new_nodes = []
+    new_links = []
+    processed_conditions = set()
+
+    for link in links:
+        sink_id = link.get("sink_id")
+        sink_name = link.get("sink_name")
+
+        # Check if this link goes to a ConditionBlock's value2
+        if sink_id in condition_node_ids and sink_name == "value2":
+            source_node = next(
+                (n for n in nodes if n["id"] == link.get("source_id")), None
+            )
+
+            # Skip if source is already a StoreValueBlock
+            if source_node and source_node.get("block_id") == STORE_VALUE_BLOCK_ID:
+                continue
+
+            # Skip if we already processed this condition
+            if sink_id in processed_conditions:
+                continue
+
+            processed_conditions.add(sink_id)
+
+            # Create StoreValueBlock
+            store_node_id = str(uuid.uuid4())
+            store_node = {
+                "id": store_node_id,
+                "block_id": STORE_VALUE_BLOCK_ID,
+                "input_default": {"data": None},
+                "metadata": {"position": {"x": 0, "y": -100}},
+            }
+            new_nodes.append(store_node)
+
+            # Create link: original source -> StoreValueBlock
+            new_links.append(
+                {
+                    "id": str(uuid.uuid4()),
+                    "source_id": link["source_id"],
+                    "source_name": link["source_name"],
+                    "sink_id": store_node_id,
+                    "sink_name": "input",
+                    "is_static": False,
+                }
+            )
+
+            # Update original link: StoreValueBlock -> ConditionBlock
+            link["source_id"] = store_node_id
+            link["source_name"] = "output"
+
+            logger.debug(f"Added StoreValueBlock before ConditionBlock {sink_id}")
+
+    if new_nodes:
+        agent["nodes"] = nodes + new_nodes
+
+    return agent
+
+
+def fix_addtolist_blocks(agent: dict[str, Any]) -> dict[str, Any]:
+    """Fix AddToList blocks by adding prerequisite empty AddToList block.
+
+    When an AddToList block is found:
+    1. Checks if there's a CreateListBlock before it
+    2. Removes CreateListBlock if linked directly to AddToList
+    3. Adds an empty AddToList block before the original
+    4. Ensures the original has a self-referencing link
+    """
+    nodes = agent.get("nodes", [])
+    links = agent.get("links", [])
+    new_nodes = []
+    original_addtolist_ids = set()
+    nodes_to_remove = set()
+    links_to_remove = []
+
+    # First pass: identify CreateListBlock nodes to remove
+    for link in links:
+        source_node = next(
+            (n for n in nodes if n.get("id") == link.get("source_id")), None
+        )
+        sink_node = next((n for n in nodes if n.get("id") == link.get("sink_id")), None)
+
+        if (
+            source_node
+            and sink_node
+            and source_node.get("block_id") == CREATELIST_BLOCK_ID
+            and sink_node.get("block_id") == ADDTOLIST_BLOCK_ID
+        ):
+            nodes_to_remove.add(source_node.get("id"))
+            links_to_remove.append(link)
+            logger.debug(f"Removing CreateListBlock {source_node.get('id')}")
+
+    # Second pass: process AddToList blocks
+    filtered_nodes = []
+    for node in nodes:
+        if node.get("id") in nodes_to_remove:
+            continue
+
+        if node.get("block_id") == ADDTOLIST_BLOCK_ID:
+            original_addtolist_ids.add(node.get("id"))
+            node_id = node.get("id")
+            pos = node.get("metadata", {}).get("position", {"x": 0, "y": 0})
+
+            # Check if already has prerequisite
+            has_prereq = any(
+                link.get("sink_id") == node_id
+                and link.get("sink_name") == "list"
+                and link.get("source_name") == "updated_list"
+                for link in links
+            )
+
+            if not has_prereq:
+                # Remove links to "list" input (except self-reference)
+                for link in links:
+                    if (
+                        link.get("sink_id") == node_id
+                        and link.get("sink_name") == "list"
+                        and link.get("source_id") != node_id
+                        and link not in links_to_remove
+                    ):
+                        links_to_remove.append(link)
+
+                # Create prerequisite AddToList block
+                prereq_id = str(uuid.uuid4())
+                prereq_node = {
+                    "id": prereq_id,
+                    "block_id": ADDTOLIST_BLOCK_ID,
+                    "input_default": {"list": [], "entry": None, "entries": []},
+                    "metadata": {
+                        "position": {"x": pos.get("x", 0) - 800, "y": pos.get("y", 0)}
+                    },
+                }
+                new_nodes.append(prereq_node)
+
+                # Link prerequisite to original
+                links.append(
+                    {
+                        "id": str(uuid.uuid4()),
+                        "source_id": prereq_id,
+                        "source_name": "updated_list",
+                        "sink_id": node_id,
+                        "sink_name": "list",
+                        "is_static": False,
+                    }
+                )
+                logger.debug(f"Added prerequisite AddToList block for {node_id}")
+
+        filtered_nodes.append(node)
+
+    # Remove marked links
+    filtered_links = [link for link in links if link not in links_to_remove]
+
+    # Add self-referencing links for original AddToList blocks
+    for node in filtered_nodes + new_nodes:
+        if (
+            node.get("block_id") == ADDTOLIST_BLOCK_ID
+            and node.get("id") in original_addtolist_ids
+        ):
+            node_id = node.get("id")
+            has_self_ref = any(
+                link["source_id"] == node_id
+                and link["sink_id"] == node_id
+                and link["source_name"] == "updated_list"
+                and link["sink_name"] == "list"
+                for link in filtered_links
+            )
+            if not has_self_ref:
+                filtered_links.append(
+                    {
+                        "id": str(uuid.uuid4()),
+                        "source_id": node_id,
+                        "source_name": "updated_list",
+                        "sink_id": node_id,
+                        "sink_name": "list",
+                        "is_static": False,
+                    }
+                )
+                logger.debug(f"Added self-reference for AddToList {node_id}")
+
+    agent["nodes"] = filtered_nodes + new_nodes
+    agent["links"] = filtered_links
+    return agent
+
+
+def fix_addtodictionary_blocks(agent: dict[str, Any]) -> dict[str, Any]:
+    """Fix AddToDictionary blocks by removing empty CreateDictionary nodes."""
+    nodes = agent.get("nodes", [])
+    links = agent.get("links", [])
+    nodes_to_remove = set()
+    links_to_remove = []
+
+    for link in links:
+        source_node = next(
+            (n for n in nodes if n.get("id") == link.get("source_id")), None
+        )
+        sink_node = next((n for n in nodes if n.get("id") == link.get("sink_id")), None)
+
+        if (
+            source_node
+            and sink_node
+            and source_node.get("block_id") == CREATEDICT_BLOCK_ID
+            and sink_node.get("block_id") == ADDTODICTIONARY_BLOCK_ID
+        ):
+            nodes_to_remove.add(source_node.get("id"))
+            links_to_remove.append(link)
+            logger.debug(f"Removing CreateDictionary {source_node.get('id')}")
+
+    agent["nodes"] = [n for n in nodes if n.get("id") not in nodes_to_remove]
+    agent["links"] = [link for link in links if link not in links_to_remove]
+    return agent
+
+
+def fix_code_execution_output(agent: dict[str, Any]) -> dict[str, Any]:
+    """Fix CodeExecutionBlock output: change 'response' to 'stdout_logs'."""
+    nodes = agent.get("nodes", [])
+    links = agent.get("links", [])
+
+    for link in links:
+        source_node = next(
+            (n for n in nodes if n.get("id") == link.get("source_id")), None
+        )
+        if (
+            source_node
+            and source_node.get("block_id") == CODE_EXECUTION_BLOCK_ID
+            and link.get("source_name") == "response"
+        ):
+            link["source_name"] = "stdout_logs"
+            logger.debug("Fixed CodeExecutionBlock output: response -> stdout_logs")
+
+    return agent
+
+
+def fix_data_sampling_sample_size(agent: dict[str, Any]) -> dict[str, Any]:
+    """Fix DataSamplingBlock by setting sample_size to 1 as default."""
+    nodes = agent.get("nodes", [])
+    links = agent.get("links", [])
+    links_to_remove = []
+
+    for node in nodes:
+        if node.get("block_id") == DATA_SAMPLING_BLOCK_ID:
+            node_id = node.get("id")
+            input_default = node.get("input_default", {})
+
+            # Remove links to sample_size
+            for link in links:
+                if (
+                    link.get("sink_id") == node_id
+                    and link.get("sink_name") == "sample_size"
+                ):
+                    links_to_remove.append(link)
+
+            # Set default
+            input_default["sample_size"] = 1
+            node["input_default"] = input_default
+            logger.debug(f"Fixed DataSamplingBlock {node_id} sample_size to 1")
+
+    if links_to_remove:
+        agent["links"] = [link for link in links if link not in links_to_remove]
+
+    return agent
+
+
+def fix_node_x_coordinates(agent: dict[str, Any]) -> dict[str, Any]:
+    """Fix node x-coordinates to ensure 800+ unit spacing between linked nodes."""
+    nodes = agent.get("nodes", [])
+    links = agent.get("links", [])
+    node_lookup = {n.get("id"): n for n in nodes}
+
+    for link in links:
+        source_id = link.get("source_id")
+        sink_id = link.get("sink_id")
+
+        source_node = node_lookup.get(source_id)
+        sink_node = node_lookup.get(sink_id)
+
+        if not source_node or not sink_node:
+            continue
+
+        source_pos = source_node.get("metadata", {}).get("position", {})
+        sink_pos = sink_node.get("metadata", {}).get("position", {})
+
+        source_x = source_pos.get("x", 0)
+        sink_x = sink_pos.get("x", 0)
+
+        if abs(sink_x - source_x) < 800:
+            new_x = source_x + 800
+            if "metadata" not in sink_node:
+                sink_node["metadata"] = {}
+            if "position" not in sink_node["metadata"]:
+                sink_node["metadata"]["position"] = {}
+            sink_node["metadata"]["position"]["x"] = new_x
+            logger.debug(f"Fixed node {sink_id} x: {sink_x} -> {new_x}")
+
+    return agent
+
+
+def fix_getcurrentdate_offset(agent: dict[str, Any]) -> dict[str, Any]:
+    """Fix GetCurrentDateBlock offset to ensure it's positive."""
+    for node in agent.get("nodes", []):
+        if node.get("block_id") == GET_CURRENT_DATE_BLOCK_ID:
+            input_default = node.get("input_default", {})
+            if "offset" in input_default:
+                offset = input_default["offset"]
+                if isinstance(offset, (int, float)) and offset < 0:
+                    input_default["offset"] = abs(offset)
+                    logger.debug(f"Fixed offset: {offset} -> {abs(offset)}")
+
+    return agent
+
+
+def fix_ai_model_parameter(
+    agent: dict[str, Any],
+    blocks_info: list[dict[str, Any]],
+    default_model: str = "gpt-4o",
+) -> dict[str, Any]:
+    """Add default model parameter to AI blocks if missing."""
+    block_map = {b.get("id"): b for b in blocks_info}
+
+    for node in agent.get("nodes", []):
+        block_id = node.get("block_id")
+        block = block_map.get(block_id)
+
+        if not block:
+            continue
+
+        # Check if block has AI category
+        categories = block.get("categories", [])
+        is_ai_block = any(
+            cat.get("category") == "AI" for cat in categories if isinstance(cat, dict)
+        )
+
+        if is_ai_block:
+            input_default = node.get("input_default", {})
+            if "model" not in input_default:
+                input_default["model"] = default_model
+                node["input_default"] = input_default
+                logger.debug(
+                    f"Added model '{default_model}' to AI block {node.get('id')}"
+                )
+
+    return agent
+
+
+def fix_link_static_properties(
+    agent: dict[str, Any], blocks_info: list[dict[str, Any]]
+) -> dict[str, Any]:
+    """Fix is_static property based on source block's staticOutput."""
+    block_map = {b.get("id"): b for b in blocks_info}
+    node_lookup = {n.get("id"): n for n in agent.get("nodes", [])}
+
+    for link in agent.get("links", []):
+        source_node = node_lookup.get(link.get("source_id"))
+        if not source_node:
+            continue
+
+        source_block = block_map.get(source_node.get("block_id"))
+        if not source_block:
+            continue
+
+        static_output = source_block.get("staticOutput", False)
+        if link.get("is_static") != static_output:
+            link["is_static"] = static_output
+            logger.debug(f"Fixed link {link.get('id')} is_static to {static_output}")
+
+    return agent
+
+
+def fix_data_type_mismatch(
+    agent: dict[str, Any], blocks_info: list[dict[str, Any]]
+) -> dict[str, Any]:
+    """Fix data type mismatches by inserting UniversalTypeConverterBlock."""
+    nodes = agent.get("nodes", [])
+    links = agent.get("links", [])
+    block_map = {b.get("id"): b for b in blocks_info}
+    node_lookup = {n.get("id"): n for n in nodes}
+
+    def get_property_type(schema: dict, name: str) -> str | None:
+        if "_#_" in name:
+            parent, child = name.split("_#_", 1)
+            parent_schema = schema.get(parent, {})
+            if "properties" in parent_schema:
+                return parent_schema["properties"].get(child, {}).get("type")
+            return None
+        return schema.get(name, {}).get("type")
+
+    def are_types_compatible(src: str, sink: str) -> bool:
+        if {src, sink} <= {"integer", "number"}:
+            return True
+        return src == sink
+
+    type_mapping = {
+        "string": "string",
+        "text": "string",
+        "integer": "number",
+        "number": "number",
+        "float": "number",
+        "boolean": "boolean",
+        "bool": "boolean",
+        "array": "list",
+        "list": "list",
+        "object": "dictionary",
+        "dict": "dictionary",
+        "dictionary": "dictionary",
+    }
+
+    new_links = []
+    nodes_to_add = []
+
+    for link in links:
+        source_node = node_lookup.get(link.get("source_id"))
+        sink_node = node_lookup.get(link.get("sink_id"))
+
+        if not source_node or not sink_node:
+            new_links.append(link)
+            continue
+
+        source_block = block_map.get(source_node.get("block_id"))
+        sink_block = block_map.get(sink_node.get("block_id"))
+
+        if not source_block or not sink_block:
+            new_links.append(link)
+            continue
+
+        source_outputs = source_block.get("outputSchema", {}).get("properties", {})
+        sink_inputs = sink_block.get("inputSchema", {}).get("properties", {})
+
+        source_type = get_property_type(source_outputs, link.get("source_name", ""))
+        sink_type = get_property_type(sink_inputs, link.get("sink_name", ""))
+
+        if (
+            source_type
+            and sink_type
+            and not are_types_compatible(source_type, sink_type)
+        ):
+            # Insert type converter
+            converter_id = str(uuid.uuid4())
+            target_type = type_mapping.get(sink_type, sink_type)
+
+            converter_node = {
+                "id": converter_id,
+                "block_id": UNIVERSAL_TYPE_CONVERTER_BLOCK_ID,
+                "input_default": {"type": target_type},
+                "metadata": {"position": {"x": 0, "y": 100}},
+            }
+            nodes_to_add.append(converter_node)
+
+            # source -> converter
+            new_links.append(
+                {
+                    "id": str(uuid.uuid4()),
+                    "source_id": link["source_id"],
+                    "source_name": link["source_name"],
+                    "sink_id": converter_id,
+                    "sink_name": "value",
+                    "is_static": False,
+                }
+            )
+
+            # converter -> sink
+            new_links.append(
+                {
+                    "id": str(uuid.uuid4()),
+                    "source_id": converter_id,
+                    "source_name": "value",
+                    "sink_id": link["sink_id"],
+                    "sink_name": link["sink_name"],
+                    "is_static": False,
+                }
+            )
+
+            logger.debug(f"Inserted type converter: {source_type} -> {target_type}")
+        else:
+            new_links.append(link)
+
+    if nodes_to_add:
+        agent["nodes"] = nodes + nodes_to_add
+        agent["links"] = new_links
+
+    return agent
+
+
+def apply_all_fixes(
+    agent: dict[str, Any], blocks_info: list[dict[str, Any]] | None = None
+) -> dict[str, Any]:
+    """Apply all fixes to an agent JSON.
+
+    Args:
+        agent: Agent JSON dict
+        blocks_info: Optional list of block info dicts for advanced fixes
+
+    Returns:
+        Fixed agent JSON
+    """
+    # Basic fixes (no block info needed)
+    agent = fix_agent_ids(agent)
+    agent = fix_double_curly_braces(agent)
+    agent = fix_storevalue_before_condition(agent)
+    agent = fix_addtolist_blocks(agent)
+    agent = fix_addtodictionary_blocks(agent)
+    agent = fix_code_execution_output(agent)
+    agent = fix_data_sampling_sample_size(agent)
+    agent = fix_node_x_coordinates(agent)
+    agent = fix_getcurrentdate_offset(agent)
+
+    # Advanced fixes (require block info)
+    if blocks_info is None:
+        blocks_info = get_blocks_info()
+
+    agent = fix_ai_model_parameter(agent, blocks_info)
+    agent = fix_link_static_properties(agent, blocks_info)
+    agent = fix_data_type_mismatch(agent, blocks_info)
+
+    return agent

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/prompts.py

@@ -0,0 +1,225 @@
+"""Prompt templates for agent generation."""
+
+DECOMPOSITION_PROMPT = """
+You are an expert AutoGPT Workflow Decomposer. Your task is to analyze a user's high-level goal and break it down into a clear, step-by-step plan using the available blocks.
+
+Each step should represent a distinct, automatable action suitable for execution by an AI automation system.
+
+---
+
+FIRST: Analyze the user's goal and determine:
+1) Design-time configuration (fixed settings that won't change per run)
+2) Runtime inputs (values the agent's end-user will provide each time it runs)
+
+For anything that can vary per run (email addresses, names, dates, search terms, etc.):
+- DO NOT ask for the actual value
+- Instead, define it as an Agent Input with a clear name, type, and description
+
+Only ask clarifying questions about design-time config that affects how you build the workflow:
+- Which external service to use (e.g., "Gmail vs Outlook", "Notion vs Google Docs")
+- Required formats or structures (e.g., "CSV, JSON, or PDF output?")
+- Business rules that must be hard-coded
+
+IMPORTANT CLARIFICATIONS POLICY:
+- Ask no more than five essential questions
+- Do not ask for concrete values that can be provided at runtime as Agent Inputs
+- Do not ask for API keys or credentials; the platform handles those directly
+- If there is enough information to infer reasonable defaults, prefer to propose defaults
+
+---
+
+GUIDELINES:
+1. List each step as a numbered item
+2. Describe the action clearly and specify inputs/outputs
+3. Ensure steps are in logical, sequential order
+4. Mention block names naturally (e.g., "Use GetWeatherByLocationBlock to...")
+5. Help the user reach their goal efficiently
+
+---
+
+RULES:
+1. OUTPUT FORMAT: Only output either clarifying questions OR step-by-step instructions, not both
+2. USE ONLY THE BLOCKS PROVIDED
+3. ALL required_input fields must be provided
+4. Data types of linked properties must match
+5. Write expert-level prompts for AI-related blocks
+
+---
+
+CRITICAL BLOCK RESTRICTIONS:
+1. AddToListBlock: Outputs updated list EVERY addition, not after all additions
+2. SendEmailBlock: Draft the email for user review; set SMTP config based on email type
+3. ConditionBlock: value2 is reference, value1 is contrast
+4. CodeExecutionBlock: DO NOT USE - use AI blocks instead
+5. ReadCsvBlock: Only use the 'rows' output, not 'row'
+
+---
+
+OUTPUT FORMAT:
+
+If more information is needed:
+```json
+{{
+  "type": "clarifying_questions",
+  "questions": [
+    {{
+      "question": "Which email provider should be used? (Gmail, Outlook, custom SMTP)",
+      "keyword": "email_provider",
+      "example": "Gmail"
+    }}
+  ]
+}}
+```
+
+If ready to proceed:
+```json
+{{
+  "type": "instructions",
+  "steps": [
+    {{
+      "step_number": 1,
+      "block_name": "AgentShortTextInputBlock",
+      "description": "Get the URL of the content to analyze.",
+      "inputs": [{{"name": "name", "value": "URL"}}],
+      "outputs": [{{"name": "result", "description": "The URL entered by user"}}]
+    }}
+  ]
+}}
+```
+
+---
+
+AVAILABLE BLOCKS:
+{block_summaries}
+"""
+
+GENERATION_PROMPT = """
+You are an expert AI workflow builder. Generate a valid agent JSON from the given instructions.
+
+---
+
+NODES:
+Each node must include:
+- `id`: Unique UUID v4 (e.g. `a8f5b1e2-c3d4-4e5f-8a9b-0c1d2e3f4a5b`)
+- `block_id`: The block identifier (must match an Allowed Block)
+- `input_default`: Dict of inputs (can be empty if no static inputs needed)
+- `metadata`: Must contain:
+  - `position`: {{"x": number, "y": number}} - adjacent nodes should differ by 800+ in X
+  - `customized_name`: Clear name describing this block's purpose in the workflow
+
+---
+
+LINKS:
+Each link connects a source node's output to a sink node's input:
+- `id`: MUST be UUID v4 (NOT "link-1", "link-2", etc.)
+- `source_id`: ID of the source node
+- `source_name`: Output field name from the source block
+- `sink_id`: ID of the sink node
+- `sink_name`: Input field name on the sink block
+- `is_static`: true only if source block has static_output: true
+
+CRITICAL: All IDs must be valid UUID v4 format!
+
+---
+
+AGENT (GRAPH):
+Wrap nodes and links in:
+- `id`: UUID of the agent
+- `name`: Short, generic name (avoid specific company names, URLs)
+- `description`: Short, generic description
+- `nodes`: List of all nodes
+- `links`: List of all links
+- `version`: 1
+- `is_active`: true
+
+---
+
+TIPS:
+- All required_input fields must be provided via input_default or a valid link
+- Ensure consistent source_id and sink_id references
+- Avoid dangling links
+- Input/output pins must match block schemas
+- Do not invent unknown block_ids
+
+---
+
+ALLOWED BLOCKS:
+{block_summaries}
+
+---
+
+Generate the complete agent JSON. Output ONLY valid JSON, no explanation.
+"""
+
+PATCH_PROMPT = """
+You are an expert at modifying AutoGPT agent workflows. Given the current agent and a modification request, generate a JSON patch to update the agent.
+
+CURRENT AGENT:
+{current_agent}
+
+AVAILABLE BLOCKS:
+{block_summaries}
+
+---
+
+PATCH FORMAT:
+Return a JSON object with the following structure:
+
+```json
+{{
+  "type": "patch",
+  "intent": "Brief description of what the patch does",
+  "patches": [
+    {{
+      "type": "modify",
+      "node_id": "uuid-of-node-to-modify",
+      "changes": {{
+        "input_default": {{"field": "new_value"}},
+        "metadata": {{"customized_name": "New Name"}}
+      }}
+    }},
+    {{
+      "type": "add",
+      "new_nodes": [
+        {{
+          "id": "new-uuid",
+          "block_id": "block-uuid",
+          "input_default": {{}},
+          "metadata": {{"position": {{"x": 0, "y": 0}}, "customized_name": "Name"}}
+        }}
+      ],
+      "new_links": [
+        {{
+          "id": "link-uuid",
+          "source_id": "source-node-id",
+          "source_name": "output_field",
+          "sink_id": "sink-node-id",
+          "sink_name": "input_field"
+        }}
+      ]
+    }},
+    {{
+      "type": "remove",
+      "node_ids": ["uuid-of-node-to-remove"],
+      "link_ids": ["uuid-of-link-to-remove"]
+    }}
+  ]
+}}
+```
+
+If you need more information, return:
+```json
+{{
+  "type": "clarifying_questions",
+  "questions": [
+    {{
+      "question": "What specific change do you want?",
+      "keyword": "change_type",
+      "example": "Add error handling"
+    }}
+  ]
+}}
+```
+
+Generate the minimal patch needed. Output ONLY valid JSON.
+"""

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/utils.py

@@ -0,0 +1,213 @@
+"""Utilities for agent generation."""
+
+import json
+import re
+from typing import Any
+
+from backend.data.block import get_blocks
+
+# UUID validation regex
+UUID_REGEX = re.compile(
+    r"^[a-f0-9]{8}-[a-f0-9]{4}-4[a-f0-9]{3}-[89ab][a-f0-9]{3}-[a-f0-9]{12}$"
+)
+
+# Block IDs for various fixes
+STORE_VALUE_BLOCK_ID = "1ff065e9-88e8-4358-9d82-8dc91f622ba9"
+CONDITION_BLOCK_ID = "715696a0-e1da-45c8-b209-c2fa9c3b0be6"
+ADDTOLIST_BLOCK_ID = "aeb08fc1-2fc1-4141-bc8e-f758f183a822"
+ADDTODICTIONARY_BLOCK_ID = "31d1064e-7446-4693-a7d4-65e5ca1180d1"
+CREATELIST_BLOCK_ID = "a912d5c7-6e00-4542-b2a9-8034136930e4"
+CREATEDICT_BLOCK_ID = "b924ddf4-de4f-4b56-9a85-358930dcbc91"
+CODE_EXECUTION_BLOCK_ID = "0b02b072-abe7-11ef-8372-fb5d162dd712"
+DATA_SAMPLING_BLOCK_ID = "4a448883-71fa-49cf-91cf-70d793bd7d87"
+UNIVERSAL_TYPE_CONVERTER_BLOCK_ID = "95d1b990-ce13-4d88-9737-ba5c2070c97b"
+GET_CURRENT_DATE_BLOCK_ID = "b29c1b50-5d0e-4d9f-8f9d-1b0e6fcbf0b1"
+
+DOUBLE_CURLY_BRACES_BLOCK_IDS = [
+    "44f6c8ad-d75c-4ae1-8209-aad1c0326928",  # FillTextTemplateBlock
+    "6ab085e2-20b3-4055-bc3e-08036e01eca6",
+    "90f8c45e-e983-4644-aa0b-b4ebe2f531bc",
+    "363ae599-353e-4804-937e-b2ee3cef3da4",  # AgentOutputBlock
+    "3b191d9f-356f-482d-8238-ba04b6d18381",
+    "db7d8f02-2f44-4c55-ab7a-eae0941f0c30",
+    "3a7c4b8d-6e2f-4a5d-b9c1-f8d23c5a9b0e",
+    "ed1ae7a0-b770-4089-b520-1f0005fad19a",
+    "a892b8d9-3e4e-4e9c-9c1e-75f8efcf1bfa",
+    "b29c1b50-5d0e-4d9f-8f9d-1b0e6fcbf0b1",
+    "716a67b3-6760-42e7-86dc-18645c6e00fc",
+    "530cf046-2ce0-4854-ae2c-659db17c7a46",
+    "ed55ac19-356e-4243-a6cb-bc599e9b716f",
+    "1f292d4a-41a4-4977-9684-7c8d560b9f91",  # LLM blocks
+    "32a87eab-381e-4dd4-bdb8-4c47151be35a",
+]
+
+
+def is_valid_uuid(value: str) -> bool:
+    """Check if a string is a valid UUID v4."""
+    return isinstance(value, str) and UUID_REGEX.match(value) is not None
+
+
+def _compact_schema(schema: dict) -> dict[str, str]:
+    """Extract compact type info from a JSON schema properties dict.
+
+    Returns a dict of {field_name: type_string} for essential info only.
+    """
+    props = schema.get("properties", {})
+    result = {}
+
+    for name, prop in props.items():
+        # Skip internal/complex fields
+        if name.startswith("_"):
+            continue
+
+        # Get type string
+        type_str = prop.get("type", "any")
+
+        # Handle anyOf/oneOf (optional types)
+        if "anyOf" in prop:
+            types = [t.get("type", "?") for t in prop["anyOf"] if t.get("type")]
+            type_str = "|".join(types) if types else "any"
+        elif "allOf" in prop:
+            type_str = "object"
+
+        # Add array item type if present
+        if type_str == "array" and "items" in prop:
+            items = prop["items"]
+            if isinstance(items, dict):
+                item_type = items.get("type", "any")
+                type_str = f"array[{item_type}]"
+
+        result[name] = type_str
+
+    return result
+
+
+def get_block_summaries(include_schemas: bool = True) -> str:
+    """Generate compact block summaries for prompts.
+
+    Args:
+        include_schemas: Whether to include input/output type info
+
+    Returns:
+        Formatted string of block summaries (compact format)
+    """
+    blocks = get_blocks()
+    summaries = []
+
+    for block_id, block_cls in blocks.items():
+        block = block_cls()
+        name = block.name
+        desc = getattr(block, "description", "") or ""
+
+        # Truncate description
+        if len(desc) > 150:
+            desc = desc[:147] + "..."
+
+        if not include_schemas:
+            summaries.append(f"- {name} (id: {block_id}): {desc}")
+        else:
+            # Compact format with type info only
+            inputs = {}
+            outputs = {}
+            required = []
+
+            if hasattr(block, "input_schema"):
+                try:
+                    schema = block.input_schema.jsonschema()
+                    inputs = _compact_schema(schema)
+                    required = schema.get("required", [])
+                except Exception:
+                    pass
+
+            if hasattr(block, "output_schema"):
+                try:
+                    schema = block.output_schema.jsonschema()
+                    outputs = _compact_schema(schema)
+                except Exception:
+                    pass
+
+            # Build compact line format
+            # Format: NAME (id): desc | in: {field:type, ...} [required] | out: {field:type}
+            in_str = ", ".join(f"{k}:{v}" for k, v in inputs.items())
+            out_str = ", ".join(f"{k}:{v}" for k, v in outputs.items())
+            req_str = f" req=[{','.join(required)}]" if required else ""
+
+            static = " [static]" if getattr(block, "static_output", False) else ""
+
+            line = f"- {name} (id: {block_id}): {desc}"
+            if in_str:
+                line += f"\n  in: {{{in_str}}}{req_str}"
+            if out_str:
+                line += f"\n  out: {{{out_str}}}{static}"
+
+            summaries.append(line)
+
+    return "\n".join(summaries)
+
+
+def get_blocks_info() -> list[dict[str, Any]]:
+    """Get block information with schemas for validation and fixing."""
+    blocks = get_blocks()
+    blocks_info = []
+    for block_id, block_cls in blocks.items():
+        block = block_cls()
+        blocks_info.append(
+            {
+                "id": block_id,
+                "name": block.name,
+                "description": getattr(block, "description", ""),
+                "categories": getattr(block, "categories", []),
+                "staticOutput": getattr(block, "static_output", False),
+                "inputSchema": (
+                    block.input_schema.jsonschema()
+                    if hasattr(block, "input_schema")
+                    else {}
+                ),
+                "outputSchema": (
+                    block.output_schema.jsonschema()
+                    if hasattr(block, "output_schema")
+                    else {}
+                ),
+            }
+        )
+    return blocks_info
+
+
+def parse_json_from_llm(text: str) -> dict[str, Any] | None:
+    """Extract JSON from LLM response (handles markdown code blocks)."""
+    if not text:
+        return None
+
+    # Try fenced code block
+    match = re.search(r"```(?:json)?\s*([\s\S]*?)```", text, re.IGNORECASE)
+    if match:
+        try:
+            return json.loads(match.group(1).strip())
+        except json.JSONDecodeError:
+            pass
+
+    # Try raw text
+    try:
+        return json.loads(text.strip())
+    except json.JSONDecodeError:
+        pass
+
+    # Try finding {...} span
+    start = text.find("{")
+    end = text.rfind("}")
+    if start != -1 and end > start:
+        try:
+            return json.loads(text[start : end + 1])
+        except json.JSONDecodeError:
+            pass
+
+    # Try finding [...] span
+    start = text.find("[")
+    end = text.rfind("]")
+    if start != -1 and end > start:
+        try:
+            return json.loads(text[start : end + 1])
+        except json.JSONDecodeError:
+            pass
+
+    return None

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/validator.py

@@ -0,0 +1,279 @@
+"""Agent validator - Validates agent structure and connections."""
+
+import logging
+import re
+from typing import Any
+
+from .utils import get_blocks_info
+
+logger = logging.getLogger(__name__)
+
+
+class AgentValidator:
+    """Validator for AutoGPT agents with detailed error reporting."""
+
+    def __init__(self):
+        self.errors: list[str] = []
+
+    def add_error(self, error: str) -> None:
+        """Add an error message."""
+        self.errors.append(error)
+
+    def validate_block_existence(
+        self, agent: dict[str, Any], blocks_info: list[dict[str, Any]]
+    ) -> bool:
+        """Validate all block IDs exist in the blocks library."""
+        valid = True
+        valid_block_ids = {b.get("id") for b in blocks_info if b.get("id")}
+
+        for node in agent.get("nodes", []):
+            block_id = node.get("block_id")
+            node_id = node.get("id")
+
+            if not block_id:
+                self.add_error(f"Node '{node_id}' is missing 'block_id' field.")
+                valid = False
+                continue
+
+            if block_id not in valid_block_ids:
+                self.add_error(
+                    f"Node '{node_id}' references block_id '{block_id}' which does not exist."
+                )
+                valid = False
+
+        return valid
+
+    def validate_link_node_references(self, agent: dict[str, Any]) -> bool:
+        """Validate all node IDs referenced in links exist."""
+        valid = True
+        valid_node_ids = {n.get("id") for n in agent.get("nodes", []) if n.get("id")}
+
+        for link in agent.get("links", []):
+            link_id = link.get("id", "Unknown")
+            source_id = link.get("source_id")
+            sink_id = link.get("sink_id")
+
+            if not source_id:
+                self.add_error(f"Link '{link_id}' is missing 'source_id'.")
+                valid = False
+            elif source_id not in valid_node_ids:
+                self.add_error(
+                    f"Link '{link_id}' references non-existent source_id '{source_id}'."
+                )
+                valid = False
+
+            if not sink_id:
+                self.add_error(f"Link '{link_id}' is missing 'sink_id'.")
+                valid = False
+            elif sink_id not in valid_node_ids:
+                self.add_error(
+                    f"Link '{link_id}' references non-existent sink_id '{sink_id}'."
+                )
+                valid = False
+
+        return valid
+
+    def validate_required_inputs(
+        self, agent: dict[str, Any], blocks_info: list[dict[str, Any]]
+    ) -> bool:
+        """Validate required inputs are provided."""
+        valid = True
+        block_map = {b.get("id"): b for b in blocks_info}
+
+        for node in agent.get("nodes", []):
+            block_id = node.get("block_id")
+            block = block_map.get(block_id)
+
+            if not block:
+                continue
+
+            required_inputs = block.get("inputSchema", {}).get("required", [])
+            input_defaults = node.get("input_default", {})
+            node_id = node.get("id")
+
+            # Get linked inputs
+            linked_inputs = {
+                link["sink_name"]
+                for link in agent.get("links", [])
+                if link.get("sink_id") == node_id
+            }
+
+            for req_input in required_inputs:
+                if (
+                    req_input not in input_defaults
+                    and req_input not in linked_inputs
+                    and req_input != "credentials"
+                ):
+                    block_name = block.get("name", "Unknown Block")
+                    self.add_error(
+                        f"Node '{node_id}' ({block_name}) is missing required input '{req_input}'."
+                    )
+                    valid = False
+
+        return valid
+
+    def validate_data_type_compatibility(
+        self, agent: dict[str, Any], blocks_info: list[dict[str, Any]]
+    ) -> bool:
+        """Validate linked data types are compatible."""
+        valid = True
+        block_map = {b.get("id"): b for b in blocks_info}
+        node_lookup = {n.get("id"): n for n in agent.get("nodes", [])}
+
+        def get_type(schema: dict, name: str) -> str | None:
+            if "_#_" in name:
+                parent, child = name.split("_#_", 1)
+                parent_schema = schema.get(parent, {})
+                if "properties" in parent_schema:
+                    return parent_schema["properties"].get(child, {}).get("type")
+                return None
+            return schema.get(name, {}).get("type")
+
+        def are_compatible(src: str, sink: str) -> bool:
+            if {src, sink} <= {"integer", "number"}:
+                return True
+            return src == sink
+
+        for link in agent.get("links", []):
+            source_node = node_lookup.get(link.get("source_id"))
+            sink_node = node_lookup.get(link.get("sink_id"))
+
+            if not source_node or not sink_node:
+                continue
+
+            source_block = block_map.get(source_node.get("block_id"))
+            sink_block = block_map.get(sink_node.get("block_id"))
+
+            if not source_block or not sink_block:
+                continue
+
+            source_outputs = source_block.get("outputSchema", {}).get("properties", {})
+            sink_inputs = sink_block.get("inputSchema", {}).get("properties", {})
+
+            source_type = get_type(source_outputs, link.get("source_name", ""))
+            sink_type = get_type(sink_inputs, link.get("sink_name", ""))
+
+            if source_type and sink_type and not are_compatible(source_type, sink_type):
+                self.add_error(
+                    f"Type mismatch: {source_block.get('name')} output '{link['source_name']}' "
+                    f"({source_type}) -> {sink_block.get('name')} input '{link['sink_name']}' ({sink_type})."
+                )
+                valid = False
+
+        return valid
+
+    def validate_nested_sink_links(
+        self, agent: dict[str, Any], blocks_info: list[dict[str, Any]]
+    ) -> bool:
+        """Validate nested sink links (with _#_ notation)."""
+        valid = True
+        block_map = {b.get("id"): b for b in blocks_info}
+        node_lookup = {n.get("id"): n for n in agent.get("nodes", [])}
+
+        for link in agent.get("links", []):
+            sink_name = link.get("sink_name", "")
+
+            if "_#_" in sink_name:
+                parent, child = sink_name.split("_#_", 1)
+
+                sink_node = node_lookup.get(link.get("sink_id"))
+                if not sink_node:
+                    continue
+
+                block = block_map.get(sink_node.get("block_id"))
+                if not block:
+                    continue
+
+                input_props = block.get("inputSchema", {}).get("properties", {})
+                parent_schema = input_props.get(parent)
+
+                if not parent_schema:
+                    self.add_error(
+                        f"Invalid nested link '{sink_name}': parent '{parent}' not found."
+                    )
+                    valid = False
+                    continue
+
+                if not parent_schema.get("additionalProperties"):
+                    if not (
+                        isinstance(parent_schema, dict)
+                        and "properties" in parent_schema
+                        and child in parent_schema.get("properties", {})
+                    ):
+                        self.add_error(
+                            f"Invalid nested link '{sink_name}': child '{child}' not found in '{parent}'."
+                        )
+                        valid = False
+
+        return valid
+
+    def validate_prompt_spaces(self, agent: dict[str, Any]) -> bool:
+        """Validate prompts don't have spaces in template variables."""
+        valid = True
+
+        for node in agent.get("nodes", []):
+            input_default = node.get("input_default", {})
+            prompt = input_default.get("prompt", "")
+
+            if not isinstance(prompt, str):
+                continue
+
+            # Find {{...}} with spaces
+            matches = re.finditer(r"\{\{([^}]+)\}\}", prompt)
+            for match in matches:
+                content = match.group(1)
+                if " " in content:
+                    self.add_error(
+                        f"Node '{node.get('id')}' has spaces in template variable: "
+                        f"'{{{{{content}}}}}' should be '{{{{{content.replace(' ', '_')}}}}}'."
+                    )
+                    valid = False
+
+        return valid
+
+    def validate(
+        self, agent: dict[str, Any], blocks_info: list[dict[str, Any]] | None = None
+    ) -> tuple[bool, str | None]:
+        """Run all validations.
+
+        Returns:
+            Tuple of (is_valid, error_message)
+        """
+        self.errors = []
+
+        if blocks_info is None:
+            blocks_info = get_blocks_info()
+
+        checks = [
+            self.validate_block_existence(agent, blocks_info),
+            self.validate_link_node_references(agent),
+            self.validate_required_inputs(agent, blocks_info),
+            self.validate_data_type_compatibility(agent, blocks_info),
+            self.validate_nested_sink_links(agent, blocks_info),
+            self.validate_prompt_spaces(agent),
+        ]
+
+        all_passed = all(checks)
+
+        if all_passed:
+            logger.info("Agent validation successful")
+            return True, None
+
+        error_message = "Agent validation failed:\n"
+        for i, error in enumerate(self.errors, 1):
+            error_message += f"{i}. {error}\n"
+
+        logger.warning(f"Agent validation failed with {len(self.errors)} errors")
+        return False, error_message
+
+
+def validate_agent(
+    agent: dict[str, Any], blocks_info: list[dict[str, Any]] | None = None
+) -> tuple[bool, str | None]:
+    """Convenience function to validate an agent.
+
+    Returns:
+        Tuple of (is_valid, error_message)
+    """
+    validator = AgentValidator()
+    return validator.validate(agent, blocks_info)

autogpt_platform/backend/backend/api/features/chat/tools/create_agent.py

@@ -0,0 +1,279 @@
+"""CreateAgentTool - Creates agents from natural language descriptions."""
+
+import logging
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+
+from .agent_generator import (
+    apply_all_fixes,
+    decompose_goal,
+    generate_agent,
+    get_blocks_info,
+    save_agent_to_library,
+    validate_agent,
+)
+from .base import BaseTool
+from .models import (
+    AgentPreviewResponse,
+    AgentSavedResponse,
+    ClarificationNeededResponse,
+    ClarifyingQuestion,
+    ErrorResponse,
+    ToolResponseBase,
+)
+
+logger = logging.getLogger(__name__)
+
+# Maximum retries for agent generation with validation feedback
+MAX_GENERATION_RETRIES = 2
+
+
+class CreateAgentTool(BaseTool):
+    """Tool for creating agents from natural language descriptions."""
+
+    @property
+    def name(self) -> str:
+        return "create_agent"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Create a new agent workflow from a natural language description. "
+            "First generates a preview, then saves to library if save=true."
+        )
+
+    @property
+    def requires_auth(self) -> bool:
+        return True
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "description": {
+                    "type": "string",
+                    "description": (
+                        "Natural language description of what the agent should do. "
+                        "Be specific about inputs, outputs, and the workflow steps."
+                    ),
+                },
+                "context": {
+                    "type": "string",
+                    "description": (
+                        "Additional context or answers to previous clarifying questions. "
+                        "Include any preferences or constraints mentioned by the user."
+                    ),
+                },
+                "save": {
+                    "type": "boolean",
+                    "description": (
+                        "Whether to save the agent to the user's library. "
+                        "Default is true. Set to false for preview only."
+                    ),
+                    "default": True,
+                },
+            },
+            "required": ["description"],
+        }
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs,
+    ) -> ToolResponseBase:
+        """Execute the create_agent tool.
+
+        Flow:
+        1. Decompose the description into steps (may return clarifying questions)
+        2. Generate agent JSON from the steps
+        3. Apply fixes to correct common LLM errors
+        4. Preview or save based on the save parameter
+        """
+        description = kwargs.get("description", "").strip()
+        context = kwargs.get("context", "")
+        save = kwargs.get("save", True)
+        session_id = session.session_id if session else None
+
+        if not description:
+            return ErrorResponse(
+                message="Please provide a description of what the agent should do.",
+                error="Missing description parameter",
+                session_id=session_id,
+            )
+
+        # Step 1: Decompose goal into steps
+        try:
+            decomposition_result = await decompose_goal(description, context)
+        except ValueError as e:
+            # Handle missing API key or configuration errors
+            return ErrorResponse(
+                message=f"Agent generation is not configured: {str(e)}",
+                error="configuration_error",
+                session_id=session_id,
+            )
+
+        if decomposition_result is None:
+            return ErrorResponse(
+                message="Failed to analyze the goal. Please try rephrasing.",
+                error="Decomposition failed",
+                session_id=session_id,
+            )
+
+        # Check if LLM returned clarifying questions
+        if decomposition_result.get("type") == "clarifying_questions":
+            questions = decomposition_result.get("questions", [])
+            return ClarificationNeededResponse(
+                message=(
+                    "I need some more information to create this agent. "
+                    "Please answer the following questions:"
+                ),
+                questions=[
+                    ClarifyingQuestion(
+                        question=q.get("question", ""),
+                        keyword=q.get("keyword", ""),
+                        example=q.get("example"),
+                    )
+                    for q in questions
+                ],
+                session_id=session_id,
+            )
+
+        # Check for unachievable/vague goals
+        if decomposition_result.get("type") == "unachievable_goal":
+            suggested = decomposition_result.get("suggested_goal", "")
+            reason = decomposition_result.get("reason", "")
+            return ErrorResponse(
+                message=(
+                    f"This goal cannot be accomplished with the available blocks. "
+                    f"{reason} "
+                    f"Suggestion: {suggested}"
+                ),
+                error="unachievable_goal",
+                details={"suggested_goal": suggested, "reason": reason},
+                session_id=session_id,
+            )
+
+        if decomposition_result.get("type") == "vague_goal":
+            suggested = decomposition_result.get("suggested_goal", "")
+            return ErrorResponse(
+                message=(
+                    f"The goal is too vague to create a specific workflow. "
+                    f"Suggestion: {suggested}"
+                ),
+                error="vague_goal",
+                details={"suggested_goal": suggested},
+                session_id=session_id,
+            )
+
+        # Step 2: Generate agent JSON with retry on validation failure
+        blocks_info = get_blocks_info()
+        agent_json = None
+        validation_errors = None
+
+        for attempt in range(MAX_GENERATION_RETRIES + 1):
+            # Generate agent (include validation errors from previous attempt)
+            if attempt == 0:
+                agent_json = await generate_agent(decomposition_result)
+            else:
+                # Retry with validation error feedback
+                logger.info(
+                    f"Retry {attempt}/{MAX_GENERATION_RETRIES} with validation feedback"
+                )
+                retry_instructions = {
+                    **decomposition_result,
+                    "previous_errors": validation_errors,
+                    "retry_instructions": (
+                        "The previous generation had validation errors. "
+                        "Please fix these issues in the new generation:\n"
+                        f"{validation_errors}"
+                    ),
+                }
+                agent_json = await generate_agent(retry_instructions)
+
+            if agent_json is None:
+                if attempt == MAX_GENERATION_RETRIES:
+                    return ErrorResponse(
+                        message="Failed to generate the agent. Please try again.",
+                        error="Generation failed",
+                        session_id=session_id,
+                    )
+                continue
+
+            # Step 3: Apply fixes to correct common errors
+            agent_json = apply_all_fixes(agent_json, blocks_info)
+
+            # Step 4: Validate the agent
+            is_valid, validation_errors = validate_agent(agent_json, blocks_info)
+
+            if is_valid:
+                logger.info(f"Agent generated successfully on attempt {attempt + 1}")
+                break
+
+            logger.warning(
+                f"Validation failed on attempt {attempt + 1}: {validation_errors}"
+            )
+
+            if attempt == MAX_GENERATION_RETRIES:
+                # Return error with validation details
+                return ErrorResponse(
+                    message=(
+                        f"Generated agent has validation errors after {MAX_GENERATION_RETRIES + 1} attempts. "
+                        f"Please try rephrasing your request or simplify the workflow."
+                    ),
+                    error="validation_failed",
+                    details={"validation_errors": validation_errors},
+                    session_id=session_id,
+                )
+
+        agent_name = agent_json.get("name", "Generated Agent")
+        agent_description = agent_json.get("description", "")
+        node_count = len(agent_json.get("nodes", []))
+        link_count = len(agent_json.get("links", []))
+
+        # Step 4: Preview or save
+        if not save:
+            return AgentPreviewResponse(
+                message=(
+                    f"I've generated an agent called '{agent_name}' with {node_count} blocks. "
+                    f"Review it and call create_agent with save=true to save it to your library."
+                ),
+                agent_json=agent_json,
+                agent_name=agent_name,
+                description=agent_description,
+                node_count=node_count,
+                link_count=link_count,
+                session_id=session_id,
+            )
+
+        # Save to library
+        if not user_id:
+            return ErrorResponse(
+                message="You must be logged in to save agents.",
+                error="auth_required",
+                session_id=session_id,
+            )
+
+        try:
+            created_graph, library_agent = await save_agent_to_library(
+                agent_json, user_id
+            )
+
+            return AgentSavedResponse(
+                message=f"Agent '{created_graph.name}' has been saved to your library!",
+                agent_id=created_graph.id,
+                agent_name=created_graph.name,
+                library_agent_id=library_agent.id,
+                library_agent_link=f"/library/{library_agent.id}",
+                agent_page_link=f"/build?flowID={created_graph.id}",
+                session_id=session_id,
+            )
+        except Exception as e:
+            return ErrorResponse(
+                message=f"Failed to save the agent: {str(e)}",
+                error="save_failed",
+                details={"exception": str(e)},
+                session_id=session_id,
+            )

autogpt_platform/backend/backend/api/features/chat/tools/edit_agent.py

@@ -0,0 +1,294 @@
+"""EditAgentTool - Edits existing agents using natural language."""
+
+import logging
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+
+from .agent_generator import (
+    apply_agent_patch,
+    apply_all_fixes,
+    generate_agent_patch,
+    get_agent_as_json,
+    get_blocks_info,
+    save_agent_to_library,
+    validate_agent,
+)
+from .base import BaseTool
+from .models import (
+    AgentPreviewResponse,
+    AgentSavedResponse,
+    ClarificationNeededResponse,
+    ClarifyingQuestion,
+    ErrorResponse,
+    ToolResponseBase,
+)
+
+logger = logging.getLogger(__name__)
+
+# Maximum retries for patch generation with validation feedback
+MAX_GENERATION_RETRIES = 2
+
+
+class EditAgentTool(BaseTool):
+    """Tool for editing existing agents using natural language."""
+
+    @property
+    def name(self) -> str:
+        return "edit_agent"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Edit an existing agent from the user's library using natural language. "
+            "Generates a patch to update the agent while preserving unchanged parts."
+        )
+
+    @property
+    def requires_auth(self) -> bool:
+        return True
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "agent_id": {
+                    "type": "string",
+                    "description": (
+                        "The ID of the agent to edit. "
+                        "Can be a graph ID or library agent ID."
+                    ),
+                },
+                "changes": {
+                    "type": "string",
+                    "description": (
+                        "Natural language description of what changes to make. "
+                        "Be specific about what to add, remove, or modify."
+                    ),
+                },
+                "context": {
+                    "type": "string",
+                    "description": (
+                        "Additional context or answers to previous clarifying questions."
+                    ),
+                },
+                "save": {
+                    "type": "boolean",
+                    "description": (
+                        "Whether to save the changes. "
+                        "Default is true. Set to false for preview only."
+                    ),
+                    "default": True,
+                },
+            },
+            "required": ["agent_id", "changes"],
+        }
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs,
+    ) -> ToolResponseBase:
+        """Execute the edit_agent tool.
+
+        Flow:
+        1. Fetch the current agent
+        2. Generate a patch based on the requested changes
+        3. Apply the patch to create an updated agent
+        4. Preview or save based on the save parameter
+        """
+        agent_id = kwargs.get("agent_id", "").strip()
+        changes = kwargs.get("changes", "").strip()
+        context = kwargs.get("context", "")
+        save = kwargs.get("save", True)
+        session_id = session.session_id if session else None
+
+        if not agent_id:
+            return ErrorResponse(
+                message="Please provide the agent ID to edit.",
+                error="Missing agent_id parameter",
+                session_id=session_id,
+            )
+
+        if not changes:
+            return ErrorResponse(
+                message="Please describe what changes you want to make.",
+                error="Missing changes parameter",
+                session_id=session_id,
+            )
+
+        # Step 1: Fetch current agent
+        current_agent = await get_agent_as_json(agent_id, user_id)
+
+        if current_agent is None:
+            return ErrorResponse(
+                message=f"Could not find agent with ID '{agent_id}' in your library.",
+                error="agent_not_found",
+                session_id=session_id,
+            )
+
+        # Build the update request with context
+        update_request = changes
+        if context:
+            update_request = f"{changes}\n\nAdditional context:\n{context}"
+
+        # Step 2: Generate patch with retry on validation failure
+        blocks_info = get_blocks_info()
+        updated_agent = None
+        validation_errors = None
+        intent = "Applied requested changes"
+
+        for attempt in range(MAX_GENERATION_RETRIES + 1):
+            # Generate patch (include validation errors from previous attempt)
+            try:
+                if attempt == 0:
+                    patch_result = await generate_agent_patch(
+                        update_request, current_agent
+                    )
+                else:
+                    # Retry with validation error feedback
+                    logger.info(
+                        f"Retry {attempt}/{MAX_GENERATION_RETRIES} with validation feedback"
+                    )
+                    retry_request = (
+                        f"{update_request}\n\n"
+                        f"IMPORTANT: The previous edit had validation errors. "
+                        f"Please fix these issues:\n{validation_errors}"
+                    )
+                    patch_result = await generate_agent_patch(
+                        retry_request, current_agent
+                    )
+            except ValueError as e:
+                # Handle missing API key or configuration errors
+                return ErrorResponse(
+                    message=f"Agent generation is not configured: {str(e)}",
+                    error="configuration_error",
+                    session_id=session_id,
+                )
+
+            if patch_result is None:
+                if attempt == MAX_GENERATION_RETRIES:
+                    return ErrorResponse(
+                        message="Failed to generate changes. Please try rephrasing.",
+                        error="Patch generation failed",
+                        session_id=session_id,
+                    )
+                continue
+
+            # Check if LLM returned clarifying questions
+            if patch_result.get("type") == "clarifying_questions":
+                questions = patch_result.get("questions", [])
+                return ClarificationNeededResponse(
+                    message=(
+                        "I need some more information about the changes. "
+                        "Please answer the following questions:"
+                    ),
+                    questions=[
+                        ClarifyingQuestion(
+                            question=q.get("question", ""),
+                            keyword=q.get("keyword", ""),
+                            example=q.get("example"),
+                        )
+                        for q in questions
+                    ],
+                    session_id=session_id,
+                )
+
+            # Step 3: Apply patch and fixes
+            try:
+                updated_agent = apply_agent_patch(current_agent, patch_result)
+                updated_agent = apply_all_fixes(updated_agent, blocks_info)
+            except Exception as e:
+                if attempt == MAX_GENERATION_RETRIES:
+                    return ErrorResponse(
+                        message=f"Failed to apply changes: {str(e)}",
+                        error="patch_apply_failed",
+                        details={"exception": str(e)},
+                        session_id=session_id,
+                    )
+                validation_errors = str(e)
+                continue
+
+            # Step 4: Validate the updated agent
+            is_valid, validation_errors = validate_agent(updated_agent, blocks_info)
+
+            if is_valid:
+                logger.info(f"Agent edited successfully on attempt {attempt + 1}")
+                intent = patch_result.get("intent", "Applied requested changes")
+                break
+
+            logger.warning(
+                f"Validation failed on attempt {attempt + 1}: {validation_errors}"
+            )
+
+            if attempt == MAX_GENERATION_RETRIES:
+                # Return error with validation details
+                return ErrorResponse(
+                    message=(
+                        f"Updated agent has validation errors after "
+                        f"{MAX_GENERATION_RETRIES + 1} attempts. "
+                        f"Please try rephrasing your request or simplify the changes."
+                    ),
+                    error="validation_failed",
+                    details={"validation_errors": validation_errors},
+                    session_id=session_id,
+                )
+
+        # At this point, updated_agent is guaranteed to be set (we return on all failure paths)
+        assert updated_agent is not None
+
+        agent_name = updated_agent.get("name", "Updated Agent")
+        agent_description = updated_agent.get("description", "")
+        node_count = len(updated_agent.get("nodes", []))
+        link_count = len(updated_agent.get("links", []))
+
+        # Step 5: Preview or save
+        if not save:
+            return AgentPreviewResponse(
+                message=(
+                    f"I've updated the agent. Changes: {intent}. "
+                    f"The agent now has {node_count} blocks. "
+                    f"Review it and call edit_agent with save=true to save the changes."
+                ),
+                agent_json=updated_agent,
+                agent_name=agent_name,
+                description=agent_description,
+                node_count=node_count,
+                link_count=link_count,
+                session_id=session_id,
+            )
+
+        # Save to library (creates a new version)
+        if not user_id:
+            return ErrorResponse(
+                message="You must be logged in to save agents.",
+                error="auth_required",
+                session_id=session_id,
+            )
+
+        try:
+            created_graph, library_agent = await save_agent_to_library(
+                updated_agent, user_id, is_update=True
+            )
+
+            return AgentSavedResponse(
+                message=(
+                    f"Updated agent '{created_graph.name}' has been saved to your library! "
+                    f"Changes: {intent}"
+                ),
+                agent_id=created_graph.id,
+                agent_name=created_graph.name,
+                library_agent_id=library_agent.id,
+                library_agent_link=f"/library/{library_agent.id}",
+                agent_page_link=f"/build?flowID={created_graph.id}",
+                session_id=session_id,
+            )
+        except Exception as e:
+            return ErrorResponse(
+                message=f"Failed to save the updated agent: {str(e)}",
+                error="save_failed",
+                details={"exception": str(e)},
+                session_id=session_id,
+            )

autogpt_platform/backend/backend/api/features/chat/tools/find_block.py

@@ -0,0 +1,192 @@
+import logging
+from typing import Any
+
+from prisma.enums import ContentType
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools.base import BaseTool, ToolResponseBase
+from backend.api.features.chat.tools.models import (
+    BlockInfoSummary,
+    BlockInputFieldInfo,
+    BlockListResponse,
+    ErrorResponse,
+    NoResultsResponse,
+)
+from backend.api.features.store.hybrid_search import unified_hybrid_search
+from backend.data.block import get_block
+
+logger = logging.getLogger(__name__)
+
+
+class FindBlockTool(BaseTool):
+    """Tool for searching available blocks."""
+
+    @property
+    def name(self) -> str:
+        return "find_block"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Search for available blocks by name or description. "
+            "Blocks are reusable components that perform specific tasks like "
+            "sending emails, making API calls, processing text, etc. "
+            "IMPORTANT: Use this tool FIRST to get the block's 'id' before calling run_block. "
+            "The response includes each block's id, required_inputs, and input_schema."
+        )
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "query": {
+                    "type": "string",
+                    "description": (
+                        "Search query to find blocks by name or description. "
+                        "Use keywords like 'email', 'http', 'text', 'ai', etc."
+                    ),
+                },
+            },
+            "required": ["query"],
+        }
+
+    @property
+    def requires_auth(self) -> bool:
+        return True
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs,
+    ) -> ToolResponseBase:
+        """Search for blocks matching the query.
+
+        Args:
+            user_id: User ID (required)
+            session: Chat session
+            query: Search query
+
+        Returns:
+            BlockListResponse: List of matching blocks
+            NoResultsResponse: No blocks found
+            ErrorResponse: Error message
+        """
+        query = kwargs.get("query", "").strip()
+        session_id = session.session_id
+
+        if not query:
+            return ErrorResponse(
+                message="Please provide a search query",
+                session_id=session_id,
+            )
+
+        try:
+            # Search for blocks using hybrid search
+            results, total = await unified_hybrid_search(
+                query=query,
+                content_types=[ContentType.BLOCK],
+                page=1,
+                page_size=10,
+            )
+
+            if not results:
+                return NoResultsResponse(
+                    message=f"No blocks found for '{query}'",
+                    suggestions=[
+                        "Try broader keywords like 'email', 'http', 'text', 'ai'",
+                        "Check spelling of technical terms",
+                    ],
+                    session_id=session_id,
+                )
+
+            # Enrich results with full block information
+            blocks: list[BlockInfoSummary] = []
+            for result in results:
+                block_id = result["content_id"]
+                block = get_block(block_id)
+
+                if block:
+                    # Get input/output schemas
+                    input_schema = {}
+                    output_schema = {}
+                    try:
+                        input_schema = block.input_schema.jsonschema()
+                    except Exception:
+                        pass
+                    try:
+                        output_schema = block.output_schema.jsonschema()
+                    except Exception:
+                        pass
+
+                    # Get categories from block instance
+                    categories = []
+                    if hasattr(block, "categories") and block.categories:
+                        categories = [cat.value for cat in block.categories]
+
+                    # Extract required inputs for easier use
+                    required_inputs: list[BlockInputFieldInfo] = []
+                    if input_schema:
+                        properties = input_schema.get("properties", {})
+                        required_fields = set(input_schema.get("required", []))
+                        # Get credential field names to exclude from required inputs
+                        credentials_fields = set(
+                            block.input_schema.get_credentials_fields().keys()
+                        )
+
+                        for field_name, field_schema in properties.items():
+                            # Skip credential fields - they're handled separately
+                            if field_name in credentials_fields:
+                                continue
+
+                            required_inputs.append(
+                                BlockInputFieldInfo(
+                                    name=field_name,
+                                    type=field_schema.get("type", "string"),
+                                    description=field_schema.get("description", ""),
+                                    required=field_name in required_fields,
+                                    default=field_schema.get("default"),
+                                )
+                            )
+
+                    blocks.append(
+                        BlockInfoSummary(
+                            id=block_id,
+                            name=block.name,
+                            description=block.description or "",
+                            categories=categories,
+                            input_schema=input_schema,
+                            output_schema=output_schema,
+                            required_inputs=required_inputs,
+                        )
+                    )
+
+            if not blocks:
+                return NoResultsResponse(
+                    message=f"No blocks found for '{query}'",
+                    suggestions=[
+                        "Try broader keywords like 'email', 'http', 'text', 'ai'",
+                    ],
+                    session_id=session_id,
+                )
+
+            return BlockListResponse(
+                message=(
+                    f"Found {len(blocks)} block(s) matching '{query}'. "
+                    "To execute a block, use run_block with the block's 'id' field "
+                    "and provide 'input_data' matching the block's input_schema."
+                ),
+                blocks=blocks,
+                count=len(blocks),
+                query=query,
+                session_id=session_id,
+            )
+
+        except Exception as e:
+            logger.error(f"Error searching blocks: {e}", exc_info=True)
+            return ErrorResponse(
+                message="Failed to search blocks",
+                error=str(e),
+                session_id=session_id,
+            )

autogpt_platform/backend/backend/api/features/chat/tools/get_doc_page.py

@@ -0,0 +1,148 @@
+"""GetDocPageTool - Fetch full content of a documentation page."""
+
+import logging
+from pathlib import Path
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools.base import BaseTool
+from backend.api.features.chat.tools.models import (
+    DocPageResponse,
+    ErrorResponse,
+    ToolResponseBase,
+)
+
+logger = logging.getLogger(__name__)
+
+# Base URL for documentation (can be configured)
+DOCS_BASE_URL = "https://docs.agpt.co"
+
+
+class GetDocPageTool(BaseTool):
+    """Tool for fetching full content of a documentation page."""
+
+    @property
+    def name(self) -> str:
+        return "get_doc_page"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Get the full content of a documentation page by its path. "
+            "Use this after search_docs to read the complete content of a relevant page."
+        )
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "path": {
+                    "type": "string",
+                    "description": (
+                        "The path to the documentation file, as returned by search_docs. "
+                        "Example: 'platform/block-sdk-guide.md'"
+                    ),
+                },
+            },
+            "required": ["path"],
+        }
+
+    @property
+    def requires_auth(self) -> bool:
+        return False  # Documentation is public
+
+    def _get_docs_root(self) -> Path:
+        """Get the documentation root directory."""
+        this_file = Path(__file__)
+        project_root = this_file.parent.parent.parent.parent.parent.parent.parent.parent
+        return project_root / "docs"
+
+    def _extract_title(self, content: str, fallback: str) -> str:
+        """Extract title from markdown content."""
+        lines = content.split("\n")
+        for line in lines:
+            if line.startswith("# "):
+                return line[2:].strip()
+        return fallback
+
+    def _make_doc_url(self, path: str) -> str:
+        """Create a URL for a documentation page."""
+        url_path = path.rsplit(".", 1)[0] if "." in path else path
+        return f"{DOCS_BASE_URL}/{url_path}"
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs,
+    ) -> ToolResponseBase:
+        """Fetch full content of a documentation page.
+
+        Args:
+            user_id: User ID (not required for docs)
+            session: Chat session
+            path: Path to the documentation file
+
+        Returns:
+            DocPageResponse: Full document content
+            ErrorResponse: Error message
+        """
+        path = kwargs.get("path", "").strip()
+        session_id = session.session_id if session else None
+
+        if not path:
+            return ErrorResponse(
+                message="Please provide a documentation path.",
+                error="Missing path parameter",
+                session_id=session_id,
+            )
+
+        # Sanitize path to prevent directory traversal
+        if ".." in path or path.startswith("/"):
+            return ErrorResponse(
+                message="Invalid documentation path.",
+                error="invalid_path",
+                session_id=session_id,
+            )
+
+        docs_root = self._get_docs_root()
+        full_path = docs_root / path
+
+        if not full_path.exists():
+            return ErrorResponse(
+                message=f"Documentation page not found: {path}",
+                error="not_found",
+                session_id=session_id,
+            )
+
+        # Ensure the path is within docs root
+        try:
+            full_path.resolve().relative_to(docs_root.resolve())
+        except ValueError:
+            return ErrorResponse(
+                message="Invalid documentation path.",
+                error="invalid_path",
+                session_id=session_id,
+            )
+
+        try:
+            content = full_path.read_text(encoding="utf-8")
+            title = self._extract_title(content, path)
+
+            return DocPageResponse(
+                message=f"Retrieved documentation page: {title}",
+                title=title,
+                path=path,
+                content=content,
+                doc_url=self._make_doc_url(path),
+                session_id=session_id,
+            )
+
+        except Exception as e:
+            logger.error(f"Failed to read documentation page {path}: {e}")
+            return ErrorResponse(
+                message=f"Failed to read documentation page: {str(e)}",
+                error="read_failed",
+                session_id=session_id,
+            )

autogpt_platform/backend/backend/api/features/chat/tools/models.py

@@ -21,6 +21,13 @@ class ResponseType(str, Enum):
     NO_RESULTS = "no_results"
     AGENT_OUTPUT = "agent_output"
     UNDERSTANDING_UPDATED = "understanding_updated"
+    AGENT_PREVIEW = "agent_preview"
+    AGENT_SAVED = "agent_saved"
+    CLARIFICATION_NEEDED = "clarification_needed"
+    BLOCK_LIST = "block_list"
+    BLOCK_OUTPUT = "block_output"
+    DOC_SEARCH_RESULTS = "doc_search_results"
+    DOC_PAGE = "doc_page"
 
 
 # Base response model
@@ -209,3 +216,121 @@ class UnderstandingUpdatedResponse(ToolResponseBase):
     type: ResponseType = ResponseType.UNDERSTANDING_UPDATED
     updated_fields: list[str] = Field(default_factory=list)
     current_understanding: dict[str, Any] = Field(default_factory=dict)
+
+
+# Agent generation models
+class ClarifyingQuestion(BaseModel):
+    """A question that needs user clarification."""
+
+    question: str
+    keyword: str
+    example: str | None = None
+
+
+class AgentPreviewResponse(ToolResponseBase):
+    """Response for previewing a generated agent before saving."""
+
+    type: ResponseType = ResponseType.AGENT_PREVIEW
+    agent_json: dict[str, Any]
+    agent_name: str
+    description: str
+    node_count: int
+    link_count: int = 0
+
+
+class AgentSavedResponse(ToolResponseBase):
+    """Response when an agent is saved to the library."""
+
+    type: ResponseType = ResponseType.AGENT_SAVED
+    agent_id: str
+    agent_name: str
+    library_agent_id: str
+    library_agent_link: str
+    agent_page_link: str  # Link to the agent builder/editor page
+
+
+class ClarificationNeededResponse(ToolResponseBase):
+    """Response when the LLM needs more information from the user."""
+
+    type: ResponseType = ResponseType.CLARIFICATION_NEEDED
+    questions: list[ClarifyingQuestion] = Field(default_factory=list)
+
+
+# Documentation search models
+class DocSearchResult(BaseModel):
+    """A single documentation search result."""
+
+    title: str
+    path: str
+    section: str
+    snippet: str  # Short excerpt for UI display
+    score: float
+    doc_url: str | None = None
+
+
+class DocSearchResultsResponse(ToolResponseBase):
+    """Response for search_docs tool."""
+
+    type: ResponseType = ResponseType.DOC_SEARCH_RESULTS
+    results: list[DocSearchResult]
+    count: int
+    query: str
+
+
+class DocPageResponse(ToolResponseBase):
+    """Response for get_doc_page tool."""
+
+    type: ResponseType = ResponseType.DOC_PAGE
+    title: str
+    path: str
+    content: str  # Full document content
+    doc_url: str | None = None
+
+
+# Block models
+class BlockInputFieldInfo(BaseModel):
+    """Information about a block input field."""
+
+    name: str
+    type: str
+    description: str = ""
+    required: bool = False
+    default: Any | None = None
+
+
+class BlockInfoSummary(BaseModel):
+    """Summary of a block for search results."""
+
+    id: str
+    name: str
+    description: str
+    categories: list[str]
+    input_schema: dict[str, Any]
+    output_schema: dict[str, Any]
+    required_inputs: list[BlockInputFieldInfo] = Field(
+        default_factory=list,
+        description="List of required input fields for this block",
+    )
+
+
+class BlockListResponse(ToolResponseBase):
+    """Response for find_block tool."""
+
+    type: ResponseType = ResponseType.BLOCK_LIST
+    blocks: list[BlockInfoSummary]
+    count: int
+    query: str
+    usage_hint: str = Field(
+        default="To execute a block, call run_block with block_id set to the block's "
+        "'id' field and input_data containing the required fields from input_schema."
+    )
+
+
+class BlockOutputResponse(ToolResponseBase):
+    """Response for run_block tool."""
+
+    type: ResponseType = ResponseType.BLOCK_OUTPUT
+    block_id: str
+    block_name: str
+    outputs: dict[str, list[Any]]
+    success: bool = True

autogpt_platform/backend/backend/api/features/chat/tools/run_block.py

@@ -0,0 +1,297 @@
+"""Tool for executing blocks directly."""
+
+import logging
+from collections import defaultdict
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+from backend.data.block import get_block
+from backend.data.execution import ExecutionContext
+from backend.data.model import CredentialsMetaInput
+from backend.integrations.creds_manager import IntegrationCredentialsManager
+from backend.util.exceptions import BlockError
+
+from .base import BaseTool
+from .models import (
+    BlockOutputResponse,
+    ErrorResponse,
+    SetupInfo,
+    SetupRequirementsResponse,
+    ToolResponseBase,
+    UserReadiness,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class RunBlockTool(BaseTool):
+    """Tool for executing a block and returning its outputs."""
+
+    @property
+    def name(self) -> str:
+        return "run_block"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Execute a specific block with the provided input data. "
+            "IMPORTANT: You MUST call find_block first to get the block's 'id' - "
+            "do NOT guess or make up block IDs. "
+            "Use the 'id' from find_block results and provide input_data "
+            "matching the block's required_inputs."
+        )
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "block_id": {
+                    "type": "string",
+                    "description": (
+                        "The block's 'id' field from find_block results. "
+                        "NEVER guess this - always get it from find_block first."
+                    ),
+                },
+                "input_data": {
+                    "type": "object",
+                    "description": (
+                        "Input values for the block. Use the 'required_inputs' field "
+                        "from find_block to see what fields are needed."
+                    ),
+                },
+            },
+            "required": ["block_id", "input_data"],
+        }
+
+    @property
+    def requires_auth(self) -> bool:
+        return True
+
+    async def _check_block_credentials(
+        self,
+        user_id: str,
+        block: Any,
+    ) -> tuple[dict[str, CredentialsMetaInput], list[CredentialsMetaInput]]:
+        """
+        Check if user has required credentials for a block.
+
+        Returns:
+            tuple[matched_credentials, missing_credentials]
+        """
+        matched_credentials: dict[str, CredentialsMetaInput] = {}
+        missing_credentials: list[CredentialsMetaInput] = []
+
+        # Get credential field info from block's input schema
+        credentials_fields_info = block.input_schema.get_credentials_fields_info()
+
+        if not credentials_fields_info:
+            return matched_credentials, missing_credentials
+
+        # Get user's available credentials
+        creds_manager = IntegrationCredentialsManager()
+        available_creds = await creds_manager.store.get_all_creds(user_id)
+
+        for field_name, field_info in credentials_fields_info.items():
+            # field_info.provider is a frozenset of acceptable providers
+            # field_info.supported_types is a frozenset of acceptable types
+            matching_cred = next(
+                (
+                    cred
+                    for cred in available_creds
+                    if cred.provider in field_info.provider
+                    and cred.type in field_info.supported_types
+                ),
+                None,
+            )
+
+            if matching_cred:
+                matched_credentials[field_name] = CredentialsMetaInput(
+                    id=matching_cred.id,
+                    provider=matching_cred.provider,  # type: ignore
+                    type=matching_cred.type,
+                    title=matching_cred.title,
+                )
+            else:
+                # Create a placeholder for the missing credential
+                provider = next(iter(field_info.provider), "unknown")
+                cred_type = next(iter(field_info.supported_types), "api_key")
+                missing_credentials.append(
+                    CredentialsMetaInput(
+                        id=field_name,
+                        provider=provider,  # type: ignore
+                        type=cred_type,  # type: ignore
+                        title=field_name.replace("_", " ").title(),
+                    )
+                )
+
+        return matched_credentials, missing_credentials
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs,
+    ) -> ToolResponseBase:
+        """Execute a block with the given input data.
+
+        Args:
+            user_id: User ID (required)
+            session: Chat session
+            block_id: Block UUID to execute
+            input_data: Input values for the block
+
+        Returns:
+            BlockOutputResponse: Block execution outputs
+            SetupRequirementsResponse: Missing credentials
+            ErrorResponse: Error message
+        """
+        block_id = kwargs.get("block_id", "").strip()
+        input_data = kwargs.get("input_data", {})
+        session_id = session.session_id
+
+        if not block_id:
+            return ErrorResponse(
+                message="Please provide a block_id",
+                session_id=session_id,
+            )
+
+        if not isinstance(input_data, dict):
+            return ErrorResponse(
+                message="input_data must be an object",
+                session_id=session_id,
+            )
+
+        if not user_id:
+            return ErrorResponse(
+                message="Authentication required",
+                session_id=session_id,
+            )
+
+        # Get the block
+        block = get_block(block_id)
+        if not block:
+            return ErrorResponse(
+                message=f"Block '{block_id}' not found",
+                session_id=session_id,
+            )
+
+        logger.info(f"Executing block {block.name} ({block_id}) for user {user_id}")
+
+        # Check credentials
+        creds_manager = IntegrationCredentialsManager()
+        matched_credentials, missing_credentials = await self._check_block_credentials(
+            user_id, block
+        )
+
+        if missing_credentials:
+            # Return setup requirements response with missing credentials
+            missing_creds_dict = {c.id: c.model_dump() for c in missing_credentials}
+
+            return SetupRequirementsResponse(
+                message=(
+                    f"Block '{block.name}' requires credentials that are not configured. "
+                    "Please set up the required credentials before running this block."
+                ),
+                session_id=session_id,
+                setup_info=SetupInfo(
+                    agent_id=block_id,
+                    agent_name=block.name,
+                    user_readiness=UserReadiness(
+                        has_all_credentials=False,
+                        missing_credentials=missing_creds_dict,
+                        ready_to_run=False,
+                    ),
+                    requirements={
+                        "credentials": [c.model_dump() for c in missing_credentials],
+                        "inputs": self._get_inputs_list(block),
+                        "execution_modes": ["immediate"],
+                    },
+                ),
+                graph_id=None,
+                graph_version=None,
+            )
+
+        try:
+            # Fetch actual credentials and prepare kwargs for block execution
+            # Create execution context with defaults (blocks may require it)
+            exec_kwargs: dict[str, Any] = {
+                "user_id": user_id,
+                "execution_context": ExecutionContext(),
+            }
+
+            for field_name, cred_meta in matched_credentials.items():
+                # Inject metadata into input_data (for validation)
+                if field_name not in input_data:
+                    input_data[field_name] = cred_meta.model_dump()
+
+                # Fetch actual credentials and pass as kwargs (for execution)
+                actual_credentials = await creds_manager.get(
+                    user_id, cred_meta.id, lock=False
+                )
+                if actual_credentials:
+                    exec_kwargs[field_name] = actual_credentials
+                else:
+                    return ErrorResponse(
+                        message=f"Failed to retrieve credentials for {field_name}",
+                        session_id=session_id,
+                    )
+
+            # Execute the block and collect outputs
+            outputs: dict[str, list[Any]] = defaultdict(list)
+            async for output_name, output_data in block.execute(
+                input_data,
+                **exec_kwargs,
+            ):
+                outputs[output_name].append(output_data)
+
+            return BlockOutputResponse(
+                message=f"Block '{block.name}' executed successfully",
+                block_id=block_id,
+                block_name=block.name,
+                outputs=dict(outputs),
+                success=True,
+                session_id=session_id,
+            )
+
+        except BlockError as e:
+            logger.warning(f"Block execution failed: {e}")
+            return ErrorResponse(
+                message=f"Block execution failed: {e}",
+                error=str(e),
+                session_id=session_id,
+            )
+        except Exception as e:
+            logger.error(f"Unexpected error executing block: {e}", exc_info=True)
+            return ErrorResponse(
+                message=f"Failed to execute block: {str(e)}",
+                error=str(e),
+                session_id=session_id,
+            )
+
+    def _get_inputs_list(self, block: Any) -> list[dict[str, Any]]:
+        """Extract non-credential inputs from block schema."""
+        inputs_list = []
+        schema = block.input_schema.jsonschema()
+        properties = schema.get("properties", {})
+        required_fields = set(schema.get("required", []))
+
+        # Get credential field names to exclude
+        credentials_fields = set(block.input_schema.get_credentials_fields().keys())
+
+        for field_name, field_schema in properties.items():
+            # Skip credential fields
+            if field_name in credentials_fields:
+                continue
+
+            inputs_list.append(
+                {
+                    "name": field_name,
+                    "title": field_schema.get("title", field_name),
+                    "type": field_schema.get("type", "string"),
+                    "description": field_schema.get("description", ""),
+                    "required": field_name in required_fields,
+                }
+            )
+
+        return inputs_list

autogpt_platform/backend/backend/api/features/chat/tools/search_docs.py

@@ -0,0 +1,208 @@
+"""SearchDocsTool - Search documentation using hybrid search."""
+
+import logging
+from typing import Any
+
+from prisma.enums import ContentType
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools.base import BaseTool
+from backend.api.features.chat.tools.models import (
+    DocSearchResult,
+    DocSearchResultsResponse,
+    ErrorResponse,
+    NoResultsResponse,
+    ToolResponseBase,
+)
+from backend.api.features.store.hybrid_search import unified_hybrid_search
+
+logger = logging.getLogger(__name__)
+
+# Base URL for documentation (can be configured)
+DOCS_BASE_URL = "https://docs.agpt.co"
+
+# Maximum number of results to return
+MAX_RESULTS = 5
+
+# Snippet length for preview
+SNIPPET_LENGTH = 200
+
+
+class SearchDocsTool(BaseTool):
+    """Tool for searching AutoGPT platform documentation."""
+
+    @property
+    def name(self) -> str:
+        return "search_docs"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Search the AutoGPT platform documentation for information about "
+            "how to use the platform, build agents, configure blocks, and more. "
+            "Returns relevant documentation sections. Use get_doc_page to read full content."
+        )
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "query": {
+                    "type": "string",
+                    "description": (
+                        "Search query to find relevant documentation. "
+                        "Use natural language to describe what you're looking for."
+                    ),
+                },
+            },
+            "required": ["query"],
+        }
+
+    @property
+    def requires_auth(self) -> bool:
+        return False  # Documentation is public
+
+    def _create_snippet(self, content: str, max_length: int = SNIPPET_LENGTH) -> str:
+        """Create a short snippet from content for preview."""
+        # Remove markdown formatting for cleaner snippet
+        clean_content = content.replace("#", "").replace("*", "").replace("`", "")
+        # Remove extra whitespace
+        clean_content = " ".join(clean_content.split())
+
+        if len(clean_content) <= max_length:
+            return clean_content
+
+        # Truncate at word boundary
+        truncated = clean_content[:max_length]
+        last_space = truncated.rfind(" ")
+        if last_space > max_length // 2:
+            truncated = truncated[:last_space]
+
+        return truncated + "..."
+
+    def _make_doc_url(self, path: str) -> str:
+        """Create a URL for a documentation page."""
+        # Remove file extension for URL
+        url_path = path.rsplit(".", 1)[0] if "." in path else path
+        return f"{DOCS_BASE_URL}/{url_path}"
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs,
+    ) -> ToolResponseBase:
+        """Search documentation and return relevant sections.
+
+        Args:
+            user_id: User ID (not required for docs)
+            session: Chat session
+            query: Search query
+
+        Returns:
+            DocSearchResultsResponse: List of matching documentation sections
+            NoResultsResponse: No results found
+            ErrorResponse: Error message
+        """
+        query = kwargs.get("query", "").strip()
+        session_id = session.session_id if session else None
+
+        if not query:
+            return ErrorResponse(
+                message="Please provide a search query.",
+                error="Missing query parameter",
+                session_id=session_id,
+            )
+
+        try:
+            # Search using hybrid search for DOCUMENTATION content type only
+            results, total = await unified_hybrid_search(
+                query=query,
+                content_types=[ContentType.DOCUMENTATION],
+                page=1,
+                page_size=MAX_RESULTS * 2,  # Fetch extra for deduplication
+                min_score=0.1,  # Lower threshold for docs
+            )
+
+            if not results:
+                return NoResultsResponse(
+                    message=f"No documentation found for '{query}'.",
+                    suggestions=[
+                        "Try different keywords",
+                        "Use more general terms",
+                        "Check for typos in your query",
+                    ],
+                    session_id=session_id,
+                )
+
+            # Deduplicate by document path (keep highest scoring section per doc)
+            seen_docs: dict[str, dict[str, Any]] = {}
+            for result in results:
+                metadata = result.get("metadata", {})
+                doc_path = metadata.get("path", "")
+
+                if not doc_path:
+                    continue
+
+                # Keep the highest scoring result for each document
+                if doc_path not in seen_docs:
+                    seen_docs[doc_path] = result
+                elif result.get("combined_score", 0) > seen_docs[doc_path].get(
+                    "combined_score", 0
+                ):
+                    seen_docs[doc_path] = result
+
+            # Sort by score and take top MAX_RESULTS
+            deduplicated = sorted(
+                seen_docs.values(),
+                key=lambda x: x.get("combined_score", 0),
+                reverse=True,
+            )[:MAX_RESULTS]
+
+            if not deduplicated:
+                return NoResultsResponse(
+                    message=f"No documentation found for '{query}'.",
+                    suggestions=[
+                        "Try different keywords",
+                        "Use more general terms",
+                    ],
+                    session_id=session_id,
+                )
+
+            # Build response
+            doc_results: list[DocSearchResult] = []
+            for result in deduplicated:
+                metadata = result.get("metadata", {})
+                doc_path = metadata.get("path", "")
+                doc_title = metadata.get("doc_title", "")
+                section_title = metadata.get("section_title", "")
+                searchable_text = result.get("searchable_text", "")
+                score = result.get("combined_score", 0)
+
+                doc_results.append(
+                    DocSearchResult(
+                        title=doc_title or section_title or doc_path,
+                        path=doc_path,
+                        section=section_title,
+                        snippet=self._create_snippet(searchable_text),
+                        score=round(score, 3),
+                        doc_url=self._make_doc_url(doc_path),
+                    )
+                )
+
+            return DocSearchResultsResponse(
+                message=f"Found {len(doc_results)} relevant documentation sections.",
+                results=doc_results,
+                count=len(doc_results),
+                query=query,
+                session_id=session_id,
+            )
+
+        except Exception as e:
+            logger.error(f"Documentation search failed: {e}")
+            return ErrorResponse(
+                message=f"Failed to search documentation: {str(e)}",
+                error="search_failed",
+                session_id=session_id,
+            )

autogpt_platform/backend/backend/api/features/store/content_handlers.py

@@ -275,8 +275,22 @@ class BlockHandler(ContentHandler):
         }
 
 
+@dataclass
+class MarkdownSection:
+    """Represents a section of a markdown document."""
+
+    title: str  # Section heading text
+    content: str  # Section content (including the heading line)
+    level: int  # Heading level (1 for #, 2 for ##, etc.)
+    index: int  # Section index within the document
+
+
 class DocumentationHandler(ContentHandler):
-    """Handler for documentation files (.md/.mdx)."""
+    """Handler for documentation files (.md/.mdx).
+
+    Chunks documents by markdown headings to create multiple embeddings per file.
+    Each section (## heading) becomes a separate embedding for better retrieval.
+    """
 
     @property
     def content_type(self) -> ContentType:
@@ -297,35 +311,162 @@ class DocumentationHandler(ContentHandler):
         docs_root = project_root / "docs"
         return docs_root
 
-    def _extract_title_and_content(self, file_path: Path) -> tuple[str, str]:
-        """Extract title and content from markdown file."""
+    def _extract_doc_title(self, file_path: Path) -> str:
+        """Extract the document title from a markdown file."""
         try:
             content = file_path.read_text(encoding="utf-8")
+            lines = content.split("\n")
 
             # Try to extract title from first # heading
-            lines = content.split("\n")
-            title = ""
-            body_lines = []
-
             for line in lines:
-                if line.startswith("# ") and not title:
-                    title = line[2:].strip()
-                else:
-                    body_lines.append(line)
+                if line.startswith("# "):
+                    return line[2:].strip()
 
             # If no title found, use filename
-            if not title:
-                title = file_path.stem.replace("-", " ").replace("_", " ").title()
+            return file_path.stem.replace("-", " ").replace("_", " ").title()
+        except Exception as e:
+            logger.warning(f"Failed to read title from {file_path}: {e}")
+            return file_path.stem.replace("-", " ").replace("_", " ").title()
 
-            body = "\n".join(body_lines)
+    def _chunk_markdown_by_headings(
+        self, file_path: Path, min_heading_level: int = 2
+    ) -> list[MarkdownSection]:
+        """
+        Split a markdown file into sections based on headings.
 
-            return title, body
+        Args:
+            file_path: Path to the markdown file
+            min_heading_level: Minimum heading level to split on (default: 2 for ##)
+
+        Returns:
+            List of MarkdownSection objects, one per section.
+            If no headings found, returns a single section with all content.
+        """
+        try:
+            content = file_path.read_text(encoding="utf-8")
         except Exception as e:
             logger.warning(f"Failed to read {file_path}: {e}")
-            return file_path.stem, ""
+            return []
+
+        lines = content.split("\n")
+        sections: list[MarkdownSection] = []
+        current_section_lines: list[str] = []
+        current_title = ""
+        current_level = 0
+        section_index = 0
+        doc_title = ""
+
+        for line in lines:
+            # Check if line is a heading
+            if line.startswith("#"):
+                # Count heading level
+                level = 0
+                for char in line:
+                    if char == "#":
+                        level += 1
+                    else:
+                        break
+
+                heading_text = line[level:].strip()
+
+                # Track document title (level 1 heading)
+                if level == 1 and not doc_title:
+                    doc_title = heading_text
+                    # Don't create a section for just the title - add it to first section
+                    current_section_lines.append(line)
+                    continue
+
+                # Check if this heading should start a new section
+                if level >= min_heading_level:
+                    # Save previous section if it has content
+                    if current_section_lines:
+                        section_content = "\n".join(current_section_lines).strip()
+                        if section_content:
+                            # Use doc title for first section if no specific title
+                            title = current_title if current_title else doc_title
+                            if not title:
+                                title = file_path.stem.replace("-", " ").replace(
+                                    "_", " "
+                                )
+                            sections.append(
+                                MarkdownSection(
+                                    title=title,
+                                    content=section_content,
+                                    level=current_level if current_level else 1,
+                                    index=section_index,
+                                )
+                            )
+                            section_index += 1
+
+                    # Start new section
+                    current_section_lines = [line]
+                    current_title = heading_text
+                    current_level = level
+                else:
+                    # Lower level heading (e.g., # when splitting on ##)
+                    current_section_lines.append(line)
+            else:
+                current_section_lines.append(line)
+
+        # Don't forget the last section
+        if current_section_lines:
+            section_content = "\n".join(current_section_lines).strip()
+            if section_content:
+                title = current_title if current_title else doc_title
+                if not title:
+                    title = file_path.stem.replace("-", " ").replace("_", " ")
+                sections.append(
+                    MarkdownSection(
+                        title=title,
+                        content=section_content,
+                        level=current_level if current_level else 1,
+                        index=section_index,
+                    )
+                )
+
+        # If no sections were created (no headings found), create one section with all content
+        if not sections and content.strip():
+            title = (
+                doc_title
+                if doc_title
+                else file_path.stem.replace("-", " ").replace("_", " ")
+            )
+            sections.append(
+                MarkdownSection(
+                    title=title,
+                    content=content.strip(),
+                    level=1,
+                    index=0,
+                )
+            )
+
+        return sections
+
+    def _make_section_content_id(self, doc_path: str, section_index: int) -> str:
+        """Create a unique content ID for a document section.
+
+        Format: doc_path::section_index
+        Example: 'platform/getting-started.md::0'
+        """
+        return f"{doc_path}::{section_index}"
+
+    def _parse_section_content_id(self, content_id: str) -> tuple[str, int]:
+        """Parse a section content ID back into doc_path and section_index.
+
+        Returns: (doc_path, section_index)
+        """
+        if "::" in content_id:
+            parts = content_id.rsplit("::", 1)
+            return parts[0], int(parts[1])
+        # Legacy format (whole document)
+        return content_id, 0
 
     async def get_missing_items(self, batch_size: int) -> list[ContentItem]:
-        """Fetch documentation files without embeddings."""
+        """Fetch documentation sections without embeddings.
+
+        Chunks each document by markdown headings and creates embeddings for each section.
+        Content IDs use the format: 'path/to/doc.md::section_index'
+        """
         docs_root = self._get_docs_root()
 
         if not docs_root.exists():
@@ -335,14 +476,28 @@ class DocumentationHandler(ContentHandler):
         # Find all .md and .mdx files
         all_docs = list(docs_root.rglob("*.md")) + list(docs_root.rglob("*.mdx"))
 
-        # Get relative paths for content IDs
-        doc_paths = [str(doc.relative_to(docs_root)) for doc in all_docs]
-
-        if not doc_paths:
+        if not all_docs:
             return []
 
+        # Build list of all sections from all documents
+        all_sections: list[tuple[str, Path, MarkdownSection]] = []
+        for doc_file in all_docs:
+            doc_path = str(doc_file.relative_to(docs_root))
+            sections = self._chunk_markdown_by_headings(doc_file)
+            for section in sections:
+                all_sections.append((doc_path, doc_file, section))
+
+        if not all_sections:
+            return []
+
+        # Generate content IDs for all sections
+        section_content_ids = [
+            self._make_section_content_id(doc_path, section.index)
+            for doc_path, _, section in all_sections
+        ]
+
         # Check which ones have embeddings
-        placeholders = ",".join([f"${i+1}" for i in range(len(doc_paths))])
+        placeholders = ",".join([f"${i+1}" for i in range(len(section_content_ids))])
         existing_result = await query_raw_with_schema(
             f"""
             SELECT "contentId"
@@ -350,76 +505,100 @@ class DocumentationHandler(ContentHandler):
             WHERE "contentType" = 'DOCUMENTATION'::{{schema_prefix}}"ContentType"
             AND "contentId" = ANY(ARRAY[{placeholders}])
             """,
-            *doc_paths,
+            *section_content_ids,
         )
 
         existing_ids = {row["contentId"] for row in existing_result}
-        missing_docs = [
-            (doc_path, doc_file)
-            for doc_path, doc_file in zip(doc_paths, all_docs)
-            if doc_path not in existing_ids
+
+        # Filter to missing sections
+        missing_sections = [
+            (doc_path, doc_file, section, content_id)
+            for (doc_path, doc_file, section), content_id in zip(
+                all_sections, section_content_ids
+            )
+            if content_id not in existing_ids
         ]
 
-        # Convert to ContentItem
+        # Convert to ContentItem (up to batch_size)
         items = []
-        for doc_path, doc_file in missing_docs[:batch_size]:
+        for doc_path, doc_file, section, content_id in missing_sections[:batch_size]:
             try:
-                title, content = self._extract_title_and_content(doc_file)
+                # Get document title for context
+                doc_title = self._extract_doc_title(doc_file)
 
-                # Build searchable text
-                searchable_text = f"{title} {content}"
+                # Build searchable text with context
+                # Include doc title and section title for better search relevance
+                searchable_text = f"{doc_title} - {section.title}\n\n{section.content}"
 
                 items.append(
                     ContentItem(
-                        content_id=doc_path,
+                        content_id=content_id,
                         content_type=ContentType.DOCUMENTATION,
                         searchable_text=searchable_text,
                         metadata={
-                            "title": title,
+                            "doc_title": doc_title,
+                            "section_title": section.title,
+                            "section_index": section.index,
+                            "heading_level": section.level,
                             "path": doc_path,
                         },
                         user_id=None,  # Documentation is public
                     )
                 )
             except Exception as e:
-                logger.warning(f"Failed to process doc {doc_path}: {e}")
+                logger.warning(f"Failed to process section {content_id}: {e}")
                 continue
 
         return items
 
+    def _get_all_section_content_ids(self, docs_root: Path) -> set[str]:
+        """Get all current section content IDs from the docs directory.
+
+        Used for stats and cleanup to know what sections should exist.
+        """
+        all_docs = list(docs_root.rglob("*.md")) + list(docs_root.rglob("*.mdx"))
+        content_ids = set()
+
+        for doc_file in all_docs:
+            doc_path = str(doc_file.relative_to(docs_root))
+            sections = self._chunk_markdown_by_headings(doc_file)
+            for section in sections:
+                content_ids.add(self._make_section_content_id(doc_path, section.index))
+
+        return content_ids
+
     async def get_stats(self) -> dict[str, int]:
-        """Get statistics about documentation embedding coverage."""
+        """Get statistics about documentation embedding coverage.
+
+        Counts sections (not documents) since each section gets its own embedding.
+        """
         docs_root = self._get_docs_root()
 
         if not docs_root.exists():
             return {"total": 0, "with_embeddings": 0, "without_embeddings": 0}
 
-        # Count all .md and .mdx files
-        all_docs = list(docs_root.rglob("*.md")) + list(docs_root.rglob("*.mdx"))
-        total_docs = len(all_docs)
+        # Get all section content IDs
+        all_section_ids = self._get_all_section_content_ids(docs_root)
+        total_sections = len(all_section_ids)
 
-        if total_docs == 0:
+        if total_sections == 0:
             return {"total": 0, "with_embeddings": 0, "without_embeddings": 0}
 
-        doc_paths = [str(doc.relative_to(docs_root)) for doc in all_docs]
-        placeholders = ",".join([f"${i+1}" for i in range(len(doc_paths))])
-
+        # Count embeddings in database for DOCUMENTATION type
         embedded_result = await query_raw_with_schema(
-            f"""
+            """
             SELECT COUNT(*) as count
-            FROM {{schema_prefix}}"UnifiedContentEmbedding"
-            WHERE "contentType" = 'DOCUMENTATION'::{{schema_prefix}}"ContentType"
-            AND "contentId" = ANY(ARRAY[{placeholders}])
-            """,
-            *doc_paths,
+            FROM {schema_prefix}"UnifiedContentEmbedding"
+            WHERE "contentType" = 'DOCUMENTATION'::{schema_prefix}"ContentType"
+            """
         )
 
         with_embeddings = embedded_result[0]["count"] if embedded_result else 0
 
         return {
-            "total": total_docs,
+            "total": total_sections,
             "with_embeddings": with_embeddings,
-            "without_embeddings": total_docs - with_embeddings,
+            "without_embeddings": total_sections - with_embeddings,
         }
 
 

autogpt_platform/backend/backend/api/features/store/content_handlers_test.py

@@ -164,20 +164,20 @@ async def test_documentation_handler_get_missing_items(tmp_path, mocker):
 
             assert len(items) == 2
 
-            # Check guide.md
+            # Check guide.md (content_id format: doc_path::section_index)
             guide_item = next(
-                (item for item in items if item.content_id == "guide.md"), None
+                (item for item in items if item.content_id == "guide.md::0"), None
             )
             assert guide_item is not None
             assert guide_item.content_type == ContentType.DOCUMENTATION
             assert "Getting Started" in guide_item.searchable_text
             assert "This is a guide" in guide_item.searchable_text
-            assert guide_item.metadata["title"] == "Getting Started"
+            assert guide_item.metadata["doc_title"] == "Getting Started"
             assert guide_item.user_id is None
 
-            # Check api.mdx
+            # Check api.mdx (content_id format: doc_path::section_index)
             api_item = next(
-                (item for item in items if item.content_id == "api.mdx"), None
+                (item for item in items if item.content_id == "api.mdx::0"), None
             )
             assert api_item is not None
             assert "API Reference" in api_item.searchable_text
@@ -218,17 +218,74 @@ async def test_documentation_handler_title_extraction(tmp_path):
     # Test with heading
     doc_with_heading = tmp_path / "with_heading.md"
     doc_with_heading.write_text("# My Title\n\nContent here")
-    title, content = handler._extract_title_and_content(doc_with_heading)
+    title = handler._extract_doc_title(doc_with_heading)
     assert title == "My Title"
-    assert "# My Title" not in content
-    assert "Content here" in content
 
     # Test without heading
     doc_without_heading = tmp_path / "no-heading.md"
     doc_without_heading.write_text("Just content, no heading")
-    title, content = handler._extract_title_and_content(doc_without_heading)
+    title = handler._extract_doc_title(doc_without_heading)
     assert title == "No Heading"  # Uses filename
-    assert "Just content" in content
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_documentation_handler_markdown_chunking(tmp_path):
+    """Test DocumentationHandler chunks markdown by headings."""
+    handler = DocumentationHandler()
+
+    # Test document with multiple sections
+    doc_with_sections = tmp_path / "sections.md"
+    doc_with_sections.write_text(
+        "# Document Title\n\n"
+        "Intro paragraph.\n\n"
+        "## Section One\n\n"
+        "Content for section one.\n\n"
+        "## Section Two\n\n"
+        "Content for section two.\n"
+    )
+    sections = handler._chunk_markdown_by_headings(doc_with_sections)
+
+    # Should have 3 sections: intro (with doc title), section one, section two
+    assert len(sections) == 3
+    assert sections[0].title == "Document Title"
+    assert sections[0].index == 0
+    assert "Intro paragraph" in sections[0].content
+
+    assert sections[1].title == "Section One"
+    assert sections[1].index == 1
+    assert "Content for section one" in sections[1].content
+
+    assert sections[2].title == "Section Two"
+    assert sections[2].index == 2
+    assert "Content for section two" in sections[2].content
+
+    # Test document without headings
+    doc_no_sections = tmp_path / "no-sections.md"
+    doc_no_sections.write_text("Just plain content without any headings.")
+    sections = handler._chunk_markdown_by_headings(doc_no_sections)
+    assert len(sections) == 1
+    assert sections[0].index == 0
+    assert "Just plain content" in sections[0].content
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_documentation_handler_section_content_ids():
+    """Test DocumentationHandler creates and parses section content IDs."""
+    handler = DocumentationHandler()
+
+    # Test making content ID
+    content_id = handler._make_section_content_id("docs/guide.md", 2)
+    assert content_id == "docs/guide.md::2"
+
+    # Test parsing content ID
+    doc_path, section_index = handler._parse_section_content_id("docs/guide.md::2")
+    assert doc_path == "docs/guide.md"
+    assert section_index == 2
+
+    # Test parsing legacy format (no section index)
+    doc_path, section_index = handler._parse_section_content_id("docs/old-format.md")
+    assert doc_path == "docs/old-format.md"
+    assert section_index == 0
 
 
 @pytest.mark.asyncio(loop_scope="session")

autogpt_platform/backend/backend/api/features/store/embeddings.py

@@ -683,20 +683,20 @@ async def cleanup_orphaned_embeddings() -> dict[str, Any]:
 
                 current_ids = set(get_blocks().keys())
             elif content_type == ContentType.DOCUMENTATION:
-                from pathlib import Path
-
-                # embeddings.py is at: backend/backend/api/features/store/embeddings.py
-                # Need to go up to project root then into docs/
-                this_file = Path(__file__)
-                project_root = (
-                    this_file.parent.parent.parent.parent.parent.parent.parent
+                # Use DocumentationHandler to get section-based content IDs
+                from backend.api.features.store.content_handlers import (
+                    DocumentationHandler,
                 )
-                docs_root = project_root / "docs"
-                if docs_root.exists():
-                    all_docs = list(docs_root.rglob("*.md")) + list(
-                        docs_root.rglob("*.mdx")
-                    )
-                    current_ids = {str(doc.relative_to(docs_root)) for doc in all_docs}
+
+                doc_handler = CONTENT_HANDLERS.get(ContentType.DOCUMENTATION)
+                if isinstance(doc_handler, DocumentationHandler):
+                    docs_root = doc_handler._get_docs_root()
+                    if docs_root.exists():
+                        current_ids = doc_handler._get_all_section_content_ids(
+                            docs_root
+                        )
+                    else:
+                        current_ids = set()
                 else:
                     current_ids = set()
             else:

autogpt_platform/backend/backend/api/features/store/hybrid_search.py

@@ -3,13 +3,16 @@ Unified Hybrid Search
 
 Combines semantic (embedding) search with lexical (tsvector) search
 for improved relevance across all content types (agents, blocks, docs).
+Includes BM25 reranking for improved lexical relevance.
 """
 
 import logging
+import re
 from dataclasses import dataclass
 from typing import Any, Literal
 
 from prisma.enums import ContentType
+from rank_bm25 import BM25Okapi
 
 from backend.api.features.store.embeddings import (
     EMBEDDING_DIM,
@@ -21,6 +24,84 @@ from backend.data.db import query_raw_with_schema
 logger = logging.getLogger(__name__)
 
 
+# ============================================================================
+# BM25 Reranking
+# ============================================================================
+
+
+def tokenize(text: str) -> list[str]:
+    """Simple tokenizer for BM25 - lowercase and split on non-alphanumeric."""
+    if not text:
+        return []
+    # Lowercase and split on non-alphanumeric characters
+    tokens = re.findall(r"\b\w+\b", text.lower())
+    return tokens
+
+
+def bm25_rerank(
+    query: str,
+    results: list[dict[str, Any]],
+    text_field: str = "searchable_text",
+    bm25_weight: float = 0.3,
+    original_score_field: str = "combined_score",
+) -> list[dict[str, Any]]:
+    """
+    Rerank search results using BM25.
+
+    Combines the original combined_score with BM25 score for improved
+    lexical relevance, especially for exact term matches.
+
+    Args:
+        query: The search query
+        results: List of result dicts with text_field and original_score_field
+        text_field: Field name containing the text to score
+        bm25_weight: Weight for BM25 score (0-1). Original score gets (1 - bm25_weight)
+        original_score_field: Field name containing the original score
+
+    Returns:
+        Results list sorted by combined score (BM25 + original)
+    """
+    if not results or not query:
+        return results
+
+    # Extract texts and tokenize
+    corpus = [tokenize(r.get(text_field, "") or "") for r in results]
+
+    # Handle edge case where all documents are empty
+    if all(len(doc) == 0 for doc in corpus):
+        return results
+
+    # Build BM25 index
+    bm25 = BM25Okapi(corpus)
+
+    # Score query against corpus
+    query_tokens = tokenize(query)
+    if not query_tokens:
+        return results
+
+    bm25_scores = bm25.get_scores(query_tokens)
+
+    # Normalize BM25 scores to 0-1 range
+    max_bm25 = max(bm25_scores) if max(bm25_scores) > 0 else 1.0
+    normalized_bm25 = [s / max_bm25 for s in bm25_scores]
+
+    # Combine scores
+    original_weight = 1.0 - bm25_weight
+    for i, result in enumerate(results):
+        original_score = result.get(original_score_field, 0) or 0
+        result["bm25_score"] = normalized_bm25[i]
+        final_score = (
+            original_weight * original_score + bm25_weight * normalized_bm25[i]
+        )
+        result["final_score"] = final_score
+        result["relevance"] = final_score
+
+    # Sort by relevance descending
+    results.sort(key=lambda x: x.get("relevance", 0), reverse=True)
+
+    return results
+
+
 @dataclass
 class UnifiedSearchWeights:
     """Weights for unified search (no popularity signal)."""
@@ -273,9 +354,7 @@ async def unified_hybrid_search(
             FROM normalized
         ),
         filtered AS (
-            SELECT
-                *,
-                COUNT(*) OVER () as total_count
+            SELECT *, COUNT(*) OVER () as total_count
             FROM scored
             WHERE combined_score >= {min_score_param}
         )
@@ -289,6 +368,15 @@ async def unified_hybrid_search(
     )
 
     total = results[0]["total_count"] if results else 0
+    # Apply BM25 reranking
+    if results:
+        results = bm25_rerank(
+            query=query,
+            results=results,
+            text_field="searchable_text",
+            bm25_weight=0.3,
+            original_score_field="combined_score",
+        )
 
     # Clean up results
     for result in results:
@@ -516,6 +604,8 @@ async def hybrid_search(
                 sa.featured,
                 sa.is_available,
                 sa.updated_at,
+                -- Searchable text for BM25 reranking
+                COALESCE(sa.agent_name, '') || ' ' || COALESCE(sa.sub_heading, '') || ' ' || COALESCE(sa.description, '') as searchable_text,
                 -- Semantic score
                 COALESCE(1 - (uce.embedding <=> {embedding_param}::vector), 0) as semantic_score,
                 -- Lexical score (raw, will normalize)
@@ -573,6 +663,7 @@ async def hybrid_search(
                 featured,
                 is_available,
                 updated_at,
+                searchable_text,
                 semantic_score,
                 lexical_score,
                 category_score,
@@ -603,8 +694,19 @@ async def hybrid_search(
 
     total = results[0]["total_count"] if results else 0
 
+    # Apply BM25 reranking
+    if results:
+        results = bm25_rerank(
+            query=query,
+            results=results,
+            text_field="searchable_text",
+            bm25_weight=0.3,
+            original_score_field="combined_score",
+        )
+
     for result in results:
         result.pop("total_count", None)
+        result.pop("searchable_text", None)
 
     logger.info(f"Hybrid search (store agents): {len(results)} results, {total} total")
 

autogpt_platform/backend/backend/api/features/store/hybrid_search_test.py

@@ -311,11 +311,43 @@ async def test_hybrid_search_min_score_filtering():
 @pytest.mark.asyncio(loop_scope="session")
 @pytest.mark.integration
 async def test_hybrid_search_pagination():
-    """Test hybrid search pagination."""
+    """Test hybrid search pagination.
+
+    Pagination happens in SQL (LIMIT/OFFSET), then BM25 reranking is applied
+    to the paginated results.
+    """
+    # Create mock results that SQL would return for a page
+    mock_results = [
+        {
+            "slug": f"agent-{i}",
+            "agent_name": f"Agent {i}",
+            "agent_image": "test.png",
+            "creator_username": "test",
+            "creator_avatar": "avatar.png",
+            "sub_heading": "Test",
+            "description": "Test description",
+            "runs": 100 - i,
+            "rating": 4.5,
+            "categories": ["test"],
+            "featured": False,
+            "is_available": True,
+            "updated_at": "2024-01-01T00:00:00Z",
+            "searchable_text": f"Agent {i} test description",
+            "combined_score": 0.9 - (i * 0.01),
+            "semantic_score": 0.7,
+            "lexical_score": 0.6,
+            "category_score": 0.5,
+            "recency_score": 0.4,
+            "popularity_score": 0.3,
+            "total_count": 25,
+        }
+        for i in range(10)  # SQL returns page_size results
+    ]
+
     with patch(
         "backend.api.features.store.hybrid_search.query_raw_with_schema"
     ) as mock_query:
-        mock_query.return_value = []
+        mock_query.return_value = mock_results
 
         with patch(
             "backend.api.features.store.hybrid_search.embed_query"
@@ -329,16 +361,18 @@ async def test_hybrid_search_pagination():
                 page_size=10,
             )
 
-            # Verify pagination parameters
+            # Verify results returned
+            assert len(results) == 10
+            assert total == 25  # Total from SQL COUNT(*) OVER()
+
+            # Verify the SQL query uses page_size and offset
             call_args = mock_query.call_args
             params = call_args[0]
-
-            # Last two params should be LIMIT and OFFSET
-            limit = params[-2]
-            offset = params[-1]
-
-            assert limit == 10  # page_size
-            assert offset == 10  # (page - 1) * page_size = (2 - 1) * 10
+            # Last two params are page_size and offset
+            page_size_param = params[-2]
+            offset_param = params[-1]
+            assert page_size_param == 10
+            assert offset_param == 10  # (page 2 - 1) * 10
 
 
 @pytest.mark.asyncio(loop_scope="session")
@@ -609,14 +643,36 @@ async def test_unified_hybrid_search_empty_query():
 @pytest.mark.asyncio(loop_scope="session")
 @pytest.mark.integration
 async def test_unified_hybrid_search_pagination():
-    """Test unified search pagination."""
+    """Test unified search pagination with BM25 reranking.
+
+    Pagination happens in SQL (LIMIT/OFFSET), then BM25 reranking is applied
+    to the paginated results.
+    """
+    # Create mock results that SQL would return for a page
+    mock_results = [
+        {
+            "content_type": "STORE_AGENT",
+            "content_id": f"agent-{i}",
+            "searchable_text": f"Agent {i} description",
+            "metadata": {"name": f"Agent {i}"},
+            "updated_at": "2025-01-01T00:00:00Z",
+            "semantic_score": 0.7,
+            "lexical_score": 0.8 - (i * 0.01),
+            "category_score": 0.5,
+            "recency_score": 0.3,
+            "combined_score": 0.6 - (i * 0.01),
+            "total_count": 50,
+        }
+        for i in range(15)  # SQL returns page_size results
+    ]
+
     with patch(
         "backend.api.features.store.hybrid_search.query_raw_with_schema"
     ) as mock_query:
         with patch(
             "backend.api.features.store.hybrid_search.embed_query"
         ) as mock_embed:
-            mock_query.return_value = []
+            mock_query.return_value = mock_results
             mock_embed.return_value = [0.1] * embeddings.EMBEDDING_DIM
 
             results, total = await unified_hybrid_search(
@@ -625,15 +681,18 @@ async def test_unified_hybrid_search_pagination():
                 page_size=15,
             )
 
-            # Verify pagination parameters (last two params are LIMIT and OFFSET)
+            # Verify results returned
+            assert len(results) == 15
+            assert total == 50  # Total from SQL COUNT(*) OVER()
+
+            # Verify the SQL query uses page_size and offset
             call_args = mock_query.call_args
             params = call_args[0]
-
-            limit = params[-2]
-            offset = params[-1]
-
-            assert limit == 15  # page_size
-            assert offset == 30  # (page - 1) * page_size = (3 - 1) * 15
+            # Last two params are page_size and offset
+            page_size_param = params[-2]
+            offset_param = params[-1]
+            assert page_size_param == 15
+            assert offset_param == 30  # (page 3 - 1) * 15
 
 
 @pytest.mark.asyncio(loop_scope="session")

autogpt_platform/backend/backend/api/features/store/routes.py

@@ -393,7 +393,6 @@ async def get_creators(
 @router.get(
     "/creator/{username}",
     summary="Get creator details",
-    operation_id="getV2GetCreatorDetails",
     tags=["store", "public"],
     response_model=store_model.CreatorDetails,
 )

autogpt_platform/backend/backend/api/rest_api.py

@@ -18,7 +18,6 @@ from prisma.errors import PrismaError
 
 import backend.api.features.admin.credit_admin_routes
 import backend.api.features.admin.execution_analytics_routes
-import backend.api.features.admin.llm_routes
 import backend.api.features.admin.store_admin_routes
 import backend.api.features.builder
 import backend.api.features.builder.routes
@@ -38,11 +37,9 @@ import backend.data.db
 import backend.data.graph
 import backend.data.user
 import backend.integrations.webhooks.utils
-import backend.server.v2.llm.routes as public_llm_routes
 import backend.util.service
 import backend.util.settings
-from backend.data import llm_registry
-from backend.data.block_cost_config import refresh_llm_costs
+from backend.blocks.llm import DEFAULT_LLM_MODEL
 from backend.data.model import Credentials
 from backend.integrations.providers import ProviderName
 from backend.monitoring.instrumentation import instrument_fastapi
@@ -112,27 +109,11 @@ async def lifespan_context(app: fastapi.FastAPI):
 
     AutoRegistry.patch_integrations()
 
-    # Refresh LLM registry before initializing blocks so blocks can use registry data
-    await llm_registry.refresh_llm_registry()
-    refresh_llm_costs()
-
-    # Clear block schema caches so they're regenerated with updated discriminator_mapping
-    from backend.data.block import BlockSchema
-
-    BlockSchema.clear_all_schema_caches()
-
     await backend.data.block.initialize_blocks()
 
     await backend.data.user.migrate_and_encrypt_user_integrations()
     await backend.data.graph.fix_llm_provider_credentials()
-    # migrate_llm_models uses registry default model
-    from backend.blocks.llm import LlmModel
-
-    default_model_slug = llm_registry.get_default_model_slug()
-    if default_model_slug:
-        await backend.data.graph.migrate_llm_models(LlmModel(default_model_slug))
-    else:
-        logger.warning("Skipping LLM model migration: no default model available")
+    await backend.data.graph.migrate_llm_models(DEFAULT_LLM_MODEL)
     await backend.integrations.webhooks.utils.migrate_legacy_triggered_graphs()
 
     with launch_darkly_context():
@@ -317,16 +298,6 @@ app.include_router(
     tags=["v2", "executions", "review"],
     prefix="/api/review",
 )
-app.include_router(
-    backend.api.features.admin.llm_routes.router,
-    tags=["v2", "admin", "llm"],
-    prefix="/api/llm/admin",
-)
-app.include_router(
-    public_llm_routes.router,
-    tags=["v2", "llm"],
-    prefix="/api",
-)
 app.include_router(
     backend.api.features.library.routes.router, tags=["v2"], prefix="/api/library"
 )

autogpt_platform/backend/backend/api/ws_api.py

@@ -77,39 +77,7 @@ async def event_broadcaster(manager: ConnectionManager):
                 payload=notification.payload,
             )
 
-    async def registry_refresh_worker():
-        """Listen for LLM registry refresh notifications and broadcast to all clients."""
-        from backend.data.llm_registry import REGISTRY_REFRESH_CHANNEL
-        from backend.data.redis_client import connect_async
-
-        redis = await connect_async()
-        pubsub = redis.pubsub()
-        await pubsub.subscribe(REGISTRY_REFRESH_CHANNEL)
-        logger.info(
-            "Subscribed to LLM registry refresh notifications for WebSocket broadcast"
-        )
-
-        async for message in pubsub.listen():
-            if (
-                message["type"] == "message"
-                and message["channel"] == REGISTRY_REFRESH_CHANNEL
-            ):
-                logger.info(
-                    "Broadcasting LLM registry refresh to all WebSocket clients"
-                )
-                await manager.broadcast_to_all(
-                    method=WSMethod.NOTIFICATION,
-                    data={
-                        "type": "LLM_REGISTRY_REFRESH",
-                        "event": "registry_updated",
-                    },
-                )
-
-    await asyncio.gather(
-        execution_worker(),
-        notification_worker(),
-        registry_refresh_worker(),
-    )
+    await asyncio.gather(execution_worker(), notification_worker())
 
 
 async def authenticate_websocket(websocket: WebSocket) -> str:

autogpt_platform/backend/backend/blocks/ai_condition.py

@@ -1,6 +1,7 @@
 from typing import Any
 
 from backend.blocks.llm import (
+    DEFAULT_LLM_MODEL,
     TEST_CREDENTIALS,
     TEST_CREDENTIALS_INPUT,
     AIBlockBase,
@@ -9,7 +10,6 @@ from backend.blocks.llm import (
     LlmModel,
     LLMResponse,
     llm_call,
-    llm_model_schema_extra,
 )
 from backend.data.block import (
     BlockCategory,
@@ -50,10 +50,9 @@ class AIConditionBlock(AIBlockBase):
         )
         model: LlmModel = SchemaField(
             title="LLM Model",
-            default_factory=LlmModel.default,
+            default=DEFAULT_LLM_MODEL,
             description="The language model to use for evaluating the condition.",
             advanced=False,
-            json_schema_extra=llm_model_schema_extra(),
         )
         credentials: AICredentials = AICredentialsField()
 
@@ -83,7 +82,7 @@ class AIConditionBlock(AIBlockBase):
                 "condition": "the input is an email address",
                 "yes_value": "Valid email",
                 "no_value": "Not an email",
-                "model": "gpt-4o",  # Using string value - enum accepts any model slug dynamically
+                "model": DEFAULT_LLM_MODEL,
                 "credentials": TEST_CREDENTIALS_INPUT,
             },
             test_credentials=TEST_CREDENTIALS,

autogpt_platform/backend/backend/blocks/llm.py

@@ -4,19 +4,17 @@ import logging
 import re
 import secrets
 from abc import ABC
-from enum import Enum
+from enum import Enum, EnumMeta
 from json import JSONDecodeError
-from typing import Any, Iterable, List, Literal, Optional
+from typing import Any, Iterable, List, Literal, NamedTuple, Optional
 
 import anthropic
 import ollama
 import openai
 from anthropic.types import ToolParam
 from groq import AsyncGroq
-from pydantic import BaseModel, GetCoreSchemaHandler, SecretStr
-from pydantic_core import CoreSchema, core_schema
+from pydantic import BaseModel, SecretStr
 
-from backend.data import llm_registry
 from backend.data.block import (
     Block,
     BlockCategory,
@@ -24,7 +22,6 @@ from backend.data.block import (
     BlockSchemaInput,
     BlockSchemaOutput,
 )
-from backend.data.llm_registry import ModelMetadata
 from backend.data.model import (
     APIKeyCredentials,
     CredentialsField,
@@ -69,117 +66,114 @@ TEST_CREDENTIALS_INPUT = {
 
 
 def AICredentialsField() -> AICredentials:
-    """
-    Returns a CredentialsField for LLM providers.
-    The discriminator_mapping will be refreshed when the schema is generated
-    if it's empty, ensuring the LLM registry is loaded.
-    """
-    # Get the mapping now - it may be empty initially, but will be refreshed
-    # when the schema is generated via CredentialsMetaInput._add_json_schema_extra
-    mapping = llm_registry.get_llm_discriminator_mapping()
-
     return CredentialsField(
         description="API key for the LLM provider.",
         discriminator="model",
-        discriminator_mapping=mapping,  # May be empty initially, refreshed later
+        discriminator_mapping={
+            model.value: model.metadata.provider for model in LlmModel
+        },
     )
 
 
-def llm_model_schema_extra() -> dict[str, Any]:
-    return {"options": llm_registry.get_llm_model_schema_options()}
+class ModelMetadata(NamedTuple):
+    provider: str
+    context_window: int
+    max_output_tokens: int | None
 
 
-class LlmModelMeta(type):
-    """
-    Metaclass for LlmModel that enables attribute-style access to dynamic models.
-
-    This allows code like `LlmModel.GPT4O` to work by converting the attribute
-    name to a slug format:
-    - GPT4O -> gpt-4o
-    - GPT4O_MINI -> gpt-4o-mini
-    - CLAUDE_3_5_SONNET -> claude-3-5-sonnet
-    """
-
-    def __getattr__(cls, name: str):
-        # Don't intercept private/dunder attributes
-        if name.startswith("_"):
-            raise AttributeError(f"type object 'LlmModel' has no attribute '{name}'")
-
-        # Convert attribute name to slug format:
-        # 1. Lowercase: GPT4O -> gpt4o
-        # 2. Underscores to hyphens: GPT4O_MINI -> gpt4o-mini
-        # 3. Insert hyphen between letter and digit: gpt4o -> gpt-4o
-        slug = name.lower().replace("_", "-")
-        slug = re.sub(r"([a-z])(\d)", r"\1-\2", slug)
-
-        return cls(slug)
+class LlmModelMeta(EnumMeta):
+    pass
 
 
-class LlmModel(str, metaclass=LlmModelMeta):
-    """
-    Dynamic LLM model type that accepts any model slug from the registry.
-
-    This is a string subclass (not an Enum) that allows any model slug value.
-    All models are managed via the LLM Registry in the database.
-
-    Usage:
-        model = LlmModel("gpt-4o")  # Direct construction
-        model = LlmModel.GPT4O      # Attribute access (converted to "gpt-4o")
-        model.value                  # Returns the slug string
-        model.provider               # Returns the provider from registry
-    """
-
-    def __new__(cls, value: str):
-        if isinstance(value, LlmModel):
-            return value
-        return str.__new__(cls, value)
-
-    @classmethod
-    def __get_pydantic_core_schema__(
-        cls, source_type: Any, handler: GetCoreSchemaHandler
-    ) -> CoreSchema:
-        """
-        Tell Pydantic how to validate LlmModel.
-
-        Accepts strings and converts them to LlmModel instances.
-        """
-        return core_schema.no_info_after_validator_function(
-            cls,  # The validator function (LlmModel constructor)
-            core_schema.str_schema(),  # Accept string input
-            serialization=core_schema.to_string_ser_schema(),  # Serialize as string
-        )
-
-    @property
-    def value(self) -> str:
-        """Return the model slug (for compatibility with enum-style access)."""
-        return str(self)
-
-    @classmethod
-    def default(cls) -> "LlmModel":
-        """
-        Get the default model from the registry.
-
-        Returns the recommended model if set, otherwise gpt-4o if available
-        and enabled, otherwise the first enabled model from the registry.
-        Falls back to "gpt-4o" if registry is empty (e.g., at module import time).
-        """
-        from backend.data.llm_registry import get_default_model_slug
-
-        slug = get_default_model_slug()
-        if slug is None:
-            # Registry is empty (e.g., at module import time before DB connection).
-            # Fall back to gpt-4o for backward compatibility.
-            slug = "gpt-4o"
-        return cls(slug)
+class LlmModel(str, Enum, metaclass=LlmModelMeta):
+    # OpenAI models
+    O3_MINI = "o3-mini"
+    O3 = "o3-2025-04-16"
+    O1 = "o1"
+    O1_MINI = "o1-mini"
+    # GPT-5 models
+    GPT5_2 = "gpt-5.2-2025-12-11"
+    GPT5_1 = "gpt-5.1-2025-11-13"
+    GPT5 = "gpt-5-2025-08-07"
+    GPT5_MINI = "gpt-5-mini-2025-08-07"
+    GPT5_NANO = "gpt-5-nano-2025-08-07"
+    GPT5_CHAT = "gpt-5-chat-latest"
+    GPT41 = "gpt-4.1-2025-04-14"
+    GPT41_MINI = "gpt-4.1-mini-2025-04-14"
+    GPT4O_MINI = "gpt-4o-mini"
+    GPT4O = "gpt-4o"
+    GPT4_TURBO = "gpt-4-turbo"
+    GPT3_5_TURBO = "gpt-3.5-turbo"
+    # Anthropic models
+    CLAUDE_4_1_OPUS = "claude-opus-4-1-20250805"
+    CLAUDE_4_OPUS = "claude-opus-4-20250514"
+    CLAUDE_4_SONNET = "claude-sonnet-4-20250514"
+    CLAUDE_4_5_OPUS = "claude-opus-4-5-20251101"
+    CLAUDE_4_5_SONNET = "claude-sonnet-4-5-20250929"
+    CLAUDE_4_5_HAIKU = "claude-haiku-4-5-20251001"
+    CLAUDE_3_7_SONNET = "claude-3-7-sonnet-20250219"
+    CLAUDE_3_HAIKU = "claude-3-haiku-20240307"
+    # AI/ML API models
+    AIML_API_QWEN2_5_72B = "Qwen/Qwen2.5-72B-Instruct-Turbo"
+    AIML_API_LLAMA3_1_70B = "nvidia/llama-3.1-nemotron-70b-instruct"
+    AIML_API_LLAMA3_3_70B = "meta-llama/Llama-3.3-70B-Instruct-Turbo"
+    AIML_API_META_LLAMA_3_1_70B = "meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo"
+    AIML_API_LLAMA_3_2_3B = "meta-llama/Llama-3.2-3B-Instruct-Turbo"
+    # Groq models
+    LLAMA3_3_70B = "llama-3.3-70b-versatile"
+    LLAMA3_1_8B = "llama-3.1-8b-instant"
+    # Ollama models
+    OLLAMA_LLAMA3_3 = "llama3.3"
+    OLLAMA_LLAMA3_2 = "llama3.2"
+    OLLAMA_LLAMA3_8B = "llama3"
+    OLLAMA_LLAMA3_405B = "llama3.1:405b"
+    OLLAMA_DOLPHIN = "dolphin-mistral:latest"
+    # OpenRouter models
+    OPENAI_GPT_OSS_120B = "openai/gpt-oss-120b"
+    OPENAI_GPT_OSS_20B = "openai/gpt-oss-20b"
+    GEMINI_2_5_PRO = "google/gemini-2.5-pro-preview-03-25"
+    GEMINI_3_PRO_PREVIEW = "google/gemini-3-pro-preview"
+    GEMINI_2_5_FLASH = "google/gemini-2.5-flash"
+    GEMINI_2_0_FLASH = "google/gemini-2.0-flash-001"
+    GEMINI_2_5_FLASH_LITE_PREVIEW = "google/gemini-2.5-flash-lite-preview-06-17"
+    GEMINI_2_0_FLASH_LITE = "google/gemini-2.0-flash-lite-001"
+    MISTRAL_NEMO = "mistralai/mistral-nemo"
+    COHERE_COMMAND_R_08_2024 = "cohere/command-r-08-2024"
+    COHERE_COMMAND_R_PLUS_08_2024 = "cohere/command-r-plus-08-2024"
+    DEEPSEEK_CHAT = "deepseek/deepseek-chat"  # Actually: DeepSeek V3
+    DEEPSEEK_R1_0528 = "deepseek/deepseek-r1-0528"
+    PERPLEXITY_SONAR = "perplexity/sonar"
+    PERPLEXITY_SONAR_PRO = "perplexity/sonar-pro"
+    PERPLEXITY_SONAR_DEEP_RESEARCH = "perplexity/sonar-deep-research"
+    NOUSRESEARCH_HERMES_3_LLAMA_3_1_405B = "nousresearch/hermes-3-llama-3.1-405b"
+    NOUSRESEARCH_HERMES_3_LLAMA_3_1_70B = "nousresearch/hermes-3-llama-3.1-70b"
+    AMAZON_NOVA_LITE_V1 = "amazon/nova-lite-v1"
+    AMAZON_NOVA_MICRO_V1 = "amazon/nova-micro-v1"
+    AMAZON_NOVA_PRO_V1 = "amazon/nova-pro-v1"
+    MICROSOFT_WIZARDLM_2_8X22B = "microsoft/wizardlm-2-8x22b"
+    GRYPHE_MYTHOMAX_L2_13B = "gryphe/mythomax-l2-13b"
+    META_LLAMA_4_SCOUT = "meta-llama/llama-4-scout"
+    META_LLAMA_4_MAVERICK = "meta-llama/llama-4-maverick"
+    GROK_4 = "x-ai/grok-4"
+    GROK_4_FAST = "x-ai/grok-4-fast"
+    GROK_4_1_FAST = "x-ai/grok-4.1-fast"
+    GROK_CODE_FAST_1 = "x-ai/grok-code-fast-1"
+    KIMI_K2 = "moonshotai/kimi-k2"
+    QWEN3_235B_A22B_THINKING = "qwen/qwen3-235b-a22b-thinking-2507"
+    QWEN3_CODER = "qwen/qwen3-coder"
+    # Llama API models
+    LLAMA_API_LLAMA_4_SCOUT = "Llama-4-Scout-17B-16E-Instruct-FP8"
+    LLAMA_API_LLAMA4_MAVERICK = "Llama-4-Maverick-17B-128E-Instruct-FP8"
+    LLAMA_API_LLAMA3_3_8B = "Llama-3.3-8B-Instruct"
+    LLAMA_API_LLAMA3_3_70B = "Llama-3.3-70B-Instruct"
+    # v0 by Vercel models
+    V0_1_5_MD = "v0-1.5-md"
+    V0_1_5_LG = "v0-1.5-lg"
+    V0_1_0_MD = "v0-1.0-md"
 
     @property
     def metadata(self) -> ModelMetadata:
-        metadata = llm_registry.get_llm_model_metadata(self.value)
-        if metadata:
-            return metadata
-        raise ValueError(
-            f"Missing metadata for model: {self.value}. Model not found in LLM registry."
-        )
+        return MODEL_METADATA[self]
 
     @property
     def provider(self) -> str:
@@ -194,11 +188,128 @@ class LlmModel(str, metaclass=LlmModelMeta):
         return self.metadata.max_output_tokens
 
 
-# MODEL_METADATA removed - all models now come from the database via llm_registry
+MODEL_METADATA = {
+    # https://platform.openai.com/docs/models
+    LlmModel.O3: ModelMetadata("openai", 200000, 100000),
+    LlmModel.O3_MINI: ModelMetadata("openai", 200000, 100000),  # o3-mini-2025-01-31
+    LlmModel.O1: ModelMetadata("openai", 200000, 100000),  # o1-2024-12-17
+    LlmModel.O1_MINI: ModelMetadata("openai", 128000, 65536),  # o1-mini-2024-09-12
+    # GPT-5 models
+    LlmModel.GPT5_2: ModelMetadata("openai", 400000, 128000),
+    LlmModel.GPT5_1: ModelMetadata("openai", 400000, 128000),
+    LlmModel.GPT5: ModelMetadata("openai", 400000, 128000),
+    LlmModel.GPT5_MINI: ModelMetadata("openai", 400000, 128000),
+    LlmModel.GPT5_NANO: ModelMetadata("openai", 400000, 128000),
+    LlmModel.GPT5_CHAT: ModelMetadata("openai", 400000, 16384),
+    LlmModel.GPT41: ModelMetadata("openai", 1047576, 32768),
+    LlmModel.GPT41_MINI: ModelMetadata("openai", 1047576, 32768),
+    LlmModel.GPT4O_MINI: ModelMetadata(
+        "openai", 128000, 16384
+    ),  # gpt-4o-mini-2024-07-18
+    LlmModel.GPT4O: ModelMetadata("openai", 128000, 16384),  # gpt-4o-2024-08-06
+    LlmModel.GPT4_TURBO: ModelMetadata(
+        "openai", 128000, 4096
+    ),  # gpt-4-turbo-2024-04-09
+    LlmModel.GPT3_5_TURBO: ModelMetadata("openai", 16385, 4096),  # gpt-3.5-turbo-0125
+    # https://docs.anthropic.com/en/docs/about-claude/models
+    LlmModel.CLAUDE_4_1_OPUS: ModelMetadata(
+        "anthropic", 200000, 32000
+    ),  # claude-opus-4-1-20250805
+    LlmModel.CLAUDE_4_OPUS: ModelMetadata(
+        "anthropic", 200000, 32000
+    ),  # claude-4-opus-20250514
+    LlmModel.CLAUDE_4_SONNET: ModelMetadata(
+        "anthropic", 200000, 64000
+    ),  # claude-4-sonnet-20250514
+    LlmModel.CLAUDE_4_5_OPUS: ModelMetadata(
+        "anthropic", 200000, 64000
+    ),  # claude-opus-4-5-20251101
+    LlmModel.CLAUDE_4_5_SONNET: ModelMetadata(
+        "anthropic", 200000, 64000
+    ),  # claude-sonnet-4-5-20250929
+    LlmModel.CLAUDE_4_5_HAIKU: ModelMetadata(
+        "anthropic", 200000, 64000
+    ),  # claude-haiku-4-5-20251001
+    LlmModel.CLAUDE_3_7_SONNET: ModelMetadata(
+        "anthropic", 200000, 64000
+    ),  # claude-3-7-sonnet-20250219
+    LlmModel.CLAUDE_3_HAIKU: ModelMetadata(
+        "anthropic", 200000, 4096
+    ),  # claude-3-haiku-20240307
+    # https://docs.aimlapi.com/api-overview/model-database/text-models
+    LlmModel.AIML_API_QWEN2_5_72B: ModelMetadata("aiml_api", 32000, 8000),
+    LlmModel.AIML_API_LLAMA3_1_70B: ModelMetadata("aiml_api", 128000, 40000),
+    LlmModel.AIML_API_LLAMA3_3_70B: ModelMetadata("aiml_api", 128000, None),
+    LlmModel.AIML_API_META_LLAMA_3_1_70B: ModelMetadata("aiml_api", 131000, 2000),
+    LlmModel.AIML_API_LLAMA_3_2_3B: ModelMetadata("aiml_api", 128000, None),
+    # https://console.groq.com/docs/models
+    LlmModel.LLAMA3_3_70B: ModelMetadata("groq", 128000, 32768),
+    LlmModel.LLAMA3_1_8B: ModelMetadata("groq", 128000, 8192),
+    # https://ollama.com/library
+    LlmModel.OLLAMA_LLAMA3_3: ModelMetadata("ollama", 8192, None),
+    LlmModel.OLLAMA_LLAMA3_2: ModelMetadata("ollama", 8192, None),
+    LlmModel.OLLAMA_LLAMA3_8B: ModelMetadata("ollama", 8192, None),
+    LlmModel.OLLAMA_LLAMA3_405B: ModelMetadata("ollama", 8192, None),
+    LlmModel.OLLAMA_DOLPHIN: ModelMetadata("ollama", 32768, None),
+    # https://openrouter.ai/models
+    LlmModel.GEMINI_2_5_PRO: ModelMetadata("open_router", 1050000, 8192),
+    LlmModel.GEMINI_3_PRO_PREVIEW: ModelMetadata("open_router", 1048576, 65535),
+    LlmModel.GEMINI_2_5_FLASH: ModelMetadata("open_router", 1048576, 65535),
+    LlmModel.GEMINI_2_0_FLASH: ModelMetadata("open_router", 1048576, 8192),
+    LlmModel.GEMINI_2_5_FLASH_LITE_PREVIEW: ModelMetadata(
+        "open_router", 1048576, 65535
+    ),
+    LlmModel.GEMINI_2_0_FLASH_LITE: ModelMetadata("open_router", 1048576, 8192),
+    LlmModel.MISTRAL_NEMO: ModelMetadata("open_router", 128000, 4096),
+    LlmModel.COHERE_COMMAND_R_08_2024: ModelMetadata("open_router", 128000, 4096),
+    LlmModel.COHERE_COMMAND_R_PLUS_08_2024: ModelMetadata("open_router", 128000, 4096),
+    LlmModel.DEEPSEEK_CHAT: ModelMetadata("open_router", 64000, 2048),
+    LlmModel.DEEPSEEK_R1_0528: ModelMetadata("open_router", 163840, 163840),
+    LlmModel.PERPLEXITY_SONAR: ModelMetadata("open_router", 127000, 8000),
+    LlmModel.PERPLEXITY_SONAR_PRO: ModelMetadata("open_router", 200000, 8000),
+    LlmModel.PERPLEXITY_SONAR_DEEP_RESEARCH: ModelMetadata(
+        "open_router",
+        128000,
+        16000,
+    ),
+    LlmModel.NOUSRESEARCH_HERMES_3_LLAMA_3_1_405B: ModelMetadata(
+        "open_router", 131000, 4096
+    ),
+    LlmModel.NOUSRESEARCH_HERMES_3_LLAMA_3_1_70B: ModelMetadata(
+        "open_router", 12288, 12288
+    ),
+    LlmModel.OPENAI_GPT_OSS_120B: ModelMetadata("open_router", 131072, 131072),
+    LlmModel.OPENAI_GPT_OSS_20B: ModelMetadata("open_router", 131072, 32768),
+    LlmModel.AMAZON_NOVA_LITE_V1: ModelMetadata("open_router", 300000, 5120),
+    LlmModel.AMAZON_NOVA_MICRO_V1: ModelMetadata("open_router", 128000, 5120),
+    LlmModel.AMAZON_NOVA_PRO_V1: ModelMetadata("open_router", 300000, 5120),
+    LlmModel.MICROSOFT_WIZARDLM_2_8X22B: ModelMetadata("open_router", 65536, 4096),
+    LlmModel.GRYPHE_MYTHOMAX_L2_13B: ModelMetadata("open_router", 4096, 4096),
+    LlmModel.META_LLAMA_4_SCOUT: ModelMetadata("open_router", 131072, 131072),
+    LlmModel.META_LLAMA_4_MAVERICK: ModelMetadata("open_router", 1048576, 1000000),
+    LlmModel.GROK_4: ModelMetadata("open_router", 256000, 256000),
+    LlmModel.GROK_4_FAST: ModelMetadata("open_router", 2000000, 30000),
+    LlmModel.GROK_4_1_FAST: ModelMetadata("open_router", 2000000, 30000),
+    LlmModel.GROK_CODE_FAST_1: ModelMetadata("open_router", 256000, 10000),
+    LlmModel.KIMI_K2: ModelMetadata("open_router", 131000, 131000),
+    LlmModel.QWEN3_235B_A22B_THINKING: ModelMetadata("open_router", 262144, 262144),
+    LlmModel.QWEN3_CODER: ModelMetadata("open_router", 262144, 262144),
+    # Llama API models
+    LlmModel.LLAMA_API_LLAMA_4_SCOUT: ModelMetadata("llama_api", 128000, 4028),
+    LlmModel.LLAMA_API_LLAMA4_MAVERICK: ModelMetadata("llama_api", 128000, 4028),
+    LlmModel.LLAMA_API_LLAMA3_3_8B: ModelMetadata("llama_api", 128000, 4028),
+    LlmModel.LLAMA_API_LLAMA3_3_70B: ModelMetadata("llama_api", 128000, 4028),
+    # v0 by Vercel models
+    LlmModel.V0_1_5_MD: ModelMetadata("v0", 128000, 64000),
+    LlmModel.V0_1_5_LG: ModelMetadata("v0", 512000, 64000),
+    LlmModel.V0_1_0_MD: ModelMetadata("v0", 128000, 64000),
+}
 
-# Default model constant for backward compatibility
-# Uses the dynamic registry to get the default model
-DEFAULT_LLM_MODEL = LlmModel.default()
+DEFAULT_LLM_MODEL = LlmModel.GPT5_2
+
+for model in LlmModel:
+    if model not in MODEL_METADATA:
+        raise ValueError(f"Missing MODEL_METADATA metadata for model: {model}")
 
 
 class ToolCall(BaseModel):
@@ -327,94 +438,19 @@ async def llm_call(
             - prompt_tokens: The number of tokens used in the prompt.
             - completion_tokens: The number of tokens used in the completion.
     """
-    # Get model metadata and check if enabled - with fallback support
-    # The model we'll actually use (may differ if original is disabled)
-    model_to_use = llm_model.value
-
-    # Check if model is in registry and if it's enabled
-    from backend.data.llm_registry import (
-        get_fallback_model_for_disabled,
-        get_model_info,
-    )
-
-    model_info = get_model_info(llm_model.value)
-
-    if model_info and not model_info.is_enabled:
-        # Model is disabled - try to find a fallback from the same provider
-        fallback = get_fallback_model_for_disabled(llm_model.value)
-        if fallback:
-            logger.warning(
-                f"Model '{llm_model.value}' is disabled. Using fallback model '{fallback.slug}' from the same provider ({fallback.metadata.provider})."
-            )
-            model_to_use = fallback.slug
-            # Use fallback model's metadata
-            provider = fallback.metadata.provider
-            context_window = fallback.metadata.context_window
-            model_max_output = fallback.metadata.max_output_tokens or int(2**15)
-        else:
-            # No fallback available - raise error
-            raise ValueError(
-                f"LLM model '{llm_model.value}' is disabled and no fallback model "
-                f"from the same provider is available. Please enable the model or "
-                f"select a different model in the block configuration."
-            )
-    else:
-        # Model is enabled or not in registry (legacy/static model)
-        try:
-            provider = llm_model.metadata.provider
-            context_window = llm_model.context_window
-            model_max_output = llm_model.max_output_tokens or int(2**15)
-        except ValueError:
-            # Model not in cache - try refreshing the registry once if we have DB access
-            logger.warning(f"Model {llm_model.value} not found in registry cache")
-
-            # Try refreshing the registry if we have database access
-            from backend.data.db import is_connected
-
-            if is_connected():
-                try:
-                    logger.info(
-                        f"Refreshing LLM registry and retrying lookup for {llm_model.value}"
-                    )
-                    await llm_registry.refresh_llm_registry()
-                    # Try again after refresh
-                    try:
-                        provider = llm_model.metadata.provider
-                        context_window = llm_model.context_window
-                        model_max_output = llm_model.max_output_tokens or int(2**15)
-                        logger.info(
-                            f"Successfully loaded model {llm_model.value} metadata after registry refresh"
-                        )
-                    except ValueError:
-                        # Still not found after refresh
-                        raise ValueError(
-                            f"LLM model '{llm_model.value}' not found in registry after refresh. "
-                            "Please ensure the model is added and enabled in the LLM registry via the admin UI."
-                        )
-                except Exception as refresh_exc:
-                    logger.error(f"Failed to refresh LLM registry: {refresh_exc}")
-                    raise ValueError(
-                        f"LLM model '{llm_model.value}' not found in registry and failed to refresh. "
-                        "Please ensure the model is added to the LLM registry via the admin UI."
-                    ) from refresh_exc
-            else:
-                # No DB access (e.g., in executor without direct DB connection)
-                # The registry should have been loaded on startup
-                raise ValueError(
-                    f"LLM model '{llm_model.value}' not found in registry cache. "
-                    "The registry may need to be refreshed. Please contact support or try again later."
-                )
+    provider = llm_model.metadata.provider
+    context_window = llm_model.context_window
 
     if compress_prompt_to_fit:
         prompt = compress_prompt(
             messages=prompt,
-            target_tokens=context_window // 2,
+            target_tokens=llm_model.context_window // 2,
             lossy_ok=True,
         )
 
     # Calculate available tokens based on context window and input length
     estimated_input_tokens = estimate_token_count(prompt)
-    # model_max_output already set above
+    model_max_output = llm_model.max_output_tokens or int(2**15)
     user_max = max_tokens or model_max_output
     available_tokens = max(context_window - estimated_input_tokens, 0)
     max_tokens = max(min(available_tokens, model_max_output, user_max), 1)
@@ -432,7 +468,7 @@ async def llm_call(
             response_format = {"type": "json_object"}
 
         response = await oai_client.chat.completions.create(
-            model=model_to_use,
+            model=llm_model.value,
             messages=prompt,  # type: ignore
             response_format=response_format,  # type: ignore
             max_completion_tokens=max_tokens,
@@ -479,7 +515,7 @@ async def llm_call(
         )
         try:
             resp = await client.messages.create(
-                model=model_to_use,
+                model=llm_model.value,
                 system=sysprompt,
                 messages=messages,
                 max_tokens=max_tokens,
@@ -543,7 +579,7 @@ async def llm_call(
         client = AsyncGroq(api_key=credentials.api_key.get_secret_value())
         response_format = {"type": "json_object"} if force_json_output else None
         response = await client.chat.completions.create(
-            model=model_to_use,
+            model=llm_model.value,
             messages=prompt,  # type: ignore
             response_format=response_format,  # type: ignore
             max_tokens=max_tokens,
@@ -565,7 +601,7 @@ async def llm_call(
         sys_messages = [p["content"] for p in prompt if p["role"] == "system"]
         usr_messages = [p["content"] for p in prompt if p["role"] != "system"]
         response = await client.generate(
-            model=model_to_use,
+            model=llm_model.value,
             prompt=f"{sys_messages}\n\n{usr_messages}",
             stream=False,
             options={"num_ctx": max_tokens},
@@ -595,7 +631,7 @@ async def llm_call(
                 "HTTP-Referer": "https://agpt.co",
                 "X-Title": "AutoGPT",
             },
-            model=model_to_use,
+            model=llm_model.value,
             messages=prompt,  # type: ignore
             max_tokens=max_tokens,
             tools=tools_param,  # type: ignore
@@ -637,7 +673,7 @@ async def llm_call(
                 "HTTP-Referer": "https://agpt.co",
                 "X-Title": "AutoGPT",
             },
-            model=model_to_use,
+            model=llm_model.value,
             messages=prompt,  # type: ignore
             max_tokens=max_tokens,
             tools=tools_param,  # type: ignore
@@ -664,7 +700,7 @@ async def llm_call(
             reasoning=reasoning,
         )
     elif provider == "aiml_api":
-        client = openai.AsyncOpenAI(
+        client = openai.OpenAI(
             base_url="https://api.aimlapi.com/v2",
             api_key=credentials.api_key.get_secret_value(),
             default_headers={
@@ -674,8 +710,8 @@ async def llm_call(
             },
         )
 
-        completion = await client.chat.completions.create(
-            model=model_to_use,
+        completion = client.chat.completions.create(
+            model=llm_model.value,
             messages=prompt,  # type: ignore
             max_tokens=max_tokens,
         )
@@ -707,7 +743,7 @@ async def llm_call(
         )
 
         response = await client.chat.completions.create(
-            model=model_to_use,
+            model=llm_model.value,
             messages=prompt,  # type: ignore
             response_format=response_format,  # type: ignore
             max_tokens=max_tokens,
@@ -758,10 +794,9 @@ class AIStructuredResponseGeneratorBlock(AIBlockBase):
         )
         model: LlmModel = SchemaField(
             title="LLM Model",
-            default_factory=LlmModel.default,
+            default=DEFAULT_LLM_MODEL,
             description="The language model to use for answering the prompt.",
             advanced=False,
-            json_schema_extra=llm_model_schema_extra(),
         )
         force_json_output: bool = SchemaField(
             title="Restrict LLM to pure JSON output",
@@ -824,7 +859,7 @@ class AIStructuredResponseGeneratorBlock(AIBlockBase):
             input_schema=AIStructuredResponseGeneratorBlock.Input,
             output_schema=AIStructuredResponseGeneratorBlock.Output,
             test_input={
-                "model": "gpt-4o",  # Using string value - enum accepts any model slug dynamically
+                "model": DEFAULT_LLM_MODEL,
                 "credentials": TEST_CREDENTIALS_INPUT,
                 "expected_format": {
                     "key1": "value1",
@@ -1190,10 +1225,9 @@ class AITextGeneratorBlock(AIBlockBase):
         )
         model: LlmModel = SchemaField(
             title="LLM Model",
-            default_factory=LlmModel.default,
+            default=DEFAULT_LLM_MODEL,
             description="The language model to use for answering the prompt.",
             advanced=False,
-            json_schema_extra=llm_model_schema_extra(),
         )
         credentials: AICredentials = AICredentialsField()
         sys_prompt: str = SchemaField(
@@ -1287,9 +1321,8 @@ class AITextSummarizerBlock(AIBlockBase):
         )
         model: LlmModel = SchemaField(
             title="LLM Model",
-            default_factory=LlmModel.default,
+            default=DEFAULT_LLM_MODEL,
             description="The language model to use for summarizing the text.",
-            json_schema_extra=llm_model_schema_extra(),
         )
         focus: str = SchemaField(
             title="Focus",
@@ -1505,9 +1538,8 @@ class AIConversationBlock(AIBlockBase):
         )
         model: LlmModel = SchemaField(
             title="LLM Model",
-            default_factory=LlmModel.default,
+            default=DEFAULT_LLM_MODEL,
             description="The language model to use for the conversation.",
-            json_schema_extra=llm_model_schema_extra(),
         )
         credentials: AICredentials = AICredentialsField()
         max_tokens: int | None = SchemaField(
@@ -1544,7 +1576,7 @@ class AIConversationBlock(AIBlockBase):
                     },
                     {"role": "user", "content": "Where was it played?"},
                 ],
-                "model": "gpt-4o",  # Using string value - enum accepts any model slug dynamically
+                "model": DEFAULT_LLM_MODEL,
                 "credentials": TEST_CREDENTIALS_INPUT,
             },
             test_credentials=TEST_CREDENTIALS,
@@ -1607,10 +1639,9 @@ class AIListGeneratorBlock(AIBlockBase):
         )
         model: LlmModel = SchemaField(
             title="LLM Model",
-            default_factory=LlmModel.default,
+            default=DEFAULT_LLM_MODEL,
             description="The language model to use for generating the list.",
             advanced=True,
-            json_schema_extra=llm_model_schema_extra(),
         )
         credentials: AICredentials = AICredentialsField()
         max_retries: int = SchemaField(
@@ -1665,7 +1696,7 @@ class AIListGeneratorBlock(AIBlockBase):
                     "drawing explorers to uncover its mysteries. Each planet showcases the limitless possibilities of "
                     "fictional worlds."
                 ),
-                "model": "gpt-4o",  # Using string value - enum accepts any model slug dynamically
+                "model": DEFAULT_LLM_MODEL,
                 "credentials": TEST_CREDENTIALS_INPUT,
                 "max_retries": 3,
                 "force_json_output": False,

autogpt_platform/backend/backend/blocks/smart_decision_maker.py

@@ -226,10 +226,9 @@ class SmartDecisionMakerBlock(Block):
         )
         model: llm.LlmModel = SchemaField(
             title="LLM Model",
-            default_factory=llm.LlmModel.default,
+            default=llm.DEFAULT_LLM_MODEL,
             description="The language model to use for answering the prompt.",
             advanced=False,
-            json_schema_extra=llm.llm_model_schema_extra(),
         )
         credentials: llm.AICredentials = llm.AICredentialsField()
         multiple_tool_calls: bool = SchemaField(

autogpt_platform/backend/backend/blocks/stagehand/blocks.py

@@ -10,13 +10,13 @@ import stagehand.main
 from stagehand import Stagehand
 
 from backend.blocks.llm import (
+    MODEL_METADATA,
     AICredentials,
     AICredentialsField,
     LlmModel,
     ModelMetadata,
 )
 from backend.blocks.stagehand._config import stagehand as stagehand_provider
-from backend.data import llm_registry
 from backend.sdk import (
     APIKeyCredentials,
     Block,
@@ -91,7 +91,7 @@ class StagehandRecommendedLlmModel(str, Enum):
         Returns the provider name for the model in the required format for Stagehand:
         provider/model_name
         """
-        model_metadata = self.metadata
+        model_metadata = MODEL_METADATA[LlmModel(self.value)]
         model_name = self.value
 
         if len(model_name.split("/")) == 1 and not self.value.startswith(
@@ -107,23 +107,19 @@ class StagehandRecommendedLlmModel(str, Enum):
 
     @property
     def provider(self) -> str:
-        return self.metadata.provider
+        return MODEL_METADATA[LlmModel(self.value)].provider
 
     @property
     def metadata(self) -> ModelMetadata:
-        metadata = llm_registry.get_llm_model_metadata(self.value)
-        if metadata:
-            return metadata
-        # Fallback to LlmModel enum if registry lookup fails
-        return LlmModel(self.value).metadata
+        return MODEL_METADATA[LlmModel(self.value)]
 
     @property
     def context_window(self) -> int:
-        return self.metadata.context_window
+        return MODEL_METADATA[LlmModel(self.value)].context_window
 
     @property
     def max_output_tokens(self) -> int | None:
-        return self.metadata.max_output_tokens
+        return MODEL_METADATA[LlmModel(self.value)].max_output_tokens
 
 
 class StagehandObserveBlock(Block):

autogpt_platform/backend/backend/data/analytics.py

@@ -104,7 +104,7 @@ async def get_accuracy_trends_and_alerts(
             AND e."executionStatus" IN ('COMPLETED', 'FAILED', 'TERMINATED')
             {user_filter}
         GROUP BY DATE(e."createdAt")
-        HAVING COUNT(*) >= 3  -- Need at least 3 executions per day
+        HAVING COUNT(*) >= 1  -- Include all days with at least 1 execution
     ),
     trends AS (
         SELECT 

autogpt_platform/backend/backend/data/block.py

@@ -25,7 +25,6 @@ from prisma.models import AgentBlock
 from prisma.types import AgentBlockCreateInput
 from pydantic import BaseModel
 
-from backend.data.llm_registry import update_schema_with_llm_registry
 from backend.data.model import NodeExecutionStats
 from backend.integrations.providers import ProviderName
 from backend.util import json
@@ -144,59 +143,35 @@ class BlockInfo(BaseModel):
 
 
 class BlockSchema(BaseModel):
-    cached_jsonschema: ClassVar[dict[str, Any] | None] = None
-
-    @classmethod
-    def clear_schema_cache(cls) -> None:
-        """Clear the cached JSON schema for this class."""
-        # Use None instead of {} because {} is truthy and would prevent regeneration
-        cls.cached_jsonschema = None  # type: ignore
-
-    @staticmethod
-    def clear_all_schema_caches() -> None:
-        """Clear cached JSON schemas for all BlockSchema subclasses."""
-
-        def clear_recursive(cls: type) -> None:
-            """Recursively clear cache for class and all subclasses."""
-            if hasattr(cls, "clear_schema_cache"):
-                cls.clear_schema_cache()
-            for subclass in cls.__subclasses__():
-                clear_recursive(subclass)
-
-        clear_recursive(BlockSchema)
+    cached_jsonschema: ClassVar[dict[str, Any]]
 
     @classmethod
     def jsonschema(cls) -> dict[str, Any]:
-        # Generate schema if not cached
-        if not cls.cached_jsonschema:
-            model = jsonref.replace_refs(cls.model_json_schema(), merge_props=True)
+        if cls.cached_jsonschema:
+            return cls.cached_jsonschema
 
-            def ref_to_dict(obj):
-                if isinstance(obj, dict):
-                    # OpenAPI <3.1 does not support sibling fields that has a $ref key
-                    # So sometimes, the schema has an "allOf"/"anyOf"/"oneOf" with 1 item.
-                    keys = {"allOf", "anyOf", "oneOf"}
-                    one_key = next(
-                        (k for k in keys if k in obj and len(obj[k]) == 1), None
-                    )
-                    if one_key:
-                        obj.update(obj[one_key][0])
+        model = jsonref.replace_refs(cls.model_json_schema(), merge_props=True)
 
-                    return {
-                        key: ref_to_dict(value)
-                        for key, value in obj.items()
-                        if not key.startswith("$") and key != one_key
-                    }
-                elif isinstance(obj, list):
-                    return [ref_to_dict(item) for item in obj]
+        def ref_to_dict(obj):
+            if isinstance(obj, dict):
+                # OpenAPI <3.1 does not support sibling fields that has a $ref key
+                # So sometimes, the schema has an "allOf"/"anyOf"/"oneOf" with 1 item.
+                keys = {"allOf", "anyOf", "oneOf"}
+                one_key = next((k for k in keys if k in obj and len(obj[k]) == 1), None)
+                if one_key:
+                    obj.update(obj[one_key][0])
 
-                return obj
+                return {
+                    key: ref_to_dict(value)
+                    for key, value in obj.items()
+                    if not key.startswith("$") and key != one_key
+                }
+            elif isinstance(obj, list):
+                return [ref_to_dict(item) for item in obj]
 
-            cls.cached_jsonschema = cast(dict[str, Any], ref_to_dict(model))
+            return obj
 
-        # Always post-process to ensure LLM registry data is up-to-date
-        # This refreshes model options and discriminator mappings even if schema was cached
-        update_schema_with_llm_registry(cls.cached_jsonschema, cls)
+        cls.cached_jsonschema = cast(dict[str, Any], ref_to_dict(model))
 
         return cls.cached_jsonschema
 
@@ -705,12 +680,23 @@ class Block(ABC, Generic[BlockSchemaInputType, BlockSchemaOutputType]):
         return False, reviewed_data
 
     async def _execute(self, input_data: BlockInput, **kwargs) -> BlockOutput:
-        # Check for review requirement and get potentially modified input data
-        should_pause, input_data = await self.is_block_exec_need_review(
-            input_data, **kwargs
+        # Check for review requirement only if running within a graph execution context
+        # Direct block execution (e.g., from chat) skips the review process
+        has_graph_context = all(
+            key in kwargs
+            for key in (
+                "node_exec_id",
+                "graph_exec_id",
+                "graph_id",
+                "execution_context",
+            )
         )
-        if should_pause:
-            return
+        if has_graph_context:
+            should_pause, input_data = await self.is_block_exec_need_review(
+                input_data, **kwargs
+            )
+            if should_pause:
+                return
 
         # Validate the input data (original or reviewer-modified) once
         if error := self.input_schema.validate_data(input_data):
@@ -884,28 +870,6 @@ def is_block_auth_configured(
 
 
 async def initialize_blocks() -> None:
-    # Refresh LLM registry before initializing blocks so blocks can use registry data
-    # This ensures the registry cache is populated even in executor context
-    try:
-        from backend.data import llm_registry
-        from backend.data.block_cost_config import refresh_llm_costs
-
-        # Only refresh if we have DB access (check if Prisma is connected)
-        from backend.data.db import is_connected
-
-        if is_connected():
-            await llm_registry.refresh_llm_registry()
-            refresh_llm_costs()
-            logger.info("LLM registry refreshed during block initialization")
-        else:
-            logger.warning(
-                "Prisma not connected, skipping LLM registry refresh during block initialization"
-            )
-    except Exception as exc:
-        logger.warning(
-            "Failed to refresh LLM registry during block initialization: %s", exc
-        )
-
     # First, sync all provider costs to blocks
     # Imported here to avoid circular import
     from backend.sdk.cost_integration import sync_all_provider_costs

autogpt_platform/backend/backend/data/block_cost_config.py

@@ -1,4 +1,3 @@
-import logging
 from typing import Type
 
 from backend.blocks.ai_image_customizer import AIImageCustomizerBlock, GeminiImageModel
@@ -24,18 +23,19 @@ from backend.blocks.ideogram import IdeogramModelBlock
 from backend.blocks.jina.embeddings import JinaEmbeddingBlock
 from backend.blocks.jina.search import ExtractWebsiteContentBlock, SearchTheWebBlock
 from backend.blocks.llm import (
+    MODEL_METADATA,
     AIConversationBlock,
     AIListGeneratorBlock,
     AIStructuredResponseGeneratorBlock,
     AITextGeneratorBlock,
     AITextSummarizerBlock,
+    LlmModel,
 )
 from backend.blocks.replicate.flux_advanced import ReplicateFluxAdvancedModelBlock
 from backend.blocks.replicate.replicate_block import ReplicateModelBlock
 from backend.blocks.smart_decision_maker import SmartDecisionMakerBlock
 from backend.blocks.talking_head import CreateTalkingAvatarVideoBlock
 from backend.blocks.text_to_speech_block import UnrealTextToSpeechBlock
-from backend.data import llm_registry
 from backend.data.block import Block, BlockCost, BlockCostType
 from backend.integrations.credentials_store import (
     aiml_api_credentials,
@@ -55,63 +55,210 @@ from backend.integrations.credentials_store import (
     v0_credentials,
 )
 
-logger = logging.getLogger(__name__)
+# =============== Configure the cost for each LLM Model call =============== #
 
-PROVIDER_CREDENTIALS = {
-    "openai": openai_credentials,
-    "anthropic": anthropic_credentials,
-    "groq": groq_credentials,
-    "open_router": open_router_credentials,
-    "llama_api": llama_api_credentials,
-    "aiml_api": aiml_api_credentials,
-    "v0": v0_credentials,
+MODEL_COST: dict[LlmModel, int] = {
+    LlmModel.O3: 4,
+    LlmModel.O3_MINI: 2,
+    LlmModel.O1: 16,
+    LlmModel.O1_MINI: 4,
+    # GPT-5 models
+    LlmModel.GPT5_2: 6,
+    LlmModel.GPT5_1: 5,
+    LlmModel.GPT5: 2,
+    LlmModel.GPT5_MINI: 1,
+    LlmModel.GPT5_NANO: 1,
+    LlmModel.GPT5_CHAT: 5,
+    LlmModel.GPT41: 2,
+    LlmModel.GPT41_MINI: 1,
+    LlmModel.GPT4O_MINI: 1,
+    LlmModel.GPT4O: 3,
+    LlmModel.GPT4_TURBO: 10,
+    LlmModel.GPT3_5_TURBO: 1,
+    LlmModel.CLAUDE_4_1_OPUS: 21,
+    LlmModel.CLAUDE_4_OPUS: 21,
+    LlmModel.CLAUDE_4_SONNET: 5,
+    LlmModel.CLAUDE_4_5_HAIKU: 4,
+    LlmModel.CLAUDE_4_5_OPUS: 14,
+    LlmModel.CLAUDE_4_5_SONNET: 9,
+    LlmModel.CLAUDE_3_7_SONNET: 5,
+    LlmModel.CLAUDE_3_HAIKU: 1,
+    LlmModel.AIML_API_QWEN2_5_72B: 1,
+    LlmModel.AIML_API_LLAMA3_1_70B: 1,
+    LlmModel.AIML_API_LLAMA3_3_70B: 1,
+    LlmModel.AIML_API_META_LLAMA_3_1_70B: 1,
+    LlmModel.AIML_API_LLAMA_3_2_3B: 1,
+    LlmModel.LLAMA3_3_70B: 1,
+    LlmModel.LLAMA3_1_8B: 1,
+    LlmModel.OLLAMA_LLAMA3_3: 1,
+    LlmModel.OLLAMA_LLAMA3_2: 1,
+    LlmModel.OLLAMA_LLAMA3_8B: 1,
+    LlmModel.OLLAMA_LLAMA3_405B: 1,
+    LlmModel.OLLAMA_DOLPHIN: 1,
+    LlmModel.OPENAI_GPT_OSS_120B: 1,
+    LlmModel.OPENAI_GPT_OSS_20B: 1,
+    LlmModel.GEMINI_2_5_PRO: 4,
+    LlmModel.GEMINI_3_PRO_PREVIEW: 5,
+    LlmModel.MISTRAL_NEMO: 1,
+    LlmModel.COHERE_COMMAND_R_08_2024: 1,
+    LlmModel.COHERE_COMMAND_R_PLUS_08_2024: 3,
+    LlmModel.DEEPSEEK_CHAT: 2,
+    LlmModel.PERPLEXITY_SONAR: 1,
+    LlmModel.PERPLEXITY_SONAR_PRO: 5,
+    LlmModel.PERPLEXITY_SONAR_DEEP_RESEARCH: 10,
+    LlmModel.NOUSRESEARCH_HERMES_3_LLAMA_3_1_405B: 1,
+    LlmModel.NOUSRESEARCH_HERMES_3_LLAMA_3_1_70B: 1,
+    LlmModel.AMAZON_NOVA_LITE_V1: 1,
+    LlmModel.AMAZON_NOVA_MICRO_V1: 1,
+    LlmModel.AMAZON_NOVA_PRO_V1: 1,
+    LlmModel.MICROSOFT_WIZARDLM_2_8X22B: 1,
+    LlmModel.GRYPHE_MYTHOMAX_L2_13B: 1,
+    LlmModel.META_LLAMA_4_SCOUT: 1,
+    LlmModel.META_LLAMA_4_MAVERICK: 1,
+    LlmModel.LLAMA_API_LLAMA_4_SCOUT: 1,
+    LlmModel.LLAMA_API_LLAMA4_MAVERICK: 1,
+    LlmModel.LLAMA_API_LLAMA3_3_8B: 1,
+    LlmModel.LLAMA_API_LLAMA3_3_70B: 1,
+    LlmModel.GROK_4: 9,
+    LlmModel.GROK_4_FAST: 1,
+    LlmModel.GROK_4_1_FAST: 1,
+    LlmModel.GROK_CODE_FAST_1: 1,
+    LlmModel.KIMI_K2: 1,
+    LlmModel.QWEN3_235B_A22B_THINKING: 1,
+    LlmModel.QWEN3_CODER: 9,
+    LlmModel.GEMINI_2_5_FLASH: 1,
+    LlmModel.GEMINI_2_0_FLASH: 1,
+    LlmModel.GEMINI_2_5_FLASH_LITE_PREVIEW: 1,
+    LlmModel.GEMINI_2_0_FLASH_LITE: 1,
+    LlmModel.DEEPSEEK_R1_0528: 1,
+    # v0 by Vercel models
+    LlmModel.V0_1_5_MD: 1,
+    LlmModel.V0_1_5_LG: 2,
+    LlmModel.V0_1_0_MD: 1,
 }
 
-# =============== Configure the cost for each LLM Model call =============== #
-# All LLM costs now come from the database via llm_registry
-
-LLM_COST: list[BlockCost] = []
+for model in LlmModel:
+    if model not in MODEL_COST:
+        raise ValueError(f"Missing MODEL_COST for model: {model}")
 
 
-def _build_llm_costs_from_registry() -> list[BlockCost]:
-    """Build BlockCost list from all models in the LLM registry."""
-    costs: list[BlockCost] = []
-    for model in llm_registry.iter_dynamic_models():
-        for cost in model.costs:
-            credentials = PROVIDER_CREDENTIALS.get(cost.credential_provider)
-            if not credentials:
-                logger.warning(
-                    "Skipping cost entry for %s due to unknown credentials provider %s",
-                    model.slug,
-                    cost.credential_provider,
-                )
-                continue
-            cost_filter = {
-                "model": model.slug,
+LLM_COST = (
+    # Anthropic Models
+    [
+        BlockCost(
+            cost_type=BlockCostType.RUN,
+            cost_filter={
+                "model": model,
                 "credentials": {
-                    "id": credentials.id,
-                    "provider": credentials.provider,
-                    "type": credentials.type,
+                    "id": anthropic_credentials.id,
+                    "provider": anthropic_credentials.provider,
+                    "type": anthropic_credentials.type,
                 },
-            }
-            costs.append(
-                BlockCost(
-                    cost_type=BlockCostType.RUN,
-                    cost_filter=cost_filter,
-                    cost_amount=cost.credit_cost,
-                )
-            )
-    return costs
-
-
-def refresh_llm_costs() -> None:
-    """Refresh LLM costs from the registry. All costs now come from the database."""
-    LLM_COST.clear()
-    LLM_COST.extend(_build_llm_costs_from_registry())
-
-
-# Initial load will happen after registry is refreshed at startup
-# Don't call refresh_llm_costs() here - it will be called after registry refresh
+            },
+            cost_amount=cost,
+        )
+        for model, cost in MODEL_COST.items()
+        if MODEL_METADATA[model].provider == "anthropic"
+    ]
+    # OpenAI Models
+    + [
+        BlockCost(
+            cost_type=BlockCostType.RUN,
+            cost_filter={
+                "model": model,
+                "credentials": {
+                    "id": openai_credentials.id,
+                    "provider": openai_credentials.provider,
+                    "type": openai_credentials.type,
+                },
+            },
+            cost_amount=cost,
+        )
+        for model, cost in MODEL_COST.items()
+        if MODEL_METADATA[model].provider == "openai"
+    ]
+    # Groq Models
+    + [
+        BlockCost(
+            cost_type=BlockCostType.RUN,
+            cost_filter={
+                "model": model,
+                "credentials": {"id": groq_credentials.id},
+            },
+            cost_amount=cost,
+        )
+        for model, cost in MODEL_COST.items()
+        if MODEL_METADATA[model].provider == "groq"
+    ]
+    # Open Router Models
+    + [
+        BlockCost(
+            cost_type=BlockCostType.RUN,
+            cost_filter={
+                "model": model,
+                "credentials": {
+                    "id": open_router_credentials.id,
+                    "provider": open_router_credentials.provider,
+                    "type": open_router_credentials.type,
+                },
+            },
+            cost_amount=cost,
+        )
+        for model, cost in MODEL_COST.items()
+        if MODEL_METADATA[model].provider == "open_router"
+    ]
+    # Llama API Models
+    + [
+        BlockCost(
+            cost_type=BlockCostType.RUN,
+            cost_filter={
+                "model": model,
+                "credentials": {
+                    "id": llama_api_credentials.id,
+                    "provider": llama_api_credentials.provider,
+                    "type": llama_api_credentials.type,
+                },
+            },
+            cost_amount=cost,
+        )
+        for model, cost in MODEL_COST.items()
+        if MODEL_METADATA[model].provider == "llama_api"
+    ]
+    # v0 by Vercel Models
+    + [
+        BlockCost(
+            cost_type=BlockCostType.RUN,
+            cost_filter={
+                "model": model,
+                "credentials": {
+                    "id": v0_credentials.id,
+                    "provider": v0_credentials.provider,
+                    "type": v0_credentials.type,
+                },
+            },
+            cost_amount=cost,
+        )
+        for model, cost in MODEL_COST.items()
+        if MODEL_METADATA[model].provider == "v0"
+    ]
+    # AI/ML Api Models
+    + [
+        BlockCost(
+            cost_type=BlockCostType.RUN,
+            cost_filter={
+                "model": model,
+                "credentials": {
+                    "id": aiml_api_credentials.id,
+                    "provider": aiml_api_credentials.provider,
+                    "type": aiml_api_credentials.type,
+                },
+            },
+            cost_amount=cost,
+        )
+        for model, cost in MODEL_COST.items()
+        if MODEL_METADATA[model].provider == "aiml_api"
+    ]
+)
 
 # =============== This is the exhaustive list of cost for each Block =============== #
 

autogpt_platform/backend/backend/data/execution.py

@@ -153,8 +153,14 @@ class GraphExecutionMeta(BaseDbModel):
     nodes_input_masks: Optional[dict[str, BlockInput]]
     preset_id: Optional[str]
     status: ExecutionStatus
-    started_at: datetime
-    ended_at: datetime
+    started_at: Optional[datetime] = Field(
+        None,
+        description="When execution started running. Null if not yet started (QUEUED).",
+    )
+    ended_at: Optional[datetime] = Field(
+        None,
+        description="When execution finished. Null if not yet completed (QUEUED, RUNNING, INCOMPLETE, REVIEW).",
+    )
     is_shared: bool = False
     share_token: Optional[str] = None
 
@@ -229,10 +235,8 @@ class GraphExecutionMeta(BaseDbModel):
 
     @staticmethod
     def from_db(_graph_exec: AgentGraphExecution):
-        now = datetime.now(timezone.utc)
-        # TODO: make started_at and ended_at optional
-        start_time = _graph_exec.startedAt or _graph_exec.createdAt
-        end_time = _graph_exec.updatedAt or now
+        start_time = _graph_exec.startedAt
+        end_time = _graph_exec.endedAt
 
         try:
             stats = GraphExecutionStats.model_validate(_graph_exec.stats)
@@ -902,6 +906,14 @@ async def update_graph_execution_stats(
 
     if status:
         update_data["executionStatus"] = status
+        # Set endedAt when execution reaches a terminal status
+        terminal_statuses = [
+            ExecutionStatus.COMPLETED,
+            ExecutionStatus.FAILED,
+            ExecutionStatus.TERMINATED,
+        ]
+        if status in terminal_statuses:
+            update_data["endedAt"] = datetime.now(tz=timezone.utc)
 
     where_clause: AgentGraphExecutionWhereInput = {"id": graph_exec_id}
 

autogpt_platform/backend/backend/data/graph.py

@@ -1483,10 +1483,8 @@ async def migrate_llm_models(migrate_to: LlmModel):
             if field.annotation == LlmModel:
                 llm_model_fields[block.id] = field_name
 
-    # Get all model slugs from the registry (dynamic, not hardcoded enum)
-    from backend.data import llm_registry
-
-    enum_values = list(llm_registry.get_all_model_slugs_for_validation())
+    # Convert enum values to a list of strings for the SQL query
+    enum_values = [v.value for v in LlmModel]
     escaped_enum_values = repr(tuple(enum_values))  # hack but works
 
     # Update each block

autogpt_platform/backend/backend/data/llm_registry/__init__.py

@@ -1,72 +0,0 @@
-"""
-LLM Registry module for managing LLM models, providers, and costs dynamically.
-
-This module provides a database-driven registry system for LLM models,
-replacing hardcoded model configurations with a flexible admin-managed system.
-"""
-
-from backend.data.llm_registry.model_types import ModelMetadata
-
-# Re-export for backwards compatibility
-from backend.data.llm_registry.notifications import (
-    REGISTRY_REFRESH_CHANNEL,
-    publish_registry_refresh_notification,
-    subscribe_to_registry_refresh,
-)
-from backend.data.llm_registry.registry import (
-    RegistryModel,
-    RegistryModelCost,
-    RegistryModelCreator,
-    get_all_model_slugs_for_validation,
-    get_default_model_slug,
-    get_dynamic_model_slugs,
-    get_fallback_model_for_disabled,
-    get_llm_discriminator_mapping,
-    get_llm_model_cost,
-    get_llm_model_metadata,
-    get_llm_model_schema_options,
-    get_model_info,
-    is_model_enabled,
-    iter_dynamic_models,
-    refresh_llm_registry,
-    register_static_costs,
-    register_static_metadata,
-)
-from backend.data.llm_registry.schema_utils import (
-    is_llm_model_field,
-    refresh_llm_discriminator_mapping,
-    refresh_llm_model_options,
-    update_schema_with_llm_registry,
-)
-
-__all__ = [
-    # Types
-    "ModelMetadata",
-    "RegistryModel",
-    "RegistryModelCost",
-    "RegistryModelCreator",
-    # Registry functions
-    "get_all_model_slugs_for_validation",
-    "get_default_model_slug",
-    "get_dynamic_model_slugs",
-    "get_fallback_model_for_disabled",
-    "get_llm_discriminator_mapping",
-    "get_llm_model_cost",
-    "get_llm_model_metadata",
-    "get_llm_model_schema_options",
-    "get_model_info",
-    "is_model_enabled",
-    "iter_dynamic_models",
-    "refresh_llm_registry",
-    "register_static_costs",
-    "register_static_metadata",
-    # Notifications
-    "REGISTRY_REFRESH_CHANNEL",
-    "publish_registry_refresh_notification",
-    "subscribe_to_registry_refresh",
-    # Schema utilities
-    "is_llm_model_field",
-    "refresh_llm_discriminator_mapping",
-    "refresh_llm_model_options",
-    "update_schema_with_llm_registry",
-]

autogpt_platform/backend/backend/data/llm_registry/model_types.py

@@ -1,11 +0,0 @@
-"""Type definitions for LLM model metadata."""
-
-from typing import NamedTuple
-
-
-class ModelMetadata(NamedTuple):
-    """Metadata for an LLM model."""
-
-    provider: str
-    context_window: int
-    max_output_tokens: int | None

autogpt_platform/backend/backend/data/llm_registry/notifications.py

@@ -1,89 +0,0 @@
-"""
-Redis pub/sub notifications for LLM registry updates.
-
-When models are added/updated/removed via the admin UI, this module
-publishes notifications to Redis that all executor services subscribe to,
-ensuring they refresh their registry cache in real-time.
-"""
-
-import asyncio
-import logging
-from typing import Any
-
-from backend.data.redis_client import connect_async
-
-logger = logging.getLogger(__name__)
-
-# Redis channel name for LLM registry refresh notifications
-REGISTRY_REFRESH_CHANNEL = "llm_registry:refresh"
-
-
-async def publish_registry_refresh_notification() -> None:
-    """
-    Publish a notification to Redis that the LLM registry has been updated.
-    All executor services subscribed to this channel will refresh their registry.
-    """
-    try:
-        redis = await connect_async()
-        await redis.publish(REGISTRY_REFRESH_CHANNEL, "refresh")
-        logger.info("Published LLM registry refresh notification to Redis")
-    except Exception as exc:
-        logger.warning(
-            "Failed to publish LLM registry refresh notification: %s",
-            exc,
-            exc_info=True,
-        )
-
-
-async def subscribe_to_registry_refresh(
-    on_refresh: Any,  # Async callable that takes no args
-) -> None:
-    """
-    Subscribe to Redis notifications for LLM registry updates.
-    This runs in a loop and processes messages as they arrive.
-
-    Args:
-        on_refresh: Async callable to execute when a refresh notification is received
-    """
-    try:
-        redis = await connect_async()
-        pubsub = redis.pubsub()
-        await pubsub.subscribe(REGISTRY_REFRESH_CHANNEL)
-        logger.info(
-            "Subscribed to LLM registry refresh notifications on channel: %s",
-            REGISTRY_REFRESH_CHANNEL,
-        )
-
-        # Process messages in a loop
-        while True:
-            try:
-                message = await pubsub.get_message(
-                    ignore_subscribe_messages=True, timeout=1.0
-                )
-                if (
-                    message
-                    and message["type"] == "message"
-                    and message["channel"] == REGISTRY_REFRESH_CHANNEL
-                ):
-                    logger.info("Received LLM registry refresh notification")
-                    try:
-                        await on_refresh()
-                    except Exception as exc:
-                        logger.error(
-                            "Error refreshing LLM registry from notification: %s",
-                            exc,
-                            exc_info=True,
-                        )
-            except Exception as exc:
-                logger.warning(
-                    "Error processing registry refresh message: %s", exc, exc_info=True
-                )
-                # Continue listening even if one message fails
-                await asyncio.sleep(1)
-    except Exception as exc:
-        logger.error(
-            "Failed to subscribe to LLM registry refresh notifications: %s",
-            exc,
-            exc_info=True,
-        )
-        raise

autogpt_platform/backend/backend/data/llm_registry/registry.py

@@ -1,372 +0,0 @@
-"""Core LLM registry implementation for managing models dynamically."""
-
-from __future__ import annotations
-
-import asyncio
-import logging
-from dataclasses import dataclass, field
-from typing import Any, Iterable
-
-import prisma.models
-
-from backend.data.llm_registry.model_types import ModelMetadata
-
-logger = logging.getLogger(__name__)
-
-
-def _json_to_dict(value: Any) -> dict[str, Any]:
-    """Convert Prisma Json type to dict, with fallback to empty dict."""
-    if value is None:
-        return {}
-    if isinstance(value, dict):
-        return value
-    # Prisma Json type should always be a dict at runtime
-    return dict(value) if value else {}
-
-
-@dataclass(frozen=True)
-class RegistryModelCost:
-    """Cost configuration for an LLM model."""
-
-    credit_cost: int
-    credential_provider: str
-    credential_id: str | None
-    credential_type: str | None
-    currency: str | None
-    metadata: dict[str, Any]
-
-
-@dataclass(frozen=True)
-class RegistryModelCreator:
-    """Creator information for an LLM model."""
-
-    id: str
-    name: str
-    display_name: str
-    description: str | None
-    website_url: str | None
-    logo_url: str | None
-
-
-@dataclass(frozen=True)
-class RegistryModel:
-    """Represents a model in the LLM registry."""
-
-    slug: str
-    display_name: str
-    description: str | None
-    metadata: ModelMetadata
-    capabilities: dict[str, Any]
-    extra_metadata: dict[str, Any]
-    provider_display_name: str
-    is_enabled: bool
-    is_recommended: bool = False
-    costs: tuple[RegistryModelCost, ...] = field(default_factory=tuple)
-    creator: RegistryModelCreator | None = None
-
-
-_static_metadata: dict[str, ModelMetadata] = {}
-_static_costs: dict[str, int] = {}
-_dynamic_models: dict[str, RegistryModel] = {}
-_schema_options: list[dict[str, str]] = []
-_discriminator_mapping: dict[str, str] = {}
-_lock = asyncio.Lock()
-
-
-def register_static_metadata(metadata: dict[Any, ModelMetadata]) -> None:
-    """Register static metadata for legacy models (deprecated)."""
-    _static_metadata.update({str(key): value for key, value in metadata.items()})
-    _refresh_cached_schema()
-
-
-def register_static_costs(costs: dict[Any, int]) -> None:
-    """Register static costs for legacy models (deprecated)."""
-    _static_costs.update({str(key): value for key, value in costs.items()})
-
-
-def _build_schema_options() -> list[dict[str, str]]:
-    """Build schema options for model selection dropdown. Only includes enabled models."""
-    options: list[dict[str, str]] = []
-    # Only include enabled models in the dropdown options
-    for model in sorted(_dynamic_models.values(), key=lambda m: m.display_name.lower()):
-        if model.is_enabled:
-            options.append(
-                {
-                    "label": model.display_name,
-                    "value": model.slug,
-                    "group": model.metadata.provider,
-                    "description": model.description or "",
-                }
-            )
-
-    for slug, metadata in _static_metadata.items():
-        if slug in _dynamic_models:
-            continue
-        options.append(
-            {
-                "label": slug,
-                "value": slug,
-                "group": metadata.provider,
-                "description": "",
-            }
-        )
-    return options
-
-
-async def refresh_llm_registry() -> None:
-    """Refresh the LLM registry from the database. Loads all models (enabled and disabled)."""
-    async with _lock:
-        try:
-            records = await prisma.models.LlmModel.prisma().find_many(
-                include={
-                    "Provider": True,
-                    "Costs": True,
-                    "Creator": True,
-                }
-            )
-            logger.debug("Found %d LLM model records in database", len(records))
-        except Exception as exc:
-            logger.error(
-                "Failed to refresh LLM registry from DB: %s", exc, exc_info=True
-            )
-            return
-
-        dynamic: dict[str, RegistryModel] = {}
-        for record in records:
-            provider_name = (
-                record.Provider.name if record.Provider else record.providerId
-            )
-            metadata = ModelMetadata(
-                provider=provider_name,
-                context_window=record.contextWindow,
-                max_output_tokens=record.maxOutputTokens,
-            )
-            costs = tuple(
-                RegistryModelCost(
-                    credit_cost=cost.creditCost,
-                    credential_provider=cost.credentialProvider,
-                    credential_id=cost.credentialId,
-                    credential_type=cost.credentialType,
-                    currency=cost.currency,
-                    metadata=_json_to_dict(cost.metadata),
-                )
-                for cost in (record.Costs or [])
-            )
-
-            # Map creator if present
-            creator = None
-            if record.Creator:
-                creator = RegistryModelCreator(
-                    id=record.Creator.id,
-                    name=record.Creator.name,
-                    display_name=record.Creator.displayName,
-                    description=record.Creator.description,
-                    website_url=record.Creator.websiteUrl,
-                    logo_url=record.Creator.logoUrl,
-                )
-
-            dynamic[record.slug] = RegistryModel(
-                slug=record.slug,
-                display_name=record.displayName,
-                description=record.description,
-                metadata=metadata,
-                capabilities=_json_to_dict(record.capabilities),
-                extra_metadata=_json_to_dict(record.metadata),
-                provider_display_name=(
-                    record.Provider.displayName
-                    if record.Provider
-                    else record.providerId
-                ),
-                is_enabled=record.isEnabled,
-                is_recommended=record.isRecommended,
-                costs=costs,
-                creator=creator,
-            )
-
-        # Atomic swap - build new structures then replace references
-        # This ensures readers never see partially updated state
-        global _dynamic_models
-        _dynamic_models = dynamic
-        _refresh_cached_schema()
-        logger.info(
-            "LLM registry refreshed with %s dynamic models (enabled: %s, disabled: %s)",
-            len(dynamic),
-            sum(1 for m in dynamic.values() if m.is_enabled),
-            sum(1 for m in dynamic.values() if not m.is_enabled),
-        )
-
-
-def _refresh_cached_schema() -> None:
-    """Refresh cached schema options and discriminator mapping."""
-    global _schema_options, _discriminator_mapping
-
-    # Build new structures
-    new_options = _build_schema_options()
-    new_mapping = {
-        slug: entry.metadata.provider for slug, entry in _dynamic_models.items()
-    }
-    for slug, metadata in _static_metadata.items():
-        new_mapping.setdefault(slug, metadata.provider)
-
-    # Atomic swap - replace references to ensure readers see consistent state
-    _schema_options = new_options
-    _discriminator_mapping = new_mapping
-
-
-def get_llm_model_metadata(slug: str) -> ModelMetadata | None:
-    """Get model metadata by slug. Checks dynamic models first, then static metadata."""
-    if slug in _dynamic_models:
-        return _dynamic_models[slug].metadata
-    return _static_metadata.get(slug)
-
-
-def get_llm_model_cost(slug: str) -> tuple[RegistryModelCost, ...]:
-    """Get model cost configuration by slug."""
-    if slug in _dynamic_models:
-        return _dynamic_models[slug].costs
-    cost_value = _static_costs.get(slug)
-    if cost_value is None:
-        return tuple()
-    return (
-        RegistryModelCost(
-            credit_cost=cost_value,
-            credential_provider="static",
-            credential_id=None,
-            credential_type=None,
-            currency=None,
-            metadata={},
-        ),
-    )
-
-
-def get_llm_model_schema_options() -> list[dict[str, str]]:
-    """
-    Get schema options for LLM model selection dropdown.
-
-    Returns a copy of cached schema options that are refreshed when the registry is
-    updated via refresh_llm_registry() (called on startup and via Redis pub/sub).
-    """
-    # Return a copy to prevent external mutation
-    return list(_schema_options)
-
-
-def get_llm_discriminator_mapping() -> dict[str, str]:
-    """
-    Get discriminator mapping for LLM models.
-
-    Returns a copy of cached discriminator mapping that is refreshed when the registry
-    is updated via refresh_llm_registry() (called on startup and via Redis pub/sub).
-    """
-    # Return a copy to prevent external mutation
-    return dict(_discriminator_mapping)
-
-
-def get_dynamic_model_slugs() -> set[str]:
-    """Get all dynamic model slugs from the registry."""
-    return set(_dynamic_models.keys())
-
-
-def get_all_model_slugs_for_validation() -> set[str]:
-    """
-    Get ALL model slugs (both enabled and disabled) for validation purposes.
-
-    This is used for JSON schema enum validation - we need to accept any known
-    model value (even disabled ones) so that existing graphs don't fail validation.
-    The actual fallback/enforcement happens at runtime in llm_call().
-    """
-    all_slugs = set(_dynamic_models.keys())
-    all_slugs.update(_static_metadata.keys())
-    return all_slugs
-
-
-def iter_dynamic_models() -> Iterable[RegistryModel]:
-    """Iterate over all dynamic models in the registry."""
-    return tuple(_dynamic_models.values())
-
-
-def get_fallback_model_for_disabled(disabled_model_slug: str) -> RegistryModel | None:
-    """
-    Find a fallback model when the requested model is disabled.
-
-    Looks for an enabled model from the same provider. Prefers models with
-    similar names or capabilities if possible.
-
-    Args:
-        disabled_model_slug: The slug of the disabled model
-
-    Returns:
-        An enabled RegistryModel from the same provider, or None if no fallback found
-    """
-    disabled_model = _dynamic_models.get(disabled_model_slug)
-    if not disabled_model:
-        return None
-
-    provider = disabled_model.metadata.provider
-
-    # Find all enabled models from the same provider
-    candidates = [
-        model
-        for model in _dynamic_models.values()
-        if model.is_enabled and model.metadata.provider == provider
-    ]
-
-    if not candidates:
-        return None
-
-    # Sort by: prefer models with similar context window, then by name
-    candidates.sort(
-        key=lambda m: (
-            abs(m.metadata.context_window - disabled_model.metadata.context_window),
-            m.display_name.lower(),
-        )
-    )
-
-    return candidates[0]
-
-
-def is_model_enabled(model_slug: str) -> bool:
-    """Check if a model is enabled in the registry."""
-    model = _dynamic_models.get(model_slug)
-    if not model:
-        # Model not in registry - assume it's a static/legacy model and allow it
-        return True
-    return model.is_enabled
-
-
-def get_model_info(model_slug: str) -> RegistryModel | None:
-    """Get model info from the registry."""
-    return _dynamic_models.get(model_slug)
-
-
-def get_default_model_slug() -> str | None:
-    """
-    Get the default model slug to use for block defaults.
-
-    Returns the recommended model if set (configured via admin UI),
-    otherwise returns the first enabled model alphabetically.
-    Returns None if no models are available or enabled.
-    """
-    # Return the recommended model if one is set and enabled
-    for model in _dynamic_models.values():
-        if model.is_recommended and model.is_enabled:
-            return model.slug
-
-    # No recommended model set - find first enabled model alphabetically
-    for model in sorted(_dynamic_models.values(), key=lambda m: m.display_name.lower()):
-        if model.is_enabled:
-            logger.warning(
-                "No recommended model set, using '%s' as default",
-                model.slug,
-            )
-            return model.slug
-
-    # No enabled models available
-    if _dynamic_models:
-        logger.error(
-            "No enabled models found in registry (%d models registered but all disabled)",
-            len(_dynamic_models),
-        )
-    else:
-        logger.error("No models registered in LLM registry")
-
-    return None

autogpt_platform/backend/backend/data/llm_registry/schema_utils.py

@@ -1,130 +0,0 @@
-"""
-Helper utilities for LLM registry integration with block schemas.
-
-This module handles the dynamic injection of discriminator mappings
-and model options from the LLM registry into block schemas.
-"""
-
-import logging
-from typing import Any
-
-from backend.data.llm_registry.registry import (
-    get_all_model_slugs_for_validation,
-    get_default_model_slug,
-    get_llm_discriminator_mapping,
-    get_llm_model_schema_options,
-)
-
-logger = logging.getLogger(__name__)
-
-
-def is_llm_model_field(field_name: str, field_info: Any) -> bool:
-    """
-    Check if a field is an LLM model selection field.
-
-    Returns True if the field has 'options' in json_schema_extra
-    (set by llm_model_schema_extra() in blocks/llm.py).
-    """
-    if not hasattr(field_info, "json_schema_extra"):
-        return False
-
-    extra = field_info.json_schema_extra
-    if isinstance(extra, dict):
-        return "options" in extra
-
-    return False
-
-
-def refresh_llm_model_options(field_schema: dict[str, Any]) -> None:
-    """
-    Refresh LLM model options from the registry.
-
-    Updates 'options' (for frontend dropdown) to show only enabled models,
-    but keeps the 'enum' (for validation) inclusive of ALL known models.
-
-    This is important because:
-    - Options: What users see in the dropdown (enabled models only)
-    - Enum: What values pass validation (all known models, including disabled)
-
-    Existing graphs may have disabled models selected - they should pass validation
-    and the fallback logic in llm_call() will handle using an alternative model.
-    """
-    fresh_options = get_llm_model_schema_options()
-    if not fresh_options:
-        return
-
-    # Update options array (UI dropdown) - only enabled models
-    if "options" in field_schema:
-        field_schema["options"] = fresh_options
-
-    all_known_slugs = get_all_model_slugs_for_validation()
-    if all_known_slugs and "enum" in field_schema:
-        existing_enum = set(field_schema.get("enum", []))
-        combined_enum = existing_enum | all_known_slugs
-        field_schema["enum"] = sorted(combined_enum)
-
-    # Set the default value from the registry (gpt-4o if available, else first enabled)
-    # This ensures new blocks have a sensible default pre-selected
-    default_slug = get_default_model_slug()
-    if default_slug:
-        field_schema["default"] = default_slug
-
-
-def refresh_llm_discriminator_mapping(field_schema: dict[str, Any]) -> None:
-    """
-    Refresh discriminator_mapping for fields that use model-based discrimination.
-
-    The discriminator is already set when AICredentialsField() creates the field.
-    We only need to refresh the mapping when models are added/removed.
-    """
-    if field_schema.get("discriminator") != "model":
-        return
-
-    # Always refresh the mapping to get latest models
-    fresh_mapping = get_llm_discriminator_mapping()
-    if fresh_mapping:
-        field_schema["discriminator_mapping"] = fresh_mapping
-
-
-def update_schema_with_llm_registry(
-    schema: dict[str, Any], model_class: type | None = None
-) -> None:
-    """
-    Update a JSON schema with current LLM registry data.
-
-    Refreshes:
-    1. Model options for LLM model selection fields (dropdown choices)
-    2. Discriminator mappings for credentials fields (model → provider)
-
-    Args:
-        schema: The JSON schema to update (mutated in-place)
-        model_class: The Pydantic model class (optional, for field introspection)
-    """
-    properties = schema.get("properties", {})
-
-    for field_name, field_schema in properties.items():
-        if not isinstance(field_schema, dict):
-            continue
-
-        # Refresh model options for LLM model fields
-        if model_class and hasattr(model_class, "model_fields"):
-            field_info = model_class.model_fields.get(field_name)
-            if field_info and is_llm_model_field(field_name, field_info):
-                try:
-                    refresh_llm_model_options(field_schema)
-                except Exception as exc:
-                    logger.warning(
-                        "Failed to refresh LLM options for field %s: %s",
-                        field_name,
-                        exc,
-                    )
-
-        # Refresh discriminator mapping for fields that use model discrimination
-        try:
-            refresh_llm_discriminator_mapping(field_schema)
-        except Exception as exc:
-            logger.warning(
-                "Failed to refresh discriminator mapping for field %s: %s",
-                field_name,
-                exc,
-            )

autogpt_platform/backend/backend/data/model.py

@@ -40,7 +40,6 @@ from pydantic_core import (
 )
 from typing_extensions import TypedDict
 
-from backend.data.llm_registry import update_schema_with_llm_registry
 from backend.integrations.providers import ProviderName
 from backend.util.json import loads as json_loads
 from backend.util.settings import Secrets
@@ -545,9 +544,7 @@ class CredentialsMetaInput(BaseModel, Generic[CP, CT]):
             else:
                 schema["credentials_provider"] = allowed_providers
             schema["credentials_types"] = model_class.allowed_cred_types()
-
-        # Ensure LLM discriminators are populated (delegates to shared helper)
-        update_schema_with_llm_registry(schema, model_class)
+        # Do not return anything, just mutate schema in place
 
     model_config = ConfigDict(
         json_schema_extra=_add_json_schema_extra,  # type: ignore
@@ -696,20 +693,16 @@ def CredentialsField(
     This is enforced by the `BlockSchema` base class.
     """
 
-    # Build field_schema_extra - always include discriminator and mapping if discriminator is set
-    field_schema_extra: dict[str, Any] = {}
-
-    # Always include discriminator if provided
-    if discriminator is not None:
-        field_schema_extra["discriminator"] = discriminator
-        # Always include discriminator_mapping when discriminator is set (even if empty initially)
-        field_schema_extra["discriminator_mapping"] = discriminator_mapping or {}
-
-    # Include other optional fields (only if not None)
-    if required_scopes:
-        field_schema_extra["credentials_scopes"] = list(required_scopes)
-    if discriminator_values:
-        field_schema_extra["discriminator_values"] = discriminator_values
+    field_schema_extra = {
+        k: v
+        for k, v in {
+            "credentials_scopes": list(required_scopes) or None,
+            "discriminator": discriminator,
+            "discriminator_mapping": discriminator_mapping,
+            "discriminator_values": discriminator_values,
+        }.items()
+        if v is not None
+    }
 
     # Merge any json_schema_extra passed in kwargs
     if "json_schema_extra" in kwargs:

autogpt_platform/backend/backend/executor/llm_registry_init.py

@@ -1,66 +0,0 @@
-"""
-Helper functions for LLM registry initialization in executor context.
-
-These functions handle refreshing the LLM registry when the executor starts
-and subscribing to real-time updates via Redis pub/sub.
-"""
-
-import logging
-
-from backend.data import db, llm_registry
-from backend.data.block import BlockSchema, initialize_blocks
-from backend.data.block_cost_config import refresh_llm_costs
-from backend.data.llm_registry import subscribe_to_registry_refresh
-
-logger = logging.getLogger(__name__)
-
-
-async def initialize_registry_for_executor() -> None:
-    """
-    Initialize blocks and refresh LLM registry in the executor context.
-
-    This must run in the executor's event loop to have access to the database.
-    """
-    try:
-        # Connect to database if not already connected
-        if not db.is_connected():
-            await db.connect()
-            logger.info("[GraphExecutor] Connected to database for registry refresh")
-
-        # Initialize blocks (internally refreshes LLM registry and costs)
-        await initialize_blocks()
-        logger.info("[GraphExecutor] Blocks initialized")
-    except Exception as exc:
-        logger.warning(
-            "[GraphExecutor] Failed to refresh LLM registry on startup: %s",
-            exc,
-            exc_info=True,
-        )
-
-
-async def refresh_registry_on_notification() -> None:
-    """Refresh LLM registry when notified via Redis pub/sub."""
-    try:
-        # Ensure DB is connected
-        if not db.is_connected():
-            await db.connect()
-
-        # Refresh registry and costs
-        await llm_registry.refresh_llm_registry()
-        refresh_llm_costs()
-
-        # Clear block schema caches so they regenerate with new model options
-        BlockSchema.clear_all_schema_caches()
-
-        logger.info("[GraphExecutor] LLM registry refreshed from notification")
-    except Exception as exc:
-        logger.error(
-            "[GraphExecutor] Failed to refresh LLM registry from notification: %s",
-            exc,
-            exc_info=True,
-        )
-
-
-async def subscribe_to_registry_updates() -> None:
-    """Subscribe to Redis pub/sub for LLM registry refresh notifications."""
-    await subscribe_to_registry_refresh(refresh_registry_on_notification)

autogpt_platform/backend/backend/executor/manager.py

@@ -702,20 +702,6 @@ class ExecutionProcessor:
         )
         self.node_execution_thread.start()
         self.node_evaluation_thread.start()
-
-        # Initialize LLM registry and subscribe to updates
-        from backend.executor.llm_registry_init import (
-            initialize_registry_for_executor,
-            subscribe_to_registry_updates,
-        )
-
-        asyncio.run_coroutine_threadsafe(
-            initialize_registry_for_executor(), self.node_execution_loop
-        )
-        asyncio.run_coroutine_threadsafe(
-            subscribe_to_registry_updates(), self.node_execution_loop
-        )
-
         logger.info(f"[GraphExecutor] {self.tid} started")
 
     @error_logged(swallow=False)

autogpt_platform/backend/backend/executor/scheduler.py

@@ -602,6 +602,18 @@ class Scheduler(AppService):
         self.scheduler.add_listener(job_max_instances_listener, EVENT_JOB_MAX_INSTANCES)
         self.scheduler.start()
 
+        # Run embedding backfill immediately on startup
+        # This ensures blocks/docs are searchable right away, not after 6 hours
+        # Safe to run on multiple pods - uses upserts and checks for existing embeddings
+        if self.register_system_tasks:
+            logger.info("Running embedding backfill on startup...")
+            try:
+                result = ensure_embeddings_coverage()
+                logger.info(f"Startup embedding backfill complete: {result}")
+            except Exception as e:
+                logger.error(f"Startup embedding backfill failed: {e}")
+                # Don't fail startup - the scheduled job will retry later
+
         # Keep the service running since BackgroundScheduler doesn't block
         super().run_service()
 

autogpt_platform/backend/backend/integrations/credentials_store.py

@@ -96,9 +96,9 @@ jina_credentials = APIKeyCredentials(
 )
 unreal_credentials = APIKeyCredentials(
     id="66f20754-1b81-48e4-91d0-f4f0dd82145f",
-    provider="unreal",
+    provider="unreal_speech",
     api_key=SecretStr(settings.secrets.unreal_speech_api_key),
-    title="Use Credits for Unreal",
+    title="Use Credits for Unreal Speech",
     expires_at=None,
 )
 open_router_credentials = APIKeyCredentials(
@@ -216,6 +216,14 @@ webshare_proxy_credentials = UserPasswordCredentials(
     title="Use Credits for Webshare Proxy",
 )
 
+openweathermap_credentials = APIKeyCredentials(
+    id="8b3d4e5f-6a7b-8c9d-0e1f-2a3b4c5d6e7f",
+    provider="openweathermap",
+    api_key=SecretStr(settings.secrets.openweathermap_api_key),
+    title="Use Credits for OpenWeatherMap",
+    expires_at=None,
+)
+
 DEFAULT_CREDENTIALS = [
     ollama_credentials,
     revid_credentials,
@@ -243,6 +251,7 @@ DEFAULT_CREDENTIALS = [
     llama_api_credentials,
     v0_credentials,
     webshare_proxy_credentials,
+    openweathermap_credentials,
 ]
 
 SYSTEM_CREDENTIAL_IDS = {cred.id for cred in DEFAULT_CREDENTIALS}
@@ -346,11 +355,17 @@ class IntegrationCredentialsStore:
             all_credentials.append(zerobounce_credentials)
         if settings.secrets.google_maps_api_key:
             all_credentials.append(google_maps_credentials)
+        if settings.secrets.llama_api_key:
+            all_credentials.append(llama_api_credentials)
+        if settings.secrets.v0_api_key:
+            all_credentials.append(v0_credentials)
         if (
             settings.secrets.webshare_proxy_username
             and settings.secrets.webshare_proxy_password
         ):
             all_credentials.append(webshare_proxy_credentials)
+        if settings.secrets.openweathermap_api_key:
+            all_credentials.append(openweathermap_credentials)
         return all_credentials
 
     async def get_creds_by_id(

autogpt_platform/backend/backend/monitoring/late_execution_monitor.py

@@ -60,8 +60,10 @@ class LateExecutionMonitor:
         if not all_late_executions:
             return "No late executions detected."
 
-        # Sort by created time (oldest first)
-        all_late_executions.sort(key=lambda x: x.started_at)
+        # Sort by started time (oldest first), with None values (unstarted) first
+        all_late_executions.sort(
+            key=lambda x: x.started_at or datetime.min.replace(tzinfo=timezone.utc)
+        )
 
         num_total_late = len(all_late_executions)
         num_queued = len(queued_late_executions)
@@ -74,7 +76,7 @@ class LateExecutionMonitor:
         was_truncated = num_total_late > tuncate_size
 
         late_execution_details = [
-            f"* `Execution ID: {exec.id}, Graph ID: {exec.graph_id}v{exec.graph_version}, User ID: {exec.user_id}, Status: {exec.status}, Created At: {exec.started_at.isoformat()}`"
+            f"* `Execution ID: {exec.id}, Graph ID: {exec.graph_id}v{exec.graph_version}, User ID: {exec.user_id}, Status: {exec.status}, Started At: {exec.started_at.isoformat() if exec.started_at else 'Not started'}`"
             for exec in truncated_executions
         ]
 

autogpt_platform/backend/backend/server/v2/llm/db.py

@@ -1,854 +0,0 @@
-from __future__ import annotations
-
-from typing import Any, Iterable, Sequence, cast
-
-import prisma
-import prisma.models
-
-from backend.data.db import transaction
-from backend.server.v2.llm import model as llm_model
-
-
-def _json_dict(value: Any | None) -> dict[str, Any]:
-    if not value:
-        return {}
-    if isinstance(value, dict):
-        return value
-    return {}
-
-
-def _map_cost(record: prisma.models.LlmModelCost) -> llm_model.LlmModelCost:
-    return llm_model.LlmModelCost(
-        id=record.id,
-        unit=record.unit,
-        credit_cost=record.creditCost,
-        credential_provider=record.credentialProvider,
-        credential_id=record.credentialId,
-        credential_type=record.credentialType,
-        currency=record.currency,
-        metadata=_json_dict(record.metadata),
-    )
-
-
-def _map_creator(
-    record: prisma.models.LlmModelCreator,
-) -> llm_model.LlmModelCreator:
-    return llm_model.LlmModelCreator(
-        id=record.id,
-        name=record.name,
-        display_name=record.displayName,
-        description=record.description,
-        website_url=record.websiteUrl,
-        logo_url=record.logoUrl,
-        metadata=_json_dict(record.metadata),
-    )
-
-
-def _map_model(record: prisma.models.LlmModel) -> llm_model.LlmModel:
-    costs = []
-    if record.Costs:
-        costs = [_map_cost(cost) for cost in record.Costs]
-
-    creator = None
-    if hasattr(record, "Creator") and record.Creator:
-        creator = _map_creator(record.Creator)
-
-    return llm_model.LlmModel(
-        id=record.id,
-        slug=record.slug,
-        display_name=record.displayName,
-        description=record.description,
-        provider_id=record.providerId,
-        creator_id=record.creatorId,
-        creator=creator,
-        context_window=record.contextWindow,
-        max_output_tokens=record.maxOutputTokens,
-        is_enabled=record.isEnabled,
-        is_recommended=record.isRecommended,
-        capabilities=_json_dict(record.capabilities),
-        metadata=_json_dict(record.metadata),
-        costs=costs,
-    )
-
-
-def _map_provider(record: prisma.models.LlmProvider) -> llm_model.LlmProvider:
-    models: list[llm_model.LlmModel] = []
-    if record.Models:
-        models = [_map_model(model) for model in record.Models]
-
-    return llm_model.LlmProvider(
-        id=record.id,
-        name=record.name,
-        display_name=record.displayName,
-        description=record.description,
-        default_credential_provider=record.defaultCredentialProvider,
-        default_credential_id=record.defaultCredentialId,
-        default_credential_type=record.defaultCredentialType,
-        supports_tools=record.supportsTools,
-        supports_json_output=record.supportsJsonOutput,
-        supports_reasoning=record.supportsReasoning,
-        supports_parallel_tool=record.supportsParallelTool,
-        metadata=_json_dict(record.metadata),
-        models=models,
-    )
-
-
-async def list_providers(
-    include_models: bool = True, enabled_only: bool = False
-) -> list[llm_model.LlmProvider]:
-    """
-    List all LLM providers.
-
-    Args:
-        include_models: Whether to include models for each provider
-        enabled_only: If True, only include enabled models (for public routes)
-    """
-    include: Any = None
-    if include_models:
-        model_where = {"isEnabled": True} if enabled_only else None
-        include = {
-            "Models": {
-                "include": {"Costs": True, "Creator": True},
-                "where": model_where,
-            }
-        }
-    records = await prisma.models.LlmProvider.prisma().find_many(include=include)
-    return [_map_provider(record) for record in records]
-
-
-async def upsert_provider(
-    request: llm_model.UpsertLlmProviderRequest,
-    provider_id: str | None = None,
-) -> llm_model.LlmProvider:
-    data: Any = {
-        "name": request.name,
-        "displayName": request.display_name,
-        "description": request.description,
-        "defaultCredentialProvider": request.default_credential_provider,
-        "defaultCredentialId": request.default_credential_id,
-        "defaultCredentialType": request.default_credential_type,
-        "supportsTools": request.supports_tools,
-        "supportsJsonOutput": request.supports_json_output,
-        "supportsReasoning": request.supports_reasoning,
-        "supportsParallelTool": request.supports_parallel_tool,
-        "metadata": request.metadata,
-    }
-    include: Any = {"Models": {"include": {"Costs": True, "Creator": True}}}
-    if provider_id:
-        record = await prisma.models.LlmProvider.prisma().update(
-            where={"id": provider_id},
-            data=data,
-            include=include,
-        )
-    else:
-        record = await prisma.models.LlmProvider.prisma().create(
-            data=data,
-            include=include,
-        )
-    if record is None:
-        raise ValueError("Failed to create/update provider")
-    return _map_provider(record)
-
-
-async def list_models(
-    provider_id: str | None = None, enabled_only: bool = False
-) -> list[llm_model.LlmModel]:
-    """
-    List LLM models.
-
-    Args:
-        provider_id: Optional filter by provider ID
-        enabled_only: If True, only return enabled models (for public routes)
-    """
-    where: Any = {}
-    if provider_id:
-        where["providerId"] = provider_id
-    if enabled_only:
-        where["isEnabled"] = True
-
-    records = await prisma.models.LlmModel.prisma().find_many(
-        where=where if where else None,
-        include={"Costs": True, "Creator": True},
-    )
-    return [_map_model(record) for record in records]
-
-
-def _cost_create_payload(
-    costs: Sequence[llm_model.LlmModelCostInput],
-) -> dict[str, Iterable[dict[str, Any]]]:
-
-    create_items = []
-    for cost in costs:
-        item: dict[str, Any] = {
-            "unit": cost.unit,
-            "creditCost": cost.credit_cost,
-            "credentialProvider": cost.credential_provider,
-        }
-        # Only include optional fields if they have values
-        if cost.credential_id:
-            item["credentialId"] = cost.credential_id
-        if cost.credential_type:
-            item["credentialType"] = cost.credential_type
-        if cost.currency:
-            item["currency"] = cost.currency
-        # Handle metadata - use Prisma Json type
-        if cost.metadata is not None and cost.metadata != {}:
-            item["metadata"] = prisma.Json(cost.metadata)
-        create_items.append(item)
-    return {"create": create_items}
-
-
-async def create_model(
-    request: llm_model.CreateLlmModelRequest,
-) -> llm_model.LlmModel:
-    data: Any = {
-        "slug": request.slug,
-        "displayName": request.display_name,
-        "description": request.description,
-        "providerId": request.provider_id,
-        "contextWindow": request.context_window,
-        "maxOutputTokens": request.max_output_tokens,
-        "isEnabled": request.is_enabled,
-        "capabilities": request.capabilities,
-        "metadata": request.metadata,
-        "Costs": _cost_create_payload(request.costs),
-    }
-    if request.creator_id:
-        data["creatorId"] = request.creator_id
-
-    record = await prisma.models.LlmModel.prisma().create(
-        data=data,
-        include={"Costs": True, "Creator": True},
-    )
-    return _map_model(record)
-
-
-async def update_model(
-    model_id: str,
-    request: llm_model.UpdateLlmModelRequest,
-) -> llm_model.LlmModel:
-    # Build scalar field updates (non-relation fields)
-    scalar_data: Any = {}
-    if request.display_name is not None:
-        scalar_data["displayName"] = request.display_name
-    if request.description is not None:
-        scalar_data["description"] = request.description
-    if request.context_window is not None:
-        scalar_data["contextWindow"] = request.context_window
-    if request.max_output_tokens is not None:
-        scalar_data["maxOutputTokens"] = request.max_output_tokens
-    if request.is_enabled is not None:
-        scalar_data["isEnabled"] = request.is_enabled
-    if request.capabilities is not None:
-        scalar_data["capabilities"] = request.capabilities
-    if request.metadata is not None:
-        scalar_data["metadata"] = request.metadata
-    # Foreign keys can be updated directly as scalar fields
-    if request.provider_id is not None:
-        scalar_data["providerId"] = request.provider_id
-    if request.creator_id is not None:
-        # Empty string means remove the creator
-        scalar_data["creatorId"] = request.creator_id if request.creator_id else None
-
-    # If we have costs to update, we need to handle them separately
-    # because nested writes have different constraints
-    if request.costs is not None:
-        # Wrap cost replacement in a transaction for atomicity
-        async with transaction() as tx:
-            # First update scalar fields
-            if scalar_data:
-                await tx.llmmodel.update(
-                    where={"id": model_id},
-                    data=scalar_data,
-                )
-            # Then handle costs: delete existing and create new
-            await tx.llmmodelcost.delete_many(where={"llmModelId": model_id})
-            if request.costs:
-                cost_payload = _cost_create_payload(request.costs)
-                for cost_item in cost_payload["create"]:
-                    cost_item["llmModelId"] = model_id
-                    await tx.llmmodelcost.create(data=cast(Any, cost_item))
-        # Fetch the updated record (outside transaction)
-        record = await prisma.models.LlmModel.prisma().find_unique(
-            where={"id": model_id},
-            include={"Costs": True, "Creator": True},
-        )
-    else:
-        # No costs update - simple update
-        record = await prisma.models.LlmModel.prisma().update(
-            where={"id": model_id},
-            data=scalar_data,
-            include={"Costs": True, "Creator": True},
-        )
-
-    if not record:
-        raise ValueError(f"Model with id '{model_id}' not found")
-    return _map_model(record)
-
-
-async def toggle_model(
-    model_id: str,
-    is_enabled: bool,
-    migrate_to_slug: str | None = None,
-    migration_reason: str | None = None,
-    custom_credit_cost: int | None = None,
-) -> llm_model.ToggleLlmModelResponse:
-    """
-    Toggle a model's enabled status, optionally migrating workflows when disabling.
-
-    Args:
-        model_id: UUID of the model to toggle
-        is_enabled: New enabled status
-        migrate_to_slug: If disabling and this is provided, migrate all workflows
-                         using this model to the specified replacement model
-        migration_reason: Optional reason for the migration (e.g., "Provider outage")
-        custom_credit_cost: Optional custom pricing override for migrated workflows.
-                           When set, the billing system should use this cost instead
-                           of the target model's cost for affected nodes.
-
-    Returns:
-        ToggleLlmModelResponse with the updated model and optional migration stats
-    """
-    import json
-
-    # Get the model being toggled
-    model = await prisma.models.LlmModel.prisma().find_unique(
-        where={"id": model_id}, include={"Costs": True}
-    )
-    if not model:
-        raise ValueError(f"Model with id '{model_id}' not found")
-
-    nodes_migrated = 0
-    migration_id: str | None = None
-
-    # If disabling with migration, perform migration first
-    if not is_enabled and migrate_to_slug:
-        # Validate replacement model exists and is enabled
-        replacement = await prisma.models.LlmModel.prisma().find_unique(
-            where={"slug": migrate_to_slug}
-        )
-        if not replacement:
-            raise ValueError(f"Replacement model '{migrate_to_slug}' not found")
-        if not replacement.isEnabled:
-            raise ValueError(
-                f"Replacement model '{migrate_to_slug}' is disabled. "
-                f"Please enable it before using it as a replacement."
-            )
-
-        # Perform all operations atomically within a single transaction
-        # This ensures no nodes are missed between query and update
-        async with transaction() as tx:
-            # Get the IDs of nodes that will be migrated (inside transaction for consistency)
-            node_ids_result = await tx.query_raw(
-                """
-                SELECT id
-                FROM "AgentNode"
-                WHERE "constantInput"::jsonb->>'model' = $1
-                FOR UPDATE
-                """,
-                model.slug,
-            )
-            migrated_node_ids = (
-                [row["id"] for row in node_ids_result] if node_ids_result else []
-            )
-            nodes_migrated = len(migrated_node_ids)
-
-            if nodes_migrated > 0:
-                # Update by IDs to ensure we only update the exact nodes we queried
-                # Use JSON array and jsonb_array_elements_text for safe parameterization
-                node_ids_json = json.dumps(migrated_node_ids)
-                await tx.execute_raw(
-                    """
-                    UPDATE "AgentNode"
-                    SET "constantInput" = JSONB_SET(
-                        "constantInput"::jsonb,
-                        '{model}',
-                        to_jsonb($1::text)
-                    )
-                    WHERE id::text IN (
-                        SELECT jsonb_array_elements_text($2::jsonb)
-                    )
-                    """,
-                    migrate_to_slug,
-                    node_ids_json,
-                )
-
-            record = await tx.llmmodel.update(
-                where={"id": model_id},
-                data={"isEnabled": is_enabled},
-                include={"Costs": True},
-            )
-
-            # Create migration record for revert capability
-            if nodes_migrated > 0:
-                migration_data: Any = {
-                    "sourceModelSlug": model.slug,
-                    "targetModelSlug": migrate_to_slug,
-                    "reason": migration_reason,
-                    "migratedNodeIds": json.dumps(migrated_node_ids),
-                    "nodeCount": nodes_migrated,
-                    "customCreditCost": custom_credit_cost,
-                }
-                migration_record = await tx.llmmodelmigration.create(
-                    data=migration_data
-                )
-                migration_id = migration_record.id
-    else:
-        # Simple toggle without migration
-        record = await prisma.models.LlmModel.prisma().update(
-            where={"id": model_id},
-            data={"isEnabled": is_enabled},
-            include={"Costs": True},
-        )
-
-    if record is None:
-        raise ValueError(f"Model with id '{model_id}' not found")
-    return llm_model.ToggleLlmModelResponse(
-        model=_map_model(record),
-        nodes_migrated=nodes_migrated,
-        migrated_to_slug=migrate_to_slug if nodes_migrated > 0 else None,
-        migration_id=migration_id,
-    )
-
-
-async def get_model_usage(model_id: str) -> llm_model.LlmModelUsageResponse:
-    """Get usage count for a model."""
-    import prisma as prisma_module
-
-    model = await prisma.models.LlmModel.prisma().find_unique(where={"id": model_id})
-    if not model:
-        raise ValueError(f"Model with id '{model_id}' not found")
-
-    count_result = await prisma_module.get_client().query_raw(
-        """
-        SELECT COUNT(*) as count
-        FROM "AgentNode"
-        WHERE "constantInput"::jsonb->>'model' = $1
-        """,
-        model.slug,
-    )
-    node_count = int(count_result[0]["count"]) if count_result else 0
-
-    return llm_model.LlmModelUsageResponse(model_slug=model.slug, node_count=node_count)
-
-
-async def delete_model(
-    model_id: str, replacement_model_slug: str
-) -> llm_model.DeleteLlmModelResponse:
-    """
-    Delete a model and migrate all AgentNodes using it to a replacement model.
-
-    This performs an atomic operation within a database transaction:
-    1. Validates the model exists
-    2. Validates the replacement model exists and is enabled
-    3. Counts affected nodes
-    4. Migrates all AgentNode.constantInput->model to replacement (in transaction)
-    5. Deletes the LlmModel record (CASCADE deletes costs) (in transaction)
-
-    Args:
-        model_id: UUID of the model to delete
-        replacement_model_slug: Slug of the model to migrate to
-
-    Returns:
-        DeleteLlmModelResponse with migration stats
-
-    Raises:
-        ValueError: If model not found, replacement not found, or replacement is disabled
-    """
-    # 1. Get the model being deleted (validation - outside transaction)
-    model = await prisma.models.LlmModel.prisma().find_unique(
-        where={"id": model_id}, include={"Costs": True}
-    )
-    if not model:
-        raise ValueError(f"Model with id '{model_id}' not found")
-
-    deleted_slug = model.slug
-    deleted_display_name = model.displayName
-
-    # 2. Validate replacement model exists and is enabled (validation - outside transaction)
-    replacement = await prisma.models.LlmModel.prisma().find_unique(
-        where={"slug": replacement_model_slug}
-    )
-    if not replacement:
-        raise ValueError(f"Replacement model '{replacement_model_slug}' not found")
-    if not replacement.isEnabled:
-        raise ValueError(
-            f"Replacement model '{replacement_model_slug}' is disabled. "
-            f"Please enable it before using it as a replacement."
-        )
-
-    # 3 & 4. Perform count, migration and deletion atomically within a transaction
-    nodes_affected = 0
-    async with transaction() as tx:
-        # Count affected nodes (inside transaction for consistency)
-        count_result = await tx.query_raw(
-            """
-            SELECT COUNT(*) as count
-            FROM "AgentNode"
-            WHERE "constantInput"::jsonb->>'model' = $1
-            """,
-            deleted_slug,
-        )
-        nodes_affected = int(count_result[0]["count"]) if count_result else 0
-
-        # Migrate all AgentNode.constantInput->model to replacement
-        if nodes_affected > 0:
-            await tx.execute_raw(
-                """
-                UPDATE "AgentNode"
-                SET "constantInput" = JSONB_SET(
-                    "constantInput"::jsonb,
-                    '{model}',
-                    to_jsonb($1::text)
-                )
-                WHERE "constantInput"::jsonb->>'model' = $2
-                """,
-                replacement_model_slug,
-                deleted_slug,
-            )
-
-        # Delete the model (CASCADE will delete costs automatically)
-        await tx.llmmodel.delete(where={"id": model_id})
-
-    return llm_model.DeleteLlmModelResponse(
-        deleted_model_slug=deleted_slug,
-        deleted_model_display_name=deleted_display_name,
-        replacement_model_slug=replacement_model_slug,
-        nodes_migrated=nodes_affected,
-        message=(
-            f"Successfully deleted model '{deleted_display_name}' ({deleted_slug}) "
-            f"and migrated {nodes_affected} workflow node(s) to '{replacement_model_slug}'."
-        ),
-    )
-
-
-def _map_migration(
-    record: prisma.models.LlmModelMigration,
-) -> llm_model.LlmModelMigration:
-    return llm_model.LlmModelMigration(
-        id=record.id,
-        source_model_slug=record.sourceModelSlug,
-        target_model_slug=record.targetModelSlug,
-        reason=record.reason,
-        node_count=record.nodeCount,
-        custom_credit_cost=record.customCreditCost,
-        is_reverted=record.isReverted,
-        created_at=record.createdAt.isoformat(),
-        reverted_at=record.revertedAt.isoformat() if record.revertedAt else None,
-    )
-
-
-async def list_migrations(
-    include_reverted: bool = False,
-) -> list[llm_model.LlmModelMigration]:
-    """
-    List model migrations, optionally including reverted ones.
-
-    Args:
-        include_reverted: If True, include reverted migrations. Default is False.
-
-    Returns:
-        List of LlmModelMigration records
-    """
-    where: Any = None if include_reverted else {"isReverted": False}
-    records = await prisma.models.LlmModelMigration.prisma().find_many(
-        where=where,
-        order={"createdAt": "desc"},
-    )
-    return [_map_migration(record) for record in records]
-
-
-async def get_migration(migration_id: str) -> llm_model.LlmModelMigration | None:
-    """Get a specific migration by ID."""
-    record = await prisma.models.LlmModelMigration.prisma().find_unique(
-        where={"id": migration_id}
-    )
-    return _map_migration(record) if record else None
-
-
-async def revert_migration(
-    migration_id: str,
-    re_enable_source_model: bool = True,
-) -> llm_model.RevertMigrationResponse:
-    """
-    Revert a model migration, restoring affected nodes to their original model.
-
-    This only reverts the specific nodes that were migrated, not all nodes
-    currently using the target model.
-
-    Args:
-        migration_id: UUID of the migration to revert
-        re_enable_source_model: Whether to re-enable the source model if it's disabled
-
-    Returns:
-        RevertMigrationResponse with revert stats
-
-    Raises:
-        ValueError: If migration not found, already reverted, or source model not available
-    """
-    import json
-    from datetime import datetime, timezone
-
-    # Get the migration record
-    migration = await prisma.models.LlmModelMigration.prisma().find_unique(
-        where={"id": migration_id}
-    )
-    if not migration:
-        raise ValueError(f"Migration with id '{migration_id}' not found")
-
-    if migration.isReverted:
-        raise ValueError(
-            f"Migration '{migration_id}' has already been reverted "
-            f"on {migration.revertedAt.isoformat() if migration.revertedAt else 'unknown date'}"
-        )
-
-    # Check if source model exists
-    source_model = await prisma.models.LlmModel.prisma().find_unique(
-        where={"slug": migration.sourceModelSlug}
-    )
-    if not source_model:
-        raise ValueError(
-            f"Source model '{migration.sourceModelSlug}' no longer exists. "
-            f"Cannot revert migration."
-        )
-
-    # Get the migrated node IDs (Prisma auto-parses JSONB to list)
-    migrated_node_ids: list[str] = (
-        migration.migratedNodeIds
-        if isinstance(migration.migratedNodeIds, list)
-        else json.loads(migration.migratedNodeIds)  # type: ignore
-    )
-    if not migrated_node_ids:
-        raise ValueError("No nodes to revert in this migration")
-
-    # Track if we need to re-enable the source model
-    source_model_was_disabled = not source_model.isEnabled
-    should_re_enable = source_model_was_disabled and re_enable_source_model
-    source_model_re_enabled = False
-
-    # Perform revert atomically
-    async with transaction() as tx:
-        # Re-enable the source model if requested and it was disabled
-        if should_re_enable:
-            await tx.llmmodel.update(
-                where={"id": source_model.id},
-                data={"isEnabled": True},
-            )
-            source_model_re_enabled = True
-
-        # Update only the specific nodes that were migrated
-        # We need to check that they still have the target model (haven't been changed since)
-        # Use a single batch update for efficiency
-        # Use JSON array and jsonb_array_elements_text for safe parameterization
-        node_ids_json = json.dumps(migrated_node_ids)
-        result = await tx.execute_raw(
-            """
-            UPDATE "AgentNode"
-            SET "constantInput" = JSONB_SET(
-                "constantInput"::jsonb,
-                '{model}',
-                to_jsonb($1::text)
-            )
-            WHERE id::text IN (
-                SELECT jsonb_array_elements_text($2::jsonb)
-            )
-            AND "constantInput"::jsonb->>'model' = $3
-            """,
-            migration.sourceModelSlug,
-            node_ids_json,
-            migration.targetModelSlug,
-        )
-        nodes_reverted = result if result else 0
-
-        # Mark migration as reverted
-        await tx.llmmodelmigration.update(
-            where={"id": migration_id},
-            data={
-                "isReverted": True,
-                "revertedAt": datetime.now(timezone.utc),
-            },
-        )
-
-    # Calculate nodes that were already changed since migration
-    nodes_already_changed = len(migrated_node_ids) - nodes_reverted
-
-    # Build appropriate message
-    message_parts = [
-        f"Successfully reverted migration: {nodes_reverted} node(s) restored "
-        f"from '{migration.targetModelSlug}' to '{migration.sourceModelSlug}'."
-    ]
-    if nodes_already_changed > 0:
-        message_parts.append(
-            f" {nodes_already_changed} node(s) were already changed and not reverted."
-        )
-    if source_model_re_enabled:
-        message_parts.append(
-            f" Model '{migration.sourceModelSlug}' has been re-enabled."
-        )
-
-    return llm_model.RevertMigrationResponse(
-        migration_id=migration_id,
-        source_model_slug=migration.sourceModelSlug,
-        target_model_slug=migration.targetModelSlug,
-        nodes_reverted=nodes_reverted,
-        nodes_already_changed=nodes_already_changed,
-        source_model_re_enabled=source_model_re_enabled,
-        message="".join(message_parts),
-    )
-
-
-# ============================================================================
-# Creator CRUD operations
-# ============================================================================
-
-
-async def list_creators() -> list[llm_model.LlmModelCreator]:
-    """List all LLM model creators."""
-    records = await prisma.models.LlmModelCreator.prisma().find_many(
-        order={"displayName": "asc"}
-    )
-    return [_map_creator(record) for record in records]
-
-
-async def get_creator(creator_id: str) -> llm_model.LlmModelCreator | None:
-    """Get a specific creator by ID."""
-    record = await prisma.models.LlmModelCreator.prisma().find_unique(
-        where={"id": creator_id}
-    )
-    return _map_creator(record) if record else None
-
-
-async def upsert_creator(
-    request: llm_model.UpsertLlmCreatorRequest,
-    creator_id: str | None = None,
-) -> llm_model.LlmModelCreator:
-    """Create or update a model creator."""
-    data: Any = {
-        "name": request.name,
-        "displayName": request.display_name,
-        "description": request.description,
-        "websiteUrl": request.website_url,
-        "logoUrl": request.logo_url,
-        "metadata": request.metadata,
-    }
-    if creator_id:
-        record = await prisma.models.LlmModelCreator.prisma().update(
-            where={"id": creator_id},
-            data=data,
-        )
-    else:
-        record = await prisma.models.LlmModelCreator.prisma().create(data=data)
-    if record is None:
-        raise ValueError("Failed to create/update creator")
-    return _map_creator(record)
-
-
-async def delete_creator(creator_id: str) -> bool:
-    """
-    Delete a model creator.
-
-    This will set creatorId to NULL on all associated models (due to onDelete: SetNull).
-
-    Args:
-        creator_id: UUID of the creator to delete
-
-    Returns:
-        True if deleted successfully
-
-    Raises:
-        ValueError: If creator not found
-    """
-    creator = await prisma.models.LlmModelCreator.prisma().find_unique(
-        where={"id": creator_id}
-    )
-    if not creator:
-        raise ValueError(f"Creator with id '{creator_id}' not found")
-
-    await prisma.models.LlmModelCreator.prisma().delete(where={"id": creator_id})
-    return True
-
-
-async def get_recommended_model() -> llm_model.LlmModel | None:
-    """
-    Get the currently recommended LLM model.
-
-    Returns:
-        The recommended model, or None if no model is marked as recommended.
-    """
-    record = await prisma.models.LlmModel.prisma().find_first(
-        where={"isRecommended": True, "isEnabled": True},
-        include={"Costs": True, "Creator": True},
-    )
-    return _map_model(record) if record else None
-
-
-async def set_recommended_model(
-    model_id: str,
-) -> tuple[llm_model.LlmModel, str | None]:
-    """
-    Set a model as the recommended model.
-
-    This will clear the isRecommended flag from any other model and set it
-    on the specified model. The model must be enabled.
-
-    Args:
-        model_id: UUID of the model to set as recommended
-
-    Returns:
-        Tuple of (the updated model, previous recommended model slug or None)
-
-    Raises:
-        ValueError: If model not found or not enabled
-    """
-    # First, verify the model exists and is enabled
-    target_model = await prisma.models.LlmModel.prisma().find_unique(
-        where={"id": model_id}
-    )
-    if not target_model:
-        raise ValueError(f"Model with id '{model_id}' not found")
-    if not target_model.isEnabled:
-        raise ValueError(
-            f"Cannot set disabled model '{target_model.slug}' as recommended"
-        )
-
-    # Get the current recommended model (if any)
-    current_recommended = await prisma.models.LlmModel.prisma().find_first(
-        where={"isRecommended": True}
-    )
-    previous_slug = current_recommended.slug if current_recommended else None
-
-    # Use a transaction to ensure atomicity
-    async with transaction() as tx:
-        # Clear isRecommended from all models
-        await tx.llmmodel.update_many(
-            where={"isRecommended": True},
-            data={"isRecommended": False},
-        )
-        # Set the new recommended model
-        await tx.llmmodel.update(
-            where={"id": model_id},
-            data={"isRecommended": True},
-        )
-
-    # Fetch and return the updated model
-    updated_record = await prisma.models.LlmModel.prisma().find_unique(
-        where={"id": model_id},
-        include={"Costs": True, "Creator": True},
-    )
-    if not updated_record:
-        raise ValueError("Failed to fetch updated model")
-
-    return _map_model(updated_record), previous_slug
-
-
-async def get_recommended_model_slug() -> str | None:
-    """
-    Get the slug of the currently recommended LLM model.
-
-    Returns:
-        The slug of the recommended model, or None if no model is marked as recommended.
-    """
-    record = await prisma.models.LlmModel.prisma().find_first(
-        where={"isRecommended": True, "isEnabled": True},
-    )
-    return record.slug if record else None

autogpt_platform/backend/backend/server/v2/llm/model.py

@@ -1,231 +0,0 @@
-from __future__ import annotations
-
-import re
-from typing import Any, Optional
-
-import prisma.enums
-import pydantic
-
-# Pattern for valid model slugs: alphanumeric start, then alphanumeric, dots, underscores, slashes, hyphens
-SLUG_PATTERN = re.compile(r"^[a-zA-Z0-9][a-zA-Z0-9._/-]*$")
-
-
-class LlmModelCost(pydantic.BaseModel):
-    id: str
-    unit: prisma.enums.LlmCostUnit = prisma.enums.LlmCostUnit.RUN
-    credit_cost: int
-    credential_provider: str
-    credential_id: Optional[str] = None
-    credential_type: Optional[str] = None
-    currency: Optional[str] = None
-    metadata: dict[str, Any] = pydantic.Field(default_factory=dict)
-
-
-class LlmModelCreator(pydantic.BaseModel):
-    """Represents the organization that created/trained the model (e.g., OpenAI, Meta)."""
-
-    id: str
-    name: str
-    display_name: str
-    description: Optional[str] = None
-    website_url: Optional[str] = None
-    logo_url: Optional[str] = None
-    metadata: dict[str, Any] = pydantic.Field(default_factory=dict)
-
-
-class LlmModel(pydantic.BaseModel):
-    id: str
-    slug: str
-    display_name: str
-    description: Optional[str] = None
-    provider_id: str
-    creator_id: Optional[str] = None
-    creator: Optional[LlmModelCreator] = None
-    context_window: int
-    max_output_tokens: Optional[int] = None
-    is_enabled: bool = True
-    is_recommended: bool = False
-    capabilities: dict[str, Any] = pydantic.Field(default_factory=dict)
-    metadata: dict[str, Any] = pydantic.Field(default_factory=dict)
-    costs: list[LlmModelCost] = pydantic.Field(default_factory=list)
-
-
-class LlmProvider(pydantic.BaseModel):
-    id: str
-    name: str
-    display_name: str
-    description: Optional[str] = None
-    default_credential_provider: Optional[str] = None
-    default_credential_id: Optional[str] = None
-    default_credential_type: Optional[str] = None
-    supports_tools: bool = True
-    supports_json_output: bool = True
-    supports_reasoning: bool = False
-    supports_parallel_tool: bool = False
-    metadata: dict[str, Any] = pydantic.Field(default_factory=dict)
-    models: list[LlmModel] = pydantic.Field(default_factory=list)
-
-
-class LlmProvidersResponse(pydantic.BaseModel):
-    providers: list[LlmProvider]
-
-
-class LlmModelsResponse(pydantic.BaseModel):
-    models: list[LlmModel]
-
-
-class LlmCreatorsResponse(pydantic.BaseModel):
-    creators: list[LlmModelCreator]
-
-
-class UpsertLlmProviderRequest(pydantic.BaseModel):
-    name: str
-    display_name: str
-    description: Optional[str] = None
-    default_credential_provider: Optional[str] = None
-    default_credential_id: Optional[str] = None
-    default_credential_type: Optional[str] = "api_key"
-    supports_tools: bool = True
-    supports_json_output: bool = True
-    supports_reasoning: bool = False
-    supports_parallel_tool: bool = False
-    metadata: dict[str, Any] = pydantic.Field(default_factory=dict)
-
-
-class UpsertLlmCreatorRequest(pydantic.BaseModel):
-    name: str
-    display_name: str
-    description: Optional[str] = None
-    website_url: Optional[str] = None
-    logo_url: Optional[str] = None
-    metadata: dict[str, Any] = pydantic.Field(default_factory=dict)
-
-
-class LlmModelCostInput(pydantic.BaseModel):
-    unit: prisma.enums.LlmCostUnit = prisma.enums.LlmCostUnit.RUN
-    credit_cost: int
-    credential_provider: str
-    credential_id: Optional[str] = None
-    credential_type: Optional[str] = "api_key"
-    currency: Optional[str] = None
-    metadata: dict[str, Any] = pydantic.Field(default_factory=dict)
-
-
-class CreateLlmModelRequest(pydantic.BaseModel):
-    slug: str
-    display_name: str
-    description: Optional[str] = None
-    provider_id: str
-    creator_id: Optional[str] = None
-    context_window: int
-    max_output_tokens: Optional[int] = None
-    is_enabled: bool = True
-    capabilities: dict[str, Any] = pydantic.Field(default_factory=dict)
-    metadata: dict[str, Any] = pydantic.Field(default_factory=dict)
-    costs: list[LlmModelCostInput]
-
-    @pydantic.field_validator("slug")
-    @classmethod
-    def validate_slug(cls, v: str) -> str:
-        if not v or len(v) > 100:
-            raise ValueError("Slug must be 1-100 characters")
-        if not SLUG_PATTERN.match(v):
-            raise ValueError(
-                "Slug must start with alphanumeric and contain only "
-                "alphanumeric characters, dots, underscores, slashes, or hyphens"
-            )
-        return v
-
-
-class UpdateLlmModelRequest(pydantic.BaseModel):
-    display_name: Optional[str] = None
-    description: Optional[str] = None
-    context_window: Optional[int] = None
-    max_output_tokens: Optional[int] = None
-    is_enabled: Optional[bool] = None
-    capabilities: Optional[dict[str, Any]] = None
-    metadata: Optional[dict[str, Any]] = None
-    provider_id: Optional[str] = None
-    creator_id: Optional[str] = None
-    costs: Optional[list[LlmModelCostInput]] = None
-
-
-class ToggleLlmModelRequest(pydantic.BaseModel):
-    is_enabled: bool
-    migrate_to_slug: Optional[str] = None
-    migration_reason: Optional[str] = None  # e.g., "Provider outage"
-    # Custom pricing override for migrated workflows. When set, billing should use
-    # this cost instead of the target model's cost for affected nodes.
-    # See LlmModelMigration in schema.prisma for full documentation.
-    custom_credit_cost: Optional[int] = None
-
-
-class ToggleLlmModelResponse(pydantic.BaseModel):
-    model: LlmModel
-    nodes_migrated: int = 0
-    migrated_to_slug: Optional[str] = None
-    migration_id: Optional[str] = None  # ID of the migration record for revert
-
-
-class DeleteLlmModelResponse(pydantic.BaseModel):
-    deleted_model_slug: str
-    deleted_model_display_name: str
-    replacement_model_slug: str
-    nodes_migrated: int
-    message: str
-
-
-class LlmModelUsageResponse(pydantic.BaseModel):
-    model_slug: str
-    node_count: int
-
-
-# Migration tracking models
-class LlmModelMigration(pydantic.BaseModel):
-    id: str
-    source_model_slug: str
-    target_model_slug: str
-    reason: Optional[str] = None
-    node_count: int
-    # Custom pricing override - billing should use this instead of target model's cost
-    custom_credit_cost: Optional[int] = None
-    is_reverted: bool = False
-    created_at: str  # ISO datetime string
-    reverted_at: Optional[str] = None
-
-
-class LlmMigrationsResponse(pydantic.BaseModel):
-    migrations: list[LlmModelMigration]
-
-
-class RevertMigrationRequest(pydantic.BaseModel):
-    re_enable_source_model: bool = (
-        True  # Whether to re-enable the source model if disabled
-    )
-
-
-class RevertMigrationResponse(pydantic.BaseModel):
-    migration_id: str
-    source_model_slug: str
-    target_model_slug: str
-    nodes_reverted: int
-    nodes_already_changed: int = (
-        0  # Nodes that were modified since migration (not reverted)
-    )
-    source_model_re_enabled: bool = False  # Whether the source model was re-enabled
-    message: str
-
-
-class SetRecommendedModelRequest(pydantic.BaseModel):
-    model_id: str
-
-
-class SetRecommendedModelResponse(pydantic.BaseModel):
-    model: LlmModel
-    previous_recommended_slug: Optional[str] = None
-    message: str
-
-
-class RecommendedModelResponse(pydantic.BaseModel):
-    model: Optional[LlmModel] = None
-    slug: Optional[str] = None

autogpt_platform/backend/backend/server/v2/llm/routes.py

@@ -1,25 +0,0 @@
-import autogpt_libs.auth
-import fastapi
-
-from backend.server.v2.llm import db as llm_db
-from backend.server.v2.llm import model as llm_model
-
-router = fastapi.APIRouter(
-    prefix="/llm",
-    tags=["llm"],
-    dependencies=[fastapi.Security(autogpt_libs.auth.requires_user)],
-)
-
-
-@router.get("/models", response_model=llm_model.LlmModelsResponse)
-async def list_models():
-    """List all enabled LLM models available to users."""
-    models = await llm_db.list_models(enabled_only=True)
-    return llm_model.LlmModelsResponse(models=models)
-
-
-@router.get("/providers", response_model=llm_model.LlmProvidersResponse)
-async def list_providers():
-    """List all LLM providers with their enabled models."""
-    providers = await llm_db.list_providers(include_models=True, enabled_only=True)
-    return llm_model.LlmProvidersResponse(providers=providers)

autogpt_platform/backend/backend/util/cache.py

@@ -16,7 +16,7 @@ import pickle
 import threading
 import time
 from dataclasses import dataclass
-from functools import wraps
+from functools import cache, wraps
 from typing import Any, Callable, ParamSpec, Protocol, TypeVar, cast, runtime_checkable
 
 from redis import ConnectionPool, Redis
@@ -38,29 +38,34 @@ settings = Settings()
 #   maxmemory 2gb                   # Set memory limit (adjust based on your needs)
 #   save ""                         # Disable persistence if using Redis purely for caching
 
-# Create a dedicated Redis connection pool for caching (binary mode for pickle)
-_cache_pool: ConnectionPool | None = None
 
-
-@conn_retry("Redis", "Acquiring cache connection pool")
+@cache
 def _get_cache_pool() -> ConnectionPool:
-    """Get or create a connection pool for cache operations."""
-    global _cache_pool
-    if _cache_pool is None:
-        _cache_pool = ConnectionPool(
-            host=settings.config.redis_host,
-            port=settings.config.redis_port,
-            password=settings.config.redis_password or None,
-            decode_responses=False,  # Binary mode for pickle
-            max_connections=50,
-            socket_keepalive=True,
-            socket_connect_timeout=5,
-            retry_on_timeout=True,
-        )
-    return _cache_pool
+    """Get or create a connection pool for cache operations (lazy, thread-safe)."""
+    return ConnectionPool(
+        host=settings.config.redis_host,
+        port=settings.config.redis_port,
+        password=settings.config.redis_password or None,
+        decode_responses=False,  # Binary mode for pickle
+        max_connections=50,
+        socket_keepalive=True,
+        socket_connect_timeout=5,
+        retry_on_timeout=True,
+    )
 
 
-redis = Redis(connection_pool=_get_cache_pool())
+@cache
+@conn_retry("Redis", "Acquiring cache connection")
+def _get_redis() -> Redis:
+    """
+    Get the lazily-initialized Redis client for shared cache operations.
+    Uses @cache for thread-safe singleton behavior - connection is only
+    established when first accessed, allowing services that only use
+    in-memory caching to work without Redis configuration.
+    """
+    r = Redis(connection_pool=_get_cache_pool())
+    r.ping()  # Verify connection
+    return r
 
 
 @dataclass
@@ -179,9 +184,9 @@ def cached(
             try:
                 if refresh_ttl_on_get:
                     # Use GETEX to get value and refresh expiry atomically
-                    cached_bytes = redis.getex(redis_key, ex=ttl_seconds)
+                    cached_bytes = _get_redis().getex(redis_key, ex=ttl_seconds)
                 else:
-                    cached_bytes = redis.get(redis_key)
+                    cached_bytes = _get_redis().get(redis_key)
 
                 if cached_bytes and isinstance(cached_bytes, bytes):
                     return pickle.loads(cached_bytes)
@@ -195,7 +200,7 @@ def cached(
             """Set value in Redis with TTL."""
             try:
                 pickled_value = pickle.dumps(value, protocol=pickle.HIGHEST_PROTOCOL)
-                redis.setex(redis_key, ttl_seconds, pickled_value)
+                _get_redis().setex(redis_key, ttl_seconds, pickled_value)
             except Exception as e:
                 logger.error(
                     f"Redis error storing cache for {target_func.__name__}: {e}"
@@ -333,14 +338,18 @@ def cached(
                 if pattern:
                     # Clear entries matching pattern
                     keys = list(
-                        redis.scan_iter(f"cache:{target_func.__name__}:{pattern}")
+                        _get_redis().scan_iter(
+                            f"cache:{target_func.__name__}:{pattern}"
+                        )
                     )
                 else:
                     # Clear all cache keys
-                    keys = list(redis.scan_iter(f"cache:{target_func.__name__}:*"))
+                    keys = list(
+                        _get_redis().scan_iter(f"cache:{target_func.__name__}:*")
+                    )
 
                 if keys:
-                    pipeline = redis.pipeline()
+                    pipeline = _get_redis().pipeline()
                     for key in keys:
                         pipeline.delete(key)
                     pipeline.execute()
@@ -355,7 +364,9 @@ def cached(
 
         def cache_info() -> dict[str, int | None]:
             if shared_cache:
-                cache_keys = list(redis.scan_iter(f"cache:{target_func.__name__}:*"))
+                cache_keys = list(
+                    _get_redis().scan_iter(f"cache:{target_func.__name__}:*")
+                )
                 return {
                     "size": len(cache_keys),
                     "maxsize": None,  # Redis manages its own size
@@ -373,10 +384,8 @@ def cached(
             key = _make_hashable_key(args, kwargs)
             if shared_cache:
                 redis_key = _make_redis_key(key, target_func.__name__)
-                if redis.exists(redis_key):
-                    redis.delete(redis_key)
-                    return True
-                return False
+                deleted_count = cast(int, _get_redis().delete(redis_key))
+                return deleted_count > 0
             else:
                 if key in cache_storage:
                     del cache_storage[key]