Fixes duplicate image files in generations folder when Gemini returns JPEG.
Changes:
- POST handler now extracts MIME type from data URL and uses correct extension
- GET handler searches multiple extensions (png, jpg, jpeg, gif, webp)
- GET handler returns correct MIME type in data URL based on file extension
Root cause: workflow-images/route.ts hardcoded .png extension regardless of
actual image format, while save-generation/route.ts correctly preserved formats.
This caused Gemini's JPEG images to be saved twice (once as .jpg, once as .png).
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
- Remove unused `useLoader` import from GLBViewerNode
- Add useEffect cleanup to revoke blob URL on node unmount (memory leak)
- Replace alert() with useToast for file validation errors
- Lazy-load GLBViewerNode via next/dynamic to avoid bundling three.js for all users
- Add glbViewer to ConnectionDropMenu IMAGE_SOURCE_OPTIONS
- Add glbViewer case to quickstart validation createDefaultNodeData (build fix)
- Restore trailing newline in gemini.ts
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- fal.ts: Use model capabilities to determine fallback content-type
(video/mp4 vs image/png) instead of defaulting to video for all models
- replicate.ts: Increase polling timeout to 10 min for video models;
filter output array to valid strings before URL validation
- nanoBananaExecutor.ts: Replace read-modify-write on outputGallery with
atomic appendOutputGalleryImage using Zustand set() callback
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Providers:
- replicate: add SSRF validation on mediaUrl before fetch
- fal/kie/replicate/wavespeed: large-video returns data="" with url set
instead of putting raw URL in data field (violates GenerationOutput.data semantic)
- kie: remove stale WaveSpeed JSDoc block from kie provider file
- route.ts: update all output handling to support data="" + url for large videos
Executors:
- generateVideoExecutor/nanoBananaExecutor: track save-generation fetch in
pendingImageSyncs via new trackSaveGeneration context method so auto-save
waits for server ID resolution
- videoProcessingExecutors: throw errors instead of returning in validation
branches (encoderSupported, insufficient videos, no video connected) so
Promise.allSettled sees rejections
- types.ts: add trackSaveGeneration to NodeExecutionContext interface
- workflowStore: wire trackSaveGeneration into both execution context sites
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move generateWithFalQueue, generateWithKie, generateWithWaveSpeed and all
their helpers into providers/fal.ts, providers/kie.ts, providers/wavespeed.ts.
Add barrel export at providers/index.ts.
route.ts reduced from 2,838 to 518 lines (POST handler + routing only).
Re-export clearFalInputMappingCache from route.ts for test backward compat.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move generateWithReplicate() into providers/replicate.ts.
It imports schema utilities from schemaUtils.ts.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move MODEL_MAP and generateWithGemini() into providers/gemini.ts.
Remove unused imports (GoogleGenAI, uploadImageForUrl, shouldUseImageUrl,
deleteImages, ProviderModel) from route.ts.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Extract INPUT_PATTERNS, InputMapping, ParameterTypeInfo, getParameterTypesFromSchema,
coerceParameterTypes, and getInputMappingFromSchema into schemaUtils.ts with 23 new tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Validate uploadUrl and fileUrl from fal CDN initiate response through
validateMediaUrl() and HTTPS check before performing the PUT request.
Throws if either URL is missing or fails validation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Validate fal-provided status_url and response_url against validateMediaUrl()
and domain allowlist (queue.fal.run) before using them for polling. Falls back
to constructed URLs on validation failure.
Also updates mockFalCdnUpload test helper to match the two-step CDN upload
flow (initiate + PUT) introduced in 3fea13e.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
WaveSpeed's CloudFront CDN returns `binary/octet-stream` which wasn't
caught by the old `!== "application/octet-stream"` check, producing
unplayable data URLs. Now only accepts content types starting with
`video/` or `image/`, falling back to `video/mp4` or `image/png`.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The old single-POST endpoint (fal.ai/api/storage/upload) returns 404.
Switched to the current fal.ai upload API: initiate via
rest.alpha.fal.ai/storage/upload/initiate, then PUT binary data to
the returned signed URL.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Delete unused generateWithFal function (replaced by generateWithFalQueue)
- Add 20MB size guard in uploadImageToFal to prevent memory spikes
- Add refresh button to ModelSearchDialog that clears model and schema
localStorage caches plus in-memory fetch cache
- Wire "Try Again" error button to also clear stale caches
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace blocking fal.run endpoint with async queue.fal.run + 1s polling
to match Replicate/WaveSpeed behavior. Upload base64 images to fal CDN
before queue submission to avoid large inline payloads. Cache schema
mappings in-memory (30min TTL) to eliminate redundant API calls per
generation. Also fix localStorage cache leaking between tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Derive provider capabilities from mediaType instead of hardcoding
text-to-image, so isVideoModel is correct for video generations.
Treat application/octet-stream as unhelpful content-type and fall back
to video/mp4 or image/png based on model type. Add safety net in
save-generation to override .bin extension when media type is known.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add Veo 3 and Veo 3 Fast models (text-to-video + image-to-video) with
custom API integration using Kie's /api/v1/veo/ endpoints and numeric
successFlag polling. Also document Kie model integration SOP in CLAUDE.md.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Refactor executeSingleNode error paths to throw instead of setting global
isRunning/currentNodeIds state, which corrupted sibling nodes running in
parallel via Promise.allSettled
- Validate WaveSpeed providedPollUrl with SSRF check before use
- Sanitize WaveSpeed taskId in provider with path-traversal validation
- Revoke blob URLs (URL.revokeObjectURL) before replacing outputVideo to
prevent memory leaks from large video files
- Move pause edge cleanup to outer executeWorkflow loop
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add validateMediaUrl() utility to block private IPs, localhost, and
non-HTTP protocols before server-side fetches (with tests)
- URL-encode Kie taskId in poll URL to prevent injection
- Validate WaveSpeed modelId against path traversal in both route.ts
and wavespeed.ts provider
- Add 500MB Content-Length check before downloading media at all 3
fetch sites (Kie, WaveSpeed route, WaveSpeed provider)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Update stale test mocks from currentNodeId to currentNodeIds in
AudioInputNode, EaseCurveNode, and VideoStitchNode tests. Replace
hard 401 return for missing FAL API key with console.warn, allowing
requests to proceed without auth (rate-limited) since
generateWithFal() already handles null keys gracefully.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add nano-banana-pro model to Kie provider with schema and image input key
- Reorder Kie generation to process dynamicInputs before fallback image
uploads, preventing double-uploads when schema-mapped connections exist
- Update test mocks for deduplicatedFetch, useProviderApiKeys, and static
handle changes across GenerateImageNode, GenerateVideoNode,
ModelParameters, and ModelSearchDialog tests
- Add test commands to CLAUDE.md
- Fix SplitGridSettingsModal grid preview count (5 → 6 options)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Remove 36 unsupported Kie.ai models (14 image, 22 video) and add 2 new
ones (kling-2.6/motion-control, topaz/video-upscale), bringing the total
to 22 models. Fix parameter schemas to match official Kie.ai docs:
seedream quality param, flux-2 input_urls, grok-imagine video mode/duration,
kling sound param, wan resolution param, and singular vs array image keys.
Remove dead Veo/Kontext/Runway helper functions and endpoint branching —
all remaining models use the standard createTask endpoint.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add ~38 new image/video models (Imagen 4, FLUX.2, Flux Kontext,
Grok Imagine, Ideogram, Qwen, Runway, Hailuo, Seedance variants,
Kling older, Wan older) with schemas and endpoint routing for
dedicated APIs (Flux Kontext, Runway). Add pageUrl to all Kie
models and update ModelSearchDialog to use it.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Added validation to ensure fal.ai API key is provided in requests, returning a 401 status with an appropriate error message if not configured.
- Updated related tests to reflect the new API key requirement and adjusted assertions for expected responses.
- Modified documentation to clarify that fal.ai API key is now mandatory for model access.
GPT Image 1.5 only supports 'aspect_ratio', not 'size'.
Remove stale 'size' values from old workflow data before sending.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Fix Kie.ai GPT Image 1.5 failing to retrieve output URL by parsing
resultJson as JSON string (was treated as object incorrectly)
- Add 1x5 grid option to SplitGridNode for vertical contact sheets
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Import extractSubgraph from subgraphExtractor
- Parse selectedNodeIds from request body
- Call extractSubgraph before building workflow context
- Pass subgraph.selectedNodes and selectedEdges to buildWorkflowContext
- Update buildEditSystemPrompt to accept optional restSummary parameter
- Add WORKFLOW CONTEXT (SELECTED SUBSET) section when scoped
- Include rest-of-workflow summary with type breakdown and boundary connections
- Add metadata placeholder note to system prompt
- Enhance error handling to return 413 for oversized payloads
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Broaden the trailing-slash exemption to recognize both C:\ and C:/ as
drive roots. Tighten the client-side absolute path check to require a
separator after the drive letter and accept UNC paths (\\server\share).
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Rewrite /api/chat route to accept workflow context from client and use
AI SDK tool calling for multi-modal agent routing.
Changes:
- Accept optional workflowState (nodes/edges) in request body
- Build workflow context using buildWorkflowContext from contextBuilder
- Generate context-aware system prompt using buildEditSystemPrompt
- Create chat tools (answerQuestion, createWorkflow, editWorkflow) with
current node IDs for validation
- Configure streamText with tools, toolChoice: 'auto', and stopWhen for
multi-step reasoning (up to 3 steps)
- Stream results back via toUIMessageStreamResponse for useChat hook
compatibility
LLM now automatically routes to the correct tool based on user intent.
Tool results stream back to client for application in Plan 03.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
On macOS, osascript's folder picker can return hostname-prefixed paths
(e.g. "HOSTNAME/Users/...") for network-mounted volumes instead of
absolute POSIX paths, causing directory validation to fail.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Update dynamicInputs type from Record<string, string> to Record<string, string | string[]> in 4 locations
- Fix getConnectedInputs to aggregate same-key values into arrays instead of overwriting
- Fix dynamicInputs.prompt access to handle array case
- Fix hasPrompt check in POST handler to handle array-typed prompt
- Update processedDynamicInputs type in both Replicate and fal blocks
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Update dynamicInputs type from Record<string, string> to Record<string, string | string[]> in 4 locations
- Fix getConnectedInputs to aggregate same-key values into arrays instead of overwriting
- Fix dynamicInputs.prompt access to handle array case
- Fix hasPrompt check in POST handler to handle array-typed prompt
- Update processedDynamicInputs type in both Replicate and fal blocks
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The proxy was downloading 80-275MB workflow files through a serverless
function, causing timeouts and failures. Now the proxy just returns the
presigned R2 URL and the client downloads directly from R2.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Kie API returns HTTP 200 even on errors, with error code in the JSON body.
Now properly checks createResult.code !== 200 and returns the actual error
message from the API instead of generic "No task ID in response".
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Some Kie.ai models (like gpt-image/1.5-text-to-image) require certain
parameters to be present even if not user-specified (e.g., aspect_ratio,
quality). Add getKieModelDefaults() function that provides model-specific
defaults which are merged with user-provided parameters.
Also skip file logging on Vercel since filesystem is read-only.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Use multipart/form-data for file upload with 'file' and 'uploadPath' fields
- Extract download URL from data.downloadUrl in upload response
- Wrap parameters in 'input' object for standard models
- Use flat structure with camelCase for Veo models
- Extract results from resultJson.resultUrls in poll response
Co-Authored-By: Claude <noreply@anthropic.com>
- Add Kie.ai section in provider settings with API key input
- Add kie to env-status API for .env detection
Co-Authored-By: Claude <noreply@anthropic.com>
Add Kie.ai as a new provider for image and video generation:
Image models (8):
- z-image, seedream/4.5-text-to-image, seedream/4.5-edit
- nano-banana-pro, gpt-image/1.5-text-to-image, gpt-image/1.5-image-to-image
- google/nano-banana, google/nano-banana-edit
Video models (11):
- sora-2-text-to-video, sora-2-image-to-video
- sora-2-pro-text-to-video, sora-2-pro-image-to-video
- veo3_fast, veo3, bytedance/seedance-1.5-pro
- kling-2.6/text-to-video, kling-2.6/image-to-video
- wan/2-6-text-to-video, wan/2-6-image-to-video
Key changes:
- Add "kie" to ProviderType union
- Add generateWithKie() with image upload and task polling
- Add hardcoded model list and schemas (no discovery API)
- Update workflow store for API key handling
- Update UI components with Kie provider badge
Co-Authored-By: Claude <noreply@anthropic.com>
Rewrite chat system prompt to be concise — bullet points over
paragraphs, 2-4 sentences per response. Make chat panel stretch
from top-16 to bottom-[220px] for near-full canvas height.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Import convertToModelMessages and UIMessage from ai
- Convert incoming messages for useChat hook compatibility
- Return toUIMessageStreamResponse() instead of toTextStreamResponse()
- POST /api/chat with Vercel AI SDK streamText
- Node Banana domain expertise in system prompt
- Conversational workflow planning assistant
- Uses createGoogleGenerativeAI with GEMINI_API_KEY
Add POST endpoint that generates WorkflowProposal JSON from a description:
- Validates description (min 3 chars)
- Calls Gemini with buildProposalPrompt
- Parses JSON with repair logic from validation.ts
- Validates response matches WorkflowProposal structure
- Returns { success: true, proposal } or { success: false, error }
Uses same patterns as /api/quickstart:
- maxDuration = 60
- Temperature 0.3 for consistent output
- Request ID logging
- Rate limit (429) handling
- gemini-3-flash-preview model
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Shrimbly