Skip to content

NODE_ISSUES.md

Latest commit

 

History

History
177 lines (139 loc) · 6.97 KB

NODE_ISSUES.md

File metadata and controls

177 lines (139 loc) · 6.97 KB
Raw

Node Implementation Gap Report

Systemic Issues

S1: Audio nodes operate on raw bytes instead of decoded samples

Nearly every TS audio node treats audio as Uint8Array instead of decoded PCM samples. Affected: OverlayAudio, RemoveSilence, Reverse, FadeIn, FadeOut, MonoToStereo, StereoToMono, AudioMixer, SliceAudio, Trim, Normalize, CreateSilence, AudioToNumpy, NumpyToAudio, ConvertToArray. Root cause: No pydub equivalent. Need ffmpeg-based or WAV-aware audio processing.

S2: Video nodes with stub byte operations

Affected: FrameIterator, Concat, Trim, Reverse, ExtractFrame, FrameToVideo, AddAudio. These concatenate/slice raw bytes instead of using ffmpeg. Fps and GetVideoInfo return hardcoded values.

S3: Video ffmpeg nodes with unused inputs

Affected: Denoise(strength), Stabilize(smoothing,crop_black), Sharpness(luma_amount,chroma_amount), Overlay(overlay_audio_volume), ColorBalance(wrong filter), AddSubtitles(timestamps,font,color,align), Transition(transition_type,offset), SetSpeed(extreme values).

S4: Model3D nodes are all stubs

All mesh operations do byte arithmetic instead of real geometry. Affected: Decimate, Boolean3D, Transform3D, RecalculateNormals, CenterMesh, FlipNormals, MergeMeshes, FormatConverter, GetModel3DMetadata, TextTo3D, ImageTo3D.

S5: Missing asset system integration

Nodes that read/write filesystem instead of context asset system: LoadImageAssets, LoadAudioAssets, LoadVideoAssets, LoadTextAssets, LoadCSVAssets, SaveAudio, SaveImage, SaveText, SaveVideo.

S6: Missing SaveUpdate/LogUpdate events

File-saving nodes don't emit SaveUpdate. Execute* nodes don't emit LogUpdate.

S7: Missing is_production() guards

All lib.os.* nodes (~25) allow unrestricted filesystem access.

Critical Agent Gaps

A1: ClaudeAgent (anthropic.ts)

  • Missing MCP/control tool support
  • Missing streaming output (never emits chunks despite declaring streaming)
  • Missing use_claude_credentials (Claude Code subscription)
  • Missing cwd (working directory)
  • Missing stderr collection for error context
  • Missing CLAUDECODE env var
  • Missing is_cacheable=False

A2: OpenAI RealtimeAgent (openai.ts)

  • Tool/function-calling support completely missing
  • Wrong default system prompt (short summary vs detailed 8-rule prompt)
  • Wrong VAD: server_vad instead of semantic_vad
  • No audio accumulation or final AudioRef emission
  • No cancellation/turn-detected handling
  • No multi-response pending counter
  • Audio transcript deltas not streamed as chunks

A3: OpenAI RealtimeTranscription (openai.ts)

  • Listens for WRONG event: conversation.item.input_audio_transcription.completed instead of response.audio_transcript.delta
  • Session modalities: text-only vs audio+text
  • No segment aggregation logic
  • No cancellation handling

A4: Agent node (agents.ts)

  • No parallel tool execution (Python uses asyncio.gather)
  • No DB-backed thread persistence (uses in-memory map)
  • No tool name normalization for tolerant matching
  • No iteration limits (Python: max_iterations=32, max_control_tool_calls=16)
  • Missing ToolCallUpdate/EdgeUpdate messages
  • Missing response_format for structured output

A5: Missing: ControlAgent, ResearchAgent (agents.ts)

Generator Stubs

G1: ChartGenerator - returns hardcoded bar chart, no LLM call

G2: SVGGenerator - returns hardcoded placeholder SVG, no LLM call

Image Processing Gaps

I1: RenderText (lib-image-draw.ts) - ignores all inputs, hardcoded SVG

I2: Canny (lib-image-filter.ts) - not real Canny, same as FindEdges/Contour

I3: AutoContrast/Equalize/AdaptiveContrast all use sharp.normalize()

I4: LiftGammaGain wrong operation order

I5: FilmLook preset values all wrong

I6: RankFilter ignores rank input

I7: Expand reads wrong property name for fill

I8: GaussianNoise wrong algorithm (uniform not Gaussian), ignores mean/stddev

Text/Document Gaps

T1: CountTokens uses naive regex instead of tiktoken

T2: FormatText missing Jinja2 filter support

T3: RegexSplit maxsplit implementation is incorrect

T4: RegexValidate uses .test() instead of re.match() (anchored)

T5: ExtractJSON JSONPath only supports basic paths

T6: SplitDocument/SplitHTML/SplitJSON/SplitMarkdown - all naive char chunking

T7: HtmlToText uses regex stripping instead of html2text library

T8: LoadWordDocument extracts plain text, loses all formatting

PDF Gaps

P1: ExtractTables - naive heuristic vs pdfplumber/pymupdf

P2: ExtractMarkdown - font-size heuristic vs pymupdf4llm

P3: ExtractImages - entirely missing

P4: ExtractTextWithStyle missing color field

KIE API Gaps

K1: 31 model ID mismatches (see full report)

K2: All Suno nodes (6) hit wrong endpoint (/api/v1/generate instead of specific)

K3: Veo31 nodes wrong endpoints and polling

K4: Runway nodes wrong endpoints and download paths

K5: Missing audio→MP3 and video→MP4 conversion before upload

K6: 7 missing audio nodes (lyrics, stems, covers, etc.)

Search Gaps

SR1: 34 of 42 search nodes missing entirely

SR2: All 8 existing nodes missing text output field

Messaging Gaps

M1: DiscordBotTrigger only validates token, doesn't listen

M2: TelegramBotTrigger only validates token, doesn't poll

M3: Gmail nodes are stubs (throw errors)

API Integration Gaps

API1: Gemini ImageGeneration uses :predict instead of :generateImages

API2: Apify API key name mismatch (APIFY_API_TOKEN vs APIFY_API_KEY)

API3: Supabase nodes don't declare url/key inputs (always undefined)

Control/Trigger Gaps

C1: If node missing streaming input handling

C2: Collect emits intermediate results instead of final list

C3: WaitNode uses setTimeout instead of suspend/resume

C4: WebhookTrigger only parses JSON (missing form data)

SVG/Misc Gaps

SVG1: SVGToImage returns raw SVG instead of rasterizing to PNG

MISC1: SpiderCrawl respect_robots_txt ignored

MISC2: ShowNotification does nothing

MISC3: ChartRenderer only 3 of 17+ plot types

Missing Namespaces (144 nodes)

  • nodetool.boolean.* (8 nodes)
  • nodetool.list.* (22 nodes)
  • nodetool.dictionary.* (20 nodes)
  • nodetool.numbers.* (2 nodes)
  • search.* 34 missing nodes
  • lib.http.* (15 nodes)
  • lib.date.* (15 nodes)
  • lib.json.* (10 nodes)
  • lib.math.* (10 nodes)
  • lib.uuid.* (7 nodes)
  • lib.ocr (1 node)

Minor Gaps

  • Missing is_cacheable=False on trigger/workspace nodes
  • Missing refresh_uri() on constant nodes
  • Missing _auto_save_asset on OpenAI/Gemini image nodes
  • Output wrapping differences ({output:...} vs flat)
  • DropNA treats empty string as NA
  • Aggregate missing std/var functions
  • ModulusArray JS % vs numpy mod for negatives
  • SliceArray stop=0 opposite behavior
  • DateTime constant wrong utc_offset conversion
  • CompareImages does byte comparison instead of preview slider
  • LoadImageFolder/LoadAudioFolder ignore include_subdirectories, extensions, pattern props
  • SaveImageFile ignores overwrite prop
  • TextToImage/ImageToImage missing provider params (guidance_scale, seed, etc.)
  • Image GetMetadata wrong output schema