workflow-execution

name	workflow-execution
description	Workflow execution and complete workflow generation tools (executeWorkflow, applyWorkflowJson). Use when the user wants to run a workflow or build a complete workflow from a description.

Workflow Execution and Generation

You have tools to run workflows and load complete workflows from JSON. Follow the decision order in the workflow-guardrails skill before creating any complete workflow.

Available tools

Tool	When to use
executeWorkflow	When the user says "run", "execute", "generate", "queue", or wants to see results from their workflow.
applyWorkflowJson	When building a complex multi-node workflow that would take many addNode/connectNodes calls. Generates and loads the entire graph in one step.

When to use executeWorkflow

• User says "run the workflow", "execute it", "generate the image", "queue it"
• After building or modifying a workflow, when the user wants results
• Accepts an optional timeout (seconds, default 300)

Result interpretation

• success — Execution completed. Mention the number of outputs and any image filenames.
• error — Execution failed. Explain the error (node type, message) and suggest fixes (check connections, missing models, etc.).
• interrupted — User or system cancelled the execution.
• timeout — Execution took too long. Suggest the user check ComfyUI's queue or increase the timeout.

Validation and retry after executeWorkflow

After every executeWorkflow call:

1. Validate the result: Treat as "problem" if the status is error, timeout, or interrupted, or if the run completed but produced no outputs when outputs were expected (e.g. empty images list for an image workflow).
2. On problem: Retry by calling executeWorkflow again. You may optionally suggest a small change (e.g. different seed, or checking a node) before retrying, but retrying the same workflow is acceptable.
3. Maximum attempts: Try up to 5 times in total (initial run + up to 4 retries). Stop as soon as you get success with valid outputs.
4. After 5 failed attempts: Do not retry further. Inform the user clearly:
   • That the workflow was executed 5 times and did not succeed.
   • The last error or status (error message, timeout, interrupted).
   • Concrete suggestions: check node connections, missing models, ComfyUI queue, or increasing the timeout.

When to use applyWorkflowJson

• User asks for a complete workflow ("create a txt2img workflow", "build me an img2img pipeline with upscale")
• The workflow requires many nodes and connections that would be tedious with individual addNode/connectNodes calls
• User provides or describes a workflow they want loaded
• User wants a full workflow built end-to-end without step-by-step edits
• When applying a workflow from getWorkflowInfo temp refs: use workflowPath with the _tempFile or fullWorkflowRef value (e.g. workflowPath: "workflow_abc123.json") instead of inline workflow — workflows are stored in user_context/temp/ to avoid payload limits

Best practice: default to JSON for complete workflows

• Prefer applyWorkflowJson for any complete workflow request, even if the workflow is small.
• Use JSON-first: generate the full workflow object and load it in one call.
• Only use addNode/connectNodes if the user explicitly asks for step-by-step construction or asks for specific node additions.

CRITICAL: Before generating a workflow

1. Always call searchInstalledNodes first to verify the node types you plan to use exist in the user's installation
2. Always call getAvailableModels to find actual model filenames (checkpoints, loras, vae)
3. Never guess model filenames — use the exact filenames from the user's installation
4. Choose the correct loader using the model-loading-rules skill (Load Diffusion Model for diffusion_models/unet; CheckpointLoaderSimple for checkpoints)
5. If a required node type is not installed, tell the user what custom node package to install

ComfyUI API format specification

The workflow parameter is an object where:

• Keys are string node IDs (e.g. "1", "2", "3")
• Values are node objects with:
  • class_type (string): The registered node type name (e.g. "CheckpointLoaderSimple")
  • inputs (object): Node inputs — either scalar values or link references
  • _meta (object, optional): { "title": "Display Name" }

Input value types

• Scalar: Direct value — "steps": 20, "text": "a cat", "seed": 42
• Link: Reference to another node's output — ["5", 0] means "output index 0 from node 5"
  • First element: source node ID (string)
  • Second element: output slot index (number)

Example: basic txt2img workflow

{
  "1": {
    "class_type": "CheckpointLoaderSimple",
    "inputs": {
      "ckpt_name": "v1-5-pruned-emaonly.safetensors"
    },
    "_meta": { "title": "Load Checkpoint" }
  },
  "2": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "text": "a beautiful sunset over the ocean, highly detailed, 4k",
      "clip": ["1", 1]
    },
    "_meta": { "title": "Positive Prompt" }
  },
  "3": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "text": "blurry, low quality, deformed, ugly, watermark",
      "clip": ["1", 1]
    },
    "_meta": { "title": "Negative Prompt" }
  },
  "4": {
    "class_type": "EmptyLatentImage",
    "inputs": {
      "width": 512,
      "height": 512,
      "batch_size": 1
    },
    "_meta": { "title": "Empty Latent" }
  },
  "5": {
    "class_type": "KSampler",
    "inputs": {
      "seed": 42,
      "steps": 20,
      "cfg": 7,
      "sampler_name": "euler",
      "scheduler": "normal",
      "denoise": 1,
      "model": ["1", 0],
      "positive": ["2", 0],
      "negative": ["3", 0],
      "latent_image": ["4", 0]
    },
    "_meta": { "title": "KSampler" }
  },
  "6": {
    "class_type": "VAEDecode",
    "inputs": {
      "samples": ["5", 0],
      "vae": ["1", 2]
    },
    "_meta": { "title": "VAE Decode" }
  },
  "7": {
    "class_type": "PreviewImage",
    "inputs": {
      "images": ["6", 0]
    },
    "_meta": { "title": "Preview Image" }
  }
}

CheckpointLoaderSimple outputs

• Output 0: MODEL
• Output 1: CLIP
• Output 2: VAE

Build-then-run pattern

After calling applyWorkflowJson to load a workflow, suggest running it with executeWorkflow if the user wants to see results. Example flow:

1. User: "Create a txt2img workflow with a cat prompt and run it"
2. Call searchInstalledNodes to verify node types
3. Call getAvailableModels to find a checkpoint
4. Call applyWorkflowJson with the complete workflow
5. Call executeWorkflow to run it
6. Validate the result: if there are problems (error, timeout, interrupted, or no outputs), retry executeWorkflow up to 5 times in total; if still failing after 5 attempts, inform the user and suggest fixes
7. Report the results (or the failure and what to check)

Guidelines

• Prefer applyWorkflowJson over addNode/connectNodes when building complete workflows
• Always verify node types and models first — never generate a workflow with guessed node types or model filenames
• Use PreviewImage instead of SaveImage by default unless the user specifically asks to save
• Set reasonable defaults: steps=20, cfg=7, sampler_name="euler", scheduler="normal", denoise=1.0
• Use random seeds: use a random number for the seed so each generation is different