workflow-node-llm

name	workflow-node-llm
description	Builds or updates frontend LLM-node data with model capability selection.
allowed-tools	["execute_node_skill_script","update_current_graph"]

Workflow LLM Node

Use type: llm for semantic understanding, classification, extraction, and content generation.

Node Schema

build_node.data only contains business fields. Do not provide id, type, position, or status; the script generates or fills them. New LLM node IDs follow the frontend rule: llm-<uuid> with a uniqueness suffix only if needed.

Field	Type	Required	Meaning	Notes
title	string	no	Display title	Defaults to 大模型
description	string	no	Node description
inputs	array	no	Variables referenced by prompts	Defaults to one input string
outputs	array	no	Values generated by the model	Defaults to one result string
config	object	no	Model, prompt, and mapping settings	Defaults match frontend LLM node

Each input/output item has:

Field	Type	Required	Meaning
name	string	yes	Stable variable name
type	string	yes	Such as String, Object, Image, or Array
description	string	no	Business meaning

Config fields:

Field	Type	Required	Meaning	Notes
prompt	string	no	Compatible prompt field	Prefer userPrompt
systemPrompt	string	no	Stable role and policy
userPrompt	string	yes	Task prompt	Variables use {{name}}
model	string	no	Configured model name	Defaults to empty; obtain through scripts below
modelProvider	string	no	Model provider	Defaults to deerflow
temperature	number	no	Sampling temperature	Low for deterministic tasks
maxTokens	integer	no	Output-token limit	Positive
enabled	boolean	no	Whether node runs	Defaults to true
responseMode	string	yes	text or json
outputKey	string	no	Main output variable	Leave empty to use the first non-reasoning output
reasoningKey	string	no	Reasoning output variable	Only with thinking
thinkingEnabled	boolean	no	Enable model reasoning	Model must support it
reasoningEffort	string	no	Reasoning effort	Only when supported
inputMappings	array	no	Input source mappings	Full ordered array

Each inputMappings item contains field, sourceType, source, and valueType. field must correspond to an input.

Input Variables

Inputs expose exactly two user-facing modes:

1. Custom string: {"field": "instruction", "sourceType": "literal", "source": "summarize", "valueType": "String"}.
2. Upstream reference: {"field": "input", "sourceType": "node", "source": "start.input", "valueType": "String"}.

context is a legacy runtime source type, not a Skill input mode. Do not emit it. The start-node input is referenced as a normal node output, such as start.input.

Before building or changing any node-reference input, execute list_input_sources. For a new node pass the IDs of its intended direct predecessors from the confirmed topology. For an existing node pass node_id. Only use an exact source returned by that script. The script includes all reachable ancestors, checks Graph direction, and returns source types.

For a node mapping, copy the returned source type into both the input type and mapping valueType. For a literal mapping, the input and valueType must both be String.

Output Variables

Supported output types are:

String, Integer, Number, Boolean, Time, Object, Image, Video, Array, Array<String>, Array<Integer>, Array<Number>, Array<Boolean>, Array<Time>, Array<Object>, Array<Image>, and Array<Video>.

Output names must be unique identifiers using letters, numbers, and underscores, and must not start with a number.

• responseMode: text: declare one primary output. Its type is normally String. outputKey may be empty to select the first non-reasoning output.
• responseMode: json: declare every expected top-level JSON field as an output with its exact type. Set outputKey only when one declared field is the primary output.
• When thinking is enabled, add reasoning_content as a String output and set reasoningKey to the same name.

Default LLM node generated by build_node:

{
  "id": "llm-<uuid>",
  "title": "大模型",
  "type": "llm",
  "description": "调用大语言模型，基于输入变量和提示词生成回复。",
  "inputs": [
    {
      "name": "input",
      "type": "String",
      "description": ""
    }
  ],
  "outputs": [
    {
      "name": "result",
      "type": "String",
      "description": ""
    }
  ],
  "config": {
    "prompt": 
     
     
     
     
     
     
     
     
     
     
     
     
      
         
         
         
         
      
    
     
     
     
     
     
     
     
     
  

Dynamic Model Selection

Execute list_models before choosing a model:

• requires_vision: true when any input is Image, ImageList, Video, or VideoList.
• requires_thinking: true only when the task needs explicit reasoning.

Never choose a model that lacks a required capability.

Then execute resolve_model_config with the selected model, capability requirements, enable_thinking, and reasoning_effort. Merge its result into the candidate node config.

Scripts

list_input_sources

• Path: workflow-node-llm/scripts/node.py
• Entry: list_input_sources
• Description: list valid upstream variables for input mappings.
• New-node arguments: {"upstream_node_ids": ["intended-direct-predecessor-id"]}
• Existing-node arguments: {"node_id": "existing-id"}
• Returns: {"sources": [{"source": "node-id.field", "nodeTitle": "...", "field": "...", "type": "...", "description": "..."}], "allowed_source_values": [...]}

list_models

• Path: workflow-node-llm/scripts/node.py
• Entry: list_models
• Description: return configured models matching vision/thinking requirements.
• Arguments: {"requirements": {"requires_vision": bool, "requires_thinking": bool}}
• Returns: {"models": [...], "recommended_model": "name"}

resolve_model_config

• Path: workflow-node-llm/scripts/node.py
• Entry: resolve_model_config
• Description: resolve the exact model and reasoning config.
• Arguments: {"request": {"model": "name", "requires_vision": bool, "requires_thinking": bool, "enable_thinking": bool, "reasoning_effort": "medium"}}
• Returns: model configuration fields to merge into data.config.

build_node

• Path: workflow-node-llm/scripts/node.py
• Entry: build_node
• Description: build a complete new LLM node and generate id and type.
• Arguments: {"data": <Node Schema business fields>, "upstream_node_ids": ["direct-predecessor-id"]}
• Returns: {"node": <complete node>}

update_node

• Path: workflow-node-llm/scripts/node.py
• Entry: update_node
• Description: find the current LLM node by ID and apply changed fields.
• Arguments: {"node_id": "existing-id", "changes": <changed fields>}
• Returns: {"node": <complete updated node>}

Objects merge recursively; arrays replace the entire old array; scalars replace directly; omitted fields stay unchanged. To modify one input or mapping, submit the complete new inputs or config.inputMappings array. Never submit only the changed array item.

Rules

• When thinking is disabled, do not declare reasoning_content as an output.
• Every prompt variable must have a matching input or input mapping.
• Never invent a node mapping source. Use list_input_sources.
• inputs and config.inputMappings must have one complete matching entry per input field.
• Keep outputKey empty when the node should use its first non-reasoning output (result by default); set it only when an explicitly planned output variable is required.
• Use only models returned by list_models.
• The runtime caches the node returned by build_node or update_node; pass only its related edges to update_current_graph.
• Do not add position or runtime status.