Dify workflow

Dify Workflow CLI Skill

Build, edit, validate, layout, and export Dify DSL files entirely from the command line. The CLI is installed as dify-workflow (or dify-workflow.exe on Windows).

If the CLI is not yet installed in the current environment, use PowerShell install script on Windows or shell install script on macOS/Linux.

When to Use

• User asks to create a Dify workflow, chatflow, chat app, agent, or completion app
• User asks to add/remove/update nodes or edges in a Dify YAML
• User asks to validate a Dify DSL file before import
• User asks to auto-layout node positions
• User asks to generate a Mermaid diagram of a workflow
• User asks to troubleshoot why a Dify import shows disconnected nodes or errors

Key Constraints

• DAG only: Dify workflows must be directed acyclic graphs. No cycles allowed — Dify's frontend runs getCycleEdges() and silently removes all edges between cycle nodes, causing disconnections.
• Variable references: Use {{#node_id.variable#}} syntax. Tool nodes output text/files/json (NOT result). LLM nodes output text.
• Node positioning: Dify uses screen coordinates (origin top-left, X→right, Y→down). Use layout command to auto-arrange.
• PowerShell JSON: On Windows PowerShell, avoid inline JSON with --data. Use --data-file with UTF-8 (no BOM) files instead.

Dify App Modes

Dify supports 5 app modes, split into two architectures:

Graph-based modes (use workflow.nodes + workflow.edges)

Mode	DSL app.mode	Description
Workflow	workflow	Single-run DAG execution, no conversation. Start → nodes → End.
Chatflow	advanced-chat	Multi-turn conversation with graph canvas. Uses Answer nodes instead of End.

Config-based modes (use model_config section)

Mode	DSL app.mode	Description
Chat	chat	Simple LLM chat with optional knowledge retrieval. No graph.
Agent	agent-chat	Chat + tool calling. agent_mode.enabled=true with strategy (function_call/react).
Completion	completion	Single-turn text generation. Supports more_like_this, requires dataset_query_variable for knowledge retrieval.

Validation coverage by mode

• workflow / chatflow: Full graph validation (node data, edges, cycles, connectivity, frontend crash prevention, publish checklist)
• chat / agent-chat / completion: model_config validation (model, prompt, variables, dataset, agent_mode, features) via model_config_validators/ package

Procedure: Create a Workflow from Scratch

Before editing a workflow, prefer this execution order:

1. Ensure the CLI is installed
2. Create or open a DSL file
3. Add/update nodes and edges
4. Run validate
5. Run checklist
6. Run layout
7. Export or inspect with Mermaid

If the user asks for a complex node payload, prefer using templates in assets.

Step 1: Create base file

dify-workflow create --mode workflow --name "My Workflow" -o workflow.yaml
# Templates: minimal (default), llm, if-else
# Modes: workflow, chatflow, chat, agent, completion

Step 2: Add nodes

# Add nodes one at a time
dify-workflow edit add-node -f workflow.yaml --type llm --title "GPT Node" --id my_llm

# With data (prefer --data-file on Windows to avoid escaping issues)
dify-workflow edit add-node -f workflow.yaml --type code --title "Process" --id processor
dify-workflow edit update-node -f workflow.yaml --id my_llm --data-file ./assets/llm-node-config.json

Step 3: Add edges

dify-workflow edit add-edge -f workflow.yaml --source start_node --target my_llm
dify-workflow edit add-edge -f workflow.yaml --source my_llm --target end_node

# For branching nodes (if-else, question-classifier), specify --source-handle
dify-workflow edit add-edge -f workflow.yaml -s ifelse_node -t happy_path --source-handle true
dify-workflow edit add-edge -f workflow.yaml -s ifelse_node -t sad_path --source-handle false

Step 4: Update node data

dify-workflow edit update-node -f workflow.yaml --id my_llm \
  -d '{"model": {"provider": "openai", "name": "gpt-4o-mini", "mode": "chat", "completion_params": {"temperature": 0.7}}}'

# Or from file (recommended on Windows)
dify-workflow edit update-node -f workflow.yaml --id my_llm --data-file llm_config.json

# Ready-to-use templates are bundled in this skill
# - ./assets/llm-node-config.json
# - ./assets/question-classifier-config.json
# - ./assets/http-request-config.json

Step 5: Validate

dify-workflow validate workflow.yaml           # Human-readable
dify-workflow validate workflow.yaml -j        # JSON output
dify-workflow validate workflow.yaml --strict  # Warnings = errors
dify-workflow checklist workflow.yaml          # Dify frontend pre-publish check

Step 6: Layout and export

dify-workflow layout -f workflow.yaml          # Auto-arrange (tree strategy, in-place)
dify-workflow layout -f workflow.yaml -o out.yaml --strategy hierarchical
dify-workflow export workflow.yaml -o final.yaml

Command Reference

See commands reference for full command details. See node types reference for all 22 node types and their data schemas. See patterns reference for common workflow patterns and PowerShell examples.

Bundled Resources

• install.ps1: editable install for Windows PowerShell
• install.sh: editable install for bash/zsh
• llm-node-config.json: LLM node payload template
• question-classifier-config.json: classifier payload template
• http-request-config.json: HTTP request node payload template

Quick Command Table

Command	Purpose
create -o file.yaml	Create new app (use --mode and --template)
edit add-node -f file.yaml --type TYPE --title TITLE	Add a node
edit add-edge -f file.yaml --source SRC --target TGT	Add an edge
edit update-node -f file.yaml --id ID -d JSON	Update node data
edit remove-node -f file.yaml --id ID	Remove node + cleanup edges
edit set-title -f file.yaml --id ID --title TEXT	Rename a node
validate file.yaml	Validate DSL
checklist file.yaml	Pre-publish checks (mirrors Dify UI)
inspect file.yaml	Tree view
inspect file.yaml -j	JSON structure
inspect file.yaml -m	Mermaid flowchart
layout -f file.yaml	Auto-layout (tree default)
diff a.yaml b.yaml	Compare two files
config set-model -f app.yaml --provider X --name Y	Set model (chat/agent/completion)
config set-prompt -f app.yaml --text "..."	Set prompt

Troubleshooting

Symptom	Cause	Fix
Nodes disconnected after Dify import	Cycle in graph → frontend strips edges	Run validate to detect cycles, remove back-edges
Checklist errors in Dify UI	Missing required fields or invalid variable refs	Run checklist locally first
Variable reference not found	Wrong output variable name (e.g. .result vs .text)	Tool nodes output text, not result
Layout looks messy in Dify	No layout applied or wrong strategy	Run layout -f file.yaml (tree default)
JSON parse error on Windows	PowerShell escaping issues	Use --data-file instead of -d

Agent Behavior

• Prefer --data-file over inline JSON on Windows
• Prefer inspect -m when the user asks for a visual explanation of the graph
• Prefer checklist in addition to validate before claiming a file is Dify-ready
• Prefer layout before export when the user cares about Dify canvas appearance
• If the user asks for a complex workflow, build it incrementally and validate after each major branch