Dify Workflow DSL Skill

Create, repair, review, migrate, and validate Dify App DSL with an AI coding agent.

中文说明 · 10 Dify 1.16 examples · Evaluation report

Important

Now supports Dify 1.16.0 and App DSL 0.7.0, including Agent v2 workflow nodes. Dify 1.16.0 is the latest official release as of 2026-07-21; this repository has been checked against its tagged source and real Console import path. See the Dify v1.16.0 release and the evaluation evidence.

🎯 Version Strategy

Important

Dify 1.16.x → App DSL "0.7.0"

This is the default for new workflows and includes portable Agent Apps and Agent v2 workflow nodes.

Target workspace	DSL version	Policy
Dify 1.16.x	"0.7.0"	Default for new generation
Dify 1.15.x	"0.6.0"	Compatibility generation and validation

If Dify reports that the DSL version is incompatible, first check the target workspace version. A 0.7.0 file cannot be made compatible with Dify 1.15.x by ignoring the warning or changing only version:

• upgrade the workspace to Dify 1.16.0 or later; or
• ask the agent to generate a real DSL 0.6.0 workflow for Dify 1.15.x.

The skill supports an explicit target version. For example:

Use $dify-workflow-dsl to create this workflow for Dify 1.15.0 using DSL 0.6.0.

Validate the result against the same target:

python3 scripts/validate_dsl.py --strict --target-version 0.6.0 workflow.yml

Tip

New apps default to workflow. Use or recommend advanced-chat only when the requirement needs multi-turn memory, sys.query, chat file upload, or answer nodes.

✨ What It Does

• Generates workflow, advanced-chat, and DSL 0.7.0 Agent App YAML.
• Supports portable Agent v2 packages, bindings, jobs, declared outputs, and omitted assets.
• Covers Start, End, Answer, LLM, Code, IF/ELSE, Question Classifier, Human Input, Iteration, Loop, Assigner v2, tools, triggers, retrieval, files, and other common nodes.
• Repairs graph wiring, branch handles, selectors, dependencies, and version-specific structures.
• Reviews existing DSL with stable diagnostic codes and machine-readable JSON.
• Keeps 0.6.0 compatibility for Dify 1.15.x.

The skill treats tagged Dify source as schema authority. Exported files from the target workspace remain the authority for dynamic plugin/tool fields.

🧪 Dify 1.16 Evaluation

Ten maintained DSL 0.7.0 scenarios cover the main static compatibility surface:

Scenario	Coverage
Text summarizer	Start → LLM → End
Multi-turn assistant	Chatflow, sys.query, memory, Answer
Excel analysis	File input, Document Extractor, Markdown conversion, LLM
Priority routing	IF/ELSE and Variable Aggregator
Question classification	Three classifier branches
Array iteration	Iteration container and internal child topology
Quality loop	Loop variables and Assigner v2
Human approval	Select/file-list inputs and action branches
Agent v2 workflow	Package, binding, job, declared output
Agent App	app.mode: agent and portable Soul

All ten pass:

1. the repository validator in strict mode; and
2. Dify 1.16.0's own production node registry, Human Input graph check, AgentPackage, and WorkflowNodeJobConfig models; and
3. Dify 1.16.0's real AppDslService.import_app against isolated PostgreSQL and Redis: 10 completed, 0 failed, 0 warnings.

This evaluation exposed and fixed a real issue: an older valid fixture omitted the LLM context object required by Dify. See the full evaluation report.

Workflow Previews

The two largest maintained graph examples each contain six nodes. Both files were imported through the real Dify 1.16.0 Console, then captured from Dify's Workflow canvas. Click an image to inspect the imported source workflow.

Priority Routing

Array Iteration

🤖 Agent v2 / DSL 0.7.0

The skill supports both 0.7.0 Agent forms:

• top-level Agent App: app.mode: agent + agent + agent_packages;
• Workflow/Chatflow Agent v2 node: version: "2" + agent_binding.package_ref + agent_job.

Agent packages are portable, but secrets and workspace-bound assets are not. After import, review warnings and reconnect model credentials, tools, contacts, skills, files, and other omitted assets.

The following screenshot is the checked-in Agent v2 example after a real import into Dify 1.16.0. The selected node and its inline Agent configuration panel are rendered by Dify itself.

📦 Install

git clone https://github.com/yzmw123/dify-workflow-dsl-skill.git
cd dify-workflow-dsl-skill
bash install.sh --platform codex

Supported targets:

bash install.sh --platform claude
bash install.sh --platform codex
bash install.sh --platform openclaw
bash install.sh --platform hermes
bash install.sh --platform opencode
bash install.sh --platform all

Use --target-dir for a custom skills directory and --force to replace an existing installation.

🚀 Use

Create a workflow:

Use $dify-workflow-dsl to create a Dify 1.16 workflow.
Read an uploaded Excel file, convert the table to Markdown, analyze it with an
LLM, and return the Markdown table plus the analysis.

Create an Agent v2 workflow:

Use $dify-workflow-dsl to create a DSL 0.7.0 workflow with a portable Agent v2
node, declared outputs, package refs, and strict validation.

Repair an existing file:

Use $dify-workflow-dsl to review and fix this YAML. Preserve its supported DSL
version and report every import blocker and remaining workspace-specific step.

For an unfamiliar plugin tool, provide a minimal export from your workspace. That export is more reliable than a marketplace page or tool name alone.

✅ Validate

Install development dependencies:

python -m pip install -r requirements-dev.txt

Run local validation and tests:

python3 scripts/validate_dsl.py --strict --target-version 0.7.0 workflow.yml
python3 scripts/validate_dsl.py --format json --strict workflow.yml
python3 -m unittest discover -s tests -v
python3 scripts/validate_dsl.py \
  --strict \
  --target-version 0.7.0 \
  examples/dify-1.16.0/*.yml

With a Dify 1.16 Python 3.12 API environment:

python scripts/validate_with_dify_source.py \
  --dify-source /path/to/dify-1.16.0 \
  examples/dify-1.16.0/*.yml

The local validator checks version/mode compatibility, graph endpoints and reachability, cycles, branch coverage, container topology, selectors, template IDs, Human Input, Agent schemas, dependencies, and common SQL risks. Any YAML PyYAML can parse returns structured diagnostics; one bad file does not stop a batch.

🗂️ Repository Map

.
├── SKILL.md                         # Agent-facing operating rules
├── examples/dify-1.16.0/            # 10 maintained DSL 0.7.0 scenarios
├── references/                      # Source-backed schemas and patterns
├── scripts/
│   ├── dify_dsl_validator/          # Reusable validator package
│   ├── validate_dsl.py              # Local deterministic CLI
│   └── validate_with_dify_source.py # Dify-source compatibility gate
├── tests/                           # Fixtures and regression tests
├── install.sh
├── README.md
└── README_CN.md

⚠️ Validation Boundary

Static validation cannot prove that a workflow will run in every workspace. Real Dify import and execution are still required for:

• installed plugin versions and dynamic tool schemas;
• model availability and credentials;
• knowledge bases, contacts, skills, files, and private assets;
• external APIs, Human Input delivery, and side effects;
• UI save/re-export behavior.

No credentials belong in public DSL. Prefer fixed, parameterized SQL over model-generated mutation or DDL.

📚 Research Basis

The project was built from Dify's tagged source, personal exports, and 262 parseable public App DSL files. Public examples are useful for graph patterns and backward compatibility, but older files are not treated as current schema authority.

Primary links:

• Dify
• Dify Marketplace
• Dify official plugins
• Agent Skills specification

Follow Silicon Scout S01 On WeChat

Welcome to follow my WeChat Official Account: 硅基斥候 S01 (Silicon Scout S01). The account is still just getting started, so if this skill helps you, a follow would mean a lot. I will keep sharing practical skills I personally find useful, plus AI news, knowledge, and notes.

Recon focus

Silicon Scout S01 follows large language models, AI agents, AI coding and toolchains, hands-on product tests, enterprise and government AI adoption, and policy, security, and compliance.

Why "Scout"?

A scout enters unknown territory first, checks what is real, and brings back useful intelligence.

In the AI context, S01 does that by personally testing new models, tools, and products, checking their risk boundaries, and explaining the actionable takeaways clearly.

My other open-source projects will also be announced on this account first.