REA ComfyUI Bridge

中文说明：README_zh.md

Use REA (Reverse Engineer Anything) to investigate a creative application you are allowed to inspect, then turn the observed behavior into an evidence-backed ComfyUI integration brief. A small read-only CLI audits ComfyUI workflow JSON against an exported /object_info snapshot.

中文简介： 面向 ComfyUI 创作者的 REA 配套项目：先记录闭源创作软件中可观察、可复现的行为，再整理成 ComfyUI 节点接入规格；同时静态盘点工作流节点和可能的文件依赖。

This is an independent companion project. It does not modify REA or claim that REA can analyze ordinary Python source code better than reading that source directly.

Why this project

Creative workflows often cross several boundaries: a desktop application, model and media files, an API or local service, and a ComfyUI workflow. When an application is opaque, its visible behavior can be investigated with REA. The resulting notes can then guide a reproducible ComfyUI integration. When the workflow is already source-visible, inspect its source directly.

The project has two small parts:

1. An agent skill that structures a permitted REA investigation and records evidence, uncertainty, and an integration contract.
2. rea-comfyui-audit, a read-only Python CLI that summarizes ComfyUI API-format or UI-format workflow JSON and, when provided, compares its node classes with an /object_info snapshot.

Install the skill

Install REA separately using its official instructions. Then install this skill from the public repository:

npx skills add chengyansen-ai/rea-comfyui-bridge --skill rea-comfyui-bridge

The skill uses whichever REA tools are actually available in the agent environment. It does not assume tool names or silently treat an inference as observed fact.

Install the workflow auditor

Python 3.10 or later; no runtime dependencies.

python -m pip install .
rea-comfyui-audit path/to/workflow.json --object-info path/to/object_info.json

To export a ComfyUI node catalog, save the JSON response from your own ComfyUI instance's /object_info endpoint. You can also omit --object-info; in that case the CLI reports node classes without judging whether they are installed.

Examples:

rea-comfyui-audit workflow.json
rea-comfyui-audit workflow_api.json --object-info object_info.json --format json --output audit.json

The auditor does not contact ComfyUI, execute nodes, resolve model files on disk, or change the workflow. File references are reported as candidates based on their names and values; confirm them against the node schema and your local model folders.

REA investigation flow

The skill guides an agent to:

1. Define the application, version, permitted scope, and concrete integration question.
2. Use available REA capabilities to gather repeatable observations from a user-visible flow or documented interface.
3. Keep evidence separate from interpretation; attach an evidence ID and confidence to each meaningful claim.
4. Map inputs, outputs, file handling, configuration, and observable service boundaries.
5. Produce a ComfyUI integration brief with proposed node inputs/outputs, dependency and path contracts, error behavior, unresolved questions, and acceptance criteria.
6. Audit a resulting ComfyUI workflow JSON with the CLI and report unresolved node classes and likely file references.

The skill never asks an agent to bypass authentication, licensing, DRM, or access controls, or to extract credentials, private keys, or user data.

Output boundaries

• REA observations are evidence only for the application version and scenario actually inspected.
• A workflow audit is a static JSON report, not proof that the workflow runs successfully.
• A generated integration brief is a design artifact, not a finished or tested ComfyUI node.
• Apps, models, and external services keep their own licenses and terms. Check them before publishing an integration.

Development

python -m pip install -e .
rea-comfyui-audit --help

License

MIT. See LICENSE. REA and third-party applications remain under their own licenses.