Python代码执行器

ComfyUI gives you a visual language, but occasionally you just want split(), a regex, or a JSON transform - without writing a whole custom node. PythonCodeExecutor is the escape hatch: a node that runs your Python snippet against up to three string inputs and returns the result. It's the "yes, you can do that" node for the fiddly data-shaping tasks this pack keeps running into.

Here's how it works. You write code into the code box. The environment pre-loads input1, input2, input3 plus a whitelist of modules (json, re, math, random, datetime, timedelta) and a set of builtins (len, str, int, list, dict, sorted, map, filter, etc.). Your code runs through exec, and whatever you assign to the output variable comes out the other side - a string stays a string, anything else gets JSON-serialized. print() output lands in the second output, logs, alongside execution notes.

The default code template is even useful: it shows the canonical "newline-separated string → JSON array" example, which is the same job StringToJsonArray does one-note.

Inputs and outputs

• code - required, multiline. The meat.
• input1 / input2 / input3 - optional string inputs (auto-JSON-ified if you wire something non-string).
• safe_mode - present, but the tooltip is a confession: security mode is forced on and this parameter is deprecated. Ignore it.
• Outputs: output (STRING) and logs (STRING).

The sandbox is a speed bump, not a wall

The safety is a regex blocklist: import os|sys|subprocess|shutil|pickle|marshal|shelve|dill, eval|exec|compile|open, dunder attribute access (__class__, __base__, ...), getattr, globals, and friends. That stops accidental footguns and casual scripts dead in their tracks - try import os and you get a clean error JSON.

But be clear-eyed: a regex filter is not a real sandbox. There are ways around string-matching filters (obfuscation, Unicode tricks, reaching restricted builtins through allowed names), so treat this as a convenience for your own trusted workflows, not as a security boundary for untrusted ones. Don't load a shared workflow that runs arbitrary code here and assume it's contained. If you need a real boundary, run untrusted workflows in a Docker sandbox, not this node.

When you'd reach for it

Building JSON for the pack's API nodes (the template with json.dumps covers 90% of that), normalizing text, parsing a batch ID out of a log line, deriving a filename. It's the utility node you discover you want after the third time you hand-format JSON. Note it outputs strings only - if you need IMAGE or LATENT transforms, this isn't your tool.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/jinchanz/ComfyUI-ADIC

Restart ComfyUI, or install "ComfyUI-ADIC" via ComfyUI Manager. Stdlib-only, no pip extras, no models. The README never documents it; the source (module/nodes_api_basic.py) carries the full security notes.