Code Config Node

Most LLM nodes just generate text. This one is different: the Code Config Node flips a switch that makes the main LLM_Node not only write Python files, but run them, read the errors, and feed them back to the model to fix - in a loop. It's easily the most interesting thing in this pack, and also the most dangerous, and you should know both halves before you touch it.

What it is

It's a config node with two inputs and one output:

• execute_code - the master switch. Off by default.
• project_folder - where finished scripts get copied. Defaults to /projects/LLM.

Wire the CodingConfig output into the CodingConfig input on the main LLM_Node. With execute_code off, nothing happens - it's the pack's default text generation. Turn it on and the behavior changes completely.

The mechanism

Reading the pack's source, once generation is done the main node regex-scrapes python code blocks out of the model's response - honoring filename headers like **main.py:** - and writes each to a tmp folder inside the pack. Then it runs every file with a subprocess under a one-second timeout with SDL_VIDEODRIVER=dummy so nothing pops a window. A file that errors triggers the "fixme" loop: the traceback is appended to the prompt, the model is told "Fix the code. Write all code. Don't take shortcut", and it regenerates and retries - up to 10 attempts. Once files pass, they're copied into project_folder and main.py is executed for real.

So the intended use is self-healing codegen: ask the model to write a small script, and let the loop converge on something that runs. It's a niche, fiddly feature - the one-second timeout means any script that actually does real work will "time out" and trip the fix loop, so it's really built for tiny self-contained snippets.

The warning

This node executes model-generated Python with your user account, no sandbox, no approval step. A malicious or simply unhinged model can write os.remove(...) or worse, and the loop will happily run it. The custom-node world has already had its malware moment, and this is the same trust problem in a different costume - it's your own model doing the writing, but the execution risk is real. Keep execute_code off unless you fully trust the model and know what you're asking it to do. Also note it only applies to the transformers model path; GGUF models return plain text and never enter the code loop.

Installing and gotchas

Same pack install: ComfyUI Manager (search "LLM Node") or cd ComfyUI/custom_nodes && git clone https://github.com/Big-Idea-Technology/ComfyUI_LLM_Node, then restart. The gotcha that will actually bite: the default project_folder of /projects/LLM is a Unix path that almost certainly doesn't exist on Windows, and the copy step errors out if the folder isn't there. Create the folder first, or set it to something real.