CLI 输出文件转图片

Every "let an external CLI do the heavy lifting" workflow needs a way back in. KSCLIImageOutputFile is the return trip: it takes the stdout, stderr, and exit code from KSAntigravityCLIExecute, plus the output_path that KSCLIImageInputFiles predicted, and loads the generated image back into the graph as a tensor. It's the node that closes the loop.

How it finds the image

Three fallbacks, in order, and it's a nicely defensive design:

1. The expected path. If expected_output_path points at an existing image file, it loads that. In the intended pipeline this is output_path from KSCLIImageInputFiles - the file the agent was told to write. This is the fast path that hits 90% of the time.
2. Parse stdout/stderr. If the expected path doesn't exist, it regex-scans the combined stdout+stderr for anything that looks like an image path (.png, .jpg, .jpeg, .webp, .bmp, .tif, .tiff, quoted or not) and tries those, resolved against working_directory if relative. This covers agents that ignore your instructions and just "print only the saved image path" somewhere in their output - which is why the default Antigravity prompt says to do exactly that.
3. Search a directory. Last resort: if search_directory is set, it loads the most-recently-modified image in that folder (recursively). Crude but effective - the newest file is usually the one your agent just made.

If all three fail, it raises - and if return_code was non-zero, the error is prefixed with "CLI failed with exit code N" plus the first chunk of output. That distinction is useful: exit 0 with no image means the agent finished but didn't save where expected; non-zero means the agent itself crashed.

The inputs that matter

• stdout / stderr / return_code - straight from the CLI execute node. All three are required.
• expected_output_path - the predicted save location. Required, but if it's blank the node skips straight to parsing.
• working_directory (optional) - used to resolve relative image paths. Blank = ComfyUI's current working directory.
• search_directory (optional) - the last-resort folder to scan.

Outputs: image (the tensor, ready for preview, saving, or more processing) and output_path (the absolute path of whatever it actually loaded, so you can confirm which file you got).

Installing

With the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/kongshan4219/ComfyUI-Kongshan-Nodes

restart ComfyUI. No keys, no models - pure local file handling.

Gotchas

• The same numeric working_directory quirk as its sibling. The source warns that a widget value mismatch can show up as a numeric working_directory; the node detects it, prints a warning, and falls back to the current directory. If paths keep resolving wrong, check whether your working_directory wire actually connected.
• It loads RGB only. The internal loader converts to RGB before returning a tensor, so a transparent PNG comes back with its alpha dropped. For this pipeline's white-background output that's fine; if you need alpha, this isn't the node.
• Wrong-file risk in fallback 3. "Newest file in the folder" is a guess. If your agent wrote several images, you may get the wrong one - check the output_path output before trusting the tensor.

The whole CLI trio (input-files → execute → output-file) is the most unusual thing in this pack: a full product-photo pipeline that passes through an autonomous coding agent. It's young, unknown (zero impressions everywhere as of now), and a little janky - but when it works, it's a genuinely different way to drive ComfyUI.