comfy-png-inspector: Read ComfyUI generation parameters from an image

A DeepSeek Harness plugin maintained by zzcyydshhh, MIT licensed, version 0.1.0, with 16 stars on GitHub. The repository provides a zero-dependency local single-page tool and a command-line parser, packaged as a DSH bundle, exposing the function `comfy_png_inspect(path, limit?, include_prompt?)` to the model. It accepts a PNG or an output directory and returns the LoRA chain and weights, seed/steps/CFG/sampler, base model/CLIP/VAE, latent size, prompt, and bypass nodes. Its distinctive feature is handling information that cannot be obtained by reading metadata directly: `CLIPTextEncode.text` is an empty string transmitted via links and requires backtracking to the source node; the LoRA order must be obtained by reverse traversal from `KSampler.model`; `mode: 4` bypass nodes exist on the canvas but do not participate in image generation; and runtime-generated negative prompts are marked separately. The Web UI also provides search, CSV/JSON export, template rebuild workflows (providing a change list followed by self-check), and batch zipping. The installation command is `dsh plugin add https://github.com/zzcyydshhh/comfy-png-inspector`, depending on `@deepseek-ai/cordis` ~4.0.5 and `@deepseek-ai/dsh-tools`. It registers patches via `cordis.patch.yml` and runs with the permissions of the current dsh process. Limitations: only supports PNG (resaving loses tEXt), cannot scan directories under `file://`, and rebuilding is limited by template node categories.