See this part of the README:

The labeler is well-suited to manage labels that categorize or reduce the semantic information of an issue. For example, labels like bug, enhancement, are self-evident from the contents of an issue. Often, a tracker will use labels that add information to an issue, e.g. wontfix, roadmap. These additive labels should be disabled in your configuration.

We can't simply modify the labeling prompt to disable these because we provide chat history of past labeling decisions. If we say in the prompt that a certain label isn't used and yet it is used in the chat history, the model will still use the labels.

So, we should use a two-part completion. The first part takes the list of labels from a tracker and categorizes each label into "descriptive" and "inscriptive", the second part processes the next label considering only descriptive labels.