AgentWatch managed to assemble a critical finding out of two unrelated plugins. One had a remote address. Another mentioned credentials. They appeared hundreds of lines apart in the same marketplace catalog, and the scanner treated the whole file as one capability.
That would be worth looking at if the signals belonged to the same installed tool. In this case, the file was a list of plugins available to install. The scanner hadn’t established that the two entries had anything to do with each other.
AgentWatch is my tool for inspecting what coding agents can reach through their configuration and installed extensions. Its individual rules flag things like remote endpoints and references to credential files. A separate step combines those signals when the combination deserves more attention. The individual matches here were real. The combined finding was wrong.
I changed that step to group nearby findings within a file. A gap of more than 50 lines starts a new group. That separated the distant entries in the report, but it left an obvious hole: minify the JSON and every plugin is on line one.
So the fix also excludes marketplace.json and plugin-catalog-cache.json from combined critical
findings. Their individual warnings still appear. I didn’t want the scanner quietly dropping
everything it found in a catalog just because the correlation was bad.
The regression cases cover the distant entries and the single-line catalog. They also keep the nearby combination that should still produce a critical finding. Otherwise I could have made the false alarm disappear by breaking the feature entirely.
The 50-line cutoff is a heuristic. It doesn’t tell me whether two settings actually participate in the same operation. For now, I can at least read the evidence without having to untangle a capability the scanner invented from a catalog.