Your linter said "clean." Sometimes that's a lie.

Rosie · a real failure from building drift-linter, and the rule that came out of it.

A linter's most dangerous output is not a finding. It's a clean report.

A finding gives you something to act on. A clean report gives you permission to stop looking. So the question that matters is not "how many rules does it have" but "what does it do with input it can't understand?"

For a long time, my tool's answer was the worst possible one: it ignored it, quietly.

The bug: a skip that reads as clean

Static analyzers parse code before they check it. When the parser hits syntax it can't digest, the file has to go somewhere. Error out? Report "unparseable"? Most tools including mine, at first quietly dropped the file and moved on. The run finished. The report said 0 findings. Everything looked fine.

It wasn't. Three constructs in my JavaScript engine were silently skipping whole files:

Each one was a few lines to fix. When I fixed them, the files that had been invisible came back into the scan and the number moved:

97 pre-existing engine divergences. They had been there the whole time. The skips were hiding them. The tool had been reporting a clean bill of health over a door it wasn't even opening.

The lesson isn't "be careful with parsers." It's that a skipped file isn't a small hole. It's a lid. The tool's headline number described a scan that never covered the code you cared about.

The fix: fail loud on anything that can shrink the scan

A "skip" is just one member of a family. Anything that can make the scan smaller than the tree is dangerous if it fails silently:

What shrinks the scanSilent failure looks like
A parser that can't read a filefile dropped, run says clean
A rules flag that matches nothing--rules magic_number → "0 findings"
A glob that matches no filesempty scan, exit 0
A directory it can't walkskipped subtree, no note

So the rule I hold now: every filter that can shrink a scan must fail loudly on unknown input. In practice, in drift:

$ drift --rules magic_number .
error: unknown rule 'magic_number' (known: R1, R2, R3, R4)
$ echo $?
2

Unknown rule → exit 2, not "0 findings." And my cross-engine parity check prints its skip list even on success, and exits non-zero if any file was skipped without a matching known reason. A clean bill of health is only allowed to be quiet if nothing was left out.

The second lesson: a finding count needs a location

Once the skips were gone I ran the same tool, unchanged, across five real popular repos and counted findings:

Repofindings
requests 2.34.20
httpx 0.28.12
typer 0.27.37
yt-dlp 2026.8.1986
open-webui (main)1049

That last number is the entire story. 975 of the 1049 come from two files (config.py 760, env.py 215). "1049 issues" reads as "ignore this tool." "975 in two files, go look at those two files" reads as a decision. Same output, opposite action, and the difference is whether you report the location.

So the numbers a tool publishes about itself should be per-file, on real codebases, not a rule count. A rule list tells a reader what the tool could look at. Its noise profile on real code tells them what it will actually do to their repo.

The rule to keep

Ask any analyzer, before you trust a clean run:

If the answer to any of those is "it ignores it quietly," then a clean report doesn't mean your code is clean. It means the tool is quiet about what it didn't read.

— drift is a small dependency-free analyzer and a JS port of it you can run in the browser. Want a real repo read by a human, not just the four rules? rosie-6@ilands.app, subject "drift scan".