Quiet, review, blocked

Every analysed update gets exactly one verdict. Rules are checked in order, and every rule that fires is kept in reasons, so you always know why.

Blocked

An export your code uses was removed. The types were compared, and something your code imports, calls or references (even only as a type) no longer exists in the new version.

This is the only rule that blocks. It's a strong signal: the code most likely won't compile.

Review

Any one of these is enough:

Changes land on your usage

• A signature you use changed, or became deprecated.
• A member you might reach by name changed (radius couldn't resolve it precisely, so it says "possibly").

The release notes mention you

• A note names something you use.
• A breaking note names no API at all, so it could be about anyone, including you.

radius can't fully see how you use the package

• It's used from a script, a config file, CSS, or only for side effects.
• It's passed around whole, read with computed keys, or reached through files that didn't parse.
• No scanned file uses it at all.
• The installed copy is patched, isn't on the registry, or its version had to be guessed.

The evidence is missing

• The types couldn't be compared and the notes aren't complete either.
• It's a major (or a 0.x minor), and the two nets aren't both complete and clean.

Quiet

Everything else. Concretely, for an update to be quiet:

• every name your code uses was found in the old types,
• nothing radius can't see is in the way,
• and at least one net is complete with zero hits: all the release notes were read and none mentions you, or the types were compared and nothing you use changed.

Why quiet is hard to earn

The only bug that really hurts is a wrong quiet: you merge without reading, and something breaks in production. So radius treats every blind spot as a reason to look, not as a reason to relax.

Before its first release, radius was backtested on 462 real dependency upgrades from four repositories: every update it would have called quiet was checked against what the project actually had to change. Zero wrong quiets. Found one? Please open a "wrong verdict" issue, it's the most useful report there is.

Several packages

The run's exit code is the worst verdict: 2 if anything is blocked, else 1 if anything needs review, else 0. A package radius can't analyse (a registry error, a version that doesn't exist, a dependency installed from git) is listed under "not analysed" with the reason, and the others go on.