You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
This release changes what PickCheck fails on and lets it judge a branch
instead of the whole file. Cognitive complexity replaces cyclomatic
complexity as the default gate, so readable flat dispatch stops failing and
deeply nested conditions start. check --base checks everything a branch
changed since it forked, and compares each violation with the merge base, so a
one-line edit in a legacy function no longer fails on debt that was already
there when --fail-on excludes it.
Changes
New cognitive metric, default limit 15, replaces cyclomatic complexity
as the default gate. It scores how hard control flow is to follow: a switch or match costs 1 however many arms it has, and every branch or
loop costs more the deeper it is nested. The algorithm is pinned per
language in docs/spec.md. (#7)
complexity is still measured but its default limit is now null, which
turns the check off. Add "limits": {"complexity": 15} to .pickcheck.json
to restore the old gate. Any limit, global or per language, now accepts null. (#7)
This changes what fails on upgrade. Across 32,805 functions in 19
repositories, the old gate failed 118 and the new one fails 220: 27 stop
failing, mostly long flat switches, and 129 start, because they nest
conditions the old count could not see. (#7)
Dart on T { } handlers without a catch now count toward complexity.
(#24)
&& inside Rust let chains (if let … && let …) now counts toward complexity and bool_ops. (#26)
check --base <ref> checks the functions a branch changed since it forked
from <ref>, committed or not, plus untracked files. It implies --changed, and a stacked branch checked against its parent reports only
its own functions. (#29)
The .pickcheck.json changed in this diff note now also appears when the
config is deleted or renamed, not only when it is edited. (#32)
Dart mixin members are now reported as Mixin.member, like class and
extension members, so their names in reports change. (#30)
check --base compares each violation with the same function at the merge
base and reports a status: new, worsened, unmatched, improved, or unchanged. --fail-on picks which statuses fail; the others print as WARN and never fail the run. Without --fail-on every violation still
fails, and hooks and plain --changed are unchanged. JSON gains base, fail_on, and per-violation status, base_metrics, and limits. (#30)
Validation
cargo test --workspace --locked --all-targets, cargo clippy --workspace --all-targets -- -D warnings, cargo run -- check crates, and cargo llvm-cov --workspace --locked --fail-under-lines 94 (94.23% lines)
pass on main at fed4279: https://github.com/pickforge/pickcheck/actions/runs/36576086532
Golden fixtures cover every language, including hand-derived cognitive
values and the Dart mixin names.
The cognitive rollout numbers above come from scanning 32,805 functions in
19 repositories (#6).
--base and the baseline comparison are covered by CLI tests over real Git
repositories: committed, staged, unstaged and untracked changes, stacked
branches, renamed files, duplicate names, closure pairing, unreadable bases,
and --fail-on exit codes.
Known limits
The baseline comparison matches functions by name within a file. A renamed
function, or one moved into a different file, counts as new; a file
rename that Git detects keeps its functions paired.
Anonymous closures pair by position under their named owner. Adding or
removing a closure there makes its siblings unmatched.
With several merge bases in a criss-cross history, Git's choice is used.
Paths that are not valid UTF-8 are decoded lossily.
Hooks still check against HEAD and never compare with a base.
