Repository navigation
fix(dist_bump): tolerate macro-based Release fields - #100
2 commits merged into
Conversation
tools/dist_bump.py raised BumpError on any Release: containing a macro,
which the CLI turned into an exit 1 that broke build-stage.yml for any
package with a dist_bump configured -- bootc, libical (%autorelease),
pipewire (%{baserelease}...), firefox (1%{?pre_tag}...) and the X11
libraries (5%{?gitdate:...}...).
spec_release now takes the literal release that precedes the first
unexpanded macro -- the stable segment a recorded baseline compares
against -- so firefox and the X11 libs support real dist bumps. A release
that is only macros has nothing literal to compare, so the bump is
skipped (empty suffix) instead of crashing.
Covered by tests/test_dist_bump.py: leading-segment extraction and the
pure-macro no-bump path replace the old refuse-to-parse test.
Signed-off-by: kylerankin <kylerankin@users.noreply.github.com>
hanthor
left a comment
There was a problem hiding this comment.
The parsing is correct and the direction is defensible, but it reverses a deliberate safety decision quietly, leaves the documented contract contradicting the code, and has a test gap over the dominant real-world case.
It does fix a real over-refusal
Checked against Release: lines that actually appear in packages/*/*.spec:
Release: line |
main |
this PR |
|---|---|---|
%autorelease |
BumpError | '' |
%autorelease -b3 |
BumpError | '' |
%{samba_release} |
BumpError | '' |
5%{?gitdate:.%{gitdate}git%{gitversion}}%{?dist} |
BumpError | '5' |
1%{?pre_tag}%{?dist} |
BumpError | '1' |
0.bootstrap%{?dist} |
'0.bootstrap' |
'0.bootstrap' |
2%{?dist} |
'2' |
'2' |
Nine specs carry that mixed literal+sub-macro shape (firefox, libXcursor, libXfixes, libXinerama, libXrandr, libinput, libratbag, pipewire, pulseaudio), and 107 have a purely-macro Release:. The old blanket refusal really was over-broad. Gates: 173 passed, 27 subtests vs 172 passed, 27 subtests on main; validated 192 source RPMs; quoting clean.
Test gap on the dominant case
I mutated split("%", 1) to split("%{", 1) and all 12 tests still passed. Under that mutant Release: %autorelease returns '%autorelease' instead of ''. test_returns_no_release_for_a_purely_macro_release only covers braced forms — %{krb5_release}, %{autorelease}, %{baserelease} — but 103 of this repo's 107 purely-macro specs use bare %autorelease, and one uses %autorelease -b3. Please add "Release: %autorelease\n" and "Release: %autorelease -b3\n" to that list.
The other mutants are caught: dropping the split fails 3, neutering the baseline comparison fails 2, turning the no-Release: error into a return fails 1.
The skip is silent
suffix() does return "" when release is "" (confirmed, as long as baseline isn't itself ""), so the skip works. But main() prints nothing and build-stage.yml:75 just captures an empty string, so a recorded dist_bump on an %autorelease package is discarded with no trace, where previously the build stopped with a message. The counter exists specifically to make a rebuild outrank a prior one; dropping it quietly produces an RPM that doesn't supersede, and nobody finds out. A print(..., file=sys.stderr) on the empty-release path costs nothing and keeps the skip observable.
A narrower version of what the old code refused still survives
5%{?dist} and 5%{?gitdate:.20260101gitabc}%{?dist} both reduce to '5'. If Fedora bolts on a snapshot suffix without moving the leading integer, the recorded baseline still matches and the stale counter is still emitted. The module's whole argument is that a moved baseline retires the counter; a moved sub-macro now doesn't. I don't think that blocks this — it's a far smaller window than refusing 107 specs — but the docstring's "stable across rebuilds of the same Fedora release" is stronger than what the code guarantees.
Docs now contradict the code
docs/targeting-hummingbird.md:239-241 still reads: "A Release: built from macros — krb5_release, nodejs, and kernel-headers's specrelease — is refused by both, rather than compared as a string that is not a release." That paragraph needs to move with the behaviour, including a note that we now diverge from Hummingbird's bump_release() on the mixed case.
Minor: .strip() after match.group(1) is dead — a (\S+) capture can't contain whitespace.
No entry in config/upstream-sources.json carries a dist_bump today, so none of this changes current output. That's an argument for fixing the docs and the diagnostic now, while it's cheap.
Generated by Claude Code
… docs - Cover bare %autorelease (and %autorelease -b3) in the purely-macro test; 103/107 purely-macro specs use the bare form, not the braced one. - Log the empty-release skip on stderr so build-stage.yml does not drop a rebuild silently (a silent "" ships an RPM that does not supersede). - Drop the dead .strip(): a (\\S+) capture carries no whitespace. - Re-target the docs at the extract-leading-literal behaviour and note the divergence from Hummingbird's bump_release().
kylerankin
left a comment
There was a problem hiding this comment.
All four addressed, tests green (12 passed):
-
Test gap — added bare
%autoreleaseand%autorelease -b3totest_returns_no_release_for_a_purely_macro_release(the dominant form; 103/107 purely-macro specs). Your mutantsplit("%{", 1)now fails, as expected. -
Silent skip —
main()now prints the empty-release skip tosys.stderr.build-stage.ymlcaptures stdout, so without this the bump was dropped with no trace. The "moved baseline" path still returns""silently, which is correct (no rebuild needed) — only the macro-skip is now logged. -
Docs — rewrote the
targeting-hummingbird.mdparagraph around the extract-leading-literal behaviour and added the explicit note that we diverge from Hummingbird'sbump_release()(it refuses macro lines; we extract the leading literal). -
.strip()— removed; a(\S+)capture cannot contain whitespace.
hanthor
left a comment
There was a problem hiding this comment.
This is the approach I'd take over #118 — the literal leading segment is the only thing a baseline can honestly be compared with, and skipping with a stderr line beats a homemade macro expander. One thing broke underneath it since it was opened.
main now has tools/rebuild_plan.py (from #70), whose expected_release() calls spec_release() and treats None as "no comparable release". With this branch merged onto main, one test fails:
FAIL: test_an_autorelease_recipe_falls_back_to_the_version (test_rebuild_plan.PublishedComparisonTests)
File "tests/test_rebuild_plan.py", line 120
self.assertIsNone(expected_release(root, entry))
AssertionError: '' is not None
Your spec_release returns "" for a macro-only release; the plan expects None. Returning None there (and keeping the if not release: skip in main(), which already handles either) makes the suite pass — 268 tests otherwise green on the merged tree, and validate still reports 340. Please rebase and make that one change and I'll merge it.
Generated by Claude Code
|
Your branch is on a fork so I couldn't push the one-line fix to it; #138 carries your two commits (authorship preserved) plus a merge of Generated by Claude Code |
a8ab8e0
Problem
tools/dist_bump.pyraisedBumpErroron anyRelease:containing a macro,which the CLI turned into an exit 1 that broke
.github/workflows/build-stage.ymlat
python3 tools/dist_bump.py "$PACKAGE". Any of these packages needing adist_bumpinconfig/upstream-sources.jsonto outrank Hummingbird or Fedorawould crash the build:
bootc,libical:Release: %{autorelease}pipewire:Release: %{baserelease}%{?snapdate:...}%{?dist}firefox:Release: 1%{?pre_tag}%{?dist}libXcursor,libXfixes,libXinerama,libXrandr,libinput,libratbag:Release: N%{?gitdate:...}%{?dist}Fix
spec_releasenow takes the literal release that precedes the first unexpandedmacro -- the stable segment a recorded baseline compares against. So
firefox(
1) and the X11 libraries (5,8,11, ...) support real dist bumps. Arelease that is only macros (
%autorelease,%{baserelease}) has nothingliteral to compare, so the bump is skipped (empty suffix) instead of crashing.
Tests
tests/test_dist_bump.py: leading-segment extraction and the pure-macrono-bump path replace the old refuse-to-parse test. Full suite,
validateandfactory_contractall pass.— hive: backend=pi model=lemonade/Ornith-1.5-35B-A3B-GGUF-Q6_K
🐝 Hive Agent:
contributor| SHA:3b8d574