Free, open-source 3D sculpting for Linux, macOS and Windows: dense meshes, remeshing, painting, built-in AI (image/text to 3D), Send to Unreal, game-ready LODs, Gaussian splats and OpenUSD.
Merge queue
Branches land on main through pull requests and scripts/merge-queue.sh; nobody pushes main by hand. (GitHub's own
merge queue exists only for organization-owned repositories; this one belongs to a user, so the script does the
same job.)
For a branch (any agent)
- Branch from
origin/main, one note fragmentchanges/<branch>.mdinstead of CHANGELOG / What's New edits (changes/README.md). - T0 green (
~/Work/.sf/t0.sh, docs/testing-flow.md), push the branch,gh pr create --fill. - The PR's CI runs the fast set (Linux x86_64 and ARM64, macOS, Windows over 4 runners; about 10 minutes). Add the
label
full-cifor the VFX Reference Platform and Blender add-on jobs too (build or packaging changes). scripts/merge-queue.sh add <pr>, write REPORT.md, tell the coordinator.
For the coordinator
scripts/merge-queue.sh run in the background (one at a time). It merges main plus every queued PR whose own checks
are green into a branch queue/batch-<time>, runs the full CI there (everything, as on main), and when that is
green moves main to exactly that commit; the PRs show as merged. A PR that conflicts or is red is dropped with a
comment and the queue-failed label; its author fixes the branch and queues it again. A red batch falls back to
--serial (update each PR from main, wait for its checks, merge), which isolates the PR that broke it.
scripts/merge-queue.sh list shows the queue.
Every landing (batch, serial, staging) first runs tools/ci/no_revert.py on its merges: a merge or squash that
throws a side's change away whole (the result keeps the other side's file or the old one, though that side's commit
still applies to it) is refused with the files listed, as 0.11.1's quick-remesh merge nearly reverted main's newer
fix. Redo it keeping both sides; a deliberate revert lands with SF_ALLOW_REVERT=<its commit>.
scripts/merge-queue.sh no-revert BASE..TIP checks a range by hand.
The merger's own batch (branches merged without pull requests, ~/Work/.sf/merge-batch.sh) lands the same way with
scripts/merge-queue.sh staging --with ~/Work/.sf/t1.sh HEAD. The batch goes to the branch staging, the full CI runs
there while the PC runs the whole suite, and main fast-forwards when both pass. Either way, CI's run on main finds the
full run of the same commit and reuses it rather than building everything a second time (docs/testing-flow.md).
Packages is dispatched on staging as well; main's Packages run reuses it. A red Packages run on staging doesn't hold
main: main's push then builds its own packages, as before.
Cut, don't wait: what is green at cut time lands; the rest rides the next batch.
Timing checks on CI
Speed checks in tests use timeLimit(ms) (tests/TestHarness.h): the given limit on a developer machine, four times it
on CI runners (shared, loaded VMs), and the strict limit again with SF_BENCH_LIMITS=1, as benchmarks are judged.