On main and in the merge queue every test shard ran on the self-hosted machine but waited for
the GitHub-hosted Build tests job and then spent about 12.5 minutes downloading its 3.2 GB
archive (four shards at once over one connection). There, each shard now builds the tests in
the runner's warm build directory (the existing build-tests one) and runs its slice directly;
Build tests runs only where the shards run on GitHub's or the untrusted runners.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Same-repo pull requests ran their heavy jobs on the three fast runners too, and the merge
queue's test shards waited behind them for hours. Pull requests now use the GitHub-hosted path.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The trusted runners' guard refuses issue and pull_request_target events by design (they carry a
write token and handle untrusted input), so the bots failed there. They go back to GitHub's runners.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
They check out main's own scripts, never pull request code, and were holding dozens of
GitHub-hosted runner slots that pull request builds need.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
#2278 dropped the golden sample from pull requests, and two geometry PRs then
broke main together. Pull requests that touch the kernel, engine, doc, sketch
or geom crates now run every seed, split over the shards, with the command to
re-record in the failure output.
Fork PRs whose number ends in 7, 8 or 9 stay on the untrusted pool; the rest
build and test on GitHub's runners, which have free slots.
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Build tests on the untrusted runners (fork PRs, gVisor on shared boxes) fail intermittently with
rustc killed mid-crate ("exit status: 2", no diagnostic) on the biggest crates: solidcraft-engine,
ui-egui. Nothing limits Cargo's parallelism, so it starts one rustc per core the sandbox reports.
scripts/ci/build-jobs.sh sets CARGO_BUILD_JOBS from the memory available (and the cgroup limit,
where lower) at 3 GB per job, capped at nproc, and logs the numbers it saw.
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
Co-authored-by: jyviko <2501648+jyviko@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Fork Build tests jobs were hitting the 45-minute limit while uploading their archive, after
the build had finished, and every one had to start over.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The trusted runs' test archive is still built on GitHub's ubuntu-24.04
runners; minisforum's newer glibc runs it unchanged, so shards 3-4 can join
1-2 on the self-hosted runners.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Fork pull requests now run on the untrusted self-hosted runners, which start
every job empty. Let them restore main's rust-cache like GitHub's runners do
(save-if stays main only, so a fork job never writes a cache).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Lists and Packaging lint each took their own runner slot on every PR push. The list checks now run
in the existing checks job; Lists stays on main. Packaging lint skips PRs that only touch icons.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
With Windows off pull requests, GitHub's 20 runners sat idle while every
fork job queued for the self-hosted pool. Select checks, repository
checks, Clippy, doc tests and the summary go back to GitHub's runners;
Build tests and the test shards stay on the isolated self-hosted pool.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Pull requests now run the repository checks, Clippy, the affected
crates' tests in 2 shards, and the doc tests. Windows, the reference
tests and the robustness goldens run on pushes to main and in the
nightly, where a break is fixed forward. This is the split Bevy, Zed
and Ruff use, and it cuts each pull request's CI to about a third.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Blacksmith isn't provisioning runners: its dashboard says the account
"requires additional verification", so every job labelled blacksmith-*
stays queued with no runner. Fork PRs' Linux jobs go back to the
self-hosted-untrusted pool (#2249/#2252) and the rest to GitHub's runners
until Blacksmith verifies the account; then #2237 can be re-applied.
This reverts commit dc40f051b1.
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
Co-authored-by: jyviko <2501648+jyviko@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Fork pull requests, the nightly run, Run workflow, the test archive and
test shards 3-4 move from ubuntu-24.04 to blacksmith-<N>vcpu-ubuntu-2404
(16 vCPU for Clippy and the test archive, 8 for test shards, doc tests,
reference tests and cross-compiles). Trusted runs keep the self-hosted
routing. Blacksmith runners count as self-hosted to GitHub, so steps now
tell our own machines apart by runner name.
TEMPORARY: this branch routes its own run like a fork pull request.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: jyviko <2501648+jyviko@users.noreply.github.com>
Every fork run waited in GitHub's shared queue just to start. The gate
and summary jobs now run on the isolated runners as well; the local
output stays false for forks so later jobs never pick the trusted pool.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Fork PRs were waiting hours for GitHub's shared runners. Their Linux
jobs (checks, Clippy, Build tests, the test shards, Doc tests, Reference
tests) now run on the ephemeral, network-jailed self-hosted-untrusted
runners, without the trusted runners' warm build directories.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Pull requests from branches of this repository share the work with the self-hosted runners like main does; fork pull requests stay on GitHub's runners. Select checks and ci run on a light self-hosted runner; the repository checks, Clippy, doc tests, reference tests and test shards 1-2 on the fast ones. The test archive is still built on GitHub's runners (glibc).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The merge queue runs the full suite on exactly the commit it then puts on
main, and main's push run then ran the full suite again on the same files.
Select checks now asks the Actions API whether a passing merge_group CI run
exists for the pushed commit; if so the push runs only the repository checks.
A push that skipped the queue, or an API that cannot answer, still runs
everything. The nightly still runs the whole suite.
Pull requests also run less: the reference tests only when solidcraft-cli (or
anything it depends on) is affected, and doc tests only for the affected
packages that have a library. The ci check requires the reference tests only
when they were selected.
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
Co-authored-by: jyviko <2501648+jyviko@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Clippy, the doc tests and test shards 1-2 of pushes to main and merge-queue
groups run on the self-hosted runner, alongside GitHub's runners for
everything else. Pull requests, the nightly run and Run workflow stay on
GitHub's runners. Clippy and the doc tests keep warm build directories
outside the checkout (scripts/ci/warm-target.sh).
The test archive is still built on GitHub's runners: the self-hosted
machine has a newer glibc, so binaries built there don't start on
ubuntu-24.04.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Only the Windows jobs still wait for GitHub's runners on main and in the
merge queue. The Vulkan packages are preinstalled on the self-hosted
runner (no sudo), so that step runs only on GitHub-hosted runners.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Pushes to main skip boolean_props (random inputs); the nightly property-tests job runs it. The three shrunk cases are ignored tests with their issues (#1647, #1648, #1649), and proptest's failure file is named outright.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Comment /claim on an issue to be assigned and get the in progress label; /unclaim releases it. A daily sweep frees claims with no PR after 72 h or with a PR idle for 14 days.
Message catalogs sorted one message per line (cargo xtask lists [--fix], checked in cargo xtask ci and the Lists workflow); the language list one per line, English first then by id; THIRD-PARTY-LICENSES.txt and docs/features.md rewritten on main by the Generated files workflow instead of in PRs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A PR that saves something new adds crates/doc/src/format/pending/<name>.rs (a note, and an upgrade step if older designs must be rewritten). The build writes the next format number while changes are pending and lists them in format_pending; scripts/fold-format-pending.sh fixes them as v<N>/ at release time. Formats 23 and 24 move into v23/ and v24/.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Clippy, the test archive and the test shards cover the packages a PR changes plus their dependents. Main, nightly and manual runs stay full.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* Windows: resize border from the first frame, DPI-correct window moves (#1381)
- The egui resize zones were laid out after the start page's early return, so a launch without
a design had no resize edges at all. They now run on every page.
- Windows: the main window's procedure answers WM_NCHITTEST with the resize edges and corners,
Windows' own border width at the window's DPI (8 px at 100 %), so the OS shows its resize
cursors and runs its own resize (snap included) from the first message; no egui zones there.
- Windows: WM_GETDPISCALEDSIZE returns the size winit then applies (same logical size), so the
rect the OS suggests while the title bar is dragged onto a monitor with another scale is the
rect the window gets.
- Zone sizes and hit test shared in titlebar.rs and unit-tested; Windows CI checks the hit test
right after launch and a real edge drag.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Windows resize smoke: find the main window by its title
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Windows resize smoke: probe the maximized window at the screen's edge
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Action matrix: nightly shards cut by measured cost, heavy designs in slices, and a determinism check
The full matrix ran as four fixed quarters of the corpus and two of them hit the 60-minute
job limit. Now action_matrix_shard runs shard k of n (ACTION_MATRIX_SHARD=k/n), planned from
tests/action_matrix/costs.json: designs that take more than five minutes are split into slices,
and the pieces go largest first to the least loaded shard. The nightly runs 12 shards and uploads
each one's outcomes and timings; cargo xtask action-matrix --from DIR --bless records them in the
baseline and costs.json without running the matrix again. action_matrix_determinism runs a sample
twice and lists any action whose outcome changed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Action matrix: costs measured on runners, 10 shards, determinism on the lighter designs
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Action matrix: baseline and costs from a full run on main, 16 shards, timeouts listed
The baseline is regenerated from the 10 shards' results of a full run on 01b1b705
(cargo xtask action-matrix --from DIR --bless): 8 new designs, 302 actions that pass now,
names that changed, and 40 actions that passed before and fail now, each listed in #1255
(22 fillets refused since #949, 4 hinge blends since #1101, 9 on the loft cut, 5 single).
costs.json now holds every design's measured time, and the nightly runs 16 shards of about
17 minutes. An action that passes in the baseline and times out is listed but not judged,
and stays passing when blessed: two runs of the same build differed only on 10 actions that
took 50-84 s against the 90 s limit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Action matrix: slice heavy designs by a hash of each action's name
Slicing by index gave some slices only fillets on oracle 41 (28 slices, a multiple of the
actions per edge), three times the planned time; two shards ran past their limit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A crash inside a graphics driver left only "an unclean exit between
commands": no backtrace, no module, no hint that it was at startup.
- The crash journal now keeps the startup step the app is in (window,
graphics instance, surface, adapter request, adapter chosen, device
creation, renderer and viewport shader compiles, surface configure,
first frame and its present, running) and the renderer as soon as the
adapter is picked. egui-wgpu tells the app about its steps through a
small hook. An unclean exit before the first frame names the step.
- Windows: an unhandled-exception filter (src/native_crash.rs, the one
Windows-only module allowed unsafe) writes the exception code, the
faulting module and offset, and a minidump next to the journal; the
next start puts them into the report.
- The Windows CI job runs a child that faults on purpose and checks the
report.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Publishing a non-prerelease release runs winget.yml, which opens the
winget-pkgs pull request for bherbruck.SolidCraft with Komac (pinned
2.16.0 binary, SHA-256 checked) from the three versioned MSIs. It skips
cleanly without the WINGET_TOKEN secret or before the package's first
version is in winget-pkgs. docs/releasing.md has the one-time token
setup; the README's Download section gains the winget install line.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Windows: start Direct3D 12 first, and Vulkan only without a D3D12 GPU
wgpu started Vulkan, D3D12 and OpenGL together and preferred Vulkan, so
every Windows start loaded all installed Vulkan drivers. An old Intel
Vulkan driver (igvk64.dll) crashed 0.10.1 at startup inside the driver.
The app now probes D3D12, then Vulkan, then OpenGL, one backend at a
time, and starts only the first with a hardware adapter (WARP when none
has one). WGPU_BACKEND still wins; graphics.json's "backend" sets the
order by hand; the software renderer (safe mode, the retry) uses D3D12
alone. A new Windows CI job starts the real app under several settings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* CI: a test filter YAML parses
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* CI: the startup smoke test writes graphics.json where the app reads it
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* CI: run the WARP shader tests one at a time
WARP's shader JIT (d3d10warp.dll) crashed twice in 124 runs of the gpu
tests with four test threads, with a null read inside its own optimizer
while several devices compiled compute pipelines at once. Serial runs
did not crash.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>