* CI: merge-queue groups run on Blacksmith, test shards on 16 vCPU
The merge queue builds up to three groups at once, but every group's
checks, Clippy, doc and reference tests and all four shards ran on the one
self-hosted machine, so the groups waited for each other. Merge groups now
run on Blacksmith like pull requests (test archive built once, shards on
16 vCPU). Pushes to main keep the self-hosted runners and their warm build
directories, and a push the queue already tested still runs only the
repository checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
* CI: every job on Blacksmith: pushes to main and the Windows jobs too
Pushes to main ran on the one self-hosted machine and the Windows jobs on
GitHub's runners. Both move to Blacksmith (Windows Server 2025, 8 vCPU),
so nothing in ci.yml depends on a single box any more. With no
self-hosted job left, the local routing goes: no warm build directories,
every run builds the test archive once and the shards run it, and main
saves the build caches every other run restores.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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>
* claim: #2621
* Sketch: dragging part of a selection moves the whole selection (#2621)
A drag that started on a selected sketch line or point moved only that
entity; the rest of the selection was left to the solver, so geometry that
kept only its length (the reporter's 5 mm tail) trailed behind like a string
and chamfers stretched.
- sketch.drag takes `with`: the rest of a selection dragged by `entity`.
Every free point of all of it slides by the cursor's move and the held
solves keep them all (circles and ellipses keep their radii), so the
selection moves as one (`plan_group`). `with` is checked and capped like
the other parameters.
- The viewport sends the rest of the selection when the drag starts on a
selected point or curve. A curve's own points, picked with it by a window,
are not a group: a circle grabbed on its edge still takes its radius.
- Snapping looks at all the moving points.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A sketch and the feature built from it may share a name, and the sample and
oracle designs do. `timeline.add` refused such a recorded name, so a design's
own exported script failed to replay. `Document::insert_feature` now keeps the
name, as `add_feature` does; a blank name still gets a fresh one.
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
Deleting a pocket or a boss from an imported part failed with "face N cannot
close: cannot attach a plane to a wire that is not on one plane". The face
around the gap only loses an inner loop, but it was rebuilt as a new plane
through its outer boundary, and an imported boundary is often a few 1e-5 mm off
the plane: within the planar test, too far for a plane fitted to it. Such a face
now keeps its surface and only drops the loop.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
Replace Face rebuilt its body with ModelBody::new and dropped the body's
lineage, so no face of the result was named exactly: a sketch on any face of
it, even one the replacement didn't touch, failed with "this face can't keep
an attachment". Each face it moves is now a lineage step, as in Offset Face
(which does the same geometry), and the result keeps the lineage.
The test of a body without history used Replace Face as its example of an
operation that reports none; it now lifts the block with Align instead, the
same geometry.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
* claim: #1985
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L15TLCjpXsQdp8ovC4ZacH
* STEP export no longer hangs on a box with a screw hole (#1985)
The export reads its own STEP text back, and meshing the read-back screw
hole never finished. On a seam cylinder the boundary walk projected points
to angles up to 5e7 rad apart and moved each by two turns at most, so the
11-point loop spanned millions of turns and the polyline closing it across
the seam had 4.3 million points. Each projected parameter is now brought to
within half a period of the point before it, whatever the gap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L15TLCjpXsQdp8ovC4ZacH
* Second #1985 regression: the reporter's flanged enclosure
The issue's reporter delta-debugged a second design whose STEP export hung
on main: a flanged enclosure with two standoffs and no screw hole. It hangs
without the get_mindiff change and reads back with 29 faces and the same
volume with it. Both tests share the read-back checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L15TLCjpXsQdp8ovC4ZacH
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
* claim: #3052
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Sketch Trim: the hover shows the piece a click takes away (#3052)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Sketch Trim: the piece is shown in the error colour, thicker
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Trim hover: preview keeps its doc line; the piece is worked out once a frame
The new function had gone between preview's doc line and preview. The
cursor's Trim piece is now kept by design revision, sketch, curve and
cursor, so the viewport's hover and the preview no longer both intersect
the curve with every other curve each frame.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Phanoob <phanoob@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* parameters: import evaluates once; parameters.change takes several names
parameters.import checked every row through change_param, which cloned the
design and evaluated every parameter per row: 2000 rows took 9.5 s, 2823 in
an assembly minutes. Rows are now checked and applied in one pass with the
values carried along, then evaluated once (2000 rows: 0.016 s), still all or
nothing; a row may name a parameter defined by a later row.
parameters.change {changes: {name: expression, ...}} sets several existing
parameters together, rebuilt once and undone in one step.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* parameters.change: the params doc names the `changed` result of `changes`
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
The self-hosted-untrusted pool is being retired for CI, so fork PRs whose
number ends in 7, 8 or 9 no longer route there. Build tests, the test shards
and the reference tests reduce to our own runners for main and the merge
queue, Blacksmith for everything else; the build's 90-minute gVisor timeout
goes. Blacksmith VMs are ephemeral and fork runs get no secrets.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
Blacksmith runners report runner.environment == 'self-hosted', and a list of
our machines' name prefixes breaks when a box is added or renamed. Select
checks already decides where each job runs; its local and untrusted outputs
(and the trusted-run condition in the repository checks) now drive the warm
build directory, cache and lavapipe steps. Same results as before on every
event.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
Re-applies #2237 (reverted in #2277 while the Blacksmith account awaited verification) on today's
ci.yml: every ubuntu-24.04 job or fallback runs on blacksmith-<N>vcpu-ubuntu-2404, sized as #2237
did (quick-compile, new since, gets 8), and steps that asked runner.environment now tell our own
machines apart by runner name, since Blacksmith runners report 'self-hosted' too. The trusted
self-hosted routing and the fork routing are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
From the review: the guidance said a pull request's run tests it merged with the latest main,
which is only main as of that run; a break between two pull requests that land close together
shows up in main's own run. It also forbade rebasing. Now: don't push only to catch up; when you
push a change anyway or GitHub reports a conflict, rebase onto main (--force-with-lease on your
own branch) or merge it in once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
From the review: the web app embeds ATTRIBUTION.md, so the Docker image's Markdown exclusion now
keeps it; and THIRD-PARTY-LICENSES.txt also reads the vendored crates' licence files and the
fonts' OFL texts, so changes there regenerate it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
From the review: git patch-id --stable ignores whitespace, so a conflict resolution that only
respaced the change (inside a string, or Python indentation) counted as no change and was
carried; --verbatim compares every byte. And a passing run is now taken only from this pull
request (by number, or for a fork, where GitHub lists none on the run, by head repository and
branch), not from another pull request with the same head commit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
From the review of this PR:
- What counts as code lives only in scripts/ci/affected.py now: nothing selected and not full
means no code changed, on pull requests and pushes to main alike. The untested bash copy in
Select checks disagreed with it. Root files and folders a crate embeds now select that crate
(the licence texts and NOTICE: the engine; skills/: the MCP server's tests), where before a
change to them ran no test at all. Other workflows and the bots' scripts select nothing.
- A push to main skips the native jobs only when nothing a build reads changed since the last
push to main whose run passed. Comparing with the push just before was unsafe: a newer push
cancels a waiting run, so that commit may never have been tested.
- Fail fast applies to pull requests only. On main, nightly and Run workflow the test build and
the shards run regardless of the repository checks, so the action matrix and property tests
that hang off Build tests keep their signal. Doc tests (main and nightly only now) no longer
wait for the checks.
- tests/robustness/ turns on the goldens, so a change to goldens.json runs the seed tests.
- tests/oracle/ no longer selects the kernel, which never reads it.
- The workflow header and CONTRIBUTING describe what main and pull requests now run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
A change that merged main twice in a row ran the full set on every other push: the change before
the second merge was itself a carried run, its shards skipped, so it had no passing shards to
carry. The script now walks down the first parents through consecutive merges of main that kept
the change's patch (each equal to the next), at most 10, to the run where the shards ran.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
The guidance told contributors and agents to keep their branch current with every upstream merge
and to re-run CI after integrating. Over three days that meant 1,241 merges of main into pull
requests and 730 re-runs: about 45,000 of 79,800 ci runner-minutes, on runners that capped how
fast everyone's checks ran. CI already tests each pull request merged with the latest main, and
a real failure fails again on the same commit. Now: merge main in only when GitHub reports a
conflict, re-run only a job runner trouble cancelled, and maintainers add approved-for-ci when a
pull request is ready to merge and merge approved ones in groups by area.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
Merging main into a pull request (Merge branch 'main', or GitHub's Update branch) restarted the
whole run. Over three days 1,241 such merges started 966 runs and about 21,000 runner-minutes, and
872 of the merges brought in nothing that touched the pull request's files. Select checks now
recognises the case (scripts/ci/main_merge.py): HEAD is a merge whose other parent is on main, the
change's own patch is unchanged (git patch-id), and a ci.yml run of the change before the merge
passed with its test shards. Then only the repository checks run, the ci job names the passing run
in its summary, and main tests the merged result after it lands. Anything else, or any doubt,
runs the full set as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
Doc tests move to main and nightly, beside the reference tests and the Windows jobs: on pull
requests they were the only failing job 3 times in about 7,900 runner-minutes over three days.
Build tests, the test shards and the doc tests now start only after the format and repository
checks pass; 212 attempts in those three days ran their heavy jobs beside a failed cheap check.
Also corrects the header, which still said pull requests run the Windows jobs and the reference
tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
- Pushes to main use the same code/no-code rule as pull requests (`touches_code` in Select
checks): a push changing only documentation, packaging, other workflows or the bots' scripts
runs the repository checks only. With the merge queue off, every push to main ran the full suite
and the Windows jobs, docs-only ones included.
- affected.py: test data under tests/ selects the packages that read it (robustness, hole matrix,
projection, taper: solidcraft-engine; oracle: the CLI, the kernel, xtask; action matrix: the
CLI, the engine, xtask), and a vendored crate the packages that use it (truck: the kernel and up;
egui-wgpu: the UI and the apps), instead of the whole workspace. tests/tools and unknown vendored
crates still select everything.
- Docker (the web app image) skips pushes touching only docs, test data or bot scripts; Pages skips
test data and bot scripts; Generated files runs only when the manifests, lockfile, command
registry, CLI or xtask change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QWhMwdqJYQKQq8RCLSL1LT
* inspect.clearance: pairs whose bodies did not change are not measured again
Each clearance check measured every close pair again, though between two
checks usually only a few bodies change or move. Results are now kept per pair,
keyed by both bodies' geometry and placement and the check's settings; a pair
whose two bodies are unchanged is taken from that cache. With 40 overlapping
parts, a check after moving one of them took 0.8 s instead of 23 s (741 of 780
pairs reused). Pairs that could not be worked out are tried again each time.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* inspect.clearance: boxes and placed keys exactly as before the reuse
The pair boxes are taken over the mesh's triangles with finite corners, as a
distance shape's are, so pruning is unchanged. A placed body's identity now
includes its model body's own frame, so two placements are never mistaken for
one another.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* inspect.clearance: a large body is meshed once for all its pairs
Each pair that touches cuts the larger body down to the pair's region with a
kernel boolean, and that boolean meshed the whole body twice: once for its
checks and once inside the intersection. A sheet with 600 faces paid about
2.5 s per pair for meshes that were the same every time.
A body can now keep its meshes for reuse (Body::keeping_meshes); its booleans
take the meshes from there, and the intersection takes its shell meshes
from there too (truck-shapeops: and_with_history_meshed). Clearance uses such
copies for one check. Results are unchanged to the bit. On a 3 x 2 m sheet
with bends and 300 holes, the clip took 0.23 s per pair instead of 2.6 s, and
24 rivets took 5.6 s instead of 16.5 s.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
* claim: #2535
* claim: #2535
* Sketch: preview all slot sides and caps while choosing width
Fix missing temporary slot geometry for all five tools and paint every preview curve with a shared sampling budget. Preserve the arc slot construction guide. Add geometry, painting, invalid-input and interactive typed-width regression coverage.
* Sketch: honor typed slot widths below the pointer minimum
---------
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
* Extrude: a profile edge short of a face is not grown through the skin beyond it (#1669)
A cut or join profile edge that lies on a face of the target is grown off it by 0.5 % of the body, so that the tool's wall does not lie in the face. Whether it lies on a face was judged half a growth step either side of it: empty just outside, material just inside. An edge 0.25 mm short of a 3.75 mm wall's far face reads the same, and the growth cut through the skin it should leave (55 050 mm³ instead of 56 130 on the test's wall). Now the space just past the edge, 1e-4 of the body out, must already be as the operation wants too.
In #1669, the pleat's window built directly at the far end of a case opened through the wall at 3.75 and 4 mm, while its mirror, built without the growth, rightly left the skin and then failed to join the loft through it. Built either way, the window now leaves the skin.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Regression tests: the issue's pleats mirror and build at the far end (#1669)
The issue's battery case at 3.5 mm: the mirror of the window and the pleat builds (it needed #1670's meshing of twisted faces), and at 3 mm the same window and pleat built directly at the far end of the finished case join it (they needed #1667's end rounds).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Far-end pleat test: the direct build matches the mirror's volume
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Erikas <erkexzcx@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
A loop made only of arcs took the all-arcs shortcut in `offset_loop`, which
moves each arc by its radius and keeps its start and sweep angles: exact for
arcs on one circle or meeting tangentially, but where arcs of different
circles meet at a corner (a circle with lobes) the moved arcs no longer
met, and the sides were built through the wrong points without an error.
`arc_corners` moves each such corner to where the two moved circles cross,
nearest the old corner (`circles_cross`, closed form), and cuts both arcs
there. Circles that no longer cross are refused ("not supported yet"), an
arc the taper turns around closes the profile. Between the sections the
corner is a straight ruling, as for every side in `taper_loop`: on the test
profile (2.5 mm at 45°) it stays within 0.015 mm of the exact corner.
Test `tapered_extrude_of_arcs_meeting_at_corners`: the volume against the
integral of the union of the shrunk circles, 18905.42 mm³ against 18907.67
(was 18975.45).
Co-authored-by: Mariusz Pacek <pacekm90@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
* Snapshot: draw a see-through appearance see-through, and give the CLI --body, --tool and --section
A headless picture dropped a body's opacity: the scene carried only a colour, so a design with
a translucent appearance was drawn opaque by `solidcraft-cli snapshot`, by MCP screenshots and
in a design's thumbnail, although the window has always drawn it as glass. There was then no
way to look at the body a cut would remove inside the material it meets: only a volume to
measure and reason about, which hides exactly the errors a picture makes obvious (a tool that
stops short, or lands in the wrong place).
- `SceneMesh` carries the body's opacity; `view::build_scene` takes it from the design
(`model_body_opacity`), and a selected body stays solid so the selection colour reads.
- The GPU capture routes a see-through body into the scene's glass at its own alpha, as the
viewport does; the CPU rasterizer draws the solid bodies first (they own the depth buffer),
then the see-through ones blended over them, far sides before near sides and without writing
depth, so what lies inside shows and a line between the two is not hidden.
- `snapshot --body NAME:opacity=0.3,color=#rrggbb`, `--tool NAME` (that body solid in the tool
colour, the rest see-through) and `--section z=27` run the commands the app itself runs
(`appearance.assign`, `inspect.section`): the options are shortcuts, not a second renderer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* Tests: measure the tool colour as it is drawn through a see-through target
The tool-colour filter was derived on paper and demanded a red more than twice either other
channel, which the target in front of it tints away (measured: (200, 108, 85) through opacity
0.3); and the colour check painted the tool, which a solid target hides.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fmt
* Snapshot options: strip the colour's # rather than slicing it; README
Slicing `value[1..]` on an input-derived string is what the house rules ask us not to write,
even where a length check has already made it safe. The error a failed assignment reports now
names the body rather than claiming it came from --body, since --tool reaches it too.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* --tool: assign the bodies around the tool in groups, not one command at a time
Each appearance.assign is an undo step with a document snapshot behind it, so one per body
made --tool cost grow with the size of the design for no reason: the bodies around the tool
all get the same opacity, and the ones that share a colour can be named in one call. A design
whose bodies all wear the default grey now costs two commands whatever its size.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* Rasterizer: say where its glass and the GPU's differ
Splitting each see-through body by facing orders its own two layers; it does not sort across
bodies, which the GPU does. Worth stating rather than leaving a reader to find it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
`joint.as_built` with `at` an origin on A or B keeps that origin on its
side (found again by its face's name on each solve) and holds the frame
where it is now on the other, so the joint follows the face or hole as
the design changes. Without `at`, or with an origin on another part,
both sides hold their frames as before.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>