Python tool that finds drawings resembling a hand sketch, photo or MS Paint
outline. It indexes the OpenNest server's drawing library (read-only) and/or
DXF folders, compares DINOv2 embeddings of normalized renderings across all
8 rotations/mirrors, and writes an HTML page of the closest matches with an
optional overall-size filter.
train.py fine-tunes the embedding contrastively on synthetic sketches of the
library itself (no labelling); evaluate.py measures recall on held-out
drawings or a folder of real sketches. On 40,757 server drawings, held-out
synthetic sketches find the right shape in the top 20 75% of the time,
up from 18% zero-shot.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The panel gains a Current plate only / All plates selector, shown when the
two scopes differ. Changing it takes the unapplied scrap cut-offs off,
cancels any running plan and replans. The toolbar button opens with the
scope last chosen there (default all plates); the Plate and Nest menu items
keep their fixed scopes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each lead tab draws a sample part with a hole and the configured leads and
pierce points, emphasizing the tab's own contour and updating as values change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Auto Nest leaves out ledger item codes already on other open server nests,
and save and post re-check a nest's codes before going on.
- Server: GET /api/nests/stock-item-codes?exclude= lists the codes on
non-deleted Draft, Quote and To Be Cut nests. Codes are stored in
nest_stock_codes (schema v14) at every archive write; the upgrade reads
each current archive once.
- Data: INestRepository.GetStockCodesInUseAsync, LedgerCodesInUse (5 s,
failures become an unknown answer), the pure LedgerCodeFilter for ledger
rows (reusable by release-time nesting) and LedgerCodeCheck for the
re-check.
- Desktop: the Auto Nest ledger summary counts "(N on other nests)" or says
other nests were not checked. Save, Save As and post open Ledger Stock
Changed when codes are taken: proceed after a per-attempt
acknowledgement, remove the codes (undoable), or cancel. A ledger that
cannot be read skips the check with a status note.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WinForms user settings lived in user.config under %LOCALAPPDATA%\OpenNest,
the Velopack install folder that setup clears on a reinstall, and the
default provider starts empty at each assembly version. Store them in
%APPDATA%\OpenNest\user-settings.json instead, carrying over any values
the old user.config still holds on first run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add a Support Bed tab beside Clamps that edits the grid or slats new
nests assigned the machine receive. A machine without a bed shows the
legacy support-grid.json values; the first edit stores them on the
machine in its units.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first background check waited the full 30-minute interval, so
sessions shorter than that never checked or downloaded an update.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Set it as the exe's ApplicationIcon and as the main and help windows'
icon (the main window showed the WinForms default). The .ico is
rendered from OpenNest.svg as 32-bit images from 16 to 256 px.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The status bar shows the running version, and the Help menu gains
Check for Updates (an immediate check sharing the background check's
lock) and About OpenNest. A copy that is not a Velopack installation
offers to open the GitHub Releases page instead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The trailing location update ran on a WinForms timer, and WM_TIMER is
only delivered once input and paint messages drain, so the label lagged
while the mouse moved and after it stopped. Moves now drive the 100 ms
throttle from a timestamp, and a 30 ms settle timer restarted by each
skipped move shows the final point. Unchanged text is not reassigned.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
VelopackApp keeps its default auto-apply on startup, so an update the
user did not restart for installs when OpenNest next opens. Also note
that nothing runs while it is closed and the first check waits 30
minutes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With GitHub as the default source, UpdaterService always built an
UpdateManager, which throws unless VelopackApp.Build().Run() has set a
locator. MainApp does, but WinForms test hosts constructing MainForm do
not, so 78 OpenNest.WinForms.Tests failed. Without a locator there is no
installation to update, so the service now stays idle.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
dev-build.yml runs after Cross-platform tests pass on a master push,
packages that commit as X.Y.Z-dev.N (the minor after the newest vX.Y.Z
tag, N the commit count) and replaces the single rolling dev-build
pre-release. Only its publish job has release write access, and a run
whose commit is no longer master's head publishes nothing.
Tools > Options > Include development builds lets an installed copy take
that pre-release; everyone else still sees full releases only.
Publish-Windows.ps1 accepts -dev.N versions.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Publish-Windows.ps1 now packs the smoke-tested folder with the pinned vpk
CLI (.config/dotnet-tools.json, 1.2.0 to match the Velopack package) into
velopack/: Setup.exe, Portable.zip, the full nupkg, releases.win.json and
RELEASES, checked for the full package and the portable layout and hashed
into OpenNest.vX.Y.Z.velopack.sha256. The release workflow uploads them
with the candidate.
With no feed URL set, installed copies now check GitHub Releases (full
releases only) instead of skipping update checks; a URL in Options still
selects a self-hosted feed. Deploy-OpenNest.ps1 uses the pinned vpk.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A sheet separator is cut after everything, so when the approach to its
pierce crossed parts already cut, the usual "cut it before them" rule
contradicted that and the search learned nothing. It fell back to the
full search, which spent the budget or found a route that climbed the
column and jumped back down past the cut parts (N26-23M).
When every "cut the blocked part first" rule would contradict a required
order, learn the other way round: cut the part the approach left from
before the parts it crossed, so the tool arrives from beyond them. A
separator blocked this way now ends the attempt at once, like any
cutoff whose rule can be learned.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At the bottom of the form the search box read as part of the Details
panel. It now sits under the Nests header, next to the list it filters,
and comes first in the tab order.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Plan Cutting ranked a convex corner first, but a corner rounded by a small
fillet had no corner vertex: its tangent joints ranked level with straight
edge midpoints, so travel decided and a large part (P230323-02 plate 8,
PT06, 0.015 corner fillets) started in the middle of a 64-inch edge. A start
there can leave a step where the cut closes after the head has put heat into
the plate; at a corner it is hidden.
A convex fillet tangent to straight edges at both ends, turning less than a
half circle and reaching no further along either edge than its length, now
makes its two joints convex-corner entries. Each carries its reach from the
sharp corner it rounds, so ranking counts it on both sides of that corner.
Longer curves, half-circle ends and inside fillets keep their tangent joints.
Every perimeter on that plate now starts at a corner; the problem-nest corpus
has no plate below its baseline.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On P230323-02 plate 7 the frame part cut after a cut-off segment took its
upper small hole first, so the rapid up from the segment's end at the
bottom of the sheet crossed the rapid from the lower hole to the window.
The free search plans parts as if segments were absent, so the part's hole
route was ranked from where the tool was before the segment (the previous
part's end, top right), not from where it really is.
After CutOffInsertion places the segments, each holed part cut straight
after one now keeps its outside entry and order but takes the hole route
and preferred hole entries planned from the segment's end, when that
chain has an entry for every hole and passes the checks. The search's
look-through of segments keeps ranking outside entries from the
pre-segment point (so routes stay as planned without the lines) but plans
hole routes from the real position, so a kept-order replan of an applied
plan picks the same holes. Plates without cut-offs are unaffected.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The support overlay was drawn inside the live-plate-only paint block, and
the Plan Cutting preview plate never copied the support bed, so opening
the cut planner hid the grid points and each part's support marks. The
overlay now follows the displayed plate, the preview carries the bed, and
tab-point clearance falls back to the live plate's tab size.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Drawing.Bends stay in the DXF's source frame while CadImporter moves the
program so its first pierce sits at the origin (Source.Offset). The plate
renderer added part.Location to raw bend points, so every bend and etch
mark was shifted by its drawing's pierce offset, visible on the drawing
library set preview and on plates alike. Remove Source.Offset before
rotating and placing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Publishing a GitHub Release, stable or pre-release, now only validates the
server image. A desktop release and a server image release are separate
owner decisions, so a stable desktop release cannot publish an image that
has not been accepted yet.
- archived library drawings: say they can't be edited or released; no
front end can restore a library drawing yet
- library filter hint names the Show all materials checkbox
- undo labels Support Bed and Automatic Scrap Cutoffs match their menus
- display-outline posting hints name both Plate > Plan Cutting... and
--plan-cutting, since the console prints them too
- MaterialLedger options are saved per Windows user, not per machine
- undo-redo: describe the heaviest nests by size instead of job numbers;
name the Nest > Edit dialog rather than the internal 'Nest Info'
- cutting-planner: clearing lead-ins and cut-offs is now undoable
- nest-storage: drop the internal D4 label and the nonexistent
Nest > Nest Info menu path
- Save to Drawing Library: the stretching row was Nesting instead of Notes,
leaving Notes one line high between empty bands.
- Nest Defaults: a read-only server copy disabled Save but left the fields
editable, so edits were silently lost; lock the fields too.
- Support Bed: name the Cutting Settings option by its actual label,
"Only tab parts the support grid can't hold level".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Clamps are physical fixtures that differ between machines, but one global
clamps.json served every machine, so a plasma table and a fiber laser could
not have different rails.
- MachineConfig.Clamps holds each machine's rail in the machine's units,
edited on a new Clamps tab in Tools > Machine Configuration.
- New and BOM import give a nest the rail of the machine they assign;
Nest > Clamps loads and saves the nest's machine's clamps. Each nest keeps
its own copy, so machine edits never move existing nests' clamps.
- At startup, while no machine has clamps, the legacy clamps.json (or the
server's shared copy) is copied once into the default or only machine;
with several machines and no default the user picks one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add OpenNest.Desktop (net8.0), the toolkit-neutral layer the WinForms app
and OpenNest.Avalonia will share, and OpenNest.Desktop.Tests with a guard
that the library references no WinForms/WindowsDesktop, System.Drawing.Common
or Avalonia assembly, project setting or (transitive) package.
Both desktop front ends reference it; OpenNest.Avalonia joins OpenNest.sln
and the new tests join the Linux CI matrix. No code moves yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
OpenNest.Gpu uses no System.Drawing, P/Invoke or Windows API: PartBitmap is a
plain int grid and ILGPU's CPU, CUDA and OpenCL backends are cross-platform.
GpuSlideContactTests pass on Linux against the net8.0 build.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
FrontEnd.Tests only matched MCP's old net8.0-windows marker; MCP and every
project it references are net8.0 and it never used WindowsDesktop.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds OpenNest.Avalonia, a throwaway Avalonia 11 app kept out of
OpenNest.sln, to gauge a Linux port. It reads nests through
OpenNest.IO and draws plates with Skia: a port of the desktop's
display paths (cuts, etches, manual lead-ins, filled material),
pan, cursor zoom, fit, hover and selection, beside a plate list and
a drawings DataGrid. --snapshot renders the window to a PNG.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Shape Library, drawing library and drawing set preview each drew
parts their own way (hardcoded green fill, steel-blue fallback, black
outlines with gray etch). They now share a library style on ColorScheme
(background, fill, outline, etch and bend colors) that ignores drawing
colors, follows the active scheme (Dark gets a dark variant) and
round-trips through saved schemes, defaulting when absent. Nest part
colors are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Replace the name list with a ListView of 96px thumbnails drawn by
Program.GetImage at each shape's preview defaults (or its first
configuration), labelled with the name and configuration count. The
left pane opens three tiles wide instead of half the window.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Server schema v12 (2d8d877) adds a units column, so migrated legacy
records now report units: "inches" and the Server image smoke's
legacy-vs-migrated manifest comparison failed (cmp differs at line 27)
on every master push since.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
FillItem_WithClampZones_PinnedLayout (Fill, Strip, Vertical Remnant) has
failed in the Engine.Tests CI job on every push since it was added in
ccd361d: the pinned digest 9BF5FC7FC47364A9 is not reproducible at that
commit or any later one, locally or on GitHub runners. The engine
deterministically yields 9B3EFF9B72243576 (Fill, Strip) and
039C0D607166D805 (Vertical Remnant) at ccd361d, 36c8de8 and master, so the
later filler consolidation is still layout-neutral for these cases.
Before repinning, all four strategies' layouts were checked: 40 parts, all
inside the work area, bounding boxes clear of both clamps, and no
overlapping parts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Small parts such as 1.5" test squares on a 2" laser grid rest on at most
one grid point, so no single support tab can bring their center of
gravity inside the contacts and Plan Cutting warned "no single tab can
steady it" on every one. In practice a part that short stays in the
sheet on its tab, so the bed's contacts don't matter.
Cutting Settings > Tabs gets "Tab alone holds parts up to N long"
(CuttingParameters.TabHeldMaxLength, 0 = off, the default). On a plate
with a bed, a part the bed can't hold level whose length (long side of
its minimum-area bounding rectangle) is within the limit is still
tabbed, but any tab clear of inside corners steadies it, so Plan
Cutting, hand-placed lead-ins and the pre-post review warn only when it
has no tab at all. The setting persists with the cutting parameters and
is part of the plan state fingerprint.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Nest drawings can require no parts, but the CAD Converter's quantity
spinner had a minimum of 1, so Edit Drawings in Converter threw
ArgumentOutOfRangeException as soon as such a drawing was selected.
Lowering the minimum to 0 lets the converter load it and hand the
drawing back with its quantity still 0.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Undo/Redo landed in phases 1-3 without a written contract, so the next
person adding a desktop edit has no rule telling them to record it, and
nothing explains why a restore must close Cutting Inspection or why
snapshots keep empty plates and skip best fits. docs/undo-redo.md states
what is and is not undoable, the BeginChange/RecordChange rule, the
"Change" safety net, how states are captured and restored, the guards
and key routing, and the measured limits. AGENTS.md links it from the
OpenNest bullet and the contract list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The nest's machine and its Nest Info fields were the last edits only the
fingerprint safety net could catch (and it misses most of them). Machine
is now one step, together with the support grid and cut-off settings the
machine brings. Nest Info is one step for what Undo can put back (units,
material, thickness, customer, notes and plate defaults); the name,
status, author and creation date stay as they are through Undo, so an
edit of only those records nothing.
The sheet strip only shows Auto Nest proposals; what it shows is
committed through Auto Nest, which is already recorded.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Drawing changes were unrecorded, so Undo skipped them or lumped them
into an anonymous "Change". Each is now one labelled step: Import
Drawings, Shape Library (menu and panel button), Drawing Library (what
the dialog added or re-quantified), Save to Drawing Library (the nest
keeps the drawing's new library link), Edit Drawings in the converter,
Drawing Properties, Delete Drawing and Clean Unused Drawings. Dropping a
drawing on a plate was already recorded as Add Part and is now tested.
The converter can change live programs while it loads, even when it is
cancelled; a cancel that did so now rolls the nest back to before the
converter opened instead of leaving those edits unrecorded.
Deleting a drawing threw when removing its parts emptied the trailing
plates, because the plate manager drops trailing empty plates while the
loop walked the live plate list; it walks a copy now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cutting edits were unrecorded, so Undo skipped them or folded them into
an anonymous "Change". Each is now one labelled step: Add Cut-Off, Move
Cut-Off (one step per drag), Automatic Cut-Offs, Auto Sequence (current
plate or all plates, only when the order changed), Set Sequence (per
click), Sequence by Path (the whole tool session; Backspace and Ctrl+Z
inside it still take back one line), Place Lead-In, Remove Lead-Ins (the
tool's menu, the plate and the nest commands), Cutting Parameters (an
edited lead-in panel when the tool closes), Plan Cutting and Clear
Lead-ins and Cut-offs.
Plan Cutting stages scrap cut-offs on the live plates before it plans,
so its scope opens with the staging and records only when the plan is
applied; closing without applying records nothing. Clearing applied work
is its own step and no longer warns that it cannot be undone.
Opening the lead-in tool used to give a plate without parameters the
defaults straight away; the plate now takes the tool's parameters with
its first lead-in, so opening and closing the tool changes nothing.
Undo no longer waits while the tool has a part picked: it ends the tool.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Plate and clamp changes went unrecorded, so Undo skipped them or lumped
them into an anonymous "Change". Each is now one labelled step: Edit
Plate (only on OK), Rotate Plate, Resize Plate, Set Nest Default Plate,
Clamps and Support Grid (applying to the nest; saving machine defaults
stays outside the history), Remove Clamp / Put Clamp Back from the
plate's clamp menu, and Add/Move/Delete Clamp in Edit Clamps.
Edit Plate used to copy an empty plate's settings into the nest defaults
even when the dialog was cancelled; that now happens only on OK, so a
cancelled dialog changes nothing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Placement edits were only caught by the fingerprint safety net, so they
undid as an anonymous "Change" or folded into the next edit. Each now
records one labelled step: Fill (at the commit, so a stopped fill that
keeps nothing records nothing), Fill Area, Pattern Tile, Auto Nest (from
the kept proposal only; Make Nests tabs too), Clear Nest, Remove Empty
Plates and Center Parts. Clear Nest no longer warns that it cannot be
undone.
Pattern Tile onto a new plate created a plate beside the empty sentinel,
which the plate manager dropped at once, so the pattern landed on a
detached plate; it now fills the sentinel.
The undo tests share one harness that compares a text description of the
nest before an edit and after Undo.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Whole-nest snapshots make undo safe without a hand-written inverse per
operation: each recorded edit stores one NestSnapshot in the document's
history, and Undo/Redo read one back and swap it into the open window.
- Document.History is reset with the opened nest; Save and database
saves mark the save point instead of clearing it.
- EditNestForm.BeginChange opens a scope: nested scopes merge into the
outermost, Complete commits one entry, and leaving it incomplete rolls
a changed nest back. No-ops are detected by the scope's changed flag,
not by comparing bytes (copies are not byte-stable).
- RestoreNest shares the per-nest binding with the constructor and keeps
the Document, the nest's identity, the plate index, the viewport and
the selection. It ends the action, closes Cutting Inspection, side
panels and the remnant viewer, rebuilds the PlateManager, reapplies
the view's cut-off settings and recolours the drawings, because each
of those holds or derives from the old nest's objects.
- An edit no scope recorded (Phase 3 sites) is detected by a cheap
fingerprint and committed as its own "Change" entry before a scope,
Undo or Redo, so it is never silently discarded; DEBUG builds also
warn about it on idle.
- Edit > Undo/Redo name the next entry. Ctrl+Z/Ctrl+Y/Ctrl+Shift+Z go
to focused text editors first, then to an action's own step-back
(Sequence by Path), then to the nest. They wait for jobs, fills,
background saves and busy actions, but not for a plain selection.
- Recorded: delete, drag, arrow nudges (coalesced), rotate, align,
push, expand spacing, add part (drop and clone), cut, add and remove
plate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Undo will keep whole-nest states and restore by reading one back.
NestSnapshot.Capture writes the same archive a save writes, with three
differences Phase 0 found necessary: it does not stamp DateLastModified
on the live nest, it leaves best fits out (they are a cache), and it
keeps empty plates, which files drop, so an undo does not shift plate
indices or lose a plate the user added. File output is unchanged.
Restore reuses each live drawing whose saved content is unchanged,
matched by a fingerprint over its nest.json entry, program, hole
sub-programs and source entities. BestFitCache and FillResultCache are
keyed by the Drawing object, so a plain read would make the next fill
recompute every best-fit pair; reuse keeps those caches, colours and
Drawing.Id, and skips parsing the reused programs. Nested counts are
recomputed for the restored nest, and given back to the live one if the
read fails.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
These round-trip tests pinned the old behaviour, where a part with a
saved lead-in program read back with null cutting parameters. Since
7abf0701 the part keeps the settings its leads were built with, as an
owned copy that stays separate from the plate's settings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The desktop app has no undo. Phase 0 showed whole-nest snapshots are
cheap enough to record after every edit, so the history keeps whole
states with a cursor instead of per-command inverses that each need
hand-written reversal code.
It drops the redo tail on commit, evicts the oldest states past 50 steps
or 128 MB without ever evicting the current one, tracks a save point for
a later dirty indicator, and merges matching commits within a second so
repeated nudges undo as one. The clock is a TimeProvider so tests control
the coalescing window. It is generic and free of IO and WinForms so it is
unit tested on its own.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PartDto had no field for Part.CuttingParameters, and restoring a saved
program cleared it, so after save and reopen the lead and pierce-zone
reviews judged manual and planned leads against the plate's settings
instead of the ones they were built with. Undo/redo will restore nests
through the same reader, so it would have carried the same loss.
The settings are written only beside a saved part program and read
back after the program installs. Older files and clean parts read null,
as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ajandClaude Opus 5.5