snaporca-ziam said gen_offer_table.py would silently delete the model-mode "Constrain sketch"
row, because that row lived in the generated header and not in tool_atlas.json. Running it found
more than that: FOUR rows existed only in the header — constrain, rename, and the three typed-
value rows sk_length / sk_radius / sk_angdist — and sk_delete's action had drifted, pointing the
sketch row at btn:delete, the FEATURE delete.
All five are now in the atlas, so the header regenerates byte-identically from it. Verbs may carry
a `note`, emitted as a C++ comment above the row: a rationale written into a generated file is
deleted by the next regeneration, which is how this started.
snaporca-z8rs (P1), found by making that true: after the atlas held all 92 verbs, the regenerated
header differed from the checked-in one by EXACTLY ONE LINE — kOfferVerbCount, 91 against 92.
Every consumer loops i < kOfferVerbCount, so the last row of the table was invisible: never listed
by show_offer_menu, never findable by mcp_run_verb. The verb that fell off the end is sk_angdist,
"Angle / distance…" — the typed-value row for a two-entity selection. On the one selection where
you would ask for the angle between two lines, the row that types it was not in the menu.
It survived because nothing compared the Sk2Ent menu against the table: sk_angdist accepts Sk2Ent
and nothing else, so an off-by-one that dropped the LAST verb was invisible from every other
selection. The vocabulary rung now covers Sk2Ent too, and picking the pair taught it one more
rig fact — shift-clicking a circle at its +X point grabs the RADIUS GRIP, which replaces the
selection with that one entity, so the pair silently collapsed to one and the offer answered
SkLine. Correctly, for the selection that actually existed.
gen_offer_table.py --check proves header == atlas and changes nothing; it is now the first step of
scripts/ladder-all.sh, and the only one that needs no rig.
Offer ladder 107/107, gesture ladder 93/93, both on the rig.
snaporca-ziam snaporca-z8rs
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MrMzTpAf78U4NG2M8jfvHY
Review request on PR #15238: "Place CAD-related files (e.g. CadDocument/
GeometryEngine) into a separate folder."
src/libslic3r/CAD/ the kernel — CadDocument, GeometryEngine, the four
Sketch* units, SketchSolver, ThreadStandards
src/slic3r/GUI/CAD/ the tab — DesignPanel, DesignCanvas, DesignSketchTool,
SketchInlineEditor, McpControl, generated DesignOffer
Pure relocation: no line of logic changes. Two include rewrites follow from it —
files that moved re-spell their own neighbours against src/ (already on the
include path), and files that did not move pick up the new folder. docs and
docs/ux/mockups/gen_offer_table.py follow the same paths.
Verified: libslic3r, libslic3r_gui and libslic3r_tests all build, CAD suite green
at 2518 assertions in 194 test cases, and the sibling fork builds identically —
17 shared sources still byte-identical, 8 diverging by their expected counts.
Mirror of snaporca 2b3e890165 (DesignPanel.cpp applied as a patch; parity 30 / 16, shared
files byte-identical).
All 86 verbs now carry a hint: 55 extracted from the C++ tool definitions so the offer and
the armed-tool hint cannot drift, 31 written by hand. One wxEVT_MENU_HIGHLIGHT binding
shows the hovered verb's hint in the status line. The generator asserts that no wired verb
lacks one.
Also: all 200 status writes go through set_status(), which wraps instead of clipping at the
panel edge; and the empty-document hint is called from on_tab_shown() as well, since
after_tree_edit() never runs on a freshly opened tab.
Verified on the rig.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LyRwbuq6fjn3VV9U9UvhBM
Mirror of snaporca bcab67f8ce (DesignPanel.cpp applied as a patch; parity 30 / 16, shared
files byte-identical).
tool_atlas.json now names an icon for 80 of 86 verbs, derived from the toolbar's own
definitions rather than invented, and every one of the 54 distinct names was checked to
exist in resources/images first.
The first attempt drew nothing despite a green build: wxGTK builds the GtkMenuItem inside
Append() and reads GetBitmap() there, so setting the bitmap on the returned item is a
silent no-op. append_offer_item() constructs, sets, then appends — the same order Orca's
own append_menu_item() uses.
Verified on the rig.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LyRwbuq6fjn3VV9U9UvhBM
Mirror of snaporca 0c83f59f13 (DesignPanel.cpp applied as a patch, not copied, so this
fork's 30 permitted divergent lines survive; parity re-checked at 30/16 with the shared
files byte-identical).
The sketch dropdown never registered "fly:<family>#<row>" addresses the way
feat_dropdown does for model verbs, so the offer could name a family but only ever arm
its first tool — Rectangle always gave a corner rectangle. Adds the registration, 14
atlas verbs (including the entire array family, which was absent, and rotate/scale), an
action for the sk_move row that previously did nothing when picked, and a second submenu
level so variants nest under their family instead of flattening 19 create tools.
Built green and verified on the rig: Oblique rectangle arms oblique, not corner.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LyRwbuq6fjn3VV9U9UvhBM
Row order is RATIFIED (charter 4.1, 2026-07-31) and this is the first working
implementation of it: right-click in the Design viewport and a vertical list
opens at the pointer with the eight families in their fixed order, the verbs
that apply live, and the ones that do not disabled IN PLACE carrying their
reason.
THE MAP EXISTS ONCE. DesignOffer.hpp is GENERATED from docs/ux/tool_atlas.json
by docs/ux/mockups/gen_offer_table.py — the same file the 113 mockups are drawn
from. A drawing and the product therefore cannot drift apart, which is the only
way row constancy survives contact with a codebase. Never hand-edit the header.
NOTHING IS RE-IMPLEMENTED. Each row routes to the code that already runs that
verb: "key:S+E" through m_keys_feature, "key:L" through m_keys_sketch,
"fly:material#4" through the feature flyout's own action, "btn:colour" through
the standalone button. The offer is a second door onto the same room, so the
toolbar, the shortcuts and the menu cannot drift into three behaviours. The 8
verbs with kernel support but no GUI path show disabled, which is honest and
matches section 10 of the charter.
Right-click only fires the offer when the canvas is IDLE. Right-click already
ends a polyline chain and finishes the move gizmo; taking those over would
break two working interactions to add a third.
Two things the running build corrected, both found by looking at screenshots:
- THE REASON MUST BE TRUE FOR WHAT IS IN FRONT OF THE USER. Taking the first
refusal in a family printed "Transform needs a body — add or import one
first" on a document that HAS a body, because the real obstacle was that
nothing was selected. Now the reason comes from a verb that accepts the
current selection and fails only on document state; if no verb in the family
accepts this selection at all, it says "select something first" or says
nothing. A menu whose whole value is telling the truth cannot ship a lie.
- Classification follows the level the pick cycle has REACHED, not the face the
ray happened to hit, so the header cannot name a face while the whole body is
lit. Sketching on the face you merely clicked is untouched — that path is
sketch_plane_from_selection (snaporca-3a2).
Verified on :11 end to end: nothing selected shows Sketch live with Shift+S and
seven greyed rows each explaining itself; a selected solid shows Move directly
with Shift+Y (one applicable verb, so no submenu and no extra click) and five
families as submenus. Both forks compile and link.
Fork parity re-checked after the port: DesignPanel.cpp 30 divergent lines,
DesignCanvas.cpp 16, every other CAD file byte-identical — the invariant exactly.
snaporca-96r.
Tommaso was not sure about the ring and proposed a vertical list: left-click
selects, right-click exposes icon / name / shortcut. Drawn, it is better, and
the reasons are visible in the renders rather than arguable.
THE DISABLED ROW CAN SPEAK. This is the one that decides it. A ring slot that
does not apply is an empty circle: it says nothing, and on a fresh document six
of the eight are empty. A list row that does not apply is greyed IN PLACE with
its own name and its own reason — "Create a sketch, or pick a solid face,
first", "Create a solid body to pattern first" — which are strings the product
already ships and which tool_atlas.json already carries. The first-run picture
stops being a mostly-empty ring and becomes a map of what the product does and
what you must do first. For the audience section 2 puts first, that is the
whole ballgame.
THE OVERFLOW DISAPPEARS. Sketch Create needs nine addresses; a ring of eight
pushed Polygon and Point behind a "More" slot. Nine rows is just nine rows. The
one measured defect in the ring design is not a defect in this one.
SHORTCUTS READ AS A COLUMN. Right-aligned in a list they stack into something
the eye learns passively, which is exactly the graduation path 4.1 claims —
and it is the mechanism by which the power user Tommaso describes stops opening
the menu at all. Around a ring the same keys are eight loose chips.
Also, unglamorously: long translated names fit, arrow keys and screen readers
work natively where a radial needs special handling, and a 324px box costs the
1366x768 machine far less than a 380px disc over the model.
What the ring keeps: equidistant targets and a future flick gesture. Since the
brief is that power users live on the keyboard, that buys less than it looks.
The invariant is untouched — same eight families, same fixed order, nothing
re-sorted, nothing compacted. Only the geometry changed, which is the point:
the map survived a change of form factor, so it was a real map.
Both forms are now rendered side by side for the same states, and the atlas
opens with the pairs.
snaporca-96r.
The charter fixed the slot-constancy invariant but carried one hand-written
eight-cell table as an illustration, and nothing of the offer exists in the
product. Before any GUI code, the group needs the map itself: what verbs there
are, what each one needs before it can be offered, and what the ring actually
looks like in every situation a user can put it in.
tool_atlas.json is the source of truth and it was extracted from the code, not
from memory: verbs, shortcuts and the exact refusal strings from DesignPanel's
six feat_dropdown call sites and the sk_key table, kernel coverage checked
against CadFeatureType, headless coverage against McpControl's dispatch, and
the selection kinds taken from the callbacks DesignCanvas actually exposes. It
carries each verb's preconditions, so "why is that slot empty" has an answer
already written in the product's own words. The eventual C++ table generates
from this file too — the map exists once.
gen_offer_mockups.py renders it: 20 selection kinds x 2 document states = 40
primary rings, plus 73 sub-rings, plus comparison sheets. 113 states, none of
them hand-drawn, because a human drawing 113 rings is exactly how an address
quietly changes. Everything is framed at 1366x768, the charter's own reach
target, so L11 is tested in the mockups before it is tested in code.
Three things the drawing found that the prose had not:
- MEAN FILL IS 3.45 OF 8. The empty-slot rule is cheap in argument and
expensive on screen; on a fresh document exactly two slots are live. That
picture is the anti-clutter thesis made literal and it is the strongest image
in the set.
- SUB-RINGS MUST ANCHOR ON THEIR PARENT. Fanning them from north put Extrude at
N, which is Create's address in the primary map, so the second level
contradicted the first. Anchored, an address is two consistent strokes: Add
material is NE and its first verb is NE again.
- ONE FAMILY OVERFLOWS, AND ONLY ONE. Sketch-mode Create needs nine addresses
on an eight-slot ring. That is the ninth-position pressure the charter
predicted, arriving on schedule and measured rather than argued: either Point
moves family, or the tail goes to a third level, or the ring is not eight.
Every model-mode family fits. The generator refuses to wrap a tenth verb onto
a first — silent collision is the one outcome worse than an ugly ring — and
reports the overflow instead.
Also fixed while looking at renders: the selection pill sat on top of the north
slot's shortcut chip and hid it, and the scrim at 0.55 swallowed the very face
the ring had been opened on, which is 4.1 failing inside its own mockup.
Fork-neutral: the generator and everything it emits name no product, so both
forks carry byte-identical copies.
snaporca-2is.
