DECISION: Refine R0.3 into a separate R0.4 exterior using the user's physical-hand photographs, preserving all five four-axis finger chains, including the little finger removed from the photographed sample.
WHY: On 2026-11-17 the user supplied five close-ups and asked for a comparison-based improvement, explicitly explaining the missing little finger. The photographs reveal curved shells, fork openings, stepped joint caps, black contact pads and a spatial thumb root more clearly than the earlier website screenshot. design-docs/lh20-r04-photo-refinement.md records the interpretation, assumptions, rejected alternatives and checks. No scale marker or internal section was supplied, so the existing dimension baseline remains an original design input and the internal drive remains pending.
ALTERNATIVE CONSIDERED AND REJECTED: Merely changing colours, duplicating the missing-finger state, or presenting a concept image would not improve the actual complete CAD model. Reusing earlier verification reports would conceal new geometry and motion failures.
WHAT CHANGES: Add curved shells and removable covers, nested fork ears, stepped three-fastener caps, wraparound black pads, segmented flex-circuit placeholders and a canted thumb mounting frame. Preserve native articulated source, binding geometry checks and previews to captured packages. A small export adapter uses the native writer's existing metadata_properties=False option and saves complete descriptive records in a hashed sidecar because >22 kB constraint strings overflow the OCCT STEP reader; geometry and assembly structure remain intact, and the shared SDK is unchanged.
Checked: Single finger and complete hand passed positive-volume, BREP validity, constraint and exact interference checks in straight and specified light-flex poses. Diagnosed and removed a 0.256 mm3 knuckle overlap without changing the 0.01 mm3 threshold. All five fingers retain four axes. STEP reimport retained 245 solids with relative volume error 4.03e-7. URDF has 21 links, 20 joints and 84 mesh files. Reviewed actual front/back/oblique/side CAD views and the R0.3 comparison. Packaging verified 102 files, relative mesh paths, source hashes and independent-finger versus nested-finger volumes. Python compilation passed. Full travel, actuator integration, continuous flex routing, manufacturing and hardware performance are not qualified.
Co-Authored-By: Codex <noreply@openai.com>
DECISION: Build a separate R0.3 exterior and articulated layout with slender fingers, rounded crowns, exposed knuckle openings and a waisted palm. Treat the older internal drive design as pending refit rather than claiming it fits the smaller envelope.
WHY: In the 2026-11-16 hand-design discussion, the user twice rejected the previous model as visually crude and supplied a WUJI product screenshot. More fasteners and edge treatments did not address the thick segmented silhouette. The second critique explicitly prioritized external form. design-docs/lh20-r03-exterior-first.md records this change in sequence, the geometry iterations and the qualification boundary. The new widths are original design inputs, not dimensions measured from the photograph.
ALTERNATIVE CONSIDERED AND REJECTED: Continuing cosmetic edits on the previous bulky layout, presenting generated images as CAD evidence, or relabelling official geometry would fail the user's request for a better actual model. The smaller exterior cannot honestly inherit the previous drive integration claims.
WHAT CHANGES: Add independent parametric finger and palm modules, a solved 20-axis assembly, exact neutral-pose BREP checks, real CAD previews, checked STEP export and a visual-only URDF. Package the exchange files with relative meshes and hashes in a new Downloads folder while preserving older work. Document the pending internal refit, full-motion assessment and manufacturing engineering.
Checked: Single finger and whole-hand geometry, positive volumes, assembly constraints and neutral exact interference checks passed. STEP reimport retained 146 solids with relative volume error 2.45e-7. URDF has 21 links, 20 joints and 89 referenced meshes. Reviewed front, back and oblique CAD views; packaging verified 99 delivered files and ZIP contents. Python compilation and staged whitespace checks passed. No claim of full-motion, actuator, manufacturing or hardware validation.
Co-Authored-By: Codex <noreply@openai.com>
DECISION: Build an original, parameterized 20-axis hand using three sizes of two-stage planetary joints, with no tendon or pull-cable transmission. Keep unselected motors and electronics explicitly represented as reserved spaces.
WHY: In the 2026-11-16 discussion the owner chose WUJI Hand 2 as the main reference, authorized exterior and internal design, clarified that electrical wiring is allowed, and prioritized precise manipulation and compliant grasping. Public WUJI assets do not disclose a manufacturable internal drivetrain. The resulting design and evidence boundary are recorded in design-docs/lh20-tendonless-prototype.md and examples/lh20_hand/DESIGN.md.
ALTERNATIVE CONSIDERED AND REJECTED: Copying simplified public reference meshes would not establish a working internal layout. Tendon drives contradict the owner's requirement. The initial MR84 support spacing gave an inadequate side-moment screening margin, so the candidate support was widened to the MR104 envelope; catalog rating is still not a module load qualification.
WHAT CHANGES: Add separately named CAD parts, real involute gears, keyed torque interfaces, bearing envelopes, five finger assemblies, palm/frame/cover/pad geometry, tactile reservations, independent kinematics, pinned build inputs and reviewed local collision reliefs. Use unique bearing assembly identities so different joints can be driven at different angles. Generated output and unrelated concurrent work remain outside this commit.
Checked: Final native build contains 1011 solid occurrences and 176 part definitions. Independent BREP validity, canonical constraint solving and external interference checks passed for neutral, four opposition poses and one intermediate state. A/B/C internal checks and an A module at 17.3 degrees passed. Full travel, minimum wall thickness, strength, selected motors, retention details and real performance are not qualified.
Co-Authored-By: Codex <noreply@openai.com>
DECISION: write the decision down in the repository, with the discussion it came from, so that a person
who was not in the conversation can take over and trace it.
WHY: the multi-body Part was a design choice made for the editor's convenience, withdrawn on 2026-11-10 after
the owner and the assistant talked it through (the owner's reasons: the editor needs several solids side by
side with no assembly constraint, e.g. pattern a tool and cut; but making that a property of the Part makes
assembly hard and was a wrong decision; the assistant's analysis: those solids are notebook intermediates,
not product items, so Part = one solid and the editor shows loose solids). The owner also asked that every
decision be its own commit with its reasons. Commit messages carry the short form; this file carries the
whole chain, the audit table of what was broken in 2.2.8, the rules after the decision, and the list of
commits.
- design-docs/part-holds-one-solid.md: new; supersedes the older document. (design-docs/ is git-ignored, but
the older design document is tracked, so this one is added the same way, with -f.)
- design-docs/multi-solid-design.md: a note at the top says it is superseded and where to read instead; the
body is left as it was (the record of what was tried).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
On top of the merge: the bindings it touched (cut_rsolid on_split, make_part_rpart body/bodies) take the typed signature
vocabulary; a kind = "part" notebook's implicit Part is made of all its open bodies (named by their variables, the pieces of a
split cut <name>_1..n) and is recorded in the definition (one result node, so its content hash is stable across cache hits);
product values are walked through their own children, not as plain dataclasses; the skill says how bodies and the implicit Part
work; the editor marks every body of the implicit Part as the product.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A test restores a cell's lineage record in a fresh process and resolves a tag and a stable reference on the
unpickled value through the restored deltas (tags and references survive the cell cache).
- References report the body (output slot / solid) of each target; element_ref.resolve reads the slot from the
shape, so each body of a split result resolves to its own piece. A cut that leaves several bodies is not a
reference error.
- kNamingVersion (scad-naming/1) is stamped into stable anchors; a reference made under another version still
resolves and says version_changed.
- graph.hpp: from_value restores the deltas (stale comment fixed, SDK_DEFECTS 22).
- Design: summary, versioning and plan updated to what is done and what is next.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A tag assignment is a graph node (apply_tag_rselection, every op-kwarg tag); the tags an entity of a node's
output carries from upstream are now computed from that node's graph, not copied per operation:
- scad::reactive tag view (reactive/tag_view.cpp): per node, from its inputs' views and its TopoDelta,
witness by witness, as each binding's lineage policy allows (continuation_fragment: every piece of a split
face); identity ops pass it on; memoised per immutable node, O(delta entries) per node. Installed as the
topo layer's history-witness provider, so has_tag / list_tags / list_tag_bindings / explain_tag and every
ql.tag see it with no API change.
- Scopes: effective (the default) = local + inherited + lineage.
- One path per binding: a recorded assignment's continuation is the view's only; the per-op copy in
autotag.cpp now carries only what shapes alone know (imported provenance, sketch topology names, tags
outside a session, replayed values) and an op's naming of what it generates (fillet patch, extrude side).
- merge_topo_deltas composes the steps of an N-ary boolean, so names and references cross the chain.
- Effect (audit, every family): plain face tags survive booleans, blends, shell, copies, patterns,
transform/scale/offset/draft; op-kwarg tags survive blends and copies; primitive role tags follow their
face. Changed expectations: nine Python-port goldens gain exactly those carried tags (only additions,
reviewed); test_tagging_refactor / test_semantic_tagging / test_tagging.cpp / test_finalize.cpp assert
effective includes lineage.
- Skill: ql-playbook ground rules 3 and 5 and the boundaries, ql-recipes, geometric-validation, ql.tag
docstring and API page; design 3.1, 3.2 (target API), 6; SDK_DEFECTS 21 fixed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A cut that splits the body is no longer refused: cut_rsolid(on_split="pieces")
(the default) returns every piece as list[Solid] ordered by centroid, recorded as
make_cut_rsolid_list with piece_order=centroid and replayed in that order;
on_split="error" keeps the strict contract. One-piece cuts are unchanged.
- Pieces become bodies: make_part_rpart(body=pieces) -> body_1..n,
bodies={"bar": pieces} -> bar_1..n.
- Overlapping bodies are allowed and warned: inspect_body_overlaps_rlist and a
BodyOverlapWarning when the definition is built.
- Multi-body definitions use profile simplecad-part-definition-3 (new schema);
profile 2 is back to its dev bytes, so profile-2 readers refuse them.
- part_from_bodies: the implicit-Part rule (all open bodies, named by variable).
- Skill pages, API docs, stubs, design doc section 9, SDK defect 24.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
query_history emitted preserved only for an input with neither Modified nor Generated images, so a shell's
kept face (unchanged, and the seed of its offset face) had no successor and shell.body_face was never proven
(owner decision on the kept-face gap: option a). Changed expectations, each reviewed:
- tracking goldens shell / extrude_rect / extrude_circle / loft / loft_ruled: the kept shell faces and rim
edges, and the profile edges that bound the start cap of a prism or loft, gain a preserved witness (they
are IsSame in the result); shell gains five shell.body_face roles.
- test_tracking_behaviour: loft toward a point 4 -> 8 entries (the base edges stay as the cap boundary);
shell.body_face 0 -> 5 (the old comment 'every kept face was re-created' was wrong).
- Python-port goldens ops_features_shell / tag_projection: face.shell.body_face on the kept faces;
blend_failures.shell_role no longer fails; test_operation_output_roles now asserts body_faces_tag tags the
five kept faces instead of raising.
- shell_rsolid docstring and API page name the four roles; design 7.2; SDK_DEFECTS 23 fixed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- modification_delta (semantic/tracking_modify.cpp) turns the retained OCCT history of the builders that
finish through finish_solid into face and edge witnesses; transform_rsolid (rigid and affine: Copy +
GTransform merged), scale_rsolid (uniform and per axis), offset_rsolid and draft_rsolid record it, so
references follow faces through them (an offset or drafted face, generated in OCCT terms, continues its
source face). The helical sweep was already covered by its tracked sweep node.
- Stable anchors: (feature, op, occurrence-among-ancestors) + params digest tie-break + kernel role;
trace_to_birth anchors a picked entity where it was born. An operation inserted earlier in the cell no
longer shifts the anchor; two indistinguishable candidates are lost(anchor_ambiguous), never a guess.
- Splits are an error with candidates (element_ref.require / ElementRefError / describe), lost references
offer guard-ranked look-alikes.
- Docstrings and API pages of the four ops say they now carry history; SDK_DEFECTS 24 fixed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>