[ManuelStaggl]

A-findings of the full audit (2026-11-12) on the rules for the assistant round the print check, on top of #353, plus the report of 2026-11-12 that the assistant does not optimise print time and material on its own, and the points of the independent review of this PR. origin/main is merged in (undo pattern of #348, block refusals of #355).

R1 – the check reaches the assistant in the turn that changed the design

Before: "Read them after your last change, before you call the part done" was a request. The check reached the model with the next user message; a turn could end on status block it never saw.

After: the work log (workLog.ts) reads SolidCraft's history before and after every changing call. When the design changed, it hands the tool's answer to withCheckOfChange (printCheckNote.ts), which waits until the mirror has delivered that revision and the renderer has reported its check, and puts the note behind the answer. It waits outside the work log's queue.

Why this and not a step at the end of the turn: Claude Code (gate proxy), ChatGPT and local models (loop.ts) all call tools through the one chain track(upstream) in index.ts. One hook there covers all three; a turn-end step would have needed three implementations and a guard against loops. A tool answer cannot loop: no change, no note.

Details, after the review:

• Two stages. With two bodies or a joint the renderer sends the mesh check at once (measuring: true) and the full check when SolidCraft has measured clearance and joints. The answer waits up to 1.5 s for the full check, then takes the mesh check marked as preliminary; after 3 s without any check it says so. Timings at SolidCraft 0.40.1 without a window (median of 5): mirror and mesh check 211 ms for a box with its lid, 141 ms for a hinge; the hinge's two joint sweeps 498 ms; inspect.clearance 1 ms for both; the renderer adds 250 ms of quiet time. Not measured: the renderer's own check time in the app.
• What counts as a change: the undo depth moved, or a change replaced what Redo held (designChanged). The revision alone does not: trials, saves and a block that was taken back move it. At the cap of SolidCraft's undo history (200; a step there leaves the depth and moves the revision) a moved revision counts, at the price of a check too many after a trial. Measured: a probe batch that is rolled back leaves revision, undo and redo as they were; body.print_face moves revision and undo depth, so it gets its check.
• Told once. CheckTold gives the full note the first time in a turn and whenever status, any finding (all of them, with faces and numbers, not only the lines shown), the preliminary mark or a failed measure change; else one line plus the measures that changed. A new user message starts over.
• Changing tools now include partbuddy_place_drawing and partbuddy_move_drawing (check, queue, lock while a question is open, a line in the work log). The other PartBuddy tools (view, standard, ask, plan, assembly, design_rules) change no geometry.
• When the history cannot be read, the answer carries no check and the log says so.

Texts to the assistant:

• VIEW_GUIDANCE, replacing the request: PartBuddy checks the part for printing after every change and shows the result to the user as a light; the findings come behind the answer of each tool call that changed the design, and partbuddy_view gives the latest ones.
• Pending behind a tool answer: [PartBuddy: the ready-to-print check of the latest change is still running; call partbuddy_view for it before you call the part done.]
• Pending from partbuddy_view itself: [PartBuddy: the ready-to-print check of the latest change is still running; it will be here in a moment.]
• Preliminary: Preliminary: clearance and joints still being measured (bodies that overlap or come closer than a push fit, joints that stop before a limit); call partbuddy_view for the full check before you call the part done. A clean preliminary check reads nothing found so far (build volume, wall and floor thickness, overhang angle, area on the bed, open bodies, failed features, bodies that touch).
• Preliminary from partbuddy_view itself: Preliminary: clearance and joints still being measured (…); the full check will be here in a moment.
• A measure that failed: NOT measured, because SolidCraft's measure failed: bodies that overlap or come closer than a push fit. Measure it yourself (inspect.clearance) before you call the part done. The clean line then leaves that part out of what it checked.
• Unchanged while measuring: [PartBuddy: ready-to-print check after this change: unchanged, nothing found; clearance and joints still being measured: call partbuddy_view for the full check before you call the part done.]
• Unchanged: [PartBuddy: ready-to-print check after this change: unchanged, still not ready to print, the same 1 finding as before.] / … unchanged, nothing found.], followed by Measures that changed: and those lines when there are any.

R3 – a clean check says what it looked at

Before: ready to print, nothing found. / The print check in partbuddy_view still judges what you build.

After (from the finding kinds of printCheck.ts, withClearance, withMotion):

• clean: … nothing found in what it checks (build volume, wall and floor thickness, overhang angle, area on the bed, open bodies, failed features, bodies that touch, bodies that overlap or come closer than a push fit, joints that stop before a limit). and at the end Not checked: strength and layer direction under load, the fits and hole allowances you chose, details and text too fine to print, bed-edge chamfers. Judge those yourself.
• with findings the last line ends with the same Not checked: … sentence.
• standard topics: None of this has been printed by PartBuddy yet, and the print check does not test these rules: measure what you built against them.

R54 – blocks measure the plastic they need

One measure for all four: a probe cylinder is laid in as a body in a batch that is rolled back, and inspect.interference says how much of it lies in the part. The cylinder starts 0.5 mm under the point, so a face that falls away that little (a curved one) is not read as a gap, and an edge or a chamfer is.

Block	Needs	Measured
keyhole	entry wide, lip + pocket + 1.2 mm deep, behind the entry and behind the screw's rest	before anything is cut
magnet	pocket + 1.2 mm round it + 0.6 mm under it	before anything is cut
insertSeat	the maker's wall round it (at least 2 mm, as a boss) + 0.6 mm under it; and the seat cuts its whole cylinder	after the cut, each hole along the way it went in (read from what that hole took out of the body: solid.hole names no direction); taken back when short
woodScrew	the screw's diameter of plastic round its clearance hole (4 mm screw: 12.7 mm wide, so a 15 mm tab takes it) + 0.6 mm of that hole under the countersink	the same, without the volume check (the plate's thickness is not given)
any cutting block	the body stays one body	after the cut; taken back

0.6 mm under a seat and under a countersink are our own choice (three layers), marked so in the code. "Round its hole" for the wood screw is the reading of the block's own sentence ("between a hole and an edge"); measured from the countersink it refused a 15 mm tab.

Refusals, as the engine gave them:

• Not built: at [220, 30, 0] P3 has 3 mm of plastic where 6.7 mm are needed (the mean over a width of 9 mm; the far side, an edge or another cut takes the rest); a keyhole needs the part 6.7 mm thick behind its entry and its slot. Thicken the part there, move the keyhole away from edges and other cuts, or use woodScrew. (an entry over the edge names the entry: at [30, -3, 0] … 1.18 mm)
• magnet: … has 3 mm of plastic where 3.3 mm are needed (the mean over a width of 8.7 mm; …); a pocket needs 1.2 mm of plastic round it and 0.6 mm under it. Thicken the part there, move the pocket away from edges and other cuts, or take a thinner magnet.
• insertSeat, the probe: … has 6.52 mm of plastic where 7.3 mm are needed (the mean over a width of 8 mm; …); the seat of an M3 insert needs 2 mm of plastic round it and 0.6 mm under it. Thicken the part there, move the seat away from edges and other cuts, or use a boss.
• insertSeat, the volume: Not built: at [4, 20, 20] the seat would cut 83.2 mm³ of plastic where its whole cylinder is 84.2 mm³: its rim breaks out of the face, into an edge, a chamfer, the far side or another cut; the seat of an M3 insert needs …
• woodScrew: … has 2.49 mm of plastic where 2.65 mm are needed (the mean over a width of 12.7 mm; …); a countersunk hole for a 4 mm wood screw needs 4 mm of plastic round its 4.7 mm hole and a plate at least 2.65 mm thick. Thicken the plate, or move the hole away from edges and other cuts.
• split: Not built: it would leave 2 bodies where Strip was one (Strip (1)): the cut goes right through the part there. Give a point with more plastic round it, or a smaller size
• not taken back: …. What was built of it could NOT be taken back: undo it before you call again (and a line in the log)

Success answers end with PartBuddy measured that plastic before building, from 0.5 mm under the face on: it is there. (keyhole, magnet), PartBuddy measured that plastic, from 0.5 mm under the face on, and that each seat cut its whole cylinder: it is there. (insertSeat), PartBuddy measured that plastic, from 0.5 mm under the face on: it is there. (woodScrew), or PartBuddy could not measure that plastic this time: check it yourself. insertSeat adds Each seat needs 2 mm of plastic round it and 0.6 mm under it.

Taking a block back. SolidCraft 0.40.1 reads the depth of its undo history as 200 at most (211 features: 200; one more: still 200, revision and feature count move). A block that must not stay is therefore undone one step at a time until the design measures as before (same bodies, same volumes, same number of features and sketches, history no deeper), with at most the steps it sent plus three; then the top component is made active again.

At the engine (SolidCraft 0.40.1, audit harness):

• The shared volume is exact, not facetted: seven probe cylinders wholly inside a block, a rod and a ball, along the axes and askew, each gave 1.0000000 of their volume; with 0.02 mm of 2.8 mm outside 0.99286; with 0.01 mm of the width over an edge 0.99993. The threshold stays at 0.999.
• The centre of mass comes with all the digits of a double: on a block of 200 × 200 × 50 mm the axis of one M3 seat was 0.000007° off. On a 45° face it read [-0.707107, 0, -0.707107].
• Built, with the measured sentence: keyhole in 8 mm, magnet in 4 mm, insert M3 in 10 mm, wood screw in 8 mm; magnet and two insert seats on a 45° face; magnet and insert seat on a rod of 20 mm radius (the seat took 0.03 % less than its cylinder there); two wood screws on a stepped plate (8 and 4 mm); a 4 mm wood screw in a 15 mm tab.
• Refused, nothing left behind: keyhole in 3 mm and with its entry over the edge; magnet with 0.3 mm under it and 1 mm from a wall; insert seat in 3 mm, in a plate exactly as thick as the seat is deep, 2.5 mm from an edge, on a 3 mm pilot hole through the part, 4 mm from an edge with a 3 mm chamfer, and one of two over a pocket from below; wood screw 5 in a 3 mm plate and 5 mm from an edge.
• A probe batch leaves revision, undo and redo untouched.
• The harness's bloecke cases (40) pass.

Limits: what is missing must be more than a thousandth of the stock's volume to be seen (a seat 6 mm from the edge with the 3 mm chamfer is built: its wall is 1 to 2 mm over the top millimetre only). The uppermost 0.5 mm are not probed. Pockets of one call are measured against the body before the cuts, not against each other. The same conditions in mechanismTool.ts are untouched (others work there).

Material and print time (reported 2026-11-12)

Rule group economy (fifth group, on by default; settings switch "Design lean" / "Sparsam konstruieren", partbuddy_design_rules { economy }). Text for a 0.4 mm nozzle and 0.2 mm layers:

• Material and print time, from the first shape on and not as a clean-up: build a shell, not a block. Slicers print 0.9, 1.35, 1.8 mm (two to four lines) as wall and fill anything thicker with sparse infill, so stiffen with ribs as thick as the wall, at least two lines (0.9 mm), instead of thickening the wall. Keep the part low as it prints: every 0.2 mm of height is one more layer, and print time grows with the layers. Shape it so it needs no support, which costs more time and material than any other choice. Leave out whole regions that do nothing (an open frame instead of a plate, a shorter body); small pockets and holes in a filled region add walls and save little or nothing. Split a part only when that saves support or much height and a joint is acceptable. Loaded regions, sizes the user names and what the user wants solid stay as they are. Infill, speed and layer height are set in the slicer: never promise grams or minutes. The print check lists each body's solid volume, height, layers and support area: compare them before and after a change.

The report line becomes After each design step, name what you added on your own, what you did for material and print time and the sizes you snapped in one short line, e.g. "…, shell 1.35 mm with two ribs instead of a solid block".

Measures with every check (tool answers, next message, partbuddy_view), computed in the renderer from the mirrored mesh with meshMetrics and overhangOf:

Per body as it prints. Solid volume overstates what thinning a thick region saves, because the slicer fills it sparsely anyway; the shell estimate (two wall lines on every surface, 15 % infill inside: an assumption, not the slicer's settings) is nearer to the filament. What counts for sure: fewer layers, less area to support, less surface. Promise neither grams nor minutes:
- "Box": 6 cm³ solid, about 2.58 cm³ as a shell, 10 mm high (50 layers), 2200 mm² of surface, 600 mm² on the bed, 0 mm² leaning past the overhang limit.
- "Pin": 90 mm³ solid, about 90 mm³ as a shell, 10 mm high (50 layers), …
- and 3 more bodies.

The shell estimate: surface × two lines, never more than the solid, plus 15 % of the rest. A block 40 × 40 × 20 mm reads 32 cm³ solid and about 9.7 cm³ as a shell; a wall of four lines reads the same in both. Six bodies are listed, the rest counted; volumes under 1 cm³ are given in mm³. No mass: the material profiles carry no density. A body that cannot be measured is left out and named in the log; a report the main process refuses is sent again without measures. packages/geometry is unchanged; laid (print face down) and smallOverhangReach are not exported there, so bodyMeasures.ts repeats the pose and the reach (marked shortcut:).

Yardstick

Topic	Rule with number	Source 1	Source 2	Contradiction
Walls as lines	Whole multiples of the line width; 0.4 nozzle: 0.45/0.9/1.35/1.8 mm for 1–4 perimeters; strength mostly from perimeters; shells typically 0.8 mm	Prusa, https://help.prusa3d.com/en/article/modeling-with-3d-printing-in-mind_164135/ , https://help.prusa3d.com/en/article/layers-and-perimeters_1748	Protolabs Network, https://www.hubs.com/knowledge-base/selecting-optimal-shell-and-infill-parameters-fdm-3d-printing/	Minimum wall differs (Hubs 0.8 mm, UltiMaker 0.5 mm, Stratasys Direct 4 × layer). No source names a thickness above which more wall stops paying off, so the rule names none.
Ribs instead of thick walls	Ribs stiffen thin walls and large flat areas; in FDM a rib needs no thinning. Rule: as thick as the wall, at least two lines	Stratasys Direct FDM guide (p. 4, 7), https://www.stratasys.com/siteassets/sdm/content---website-storage/design-guides/dg_sdm_fdm_0725a.pdf	Protolabs Network, https://www.hubs.com/knowledge-base/dfm-tips-for-3d-printed-parts-with-thin-walls/	Yes: Hubs gives 50–75 % of the wall, the rule of injection moulding against sink marks; on a 0.9 mm wall that is one line, under the thinnest wall that prints. Not taken over. Stratasys Direct also says solid can build faster than hollow with ribs.
Height in Z	Print time follows the layer count: a 10 × 30 mm cylinder is about 100 layers lying, 300 standing (0.1 mm)	Protolabs Network, https://www.hubs.com/knowledge-base/how-does-part-orientation-affect-3d-print/	Xometry, https://www.xometry.com/resources/3d-printing/5-easy-ways-to-reduce-3d-printing-costs/ (wording not verified)	Partly: Markforged names wall length, not height, as the classic driver.
Support	Self-supporting to 45°; bridges 5 mm (Hubs) or 10 mm (UltiMaker); support adds more time than any other factor	Protolabs Network, https://www.hubs.com/knowledge-base/supports-3d-printing-technology-overview/	Markforged, https://www.markforged.com/blog/design-for-3d-printing-part-3-decreasing-print-time ; UltiMaker, https://ultimaker.com/learn/design-for-fff-3d-printing-maximize-your-success/	Bridge 5 vs 10 mm; overhang 45° vs 45–60° (Prusa).
Pockets and holes	Holes may save material but not time: a ring took 10 % longer than the full disc; solid can build faster than hollowed	Markforged, same article	Stratasys Direct, same PDF (p. 7)	Yes, by process: Protolabs gives −36 % volume for hollowing, for SLS/MJF/metal, not FDM.
Bed contact	Largest face on the bed; more contact, less warping	Markforged, https://markforged.com/learn/composites-3d-printing-design-tips	UltiMaker, as above	None; Hubs: strength in XY is 4–5 × that in Z, so load can outrank it.
Splitting	Fewer supports, fits the bed, shorter build; costs a seam, pins and fit allowance	Prusa, as above	Markforged, Stratasys Direct, as above	Clearance figures differ and are printer-specific.
Large flat faces	Prone to warping; break up with ribs, round corners	Protolabs Network, https://www.hubs.com/knowledge-base/key-design-considerations-3d-printing/	Stratasys Direct, same PDF (p. 4)	Severity depends on the machine. Pillowing of large tops: one source only (Polymaker), not in the rule.
The slicer's part	Infill 10–20 % usual, above 40 % rarely sensible; layer height is the main factor of print time (0.1 → 0.2 mm halves it)	Prusa, https://blog.prusa3d.com/everything-you-need-to-know-about-infills_43579/	Protolabs Network (infill article above); Markforged, as above	Hubs contradicts itself on infill vs shells for strength; Prusa sides with shells.

Test expectations changed

• test/printCheckNote.test.ts: ready to print, nothing found → the new clean text and Not checked: ….
• test/standardTool.test.ts: closing sentence of a standard topic.
• test/blockTool.test.ts: the keyhole sends a trial batch (solid.cylinder inspect.interference … sketch.inspect) before its cuts; MAGNET has around and under; the R54 cases run on a plate that answers probes and holes as the engine does.
• test/designRules.test.ts: rule list and defaults carry economy; the report line names material and print time; ribs.
• e2e/app.spec.ts: the list of rule switches carries economy.
• e2e/solidcraft.contract.spec.ts (wall plate): the answers must carry PartBuddy measured that plastic before building, from 0.5 mm under the face on: it is there. after the keyholes and the magnets and PartBuddy measured that plastic, from 0.5 mm under the face on: it is there. after the wood screws.
• New: test/checkOfChange.test.ts (both routes, no change, bounded wait, two stages, moved revision without change, a failing call that changed the design, undo, two calls at once, unreadable history, drawing tools, told once), test/bodyMeasures.test.ts.

Checked

pnpm run typecheck, eslint . and vitest run in apps/desktop. Audit harness against SolidCraft 0.40.1: bloecke (40 cases), the block cases on flat, sloping and curved faces, the probe shares and the timings above. Not run locally: the app, Playwright E2E and contract tests (CI runs them), a real chat.

Contract run 43261598187: eight red tests (fixed in e0f821e, tests only)

• Six tests (standard screw, sketched blocks, snap fit, SVG logo, colour logo, built text) failed with the assistant is working on the design. Cause: the tool's answer now waits for the print check of its change (withCheckOfChange), so the part shows before the assistant's turn ends, and the tests called the user's action as soon as the mesh showed. Log of the screw test: session started 16:50:13.770, design:apply-action failed 16:50:14.710. The tests now wait for the Stop button to go (turnEnded) after each first build; no fixed waits.
• Wall plate and clipped cover: nothing was refused. The page snapshot of both shows the full answer with every expected sentence and it is there.; the print check behind each block ends in a list of measures, which the chat renders as a list, so the answer is several elements and a getByText across two blocks no longer matches. The tests read the whole answer (assistant-text) instead.
• The lock: bounded and always released. withCheckOfChange returns after at most CHECK_WAIT_MS (3 s) per changing call, also when the check never comes or the design changed (test/checkOfChange.test.ts: bounded wait, check of another design). Measured locally on SolidCraft 0.40.1 with the fake assistant: the turn ended 150 to 350 ms after the part showed in seven builds, 1.7 s in the first one after a cold start. Before this PR the answer did not wait at all. No product change.
• All eight tests passed locally one by one against SolidCraft 0.40.1. Not run: the other 49 contract tests, contract.yml in CI.

🤖 Generated with Claude Code