Aixle

Appearance

Frog on a lily pad, again: the agent's report ​

Round 3: the round-2 brief re-run by a fresh agent against the tool as fixed after round 2, with three additions to the brief: use at least one of the new features, ship a GLB, and verify it by a viewer screenshot and a re-import round trip. The agent was forbidden to read the earlier attempt. Seven model renders plus six four-line probe renders.

What follows is the agent's report, verbatim. The "fixed" notes in the index (dogfood/README.md) say what changed in the tool because of it. frog_reimport_probe.aix is the round-trip program.

1. The model ​

A cartoon frog, 4.4 × 2.77 × 4.4 units (frog itself 2.7 tall, pad 0.2), facing +z on a notched lily pad. Torso and head are two ellipsoids in union(k=0.35) with displace(0.02, 0.18) and a low-contrast speckle skin; a cream belly, a smile, and two nostrils are decals; the eyes are two white spheres with decal pupils; four legs are three-point tubes (hip–knee–ankle, so the back legs fold) with a def foot() of a flat ellipsoid, three toe capsules and an extruded web polygon; the pad is a cylinder minus a rotated wedge, with nine radial veins as a ring of thin boxes applied as a decal. It matches the plan closely; the only compromise was thickening the pad from 0.12 to 0.2 (see friction). New features used: decal (pupils, mouth, belly, nostrils, veins), curve (mouth path, into tube, into decal), set light_azimuth/set light_elevation, and --focus eyes --beauty. All did what the docs said: decals added no geometry (the re-import confirms it), the curve tube was smooth, the focus render re-extracted at cell 0.018 and framed the eyes.

2. Friction log ​

  • Thin flat base silently clipped from the mesh (the big one). Lines 17/19 (original: pad_disc = cylinder(2.2, 0.12, round=0.04) | move(0, 0.06, 0)), line 61 model = frog + pad. Expected: a 4.4-wide disc, as check printed (model 4.4 × 2.69 × 4.4) and as step 3 in steps.png showed. What happened: sheet title said 3.61 × 2.69 × 3.28, the top view showed the pad clipped to a rounded square, report Surface extent x -1.807..1.806, no warning. Reproduced with 4 lines (cylinder(2.2, 0.12) + sphere(1)|move(0,1,0) → extent ±1.366); --grid 200 does not help (±1.33); a 0.3-thick disc is complete; adding a sphere(0.1) at the rim "rescues" the whole disc. So a coarse pre-pass that sizes the extraction domain misses a plate thinner than its spacing lying at y = 0 and independent of --grid. Workaround: pad 0.2 thick. Nothing in the docs mentions this; it is a wrong-size that check cannot see and the sheet reports without comment.
  • Feet floated above the pad and the "separate pieces" warning was misattributed. Line 61 originally (frog | move(0, 0.12, 0)) + pad: feet bottom at 0.10–0.13 plus 0.12 landed at 0.22 over a pad top of 0.12. The warning said the loose piece (vol 0.546 at (-0.2, 0.078, 0.41)) was in 'mass', 'skinned'; it was the pad (or the whole frog vs pad), and neither pad nor legs was named. Workaround: frog | ground() | move(0, pad_thick - 0.01, 0). The probe version attributed correctly ('pad_disc', 'model'), so the attribution is unreliable with blended/decalled steps whose loose bounds contain the point.
  • Spurious thin-part warning on a decal region. Line 20 vein = box(2.0, 0.3, 0.035): check warned "'vein' is only 0.035 units thin ... may be missing or broken in the mesh ... set grid 315". vein is only ever a decal region (line 22) and adds no geometry; the warning is wrong and its advice (grid 315) would cost 15× the samples. I widened it to 0.05 to silence it.
  • Watertight warning names decal regions and blend inputs. Report: "37 edges ... in 'belly_region', 'torso', 'head'". belly_region is not geometry (line 29–30), and torso/head were blended into mass. The location is right (the belly/skin boundary), the step names are not helpful.
  • Slices draw a decal region as a filled interior solid. slices.png X=0 shows the belly (line 29 region) as a tan blob inside the green body, as if the belly were a separate solid inside the frog. The doc says "filled means solid"; here the fill colour distinguishes materials by whichever region contains the point, so a decal reads as an internal part. I could interpret it only because I wrote it.
  • Z=0 slice shows two unexplained black dots above the head: the eye spheres (centre z 0.45, r 0.45) graze z = 0 and appear as black dots rather than white; nothing on the sheet says they are tangencies.
  • Speckle at 0.07 units looked like camouflage blocks on the sheet. Line 9 originally scale=0.07 with cell 0.034: the sheet colours per vertex, so a pattern near the cell size becomes a blotchy mosaic; the beauty render was fine. The docs warn about patterns larger than the part, not smaller than a cell. Fixed with scale 0.11 and closer colours.
  • Pad veins are jagged on the sheet, clean in beauty (lines 20–22): same per-vertex colouring; a 0.05-wide decal is a stair-step on the sheet. Not documented that sheet materials are vertex-sampled.
  • Focus render clips decal regions into dark blobs. --focus eyes sheet shows the mouth decal as dark patches at the clip frame's bottom corners (the clip box cuts through the region). Interpretable, but it looks like a defect until you know the frame is a clip.
  • Beauty framing is loose. At --beauty-size 768 the model fills ~45% of the frame; no documented way to zoom the beauty camera (only azimuth/elevation), short of --focus.
  • check does not list def bodies (toe, foot, lines 46–50), so a def's sizes are only visible once called and named; foot() had to be judged in back_foot's thumbnail.
  • Import is slow: import("../out/frog/model.glb") took 18.3 s to sample 115 k triangles at cell 0.046 on both check and render (it is not cached between the two), longer than meshing the original from scratch (2.2 s).
  • ground() after union(k=) leaves the bounds at y −0.01 (line 61: model ... y -0.01..2.76) while the surface is at 0: expected from the docs, but it means every check line for the model looks a hair below the floor.
  • Viewer could not be verified (see 3); the page shows no message when three.js fails to load, just a blank canvas under a correct header.

3. Export verification ​

  • Viewer screenshot: frog_viewer.png shows the page header rendered correctly ("model 4.40 × 2.77 × 4.40 units · 115400 triangles", Perspective/Front/Right/Top/Wireframe/Grid buttons) over a blank grey canvas. shot.mjs reported webgl: true, errors: [] after 8 s; one retry with a 30 s wait and full console capture gave net::ERR_CONNECTION_RESET and TypeError: Failed to fetch dynamically imported module: https://cdn.jsdelivr.net/npm/three@0.160.0/build/three.module.js. curl reaches that URL (HTTP 200 through the proxy), so it is the headless browser's path to the CDN in this sandbox, not the page. Not pursued further, per the brief.
  • GLB round trip: model.glb is 3.69 MB with a 1024 px atlas (model.png, 610 charts). Original: 4.4 × 2.765 × 4.4, 115,400 triangles, volume 9.139. Re-import (frog_reimport.aix): 4.401 × 2.761 × 4.401, 111,368 triangles, volume 9.127, surface extent identical (±2.2, y −0.001..2.76), one piece, no warnings. The re-import sheet is the same frog in untextured clay: pad, notch, eyes, folded legs, toes and web all present; skin slightly smoother (sampled at cell 0.046); all decals gone (belly, mouth, nostrils, pupils, veins), as expected for a geometry-only import.

4. What worked well ​

check in ~1 s with every step's box; the sheet/slices/steps trio (the steps sheet let me see pad intact and thereby pin the clipping to the final extraction); decal for every facial feature with no geometry cost; curve + tube as a decal region for the mouth; tube through three points for folded legs; mirror for symmetry; def for the foot; union(k=) + displace for a soft body; --focus ... --beauty and --out for a close-up without clobbering the main render; the beauty render's shading and shadows made the character presentable in three iterations; --no-export --no-viewer for fast diagnostic renders.

5. Three changes that would have helped most ​

  1. Fix or warn on the domain-sizing pre-pass: the extraction domain must never be smaller than the output's bounds when the field is negative there; at minimum, warn when "Surface extent" is smaller than a primitive's exact bounds ("'pad_disc' reaches x ±2.2 but the mesh stops at ±1.81"). check said 4.4 and the sheet said 3.61 with no comment.
  2. Treat decal regions as regions: no thin-part warning, no watertight/loose-piece attribution to them, and draw them in slices as a surface-only colour rather than a filled interior solid (or mark region-only steps on steps.png).
  3. Document sheet-versus-beauty material sampling and a beauty zoom: say that the sheet colours per vertex so patterns and decals near a cell look blocky while the beauty render and atlas are exact; and give the beauty camera a framing/zoom setting (set zoom or a fill percentage) so the presentation image is not half empty.