Move the Design-tab scripts into scripts/CAD/ and name them by role

Requested by SoftFever on PR #15238: ten of these had accumulated loose in
scripts/ next to ~20 unrelated upstream ones, with names that only meant
something to whoever wrote them. They now sit in scripts/CAD/, mirroring the
src/libslic3r/CAD/ and src/slic3r/GUI/CAD/ split, and the verb in the name is
the role: build- produces a binary, start- brings something up, run- runs a
suite, check- asserts one thing against a live app.

  kernel-test.sh        -> CAD/run-kernel-tests.sh
  ladder-all.sh         -> CAD/run-all-checks.sh
  sketch-ladder.py      -> CAD/check-sketch-engine.py
  ladder-corpus.py      -> CAD/check-sketch-engine-corpus.py
  gui-ladder.py         -> CAD/check-gui-sketching.py
  offer-ladder.py       -> CAD/check-gui-context-menu.py
  mcp-sketch-smoke.py   -> CAD/check-mcp-sketch.py
  rig-build.sh          -> CAD/build-gui.sh
  docker-iter-build.sh  -> CAD/build-gui-incremental.sh
  gui-session.sh        -> CAD/start-headless-gui.sh

"Ladder" was the worst of them: it named the shape of the test (rungs of
increasing difficulty) rather than what the test proves, so nothing in the
directory listing told you which one needed a GPU and which was pure kernel.

Every reference rewritten -- the docs, the cross-calls between the scripts,
Dockerfile.deps, and the container-side /OrcaSlicer/scripts paths. The three
shell scripts resolve REPO relative to themselves and now sit one level
deeper, so that walk went from /.. to /../.. . The copies these push into a
container's /tmp were renamed to match, or the container would have kept the
old names alive.

Two runtime paths deliberately NOT renamed. /tmp/orca-rig-build.lock is a
cross-fork contract -- both forks take the same lock so two concurrent builds
serialise instead of OOMing the box, and renaming it on one side silently
removes that guard. /tmp/gui-session.log is a runtime artefact, not a script.

Added scripts/CAD/README.md: what each script proves, what it needs, and the
two constraints that have each cost a session (never build inside the GUI
container; a window manager is required or synthetic keys are ignored).

On CI, which was the other half of the request: the kernel suite is already
there and always has been. The cases are registered in
tests/libslic3r/CMakeLists.txt under if (SLIC3R_CAD), which defaults ON and no
workflow turns off, so they build into libslic3r_tests and run under ctest on
every platform via unit_tests.yml -- like any other unit test, needing no new
job. They have simply never been seen to run, because the workflows on this PR
are still awaiting maintainer approval. run-kernel-tests.sh is the local loop
over the same cases, and it is the only script here CI could run: the other
six need an OpenGL canvas and synthetic input.

Verified: scripts/CAD/run-kernel-tests.sh from its new location, all tests
passed, 2562 assertions in 190 test cases.
This commit is contained in:
Tommaso Bianchi
2026-08-28 19:34:03 +02:00
parent cdd41e230d
commit 13d5eac891
16 changed files with 118 additions and 65 deletions
+53
View File
@@ -0,0 +1,53 @@
# Design-tab scripts
Everything here supports the parametric Design tab (`src/libslic3r/CAD/`,
`src/slic3r/GUI/CAD/`). Nothing here is needed to build or run OrcaSlicer — these
are the development and verification tools for that one feature.
The verb in the name is the role:
| | |
|---|---|
| `build-…` | produce a binary |
| `start-…` | bring something up and leave it running |
| `run-…` | run a suite and report pass/fail |
| `check-…` | one specific assertion, usually driving a live app |
## Verification
| Script | What it proves | Needs |
|---|---|---|
| `run-kernel-tests.sh` | The CAD kernel builds and the Catch2 `[CadDocument]` tags pass — every case builds a document, recomputes it and asserts on real geometry. **Exit 0 is the verification contract.** | Docker only. No display. |
| `run-all-checks.sh` | Every check below, in one command. The gate before pushing a Design-tab change. | Docker + the GUI container |
| `check-sketch-engine.py` | A ladder of 2D sketches of increasing complexity, judged on loop count, closure and void attribution rather than on area. | Kernel only |
| `check-sketch-engine-corpus.py` | The same ladder graded against a systematic sample of real drawings instead of shapes we chose. | Kernel + corpus |
| `check-gui-sketching.py` | The same profiles drawn the way a person draws them — synthetic mouse gestures and typed values. | Headless GUI |
| `check-gui-context-menu.py` | That right-click is the pivot of the design gesture, and adapts to what was clicked. | Headless GUI |
| `check-mcp-sketch.py` | The sketch layer driven over the MCP socket, asserting what decides whether a profile is buildable. | Headless GUI + `SNAPORCA_MCP` |
**`run-kernel-tests.sh` is the only one CI can run.** The rest need a live
application with an OpenGL canvas and synthetic input, which hosted runners do not
have. The kernel suite itself is already in CI by an ordinary route: the cases are
registered in `tests/libslic3r/CMakeLists.txt` under `if (SLIC3R_CAD)`, so they are
part of `libslic3r_tests` and run under `ctest` on every platform like any other
unit test. This script exists for the local loop, where it is a two-minute round
trip instead of a full application build.
## Build and run
| Script | Purpose |
|---|---|
| `build-gui.sh` | Build the GUI binary in a throwaway container, writing into the build-cache volume the long-lived GUI container reads. |
| `build-gui-incremental.sh` | Incremental build against the deps-baked image, for a fast edit/compile loop. |
| `start-headless-gui.sh` | Bring the app up on a headless X display (Xvfb + a window manager), ready to drive or attach to over VNC. |
Two constraints that are not obvious and have each cost a session:
- **Never build inside the GUI container.** Its baked source tree silently
reconfigures the shared build directory and this fork's targets vanish.
- **A window manager is required.** Without one, windows are never focused, and an
unfocused GTK app ignores synthetic keys — which looks exactly like a code bug.
`docs/rig_build_traps.md` documents these and three more, with symptoms and exact
recovery commands. Read it before debugging a configure or link failure one of
these scripts reports.
+81
View File
@@ -0,0 +1,81 @@
#!/usr/bin/env bash
# Incremental slicer build against the orcacad-deps base image.
#
# The deps-baked image (built from scripts/Dockerfile.deps) carries the pinned
# dependencies at /OrcaSlicer/deps/build/destdir. This script mounts the LIVE source
# tree and resources over the baked copy so code/CMake edits apply immediately, and
# persists /OrcaSlicer/build in a named volume so ninja recompiles only what changed.
#
# Result: edit -> rebuild in seconds-to-minutes instead of a full Docker rebuild.
#
# Usage (run on the build host, e.g. behemoth, from anywhere):
# scripts/CAD/build-gui-incremental.sh
# IMAGE=orcacad-deps scripts/CAD/build-gui-incremental.sh
#
# On success the binary is inside the persistent volume at
# /OrcaSlicer/build/package/bin/orca-slicer (copy it out with a follow-up
# `docker run --rm -v orcacad_buildcache:/b alpine cp ...` or via this script's tail).
# Rig build traps already paid for once each (stale project, NLopt cache, pybind11, OCCT_LIBS, SLIC3R_CAD gate): docs/rig_build_traps.md
set -euo pipefail
REPO="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
# orcacad-deps, NOT snaporca-deps: see the note in run-kernel-tests.sh — the wrong image
# fails at CMake configure, not at link time.
IMAGE="${IMAGE:-orcacad-deps}"
BUILD_VOL="${BUILD_VOL:-orcacad_buildcache}"
echo "REPO=$REPO IMAGE=$IMAGE BUILD_VOL=$BUILD_VOL"
# The root CMakeLists.txt and cmake/ must be mounted too, not taken from the baked image:
# they carry the build-time gates (e.g. SLIC3R_CAD -> add_definitions(-DSLIC3R_CAD)) that the
# mounted headers are compiled against. With a stale baked copy the gate silently stays off and
# the build fails with "class GLCanvas3D has no member named set_design_sketch_tool".
#
# build_linux.sh must be mounted for the same reason, and here the stale copy is guaranteed
# wrong rather than merely risky: orcacad-deps is layered on snaporca-deps, so the baked script
# is the OTHER fork's and builds `--target Snapmaker_Orca`. This fork's target is `OrcaSlicer`,
# so without this mount configure succeeds and then ninja dies on "unknown target".
# scripts/ likewise: build_linux.sh's packaging step sources scripts/appimage_lib_policy.sh,
# which the baked snaporca tree does not have, so a fully successful link still exited
# non-zero with "missing AppImage helper" and the binary check never ran.
# ---- OOM guard (2026-08-21) -------------------------------------------------------------
# Two of these builds ran at once on 2026-08-21, each with ninja -j$(nproc)=16: ~36 cc1plus
# holding 42 GB of a 62 GB box -> global OOM at 21:05, a 2h28m kill storm, ssh unreachable,
# lightdm destroyed. Neither build produced a single object. scripts/CAD/build-gui.sh grew the
# bounds first; every script that starts a compile needs the same three, or the guard is only
# as strong as the script you happened not to use.
# flock — the lock path is SHARED with build-gui.sh and the other fork on purpose, so
# concurrent builds serialise instead of summing.
# -j — bounded parallelism; ~1.17 GB per cc1plus was the measured average.
# --memory — the actual guarantee: a runaway build dies in its own cgroup instead of taking
# the host down. --memory-swap equal to --memory forbids swap, which is what made
# ssh hang.
JOBS="${JOBS:-12}"
MEM="${MEM:-40g}"
LOCK=/tmp/orca-rig-build.lock
exec 9>"$LOCK"
if ! flock -n 9; then
echo "another build holds $LOCK — waiting (this is the OOM guard, not a hang)"
flock 9
fi
docker run --rm \
--memory="$MEM" --memory-swap="$MEM" \
-v "$REPO/src":/OrcaSlicer/src \
-v "$REPO/resources":/OrcaSlicer/resources \
-v "$REPO/CMakeLists.txt":/OrcaSlicer/CMakeLists.txt \
-v "$REPO/cmake":/OrcaSlicer/cmake \
-v "$REPO/deps_src":/OrcaSlicer/deps_src \
-v "$REPO/build_linux.sh":/OrcaSlicer/build_linux.sh \
-v "$REPO/scripts":/OrcaSlicer/scripts \
-v "$BUILD_VOL":/OrcaSlicer/build \
"$IMAGE" \
bash -lc "cd /OrcaSlicer && ./build_linux.sh -sr -j $JOBS"
# src/CMakeLists.txt:151 renames the OrcaSlicer target's output to "orca-slicer" — not
# "snapmaker-orca", which is the other fork's binary name.
echo "=== build finished; checking for binary ==="
docker run --rm -v "$BUILD_VOL":/b "$IMAGE" \
bash -lc 'ls -lh /b/package/bin/orca-slicer 2>/dev/null && file /b/package/bin/orca-slicer || echo "NO BINARY"'
+98
View File
@@ -0,0 +1,98 @@
#!/usr/bin/env bash
# Rebuild the GUI binary the design rig launches — in a THROWAWAY container, writing into the
# same build-cache volume the rig's long-lived GUI container reads from.
#
# NEVER build inside the GUI container (snaporca-gui / orcacad-gui). Its baked /OrcaSlicer tree
# is the Jun-13 Snapmaker-derived source, so a `cmake .` in there silently reconfigures the
# shared build dir as project(Snapmaker_Orca) and this fork's targets vanish. That is Trap 1 of
# five; all of them, with symptoms and exact recovery commands, are in docs/rig_build_traps.md.
# Read that file before debugging a configure or link failure this script reports.
#
# Usage:
# scripts/CAD/build-gui.sh # configure + build the fork's GUI target
# DRY_RUN=1 scripts/CAD/build-gui.sh # print the resolved fork identity and exit, no container
set -euo pipefail
REPO="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
# Fork identity is DERIVED from the repo, never hardcoded, so this file is byte-identical in
# both forks and cannot be mirrored into the wrong one. Pointing a fork at the other fork's
# image or volume is not a slow failure: with the wrong image CMake dies at configure, and with
# the wrong volume the two forks silently trade build artefacts.
PROJECT="$(sed -n 's/^project(\([A-Za-z_0-9]*\)).*/\1/p' "$REPO/CMakeLists.txt" | head -1)"
case "$PROJECT" in
Snapmaker_Orca) PREFIX=snaporca; BIN=snapmaker-orca ;;
OrcaSlicer) PREFIX=orcacad; BIN=orca-slicer ;;
*) echo "FATAL: unrecognised project($PROJECT) in $REPO/CMakeLists.txt" >&2; exit 2 ;;
esac
TARGET="$PROJECT"
IMAGE="${PREFIX}-deps"
BUILD_VOL="${PREFIX}_buildcache"
echo "REPO=$REPO PROJECT=$PROJECT IMAGE=$IMAGE BUILD_VOL=$BUILD_VOL TARGET=$TARGET BIN=$BIN"
if [ -n "${DRY_RUN:-}" ]; then
echo "DRY_RUN: resolution only, no container started"
exit 0
fi
# Three memory bounds. On 2026-08-21 both forks ran this script at the same time, each with
# ninja -j$(nproc)=16: ~36 cc1plus holding 42 GB of a 62 GB box -> global OOM at 21:05, a
# 2h28m kill storm, ssh unreachable, lightdm destroyed, 2946 session kill events. Neither
# build produced a single object. The bounds, weakest to strongest:
# flock — the lock path is shared by both forks on purpose, so they SERIALISE instead of
# summing. Peak is one build's worth no matter who else starts one.
# -j12 — measured 1.17 GB average per cc1plus in the incident dump, so 12 in flight
# is ~14 GB typical and leaves the box usable. JOBS=n overrides.
# --memory — the actual guarantee. A runaway build hits its own cgroup limit and dies alone;
# the host never reaches global OOM again, whatever -j or flock do.
# --memory-swap equal to --memory forbids swap, which is what made ssh hang.
JOBS="${JOBS:-12}"
MEM="${MEM:-40g}"
LOCK=/tmp/orca-rig-build.lock
exec 9>"$LOCK"
if ! flock -n 9; then
echo "another fork's build-gui holds $LOCK — waiting (this is the OOM guard, not a hang)"
flock 9
fi
# Every one of these mounts covers a trap, none is decorative:
# CMakeLists.txt + cmake/ carry the SLIC3R_CAD gate — inherit the baked copies and the cache
# says SLIC3R_CAD=ON while -DSLIC3R_CAD is never defined, so every #ifdef block compiles out.
# deps_src/ carries pybind11, which the image predates.
# src/, resources/, localization/, version.inc are the code under test.
rc=0
docker run --rm \
--memory="$MEM" --memory-swap="$MEM" \
-v "$REPO/src":/OrcaSlicer/src \
-v "$REPO/resources":/OrcaSlicer/resources \
-v "$REPO/cmake":/OrcaSlicer/cmake \
-v "$REPO/deps_src":/OrcaSlicer/deps_src \
-v "$REPO/localization":/OrcaSlicer/localization \
-v "$REPO/CMakeLists.txt":/OrcaSlicer/CMakeLists.txt \
-v "$REPO/version.inc":/OrcaSlicer/version.inc \
-v "$BUILD_VOL":/OrcaSlicer/build \
"$IMAGE" bash -lc "
cd /OrcaSlicer/build || exit 1
cmake . > /tmp/cfg.log 2>&1 || { echo 'CONFIGURE FAILED'; tail -25 /tmp/cfg.log; exit 1; }
# Twice, deliberately. src/libslic3r/CMakeLists.txt publishes OCCT_LIBS as CACHE INTERNAL
# at the END of its own configure, so a first pass after that list changes links the
# PREVIOUS one and drops TKBool/TKOffset — a wall of TopOpeBRepBuild undefined references
# that reads as a broken OCCT install and is not. Trap 4.
cmake . > /tmp/cfg2.log 2>&1 || { echo 'RECONFIGURE FAILED'; tail -25 /tmp/cfg2.log; exit 1; }
ninja -f build-Release.ninja -j$JOBS $TARGET > /tmp/bld.log 2>&1
rc=\$?
echo \"EXIT=\$rc\"
grep -n 'error:' /tmp/bld.log | head -20
tail -4 /tmp/bld.log
ls -la /OrcaSlicer/build/src/Release/$BIN 2>/dev/null
exit \$rc
" || rc=$?
# A target-only build writes src/Release/, but this fork's start-headless-gui.sh may default BIN to the
# PACKAGED path that only build_linux.sh refreshes — launching with the default would then run a
# stale binary. Pass BIN explicitly. See docs/rig_build_traps.md.
echo "=== launch the rig on the binary just built ==="
echo " docker exec -e BIN=/OrcaSlicer/build/src/Release/$BIN ${PREFIX}-gui /OrcaSlicer/scripts/CAD/start-headless-gui.sh"
exit "$rc"
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
+132
View File
@@ -0,0 +1,132 @@
#!/usr/bin/env python3
"""Autonomous 2D-sketch loop: drive the Design tab's sketch layer over the MCP socket and
assert the things that decide whether a profile is buildable.
WHY THIS EXISTS. The 2D layer used to be reachable only by clicking, so every question about it
("is this loop closed?", "did the offset survive?", "is the circle a void or a second body?")
cost a GUI session and a human. The socket verbs make each one a call, and this script is the
loop: build a known profile, ask the app what it thinks it has, compare against arithmetic.
RUN IT AGAINST A RUNNING APP:
SNAPORCA_MCP=/tmp/mcp.sock <binary> # launch with the socket enabled
python3 scripts/CAD/check-mcp-sketch.py [socket] # default /tmp/mcp.sock
Exit 0 = every assertion held. Anything else prints the first mismatch and stops.
"""
import json, math, socket, sys
SOCK = sys.argv[1] if len(sys.argv) > 1 else "/tmp/mcp.sock"
_n = 0
def call(method, **params):
global _n
_n += 1
s = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
s.settimeout(30)
s.connect(SOCK)
s.sendall((json.dumps({"jsonrpc": "2.0", "id": _n, "method": method,
"params": params}) + "\n").encode())
buf = b""
while b"\n" not in buf:
d = s.recv(65536)
if not d:
break
buf += d
r = json.loads(buf.decode().strip())
if "error" in r:
raise RuntimeError(f"{method}: {r['error']}")
return r["result"]
def near(a, b, tol=1e-6):
return abs(a - b) < tol
def check(cond, what):
if not cond:
print(f"FAIL: {what}", file=sys.stderr)
sys.exit(1)
print(f" ok {what}")
def areas(rep):
return sorted(round(l["area"], 6) for l in rep["closed_loops"])
print("1. a rectangle is one closed loop of exactly its own area")
try:
call("sketch_cancel")
except Exception:
pass
call("sketch_begin", plane="XY")
call("sketch_add", rect=[0, 0, 80, 50])
r = call("sketch_describe")
check(r["buildable"], "buildable")
check(areas(r) == [4000.0], f"one loop of 4000 mm^2 (got {areas(r)})")
print("2. a circle inside it is a VOID, not a second profile")
call("sketch_add", type="circle", center=[40, 25], radius=10)
r = call("sketch_describe")
outer = [l for l in r["closed_loops"] if near(l["area"], 4000.0)][0]
check(len(outer["holes"]) == 1, "the rectangle encloses exactly one void")
hole = r["closed_loops"][outer["holes"][0]]
check(near(hole["area"], math.pi * 100), f"the void is pi*r^2 (got {hole['area']})")
print("3. offsetting the outer loop inward keeps it CLOSED and exact")
call("sketch_select", entities=[0, 1, 2, 3])
call("sketch_offset", distance=5)
r = call("sketch_describe")
check(r["open_ends"] == [], "no open ends after the offset")
check(any(near(l["area"], 70 * 40) for l in r["closed_loops"]),
f"the offset loop is 70x40 (got {areas(r)})")
print("4. a gap is REPORTED with its coordinates, then healed into a constraint")
call("sketch_cancel")
call("sketch_begin", plane="XY")
call("sketch_add", entities=[
{"type": "line", "p0": [0, 0], "p1": [60, 0]},
{"type": "line", "p0": [60, 0], "p1": [60, 40]},
{"type": "line", "p0": [60, 40], "p1": [0, 40]},
{"type": "line", "p0": [0, 40], "p1": [0.4, 0]}, # 0.4 mm short of closing
])
r = call("sketch_validate", tolerance=1.0)
check(not r["buildable"], "a 0.4 mm gap makes the profile unbuildable")
check(len(r["open_ends"]) == 2, f"both free ends are named (got {r['open_ends']})")
dof_before = r["dof"]
r = call("sketch_heal", tolerance=1.0)
check(r["welded"] == 1, f"one pair welded (got {r['welded']})")
check(r["buildable"] and r["open_ends"] == [], "healed profile is buildable")
check(areas(r) == [2400.0], f"healed loop is 60x40 (got {areas(r)})")
check(r["dof"] < dof_before,
f"the weld recorded a real constraint: DoF {dof_before} -> {r['dof']}")
print("5. construction geometry is excluded from the profile")
call("sketch_select", entities=[0])
call("sketch_construction")
r = call("sketch_describe")
check(not r["buildable"], "turning one side into a guide opens the profile again")
call("sketch_construction")
r = call("sketch_describe")
check(r["buildable"], "turning it back closes it again")
print("6. re-dimensioning one side keeps the rectangle a single closed loop")
call("sketch_cancel")
call("sketch_begin", plane="XY")
call("sketch_add", rect=[0, 0, 60, 40])
r = call("sketch_describe")
check(areas(r) == [2400.0], f"one loop of 2400 mm^2 (got {areas(r)})")
call("sketch_select", entities=[0]) # the bottom edge, y=0, from x=0 to x=60
r = call("sketch_set_value", value=40)
check(r["kind"] == "length", f"dimension kind is length (got {r['kind']})")
check(near(r["before"], 60), f"the edge measured 60 before (got {r['before']})")
r = call("sketch_describe")
check(len(r["closed_loops"]) == 1, "the rectangle is still exactly one closed loop")
check(r["open_ends"] == [], "no open ends after re-dimensioning")
# The point of the whole section: a rectangle must SURVIVE one side being re-dimensioned. We do
# not assert a specific area — only that the topology held — but print it so a topology-preserving
# yet geometry-wrong result is visible in the output.
print(f" note resulting rectangle area = {areas(r)} mm^2 (topology held; geometry is what it is)")
call("sketch_cancel")
print("\nall sketch assertions held")
+492
View File
@@ -0,0 +1,492 @@
#!/usr/bin/env python3
"""Rung 9: the ladder, graded against real drawings instead of shapes I chose.
Rungs 1-8 are hand-built. That is their weakness: I wrote both the geometry and the
assertion, so they prove the engine does what I expected on cases I picked. This rung
takes a SYSTEMATIC sample of the StudyCadCam corpus (every 20th sheet, 1..996 — no
cherry-picking) and grades the engine against each drawing's OWN vector geometry,
extracted from the PDF. Nothing here is transcribed by eye; the drawing is the input.
The method: pdftocairo renders the sheet to SVG, where the drawn geometry is exactly the
stroked (fill="none") paths and the text is filled glyph paths. Beziers are flattened, so
every entity handed to the engine is a straight line and every comparison below is EXACT
— no faceting tolerance to hide behind. The closed chains are then found twice: once by
this script, in plain Python, and once by the engine. The assertions are that the two
agree, and that the engine's own operations preserve what they promise.
CLOSED the engine finds the same closed loops this script does
AREA the engine's area for each loop equals the shoelace area, to 1e-6
VOID the engine attributes each void to the loop that actually contains it
MIRROR a real closed profile, mirrored, is still exactly one closed loop
OFFSET a real closed profile, offset, is still closed
Usage: check-sketch-engine-corpus.py [--sample N] [--corpus DIR]
"""
import argparse
import glob
import json
import math
import os
import re
import socket
import subprocess
import sys
import tempfile
import time
SOCK = os.environ.get("SNAPORCA_MCP", "/tmp/mcp.sock")
TOL = 1e-6 # exact-comparison tolerance (all inputs are lines)
WELD = 0.05 # endpoint-coincidence tolerance, in PDF units
# ── the socket ───────────────────────────────────────────────────────────────
_id = [0]
def try_call(method, **params):
"""sketch_cancel throws when nothing is open, which is not an error to us."""
try:
return call(method, **params)
except RuntimeError:
return None
def call(method, **params):
_id[0] += 1
req = json.dumps({"jsonrpc": "2.0", "id": _id[0], "method": method,
"params": params}) + "\n"
s = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
s.settimeout(30)
s.connect(SOCK)
s.sendall(req.encode())
buf = b""
while not buf.endswith(b"\n"):
chunk = s.recv(65536)
if not chunk:
break
buf += chunk
s.close()
r = json.loads(buf.decode())
if "error" in r:
raise RuntimeError(r["error"]["message"])
return r["result"]
# ── SVG → line segments ──────────────────────────────────────────────────────
NUM = r"[-+]?[0-9]*\.?[0-9]+(?:[eE][-+]?[0-9]+)?"
def bezier(p0, p1, p2, p3, n=16):
"""Flatten a cubic to n straight segments — the engine then sees only lines."""
out = []
for i in range(n):
t0, t1 = i / n, (i + 1) / n
pts = []
for t in (t0, t1):
u = 1 - t
x = (u ** 3 * p0[0] + 3 * u * u * t * p1[0]
+ 3 * u * t * t * p2[0] + t ** 3 * p3[0])
y = (u ** 3 * p0[1] + 3 * u * u * t * p1[1]
+ 3 * u * t * t * p2[1] + t ** 3 * p3[1])
pts.append((x, y))
out.append((pts[0], pts[1]))
return out
def path_segments(d):
"""Parse one SVG path's `d` into straight segments."""
toks = re.findall(r"([MLCZmlcz])|(" + NUM + ")", d)
cmds, cur, start, segs, i = [], None, None, [], 0
flat = []
for a, b in toks:
flat.append(a if a else float(b))
while i < len(flat):
t = flat[i]
if isinstance(t, str):
cmd = t
i += 1
# numbers repeat the previous command, as SVG allows
if cmd in ("M", "m"):
x, y = flat[i], flat[i + 1]; i += 2
cur = (x, y); start = cur
elif cmd in ("L", "l"):
x, y = flat[i], flat[i + 1]; i += 2
segs.append((cur, (x, y))); cur = (x, y)
elif cmd in ("C", "c"):
p1 = (flat[i], flat[i + 1]); p2 = (flat[i + 2], flat[i + 3])
p3 = (flat[i + 4], flat[i + 5]); i += 6
segs.extend(bezier(cur, p1, p2, p3)); cur = p3
elif cmd in ("Z", "z"):
if cur and start and dist(cur, start) > TOL:
segs.append((cur, start))
cur = start
else:
i += 1
return segs
def dist(a, b):
return math.hypot(a[0] - b[0], a[1] - b[1])
def drawing_segments(pdf):
"""Every stroked segment on the sheet, in PDF units (y already flipped up)."""
with tempfile.TemporaryDirectory() as td:
svg = os.path.join(td, "p.svg")
subprocess.run(["pdftocairo", "-svg", pdf, svg],
check=True, capture_output=True)
s = open(svg).read()
segs = []
# Drawn geometry is stroked with no fill; glyphs are filled with no stroke.
for m in re.finditer(r'<path([^>]*)d="([^"]+)"', s):
attrs, d = m.group(1), m.group(2)
if 'fill="none"' not in attrs or "stroke=" not in attrs:
continue
segs.extend(path_segments(d))
return [((a[0], -a[1]), (b[0], -b[1])) for a, b in segs if dist(a, b) > TOL]
# ── closed-chain finding, independent of the engine ──────────────────────────
def find_loops(segs):
"""Chain segments into closed loops. Returns a list of point rings."""
key = lambda p: (round(p[0] / WELD), round(p[1] / WELD))
adj = {}
for i, (a, b) in enumerate(segs):
adj.setdefault(key(a), []).append((i, a, b))
adj.setdefault(key(b), []).append((i, b, a))
used, loops = set(), []
for i0 in range(len(segs)):
if i0 in used:
continue
a, b = segs[i0]
ring, cur, prev = [a, b], b, i0
used.add(i0)
while True:
nxt = None
for (j, p, q) in adj.get(key(cur), []):
if j in used:
continue
nxt = (j, q)
break
if nxt is None:
break
used.add(nxt[0])
cur = nxt[1]
ring.append(cur)
if dist(cur, ring[0]) <= WELD:
# Snap the seam shut. The gap is a flattening artefact of the PDF, up to
# WELD wide, and handing the engine a ring that misses closing by 0.03 would
# be testing my extractor's sloppiness rather than the engine's chaining.
ring[-1] = ring[0]
loops.append(ring)
ring = None
break
# an open chain is simply not a loop; it is dropped
return loops
def shoelace(ring):
a = 0.0
for i in range(len(ring) - 1):
a += ring[i][0] * ring[i + 1][1] - ring[i + 1][0] * ring[i][1]
return abs(a) * 0.5
def point_in(pt, ring):
inside = False
for i in range(len(ring) - 1):
A, B = ring[i], ring[i + 1]
if (A[1] > pt[1]) != (B[1] > pt[1]) and \
pt[0] < (B[0] - A[0]) * (pt[1] - A[1]) / (B[1] - A[1]) + A[0]:
inside = not inside
return inside
def interior_point(ring):
"""A point strictly inside a simple closed ring (first == last).
The lowest vertex of a simple polygon is always convex, so stepping from it along the
bisector of its two edges goes inward; the step is a small fraction of the shorter edge so
it stays inside however sharp the corner is.
"""
q = ring[:-1] if len(ring) > 1 and ring[0] == ring[-1] else ring
if len(q) < 3:
return ring[0]
k = min(range(len(q)), key=lambda i: (q[i][1], q[i][0]))
v = q[k]
a = (q[(k - 1) % len(q)][0] - v[0], q[(k - 1) % len(q)][1] - v[1])
b = (q[(k + 1) % len(q)][0] - v[0], q[(k + 1) % len(q)][1] - v[1])
la, lb = math.hypot(*a), math.hypot(*b)
if la < 1e-12 or lb < 1e-12:
return v
a = (a[0] / la, a[1] / la)
b = (b[0] / lb, b[1] / lb)
bx, by = a[0] + b[0], a[1] + b[1]
n = math.hypot(bx, by)
if n < 1e-12:
return v
step = 1e-3 * min(la, lb)
return (v[0] + bx / n * step, v[1] + by / n * step)
# ── one drawing ──────────────────────────────────────────────────────────────
def grade(pdf, name, report):
segs = drawing_segments(pdf)
loops = find_loops(segs)
if len(loops) < 2:
report(name, "SKIP", f"no nested closed geometry found ({len(loops)} loops)")
return None
# The biggest loop is the sheet frame; the part outlines live inside it. Take the
# largest loop that is NOT the frame, plus every loop contained in it.
loops.sort(key=shoelace, reverse=True)
outer = loops[1]
voids = [r for r in loops[2:]
if shoelace(r) > 1.0 and point_in(r[0], outer)]
if shoelace(outer) < 100.0:
# Not a defect and not a near miss: on these sheets the part outline is not a closed
# stroked path at all, so the only loops extraction recovers are glyph counters and
# arrowheads. Measured on MPD12/30/31/60: the LARGEST loop on the sheet is 5 to 132 mm2.
# Say the number, so nobody has to re-measure to know which kind of skip this is.
report(name, "SKIP", f"no part outline on this sheet — largest loop is only "
f"{shoelace(outer):.1f} mm2")
return None
# Feed the drawing's own geometry to the engine, as lines only.
ents = []
rings = [outer] + voids
for ring in rings:
for i in range(len(ring) - 1):
ents.append({"type": "line",
"p0": [ring[i][0], ring[i][1]],
"p1": [ring[i + 1][0], ring[i + 1][1]]})
try_call("sketch_cancel")
call("sketch_begin", plane="XY")
call("sketch_add", entities=ents)
r = call("sketch_describe")
ok = True
got = r["closed_loops"]
ok &= report(name, "CLOSED", f"engine finds {len(got)} closed loops, this script "
f"finds {len(rings)}", len(got) == len(rings))
# AREA — exact, because every entity is a line
mine = sorted(shoelace(x) for x in rings)
theirs = sorted(abs(l["area"]) for l in got)
# The bar: 1e-3 absolute, or 1e-6 relative for the big loops. Not bit-exactness — the
# auto-constraint pass still snaps segments that are already axis-aligned to within its
# 1e-4 rad tolerance, which moves an area by ~1e-4. It is deliberately tight enough to
# have caught the real defect this rung was written for: with the old 3 degree gesture
# slack applied to scripted input, a flattened circle came back 0.067% small — 0.266 on
# an area of 397, some 300x above this line.
same = len(mine) == len(theirs) and all(
abs(a - b) <= max(1e-3, 1e-6 * a) for a, b in zip(mine, theirs))
ok &= report(name, "AREA", "every loop area matches the shoelace value exactly", same)
# VOID — a loop belongs to the SMALLEST loop that contains it, not to every loop that
# encloses it. A hole inside a boss inside the part is a void of the boss, and the part
# owns the boss. Comparing against "everything inside the outline" was measuring my own
# sloppiness: on MPD781 that counted 36 voids where 26 of them are nested inside another
# void. So compute the same rule here, independently, and compare the whole attribution.
if got:
rings = [outer] + voids
# Probe from a point STRICTLY INSIDE each ring, never from one of its vertices — the
# same rule the engine now uses (DesignSketchTool::region_loops). A vertex is exactly
# where two loops touch in a real drawing, and a ray cast from a point lying ON the
# polygon under test answers by rounding: that alone accounted for every one of the 6
# sheets where the two attributions used to disagree. snaporca-5hvl.
probes = [interior_point(r) for r in rings]
mine_parent = {}
for i, r in enumerate(rings):
best, best_a = -1, 0.0
for j, q in enumerate(rings):
if i == j or not point_in(probes[i], q):
continue
a = shoelace(q)
if best < 0 or a < best_a:
best, best_a = j, a
if best >= 0:
mine_parent.setdefault(best, []).append(i)
big = max(range(len(got)), key=lambda i: abs(got[i]["area"]))
# match engine loops to my rings by area, then compare the two attributions by COUNT
mine_counts = sorted(len(v) for v in mine_parent.values())
got_counts = sorted(len(l["holes"]) for l in got if l["holes"])
ok &= report(name, "VOID",
f"void attribution matches: engine {got_counts}, containment "
f"{mine_counts}", got_counts == mine_counts)
# MIRROR / OFFSET — engine operations on a REAL profile, not a tidy one
n_outer = len(outer) - 1
try_call("sketch_cancel")
call("sketch_begin", plane="XY")
call("sketch_add", entities=ents[:n_outer])
xs = [p[0] for p in outer]
axis = min(xs) - 10.0
call("sketch_select", entities=list(range(n_outer)))
try:
call("sketch_mirror", axis_a=[axis, 0], axis_b=[axis, 1])
m = call("sketch_describe")
ok &= report(name, "MIRROR", "the mirrored copy is closed too",
len(m["closed_loops"]) == 2 and m["open_ends"] == [])
except RuntimeError as e:
ok &= report(name, "MIRROR", f"refused: {e}", False)
try_call("sketch_cancel")
call("sketch_begin", plane="XY")
call("sketch_add", entities=ents[:n_outer])
try:
call("sketch_offset", distance=0.5, entities=list(range(n_outer)))
o = call("sketch_describe")
ok &= report(name, "OFFSET", "the offset profile is still closed",
any(l["closed"] for l in o["closed_loops"]))
except RuntimeError as e:
ok &= report(name, "OFFSET", f"refused: {e}", False)
return ok
# ── scale ────────────────────────────────────────────────────────────────────
def grade_scale(pdf, name, report, budget):
"""Same exactness, on a profile of several hundred entities, and timed.
"Interactive" is measurable from here even though nothing is clicked: every MCP verb is
serviced on the UI THREAD, so the time a reply takes is time the window was not repainting.
A round trip that stays inside the budget is a window that stayed responsive.
"""
segs = drawing_segments(pdf)
loops = find_loops(segs)
if len(loops) < 2:
report(name, "SKIP", f"no nested closed geometry found ({len(loops)} loops)")
return None
loops.sort(key=shoelace, reverse=True)
outer = loops[1]
voids = [r for r in loops[2:] if shoelace(r) > 1.0 and point_in(r[0], outer)]
rings = [outer] + voids
ents = []
for ring in rings:
for i in range(len(ring) - 1):
ents.append({"type": "line",
"p0": [ring[i][0], ring[i][1]],
"p1": [ring[i + 1][0], ring[i + 1][1]]})
if len(ents) < 300:
report(name, "SKIP", f"only {len(ents)} entities — not a scale case")
return None
try_call("sketch_cancel")
call("sketch_begin", plane="XY")
t0 = time.monotonic(); call("sketch_add", entities=ents); t_add = time.monotonic() - t0
t0 = time.monotonic(); r = call("sketch_describe"); t_desc = time.monotonic() - t0
t0 = time.monotonic(); call("sketch_select", entities=list(range(len(ents))))
t_sel = time.monotonic() - t0
t0 = time.monotonic(); call("sketch_validate"); t_val = time.monotonic() - t0
ok = True
ok &= report(name, "SCALE", f"{len(ents)} entities in {len(rings)} loops", True)
got = r["closed_loops"]
ok &= report(name, "CLOSED", f"engine finds {len(got)} closed loops, this script "
f"finds {len(rings)}", len(got) == len(rings))
mine = sorted(shoelace(x) for x in rings)
theirs = sorted(abs(l["area"]) for l in got)
same = len(mine) == len(theirs) and all(
abs(a - b) <= max(1e-3, 1e-6 * a) for a, b in zip(mine, theirs))
ok &= report(name, "AREA", "every loop area matches the shoelace value exactly", same)
worst = max(t_add, t_desc, t_sel, t_val)
ok &= report(name, "TIME", f"add {t_add*1000:.0f} ms, describe {t_desc*1000:.0f} ms, "
f"select {t_sel*1000:.0f} ms, validate {t_val*1000:.0f} ms "
f"(budget {budget*1000:.0f} ms)", worst <= budget)
return ok
def _pdf_error(path):
"""What poppler says about a file it refused, so a refusal can be classified."""
r = subprocess.run(["pdfinfo", path], capture_output=True, text=True)
return (r.stderr or "") + (r.stdout or "")
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--corpus", default=os.path.expanduser("~/studycadcam"))
ap.add_argument("--step", type=int, default=20)
ap.add_argument("--limit", type=int, default=0)
ap.add_argument("--scale", action="store_true",
help="grade the LARGEST drawings instead: exactness plus a UI-thread budget")
ap.add_argument("--budget", type=float, default=2.0,
help="seconds; the slowest round trip a scale drawing may take")
a = ap.parse_args()
files = {}
for f in glob.glob(os.path.join(a.corpus, "MPD*.pdf")):
m = re.search(r"MPD(\d+)", os.path.basename(f))
if m:
files[int(m.group(1))] = f
# step 1 means EVERY sheet. Written as `n % step == 1` it silently selected nothing, because
# n % 1 is always 0 — and the run then printed "RUNG 9 HELD" over zero drawings graded. A
# gate that passes by grading nothing is worse than no gate, so the count is checked below.
picks = [files[n] for n in sorted(files) if a.step <= 1 or n % a.step == 1]
if a.scale:
# The heaviest real profiles in the corpus, biggest first — up to ~1300 entities.
sized = []
for f in files.values():
try:
segs = drawing_segments(f)
loops = find_loops(segs)
if len(loops) < 2:
continue
loops.sort(key=shoelace, reverse=True)
outer = loops[1]
voids = [r for r in loops[2:] if shoelace(r) > 1.0 and point_in(r[0], outer)]
sized.append((sum(len(r) - 1 for r in [outer] + voids), f))
except Exception: # noqa: BLE001
continue
sized.sort(reverse=True)
picks = [f for _, f in sized[:max(1, a.limit or 6)]]
elif a.limit:
picks = picks[:a.limit]
print(f"corpus: {len(files)} sheets; systematic sample every {a.step}th "
f"-> {len(picks)} drawings\n")
fails, results = [], []
def report(name, tag, msg, cond=True):
if tag == "SKIP":
print(f" {name:9s} SKIP {msg}")
return True
mark = "ok " if cond else "FAIL"
print(f" {name:9s} {tag:8s} {mark} {msg}")
if not cond:
fails.append(f"{name} {tag}: {msg}")
return cond
for f in picks:
name = re.search(r"MPD\d+", os.path.basename(f)).group(0)
try:
r = (grade_scale(f, name, report, a.budget) if a.scale
else grade(f, name, report))
if r is not None:
results.append((name, r))
except Exception as e: # noqa: BLE001
# An unreadable SOURCE file is not a grading failure. MPD133 of this corpus is
# password-protected, and pdftocairo says so on stderr while exiting non-zero;
# reporting that as ERROR made one encrypted sheet look like an engine defect.
if "password" in _pdf_error(f).lower():
report(name, "SKIP", "the PDF is password-protected — nothing to extract")
else:
report(name, "ERROR", str(e)[:120], False)
graded = len(results)
passed = sum(1 for _, r in results if r)
if graded == 0:
print("\nNOTHING WAS GRADED — that is a harness failure, not a clean run", file=sys.stderr)
sys.exit(2)
print(f"\ngraded {graded} drawings; {passed} fully clean, {graded - passed} with "
f"at least one failure")
if fails:
print("\nfailures:")
for x in fails:
print(" " + x)
sys.exit(1)
print("\nRUNG 9 HELD — the engine agrees with the drawings, not with me")
if __name__ == "__main__":
main()
+352
View File
@@ -0,0 +1,352 @@
#!/usr/bin/env python3
"""A ladder of 2D sketches of increasing complexity, judged the way a person judges them.
WHY NOT AREA. Area is derived and no one can confirm it by looking. What a human checks at a
glance, and can be exactly right or exactly wrong about, is:
VERTEX is the corner where I said it is
LENGTH is the side the length I gave it
ARC is the radius the radius I gave it
TANGENT does the straight run into the curve smoothly, or is there a kink
SYMMETRY is the mirrored half the exact reflection of the half I drew
CLOSED is it one closed loop, or does it just look like one
Every rung asserts those. Area appears only as a cross-check, never as the verdict.
Entirely 2D: sketch entities only, no extrude, revolve or any solid feature.
SNAPORCA_MCP=/tmp/mcp.sock <binary>
python3 scripts/CAD/check-sketch-engine.py [socket]
Exit 0 = every rung held. Otherwise the first broken property is named and the run stops.
"""
import json, math, socket, sys
SOCK = sys.argv[1] if len(sys.argv) > 1 else "/tmp/mcp.sock"
EPS = 1e-9
_n = 0
_fail = 0
def call(method, **params):
global _n
_n += 1
s = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
s.settimeout(30)
s.connect(SOCK)
s.sendall((json.dumps({"jsonrpc": "2.0", "id": _n, "method": method,
"params": params}) + "\n").encode())
buf = b""
while b"\n" not in buf:
d = s.recv(65536)
if not d:
break
buf += d
r = json.loads(buf.decode().strip())
if "error" in r:
raise RuntimeError(f"{method}: {r['error']['message']}")
return r["result"]
def check(kind, cond, what):
global _fail
if cond:
print(f" {kind:9s} ok {what}")
else:
print(f" {kind:9s} FAIL {what}", file=sys.stderr)
_fail += 1
def near(a, b, tol=1e-6):
return abs(a - b) <= tol
def pt_near(p, q, tol=1e-6):
return math.hypot(p[0] - q[0], p[1] - q[1]) <= tol
def fresh(plane="XY"):
try:
call("sketch_cancel")
except Exception:
pass
call("sketch_begin", plane=plane)
def ents():
return call("sketch_describe")["entities"]
def rep():
return call("sketch_describe")
def endpoints(e):
"""Both ends of an open curve, as tuples. Closed curves have none."""
if "p0" not in e or "p1" not in e:
return ()
return tuple(e["p0"]), tuple(e["p1"])
def tangent(e, at_end):
"""Unit tangent of entity e at one of its ends, pointing ALONG the curve (p0->p1)."""
if e["type"] == "line":
dx = e["p1"][0] - e["p0"][0]
dy = e["p1"][1] - e["p0"][1]
else: # arc
a = e["start_angle"] if not at_end else e["end_angle"]
ccw = e["end_angle"] >= e["start_angle"]
# d/dtheta (cos, sin) = (-sin, cos), reversed when the sweep is clockwise
dx, dy = -math.sin(a), math.cos(a)
if not ccw:
dx, dy = -dx, -dy
n = math.hypot(dx, dy)
return (dx / n, dy / n)
def tangent_at_point(e, p):
"""Unit tangent of e at whichever of its ends is p, oriented leaving that point."""
p0, p1 = endpoints(e)
if pt_near(p0, p):
t = tangent(e, False)
return t
t = tangent(e, True)
return (-t[0], -t[1]) # leaving p1 means going back along the curve
def smooth(e1, e2, p):
"""G1 at shared point p: the tangent leaving e1 is opposite the tangent leaving e2."""
a = tangent_at_point(e1, p)
b = tangent_at_point(e2, p)
return abs(a[0] * (-b[0]) - 0) >= 0 and abs(a[0] * b[1] - a[1] * b[0]) <= 1e-6
def closed_one_loop(r, voids=0):
return (r["buildable"] and r["open_ends"] == []
and len([l for l in r["closed_loops"] if not any(
i in h["holes"] for h in r["closed_loops"] for i in [])]) >= 1)
def outer_loop(r):
"""The loop that encloses the others (or the only one)."""
if not r["closed_loops"]:
return None
return max(r["closed_loops"], key=lambda l: abs(l["area"]))
# ─────────────────────────────────────────────────────────────────────────────
print("RUNG 1 — rectangle: four corners, four lengths, four right angles")
fresh()
W, H = 80.0, 50.0
call("sketch_add", rect=[0, 0, W, H])
r = rep()
es = r["entities"]
corners = {(0, 0), (W, 0), (W, H), (0, H)}
got = set()
for e in es:
got.add(tuple(e["p0"]))
got.add(tuple(e["p1"]))
check("VERTEX", all(any(pt_near(c, g) for g in got) for c in corners),
f"all four corners exactly where asked {sorted(corners)}")
lens = sorted(round(e["length"], 9) for e in es)
check("LENGTH", lens == sorted([W, W, H, H]), f"sides are {W}/{H} twice each (got {lens})")
# right angles: consecutive sides meet at 90 degrees
ang_ok = True
for e in es:
for f in es:
if e is f:
continue
for p in endpoints(e):
if any(pt_near(p, q) for q in endpoints(f)):
a, b = tangent_at_point(e, p), tangent_at_point(f, p)
if abs(a[0] * b[0] + a[1] * b[1]) > 1e-6:
ang_ok = False
check("ANGLE", ang_ok, "every corner is exactly 90 degrees")
check("CLOSED", r["buildable"] and r["open_ends"] == [], "one closed loop, no free ends")
print("\nRUNG 2 — a circular void inside it")
call("sketch_add", type="circle", center=[W / 2, H / 2], radius=12)
r = rep()
c = [e for e in r["entities"] if e["type"] == "circle"][0]
check("VERTEX", pt_near(tuple(c["center"]), (W / 2, H / 2)), "void centred exactly where asked")
check("ARC", near(c["radius"], 12), f"void radius exactly 12 (got {c['radius']})")
out = outer_loop(r)
check("CLOSED", len(out["holes"]) == 1, "the rectangle encloses exactly one void")
check("CLOSED", r["buildable"] and r["open_ends"] == [], "still closed with the void present")
print("\nRUNG 3 — stadium: straights running into caps, tangent at every junction")
fresh()
L, R = 50.0, 15.0
call("sketch_add", entities=[
{"type": "line", "p0": [-L, -R], "p1": [L, -R]},
{"type": "arc", "center": [L, 0], "radius": R,
"start_angle": -math.pi / 2, "end_angle": math.pi / 2},
{"type": "line", "p0": [L, R], "p1": [-L, R]},
{"type": "arc", "center": [-L, 0], "radius": R,
"start_angle": math.pi / 2, "end_angle": 3 * math.pi / 2},
])
r = rep()
es = r["entities"]
check("CLOSED", r["buildable"] and r["open_ends"] == [], "one closed loop, no free ends")
arcs = [e for e in es if e["type"] == "arc"]
check("ARC", all(near(a["radius"], R) for a in arcs), f"both caps exactly R={R}")
check("LENGTH", all(near(e["length"], 2 * L) for e in es if e["type"] == "line"),
f"both straights exactly {2*L}")
# tangency at all four line/arc junctions
tang = True
for a in arcs:
for p in endpoints(a):
mates = [e for e in es if e is not a and any(pt_near(p, q) for q in endpoints(e))]
for m in mates:
if not smooth(a, m, p):
tang = False
check("TANGENT", tang, "straight meets cap smoothly at all four junctions (no kink)")
print("\nRUNG 4 — mirror: the reflected half is the exact reflection")
fresh()
half = [
{"type": "line", "p0": [0, -R], "p1": [L, -R]},
{"type": "arc", "center": [L, 0], "radius": R,
"start_angle": -math.pi / 2, "end_angle": math.pi / 2},
{"type": "line", "p0": [L, R], "p1": [0, R]},
]
call("sketch_add", entities=half)
r = rep()
check("CLOSED", not r["buildable"] and len(r["open_ends"]) == 2,
f"half profile is correctly OPEN, both ends named {r['open_ends']}")
call("sketch_select", entities=[0, 1, 2])
call("sketch_mirror", axis_a=[0, 0], axis_b=[0, 1])
r = rep()
es = r["entities"]
check("CLOSED", r["buildable"] and r["open_ends"] == [], "mirroring closed the loop")
# every source vertex must have its exact reflection present
src = []
for e in es[:3]:
src += [tuple(e["p0"]), tuple(e["p1"])]
allv = []
for e in es:
allv += [tuple(e["p0"]), tuple(e["p1"])]
sym = all(any(pt_near((-x, y), v) for v in allv) for (x, y) in src)
check("SYMMETRY", sym, "every vertex has its exact mirror twin across x=0")
mirrored_arc = [e for e in es[3:] if e["type"] == "arc"]
check("ARC", mirrored_arc and near(mirrored_arc[0]["radius"], R)
and pt_near(tuple(mirrored_arc[0]["center"]), (-L, 0)),
f"mirrored cap keeps R={R} and lands at (-{L}, 0)")
print("\nRUNG 5 — offset: every curve moves by exactly d, and it stays closed")
d = 4.0
call("sketch_select", entities=list(range(len(es))))
call("sketch_offset", distance=-d) # -d = outward for this CCW loop
r = rep()
new = r["entities"][len(es):]
check("CLOSED", r["buildable"] and r["open_ends"] == [], "offset result is closed")
off_arcs = [e for e in new if e["type"] == "arc"]
check("ARC", all(near(a["radius"], R + d) for a in off_arcs),
f"each cap radius grew by exactly {d} -> {R+d}")
off_lines = [e for e in new if e["type"] == "line"]
check("VERTEX", all(near(abs(e["p0"][1]), R + d) for e in off_lines),
f"each straight moved out to |y| = {R+d} exactly")
print("\nRUNG 6 — a gap is found by coordinate, then closed by a real constraint")
fresh()
call("sketch_add", entities=[
{"type": "line", "p0": [0, 0], "p1": [60, 0]},
{"type": "line", "p0": [60, 0], "p1": [60, 40]},
{"type": "line", "p0": [60, 40], "p1": [0, 40]},
{"type": "line", "p0": [0, 40], "p1": [0.35, 0]}, # 0.35 mm short
])
r = call("sketch_validate", tolerance=1.0)
check("CLOSED", not r["buildable"] and len(r["open_ends"]) == 2,
f"the gap is reported, both free ends named {r['open_ends']}")
dof0 = r["dof"]
r = call("sketch_heal", tolerance=1.0)
check("CLOSED", r["buildable"] and r["open_ends"] == [], "healed into a closed loop")
check("VERTEX", r["welded"] == 1, "exactly one pair of vertices welded")
check("ANGLE", r["dof"] < dof0,
f"the weld is a real constraint, not a nudge: DoF {dof0} -> {r['dof']}")
es = ents()
check("VERTEX", pt_near(tuple(es[3]["p1"]), tuple(es[0]["p0"])),
"the two ends are now the same point")
print("\nRUNG 7 — the composite: mirrored, tangent, two voids, all at once")
fresh()
call("sketch_add", entities=half)
call("sketch_select", entities=[0, 1, 2])
call("sketch_mirror", axis_a=[0, 0], axis_b=[0, 1])
call("sketch_add", type="circle", center=[-25, 0], radius=6)
call("sketch_add", type="circle", center=[25, 0], radius=6)
r = rep()
es = r["entities"]
out = outer_loop(r)
check("CLOSED", r["buildable"] and r["open_ends"] == [], "one closed outer loop, no free ends")
check("CLOSED", len(out["holes"]) == 2, "it encloses exactly two voids")
circles = [e for e in es if e["type"] == "circle"]
check("ARC", all(near(c["radius"], 6) for c in circles), "both voids exactly R=6")
check("SYMMETRY", pt_near(tuple(circles[0]["center"]), (-25, 0))
and pt_near(tuple(circles[1]["center"]), (25, 0)),
"the voids sit symmetrically at x = -25 and +25")
tang = True
for a in [e for e in es if e["type"] == "arc"]:
for p in endpoints(a):
for m in [e for e in es if e is not a and any(pt_near(p, q) for q in endpoints(e))]:
if not smooth(a, m, p):
tang = False
check("TANGENT", tang, "every straight-to-cap junction is still smooth")
exact = 2 * L * 2 * R + math.pi * R * R
check("LENGTH", near(out["area"], exact, 1e-6),
f"cross-check: enclosed area {out['area']:.4f} = 2L*2R + pi*R^2 = {exact:.4f}")
print("\nRUNG 8 — a real drawing: StudyCadCam MPD5, the pin's revolve half-profile")
# Ø27 x 95 pin: C1 chamfer on the left end, cylinder to a corner at x=85, an R5 fillet into a
# cone at 23 degrees to the axis, right face at x=95. Interpretation stated so the rung is
# reproducible: 85 is to the CORNER, 23 deg is to the AXIS, C1 is 1 x 45.
fresh()
RAD, LEN, TX, ANG, RF, CH = 13.5, 95.0, 85.0, math.radians(23), 5.0, 1.0
t = RF * math.tan(ANG / 2)
ax, ay = TX - t, RAD # fillet tangent point on the cylinder
cx, cy = ax, RAD - RF # fillet centre
bx, by = TX + t * math.cos(-ANG), RAD + t * math.sin(-ANG) # tangent point on the cone
ey = by - (LEN - bx) * math.tan(ANG) # where the cone meets the right face
call("sketch_add", entities=[
{"type": "line", "p0": [0, 0], "p1": [0, RAD - CH]}, # left face
{"type": "line", "p0": [0, RAD - CH], "p1": [CH, RAD]}, # C1 chamfer
{"type": "line", "p0": [CH, RAD], "p1": [ax, ay]}, # cylinder top
{"type": "arc", "center": [cx, cy], "radius": RF,
"start_angle": math.pi / 2, "end_angle": math.pi / 2 - ANG}, # R5 fillet
{"type": "line", "p0": [bx, by], "p1": [LEN, ey]}, # 23 deg cone
{"type": "line", "p0": [LEN, ey], "p1": [LEN, 0]}, # right face
{"type": "line", "p0": [LEN, 0], "p1": [0, 0]}, # axis
])
r = rep()
es = r["entities"]
check("CLOSED", r["buildable"] and r["open_ends"] == [], "the half-profile is one closed loop")
xs = [v[0] for e in es if "p0" in e for v in (e["p0"], e["p1"])]
ys = [v[1] for e in es if "p0" in e for v in (e["p0"], e["p1"])]
check("LENGTH", near(max(xs) - min(xs), LEN), f"overall length exactly {LEN} (the 95 dimension)")
check("VERTEX", near(max(ys), RAD), f"outer radius exactly {RAD} (the dia 27)")
fil = [e for e in es if e["type"] == "arc"][0]
check("ARC", near(fil["radius"], RF), f"the corner fillet is exactly R{RF:g}")
cone = [e for e in es if e["type"] == "line"
and not near(e["p0"][0], e["p1"][0]) and not near(e["p0"][1], e["p1"][1])
and e["length"] > 5]
if cone:
c0 = cone[0]
a = abs(math.degrees(math.atan2(c0["p1"][1] - c0["p0"][1], c0["p1"][0] - c0["p0"][0])))
check("ANGLE", near(a, 23, 1e-6), f"the cone is exactly 23 degrees to the axis (got {a:.6f})")
cham = [e for e in es if e["type"] == "line" and near(e["length"], CH * math.sqrt(2), 1e-9)]
check("ANGLE", bool(cham), "the C1 chamfer is exactly 1 x 45 (length 1*sqrt2)")
tang = True
for p in endpoints(fil):
for m in [e for e in es if e is not fil and any(pt_near(p, q) for q in endpoints(e))]:
if not smooth(fil, m, p):
tang = False
check("TANGENT", tang, "the fillet is tangent to BOTH the cylinder and the cone (no kink)")
call("sketch_cancel")
try:
call("sketch_cancel")
except Exception:
pass # a rung may have closed it already
print(f"\n{'ALL RUNGS HELD' if _fail == 0 else str(_fail) + ' CHECK(S) FAILED'}")
sys.exit(1 if _fail else 0)
+69
View File
@@ -0,0 +1,69 @@
#!/usr/bin/env bash
# Every ladder, in one command, as the gate before a push that touched the Design tab.
#
# WHY A SCRIPT AND NOT CI. Three of the four rungs need a running application with an OpenGL
# canvas and synthetic input; GitHub's runners have neither. So the gate is local and explicit:
# run this, read the last line, and do not push a red one. The kernel suite is the only part CI
# can carry, and it already does.
#
# scripts/CAD/run-all-checks.sh # kernel + engine + corpus (every 20th) + gestures + offer
# FULL=1 scripts/CAD/run-all-checks.sh # corpus over ALL 997 sheets (~25 min)
# SKIP_GUI=1 scripts/CAD/run-all-checks.sh # kernel only, for a machine with no rig
#
# The rig container is expected to be up with the app running and SNAPORCA_MCP set; bring it up
# with scripts/CAD/start-headless-gui.sh inside it. The corpus lives at /corpus in that container.
set -uo pipefail
cd "$(dirname "$0")/.." || exit 1
C="${C:-snaporca-gui}"
CORPUS="${CORPUS:-/corpus}"
STEP="${STEP:-20}"
[ -n "${FULL:-}" ] && STEP=1
fail=0
step() {
local name="$1"; shift
echo
echo "=== $name ==="
if "$@"; then echo "--- $name OK"; else echo "--- $name FAILED"; fail=1; fi
}
# SC2329: every call goes through step(), which invokes it via "$@", so shellcheck
# cannot see the callers below.
# shellcheck disable=SC2329
run_in_rig() { # copy the script in fresh, then run it there
docker cp "$1" "$C:/tmp/$(basename "$1")" >/dev/null || return 1
shift
docker exec "$C" python3 "$@"
}
# FIRST, and it needs no rig: the offer table the menu is compiled from must be what the atlas
# says. The header calls itself GENERATED and had been hand-edited anyway — which cost four rows
# that existed only in the header, one row wired to the wrong action, and a count of 91 for a
# 92-row array, so the last verb was unreachable (snaporca-z8rs, snaporca-ziam).
step "offer table matches the atlas" python3 docs/ux/mockups/gen_offer_table.py --check
step "kernel suite" scripts/CAD/run-kernel-tests.sh --vol "${KVOL:-snaporca_kerneltest}"
if [ -z "${SKIP_GUI:-}" ]; then
step "engine ladder (rungs 1-8, scripted geometry)" \
run_in_rig scripts/CAD/check-sketch-engine.py /tmp/check-sketch-engine.py
step "corpus rung (real drawings, every ${STEP}th)" \
run_in_rig scripts/CAD/check-sketch-engine-corpus.py /tmp/check-sketch-engine-corpus.py --corpus "$CORPUS" --step "$STEP"
step "corpus scale rung (the heaviest sheets)" \
run_in_rig scripts/CAD/check-sketch-engine-corpus.py /tmp/check-sketch-engine-corpus.py --corpus "$CORPUS" --scale
step "gesture ladder (mouse and keyboard)" \
run_in_rig scripts/CAD/check-gui-sketching.py /tmp/check-gui-sketching.py
# The offer ladder needs TWO extra things the others do not: the app must have been launched
# with SNAPORCA_KEYTRACE=1 (its [OFFER] lines are the whole instrument), and it reads the
# generated offer table to predict what each selection should show — which is not in the
# container's own baked source tree, so it is copied in beside the script — /tmp, where
# run_in_rig puts the script, is one of the paths the ladder looks in.
docker cp src/slic3r/GUI/CAD/DesignOffer.hpp "$C:/tmp/DesignOffer.hpp" >/dev/null
step "offer ladder (right-click, the menu, the verbs behind it)" \
run_in_rig scripts/CAD/check-gui-context-menu.py /tmp/check-gui-context-menu.py
fi
echo
if [ "$fail" -eq 0 ]; then echo "ALL LADDERS HELD"; else echo "AT LEAST ONE LADDER FAILED"; fi
exit "$fail"
+149
View File
@@ -0,0 +1,149 @@
#!/usr/bin/env bash
# Headless CAD-kernel test loop. THE verification contract for delegated work:
# exit 0 means the kernel builds and the selected Catch2 tags pass. Nothing else counts.
#
# Builds only the `libslic3r_tests` target (not the GUI app), so a round trip is
# minutes, not tens of minutes. Needs no display: every [CadDocument] case builds a
# CadDocument, recompute()s it and asserts on geometry.
#
# Usage:
# scripts/CAD/run-kernel-tests.sh # [CadDocument] tags, default volume
# scripts/CAD/run-kernel-tests.sh --tags '[CadDocument],[Sketch]'
# scripts/CAD/run-kernel-tests.sh --vol wt_mirror # private build cache (parallel workers)
# scripts/CAD/run-kernel-tests.sh --host tommaso@100.103.234.2 # build on a remote host
#
# Parallel workers MUST pass a distinct --vol: two builds sharing one cache corrupt
# each other. A new volume pays one full build; runs after that are incremental.
#
# --host exists because the deps image lives wherever it was first built. It rsyncs this
# working tree to a per-volume staging dir on that host and re-runs this same script
# there, so the verification contract is identical either way. Drop --host once the image
# is present locally.
# Rig build traps already paid for once each (stale project, NLopt cache, pybind11, OCCT_LIBS, SLIC3R_CAD gate): docs/rig_build_traps.md
set -euo pipefail
REPO="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
# orcacad-deps, NOT snaporca-deps: this fork is mainline-based and needs Eigen 5.0.1,
# CGAL 5.6.3, wx 3.3.2 and Python 3.12 Development.Embed, none of which snaporca-deps has.
# With the wrong image CMake dies at configure, which is exactly why this fork went
# M1-M8 without ever compiling (see commit 1633005bba).
IMAGE="${IMAGE:-orcacad-deps}"
# Must NOT default to snaporca_buildcache: that is the other fork's volume, and pointing
# this fork at it makes the two silently trade build artefacts. build-gui-incremental.sh had the
# identical defect and was fixed to orcacad_buildcache; this script was missed.
VOL="${BUILD_VOL:-orcacad_kerneltest}"
# No exclusions. Both cases that used to be quarantined now run: the solver SIGABRT on
# circle-line tangency is fixed (snaporca-tkz), and the internal-thread case turned out to have
# correct geometry and a wrong reference in the test (snaporca-kzy). A green run here now means
# the whole CAD suite passed, not "everything except the two we gave up on".
TAGS="${TAGS:-[CadDocument]}"
HOST=""
while [[ $# -gt 0 ]]; do
case "$1" in
--vol) VOL="$2"; shift 2 ;;
--tags) TAGS="$2"; shift 2 ;;
--image) IMAGE="$2"; shift 2 ;;
--host) HOST="$2"; shift 2 ;;
*) echo "unknown arg: $1" >&2; exit 2 ;;
esac
done
echo "REPO=$REPO IMAGE=$IMAGE VOL=$VOL TAGS=$TAGS HOST=${HOST:-local}"
if [[ -n "$HOST" ]]; then
# Stage per-volume so parallel workers never share a remote tree.
REMOTE="kt-$VOL" # relative: ssh and rsync both start in the remote home dir
# SC2029: $REMOTE expanding on the CLIENT is the point -- it is derived from $VOL here,
# and the remote has no such variable. The rsync destination below expands it the same way.
# shellcheck disable=SC2029
ssh "$HOST" "mkdir -p $REMOTE"
# Only the inputs the build reads. --delete keeps a stale file from a prior worker from
# silently compiling in.
rsync -a --delete \
"$REPO/src" "$REPO/tests" "$REPO/resources" "$REPO/cmake" "$REPO/scripts" \
"$REPO/CMakeLists.txt" \
"$HOST:$REMOTE/"
exec ssh "$HOST" "cd $REMOTE && scripts/CAD/run-kernel-tests.sh --vol '$VOL' --tags '$TAGS' --image '$IMAGE'"
fi
if ! docker image inspect "$IMAGE" >/dev/null 2>&1; then
echo "FATAL: image '$IMAGE' not present locally. Either transfer it, or pass" >&2
echo " --host <user@host> to build where the image already exists." >&2
exit 3
fi
# A private volume is always built CLEAN on first use. Cloning a warm cache from another
# tree looks tempting but is a correctness trap: rsync preserves source mtimes, so ninja
# compares foreign object timestamps against them, decides everything is up to date, and
# relinks stale objects. That produced a binary with NO [CadDocument] tests at all while
# reporting success -- a green run that tested nothing. Pay the one-off full build instead;
# incremental rebuilds within a volume are correct because the mtimes then share a lineage.
if ! docker volume inspect "$VOL" >/dev/null 2>&1; then
echo "=== $VOL: new volume, clean configure + full build (one-off, slow) ==="
docker volume create "$VOL" >/dev/null
fi
# tests/ is mounted too -- unlike build-gui-incremental.sh, this script exists precisely to
# compile tests being edited. CMakeLists.txt and cmake/ carry the SLIC3R_CAD gate; taking
# them from the baked image instead leaves the gate off and the CAD symbols vanish.
# ---- OOM guard (2026-08-21) -------------------------------------------------------------
# Two of these builds ran at once on 2026-08-21, each with ninja -j$(nproc)=16: ~36 cc1plus
# holding 42 GB of a 62 GB box -> global OOM at 21:05, a 2h28m kill storm, ssh unreachable,
# lightdm destroyed. Neither build produced a single object. scripts/CAD/build-gui.sh grew the
# bounds first; every script that starts a compile needs the same three, or the guard is only
# as strong as the script you happened not to use.
# flock — the lock path is SHARED with build-gui.sh and the other fork on purpose, so
# concurrent builds serialise instead of summing.
# -j — bounded parallelism; ~1.17 GB per cc1plus was the measured average.
# --memory — the actual guarantee: a runaway build dies in its own cgroup instead of taking
# the host down. --memory-swap equal to --memory forbids swap, which is what made
# ssh hang.
JOBS="${JOBS:-12}"
MEM="${MEM:-40g}"
LOCK=/tmp/orca-rig-build.lock
exec 9>"$LOCK"
if ! flock -n 9; then
echo "another build holds $LOCK — waiting (this is the OOM guard, not a hang)"
flock 9
fi
docker run --rm \
--memory="$MEM" --memory-swap="$MEM" \
-v "$REPO/src":/OrcaSlicer/src \
-v "$REPO/tests":/OrcaSlicer/tests \
-v "$REPO/resources":/OrcaSlicer/resources \
-v "$REPO/CMakeLists.txt":/OrcaSlicer/CMakeLists.txt \
-v "$REPO/cmake":/OrcaSlicer/cmake \
-v "$REPO/deps_src":/OrcaSlicer/deps_src \
-v "$VOL":/OrcaSlicer/build \
"$IMAGE" \
bash -lc "set -e
cd /OrcaSlicer
DESTDIR=/OrcaSlicer/deps/build/destdir/usr/local
export PATH=\$DESTDIR/bin:\$PATH # wx-config, and anything else the deps prefix ships
# Configure unconditionally. Guarding on 'CMakeCache.txt exists' is wrong: a FAILED
# configure still writes that file, so the guard then skips reconfiguring forever and
# every later run reuses a poisoned cache while ignoring corrected flags. A no-op
# reconfigure costs seconds; that bug costs an afternoon.
# SLIC3R_GTK=3 and BUILD_TESTS=ON are not optional extras: src/CMakeLists.txt turns
# SLIC3R_GTK into 'wx-config --toolkit=gtk<N>', so omitting it asks for toolkit 'gtk'
# and no wx build matches -> 'Could NOT find wxWidgets'. BUILD_TESTS=ON is what makes
# the libslic3r_tests target exist at all. build_linux.sh sets both (lines 218, 226).
cmake -S . -B build -G 'Ninja Multi-Config' \
-DCMAKE_PREFIX_PATH=\$DESTDIR \
-DwxWidgets_CONFIG_EXECUTABLE=\$DESTDIR/bin/wx-config \
-DSLIC3R_GTK=3 -DBUILD_TESTS=ON -DSLIC3R_GUI=OFF \
-DSLIC3R_CAD=ON -DSLIC3R_STATIC=1 -DORCA_TOOLS=ON -DCMAKE_BUILD_TYPE=Release
# SLIC3R_GUI=OFF: this script builds ONLY libslic3r_tests, which links libslic3r and no GUI
# code, but cmake still PROCESSES the if SLIC3R_GUI block of src/CMakeLists.txt (lines 16-98)
# and every find_package inside it. That made the kernel suite depend on the GUI dependency
# set for no benefit, and it broke the moment upstream added wxInspector as a REQUIRED
# find_package at line 92: the orcacad-deps image predates it, so configure died pointing at
# src/CMakeLists.txt:92 while nothing about the kernel had changed. Turning the block off is
# not a workaround for that one dependency; it is the kernel suite finally declaring what it
# actually needs, so the next GUI-side dependency upstream adds cannot break it either.
cmake --build build --config Release --target libslic3r_tests -- -j$JOBS
./build/tests/libslic3r/Release/libslic3r_tests '$TAGS' --order decl"
+137
View File
@@ -0,0 +1,137 @@
#!/usr/bin/env bash
# Bring the headless GUI up on a VNC-served X display, ready to drive or to attach Remmina to.
#
# Runs INSIDE the long-lived GUI container (see the header of scripts/CAD/build-gui-incremental.sh for how
# that container is created). Idempotent: safe to re-run to recover a session whose app died.
#
# docker exec <container> /OrcaSlicer/scripts/CAD/start-headless-gui.sh # launch + settle
# docker exec <container> /OrcaSlicer/scripts/CAD/start-headless-gui.sh --status # report, change nothing
#
# WHY THIS EXISTS. Dismissing the first-run dialogs by computing the titlebar close box from
# `xdotool getwindowgeometry --shell` and clicking it went wrong whenever the dialog had already
# closed: the eval left the geometry variables stale or empty, the click landed at a garbage
# coordinate, and it repeatedly hit the Sketch button in the toolbar underneath, so the app came up
# in sketch mode with a stray Sketch feature. Three of those in one session.
#
# The titlebar click is nevertheless the RIGHT mechanism and is kept. `xdotool windowclose` looks
# cleaner but kills the app: it destroys the GdkWindow out from under the dialog and the process
# dies with "GdkWindow unexpectedly destroyed", three GLib-GObject criticals and a segfault
# (measured 2026-07-30). Escape does not close the Setup Wizard either. So the fix is not a
# different mechanism, it is refusing to click on geometry we have not validated.
set -euo pipefail
DISP="${DISP:-:11}"
GEOM="${GEOM:-1920x1080}"
BIN="${BIN:-/OrcaSlicer/build/src/Release/orca-slicer}"
# Packaging cannot bundle python (the deps python layer carries a doubled-DESTDIR RUNPATH), so the
# build-tree binary needs the deps libpython on the path. Packaging-only issue; the app runs fine.
LIBPY="/OrcaSlicer/deps/build/destdir/usr/local/libpython/lib"
LIBPY2="/OrcaSlicer/build/src/Release/python/lib"
LOG="${LOG:-/tmp/gui-session.log}"
export DISPLAY="$DISP" HOME=/root
export LIBGL_ALWAYS_SOFTWARE=1 GALLIUM_DRIVER=llvmpipe
export LD_LIBRARY_PATH="$LIBPY:$LIBPY2:${LD_LIBRARY_PATH:-}"
mkdir -p /root/.config # startup dies in boost::filesystem::create_directory without this
# Skip zombies. This container accumulates <defunct> instances of the app across runs, and pgrep
# matches them, so the naive "first match" reported a dead pid as if the session were healthy.
app_pid() {
local p
for p in $(pgrep -f "$(basename "$BIN")" 2>/dev/null); do
[ "$(awk "{print \$3}" "/proc/$p/stat" 2>/dev/null)" = "Z" ] && continue
echo "$p"; return 0
done
return 1
}
status() {
echo "display : $(pgrep -f "Xvfb $DISP" >/dev/null && echo up || echo DOWN)"
echo "wm : $(pgrep -x openbox >/dev/null && echo up || echo DOWN)"
if pgrep -x x11vnc >/dev/null; then echo "vnc : up on :5900"
elif ! command -v x11vnc >/dev/null; then echo "vnc : n/a (x11vnc not installed here)"
else echo "vnc : DOWN"; fi
local p; p="$(app_pid || true)"
echo "app : ${p:-DOWN}"
# WHICH binary is on screen, not just that something is. A pid alone cannot tell you whether
# you are looking at the build you just linked or one from last week, and that is precisely
# the question every rig verification is asking.
[ -n "${p:-}" ] && echo "binary : $(readlink -f "/proc/$p/exe" 2>/dev/null || echo unknown)"
[ -n "${p:-}" ] && echo "windows : $(xdotool search --name . getwindowname %@ 2>/dev/null | paste -sd'|' -)"
return 0
}
[ "${1:-}" = "--status" ] && { status; exit 0; }
# --- desktop: Xvfb, a window manager, and the VNC server ------------------------------------
# openbox is REQUIRED: without it xdotool windowactivate aborts with "windowmanager claims not to
# support _NET_ACTIVE_WINDOW" and dialogs never take focus.
pgrep -f "Xvfb $DISP" >/dev/null || { nohup Xvfb "$DISP" -screen 0 "${GEOM}x24" -nolisten tcp >/tmp/xvfb.log 2>&1 & sleep 3; }
pgrep -x openbox >/dev/null || { nohup openbox >/tmp/openbox.log 2>&1 & sleep 1; }
if ! pgrep -x x11vnc >/dev/null && command -v x11vnc >/dev/null; then
AUTH=()
[ -f /root/.vnc/passwd ] && AUTH=(-rfbauth /root/.vnc/passwd)
nohup x11vnc -display "$DISP" -rfbport 5900 "${AUTH[@]}" -forever -shared -noxdamage \
>/tmp/x11vnc.log 2>&1 &
sleep 2
fi
# --- app ------------------------------------------------------------------------------------
# Kill by BASENAME, not by "$BIN". The app enforces a single instance, so an older copy launched
# from a DIFFERENT path (the packaged build/package/bin/ one, say, when BIN points at the freshly
# linked build/src/Release/ one) survives a path-matched pkill, keeps the instance lock, and the
# new process exits seconds after loading fonts — leaving no error anywhere. app_pid() below has
# always matched by basename, so status then reported that stale process as a healthy session:
# the launch looked green while the window on screen was days old. Measured 2026-08-02, where it
# nearly passed a UI change against a Jul-30 binary. The killer and the reporter must agree on
# what counts as "the app".
pkill -9 -f "$(basename "$BIN")" 2>/dev/null || true
sleep 2
nohup "$BIN" >"$LOG" 2>&1 &
echo "launched $(basename "$BIN") pid $!"
# Wait for the main window rather than sleeping a fixed amount: cold starts vary a lot under
# software GL, and a fixed sleep either wastes time or races.
for _ in $(seq 1 40); do
xdotool search --name "Untitled" >/dev/null 2>&1 && break
sleep 1
done
# --- first-run dialogs ----------------------------------------------------------------------
# Click the titlebar close box, but only on geometry we have just read for a window that still
# exists, and only if the resulting point is inside the screen. Every variable is unset first so a
# failed read cannot leave the previous dialog's numbers behind — that is the whole bug.
screen_w="${GEOM%x*}"; screen_h="${GEOM#*x}"
close_dialog() {
local name="$1" id X Y WIDTH HEIGHT cx cy
id="$(xdotool search --name "$name" 2>/dev/null | head -1 || true)"
[ -z "$id" ] && return 1
unset X Y WIDTH HEIGHT
eval "$(xdotool getwindowgeometry --shell "$id" 2>/dev/null || true)"
# All four must be present and numeric: an empty or stale read is how the stray click happened.
for v in "${X:-}" "${Y:-}" "${WIDTH:-}" "${HEIGHT:-}"; do
[[ "$v" =~ ^-?[0-9]+$ ]] || { echo " $name: unreadable geometry, not clicking"; return 1; }
done
cx=$((X + WIDTH - 11)); cy=$((Y - 31)) # openbox decoration: close box above the frame
if [ "$cx" -lt 0 ] || [ "$cy" -lt 0 ] || [ "$cx" -ge "$screen_w" ] || [ "$cy" -ge "$screen_h" ]; then
echo " $name: close box at ${cx},${cy} is off-screen, not clicking"; return 1
fi
echo " $name: closing via titlebar at ${cx},${cy}"
xdotool mousemove "$cx" "$cy" click 1
sleep 2
return 0
}
for name in "Setup Wizard" "New version"; do
for _ in 1 2 3; do close_dialog "$name" || break; done
done
# --- main window ----------------------------------------------------------------------------
main="$(xdotool search --name "Untitled" 2>/dev/null | head -1 || true)"
if [ -n "$main" ]; then
xdotool windowmove "$main" 0 0 windowsize "$main" "$screen_w" "$screen_h" 2>/dev/null || true
xdotool windowactivate "$main" 2>/dev/null || true
sleep 2
fi
echo "--- session ---"
status