Written by an AI agent working in this repository. Shed Designer is not engineered or code-checked; its structure checks are screening, not a sign-off.
Shed Designer frames a gable shed board by board. One of its tests, fuzz.test.mjs, builds random designs and asserts what every design must keep. Every member must have a real length and every shopping line a price. The DXF must parse. Nothing may float or clash. Every connection check must pass. The random designs include odd sizes, junk names and windows hanging off the end of a wall.
On each pull request it builds 150 designs with seed 1. That keeps the suite fast, but it also means every pull request checks the same 150 designs. Bugs outside that set stay hidden until someone runs a wider sweep by hand, which round 1 did a few times.
What changed
shed-designer-fuzz.yml runs each night at 05:47 UTC. It builds 2,000 designs, using the number of days since 1970 as the seed, so each night covers different designs. If a night goes red, the job summary prints the command that rebuilds the same designs:
cd ventures/shed-designer && FUZZ_N=2000 FUZZ_SEED=<seed> node --test fuzz.test.mjsThe failing designs are uploaded as an artifact. The job is read-only: it pushes nothing and calls no API. A 2,000-design sweep took 112 seconds on a 4-core container.
The first seed failed
Tonight's seed is 20756. Before the job ever ran, I ran the sweep locally with that seed: 1 of 2,000 designs broke an invariant. The design is a 19 ft by 15 ft 6 in shed allowed only 8 ft boards, with a 64 in window on the front wall at 114 in. The front wall's double top plate got joints that the check rejects:
plate splices: 2 of 5 plate joints on a stud, cap joints at least 24" from top-plate joints (IRC R602.3.2): cap joint at 111-3/4"; cap joint at 192"
Without the window, the same wall passes. So the splice placement doesn't yet handle a long header when only short boards are allowed. The connection check caught the bad layout, so the plan doesn't claim a splice the code would reject. Still, the layout search should have found joints that pass, or said why it couldn't. The fix belongs to the model, and it has been handed to the product side as a pinned case to add.
None of the 150 designs on the pull-request seed hit this case. That gap is why the nightly sweep exists.
Test time
I timed the default suite to see if it could be cut. Its 133 tests run in 20 seconds of wall time on four cores. The slowest are the fuzz test (6.5 s and 3.7 s) and the simplicity and budget searches (4.7 s and 2.4 s). Each slow test is slow because it searches over many designs, and that search is what it tests. Shrinking them would trade coverage for seconds, so I left them alone. The wide sweep moved to the nightly job, not into the pull-request suite.