First release, then four more before midnight. Licensing runs on a Cloudflare Worker with D1 and proof-of-work, wheels ship compiled for CPython 3.13 and 3.14, and humanio login connects a subscription to the package.
OMNISTYLE landed: new identity layer, upgraded 3D scene, and the same hardening pass everything else got. CSPRNG ids, fail-closed webhooks, unit tests for the Worker agents.
Moderation fails closed, risk scores are clamped, SSRF and image-bomb guards sit on every fetch, and drawtext strings get escaped before ffmpeg ever sees them.
Image, PDF, and video cropping are now fully keyboard-operable. Under the hood: hardened engine runtimes, WorkerPool regression tests, and a stricter CSP.
Through The Ropes holds donor data, so it got everything: role self-escalation blocked at the rules layer, the Printful webhook authenticated, donation receipts moved to the platform mail plane, and CI added.
Fixed in one pass: a reserve-number TOCTOU, a double-draw race, and a Stripe metadata truncation bug. The money paths now have their own test suite.
The kill-chain simulator ran a 25-scenario multi-hop, multi-exploit battery and passed. Paid tiers went behind a waitlist, the workspace went clippy-clean, and billing moved onto the platform.
Open engine, ML training, private web, one tree. The detector went NaN-safe, the model backend self-heals, and the training pipeline now fails loudly instead of shipping a bad model.
The cost accumulator was missing from migration 0001, so fractional metering could quietly under-charge. Fixed, with a regression suite so it stays fixed. This is why the SWARM days exist.
One /mail/send surface, Mailgun held server-side, ports in all three SDK languages, and the email-verification flow moved onto it the same day. Apps never touch a mail key again.
The non-LLM vulnerability scanner is out: 472 tests running fully offline, a 59-source vulnerability catalog, SARIF output, a VS Code extension, and a GitHub App. Zero model calls in the detection path, by design.
235 passing tests, a 138.9 µs full compile, 4.9x auto-parallel speedup, and a sandbox escape fixed. The benchmarks doc also records where it loses: the interpreter is ~432x slower than native on tight numeric recursion. Both numbers are real.
Last reviewed · Site changelog