Skip to content

v0.43.1 — First Impressions

Choose a tag to compare

@gizmax gizmax released this 22 Jul 18:15
· 91 commits to main since this release
0366063

A live-usage batch from running Sandcastle hard on a DGX Spark: what a new user sees first, and what a real workload breaks.

Changed

  • A fresh install starts with an empty Workflows page. The bundled sample workflows moved into the template hub - the Workflows page holds only what you create (your wizard-built workflow included), and the samples stay one click away in the hub. Three samples that already existed in the hub were deduplicated rather than moved.
  • The header Spark badge shows the detected GPU (e.g. NVIDIA GB10) via the new /api/runtime spark_gpu field.
  • Workflow health badges judge the last run first - a workflow whose latest run completed is no longer branded "Failing" forever by historical, since-fixed engine failures.

Fixed

  • Onboarding wizard's first run actually runs - correct input field name (was silently dropped), the workflow is saved so it appears on the Workflows page, and the queued run is polled to a terminal state instead of pretending instant success.
  • Large step outputs no longer vanish. aiodocker's follow stream yields chunks, not lines; multi-KB results fragmented and were silently dropped while small outputs worked. The docker backend now line-buffers with a trailing flush - verified with a 33KB FAQ generation that previously came back empty.
  • Sandbox health constructs the runtime with the configured backend - docker-backend deployments no longer report "Runtime is down" forever, and the insight card links to a page that exists.
  • Steps record the model that actually ran (after default/autoroute/rescue) - local-model runs no longer all claim "sonnet".
  • Czech and other diacritics render correctly in PDF exports (DejaVu ships in the image; latin-1 degradation only without a Unicode font), and the markdown renderer survives numbered lists of long links.

Every change was developed and verified live on GB10 hardware.


pip install sandcastle-ai==0.43.1