Skip to content

Changelog

What changed in the API, the package and this site, newest first. A feature of the package has shipped when a published version carries it. Install and prepare a slide names the version these pages describe, and says when it is not published yet.

October 2026: the API

  • A quota refusal's Retry-After is the seconds until its counter's window resets. A request over a whole window's limit is told to send less.
  • The schema describes every refusal as the Problem document, and names each operation's credential and scope.

September 2026: the API

  • A job is created at POST /v1/jobs. POST /v1/predictions takes the same request and answers the same way, and is deprecated.
  • A model's card carries output.transform, the scale a result's values are on. What you get back says where to read it.

0.1.0.dev3, October 2026: the jobs endpoint and clearer messages

  • The API client creates jobs at POST /v1/jobs.
  • Opening a slide reads the scale more slides record, including units written by ImageJ.
  • Messages and help text name the model DeepSpot-M and say what to do next.

0.1.0.dev2, September 2026: the package reads real slides

  • Opening a slide reads the JPEG-compressed slides most scanners write, and warns when a slide's layout will make reading slow.
  • Packing tiles: the pre-flight checks accept every archive form, and compare a declared scale with the slide's own.
  • Embedding tiles locally is faster, with identical output.
  • The command line: doctor --tiles N times a slide of N tiles, and doctor --allow-download fetches missing encoder weights.
  • The API client refuses, before uploading, an embeddings file the service would refuse.

0.1.0.dev1, September 2026: the package beyond the core

  • Opening a slide: open a whole-slide image, and read its resolution from the file.
  • The API client: submit a prediction, follow the job with a wait() that honours Retry-After, resume an interrupted upload, download the result.
  • The command line over the same client, with --json on every command, and the same client exposed as MCP tools for an agent, installed by the mcp extra.
  • Local embedding: run DeepSpot-H over your tiles and write the embeddings container. The embed extra.
  • Gene identity: names resolve to Ensembl identifiers through the service, and a downloaded result names its own genes offline.
  • After prediction: what a model's card asks to be applied to a result file, applied the same way the service does. A gene the model never measured is blank in the result, not zero.
  • Add a bulk RNA profile. Add the sample's bulk RNA profile to the prediction for its slide, and keep the same genes and file format. Without a bulk profile, Aurora predicts from the slide alone.
  • Spatial structure: a per-gene measure of whether a gene varies with place, from a result and its coordinates.
  • What the service runs: the list of models from the public cards, a pre-flight tile-size check, and a predict_local that refuses and states where prediction runs.

September 2026: the API opened

  • The hosted API accepts work, and access is granted per account. Get access is the flow.

0.1.0.dev0, September 2026: the core

  • Tile quality control, the tile packer and archive validator, the result writer, and spatial subsampling, all on your own machine.

September 2026: this documentation site