Skip to content

Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog and the project uses Semantic Versioning.

0.3.0 (2026-10-17)

Fixed

  • docs: copy aliases for Pages, drop old landing page, no pip bump PRs (aa996b9)
  • release: bump _version.py and CITATION.cff through release-please (ce80de9)

Documentation

  • carousel below the title at 560 px, posters ordered JEPA, DINOv3, echo, splats, representations, atlas, PointNet, VAE, CT/MR, ICP (eea9723)
  • contributor funnel - issue forms, templates, support, issues to create (5fbbc04)
  • documentation site, documented defaults, benchmarks, docstrings (6e5e057)
  • front-page README with poster carousel, docs pages, Pages landing, CITATION.cff (55e7d74)
  • launch kit, content calendar, final report (44920e9)
  • README as a landing page, twelve runnable examples, small-archive fast path (336fa2d)
  • relative image paths in the README (absolute raw URLs do not render on a private repository) (97c8c33)
  • SEO foundation - keyword map, GitHub metadata, PyPI metadata (fe4cd1f)
  • showcase of the ten pipelines; ci: pin ruff 0.17 (fa53d61)
  • stamp generated pages with the source's last commit date (0642797)

Infrastructure

  • scorecard-action 2.4.4, docs default version, branch protection option (1b50ee7)

Unreleased

Added

  • examples/: twelve runnable examples, one folder each, titled " on ", with the images they produce.
  • Remote archives up to 64 MB (PARDESLINE_WHOLE_ARCHIVE_MB) are downloaded once and indexed locally instead of one Range request per member: the first pl.load("msd_task04_hippocampus") goes from 194 s to a few seconds.
  • docs/_gen/datasets_table.py generates the README dataset table from the manifests.
  • Documentation site (MkDocs Material, mkdocs.yml, .github/workflows/docs.yml with mike versioning): one page per dataset (43) and per algorithm family (19) generated from the manifests and the catalog, model pages, the twenty skills as guides, 18 how-to pages titled as the questions people search for, "Why our defaults work", comparison, FAQ, API reference from the docstrings, llms.txt.
  • Every catalog parameter now carries rationale (why the default, when to change it) and every method validated_on (src/pardesline/catalog/data/_defaults.json); pl.describe() prints them; tests/test_catalog.py enforces full coverage.
  • benchmarks/: package defaults vs upstream defaults on the Stanford bunny (registration, Poisson reconstruction), tables generated into benchmarks/results.md.
  • tests/test_docs_snippets.py executes the code blocks of the README, the tutorial and the how-to pages (network marker).
  • Docstrings on every public function and class (31 added).
  • CI: type-check job (mypy), coverage uploaded to Codecov, native matrix on Python 3.10 and 3.13, a job that executes the README demo and the documented code blocks against the live datasets, a nightly workflow that installs from PyPI on three OSes and runs the demo, OpenSSF Scorecard, Dependabot, pre-commit config, release-please (version bump in _version.py, the plugin manifests and CITATION.cff, CHANGELOG section, tag and GitHub release), .zenodo.json for a DOI per release, docs/assets/make_social_preview.py.
  • Contributor funnel: issue forms (bug, feature, dataset request, algorithm or model request, real case), pull request template, SUPPORT.md, CODEOWNERS, .all-contributorsrc, templates/new_dataset, templates/new_algorithm, templates/new_model, templates/new_pipeline with tests, a README per example, docs/seo/issues-to-create.md (tracking issues, 16 good first issues, 11 help wanted) and docs/seo/apply_github.py to apply labels, metadata and issues through the API.

Changed

  • [local] accepts Open3D 0.20, which ships Python 3.13 wheels: pip install "pardesline[all]" now works on 3.10-3.13 (the native test suite passes on 3.13 with 0.20).
  • PyPI metadata: description, 26 keywords, classifiers and project URLs unified on the documentation site.

Fixed

  • Docs deployment: mike aliases are copies instead of symbolic links, which the GitHub Pages builder rejects; the old hand-written docs/index.html landing page is removed now that Pages serves the gh-pages branch.
  • Archive members whose names would escape the dataset cache (../evil.bin) are ignored at index time and refused by fetch.
  • .nii.gz files are accepted by every module that accepts .nii (Volume.segment() on a Decathlon volume raised "does not accept '.gz' files").

Changed

  • README rewritten as a short front page: animated poster carousel, the promise in three lines, install, a 30-second example, a Mermaid diagram of how the pieces fit, the showcase, the MCP tools. The detailed sections moved to docs/datasets.md, docs/catalog.md, docs/running-modes.md, docs/mcp.md (text unchanged).

Added

  • docs/showcase/: the ten pipeline posters with, for each, the snippet, the expert skill and the MCP tool; posters.gif (carousel), mosaic.jpg, social-preview.png for the repository card; docs/index.html, a landing page with a real carousel for GitHub Pages; CITATION.cff.

0.2.0 — 2026-10-14

First stable release of the open package: everything below is on top of 0.2.0a1.

Added

  • Eleven expert skills with reference scripts, served by the MCP server (list_skills, get_skill, pardesline://skill/{name}) and shipped in the wheel: 3-D representations, point-cloud registration (RANSAC + FPFH + ICP, CPD), multimodal rigid registration (SimpleITK Mattes MI), atlas label transfer, PointNet segmentation, DINOv3 dense features (Gram anchoring, frozen backbone + Faster R-CNN), JEPA pretraining (3-D ViT reference), VAE anomaly detection, Gaussian splatting (points -> 3DGS PLY), medical segmentation models (nnU-Net export) and 3-D evaluation metrics. The NumPy / Open3D / SimpleITK scripts run on the bundled datasets; the PyTorch ones are reference implementations.
  • Local MCP server (pardesline mcp, extra [mcp]): twelve tools (datasets, samples, scan / confirm, export, catalog, run, describe_object, to_mesh), three resources and a start_here prompt over stdio or streamable HTTP, for Claude, Gemini, Cursor, Codex and any MCP client, without an account. Questions come back as JSON for the agent to relay; the repo ships a .mcp.json and the plugin declares the server. New skill mcp-server.
  • Medical 3D datasets, one line each. 27 new manifests: the ten Medical Segmentation Decathlon tasks, TotalSegmentator CT (117 structures) and MRI (56), LUNA16, IXI, VerSe 2019, SPIDER, CHAOS (DICOM), four TCIA collections (Pancreas-CT, LIDC-IDRI, NSCLC-Radiomics, Prostate-3T), OpenNeuro ds000228, KiTS23, AbdomenAtlas Mini and CT-RATE (gated), the BiomedSegFM 3-D validation set, MedShapeNet (106 101 STL shapes) and the C3VD endoscopy Gaussian splats - 43 manifests in all, each listed and sampled live.
  • New sources: s3 (public bucket, ListObjectsV2, pl.load("s3://...")), tcia (NBIA REST API, one DICOM series per entry, pl.load("tcia://Collection/CT")), urllist (a text index of URLs); archive sources may span several archives (urls).
  • DICOM series as folder entries (fetched whole), MetaImage .raw siblings fetched with their header, repeated pattern placeholders ({id}/anat/{id}_T1w.nii.gz) matched as backreferences.
  • Hugging Face datasets, for real. pl.load("hf://org/name[/sub/folder]") loads a bundled manifest or, for any other repo, lists the folder, drafts a manifest and asks what it cannot establish (licence from the Hub card, units, labels, array layouts) - the same four asking modes as a private folder. Listings go per folder (source.paths), are cached on disk per commit and capped (source.max_files), with a clear message instead of an HTTP 429.
  • Remote archives read by HTTP Range (source.scheme: "archive"): zip (incl. zip64) and uncompressed tar on S3 / Zenodo are indexed for a few KB and read member by member; whole download with checksum when the server ignores ranges.
  • Raw arrays through a layout: .npy / .npz / .bin roles declare layout (dtype, shape, columns, key, spacing); pl.load("sweep.bin", layout=...), RawArray.as_points() / as_volume(); decoded once into PLY / MetaImage so the 73 algorithms run unchanged; wrong layouts fail loudly.
  • Nine manifests: msd_task02_heart, amos22, kitti_raw_2011_09_26_drive_0001, hf_openmind, hf_nsclc_radiomics, hf_objaverse_glb_shard, hf_3d_arena, hf_uniocc_carla_mini, hf_kitti360_gs_subset, each checked against the live source.
  • Dataset.export(folder): a standard copy (PLY / MetaImage / NIfTI, one sidecar per file, a local manifest) of any dataset or subset.
  • CLI: pardesline datasets, describe <dataset>, fetch, export.
  • https files accept an upstream md5 and an archive_root; pattern placeholders other than {id} join into the sample id and become fields (ds.filter(split="Tr")).

0.2.0a1 — 2026-10-10

First open-source release of the package. Licence: Apache-2.0.

Added

  • Local execution of the 73 algorithms that power the PardesLine platform (19 modules: point clouds, meshes, surface reconstruction, registration, deformable registration, segmentation, SDF, LiDAR, stereo, collision, PyVista analysis and volume processing, SimpleITK registration and segmentation). Every algorithm runs on your machine with pl.run(module, method, input, **params) or as a CLI module.
  • Typed catalog (pl.catalog) generated from the platform's module manifests: 371 parameters with type, default, range, choices and description. Parameters are validated and coerced before a run.
  • pl.Object: a local file wrapped with its kind (point cloud, mesh, volume, transform, image), its sidecar metadata (registration transforms) and the provenance of every run that produced it.
  • pl.load(path) for local files; obj.run(...) and obj.methods().
  • pardesline CLI: list, describe, run, plus the cloud commands.

  • pardesline.medical: the SimpleITK service (volume I/O, DICOM series, rigid / affine / B-spline / demons registration, segmentation).

  • Object model: PointCloud, Mesh, Volume, LabelMap, Transform, Image with frame (axes, units, affine, source), labels, license and provenance; with_frame / with_labels / with_license; save() writes a .pardesline.json sidecar that load() reads back. Volume headers give LPS/mm and the affine; glTF is Y-up/m by specification; other formats stay unknown until declared - never guessed.

  • Conversions: Volume.segment() -> LabelMap; LabelMap.to_mesh(label) exact isosurface in world coordinates (in-process, direction cosines applied); Volume.to_mesh(isovalue); Mesh.to_points(); PointCloud.to_mesh(); align() and clean() on clouds and volumes.
  • Datasets: manifest schema (source, licence, citation, roles, samples, questions); Dataset / Sample with lazy fetch, split() by group, filter, take, sample; sources local, https (SHA-512, mirrors, tar/zip archives, cache), hf:// ([hf] extra) and Open3D demo data.
  • Bundled manifests: rire_brain, popi_lung_4dct, liver_tumor_patient01, visible_human_head_mri, open3d_bunny, open3d_armadillo, open3d_icp_demo.

  • pl.scan(folder) drafts a manifest for a private folder and lists its open questions; pl.load(folder) raises NeedsConfirmation until they are answered (in the call, in the saved draft, interactively, or assume=True).

  • Skills for AI agents under skills/ (eight Claude Code skills) with a plugin manifest, so /plugin marketplace add pardesline/pardesline teaches an agent the package; tests/test_skills.py keeps them consistent with the catalog, the datasets and the public API.

Changed

  • The proprietary cloud client 0.1.0 is now the pardesline.cloud subpackage, re-licensed under Apache-2.0. The top-level helpers (pl.login, pl.upload, pl.share, pl.o3d, ...) keep working.

Not yet in this release

  • Manifests for the large public hubs (OpenMind, MSD, Objaverse, KITTI).
  • Splats as a first-class kind.
  • The local MCP server.