Release 2.2.0

🏅 Highlights

• Richer operation visibility — Ray job submission IDs are now recorded in the operation resource, experiment and actuator details are logged and stored, and the live progress table no longer crashes the operation on display errors. Together these significantly improve observability of remote runs.
• CLI improvements across the board — New ado show measurements subcommand for samplestores, --from-experiment filter on measurement display, --include-properties on ado get space, and cascade-delete of datacontainers when deleting an operation make day-to-day workflows smoother.
• Correctness fixes for property domains — NumPy scalars and generic numeric types are now handled correctly in valueInDomain and PropertyDomain, eliminating subtle type errors that could surface when processing measurement data.
• Performance boosts in core — Value-in-domain checks now use arithmetic comparisons, and the YAML stack now uses the C loader/dumper when available, both reducing overhead on large spaces.

✨ Features

• CLI enhancements — ado show measurements is now available for samplestores directly; the command accepts a new --from-experiment flag to filter measurements by experiment. ado get space gains --include-properties to include property values in the output. ado delete operation now cascade-deletes associated datacontainers automatically.
• Core improvements — Properties and domains now have a compact string representation for cleaner logging and display. Operations record the Ray job submission ID and associated metadata when jobs are submitted remotely, and full experiment and actuator details are stored alongside the operation for auditability. Input size checks on JSON columns prevent silently oversized payloads from causing hard-to-debug errors.
• Discovery space templates — Discoveryspace templates and the corresponding website documentation have been improved to better guide users in defining entity spaces.

🐛 Fixes

• CLI — The progress spinner no longer leaks into error messages. Edge-case handling when rendering time deltas (e.g. very short or zero durations) has been improved.
• Core — NumPy generic subtypes are now recognized in numeric checks for PropertyDomain. NumPy scalars are correctly handled in valueInDomain. Internal range values now return native Python types, preventing downstream serialization issues. The live table display no longer propagates its own errors to fail the running operation. Parameterized experiments now use the correct identifier in progress prints.

⚡ Performance

• Core — valueInDomain now uses arithmetic comparison instead of set membership, reducing overhead for large domains. The YAML loader and dumper now prefer the C-backed implementation when available, speeding up serialization of large spaces and results.

🧰 Build

• Dirty wheels — Wheels built from an untagged commit now embed a build timestamp, making it easier to distinguish development builds.

📝 Docs

• Agent skills — Multiple skill updates: remote-execution now clarifies how to upload Python packages; define-experiment-campaign clarifies experimentVersion field requirements; a new standalone run-experiment skill has been added; AGENTS.md now recommends pre-commit for linting; guidance on manually cloning git+ssh dependencies for fromSource has been added; and a tip on writing tool outputs to file to avoid truncation has been included.
• Website — A new section covering library skills has been added to the Getting Started guide. The Resources section has been moved under User Guide for better information architecture. The Videos and Demo page has been removed. An outdated-version warning banner has been added to the website.
• General — The README and ROADMAP have been updated to reflect current project direction.

🧹 Refactoring

• Core — The SampleStore now uses SQLAlchemy ORM constructs instead of raw SQL. Experiment identifier parsing is simplified to support only the common case.
• CLI — The default max hops for show related has been changed from a higher value to 1, making the output more focused by default.

---

⚙️ Autoconf 2.0.0

• Models are now built from data on Hugging Face, enabling a broader and more up-to-date model corpus for autoconf recommendations.

🧠 SFTTrainer 1.8.8

• The Ray runtime environment setup timeout has been disabled, preventing spurious failures during environment initialization on slow clusters.

📈 vLLM Performance 1.15.2

• gpu_memory_utilization now accepts values up to 0.98, unlocking higher GPU memory allocation for large model deployments. Geospatial dataset names in BenchmarkParameters have been restored. Deployments are now fully deleted before re-creation, preventing stale deployment conflicts.
• The minimum required vLLM version has been raised to 0.27.0.

✂️ TRIM 2.2.1

• Support has been added for missingTargetMeasurement on entities whose targetOutput has not yet been measured, enabling more flexible prior-free optimization. NoPriorsParameters.targetOutput has been removed — targetOutput is now specified exclusively at the top level. Early validation and resolution of targetOutput has been introduced to catch configuration errors before a run starts. Several correctness fixes: prior parameter names are now validated correctly, the final model is always saved even after the sampling budget is exhausted, and the iterative modeling entity count (numberEntities_iterative_modeling) in the TRIM operator is fixed.

🧮 CPLEX-MIP 1.1.0

• Invalid warm-start configurations are now handled gracefully, and the best bound found so far is returned even when the solver terminates early.

🎛️ Profile Space 2.0.7

• Python 3.14 is now supported.

---

Full Changelog: 2.1.0...2.2.0