Release 2.3.0

🏅 Highlights

• Python 3.10 is no longer supported. The core package now requires a newer Python version.
• Bulk operations in the resource store and CLI. The resource store can check identifiers and update resources in bulk. ado edit builds on this to patch many resources at once.
• Clearer feedback from the CLI and core. Space statistics are now correct and clearly named. Measurement space comparisons are shown with --matching-space. Inconsistent spaces and wrongly typed property values produce readable errors.
• Ray Tune defaults and stopping. Optuna TPE is now the default sampler, and a new repeated-sample stopper is available.

✨ Features

• CLI enhancements
  • --matching-space now displays a comparison of the measurement spaces.
  • ado edit supports bulk patching of resources.
  • Space inconsistency errors are handled gracefully instead of surfacing as raw failures.
• Core improvements
  • The resource store gains a bulk update method.
  • The resource store gains a bulk check for whether identifiers exist.
  • Passing the wrong value type to a PropertyValue now gives a clearer error message.
• Ray Tune
  • A repeated sample stopper ends an optimization run once it keeps producing samples that have already been seen.

🐛 Fixes

• CLI
  • Space statistics are now correct and have clear names.
• Core
  • ray_options are now applied correctly when custom experiments are dispatched as remote tasks.
  • The random walk selector now raises an error when there are not enough entities, instead of continuing.
  • Properties collections are sorted by identifier, so their ordering is deterministic.

🧰 Build

• Python support: Python 3.10 is no longer supported for the core package.
• Dependencies: AutoGluon was updated to 1.6.3.

📝 Docs

• Agent skills
  • Skills are renamed to follow the ado-{verb}-{name} pattern.
  • A new skill helps with choosing a sampling strategy.
  • The remote-execution skill now autodetects the remote Ray version.
  • Basic skills are loaded reliably, with clarifications added.
• Security: The security policy was updated.

🧹 Refactoring

• CLI: The RESULTS statistics columns are renamed to MEASUREMENTS.
• Core
  • The SQL resource store now uses the SQLAlchemy ORM and table reflection.
  • Only the provenance information appropriate to the operation type is recorded.
• Ray Tune: The default sampler is now Optuna TPE.
• Testing: The mock actuator was removed and replaced with a test actuator.
• No changes in this release are flagged as breaking.

---

⚙️ Autoconf 2.1.2

• Throughput recommender: A new recommender suggests configurations based on throughput.
• Python support: Python 3.10 is no longer supported.

🧠 SFTTrainer 1.8.11

• Python support: Python 3.10 is no longer supported.

📈 vLLM Performance 1.15.4

• Python support: Python 3.10 is no longer supported.

✂️ TRIM 2.2.4

• Python support: Python 3.10 is no longer supported.

🧮 CPLEX-MIP 1.1.1

• Python support: Python 3.10 is no longer supported.

🎛️ Profile Space 2.0.9

• Python support: Python 3.10 is no longer supported.

---

Full Changelog: 2.2.0...2.3.0