v1.15.0

v1.15.0 (2026-09-30)

This release is published under the MIT License.

Bug Fixes

• deps: Upgrade locked anyio, PyJWT, urllib3 and virtualenv past advisories (20eb8ae)

• optimize: Accept every metric protocol everywhere; hold the test split out of selection (868755f)

• optimize: Data augmentation actually adds usable, leak-free rows (7a46ae7)

• optimize: Judge scores and metric aggregation mean what they say (950baa3)

• studio: Chat, debug, tasks, pipelines and feedback work with template agents (8b45d7d)

• studio,tasks: Make template agents chat, thread and run as tasks out of the box (8cc6ac0)

Chores

• Ignore ModelCheckpoint's default output directory (921a97d)

• Sync uv.lock with the 1.14.0 project version (9b88a71)

• studio: Rebuild the Studio UI 0.4.1 bundle with patched dependencies (1428e2d)

• studio: Ship the Studio UI 0.4.1 build from source (b5b60d1)

Continuous Integration

• Keep pre-commit fixers off the synced Studio bundle (54ce4a0)

• sync-studio: Do not fail the sync when PR labels are missing (69e1636)

Documentation

• Adding prompt optimization low level example (e94084a)

• Adding report generation (9c45d4e)

• Cleaning tmp files (f82af42)

• Improving examples (93863d3)

• optimization: A dedicated section covering the whole mechanism (dcc74af)

Features

• examples: Working prompt-optimization examples, low to high level (7d04c78)

• optimize: An anti-overfitting mechanism for prompt optimization (a39b603)

• optimize: Fit reports show what changed and whether it helped (735f9c5)

---

Detailed Changes: v1.14.0...v1.15.0