Highlights
- Improve source alignment for repeated mentions, with an option to retain the previous alignment algorithm (#485).
- Fix provider-plugin generation, generated schemas, and packaging (#513).
- Forward Gemini generation settings, including output-token limits (#509).
- Surface blocked, refused, or missing-text Gemini/OpenAI responses as inference errors (#534).
- Route OpenAI reasoning and GPT-3.5 model IDs correctly and discover reasoning-model API keys automatically (#494).
Compatibility notes
- Repeated-mention source spans may change. Set
resolver_params={"exact_alignment_algorithm": "difflib"}to retain the previous exact-alignment behavior. - Callers may now receive an inference error where an unusable model response previously appeared to succeed. Valid empty strings with normal completion remain allowed.
- GPT-3.5 requires
use_schema_constraints=False. Explicit output-token limits for OpenAI reasoning models remain a separate follow-up.
Full Changelog: v1.6.0...v1.7.0