v0.9.3

0.9.3 (2026-05-06)

Bug Fixes

• core: tune provider generation and drift coverage (#170) (6d37ad3)

Notes

• Provider calls now use shared generation profiles instead of raw stage-level token counts. Each provider maps the same intent to its own output budget and reasoning settings.
• OpenAI now uses the Responses API with gpt-5-mini and gpt-5 as the default fast and advanced models. Anthropic defaults were updated to claude-haiku-4-5 and claude-sonnet-4-6.
• Gemini and OpenAI now tune reasoning by phase. Extraction stays cheap, model updates get more room, and drift analysis keeps the stronger model without spending extra reasoning by default.
• Drift analysis now has to account for every added dependency and includes code evidence in the dependency section. This makes missed relationships and new component links easier to spot in review.
• Provider truncation errors now name the output budget that ran out, so failures are easier to diagnose and retry with a smaller change.

Build System

• deps-dev: bump @commitlint/config-conventional from 20.5.0 to 20.5.3 (#165) (0857f24)
• deps-dev: bump eslint from 10.2.1 to 10.3.0 (#167) (8175c0f)
• deps: bump @anthropic-ai/sdk from 0.91.1 to 0.92.0 (#164) (b903b1e)
• deps: bump @google/genai from 1.50.1 to 1.51.0 (#168) (18d8a4e)
• deps: bump likec4 from 1.54.0 to 1.56.0 (#166) (33e5433)