Skip to content

v0.29.0 — R23: the AI overhaul round

Choose a tag to compare

View all tags
@skymanbp skymanbp released this 17 Aug 21:17
· 169 commits to main since this release
v0.29.0
6910449

R23 of the seventeen-item feedback triage (round 2 of 3). Full ledger: docs/ROADMAP.md 当前状态 first entry.

What's new

  • The model knows every tool (#12): a compiler-enforced control registry drives the schema, the prompt catalogue and the eval ruler from one source. HSL replies with short axes are repaired instead of discarded. Local masks gain hue rotation and signed sharpness.
  • Style library, visible (#6): build/inspect it from the AI panel; the proposal disclosure now names which library and which neighbour shots fed it, and survives judge revisions.
  • Strength slider (#5): one 0-1 dial through all six restraint gates. Default 0.65; 0.5 restores the calibrated guardrail NUMBERS bit-for-bit (the old restraint wording is the ≤0.4 band — no single setting reproduces an old release in full). White-point protection is now a floor the soft caps can never shave, at any strength.
  • Deep thinking (#13): scene → per-family tool plan → intended look → recipe → self-critique, with multi-round visual review (2 rounds balanced / 3 committed). Off = request bytes identical to v0.28.0. Batch and eval never iterate.
  • Reverse-fit overhaul (#3/#16): honest confidence (one calibrated ladder, named constants), any same-frame rendition as reference, and a second independent reading over 8 joint value-range buckets that finally exposes the 'nonsense fit' in numbers (report + confidence cap + one fail-open terminal veto). Opt-in deep fit judges BEFORE saving and may buy one guided retry from a closed action list — the reviewer's text can never set a slider.

Behaviour notes

  • Recipes whose masks use the new hue/sharpness fields are rejected by older exes (forward-compat contract; mask-free recipes are byte-identical).
  • Deleted-version snapshots stay deleted: a schema-drifted registry arm re-archives once under NEW numbers, never resurrecting an old one.
  • Strength above 70% raises the visual-review ceiling to 3 rounds (worst case 17 API calls) — same ceiling as Deep thinking, either one alone does it; every tooltip says so.
  • Eval scores are not comparable across R23 (the proposer prose changed at every strength); the 147-photo re-baseline is pending.

Gates: clippy 0 (both configs); 518 lib + 8 CLI + 99 GUI + 2 contract tests green (both configs); audit_i18n 0.

Assets 4
Loading