# Changelog

## Baseline

- First survey — no prior run to compare against. CAI 52.6.

## Lenses

- Code Health 62.1
- Architecture 91.2
- Maturity 65.0
- Readiness 47.8
- Security 60.9
- Domain Modelling 54.3
- Accessibility 54.5
- Performance 60.0

## Changes since last survey

- 100 commits — 70 feature/other, 30 fixes

## By area

- app/src — 71 commits
- (root) — 12 commits
- (repo) — 11 commits
- .github/workflows — 2 commits
- app/build.gradle.kts — 2 commits
- feature_llm_armv9/src — 2 commits

## Notable commits

- fix: Added shimeji mode to the pet, changed farm style. fixed FGS error, fixed lingbot, wan errors when decoding video
- fix: Added support for Stable diffusion distributed inference, fixed model downloading stale downloads and half downloaded files taking up space. Added initial experimental support for GPU acceleration in llamacpp and sdcpp. Added backend, backend params and max VRAM to sdcpp UIs to allow sequantial loading of the models. Tried to fix issue can't enable GNU RELRO protection
- fix: Added support for voice clonning and TTS through llamacpp and audio and music generation with LiteRT, fixed repo navigation bug
- fix: Cleaning code, fixed codeql
- fix: Easy chat mode + fixes for AI agent + improved pet simulation
- fix: Final fixes for the release
- fix: Final redesign. Fixed some issues in chat llama native delegating work to the harness and passing files from the main chat
- fix: Fix for Qwen Image + Minor deepseek ai agent fixes
- fix: Fix for multiple runtimes not showing ADT models correctly
- fix: Fix to AI agents implementation plan buttons not working
- fix: Fixed app building up trash and getting laggy, specially after heavy AI usage, Added in-paint to manga/comic translation to avoid complete OCR. Fixed auto detection of sdcpp model types
- fix: Fixed aproval button error in AI Agent, added calculation of actual context budget when using parallel >1
- fix: Fixed audio gen bugs and model type override by users
- fix: Fixed audio generation error and minor bugs in the download queue
- fix: Fixed error OCR ggufs not working
- fix: Fixed error on the finish task tool in AI Agent
- fix: Fixed error with VAD 5.1.2 download, added a curated list of bundles to download for using sdcpp
- fix: Fixed max context error where the harness allowed prompts bigger than max ctx
- fix: Fixed some chat llama native errors when handling the deepseek harness. Removed the pet simulated world. Fixed an issue with lingbot video generation. Updated binaries and in-app guide
- fix: Fixed upscale compose error, fixed PDF intake
- …and 80 more

## Architecture

- 0 containers · 3 bounded contexts · 3 dependency edges (baseline)
