LettuceAI/app desktop-dev-196-1-67553e2

Desktop Dev Build #196.1 (67553e2)

on GitHub

Automated desktop build artifacts for commit 67553e2fdc03803e07ea6e0ab4043dece06cfa15.
Workflow run: https://github.com/LettuceAI/app/actions/runs/31650317025

Changes since previous Desktop build desktop-dev-195-1-5029153

Compare: desktop-dev-195-1-5029153...67553e2

• 67553e2 docs(changelog): add the 2.2.0 release entry
• 41c4e9b feat(group-chats): add per-character model and system prompt overrides
• 22b5267 fix: align companion timelines with the session clock
• 4e49f5f i18n: translate the model duplicate and safetensors format strings
• e97bf0b feat(image-gen): accept safetensors in the sd.cpp model selector and repo search
• 0b8d722 fix(widgets): hide companion-only widgets in roleplay and group chats
• 117cd58 fix(chats): show total session message counts in widgets
• 356a975 feat(models): add a duplicate action to the model menu
• 37d473e fix(navigation): preserve settings return path through playground
• a679cf5 fix(images): honor model storage and playground exit
• 145915f fix(images): stop sd.cpp when Lettuce exits
• 04e7b6f fix(prompts): drop the unused scene protocol predicate
• 052e9eb fix(images): treat unset scene generation as disabled
• 5c4e2a8 fix(prompts): gate each scene image protocol on its own model kind
• b69837c fix(storage): repair group chat schema drift
• 9b34d2a feat(images): support literouter and custom openai-format image providers
• 90b97ae feat(llama): measure the device compute reserve instead of guessing it
• a328ebb fix(llama): report kv for offloaded blocks only, add a real-model plan probe
• 2f2d672 fix(llama): derive the mtp draft reserve from real weights and kv
• b688f7f fix(llama): reserve attention scratch for query heads not kv heads
• 766f8ea fix(llama): distribute multi-gpu layers using real per-unit costs
• fb20edd fix(llama): subtract resident weights when recommending a vram context
• 89e603f fix(llama): size kv from per-layer geometry including sliding windows
• a03be64 style(llama): drop inline comments from the offload changes
• 072ada8 fix(llama): derive kv size from declared head dims and ggml block sizes
• dc2fe98 fix(llama): price gpu offload from real per-unit tensor sizes
• 47f5c75 fix(llama): load and fit bundled mtp tensors for embedded mtp models
• 1257c8d chore(llama): bump llama-cpp-rs to b10327 and opt into mtp tensor loading
• a6c99cb fix(ollama): normalize stray system messages to satisfy strict templates (#85)
• 04b97c7 fix(ci): require Ada and Blackwell CUDA kernels
• 1a97db2 fix(companion): preserve shared memory when creating chats
• d444236 fix(nanogpt): show exhausted quota as fully used
• 04c8e48 feat(companion): preserve canonical message timeline
• b7f2def feat(companion): add evidence-based continuity across chat episodes
• 3e1ef0c feat(companion): preserve soul and relationship continuity across chats
• 4350526 fix(companion): make message timestamps unmistakably system-authored
• 53e8426 feat(sync): replace legacy replication with transactional sync v2 (#84)
• 6a5eca7 fix(sync): reject mismatched builds and repair stale state