Skill v0.0.116

Changelog

Features

• Minor updates and improvements

Contributors

• Eugene Hauptmann

Skill v0.0.115

Changelog

Features

• fixed a bug with the labels, libomp, cpu/gpu backend choice for llm

Contributors

• Eugene Hauptmann

Skill v0.0.114

Changelog

Features

• migrated to nomic-ai/nomic-embed-text-v1.5 model for text and image embeddings
• cleanup
• Task 8 — Removed mirror state (~150 lines)
• Task 9 — Removed device upsert helpers (~90 lines)
• Task 10 — Removed Cortex WS state (~55 lines)
• Task 11 — Refactored settings persistence
• Task 12 — Removed calibration profile local cache

Contributors

• Eugene Hauptmann

Skill v0.0.113

Changelog

Features

• updated CI
• Daemon side (business logic added)
• crates/skill-daemon/src/background.rs:
• crates/skill-daemon/src/handlers.rs:
• crates/skill-daemon/src/routes/settings_calibration.rs:
• Tauri side (logic removed)
• src-tauri/src/background.rs:
• src-tauri/src/lifecycle.rs:
• Moved to daemon / removed from Tauri
• 1. detect_device_kind() — Added DeviceKind::from_id_and_name() to skill-data/src/device.rs with the ID-prefix logic (cortex:, usb:, cgx:, etc.) that was missing from the existing from_name(). Removed the dead function + 250

Contributors

• Eugene Hauptmann

Skill v0.0.111

Changelog

Features

• updated DMG version
• fixed the build
• llama linking fixed

Contributors

• Eugene Hauptmann

Skill v0.0.110

Changelog

Features

• feat: add CPU backend support to skill-router for CI coverage
• Add a new feature to skill-router that uses CPU-based UMAP
• Changes:
• The GPU backend remains the default for normal builds, while the
• fix: disable all GPU-dependent features in coverage workflow
• The coverage CI was still failing because the embedding features
• Solution: Disable all embedding features in the coverage workflow
• The embedding functionality will still be tested in local
• fix: add llm-native feature and use it in coverage workflow
• The coverage CI was failing because is not a valid
• Solution:
• This ensures that llama-cpp-4 is statically linked in the coverage
• fix: add native feature to llm in coverage workflow
• The coverage CI was failing because it couldn't find libllama.so.0.
• Solution: Change to in the coverage workflow's
• This matches the Cargo.toml configuration where we added the native
• fix: remove aggressive -static flag from macOS config
• The -static flag in the macOS cargo config was causing issues and
• Instead, we rely on:
• This change ensures that:
• Verified that skill-daemon runs without libllama.0.dylib present.
• chore: enable static linking for llama-cpp-4
• Add the feature to llama-cpp-4 dependencies to enable static
• This change:
• Static linking improves:
• The binary size will increase, but this is acceptable for a desktop
• chore: update llama-cpp-4 to 0.2.43
• Update llama-cpp-4 dependency from version 0.2.42 to 0.2.43.
• This brings in the latest improvements and bug fixes from the
• The update is applied to all feature variants (ggml, metal, vulkan).
• fix: disable GPU features in coverage CI
• The coverage CI was failing because tests that use GPU acceleration
• Solution: Run coverage tests with CPU-only features by:
• This allows the coverage workflow to complete successfully while still
• revert: remove CI checks from tests
• The previous approach of checking CI environment inside tests didn't work
• This commit reverts the CI checks and we'll try a different approach
• fix: use CPU-only backend in CI tests
• The coverage CI was still failing because even though we added CI checks,
• Solution: Modify test_state() to use the LUNA CPU-only backend instead of
• LUNA is a topology-agnostic encoder that runs entirely on CPU, making it
• fix: skip GPU tests in CI environment
• Solution: Add a check for the CI environment variable at the beginning
• Tests affected:

Contributors

• Eugene Hauptmann

Skill v0.0.109

Changelog

Features

• fix: disable dynamic linking for llama-cpp-4
• The llama-cpp-4 crate doesn't have a 'static' feature. Instead, it has
• To get static linking, we need to:
• This will embed the llama.cpp libraries directly into the skill-daemon
• Note: We keep the 'ggml' feature explicitly enabled since it's not in
• fix: force static linking for llama-cpp-4 in skill-daemon
• The skill-daemon binary was failing to launch on macOS with:
• This occurred because llama-cpp-4 was dynamically linking to its
• Solution: Configure llama-cpp-4 to use static linking by adding
• Changes:
• This makes the skill-daemon binary larger but more portable and

Contributors

• Eugene Hauptmann

Skill v0.0.106

Changelog

Features

• fix: add APPLE_SIGNING_IDENTITY to DMG creation step
• The DMG creation step was missing the APPLE_SIGNING_IDENTITY environment
• The script already supports using APPLE_SIGNING_IDENTITY via the SIGN_ID

Contributors

• Eugene Hauptmann

Skill v0.0.104

Changelog

Features

• feat: implement cross-platform daemon update hooks
• • Add pre-update and post-update hooks for Tauri updater
• The daemon is now properly stopped before updates and restarted afterward,
• Fix Windows release: check both target paths for skill.exe
• The release workflow was failing because the 'Log binary dependencies' step
• This change makes the step check both locations (target-specific first, then

Contributors

• Eugene Hauptmann

Skill v0.0.103

Changelog

Features

• chore(bump): improve release notes generation from git history
• Updates the bump script to generate better release notes from commit history:
• The script now always generates release notes based on actual commit history
• feat(llm): add Nemotron-3-Nano-4B model to catalog
• Adds NVIDIA's Nemotron-3-Nano-4B model with the available Q4_K_M quant.
• Also simplifies pre-commit hook to run only basic validation,
• chore: update llama-cpp-4 from 0.2.36 to 0.2.38
• Updates llama-cpp-4 and llama-cpp-sys-4 to latest versions.
• bump
• fix(calibration): improve daemon integration and error handling
• • Fix EEG window submission to daemon with proper eeg_start/eeg_end timestamps
• This ensures calibration labels are properly recorded with correct EEG windows

Contributors

• Eugene Hauptmann