CHANGELOG.md

# Changelog

All notable changes to NeuroSkill™ are documented here.
Pending changes live as fragments in [`changes/unreleased/`](changes/unreleased/).
Past releases are archived in [`changes/releases/`](changes/releases/).

---

## [Unreleased]

## [0.0.129] — 2026-04-24

### Features

- **Grayscale display mode**: optional system-wide grayscale display that activates and deactivates in sync with Do Not Disturb. Reduces visual distraction during deep work. macOS only, default off — the default focus action remains Do Not Disturb only.

### i18n

- **Grayscale translations**: added `dnd.grayscale` and `dnd.grayscaleDesc` keys to all 9 locales (en, de, es, fr, he, ja, ko, uk, zh).

### Dependencies

- **Fix rand 0.7.3 vulnerability**: vendored `phf_generator 0.8.0` locally with `rand` bumped from 0.7 to 0.8, eliminating the vulnerable `rand 0.7.3` transitive dependency pulled in by `selectors 0.24 → phf_codegen 0.8`.
- **Remove atty**: migrated `iroh_test_client` from `structopt` (clap v2) to `clap v4` derive, dropping the unmaintained `atty` crate.
- **Dismiss stale Dependabot alerts**: dismissed alerts for `crossbeam-deque`, `crossbeam-utils`, `crossbeam-queue`, `memoffset`, and `glib` — all already at patched versions or fixed in fork.

- **Update kittentts to 0.4.0**: bumped `kittentts` from 0.3.0 to 0.4.0 in both `skill-tts` and `src-tauri`.

- **Update llama-cpp-4 to 0.2.50**: bumped `llama-cpp-4` and `llama-cpp-sys-4` from 0.2.47 to 0.2.50, picking up upstream llama.cpp fixes including `common_*` symbol renames and `llama-common-base` static library linking.

- **Add grayscale crate**: added `grayscale 0.0.1` as a macOS-only dependency in `skill-data`.

## [0.0.128] — 2026-04-23

### Features

- llmfit
- glib
- notify
- linter
- updTed settings and engine
- muda
- fixed glib with a patch
- refactored UI
- Added unsloth/Qwen3.6-27B-GGUF

## [0.0.127] — 2026-04-22

### Features

- Add ANT Neuro eego amplifier family support via the `antneuro` crate (native USB backend, no vendor SDK required)
- Supported models: eego 8, eego 24, eego 64, eego mylab, eego sport, eego rt, eego hub
- Auto-detect channel count and sample rate from the amplifier on first data block
- Configurable cap layout via Device API settings (`antneuro_cap`: auto, 10-20, 10-10, or explicit channel count)
- Configurable sample rate via Device API settings (`antneuro_sample_rate`: 500, 512, 1000, 2048, etc.)
- Waveguard electrode name mappings for 8, 21, 24, 25, 32, and 64-channel cap configurations
- USB scanner auto-detects connected eego amplifiers every other tick
- Device images and company logo for all 7 amplifier models in the Supported Devices catalog

### Bugfixes

- Fix PPG and IMU sample counts not being tracked in daemon status (ppg_sample_count/imu_sample_count always stayed at 0)
- Fix `cmd_status` device block missing `firmware`, `ppg_samples`, and `imu_samples` fields (neuroskill CLI showed nulls)
- Fix `format_status_as_text` reading stale key names (`sample_count` instead of `eeg_samples`)

### UI

- Improved supported devices card grid: wider columns, taller images with `object-contain`, text wraps instead of truncating

## [0.0.126] — 2026-04-20

### Features

- daemon restart on launch

## [0.0.125] — 2026-04-20

### Features

- fixed auto connect

## [0.0.124] — 2026-04-20

### Features

- fixed daemon invoke
- linter
- fixed calibration
- fixed `set default`
- better connect
- fixed connectivity

## [0.0.123] — 2026-04-19

### Features

- Add daemon calibration session HTTP routes: `POST /v1/calibration/session/start`, `POST /v1/calibration/session/cancel`, `GET /v1/calibration/session/status`.
- Add `POST /v1/calibration/record-completed` and `GET /v1/calibration/active-profile` daemon routes.
- Daemon broadcasts `calibration-tts` events with spoken text before each phase, giving the frontend a 4-second window to play TTS audio before the countdown begins.

- Auto-install daemon as persistent OS service on first app launch. The Tauri app calls the daemon's `/service/install` endpoint after startup, registering a LaunchAgent (macOS), systemd user unit (Linux), or Windows service. The daemon now persists across app restarts and reboots.
- Add `--uninstall` CLI flag to skill-daemon for clean service removal.

- **LLM catalog**: Added Qwen3.6 35B-A3B (MoE), MiniMax M2.7, and GLM 5.1 models to the local LLM catalog.
- **HuggingFace GGUF search**: New UI section in Settings → LLM to search HuggingFace Hub for GGUF models. Browse repos by downloads/likes, expand to see individual quant files with sizes, and add models to the catalog with one click (optionally starting download immediately). Each repo shows a collapsible README (rendered as markdown with proper table and heading styling, YAML frontmatter stripped) and a direct link to the HuggingFace page. Two new daemon endpoints: `GET /llm/catalog/search` and `GET /llm/catalog/search/files` (the latter also returns the repo README, capped at 32 KB). Fully localized (9 languages).

- **Interactive search: AI summary with streaming.** "Generate" button in the Insights panel sends search context (labels, EEG metrics, sessions, screenshots) to the local LLM via `chat_completions_ipc` with a channel callback. Response streams token-by-token with live markdown rendering via `MarkdownRenderer`. Thinking blocks are shown in a collapsible `<details>` element.

- **Interactive search: rich AI prompt with full EEG context.** The LLM prompt now includes:
  - Time range and epoch count
  - Labels with timestamps and similarity distances
  - EEG epoch metrics: engagement, relaxation, SNR, α/β/θ band powers, heart rate, relevance score, session ID
  - On-demand timeseries fetch when graph nodes lack stored metrics (computes averages from raw data)
  - Session metrics with stddev, best-session flag
  - Screenshot context: app name, window title, timestamp, OCR similarity
  - Fallback session derivation from node session_ids when backend sessions are empty

- **Interactive search: prompt visibility.** Collapsible "Prompt" section shown immediately when "Generate" is clicked (open by default during generation, collapses after). Shows the exact data sent to the LLM in monospace font.

- **Interactive search: LLM auto-start.** When the LLM is not running, shows "Start LLM and try again" button that calls `start_llm_server`, waits for initialization, and retries the summary automatically.

- **Interactive search: Continue in Chat.** "Continue in Chat →" button after the AI summary creates a new chat session named "Search: {query}", saves the prompt and response as messages, and opens the chat window. Users can ask follow-up questions about their search results.

- **Interactive search: auto-save to chat history.** Every AI summary is automatically saved as a chat session, making all LLM conversations from search discoverable in the chat history regardless of whether the user clicks "Continue in Chat".

- **Interactive search: insights panel.** Collapsible "Insights & Patterns" card with:
  - Optimal conditions: auto-detected peak engagement hour and best app
  - App × Engagement correlation: bar chart showing avg engagement per app (uses `window_title` fallback)
  - Hour-of-day engagement pattern: bar chart across 24 hours
  - AI summary with streaming, thinking blocks, markdown rendering, and chat integration

- **Interactive search: bookmark/findings system.** Star button in node detail panel saves nodes to localStorage (max 50). Saved findings appear in the empty state with quick re-search and delete controls.

- **Interactive search: graph color mode switcher.** Dropdown to recolor EEG nodes by timestamp (turbo), engagement level, SNR quality, or session (hash-based). Graph reactivity fixed with explicit dependency tracking.

- **Interactive search: timeline scrubber.** Horizontal SVG timeline below the graph showing all timestamped nodes as color-coded clickable dots. Clicking selects the node in the detail panel.

- **Interactive search: EEG sparkline.** "Load EEG bands ±60s" button on EEG point nodes fetches actual timeseries data and renders α (blue), β (amber), θ (green) band power chart with a red marker at the epoch timestamp.

- **Interactive search: screenshot preview.** Inline screenshot thumbnail in the detail panel when a screenshot node is selected.

- **Interactive search: compare mode.** Select two EEG points for side-by-side metrics comparison with green/red diff highlighting and time gap display.

- **Interactive search: node detail panel improvements.** Moved to a separate card below the graph with colored left border, generous spacing, text-sm font sizes, responsive metrics grid, breadcrumb trail, and "More like this" / bookmark buttons.

- **Interactive search: 3D graph enhancements.** Double-click zoom with smooth camera tween, minimap, node kind filtering, subtle grid floor, reset view button, enhanced tooltips with metrics/session/relevance.

- Add device filter to EEG embedding search. A dropdown in the search controls lets users filter by device (e.g. MuseS-F921, AttentivU-053) or search across all devices. Backed by `GET /v1/search/devices` and `device_name` parameter on the streaming search endpoint.

- **Per-epoch EEG metrics in AI summary prompt.** The interactive search AI summary now includes per-epoch metrics (engagement, relaxation, SNR, α/β/θ band powers, FAA, mood, TAR, meditation, cognitive load, drowsiness, heart rate) instead of just "(no EEG metrics stored)". Each 5-second epoch shows its own brain state data.

- **On-the-fly CSV metrics fallback.** When embeddings lack `metrics_json`, the interactive search now lazily loads session `_metrics.csv` files and matches epochs by timestamp (±3s tolerance) to populate metrics from CSV data. This is transparent to the user — metrics appear without needing a backfill.

- **Background metrics backfill from CSV.** New `POST /v1/analysis/backfill-metrics` endpoint patches `metrics_json` in the SQLite embeddings table for rows where it is NULL, matching against `_metrics.csv` by timestamp. Fires automatically on search page load (fire-and-forget). After first run, all future searches find metrics directly in the DB.

- **Recomputation of derived EEG metrics.** When `meditation`, `cognitive_load`, `drowsiness`, or `stress_index` are null in stored `metrics_json` but band powers (α, β, θ, δ) are present, they are recomputed on-the-fly using the same formulas as the live pipeline.

- **Shared EEG score formulas (`skill-data/eeg_scores.rs`).** New module with pure-math functions for computing derived EEG metrics from averaged band powers.

- **Interactive search: EEG metrics on graph nodes.** Each EEG epoch node now carries its full metrics (engagement, relaxation, SNR, alpha/beta/theta, FAA, heart rate, etc.) via the `NeighborMetrics` struct, visible in the tooltip and detail panel.

- **Interactive search: session grouping & cross-session comparison.** Every node gets a `session_id` derived from its timestamp. The results include a session summary with averaged metrics, stddev, min/max engagement/SNR, duration, and band power ratios. The best session (highest avg engagement) is auto-flagged with a ★ marker.

- **Interactive search: cross-session trend chart.** SVG sparkline rendered above the session summary, showing engagement trend across sessions. Best session dot highlighted in green.

- **Interactive search: relevance scoring.** Each EEG node receives a composite `relevance_score` combining text similarity (50%), temporal distance (30%), and engagement (20%). Nodes are optionally sortable by relevance in the UI and rendered larger in the 3D graph for high-relevance matches.

- **Interactive search: SNR quality filter.** New `snrPositiveOnly` toggle (step 6) excludes EEG epochs with non-positive SNR (bad signal quality). Backed by server-side filtering before graph construction.

- **Interactive search: device filter.** Device dropdown (step 7) in the interactive pipeline, filtering EEG epochs by device name at the SQL query level via `get_session_timeseries_filtered()`.

- **Interactive search: date-range filter with presets.** Date-range inputs (step 8) constrain results to a time window. Quick preset buttons (24h, 7d, 30d) with a clear button.

- **Interactive search: EEG epoch ranking.** Rank-by dropdown (step 9) sorts EEG epochs by engagement, SNR, or relaxation before selecting top-k, surfacing the most interesting epochs first.

- **Interactive search: collapsible advanced filters.** Steps 6–9 are grouped behind an "Advanced filters" toggle to keep the default view clean. An amber dot indicator shows when any filter is active.

- **Interactive search: performance timing & system load.** Each search returns `embed_ms`, `graph_ms`, `total_ms` timing plus CPU usage and memory stats via `sysinfo`. Displayed in a compact perf bar.

- **Interactive search: search result caching.** LRU-style cache (8 entries, 5-minute TTL) in the daemon. Repeated searches with identical parameters return instantly without re-embedding.

- **Interactive search: CSV export.** "Export CSV" button in the sessions summary panel downloads a CSV with all session metrics.

- **Interactive search: search history.** Recent queries persisted in localStorage (max 10). Displayed as clickable chips in the empty state for quick re-runs. ArrowUp in empty textarea recalls the last query.

- **Interactive search: pipeline settings persistence.** All pipeline parameters (kText, kEeg, kLabels, reachMinutes, SNR, rankBy, advanced toggle) saved to localStorage and restored on page load.

- **Interactive search: 3D graph improvements.**
  - **Double-click to zoom:** smoothly flies the camera to the selected node with easeInOutQuad tweening.
  - **Minimap:** 80×80px SVG overlay in the bottom-right showing all node positions projected to 2D.
  - **Node kind filtering:** toggle checkboxes to hide/show EEG, Labels, and Screenshot nodes.
  - **Color mode switcher:** dropdown to recolor EEG nodes by timestamp, engagement, SNR, or session.
  - **Node sizing by relevance:** higher-relevance EEG nodes render 1.0–1.3× larger.
  - **Enhanced tooltips:** show relevance score, session ID, and full EEG metrics summary on hover.
  - **Edge kind labels:** selected node tooltip shows connected edge types (→ text_sim, ← eeg_bridge).
  - **Subtle grid floor:** transparent grid pattern at y=-15, adapts to dark/light theme.
  - **Reset view button:** "⌂ Reset" button + click-empty-space to fly camera back to default position.

- **Interactive search: node detail panel.**
  - Separate card below the graph with generous spacing and readable typography.
  - Colored left border matching node kind.
  - Breadcrumb trail showing the full path from query → text_label → eeg_point → found_label.
  - "More like this" button to re-search using the selected node's text.
  - Bookmark button to save interesting nodes as findings.
  - EEG sparkline: "Load EEG bands ±60s" fetches actual timeseries and renders α/β/θ band chart.
  - Screenshot preview: inline thumbnail for screenshot nodes.
  - Compare mode: select two EEG points for side-by-side metrics comparison with diff highlighting.

- **Interactive search: timeline scrubber.** Horizontal SVG timeline below the graph showing all timestamped nodes as color-coded dots. Click any dot to select that node.

- **Interactive search: insights & patterns panel.**
  - Auto-computed activity-engagement correlation: groups screenshots by app and shows avg engagement per app as a bar chart.
  - Hour-of-day engagement pattern: bar chart showing which hours have highest engagement.
  - Optimal conditions report: identifies peak engagement time and best app.
  - AI summary: sends search context to the local LLM for a natural-language analysis of patterns and recommendations.
  - Bookmark/findings system: save interesting nodes to localStorage, displayed in the empty state for future reference.

- **Interactive search: loading skeleton.** Graph-shaped SVG skeleton with animated nodes and edges replaces the simple spinner during search.

- **Interactive search: animated session bars.** Engagement and SNR progress bars in the session summary have smooth CSS transitions on load.

- **Interactive search: EEG epoch deduplication.** `seen_eeg_ts` HashSet prevents duplicate EEG nodes when multiple text labels have overlapping time windows.

- **Interactive search: parallel compare_search.** Both range queries in `/search/compare` now run concurrently via `tokio::join!`.

- **Interactive search: `get_session_timeseries_filtered()`.** New function in `skill-history` that filters timeseries by device name at the SQL level, used by the interactive search device filter.

- Add text embedding model selection. Settings → Embeddings now shows 16 fastembed models (nomic, BGE, MiniLM, E5, MxBAI, GTE) with dimensions and descriptions. Selecting a model downloads weights and hot-swaps the ONNX runtime in the daemon. Persisted to settings and loaded on startup. Backed by `GET/POST /v1/models/text-embedding` daemon API. Unknown model codes are rejected.

### Bugfixes

- Fix calibration data not being recorded because `record_calibration_completed` wrote to stale local state instead of the daemon.
- Fix `list_calibration_profiles` returning default profiles instead of daemon-authoritative data.

- Fix prebuilt llama archives missing `libmtmd` (multimodal library). The collect step in the prebuilt CI workflow now includes `libmtmd*` alongside `libllama*`/`libggml*`/`libcommon*`.
- Bump llama-cpp-4 from 0.2.45 to 0.2.46.
- Fix double daemon spawn on service install. The installer now checks `/service/status` before calling `/service/install`, and `install_launchagent` skips if the plist already exists with a matching binary path.
- Fix LaunchAgent label mismatch. Update hooks (`pre-update.cjs`, `post-update.cjs`) and uninstall script now use `com.skill.daemon` instead of stale `com.neuroskill.skill-daemon`.
- Remove stale `com.neuroskill.skill-daemon.plist` from app bundle Resources.
- Fix daemon logs written to `/tmp/` (wiped on reboot). LaunchAgent now logs to `~/Library/Logs/NeuroSkill/`.
- Add rollback binary as last-resort candidate in daemon path resolver.

- Fix EXG reembedding: resolve GPU f16 TypeMismatch bug by using f32 GPU encoder for batch reembed, fix per-segment channel metadata for mixed-device sessions, require full 5s epoch extraction, skip channel counts with consecutive failures, and process most recent days first.
- Fix reembed progress events not reaching UI (Tauri `listen` → daemon WebSocket `onDaemonEvent`), fix `"day"` → `"date"` field in progress payload, and emit `loading_encoder`/`scanning` status immediately for responsive feedback.
- Rebuild label EEG HNSW index after manual and idle reembed so interactive search can bridge text labels to EEG epochs.
- Add 32 unit tests for reembed edge cases: mixed devices, relative timestamps, partial rows, empty files, boundary extraction, and channel count handling.

- Fix daemon binary missing from DMG when prebuilt llama retry triggers on CI. The failed link left partial build state that cargo considered "fresh", so the retry silently skipped rebuilding the daemon.

- **Fixed EEG metrics not appearing in search results.** The `MetricsBlob` deserializer in `skill-history/cache.rs` used `#[serde(default)]` on `f64` fields, which caused the entire JSON deserialization to fail silently when any field was `null` (e.g., `cognitive_load: null`, `drowsiness: null`). All metrics defaulted to zero, causing every epoch to show "(no EEG metrics stored)" in the AI summary prompt. Fixed by adding a `null_as_zero` custom deserializer that treats JSON `null` as `0.0` without failing the parse.

- **Fixed search cache serving stale results after daemon restart.** Added `cache_clear()` to the search module and call it after metrics backfill completes, ensuring enriched metrics appear immediately.

### Refactor

- Move calibration timing loop from frontend to daemon. The daemon now drives all calibration phases (action countdowns, breaks, label submission) via a session runner with wall-clock-aligned timing. The frontend subscribes to WebSocket events for UI updates.
- Route all calibration commands through daemon HTTP endpoints. Profile CRUD, active profile selection, and calibration completion recording are now daemon-authoritative — the Tauri app is a thin proxy.
- Remove stale calibration state from Tauri `AppState`. Calibration profiles and active profile ID are no longer cached locally, preventing the save overlay from clobbering daemon-written timestamps.
- Delete `calibration_service.rs` — inline daemon HTTP calls directly in `window_cmds.rs`.
- Migrate onboarding calibration wizard to use daemon-driven sessions instead of a local timing loop.
- Remove unused `emit_calibration_event` Tauri command.

- Make `ensure_daemon_running` non-blocking. Window now appears immediately while daemon connection is established on a background thread.
- Share a single `ureq::Agent` via `OnceLock` instead of creating one per HTTP call.

### Build

- Fix macOS CI smoke test checking wrong path for daemon binary (`Contents/MacOS/skill-daemon` → `Contents/MacOS/skill-daemon.app/Contents/MacOS/skill-daemon`).
- Fix Linux CI Vulkan SDK cache restoring system library paths with permission errors. Vulkan SDK is now installed via `cache-apt-pkgs-action` alongside other system dependencies.
- Fix Linux CI daemon link failure: add `cargo:rustc-link-lib=vulkan` and openblas search paths in daemon `build.rs` for prebuilt llama archives.
- Fix Linux CI Discord notification firing on manual dispatch builds.
- Remove duplicate unguarded portable tarball upload step from Linux release workflow.
- Exclude E2E test files from `vitest related` in pre-push hook.
- Add all calibration commands to `check-daemon-invokes.js` validation script.

- Pin prebuilt llama download to specific tag (`0.2.46`) in `ci.mjs` instead of floating `latest`.
- Add explicit error and exit on daemon binary missing after prebuilt retry in release workflow.
- Add smoke-test step in release workflow: `codesign --verify`, architecture check with `file`/`lipo`.
- Add `scripts/test-daemon-e2e.sh` — 19 end-to-end tests covering fresh install, update hooks, uninstall, degraded states, edge cases, and connection reuse.

- Add `mtmd` (multimodal) to llama-cpp-4 target-specific dependency features in skill-llm, ensuring prebuilt archives include multimodal symbols.
- Copy Frameworks directory (dynamic llama dylibs) into daemon `.app` bundle during assembly.
- Fail app bundle assembly if daemon binary is missing instead of silently continuing.
- Clean daemon build artifacts during prebuilt llama retry to force a full re-link.

- Add `embed-zuna-gpu` (f32) feature to `embed-exg` for reliable GPU batch reembed.
- Fix all clippy warnings (dead code, large enum variant, too-many-arguments).
- Ignore Node.js DeprecationWarning in bump script warning detection.

### UI

- Add embedding health banner to search UI. Shows color-coded progress bars for EXG epochs, screenshots, and label EEG index when any modality is incomplete. Includes inline "Rebuild EEG index" button when the label-to-EEG bridge is empty.

- Move Embeddings tab directly below EXG in the settings sidebar.

### i18n

- Added 80+ translation keys across all 9 locales (en, de, es, fr, he, ja, ko, uk, zh) for AI summary, insights, bookmarks, color modes, timeline, compare, LLM controls, and chat integration.

- Added 60+ new translation keys across all 9 locales (en, de, es, fr, he, ja, ko, uk, zh) covering all new search features: filters, insights, timeline, compare, bookmarks, color modes, and performance stats.

### Dependencies

- Bump llama-cpp-4 / llama-cpp-sys-4 from 0.2.44 to 0.2.45.

- Bump zuna-rs to 0.1.3 (fixes burn f16→f32 tensor extraction bug).

## [0.0.122] — 2026-04-16

### Features

- Add recent/frequent command ranking to Cmd-K. Commands you use most often and most recently appear at the top in a "Recent" section, with frequency and recency boosts applied to search scoring.
- Add typo tolerance to Cmd-K search. Misspellings like "drak mode" or "calbriation" now find the right result using Damerau-Levenshtein edit distance as a fallback when fuzzy matching fails.
- Add contextual boosting to Cmd-K. Device, EEG, Calibration, and LSL commands are ranked higher when a device is connected; Devices settings are boosted when no device is connected.
- Add keyboard-driven toggles in Cmd-K. Toggleable settings (dark mode, high contrast) show an ON/OFF indicator and can be flipped with the Tab key without closing the palette.
- Add prefix-based command filtering to Cmd-K. Type `>` to show only system commands or `@` to show only settings.
- Add semantic search fallback to Cmd-K via fastembed embeddings. When fuzzy results are sparse and the query is 6+ characters, a debounced request to the daemon's `/v1/search/commands` endpoint returns semantically similar commands in a "Suggested" section. Handles natural language queries like "make text bigger" or "reduce eye strain".
- Add toggle dark/light mode command to the Cmd-K palette.

- Add deep-link search for individual settings in the Command Palette (Cmd-K). 154 settings across all 20 tabs are now searchable by name, description, and synonyms — typing "DND", "dark mode", "GPU", "OCR", etc. instantly finds the right setting.
- Auto-generate per-locale search indexes from i18n translation files and Tab components. Indexes are built for all 9 languages (EN, DE, ES, FR, HE, JA, KO, UK, ZH) and stay in sync automatically via a Vite plugin that watches for changes during development.
- Add synonym expansion to Cmd-K fuzzy search. Common abbreviations like DND, BT, GPU, OCR, LLM, and 30+ others are expanded before matching so users can search using shorthand.
- Add Command Palette button to the title bar of every window. Clicking the command-key icon opens Cmd-K, matching the keyboard shortcut behavior.
- Selecting a setting in Cmd-K now opens the Settings window, switches to the correct tab, scrolls to the exact setting, and flashes a blue highlight to draw attention to it.
- Wire settings index generation into `npm run dev`, `npm run build`, and `npm run bump` so indexes are always up to date before commits and builds.
- Add Chat and Compare shortcuts to the Shortcuts settings tab, making all 11 global shortcuts user-configurable. The Compare shortcut (previously hardcoded to Cmd+Shift+M) is now customizable like all others.
- Tray menu now reflects user-customized shortcuts for all actions including Chat and Compare.
- The keyboard shortcuts overlay (? key) now shows all 11 global shortcuts, matching the Shortcuts settings tab and tray menu.
- Global keyboard shortcuts no longer steal keystrokes from other apps. Shortcuts are now only registered when a Skill window is focused and unregistered when all windows lose focus, so the focused app always gets the keystroke first. The "Open NeuroSkill" shortcut (Cmd+Shift+O) remains always-on so users can bring the app to the foreground from anywhere.

- **Session sidecar `device_kind`**: JSON sidecars now include `device_kind` (e.g. "muse", "awear", "openbci") alongside `device_name` and `channel_names`. Enables reliable device identification during reembedding when a day directory contains sessions from different devices. Backward compatible — old sidecars without this field are handled gracefully.

- **Embedding coverage dashboard** in Settings → EXG → Re-embed. Shows a color-coded coverage bar (green ≥95%, amber ≥50%, red <50%), total/embedded/missing epoch counts, and estimated time remaining based on current embed speed.
- **Per-day breakdown table** in the re-embed section. Expandable table showing each day's total epochs, embedded count, missing count, and a mini coverage bar — makes it easy to identify which days have gaps.
- **Idle reembed status indicator**. When background embedding is enabled, the UI now shows whether it's actively processing (with day + progress), or waiting for the idle timeout to elapse, instead of running silently.
- **Re-embed progress improvements**. Progress bar now shows indeterminate animation during encoder loading, percentage + ETA during processing, amber bar when paused (device reconnected), red bar with error detail on failure.
- **EEG embedding coverage in Search page**. Corpus stats banner and empty-state panel now show epoch embedding coverage (e.g. "12,400/14,200 (87%)") with color coding, so you know at a glance whether your data is fully searchable.

- **Multi-modal search**: Interactive search now returns screenshots (proximity + OCR text similarity) and EEG epochs alongside text labels. New `kScreenshots`, `kLabels`, and `reachMinutes` parameters control result depth.
- **Search corpus stats**: New `GET /search/stats` endpoint and SSE streaming variant (`/search/stats/stream`) show EEG days, sessions, recording time, label counts, screenshot counts, and embedding model info. Stats appear in the search UI for all modes.
- **Daemon test mode**: `POST /v1/test/begin` pauses background work (screenshots, OCR, re-embed) for stable E2E testing. `POST /v1/test/end` resumes. Debug builds only (`#[cfg(debug_assertions)]`).
- **Daemon readiness**: `/readyz` now returns `{ok, ready, test_mode}` — `ready` is set after the TCP listener binds. `SKILL_DAEMON_TOKEN` env var allows injecting auth tokens for isolated test daemons.
- **Interactive test picker**: `npm test` shows a TUI picker (Node.js, cross-platform) with 17 suites. Arrow keys + space to toggle, enter to run.
- **Rust version guard**: `npm run tauri dev` and `npm run test:clippy` check Rust >= 1.95 and prompt `rustup update stable` if outdated.

### Bugfixes

- Fix Chat shortcut not being loaded from saved settings on app startup.

- **Reembed silent failures**: Added diagnostic logging on first extract failure and first encode failure per batch. Previously all epoch failures were completely silent, making it impossible to debug why 27K+ epochs would fail with zero output.
- **CSV channel length mismatch**: The CSV parser now skips entire rows when any channel value fails to parse, preventing channels from accumulating different sample counts. Previously, individual parse failures were silently skipped per-channel, creating mismatched-length arrays that the ZUNA encoder rejected.
- **Mixed-device reembedding**: CSVs with fewer columns than the metadata expects (e.g. different device in the same day directory) now use available columns instead of skipping all rows. The column count is detected per-file from the CSV header.
- **Reconnect respects test mode**: The BLE reconnect loop now pauses when `test_mode` is active, preventing background connection attempts from interfering with E2E tests.
- **`cpu-only` feature build**: Fixed compilation errors in `skill-daemon` when built with `--no-default-features --features cpu-only`. The `#[cfg(not(feature = "llm"))]` fallback paths referenced `AppState` fields that don't exist without the `llm` feature. Replaced with static stubs. This fixes the CI coverage job.
- **Clippy 1.95**: Fixed `sort_by` → `sort_by_key(Reverse)` in skill-history, `collapsible_match` in connect_wired, `checked_div` in iroh_remote.
- **Prebuilt llama naming**: Updated download URL from `q1-metal`/`q1-vulkan` to `metal-static`/`vulkan-static` to match current release asset names.

- **Fix macOS DMG creation stream error**: Added retry logic (up to 3 attempts) to the `appdmg` invocation in `scripts/create-macos-dmg.sh`. The `ERR_STREAM_WRITE_AFTER_END` race condition in appdmg's internal file-copy stream is now caught and retried automatically, with partial DMG cleanup between attempts.

- **Search hang fixed**: Removed expensive `collect_search_meta` (session JSON parsing) from the search hot path. Corpus stats are now fetched separately via a dedicated endpoint.
- **Label stale count**: `count_needing_embed()` uses `SELECT COUNT(*)` instead of loading all stale rows into memory.
- **Heredoc terminators**: Fixed indented heredoc terminators (`DPLIST`, `PYEOF`) in release-mac, release-linux, and pr-build workflows that caused `syntax error: unexpected end of file`.
- **Flaky env var test**: Added `Mutex` guard to `enforce_path_integrity` tests in skill-tools to prevent parallel env var races.
- **A11y errors**: Added `aria-label` to 3 unlabeled inputs in `EegModelTab.svelte`.
- **Smoke test auth**: Added Bearer token to all HTTP requests in `test.ts` (was sending unauthenticated). Updated 20+ REST paths from old Tauri-era shortcuts to `/v1/` daemon endpoints. Score: 83/275 → 341/0.
- **E2E test isolation**: Daemon token and data E2E tests now spawn isolated daemons on random ports with temp skill dirs. Can run in parallel.
- **Pre-checkout CI step**: `free-disk-space` step inlined for workflows that run it before `actions/checkout`.
- **ESM import fix**: Removed Node 22-only `glob` import and `require()` calls from `ci.mjs` (broke on Node 20 CI runners).

### Refactor

- **Dead scripts removed**: Deleted `prepare-daemon-sidecar.sh` (superseded by `.js`), `adopt_anyhow.py` (unused), `check_windows_manifest.py` (ported to Node).
- **Screenshot store reuse**: Interactive search opens the screenshot store once and reuses it for both proximity and OCR searches.
- **E2E helpers**: New `src/tests/e2e-helpers.ts` with `spawnDaemon()`, `testBegin()`/`testEnd()`, `isDaemonReady()`, `freePort()` shared across all E2E test files.

### Build

- **CI scripts consolidated**: Replaced ~960 lines of duplicated inline YAML with a single `scripts/ci.mjs` (Node.js, 12 subcommands). Zero Python dependency for builds — all scripts are Node.js + bash.
- **Release workflows refactored**: All 3 release workflows (macOS, Linux, Windows) support `workflow_dispatch` for on-demand builds. On-demand runs upload artifacts (14-day retention) without touching GitHub Releases.
- **macOS .app bundling**: Extracted 120-line inline YAML into `scripts/assemble-macos-app.sh`. Fixed heredoc terminator bug that caused CI failures.
- **Pre-push hook**: Runs `ci.mjs self-test` when CI scripts change, and `daemon-client.test.ts` when daemon proxy files change.
- **CI dry-run**: `npm run ci:dry-run` runs the full release pipeline locally without signing/pushing. `--skip-compile` for fast iteration.
- **Clippy 1.95 fixes**: `sort_by` → `sort_by_key(Reverse)`, collapsible match arms, `checked_div`.

### Server

- Add `POST /v1/search/commands` endpoint for semantic command search. Accepts a query and candidate list, embeds both with nomic-embed-text-v1.5 via fastembed, and returns the top 5 results ranked by cosine similarity.

- `GET /v1/models/estimate-reembed` response now includes `embedded`, `missing`, `coverage_pct`, `avg_embed_ms`, `eta_secs`, `per_day` (per-day breakdown array), and `idle_reembed` (background reembed status with active/idle_secs/done/total/current_day).
- `GET /v1/search/stats/stream` slow-tier stats now include `eeg_total_epochs`, `eeg_embedded_epochs`, and `eeg_missing_epochs`.
- New `IdleReembedStatus` struct in daemon state tracks background reembed progress and idle countdown, updated in real-time by the idle reembed loop.

## [0.0.121] — 2026-04-15

### Features

- update failing apple notary tool

## [0.0.120] — 2026-04-15

### Features

- updated API docs

## [0.0.119] — 2026-04-14

### Features

- added openblas to ci
- fixed tests limits
- missing keys

## [0.0.118] — 2026-04-14

### Features

- updated neuroloop
- neuroloop 0.1.1
- added new logo
- updated neuroloop and neuroskill
- added korean and a11y + fixed i18n
- merged PRs
- deps(cargo): bump the cargo-all group across 1 directory with 8 updates
- deps(npm): bump the npm-all group with 11 updates
- ndarray optimization
- updated embeddings and hnsw
- zuna-rs 0.0.4 — zero-cost model init (no random waste)
- fixed labels

### Compare Window Performance

- **SQLite timestamp index** — Added `CREATE INDEX idx_embeddings_timestamp ON embeddings(timestamp)` to all day databases. Range queries on the `embeddings` table went from full table scan to indexed lookup (40ms → 5ms for COUNT on 146K rows).
- **Directory date filtering** (`dirs_for_range`) — Compare endpoints now only open SQLite files for days that overlap the requested time range, skipping irrelevant directories (e.g., 363 of 365 day dirs skipped for a 2-day comparison).
- **Single JSON parse per row** — Replaced 49-50 `json_extract()` calls per SQLite row with `SELECT metrics_json` + one serde deserialize in Rust. For `get_session_metrics` on 60K rows: 1.03s → 82ms (12x faster).
- **SQL-level timeseries downsampling** — `get_session_timeseries` now uses `ROW_NUMBER() OVER ... WHERE rn % step = 1` to return ~800 evenly-spaced rows instead of all 145K. Transfer + serde reduced from ~3s to ~40ms.
- **Read-only SQLite connections** — Analysis endpoints now use `SQLITE_OPEN_READ_ONLY` to eliminate write-lock contention when 6+ parallel queries run during compare.
- **Two-phase progressive loading** — Compare UI shows metrics first (fast SQL query), then loads sleep and timeseries in the background. Results appear in <100ms instead of waiting 3+ seconds.
- **CSS `.spin` animation** — Added missing `@keyframes spin` and `.spin` class to `app.css`. Loading spinners across the app were using the class but it was never defined.

### UMAP / Brain Nebula Fixes

- **Poll handler: "running" status** — Fixed `pollUmapJob` to treat `status: "running"` same as `"pending"` (was silently calling `finishUmap()` and aborting the poll loop).
- **HTTP timeout for UMAP** — UMAP computation request now uses 5-minute timeout instead of the default 10 seconds (which killed the request mid-computation).
- **Empty embedding detection** — `umap_compute_inner` now filters out 0-length embedding BLOBs, reports `total_a`/`total_b` (total epochs found) vs `n_a`/`n_b` (epochs with actual embeddings), and returns a `reason` field explaining why UMAP can't compute.
- **User-facing error UI** — When UMAP returns empty (no embeddings), shows a warning panel with explanation ("Brain Nebula requires EEG embedding vectors…") and a Retry button.
- **UMAP progress bar** — Added estimated progress bar during UMAP computation based on elapsed/estimated time, with countdown. Shows epoch-level progress when available from the daemon.

### Embedding Pipeline & Batch Reembed

- **Encoder failure logging** — Upgraded encoder load failure from `warn!` to `error!` with actionable message. Added per-epoch `metrics_only_streak` counter that logs first failure, every 100th consecutive failure, and recovery.
- **Slow encode warning** — Logs when a single epoch encode exceeds 2 seconds.
- **Batch reembed implementation** — `POST /v1/models/trigger-reembed` was a stub returning "queued". Now fully implemented: loads encoder, scans all day directories, reads raw EEG CSVs, encodes epochs, writes embeddings back to SQLite.
- **`estimate_reembed` fixed** — Now correctly counts epochs with empty `eeg_embedding` BLOBs (was always returning 0).
- **Relative CSV timestamp fix** — Batch reembed now handles both absolute Unix timestamps and relative device-internal timestamps in CSV files by using the session start time from the filename.
- **Transaction batching** — SQLite writes during reembed use batched transactions (50 epochs per COMMIT) for write performance.
- **Progress events** — Emits `reembed-progress` WebSocket events every batch with status, total, done, failed, and current day.

### GPU-Accelerated Encoding

- **zuna-rs 0.1.0** — Upgraded from 0.0.4. Brings parallel processing (rayon), CPU preprocessing pipeline, and `wgpu-f16` feature for half-precision GPU inference.
- **GPU encoder (Metal/Vulkan)** — New `embed-zuna-gpu` and `embed-zuna-gpu-f16` features. `ZunaGpuState` loads ZUNA on wgpu (Metal on macOS). Encoding: ~120ms/epoch on GPU vs 5,600ms on CPU (47x speedup).
- **f16 half-precision** — Default GPU precision is f16 (`embed-zuna-gpu-f16`), faster than f32 on Apple Silicon. Configurable via `gpu_precision` setting.
- **Encoder fallback chain** — `load_encoder_public` tries GPU f16 → GPU f32 → CPU (NdArray + Accelerate BLAS).

### Idle Background Reembed

- **`idle_reembed` background loop** — New `idle_reembed.rs` module spawned at daemon startup. Monitors device connection state every 10 seconds. When device is disconnected for `idle_reembed_delay_secs` (default 30 minutes), starts background embedding of unprocessed epochs.
- **Backpressure on device connect** — `idle_reembed_cancel` AtomicBool flag set to `true` in `DeviceEvent::Connected` handler. Batch reembed checks cancel flag between every batch and pauses immediately, emitting a "paused" status event.
- **Configurable settings** — `ReembedConfig` extended with `idle_reembed_enabled`, `idle_reembed_delay_secs`, `idle_reembed_gpu`, `gpu_precision`, `idle_reembed_throttle_ms`. All exposed via `GET/POST /v1/settings/reembed-config`.

### Embedding Status UI

- **Timeline picker indicator** — Small colored dot next to session status: green (90%+ embedded), amber + percentage (partial), red (0% embedded).
- **UMAP section indicator** — "Calculate UMAP" button disabled when <5 embeddings available, with explanation text. Shows reembed progress bar when background embedding is running.
- **Reembed progress in compare** — Live progress bar with rate estimate (e.g., "6.3 embeddings/sec") and ETA (e.g., "~11h 32m remaining") from WebSocket `reembed-progress` events.
- **EEG Model settings** — Added GPU Precision dropdown (f16/f32), Background Embedding toggle, and Idle Delay selector (5min/15min/30min/1hr) to the EEG Model tab.

### Search Improvements

- **Interactive search implemented** — `POST /v1/search/eeg` now handles `interactive_search` requests: embeds query text → searches label HNSW → finds nearby EEG epochs → discovers temporal neighbor labels. Returns full graph with nodes and edges.
- **Text search routing fixed** — `search_labels_by_text` was routed to the wrong endpoint (`/v1/search/eeg` which returned `{ results: [] }`). Now correctly routes to `/v1/labels/search` with response unwrapping.
- **EEG search: empty embedding filter** — `read_embeddings_in_range` now filters `WHERE length(eeg_embedding) >= 4` at the SQL level, avoiding loading and searching 65K+ empty BLOBs. Search time for 66K epoch range: >60s → 3.6s.
- **SSE streaming search** — New `POST /v1/search/eeg/stream` endpoint sends results as Server-Sent Events as they're found. Frontend reads via `ReadableStream` with fallback to full-body read. Results appear progressively in the UI.
- **Cancel search** — `cancelSearch()` exported from invoke-proxy aborts the SSE fetch. Cancel button in search UI now actually stops the streaming request (not just sets a flag).
- **Channel command timeouts** — Search uses 2-minute timeout, chat uses 5-minute timeout (was 10 seconds for both).
- **Null-safety** — Added comprehensive null guards for `nb.labels`, `nb.distance`, `nb.neighbors`, `q.neighbors`, `x.text`, `x.context` across search page and search-types. Prevents crashes when daemon returns entries with null/undefined fields.
- **Streaming performance** — `$derived` computations (`searchAnalysis`, `temporalHeatmap`, `filtered`) skip during active streaming to avoid O(n^2) recomputation on every incoming result.

### Screenshot Image Serving

- **Static file serving** — Daemon now serves screenshot images at `/screenshots/{filename}` and `/screenshots/{date}/{filename}`. Infers the `YYYYMMDD` subdirectory from the first 8 characters of the filename.
- **Auth required** — Screenshot routes go through the auth middleware. `<img>` tags use `?token=` query parameter for authentication. Token loaded from daemon bootstrap at page mount.

### Search Tab UX

- **Active tab styling** — Search mode tabs now have clear visual distinction: inactive tabs are dimmed (45% opacity), active tab has bold text, accent-colored background tint, and a solid 2px bottom border.

### Crash Fix

- **muda ZeroWidth panic** — Patched `muda 0.17.2` locally to replace `.unwrap()` with safe fallbacks in `to_png()` (macOS icon.rs). Prevents crash when macOS event loop processes a zero-width icon image during window creation.

### i18n

- Added embedding status translations (`compare.embeddingsReady`, `compare.embeddingsPartial`, `compare.embeddingsNone`, `compare.embeddingsProcessing`) to all 9 locales (en, de, es, fr, he, ja, ko, uk, zh).
- Added GPU/idle reembed translations (`model.gpuPrecision`, `model.gpuPrecisionDesc`, `model.gpuPrecisionF16`, `model.gpuPrecisionF32`, `model.idleReembed`, `model.idleReembedDesc`, `model.idleDelay`) to all 9 locales.

### Tests

- **compare-analysis-e2e.test.ts** — 18 end-to-end tests against live daemon covering metrics, timeseries, sleep, UMAP, parallel requests, and invoke-proxy routing.
- **search-null-safety.spec.ts** — Playwright test for EEG/text/interactive search with null/undefined fields in results.

## [0.0.117] — 2026-04-13

### Features

- neuroskill
- fixed cargo clippy
- fixed svelte errors
- updated main app for neuroskill and neuroloop
-  Neuroloop TUI:
-  Daemon:
-  Neuroloop provider registration:
-  No Tauri changes needed — the desktop UI already handles <think> tags natively with collapsible thinking blocks.
- fixed neuroskill + added japanese
- fixed neuroskill
- updated neuroloop
- updated iroh
- updated neuroskill
- tests
- updated cli
- updated faq
- refactored faq and help
- updated disconnect
- added submodules
- fixed tests
- biome fixes

## [0.0.116] — 2026-04-11

### Features

- Minor updates and improvements

## [0.0.115] — 2026-04-11

### Features

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

## [0.0.114] — 2026-04-11

### 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

## [0.0.113] — 2026-04-11

### 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

## [0.0.112] — 2026-04-11

### Features

- New daemon files created:
- - crates/skill-daemon/src/reconnect.rs — Reconnect state machine (backoff, retry countdown, MAX_RETRY_ATTEMPTS) with background loop and REST endpoints
- Daemon state.rs changes:
- - Added broadcast() convenience method on AppState
- Daemon main.rs changes:
- - Registered new modules (reconnect, monitor, background)
- Daemon session metadata enriched:
- - write_session_meta_full() now accepts optional StatusResponse for richer metadata (battery, signal quality, device identity, phone info)
- Tauri files simplified:
- - lifecycle.rs — Removed reconnect state machine (130+ lines). go_disconnected now just cleans local state and delegates to daemon
- updated coverage.ci
-   Removed direct deps (37 total):
- lean tauri
- decouple tauri from wgpu
- chinese
- exclude skill from coverage ci
- Both GPU and CPU modes compile. Here's a summary of all changes made:
-   1. scripts/create-macos-dmg.sh — Fixed unbound CLEANUP_DIRS[@] under set -u
- fixed ci
- updated CI

## [0.0.111] — 2026-04-11

### Features

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

## [0.0.110] — 2026-04-11

### 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:

## [0.0.109] — 2026-04-10

### 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

## [0.0.106] — 2026-04-10

### 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

## [0.0.104] — 2026-04-10

### 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

## [0.0.103] — 2026-04-10

### 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

## [0.0.102] — 2026-04-10

### Features

- mtmd fix
- improved windows TUI
- added postinstall
- minor update
- win + linux
- mac
- better coverage
- improved coverage
- try_send instead of blocking_send
- fix: spawn llm_chat streaming as tokio task to prevent deadlock
- feat: draft prefill, LLM start UX fixes, chat input multiline alignment
- fix: align dev TUI header art
- fix: stabilize iroh remote sessions and dashboard streaming status
- chore: silence non-actionable cargo-deny license metadata warnings
- chore: commit all pending changes
- ci: run coverage without all-features on non-CUDA runners
- chore: update tests, config, and clients tab
- Fix iroh session stability, revoke/scope URLs, and metrics continuity
- Add start_session / cancel_session WS commands; exempt peer: from pairing check
- Wire IrohRemoteAdapter into daemon session runner
- Authenticate iroh peers via tunnel identity instead of bearer token

## [0.0.101] — 2026-04-10

### Features

- mtmd fix
- improved windows TUI
- added postinstall
- minor update
- win + linux
- mac
- better coverage
- improved coverage
- try_send instead of blocking_send
- fix: spawn llm_chat streaming as tokio task to prevent deadlock
- feat: draft prefill, LLM start UX fixes, chat input multiline alignment
- fix: align dev TUI header art
- fix: stabilize iroh remote sessions and dashboard streaming status
- chore: silence non-actionable cargo-deny license metadata warnings
- chore: commit all pending changes
- ci: run coverage without all-features on non-CUDA runners
- chore: update tests, config, and clients tab
- Fix iroh session stability, revoke/scope URLs, and metrics continuity
- Add start_session / cancel_session WS commands; exempt peer: from pairing check
- Wire IrohRemoteAdapter into daemon session runner
- Authenticate iroh peers via tunnel identity instead of bearer token

## [0.0.100] — 2026-04-10

### Features

- mtmd fix
- improved windows TUI

## [0.0.99] — 2026-04-10

### Features

- added postinstall
- minor update
- win + linux
- mac
- better coverage
- improved coverage
- try_send instead of blocking_send
- fix: spawn llm_chat streaming as tokio task to prevent deadlock
- feat: draft prefill, LLM start UX fixes, chat input multiline alignment
- fix: align dev TUI header art
- fix: stabilize iroh remote sessions and dashboard streaming status
- chore: silence non-actionable cargo-deny license metadata warnings
- chore: commit all pending changes
- ci: run coverage without all-features on non-CUDA runners
- chore: update tests, config, and clients tab
- Fix iroh session stability, revoke/scope URLs, and metrics continuity
- Add start_session / cancel_session WS commands; exempt peer: from pairing check
- Wire IrohRemoteAdapter into daemon session runner
- Authenticate iroh peers via tunnel identity instead of bearer token
- Fix TUI line truncation caused by unstripped OSC hyperlink sequences
- Fix axum 0.7 path param syntax in iroh routes (:{id} → {id})
- Use dedicated port 18445 for dev daemon to avoid system service conflict
- Kill stale daemon on dev startup and pin SKILL_DAEMON_BIN to local build

## [0.0.98] — 2026-04-10

### Features

- added postinstall
- minor update
- win + linux
- mac
- better coverage
- improved coverage
- try_send instead of blocking_send
- fix: spawn llm_chat streaming as tokio task to prevent deadlock
- feat: draft prefill, LLM start UX fixes, chat input multiline alignment
- fix: align dev TUI header art
- fix: stabilize iroh remote sessions and dashboard streaming status
- chore: silence non-actionable cargo-deny license metadata warnings
- chore: commit all pending changes
- ci: run coverage without all-features on non-CUDA runners
- chore: update tests, config, and clients tab
- Fix iroh session stability, revoke/scope URLs, and metrics continuity
- Add start_session / cancel_session WS commands; exempt peer: from pairing check
- Wire IrohRemoteAdapter into daemon session runner
- Authenticate iroh peers via tunnel identity instead of bearer token
- Fix TUI line truncation caused by unstripped OSC hyperlink sequences
- Fix axum 0.7 path param syntax in iroh routes (:{id} → {id})
- Use dedicated port 18445 for dev daemon to avoid system service conflict
- Kill stale daemon on dev startup and pin SKILL_DAEMON_BIN to local build

## [0.0.97] — 2026-04-10

### Bugfixes

- **Fix `spawnSync npm.cmd EINVAL` on Windows**. Added `shell: true` to `execFileSync` calls when the resolved Tauri CLI command is a `.cmd` shim. Windows `.cmd` files are batch scripts that require shell execution.

- **Fix coverage CI: skip EXG embedding worker in tests**. Session runner tests panicked on headless CI (no GPU) because `cubecl-wgpu` tried to initialize a Vulkan adapter. Added `cfg!(test)` check to skip the embedding worker during test builds. Also supports `SKILL_SKIP_EMBED=1` env var for headless production environments.

### Build

- **Fix release CI: build daemon before Tauri app**. Reversed the build order so `skill-daemon` compiles first using its default features (which resolve `llama-cpp-sys-4` with `mtmd` + Metal/Vulkan + `q1` via `skill-llm`'s target-specific deps). The Tauri app build runs second and reuses cached native libs. Removed explicit `--features llm-metal`/`llm-vulkan` flags — daemon defaults handle GPU backend selection per platform automatically.

## [0.0.96] — 2026-04-10

### Bugfixes

- **LLM server enabled state now persists across daemon restarts**. When the user starts the LLM server, `config.enabled = true` is saved to disk. When stopped, `config.enabled = false` is saved. On daemon startup, if `enabled` is `true`, the LLM server auto-starts. Previously, stopping the server did not persist the disabled state, and the daemon never auto-started on boot.

- **Fix TUI dev mode on Windows**. Enabled TUI by default on all platforms. Fixed `killChildTree` to use `taskkill /T /F /PID` on Windows instead of unsupported negative-PID process group kill. Set `detached: false` on Windows to prevent new console windows. Added robust try/catch around terminal escape sequences and raw mode setup. Graceful fallback to standard dev mode on any platform if TUI fails.

- **Fix `npm run tauri dev` ENOENT on Windows**. The Tauri CLI resolver used bare `npm` which fails with `spawnSync npm ENOENT` because Windows requires `npm.cmd` for `execFileSync`. Now uses `npm.cmd`/`npx.cmd` on Windows.

### Refactor

- **Thin Tauri app: LLM engine, NeuTTS, and EXG embedding removed from the UI binary**. The Tauri app is now a thin UI shell that communicates with `skill-daemon` over HTTP/WebSocket. `llama-cpp-4`, `neutts`, and all EXG encoder crates (`zuna-rs`, `luna-rs`, `tribev2`, `burn`, `fastembed`, `ort`, etc.) are no longer compiled into the Tauri binary — they live exclusively in the daemon. This dramatically reduces Tauri compile time and binary size. All LLM Tauri commands (catalog, downloads, server lifecycle, chat completions, streaming) continue to work unchanged as daemon HTTP proxies. The `llm` module is always compiled but engine-specific code (`init`, `shutdown`, direct inference) is gated behind `#[cfg(feature = "llm")]` with lightweight stub types (`LlmStatus`, `LlmLogBuffer`) provided when the feature is off.

### Build

- **Fix release CI: daemon and Tauri app built and bundled together across all OS**. The daemon sidecar is now built alongside the Tauri app in release workflows on all three platforms. macOS uses `--features llm-metal`, Windows and Linux use `--features llm-vulkan`. The Tauri app itself only needs `--features custom-protocol` (no LLM features). Fixed macOS PKG step ordering (`.app` bundle assembly now runs before PKG staging). Removed conflicting system-level service installations — the daemon self-registers as a user LaunchAgent/systemd user unit/Windows service at runtime. Added `scripts/pkg-scripts/postinstall` for macOS PKG. Windows NSIS now includes daemon process/service cleanup and firewall rules on install/uninstall. Linux `package-linux-system-bundles.sh` and `package-linux-dist.sh` updated to find and bundle the daemon binary.

- **Fix `npm run bump` changelog compilation**. Uncommented the `compileChangelog` import and fragment validation that were disabled. Bump now compiles `changes/unreleased/*.md` fragments into `changes/releases/<version>.md` and rebuilds `CHANGELOG.md`. When no fragments exist, auto-generates one from git commit history since the last version. All file mutations are wrapped in try/catch with automatic `git checkout` rollback on failure. Commit uses `git add -A` and runs pre-commit hooks (no `--no-verify`).

## [0.0.95] — 2026-04-10

### Features

- added postinstall
- minor update
- win + linux
- mac
- better coverage
- improved coverage
- try_send instead of blocking_send
- fix: spawn llm_chat streaming as tokio task to prevent deadlock
- feat: draft prefill, LLM start UX fixes, chat input multiline alignment
- fix: align dev TUI header art
- fix: stabilize iroh remote sessions and dashboard streaming status
- chore: silence non-actionable cargo-deny license metadata warnings
- chore: commit all pending changes
- ci: run coverage without all-features on non-CUDA runners
- chore: update tests, config, and clients tab
- Fix iroh session stability, revoke/scope URLs, and metrics continuity
- Add start_session / cancel_session WS commands; exempt peer: from pairing check
- Wire IrohRemoteAdapter into daemon session runner
- Authenticate iroh peers via tunnel identity instead of bearer token
- Fix TUI line truncation caused by unstripped OSC hyperlink sequences
- Fix axum 0.7 path param syntax in iroh routes (:{id} → {id})
- Use dedicated port 18445 for dev daemon to avoid system service conflict
- Kill stale daemon on dev startup and pin SKILL_DAEMON_BIN to local build
- Move iroh tunnel to daemon and add /v1/iroh/* HTTP routes
- Handle Windows STATUS_CONTROL_C_EXIT as graceful dev shutdown

## [0.0.94] — 2026-04-10

### Features

- added postinstall
- minor update
- win + linux
- mac
- better coverage
- improved coverage
- try_send instead of blocking_send
- fix: spawn llm_chat streaming as tokio task to prevent deadlock
- feat: draft prefill, LLM start UX fixes, chat input multiline alignment
- fix: align dev TUI header art
- fix: stabilize iroh remote sessions and dashboard streaming status
- chore: silence non-actionable cargo-deny license metadata warnings
- chore: commit all pending changes
- ci: run coverage without all-features on non-CUDA runners
- chore: update tests, config, and clients tab
- Fix iroh session stability, revoke/scope URLs, and metrics continuity
- Add start_session / cancel_session WS commands; exempt peer: from pairing check
- Wire IrohRemoteAdapter into daemon session runner
- Authenticate iroh peers via tunnel identity instead of bearer token
- Fix TUI line truncation caused by unstripped OSC hyperlink sequences
- Fix axum 0.7 path param syntax in iroh routes (:{id} → {id})
- Use dedicated port 18445 for dev daemon to avoid system service conflict
- Kill stale daemon on dev startup and pin SKILL_DAEMON_BIN to local build
- Move iroh tunnel to daemon and add /v1/iroh/* HTTP routes
- Handle Windows STATUS_CONTROL_C_EXIT as graceful dev shutdown

## [0.0.93] — 2026-04-10

### Bugfixes

- **Fix TUI dev mode on Windows**. Enabled TUI by default on all platforms (was disabled on Windows). Fixed `killChildTree` to use `taskkill /T /F /PID` on Windows instead of unsupported negative-PID process group kill. Set `detached: false` on Windows to prevent new console windows. Added robust try/catch around terminal escape sequences and raw mode setup. Graceful fallback to standard dev mode on any platform if TUI fails.

- **Fix `npm run tauri dev` ENOENT on Windows**. The Tauri CLI resolver fell through to bare `npm` which fails with `spawnSync npm ENOENT` because Windows requires `npm.cmd` for `execFileSync`. Now uses `npm.cmd`/`npx.cmd` on Windows.

### Build

- **Fix release CI: daemon + Tauri app bundled together across all OS**. The `skill-daemon` sidecar is now built alongside the Tauri app in a single `cargo build -p skill -p skill-daemon` invocation on all three platforms. This ensures cargo's feature unification produces one consistent `llama-cpp-sys-4` build with `mtmd` + GPU backend + `q1` features, fixing undefined `_mtmd_*` linker errors that occurred when building the daemon separately. macOS uses `--features llm-metal`, Windows and Linux use `--features llm-vulkan,screenshots`.

- **Fix macOS PKG step ordering**: moved `.app` bundle assembly before PKG staging so the `.app` exists when the installer copies it. Removed conflicting system-level `/Library/LaunchDaemons` installation — the daemon self-registers as a user LaunchAgent at runtime via its `/service/install` HTTP endpoint.

- **Fix Windows NSIS: premature compile step removed**. A "Compile (frontend + Rust + daemon)" step ran before LLVM 19 was installed, causing `__builtin_ia32_*` bindgen errors from VS2022's clang-cl headers. Daemon build now runs in the existing Compile step after LLVM 19 is available. Added daemon process/service cleanup and firewall rules to NSIS install/uninstall.

- **Fix Linux packaging: daemon bundled in deb/rpm/portable tarball**. The staging step no longer references a nonexistent AppImage. Removed conflicting `/etc/systemd/system` service file — the daemon self-registers via `systemctl --user` at runtime. Updated `package-linux-system-bundles.sh` and `package-linux-dist.sh` to find and bundle the daemon binary from the release target directory.

- **Add `scripts/pkg-scripts/postinstall`** for macOS PKG installer (required by `pkgbuild --scripts`).

## [0.0.88] — 2026-04-09

### Bugfixes

- **Fix daemon sidecar copy warning during build prep**: updated `scripts/prepare-daemon-sidecar.sh` to skip self-copy when source and destination are identical, removing false warning noise during release builds.
- **Daemon startup upgrade resilience**: added runtime readiness flow that performs protocol compatibility checks, recovery restart attempts, rollback snapshot fallback, and background-service repair.

### Build

- **Bundle daemon sidecar in packaging workflows**: ensured packaging scripts include `skill-daemon`/`skill-daemon.exe` alongside the app binary across manual bundle flows (macOS app/DMG assembly and Windows NSIS staging/install/uninstall paths), and added cross-platform packaging validation scripts (`scripts/test-daemon-packaging.sh` and `scripts/test-daemon-packaging.ps1`).

### Server

- **Daemon runtime service hardening**: app startup now ensures daemon runtime readiness (run → protocol gate → restart/rollback recovery) and auto-heals background service registration/status via daemon service endpoints.

### Docs

- **Document daemon packaging + service lifecycle checks**: expanded `docs/DEVELOPMENT.md`, `docs/WINDOWS.md`, `docs/LINUX.md`, and `docs/architecture.md` with daemon bundling validation commands, service behavior, and rollback/runtime-readiness architecture notes.

## [0.0.87] — 2026-04-08

### Features

- **Daemon logs in dev terminal**: `npm run tauri dev` now streams daemon
  `tracing` output into the same terminal as Tauri logs.  A custom
  `TeeWriter` tees every log event to both stderr and a 512-line ring buffer
  in `AppState`; `GET /v1/log/recent?since=<seq>` exposes the buffer; the
  Tauri process polls it every second and `eprintln!`s new lines, which the
  existing stderr tee also saves to the session log file.

- **Event-driven BLE device discovery** replaces the old polling approach
  that created a new `BtManager` (and thus a new `CBCentralManager`) on
  every 5-second scanner tick.  A single `run_ble_listener_task` now holds
  the adapter alive for the lifetime of the scanner, subscribes to
  `CentralEvent::DeviceDiscovered` / `DeviceUpdated`, and updates a shared
  `ble_device_cache` as advertisements arrive.  This ensures `local_name`
  is populated (CoreBluetooth fills it asynchronously; it was often `None`
  in a short poll window) and means the Muse S shows up by name instead of
  as an anonymous UUID.

- **EEG-only BLE filter**: the BLE listener now only surfaces devices whose
  advertising name matches a known EEG/neurofeedback pattern (`muse*`,
  `ganglion*`, `mw75`, `brainbit`, `unicorn`, `hermes`, `mendi`, `idun`,
  etc.).  Previously every BLE peripheral in range — phones, speakers,
  watches, UUIDs with no name — appeared in the Discovered Devices list.

- **Fast BLE connect path** for all five BLE device families (Muse, MW75,
  Hermes, Mendi, IDUN Guardian): replaced the fixed-sleep `scan_all()` call
  (3–5 s) with `connect()`, which polls `adapter.peripherals()` every 250 ms
  and exits as soon as the target device appears.  For a paired device that
  is already advertising this drops the scan phase from ~5 s to ~250 ms.
  The paired device's name is looked up from `status.paired_devices` and
  passed as `name_prefix` so `connect()` matches the exact headset rather
  than the first device of that family.  A slow `scan_all()` fallback is
  kept for Muse only when no name is known (first-time unpaired connect).

- **PPG data now written to session CSV/Parquet**: `DeviceEvent::Ppg` frames
  are now passed to `pipe.writer.push_ppg()` in the session runner.
  Previously PPG was only broadcast over WebSocket; it was never recorded.
  Muse PPG (optical heart-rate channels: Ambient, Infrared, Red) now
  appears in session files alongside EEG, IMU, and band metrics.  PPG
  heart-rate / HRV metrics (`PpgAnalyzer`) are now computed by the daemon
  and written alongside raw PPG samples in each epoch.

- **fNIRS recording pipeline**: full end-to-end fNIRS data persistence for
  devices like Mendi that provide optical brain-imaging data instead of EEG.

  - `DeviceEvent::Fnirs(FnirsFrame)` — new event variant carrying raw
    multi-channel photodetector ADC values + device timestamp.
  - `DeviceCaps::FNIRS` flag added to the capability bitflags.
  - `MendiAdapter` now emits `DeviceEvent::Fnirs` (9 channels: IR/Red/Ambient
    × Left/Right/Pulse) instead of packing optical data into a raw-JSON
    `DeviceEvent::Meta`.  Temperature-only diagnostics remain as `Meta`.
  - `SessionCsvWriter::push_fnirs()` — lazily creates `exg_*_fnirs.csv` with
    `timestamp_s` + per-channel columns using the device's
    `fnirs_channel_names`.  Auto-flushes every 64 rows.
  - `SessionParquetWriter::push_fnirs()` — lazily creates `exg_*_fnirs.parquet`
    with a Snappy-compressed schema built dynamically from channel names.
    Batches 256 rows before flushing.  Properly integrated into `flush()` and
    `close()` lifecycle.
  - `SessionWriter::push_fnirs()` dispatches to CSV, Parquet, or both based on
    the user's `storage_format` setting.
  - Session runner: `DeviceEvent::Fnirs` is written to disk and broadcast as a
    `FnirsSample` WebSocket event.

- **Firmware version shown live on the dashboard**: the connected-device
  info row (below the device name) now includes `fw <version>` when the
  Muse sends its firmware version string via a Control JSON event, e.g.
  `fw 3.4.5`.  Previously the value was extracted and stored in
  `state.status.firmware_version` but never surfaced in the UI.

- **`firmware_version` and `bootloader_version` added to the TypeScript
  `DeviceStatus` type**: both fields were serialised by Tauri and arrived
  in every status payload but were absent from the interface declaration in
  `types.ts`, so the TypeScript compiler would silently discard them.

- **Session sidecar (`exg_*.json`) now includes `firmware_version` and
  `serial_number`**: the `Pipeline` struct stores both values when
  `DeviceEvent::Connected` arrives and updates `firmware_version` again
  when the Muse's Control JSON event fires (which arrives a few seconds
  after connect, after the initial `DeviceEvent::Connected`).
  `write_session_meta()` in `shared.rs` now writes both fields.

- **LSL dashboard integration**: The dashboard now shows connected state,
  signal quality, live EEG channel values, band powers, and waveforms for
  LSL streams — previously only BLE devices updated the dashboard.
- **Generic device badge**: Replaced 6 device-specific badge branches with
  a single badge showing channel count, sample rate, and transport type.
- **Collapsible signal quality**: For high channel counts (>8), the quality
  grid collapses behind a summary row (e.g. "32✓ 0~") to save space.
- **LSL fast resolve**: Named LSL streams now connect in ~500 ms instead of
  waiting the full 5-second discovery timeout.
- **Dynamic channel support**: EEG waveforms, band power chart, and channel
  grid now support any channel count (2–1024+) with a 64-channel rendering
  cap for performance.
- **Disconnected view**: Added "LSL / Settings" button alongside "Scan for
  Device" for quick access to LSL configuration.

- **Paired devices persisted by the daemon** (`paired_devices.json`):
  `pair_device` and `forget_device` write a dedicated
  `~/.skill/paired_devices.json` on every change.  The file is written
  atomically (temp-file + rename) so a crash mid-write never corrupts it.
  The daemon reloads it at startup to populate `status.paired_devices`
  before the first scanner tick, meaning paired devices survive restarts
  without Tauri needing to re-sync them.  `settings.json` is also kept in
  sync via a background task for backward compatibility with older builds and
  the Tauri side.  `skill-constants::PAIRED_DEVICES_FILE` is the shared
  constant used by both daemon and Tauri.

- **Tauri reads `paired_devices.json` directly at startup**:
  `load_and_apply_settings` now prefers `paired_devices.json`
  (daemon-authoritative) over `settings.json` (which may lag by the async
  write).  Paired devices appear correctly in the UI on the very first render
  frame, before the daemon status poll completes.

- **Muse firmware version surfaced at runtime**: the session runner now
  handles `DeviceEvent::Meta` (previously silently discarded in the `_ => {}`
  branch).  Muse Control JSON responses contain a `"fw"` field (e.g.
  `"3.4.5"`); when one arrives, `status.firmware_version` is updated and a
  `StatusUpdate` WebSocket event is broadcast to connected clients.

### Bugfixes

- **Ctrl+C now reliably terminates both daemon and Tauri from the TUI**: switched TUI child processes to process groups and implemented graceful shutdown (`SIGTERM`) with forced fallback (`SIGKILL`) so orphaned dev processes are not left running.

- **Daemon pane shutdown handling fixed**: replaced blocking `execSync` in TUI daemon pane mode with `spawn` + signal-aware exit handling, so termination signals propagate correctly.

- **Muse (and all BLE devices) stuck at "Scanning for device…" forever**:
  when a connection attempt launched its own `CBCentralManager` scan while
  `run_ble_listener_task` was already holding a concurrent
  `CBCentralManager` scan, macOS suppressed the
  `centralManager(_:didConnect:)` delegate callback, causing
  `peripheral.connect()` to hang regardless of the timeout.  Fixed with a
  `ble_scan_paused` `AtomicBool` in `AppState`: the listener task stops its
  scan and parks whenever the flag is set; `connect_device()` sets the flag
  before delegating to the actual connect function and clears it
  unconditionally on return (success or failure).

- **`needs_ble_pause` covers all BLE-scanning device families**: Muse,
  MW75 Neuro, Hermes V1, Mendi fNIRS, IDUN Guardian (`idun`/`guardian`/
  `ige`), and OpenBCI Ganglion were all affected by the two-concurrent-scan
  bug.  The pause logic is now centralised in `connect_device()` rather than
  duplicated in each connect function.

- **Stale `ble_scan_paused` flag after scanner restart**: if a connection
  was in progress when the scanner was stopped, the flag could be left `true`
  and the freshly spawned listener task would stall indefinitely.
  `control_scanner_start()` now clears the flag unconditionally.

- **BLE discovery stops after pressing Cancel mid-connect**:
  `control_cancel_retry()` cancelled the session handle but did not clear
  `ble_scan_paused`.  The BLE listener remained parked with its scan stopped,
  so no BLE devices appeared in the Discovered list until the user either
  restarted the scanner or started another connection attempt.  Fixed:
  `control_cancel_retry()` now clears the flag immediately.

- **BLE event-loop timeout reduced from 2 s to 300 ms** so the
  `ble_scan_paused` flag is detected quickly; the corresponding settling
  pause was reduced from 600 ms to 400 ms.

- **BLE cache TTL extended from 60 s to 120 s** to cover the worst-case
  connection window (400 ms pause + 5 s scan + 10 s connect + 15 s service
  discovery) without a cached device expiring mid-attempt.

- **Daemon log `TeeWriter` never committed to ring buffer**:
  `tracing-subscriber` calls `write_all()` once per event but never calls
  `flush()`.  The original implementation only committed lines in `flush()`,
  so the buffer was always empty.  Fixed: `write()` now detects the trailing
  `\n` that marks a complete tracing event and commits immediately; `flush()`
  is kept as a fallback for partial lines.

- **Muse target UUID ignored on connect**: `connect_muse()` previously
  scanned for the first Muse in range regardless of which device was paired.
  In multi-headset environments the wrong device could be selected.  The BLE
  UUID from `ble:<uuid>` targets is now used to filter `scan_all()` results
  in the slow-path fallback.

- **Muse connected but no EEG data streamed**: daemon Muse connect paths now
  explicitly start the stream after BLE/GATT connect by calling
  `MuseHandle::start(true, false)` and then best-effort
  `request_device_info()`.  Previously the session could report
  "connected" while receiving zero samples (flat waveforms, CSV header only).

- **Retry connect could silently no-op** when `status.target_name` was empty.
  Daemon retry target resolution now uses paired-first policy and falls back
  through: `target_id` → legacy `target_name` → preferred paired discovered
  device → first paired device.

- **Auto-connect fallback improved for continuous recording**:
  when the default/previous target is unavailable, retry falls back to the
  next paired device that is currently available. If no valid target is set
  but a paired device is available, reconnect auto-selects that paired target.

- **Continuous reconnect loop on startup**:
  after an initial failed connect, auto-reconnect now keeps trying every
  **3 seconds** (instead of stopping after one failure). The dashboard shows
  the countdown timer, users can cancel reconnect, and can resume by pressing
  connect/retry again. The cadence is now centralized in a single constant
  (`AUTO_RECONNECT_CADENCE_SECS`) to keep runtime behavior/tests/docs aligned.

- **Connecting target naming is now daemon-authoritative**:
  `StatusResponse` gained canonical fields:
  - `target_id` (stable device id like `ble:...`)
  - `target_display_name` (human-readable paired name)

  Session control and routing paths (`start/switch/retry/spawn`) now set these
  fields from paired metadata, instead of forcing UI heuristics.

- **Potential daemon deadlock in connect status updates fixed**:
  target-field resolution previously re-locked `status` while already holding
  `status.lock()`.  Resolution is now performed outside lock scopes.

- **Pair-before-connect enforcement**: daemon start/switch session control now
  rejects unpaired scanner targets (`ble:*`, `usb:*`, `wifi:*`, `cortex:*`,
  etc.) with a clear user-facing error, preventing accidental connects to
  non-paired devices.

- **Defense-in-depth pairing guard**: session connect path also enforces the
  same paired-only policy before attempting transport-specific connections.

- **Settings → Devices paired list could appear empty** if paired devices were
  not currently in the discovered list.  Devices tab now reads authoritative
  `status.paired_devices`, merges them into local rows, and keeps them visible
  even when not actively advertising.

- **Manual-connect hints were mixed into live discovered hardware** in
  Settings → Devices.  The built-in `neurosky` and
  `brainvision:127.0.0.1:51244` rows are now rendered under a dedicated
  **Manual connection hints** subsection instead of the normal discovered
  hardware list.

- **i18n cleanup**: the new manual-hints subsection uses translation keys
  (`devices.manualHints`, `devices.manualHintsHint`) instead of hardcoded
  English text.

- **Recording-safety alerts** now run in both dashboard and backend poller:
  - low battery warning notification (once per connected device),
  - signal quality degradation warning when EEG quality drops from good to
    sustained poor/no-signal during an active recording stream.

- **Alert i18n**: dashboard alert titles/bodies (low battery + signal drop)
  now use translation keys (`alerts.*`) instead of hardcoded English text.

- **Dashboard pairing UX**: disconnected dashboard now shows nearby unpaired
  compatible devices and supports one-click **Pair** (which immediately sets
  default and starts connect) so EXG streaming can start without navigating to
  Settings → Devices.

- **Dashboard quick actions**: paired-device rows now include a dedicated
  **Set default** action (without immediate connect) and an **open settings**
  shortcut for advanced device management.

- **Tray/dashboard connect labels now prefer canonical fields**:
  runtime rendering uses `target_display_name` + `target_id` instead of legacy
  `target_name` fallbacks.

- **Runtime e2e schema snapshot updated** for the new daemon status fields
  (`target_id`, `target_display_name`).

- **Dead `hardware_version === "p21"` check in `isMuse2` detection**:
  `"p21"` is the Classic Muse startup preset command string, not a
  hardware version identifier.  The Muse adapter never sets
  `hardware_version`, so this condition was permanently `false` dead code.
  Removed; `isMuse2` now relies solely on the advertising name containing
  `"muse-2"` or `"muse 2"`.

- **Unreachable `_ => {}` arm in session runner event loop**: after adding
  the `DeviceEvent::Meta` handler, all seven `DeviceEvent` variants were
  explicitly matched, making the catch-all arm unreachable (Rust warned
  about it).  Removed.

- **Dashboard showed "DISCONNECTED" for daemon-managed sessions**: The Tauri
  status mirror was overwriting the daemon's authoritative state.  Introduced
  `emit_status_from_daemon()` to prevent mirror-back when data originates
  from the daemon.
- **Tray icon never updated for LSL sessions**: `refresh_tray()` was not
  called from the daemon status poll path.
- **LSL connect button re-enabled instantly**: The button now stays disabled
  with a "Connecting…" spinner until the daemon confirms the session state.
- **Band chart empty for >12 channels**: Fixed hardcoded `MAX_CH=12` buffer
  limit in BandChart; now dynamically sized.
- **EEG waveform capped at 12 channels**: Frontend `EEG_CHANNELS` constant
  updated from 12 to 32; chart buffers now sized from the actual channel
  count prop.
- **`StatusResponse` missing device fields**: Expanded with channel names,
  sample rate, signal quality, device identity, and capability flags so the
  daemon can fully describe the connected device.

- **`lsl_e2e` test ran for 30 s instead of ~8 s**: the receive loop had a
  30-second deadline that was always hit in full because the final 32 samples
  (1 LSL chunk) were not delivered within the expected window.  The test now
  runs in ~8 s.

- **`cargo test` broken by frontend test asserting dead behaviour**: the
  Vitest test `"detects Muse S Athena by hw version"` called
  `museImage("Muse-S", "p50")` and expected the Athena image path.  After
  removing the dead `hw === "p50"` check the test failed.  The test is
  rewritten to assert the CORRECT behaviour: `museImage("MuseS-F921")` (real
  Athena advertising name) → Athena image; `museImage("Muse-S", "p50")`
  (dead hw check) → Classic Muse S gen1 image.

- **`cancel_retry` left `ble_scan_paused` set**: `control_cancel_retry()`
  cancelled the session handle but did not clear `ble_scan_paused`.  After
  pressing Cancel mid-connection the BLE listener remained parked with its
  scan stopped, so no BLE devices appeared in the Discovered list until the
  next connection attempt or scanner restart.  Fixed: `control_cancel_retry`
  now unconditionally clears the flag.

- **Paired devices showed as unpaired after daemon restart**: the scanner
  tick's merge loop copied `is_paired` from `old` (previous in-memory device
  list), which is empty on the first tick after a restart.  Every
  re-discovered device was therefore marked `is_paired = false`.  Fixed: the
  merge builds a `paired_ids` set from `status.paired_devices` (authoritative,
  restored from disk) and checks it first; `old` is used only as a fallback
  for the `is_preferred` flag, which is not persisted.

- **`firmware_version` not synced from daemon to Tauri**: `apply_daemon_status()`
  copied `hardware_version` from `StatusResponse` but silently skipped
  `firmware_version`.  After the new Meta handler sets it, Tauri now receives
  it via the status poll.

- **Dead `hardware_version === "p50"` Athena detection**: both
  `devices-logic.ts` and `+page.svelte` checked `hw === "p50"` /
  `hardware_version === "p50"` to identify the Muse S Athena firmware.
  `p50` is a BLE preset command string, not a hardware version identifier,
  and the Muse adapter never sets `hardware_version`.  The condition was
  permanently `false` dead code.  Removed; the correct name-based detection
  (`"muses"` substring in the advertising name) is now the sole check.
  The corresponding Vitest test was updated to document the correct
  behaviour (`"Muse-S"` + `hw="p50"` → Classic gen1 image; `"MuseS-F921"`
  with no hw arg → Athena image).

- **Non-atomic JSON writes could leave corrupt files**: `persist_paired_devices()`
  previously used `std::fs::write()` directly for both `paired_devices.json`
  and the `settings.json` sync.  A crash or OS signal between truncation and
  completion would leave a zero-byte or partially-written file.  Both writes
  now use `write_json_atomic()` (write to `.tmp` sibling, then `rename()`),
  which is atomic on POSIX systems.

- **Catch daemon HTTP negative paths in frontend client**: added Vitest coverage for bootstrap protocol mismatch, non-2xx error propagation, and `{ ok: false }` payload failures in `src/lib/daemon/http.ts`.
- **Verify invoke-proxy fallback behavior**: added runtime tests for `daemonInvoke` fallback to Tauri `invoke` on daemon HTTP failures and unknown commands.

- **`cargo test` failed with 37+ compile errors** (invisible to
  `cargo check -p skill-daemon` because the test profile compiles
  `skill-iroh` as a direct dependency):

  - **`skill-iroh`** (`auth.rs`, `commands.rs`, `scope.rs`, `tunnel.rs`,
    `device_receiver.rs`, `device_proto.rs`): all functions returning
    `anyhow::Result` used `ok_or_else(|| "…".to_string())?`,
    `Err(format!(…))`, and `.context(…)` without `use anyhow::Context`.
    `String` does not implement `std::error::Error` so `?` could not convert
    it.  Fixed by replacing all string-error patterns with
    `anyhow::anyhow!(…)` / `anyhow::bail!`, adding
    `use anyhow::Context as _` where `.context()` is used, and changing
    `Ok::<(), String>(())` closures in `tunnel.rs` to
    `Ok::<(), anyhow::Error>(())`.

  - **`skill-daemon/src/session/connect.rs`**: same `ok_or("…")?` and
    `map_err(|e| format!(…))?` patterns throughout all `connect_*`
    functions.  Replaced with `anyhow` equivalents.

  - **`skill-daemon/src/session_runner.rs`**: test assertions called
    `.contains("…")` on `anyhow::Error` (which has no such method).
    Updated to `.to_string().contains("…")`.

  - **`skill-daemon/src/service_installer.rs`**: `#[allow(unreachable_code)]`
    applied to `anyhow::bail!` macro invocations (attributes don't apply to
    macros); caused `unused_attributes` warnings and `unreachable_expression`
    on macOS where the preceding `return` always fires.  Fixed by gating
    the fallback `bail!` behind
    `#[cfg(not(any(target_os = "macos", target_os = "linux", target_os = "windows")))]`.

  - **`skill-iroh` / `iroh_test_client` / `iroh-example-client` test and
    example code**: `serde_json::json!({"error": e})` where `e` is
    `anyhow::Error` (not `Serialize`); `err.contains(…)` on `anyhow::Error`.
    Fixed with `.to_string()` at each call site.

  - **`skill-headless/src/engine.rs`**: used `anyhow::Result` in the
    `ExternalRendererFn` type and `external_fetch_page` without `anyhow`
    in `Cargo.toml`.  Added `anyhow = { workspace = true }` to
    `skill-headless` and `use anyhow;` in `engine.rs`.

  - **`skill-lsl/src/virtual_source.rs`**: `use anyhow::Context as _` was
    placed inside the module doc-comment block, making it an inner doc
    comment rather than a use declaration.  Moved above the doc block.

- **`setup_app` return type incompatible with Tauri's `.setup()` closure**:
  the function returned `anyhow::Result<()>` but Tauri expects
  `Result<(), Box<dyn std::error::Error>>`.  Kept the function signature as
  `anyhow::Result<()>` (preserving `.context()` chaining throughout the
  body) and adapted the single call site with `.map_err(Into::into)`.

- **Protect EXG route extraction with tests**: added smoke + edge coverage for extracted EXG model routes (config/status/catalog/reembed/rebuild/estimate, duplicate-download rejection, cancel flag handling) to prevent regressions during modularization.
- **Harden runtime/dispatch/connect contracts**: added command-matrix coverage in `cmd_dispatch`, deterministic and overlap checks for connect route selection, and additional LLM runtime state transition tests to improve reliability under refactors.

### Refactor

- **`is_known_eeg_ble_name` unit tests** (`skill-daemon`): 11 cases covering
  all accepted EEG device families (Muse, Ganglion, MW75, Hermes, Mendi,
  IDUN, BrainBit, Unicorn) and 5 rejection cases (JBL, Apple Watch, iPhone,
  AirPods, empty string, anonymous UUID names).  `tempfile` added as a
  `[dev-dependencies]` entry for `skill-daemon`.

- **`write_json_atomic` round-trip test**: verifies the file is written with
  the correct content and no stray `.tmp` file is left behind.

- **Central `apply_daemon_status()` helper**: Replaced 4 copy-paste
  cherry-pick sites with a single function that maps all `StatusResponse`
  fields to `DeviceStatus`.
- **`StatusResponse::clear_device()`**: Clean disconnect reset instead of
  manual field clearing.
- **BLE-centric i18n strings**: Updated to be device-agnostic (BCI headset
  → device, added LSL references).

- **LSL E2E test redesigned for correctness and speed**:

  - **Root cause of 30-second test duration**: the rlsl TCP session thread
    checks `shutdown` at the *top* of its loop; `Drop for StreamOutlet` sets
    `shutdown = true` and pushes a sentinel, which causes the thread to exit
    immediately and discard any samples still in `chunk_buf` — reliably losing
    the last 32-sample chunk on every run.  The outlet is now kept alive in
    the push thread until the async receive loop finishes by sending it through
    a rendezvous channel; the async side drops it only after the drain pass.

  - **Receive deadline tightened**: 30 s → 8 s (5 s data + 3 s headroom).
    The old deadline was long enough to mask all timing issues silently.

  - **Sample-loss threshold tightened**: minimum accepted sample count raised
    from `TOTAL_SAMPLES / 2` (50%) to `TOTAL_SAMPLES * 95 / 100` (95%), so a
    lost chunk now actually fails the test.

  - **Post-loop drain pass**: a 500 ms drain loop after the timed receive
    window collects any EEG frames that landed in the adapter channel exactly
    as the deadline fired, ensuring they are counted in the final tally.

- **Standardise error handling on `anyhow` across the workspace**: all
  internal library functions that previously returned `Result<T, String>`
  migrated to `anyhow::Result<T>`, enabling `?` propagation, structured
  context via `.context()`, and `anyhow::bail!` at error sites.  Affected
  crates: `skill-iroh` (auth, tunnel, scope, device-proto, device-receiver,
  commands), `skill-daemon` (service installer, session runner, `Pipeline::open`,
  all `connect_*` device functions, `routes/labels::open_labels_db`,
  `cortex_probe_headsets`).  Tauri `#[command]` handlers and WebSocket
  dispatch functions in `cmd_dispatch.rs` intentionally retain
  `Result<T, String>` at their serialisation boundaries.

- **Workspace-pinned `anyhow` in all crates**: `crates/skill-headless` and
  `crates/skill-lsl` were missing the `anyhow` dependency entirely.  Added
  `anyhow = { workspace = true }` so all crates share the same resolved
  version.

- **Modularize daemon settings routes by domain**: split `crates/skill-daemon/src/routes/settings.rs` into focused modules (`settings_io`, `settings_lsl`, `settings_llm`, `settings_llm_runtime`, `settings_llm_chat`, `settings_exg`, `settings_hooks_activity`) while preserving existing route paths and handler behavior.
- **Compose settings router from subrouters**: introduced grouped route composition (`exg_routes`, `llm_routes`, `lsl_routes`) to reduce `settings.rs` complexity and make route ownership auditable.
- **Simplify command dispatch structure**: refactored `crates/skill-daemon/src/cmd_dispatch.rs` to use grouped family dispatch for hooks/sleep-schedule/dnd/iroh/llm commands, reducing top-level `match` size.
- **Make device connector routing table-driven**: updated `session/connect.rs` to use an explicit `ConnectRoute` selector path and added route-selection coverage for aliases/prefixes.

### Build

- **Ignore dev log artifacts**: added `logs/` to `.gitignore` so persisted TUI session logs are not accidentally committed.

- **Add audit script command**: added `npm run audit:daemon-routes` to regenerate the Tauri→daemon route audit report.

### UI

- **Custom dev TUI replaces tmux for `npm run tauri dev`**: removed tmux-based pane orchestration and added a built-in terminal UI (`scripts/tauri-dev-tui.js`) that shows a pink NeuroSkill banner, daemon logs on the left, and Tauri logs on the right.

- **Improved TUI usability**: added active-pane highlighting, footer status, per-pane follow mode, reduced flicker via throttled rendering, and richer key help (`Tab`, `j/k`, arrows, `PgUp/PgDn`, `g/G`, `f`, `?`, `q`).

- **Mouse wheel scroll support**: enabled SGR mouse tracking and mapped wheel events to scroll the pane under the cursor.

### Server

- **Harden Tauri→daemon API contracts end-to-end**: added runtime contract tests for `src-tauri/src/daemon_cmds.rs` that verify method/path/auth-header correctness across sync, control/LLM, and async proxy routes against a mock HTTP daemon.
- **Stabilize daemon contract test harness**: unified test locking for env-var/token-path based tests and made mock servers tolerate TCP readiness probes so parallel test execution no longer causes flaky failures.

- **Keep Tauri daemon proxy behavior stable during cleanup**: reduced boilerplate in `src-tauri/src/daemon_cmds.rs` by adding shared async proxy helpers while preserving existing endpoint contracts.

### Docs

- **Add command-to-route coverage audit**: introduced `scripts/audit-daemon-routes.js` and generated `docs/testing/tauri-daemon-route-audit.md` with full `daemonInvoke` command → method → path matrix and summary counts.

## [0.0.86] — 2026-04-06

### Features

- **Daemon CLI command dispatch**: universal JSON command dispatcher mapping ~70 CLI commands to daemon REST handler logic. `POST /` root tunnel and `POST /v1/cmd` authenticated endpoints. WebSocket handler upgraded to bidirectional — dispatches incoming JSON commands alongside broadcasting events.
- **CLI daemon connectivity**: auto-discovers daemon port 18444, loads auth token from `~/.config/skill/daemon/auth.token`, includes Bearer token in WS and HTTP requests.
- **Full session recording pipeline**: CSV and Parquet recording via `SessionWriter`, `BandAnalyzer` DSP computing band power at ~4 Hz (delta through gamma, FAA, TAR, coherence, entropy, 30+ derived metrics), `EegBands` WS events broadcast, session metadata JSON sidecar, epoch metrics stored in `eeg.sqlite`.
- **EXG embedding pipeline**: sliding-window 5s epoch accumulator with resampling, per-day HNSW + SQLite store, background embed worker thread with 9 encoder backends (ZUNA, LUNA, REVE, OSF, SleepFM, SleepLM, ST-EEGFormer, TRIBEv2, NeuroRVQ). All enabled by default via `embed-exg` feature flag.
- **Hook triggers**: `HookMatcher` with fastembed BGE-small-en-v1.5 for keyword→vector, label index HNSW search, cosine distance against live EEG embeddings, scenario filtering, rate limiting, audit log in `hooks.sqlite`.
- **LLM streaming over WebSocket**: incremental delta tokens via mpsc bridge, protocol matches CLI expectations (session → delta* → done). LLM inference enabled by default.
- **Generic adapter session runner**: single `run_adapter_session` function drives any `Box<dyn DeviceAdapter>` through the full daemon pipeline. Replaced device-specific event loops.
- **Enriched band snapshots**: focus, relaxation, engagement composite scores computed from raw band power data, matching the old Tauri session runner formulas.
- **Parquet storage support**: `SessionWriter` dispatches to CSV, Parquet, or both based on user's `storage_format` setting. IMU frames also recorded.

- **NeuroField Q21 support**: 20-channel FDA-approved EEG amplifier via PCAN-USB (CAN bus). Scanner probes online PCAN interfaces, session runner with blocking reader thread, standard 10-20 electrode names (F7, T3, T4, T5, T6, Cz, Fz, Pz, F3, C4, C3, P4, P3, O2, O1, F8, F4, Fp1, Fp2, HR). Crate: `neurofield 0.0.1`.
- **BrainBit support**: BrainBit, BrainBit 2, Pro, Flex 4/8 EEG headbands via NeuroSDK2 BLE. 4 channels (O1, O2, T3, T4) at 250 Hz. Callback-based streaming via `on_signal()`. Crate: `brainbit 0.0.1`.
- **g.tec Unicorn Hybrid Black support**: 8-channel EEG headset via Unicorn C API (BLE). 250 Hz, blocking `get_single_scan()` reader thread. Crate: `gtec 0.0.2`.
- **Restored all BLE devices**: Muse, MW75 Neuro, Hermes V1, IDUN Guardian, Mendi fNIRS — all re-wired through the generic adapter session runner with full pipeline (CSV/Parquet, DSP, embeddings, hooks).
- **Restored Emotiv**: EPOC X, Insight, Flex, MN8 via Cortex WebSocket API.
- **Restored Cognionics CGX**: Quick-20r and other CGX headsets via USB serial.
- **LSL stream sessions**: `connect_lsl()` resolves EEG streams, creates `LslAdapter`, feeds into generic pipeline. Discovery was working; session recording was broken and is now fixed.

### Bugfixes

- **Windows COM port ≥ 10**: added `\\.\COMxx` prefix normalization for COM10+ ports that otherwise fail with "file not found".
- **FTDI dongle auto-detection**: 3-pass serial port detection (VID/PID → product/manufacturer string → heuristic fallback) catches dongles that Windows reports without USB metadata.
- **Hot-plug retry**: 3 attempts with 1.5s/3s backoff delays when the serial port is temporarily locked after USB replug.
- **Device ID → session start**: `start_session` now accepts `usb:COM3`, `cgx:*`, `neurofield:*`, `brainbit:*`, `gtec:*`, `lsl:*` device IDs and routes to the correct connect function.
- **LLM inference disabled by default**: daemon `Cargo.toml` had `default = []` — all `#[cfg(feature = "llm")]` code was dead. Fixed: `default = ["llm", "embed-exg"]`.
- **LSL sessions broken**: `lsl:` targets were not handled in `connect_device`. Now wired through `connect_lsl()` with stream resolution and `LslAdapter`.

### Docs

- **DEVICES.md**: complete device support matrix — 13 device families, 19 hardware variants, 3 virtual/network sources. Transport summary, platform support table, device ID format reference, guide for adding new devices.

## [0.0.85] — 2026-04-03

### Docs

- **Add changelog fragment scaffold**: create an unreleased entry so `changes/unreleased` is no longer empty.

## [0.0.84] — 2026-04-03

### Bugfixes

- **Windows CI lockfile stability**: pinned `Cargo.lock` to prevent intermittent Windows CI breakage.

### Docs

- **Release archive correction**: rebuilt archived release notes to replace previously empty entries with history-derived notes.

## [0.0.83] — 2026-04-03

### Build

- **Prebuilt llama fallback in CI/release**: added prebuilt llama artifact usage with safe source-build fallback in CI and mac release flows.

### Bugfixes

- **iroh-remote reconnect fixes**: corrected reconnect routing and prevented unwanted BLE reconnect loops during iroh-remote session startup.

## [0.0.82] — 2026-04-02

### Build

- **Linux release packaging fix**: stopped requiring removed espeak-ng resource bundle in Linux release path.

### CLI

- **Changed-crates detection hardening**: fixed metadata parsing with safer fallback behavior for crate change detection.

## [0.0.81] — 2026-04-02

### Features

- **EXG/LLM model catalog expansion**: added NeuroRVQ integration and restored Bonsai model options with related inference controls.
- **LSL and EXG UX updates**: added LSL idle-timeout configurability, virtual test source controls, and channel inspector/stream handling improvements.
- **Inference/device settings growth**: expanded CPU/GPU preference and runtime tuning paths for llama-cpp-based inference.

### Bugfixes

- **Updater packaging fixes**: corrected ZIP compression behavior across Windows/desktop updater flows.
- **Location + backend stability**: fixed location-enable hangs and improved Neutts backend init/lifetime reliability.
- **Dependency/runtime fixes**: resolved llama-cpp source/version mismatches and related runtime integration issues.

### Dependencies

- **Dependency alignment**: moved to newer llama-cpp stack revisions and fixed metadata/licensing fields required by dependency checks.

### Docs

- **Backfilled release notes**: reconstructed from Git history because archived fragment entries were missing for this version.

## [0.0.80] — 2026-03-30

### Build

- **Release metadata update**: version/tag release step (`v0.0.80`) with no additional feature-level deltas beyond changes already landed before tagging.

### Docs

- **Backfilled release notes**: reconstructed from Git history because archived fragment entries were missing for this version.

## [0.0.79] — 2026-03-30

### Test

- **Playwright E2E expansion**: added broad E2E coverage for static pages plus `/chat`, `/settings`, `/search`, `/compare`, calibration/labels/API flows.

### Build

- **CI release workflow cleanup**: removed stale espeak C build steps and fixed invalid empty `env:` blocks in release-linux workflow.

### Docs

- **Backfilled release notes**: reconstructed from Git history because archived fragment entries were missing for this version.

## [0.0.78] — 2026-03-30

### Features

- **Location permission + gating flow**: added `location_enabled` setting with permission handling and LLM tool gating.
- **Health diagnostics**: introduced deeper `npm run health`/session integrity inspection support.

### Build

- **TTS/espeak build simplification**: removed legacy espeak C build infrastructure and aligned build scripts with newer Neutts/Kittentts flow.

### Bugfixes

- **Stability pass**: resolved clippy + svelte-check issues and fixed dependabot/CI breakages impacting release readiness.

### Docs

- **Backfilled release notes**: reconstructed from Git history because archived fragment entries were missing for this version.

## [0.0.77] — 2026-03-30

### Features

- **History map pipeline**: added unified history aggregation for GPS/HNSW/embeddings and PMTiles-based map rendering in history views.
- **GPS support path hardening**: enabled GPS sampling pipeline with feature gating and improved location tooling docs/CLI wiring.

### Bugfixes

- **History reliability fixes**: unified session listing/day-boundary behavior and fixed stale cache/deserialization issues causing empty or inconsistent history views.

### Docs

- **Backfilled release notes**: reconstructed from Git history because archived fragment entries were missing for this version.

## [0.0.76] — 2026-03-30
### Features

- **Oura Ring V2 integration**: New `skill-oura` crate fetches sleep, activity, readiness, heart rate, SpO2, workouts, and mindfulness data from the Oura Cloud API and stores it in the unified `health.sqlite` pipeline alongside Apple HealthKit data. Includes `oura_sync` and `oura_status` WS/HTTP commands, CLI `oura` command with `sync` and `status` subcommands, REST endpoints (`/v1/oura/sync`, `/v1/oura/status`), token management via system keychain, Settings UI with token input + Sync Now button, history view health overlay card showing Oura sleep/readiness/activity scores, 11 Rust tests with full round-trip integration test, E2E tests in test.ts, i18n in 6 languages, supported devices catalog entry, and comprehensive SKILL.md documentation.

## [0.0.75] — 2026-03-29

### Features

- **187 new Rust tests** across 17 crates covering: session history listing/deletion/stats, CSV timestamp parsing, screenshot store (search, OCR, embeddings, pagination, app grouping), bash/path safety checks, tool output truncation, LLM argument coercion (bool/number/string/null/object/array), tool-call extraction and stripping, path resolution and retry logic, DOT graph generation, LLM config serde, screenshot config interval logic, catalog persistence, and tool orchestration (stream sanitizer, result summarization, context condensation).

### Bugfixes

- **Fixed iroh LSL sink hardcoded descriptor**: `IrohLslAdapter` was hardcoding 4 channels at 256 Hz regardless of the actual remote stream. The DSP pipeline, CSV writer, and embedding encoder now use the correct channel count, sample rate, and channel labels discovered from the connected source. Added 120-second timeout on stream resolution (was infinite loop) and two-phase start API so the endpoint ID is available immediately.
- **Fixed `skill-data` screenshot_store tests**: added missing `source`, `chat_session_id`, and `caption` fields to test helper after struct update.
- **Fixed `skill-jobs` queue test**: captured second `submit()` return value and used correct closure variable in `sequential_execution` test.
- **Fixed `skill-llm` E2E test model selection**: `best_test_model()` now filters out mmproj filenames, preventing selection of a vision projector (0.54 GB) instead of the actual language model (1.16 GB).
- **Fixed `skill-lsl` test warnings**: removed unused import and shadowed variable bindings.

### Refactor

- **Exported `coerce_value`** from `skill-tools::parse` for external test access.
- **Added `rustls` workspace dep** with `ring` backend to avoid `aws-lc-rs` where possible.

### Build

- **Tiered Rust test runner** (`scripts/test-fast.sh`): split workspace tests into 3 tiers by compilation cost — tier 1 runs ~400 tests in 5 s (warm) vs 80 s for the full workspace. Added `npm run test:rust` and `npm run test:rust:all` scripts.
- **Disabled `jsonschema` default features** in `skill-tools`: removed transitive `reqwest` → `rustls` → `aws-lc-sys` dependency (45 s build) that was only used for remote JSON schema fetching we never use.

### Docs

- **Updated `CONTRIBUTING.md`** with tiered test runner documentation and timing table.
- **Added `docs/TEST-COVERAGE.md`**: detailed per-crate coverage analysis with prioritized gap list.

## [0.0.72] — 2026-03-25

### Features

- **Calendar event fetching**: New `skill-calendar` crate adds cross-platform OS calendar support.
  - **macOS**: Reads all calendars via Apple EventKit (`EKEventStore`) using Objective-C FFI — covers iCloud, Google, Exchange, and local calendars synced to Calendar.app. Handles the macOS 14+ `requestFullAccessToEventsWithCompletion:` API (requires `NSCalendarsFullAccessUsageDescription`) with fallback to the legacy API on macOS 10.15–13.
  - **Linux**: Scans XDG locations for `.ics` files: GNOME Calendar, Evolution, KOrganizer, Thunderbird Lightning (targeted `calendar-data/` subdir only — not the full mail profile), and `~/Calendars/`.
  - **Windows**: Scans Outlook / Windows Calendar paths under `%APPDATA%`, `%LOCALAPPDATA%`, and UWP package directories for `.ics` files.
  - Shared RFC 5545 iCal parser: line folding (CRLF/LF/tab), `VTIMEZONE` UTC-offset extraction, `VALUE=DATE` all-day events, UTC (`Z`) timestamps, iCal escape sequences (`\n`, `\,`, `\;`, `\\`), `X-WR-CALNAME` (Google Calendar name at VCALENDAR level), recurrence rule passthrough. 46 unit tests.
  - **WS commands**: `calendar_events` (fetch by range), `calendar_status` (auth state + platform), `calendar_request_permission` (macOS system dialog). Both potentially-blocking commands run via `spawn_blocking`.
  - **HTTP REST**: `POST /v1/calendar/events`, `GET /v1/calendar/status`, `POST /v1/calendar/permission` (+ unversioned aliases).
  - **CLI**: `calendar [--start --end]`, `calendar status`, `calendar permission`.
  - **LLM tools**: `calendar_events`, `calendar_status`, `calendar_request_permission` in the `skill` tool enum and `is_skill_api_command` registry. `"calendar"` alias in `resolve_skill_alias`.
  - **Skill markdown**: `skills/skills/neuroskill-calendar/SKILL.md` with LLM tool examples, timestamp arithmetic, common query patterns, and access troubleshooting.
  - **macOS entitlements**: `com.apple.security.personal-information.calendars` added to `entitlements.plist`; `NSCalendarsUsageDescription` and `NSCalendarsFullAccessUsageDescription` added to `Info.plist`.
  - `calendar_events`, `calendar_status`, `calendar_request_permission` added to `skill-router::COMMANDS` public registry.

### Bugfixes

- **Calendar Linux dedup (critical)**: `linux.rs` two-pass deduplication silently dropped every event with a non-empty UID — the first loop pre-inserted all UIDs into `seen`, causing the second loop's `!seen.contains` check to always be false. Both `linux.rs` and `windows.rs` now use single-pass atomic check-and-insert. Anonymous-event keys use NUL separator (`"\0"`) to prevent hash collisions.
- **Calendar `X-WR-CALNAME` scope**: Google Calendar exports place the calendar name at the `VCALENDAR` level, not inside `VEVENT`. The property was only scanned inside `VEVENT` and therefore never populated. Fixed with a top-level pass in `parse_ical()`.
- **Calendar async blocking**: `calendar_events` and `calendar_request_permission` called blocking ObjC EventKit code directly on the tokio async executor thread (up to 30 s for the macOS permission dialog). Both are now `async` and dispatch to `spawn_blocking`.
- **Calendar macOS 14 privacy key**: `NSCalendarsFullAccessUsageDescription` was missing; without it `requestFullAccessToEventsWithCompletion:` silently fails on Sonoma/macOS 14+.
- **Calendar Thunderbird scan**: `~/.thunderbird` root was searched to depth 6, walking the entire mail profile (potentially GB of data). Now only `~/.thunderbird/<profile>/calendar-data/` is scanned.
- **Calendar EventKit error detection**: `macos.rs` used `json_str.contains("\"error\"")` to detect access-denied responses — a false positive for any event whose title or description contains the word `"error"`. Replaced with proper JSON type-based dispatch: `Array` → success, `Object` with `"error"` key → propagate error string.
- **Calendar EventKit write-only access**: When the user chose "Add Events Only" (`EKAuthorizationStatusWriteOnly`, status 4), the fetch proceeded and returned an empty array instead of an error. Now returns `{"error":"calendar_write_only_access"}` so clients can prompt the user to grant full read access.
- **LLM e2e test non-exhaustive match**: `llm_e2e.rs` match on `ToolEvent` was non-exhaustive after `RoundComplete { .. }` was added to the enum; added the missing arm.
- **`skill-calendar` unused dependency**: `anyhow` was declared in `Cargo.toml` but never used in the crate; removed.

- **Cleared all Biome lint errors** across `src/` and `scripts/`: fixed `noNonNullAssertion`, `noExplicitAny`, `useIterableCallbackReturn`, `noAssignInExpressions`, `useImportType`, and `noUnusedFunctionParameters` in ~50 files using safe rewrites or targeted `biome-ignore` suppressions with justification comments.
- **Restored blocking Biome lint in CI**: `Biome lint` step in `.github/workflows/ci.yml` is now a hard failure again (0 errors, 0 warnings).
- **Suppressed false-positive `useImportType` warnings for `.svelte` files**: added a `biome.json` override so Biome no longer flags Svelte component runtime imports as type-only — converting them to `import type` breaks `svelte-check`.
- **Fixed incorrect auto-conversion of Svelte component imports**: reverted `import type` from bits-ui and Svelte component imports that must remain runtime imports (`PromptLibrary`, `ChatInputBar`, `ChatMessageList`, `ChatSidebar`, `DialogPrimitive`, `ProgressPrimitive`, `SeparatorPrimitive`, `DialogPortal`).

- **Security audit CI**: Fixed `cargo deny` failing due to breaking schema changes in cargo-deny 0.16+ (used by `EmbarkStudios/cargo-deny-action@v2`). Removed deprecated `[advisories]` fields (`vulnerability`, `notice`), updated `unmaintained` from a lint level to the new scope value (`"workspace"`), removed deprecated `[licenses]` fields (`unlicensed`, `copyleft`), added missing allowed licenses (`Apache-2.0 WITH LLVM-exception`, `MIT-0`, `Unlicense`, `CDLA-Permissive-2.0`, `NCSA`, `LicenseRef-AI100`), fixed workspace crate exception name (`neuroskill` → `skill`) and SPDX identifier (`GPL-3.0` → `GPL-3.0-only`), added `clarify` entries for `exg`/`exg-luna` (AI100 license), `unescaper` (non-SPDX `GPL-3.0/MIT`), and ignored RUSTSEC-2024-0415 (gtk via Tauri, not directly upgradeable).

- **Fix cargo-deny CI failure**: Updated `deny.toml` for cargo-deny v0.19+ breaking schema changes — removed deprecated fields (`vulnerability`, `notice`, `unlicensed`, `copyleft`), fixed `unmaintained` scope value, corrected `GPL-3.0` → `GPL-3.0-only` in license exceptions, added missing allowed licenses (`MIT-0`, `Apache-2.0 WITH LLVM-exception`, `Unlicense`, `NCSA`, `CDLA-Permissive-2.0`, `LicenseRef-AI100`), added clarification entries for `exg`/`exg-luna` custom-licensed crates, added exceptions for new crates (`skill-calendar`, `hermes-ble`, `mw75`), and ignored the gtk-rs GTK3 unmaintained advisory (RUSTSEC-2024-0415).

- **jsonschema 0.45 compatibility in tool argument validation**: updated `skill-tools` validation error path extraction to use `ValidationError::instance_path()` so the workspace compiles and validation errors still report schema paths correctly after the dependency upgrade.

### Build

- **Faster local hooks**: enabled automatic `sccache` usage in `.githooks/pre-commit` and `.githooks/pre-push` when available, including C/C++ launcher integration for cmake-based crates.
- **Docs-only fast path**: both local hooks now skip expensive frontend/Rust checks for docs/changelog-only changes.

- **Faster pre-commit checks**: updated `.githooks/pre-commit` to run frontend checks only on changed files (`biome check` + `vitest related`) instead of full project-wide frontend checks on every commit.

- **Faster pre-push checks**: updated `.githooks/pre-push` to run changed-files scoped frontend and Rust checks by default.
- **Optional full gate**: full pre-push validation can still be forced with `PREPUSH_FULL=1 git push`.

- **CI frontend lint stability**: switched `Biome lint` in `.github/workflows/ci.yml` back to advisory (`continue-on-error: true`) so existing lint debt no longer hard-fails the frontend job while cleanup is ongoing.

### UI

- **Mixed browser User-Agent pool**: replaced the outdated and browser-mixed UA list in `skill-tools` with current-version strings spanning Chrome 133–134, Firefox 128 ESR / 135–136, Safari 17–18, and Edge 133–134 across Windows, macOS, and Linux.

### i18n

- **sync-i18n lint fix**: removed a non-null assertion in `scripts/sync-i18n.ts` when collecting extra locale keys, using an explicit `undefined` guard instead.

### Dependencies

- **Align Arrow stack in `skill-data`**: upgraded `parquet` and `arrow-array` to `58` alongside `arrow-schema` so Parquet writer schemas and `RecordBatch` types use a consistent Arrow version and compile cleanly.

- **Align Arrow stack to v58**: bumped `parquet`, `arrow-array`, and `arrow-schema` in `skill-data` all to version 58, and updated Parquet EEG column casting in `src-tauri` to use `downcast_ref::<Float64Array>()` (replacing the removed `as_primitive_opt`).

- **Block wgpu Dependabot upgrades**: added `ignore` rule in `.github/dependabot.yml` for `wgpu >= 27`. The workspace pins wgpu to 26.x via `burn-wgpu`, `gpu-fft`, `zuna-rs`, and `luna-rs`; bumping wgpu independently causes type mismatches in the `WgpuSetup` pipeline. The whole Burn/GPU stack must be upgraded together.

- **ureq 3 migration**: updated all ureq usages across `skill-tools`, `skill-llm`, `skill-exg`, `skill-screenshots`, and `skill-skills` for the ureq 2→3 breaking API changes:
  - `AgentBuilder::new()...build()` → `Agent::config_builder()...build().into()`
  - `.timeout(d)` → `.timeout_global(Some(d))`, `.timeout_read(d)` → `.timeout_recv_body(Some(d))`
  - `.set("K", v)` → `.header("K", v)`, `.send_string(s)` → `.send(s)`
  - `resp.into_string()` → `resp.into_body().read_to_string()`
  - `resp.into_json()` → `resp.into_body().read_json()`
  - `resp.into_reader()` → `resp.into_body().into_reader()`
  - `resp.status()` → `resp.status().as_u16()` (now returns `StatusCode`)
  - `resp.header("K")` → `resp.headers().get("K").and_then(|v| v.to_str().ok())`
  - `Error::Status(code, _)` → `Error::StatusCode(code)`

## [0.0.71] — 2026-03-25

### Bugfixes

- **Latest CI failures resolved locally**: fixed frontend Biome formatting drift and removed a Windows-only unnecessary raw-pointer cast in `src-tauri/src/skill_log.rs`.
- **Hook shell compatibility**: replaced `mapfile` usage in `.githooks/pre-commit` with a bash-3-compatible dedupe loop so commits work on macOS default bash.

- **Auto-bootstrap chat LLM on first start**: when starting the chat server with no downloaded text model, the app now downloads the smallest `LFM2.5-VL 1.6B` variant first and then starts the server once download completes.
- **Onboarding LLM priority aligned to bootstrap**: first-run model targeting now prefers the smallest `LFM2.5-VL 1.6B` option before other recommendations.

- **CI frontend format failures**: formatted Svelte files flagged by Biome in the `frontend-check` job (`LlmTab`, `ScreenshotsTab`, `LlmInferenceSection`, `ScreenshotPerformanceSection`).
- **Windows clippy failure**: removed an unnecessary raw-pointer cast in `src-tauri/src/skill_log.rs` when calling `SetStdHandle`, fixing `clippy::unnecessary_cast` under `-D warnings`.

- **Default LLM switched to LFM2.5 1.2B Instruct**: first-run chat/bootstrap now prefers `LFM2.5 1.2B Instruct` (Q4_K_M first) instead of `LFM2.5-VL 1.6B`.
- **Default activation on bootstrap**: when no local text model exists, the selected default is set as active immediately and downloaded automatically before server start.

- **Hard-default LLM activation**: starting the LLM server now always prefers `LFM2.5 1.2B Instruct` as the active text model.
- **Default model enforcement with existing downloads**: even if other models are already present, the app ensures the default `LFM2.5 1.2B Instruct` is selected and downloaded before starting.

- **LLM autolaunch memory guard**: before auto-downloading/auto-launching the default `LFM2.5 1.2B Instruct` model, the backend now checks hardware memory fit and blocks autolaunch when RAM/VRAM is insufficient.
- **Clear startup feedback for low-memory systems**: when memory is too tight, startup now returns a descriptive stopped-state error with required vs available memory.

- **`skill-eeg` test build with GPU feature**: gated CPU-FFT property tests behind `not(feature = "gpu")` so workspace library test runs do not fail when `gpu` is enabled.

- **Hide Windows DND helper console windows**: run `reg query` / `reg add` with `CREATE_NO_WINDOW` so background DND polling no longer flashes a terminal window every 5 seconds on Windows.

- **Add stable Windows log mirror**: create and append to `%LOCALAPPDATA%\NeuroSkill\latest.log` as a fallback mirror. If date-based session log creation fails, stderr is redirected to `latest.log` so logs are still captured.

- **Create Windows session logs reliably**: write an explicit startup log line after logger initialization so `log_<unix>.txt` is created immediately in `%LOCALAPPDATA%\NeuroSkill\YYYYMMDD`.
- **Improve Windows quit confirmation ownership**: show the quit dialog on the caller thread (instead of a detached worker) with a parent window, improving native ownership/icon behavior.

- **Capture all stderr logs on Windows**: route process `stderr` to the session log file during startup (`SetStdHandle`) and keep a file-handle fallback sink so both `skill_log!` output and generic `eprintln!` logs are written to `%LOCALAPPDATA%\NeuroSkill\YYYYMMDD\log_<unix>.txt`.

- **Added integration tests for settings/history contracts**: introduced new integration tests for `skill-settings` defaults/path behavior and `skill-history` metrics/PPG sidecar resolution (`exg_` and legacy `muse_` paths).

- **Added `skill-history` session-prefix integration coverage**: new integration tests verify that session day listing and per-day session loading accept both modern `exg_` and legacy `muse_` prefixes for JSON sidecars and orphan CSV files.

- **Added router integration tests**: new `skill-router` integration tests cover UMAP cache path/store/load round-trips and inclusive epoch-label matching behavior.
- **Added label-index integration tests**: new `skill-label-index` integration tests verify empty-window EEG mean behavior and startup index initialization via `LabelIndexState::load()`.

- **Added `skill-skills` integration tests**: new tests cover explicit-path skill loading, required-description validation diagnostics, and system-prompt formatting that omits `disable-model-invocation` skills.

### Refactor

- **Parse module test split**: moved large inline tests from `crates/skill-tools/src/parse/mod.rs` into dedicated `crates/skill-tools/src/parse/tests.rs` and kept `mod.rs` focused on module exports.

- **Simplified `LlmTab` state surface**: removed local `showAdvanced`, `apiKeyVisible`, and `ctxSizeInput` ownership from `LlmTab` and delegated these UI concerns to `LlmInferenceSection`, while keeping config persistence centralized via `saveConfig()`.

- **Slimmed down `LlmTab` orchestration layer**: `LlmTab` now delegates server, model picker, inference, and log rendering to focused child components and keeps only shared state/event wiring.

- **Reduced `LlmTab` UI state surface**: removed log-view-local state (`logFilter`, `logSearch`, `logAutoScroll`, scroll element management) from `LlmTab` and delegated it to `LlmServerLogSection` while keeping event-driven log ingestion in the parent.

- **Simplified capture-setting updates in `ScreenshotsTab`**: centralized capture field patching via a small `onUpdate` bridge passed to `ScreenshotCaptureSettingsSection`, including recommended-size adoption on backend/model changes.

- **Further slimmed `ScreenshotsTab` composition**: `ScreenshotsTab` now delegates toggle, OCR, and performance rendering to focused child components while keeping shared state and persistence logic in one place.

- **Reduced `ScreenshotsTab` template complexity**: removed in-file chart formatting/render helpers from `ScreenshotsTab` and delegated rendering to `ScreenshotPerformanceSection`, keeping tab state/update logic focused.

- **Further streamlined `ScreenshotsTab` composition**: `ScreenshotsTab` now delegates permission and privacy presentation to focused child components, keeping the tab file centered on state orchestration.

- **Further reduced `ScreenshotsTab` orchestration scope**: removed local ETA formatting/render logic from `ScreenshotsTab` and delegated re-embed presentation details to `ScreenshotReembedSection`.

- **Simplified ToolsTab composition**: `ToolsTab` now composes `SuggestSkillCta`, `AgentSkillsSection`, and `SkillsRefreshSection`, reducing in-file template complexity and local state surface.

- **Reduced `ToolsTab` to orchestration composition**: `ToolsTab` now delegates chat-tools rendering to `ChatToolsSection` and keeps only state loading/saving plus skill-refresh/skills orchestration.

- **Split `skill-settings` internal tests into a dedicated module file**: moved in-file tests from `src/lib.rs` into `src/tests.rs` to keep the primary module focused.

### Build

- **CI lint gate enabled**: made Biome lint blocking in CI by removing advisory `continue-on-error` behavior.
- **Pinned apt-cache action version**: replaced floating `awalsh128/cache-apt-pkgs-action@latest` with `@v1` in CI/release workflows.
- **Reduced CI duplication for ONNX setup**: extracted Linux ONNX Runtime installation/export logic into `scripts/install-onnxruntime-linux.sh` and wrapped it in reusable action `.github/actions/setup-onnxruntime-linux`, now used by both Linux CI jobs.
- **Reduced CI duplication for Vulkan setup**: added reusable action `.github/actions/setup-vulkan-linux` (cache + install) and adopted it in Linux CI and release workflows.
- **Reduced CI duplication for Rust bootstrap**: added reusable action `.github/actions/setup-rust-bootstrap-linux` and adopted it in Linux CI/release jobs for Rust toolchain + sccache setup.
- **Engineering health metrics in CI/release**: added per-job metrics artifacts and step summaries for Rust/frontend test durations, sccache stats, cargo compile timings upload, and Linux release binary/package sizes.
- **Automated dependency updates**: added `.github/dependabot.yml` to schedule weekly npm and Cargo dependency update PRs.
- **Repository ownership map**: added `.github/CODEOWNERS` assigning crates, frontend, CI workflows, and release scripts to `@eugenehp`.
- **Rust lint hardening (phase 1)**: promoted critical clippy lints in workspace config (`unwrap_used`, `panic`, `undocumented_unsafe_blocks`) from `warn` to `deny`.

- **Added i18n guard steps to PR build workflow**: `pr-build.yml` now runs `sync:i18n:check`, `audit:i18n:check`, and `check:i18n:critical` (de/he) after dependency install so translation drift is caught before expensive packaging/signing steps.

- **Standardized local quality gates**: expanded `.githooks/pre-commit` to run i18n checks, frontend checks (`svelte-check` + unit tests), and targeted Rust clippy/tests for touched crates.
- **Added heavy pre-push gate**: new `.githooks/pre-push` runs full frontend checks, full Rust clippy (workspace + app features), and workspace Rust library tests before push.

- **Updated CI workflows to check all locales**: both `ci.yml` and `pr-build.yml` now run `check:i18n:locales` (all non-`en` locales) instead of a de/he-only check.

- **Wired critical-locale i18n guard into CI**: frontend CI now runs `check:i18n:critical` after sync/audit checks to catch untranslated fallback markers early.

- **Expanded integration-test crate detection**: CI now includes `skill-history` and `skill-settings` in integration-test crate selection.

- **Broadened integration-test targeting in CI**: added `skill-router` and `skill-label-index` to the integration-test crate detection list.

- **Expanded integration-test crate detection in CI**: added `skill-skills` to the integration-test crate selection list in `ci.yml`.

### UI

- **Extracted LLM advanced inference settings into a dedicated component**: moved the collapsible inference/settings panel from `src/lib/LlmTab.svelte` into `src/lib/llm/LlmInferenceSection.svelte`.

- **Extracted LLM model picker into a dedicated component**: moved family selection, quant list rendering, hardware-fit badges, and vision-projector controls from `src/lib/LlmTab.svelte` into `src/lib/llm/LlmModelPickerSection.svelte`.

- **Extracted LLM server log viewer into a dedicated component**: moved filtering/search/auto-scroll log UI from `src/lib/LlmTab.svelte` into `src/lib/llm/LlmServerLogSection.svelte`.

- **Extracted screenshots capture settings panel into a dedicated component**: moved interval/image-size/quality controls and embedding backend/model selectors from `src/lib/ScreenshotsTab.svelte` into `src/lib/screenshots/ScreenshotCaptureSettingsSection.svelte`.

- **Extracted screenshots OCR panel into a dedicated component**: moved OCR engine selection, model info, and search-hint UI from `src/lib/ScreenshotsTab.svelte` into `src/lib/screenshots/ScreenshotOcrSection.svelte`.

- **Extracted screenshot pipeline performance UI into a dedicated component**: moved the large performance charts and breakdown panel from `src/lib/ScreenshotsTab.svelte` into `src/lib/screenshots/ScreenshotPerformanceSection.svelte`.

- **Extracted screenshots permission notice into a dedicated component**: moved macOS screen-recording warning/success UI into `src/lib/screenshots/ScreenshotPermissionNotice.svelte`.
- **Extracted screenshots privacy note into a dedicated component**: moved the storage/privacy footer block into `src/lib/screenshots/ScreenshotPrivacyNote.svelte`.

- **Extracted screenshots re-embed/status block into a dedicated component**: moved model-change warning, re-embed controls, progress bar, and stats from `src/lib/ScreenshotsTab.svelte` into `src/lib/screenshots/ScreenshotReembedSection.svelte`.

- **Extracted Agent Skills UI into a dedicated component**: moved the large Agent Skills block out of `src/lib/ToolsTab.svelte` into `src/lib/tools/AgentSkillsSection.svelte`, including markdown skill description rendering and license panel behavior.

- **Extracted Chat Tools settings panel into a dedicated component**: moved the large tools configuration UI from `src/lib/ToolsTab.svelte` into `src/lib/tools/ChatToolsSection.svelte`, including tool toggles, provider settings, web-cache controls, execution limits, compression, and retry settings.

- **Split large settings tabs into sub-components**: extracted reusable sections from `LlmTab`, `ToolsTab`, and `ScreenshotsTab` into `LlmServerSection`, `SkillsRefreshSection`, `SuggestSkillCta`, and `ScreenshotToggleCard` to reduce template size and improve maintainability.

### i18n

- **All-locale i18n checks**: updated `scripts/sync-i18n.ts` and `scripts/audit-i18n.ts` to discover non-English locales dynamically from `src/lib/i18n`, so CI checks automatically cover every locale instead of a hardcoded subset.

- **Replaced de/he-only fallback-marker guard with all-locale guard**: `scripts/check-critical-i18n-locales.js` now validates every discovered non-English locale directory (currently `de`, `fr`, `he`, `uk`) for `TODO: translate` fallback markers.

- **Added critical-locale translation guard for German/Hebrew**: introduced `scripts/check-critical-i18n-locales.js` and `npm run check:i18n:critical` to fail when de/he locale files contain auto-sync TODO fallback markers.

- **Stronger translation drift guard in CI**: added `npm run audit:i18n:check` to frontend CI so missing locale coverage is caught alongside sync checks.

### Docs

- **LLM bootstrap requirements documented**: added a new section to `docs/LLM.md` describing the default bootstrap model (`LFM2.5 1.2B Instruct`), approximate memory needs, and the backend autolaunch memory guard behavior.

- **Updated `TODO.md`**: removed completed item for decomposing `LlmTab`, `ToolsTab`, and `ScreenshotsTab` templates into sub-components.

## [0.0.70] — 2026-03-25

### Bugfixes

- **Windows manifest root architecture**: set `src-tauri/manifest.xml` root `assemblyIdentity.processorArchitecture` to `amd64` (instead of wildcard) so `mt.exe -validate_manifest` passes and Windows CI proceeds to Rust compilation.

## [0.0.69] — 2026-03-25

### Bugfixes

- **Windows startup side-by-side fix**: simplified and normalized the embedded `src-tauri/manifest.xml` to a schema-safe structure (compatibility, common controls, DPI settings) to avoid side-by-side startup failures after manifest embedding.

### Build

- **Validate Windows manifest in CI**: added `scripts/check_windows_manifest.py` and wired it into `ci.yml` and `release-windows.yml` so malformed `src-tauri/manifest.xml` fails fast before Windows build/release.

- **Native manifest lint on Windows CI**: added `mt.exe -validate_manifest` checks in Windows CI and release workflows to catch side-by-side manifest issues with the native Microsoft validator.

## [0.0.68] — 2026-03-24

### Performance

- **Selective crate testing in CI**: `cargo test` now only runs for crates affected by the current changeset. A new `scripts/changed-crates.sh` script computes the transitive closure of reverse dependencies from changed files, so PRs that touch a single crate skip unrelated test suites. Both unit tests (`--lib`) and integration tests (`--test '*'`) are selectively run for affected crates. Workspace-wide changes (Cargo.lock, .cargo/config.toml, ci.yml) still trigger a full test run as a safety net.

### Bugfixes

- **Stabilize CI ORT linking**: preinstall ONNX Runtime 1.23.2 from Microsoft releases on Linux/Windows CI jobs, then export `ORT_LIB_LOCATION` and `ORT_PREFER_DYNAMIC_LINK=1` so `ort-sys` does not depend on flaky `cdn.pyke.io` downloads during clippy/test runs.

- **Cache ONNX Runtime binaries in CI**: `ort-sys` downloads ~200 MB static libraries from `cdn.pyke.io` during build. Added `actions/cache` for the download directory (`~/.cache/ort.pyke.io` on Linux, `%LOCALAPPDATA%/ort.pyke.io` on Windows) across CI, release-linux, and release-windows workflows to prevent recurring build failures caused by CDN flakiness.

- **Ensure Windows release bundles ONNX Runtime DLL**: release workflow now installs ONNX Runtime 1.23.2 from Microsoft releases, exports `ORT_LIB_LOCATION` / `ORT_PREFER_DYNAMIC_LINK`, and stages `onnxruntime.dll` into release output directories before NSIS packaging.

- **Windows: switch to dynamic CRT with app-local DLL bundling**: Removed `+crt-static` which failed for native dependencies (DirectML, ONNX Runtime, Vulkan loader) that ship as pre-built `/MD` DLLs. The build now uses `/MD` (dynamic CRT) consistently, and the NSIS installer bundles `vcruntime140.dll`, `msvcp140.dll`, and related CRT DLLs alongside `skill.exe` for app-local deployment. The VC++ Redistributable download section is also enabled by default as a system-wide fallback. This eliminates the "side-by-side configuration is incorrect" errors on Windows machines without the VC++ Redistributable.

- **Fix Windows NSIS packaging strict-mode crash**: normalize the `$missingCrt` result to an array before reading `.Count` in `scripts/create-windows-nsis.ps1`. This prevents a PowerShell strict-mode failure when `Where-Object` returns `$null` (no missing CRT DLLs).

### Refactor

- **Normalized LLM catalog format**: Refactored `llm_catalog.json` from a flat array of 389 entries to a normalized `families` + `models` structure. Family metadata (name, description, repo, tags, params_b, max_context_length) is stored once per family instead of duplicated across every quant. File size reduced from 264 KB to 136 KB (48% smaller). The persisted user catalog in `skill_dir` auto-migrates from the old flat format on first load — no user action required. Runtime in-memory representation (`LlmCatalog` with flat `Vec<LlmModelEntry>`) is unchanged, so all downstream Rust and frontend code continues to work without modification.

### Build

- **Windows CI: replace static CRT verification with dependency logging**: The release workflow now logs binary dependencies and verifies the CRT redist DLLs can be located for bundling, instead of failing on dynamic CRT linkage.

### LLM

- **Qwen3 30B-A3B Instruct full quant coverage**: Added all 9 missing GGUF quants (IQ3_S, IQ4_XS, Q3_K_S variants) from `byteshape/Qwen3-30B-A3B-Instruct-2507-GGUF` to the LLM catalog, bringing the total to 15 entries covering the full range of available quantisations.

## [0.0.67] — 2026-03-24

### Bugfixes

- **CI: remove skill-screenshots from cargo test step**: `skill-screenshots` has no `#[test]` items but pulling it into the test step forced a link of `ort-sys`/`libonnxruntime`, which fails on Linux because `cargo clippy` type-checks without final linking while `cargo test` must produce a real binary. Removing it from the `-p` list eliminates the spurious ORT link failure with zero loss of test coverage; `cargo clippy --workspace` continues to verify the crate compiles correctly.

- **CI: fix `ort-sys` link error on Linux/Windows**: `ort` was declared with `default-features = false` and no download strategy in the base `[dependencies]`, so `ort-sys` could not find `libonnxruntime` on Linux/Windows CI runners. Moved `ort` into target-specific sections — `download-binaries` for Linux/Windows (fetches the pre-built shared library at build time), `coreml` for macOS (unchanged).

- **Release: bundle ONNX Runtime library on Linux and Windows**: `ort-sys` downloads `libonnxruntime.so` / `onnxruntime.dll` into Cargo's `OUT_DIR` at build time. The binary links against it dynamically, but the packaging scripts did not include it — deployed binaries would fail to start on users' machines. Fixed by finding the library in the build tree and bundling it with every release format: portable tarball, `.deb`, `.rpm` (Linux) and NSIS installer (Windows). On Linux, `patchelf --add-rpath '$ORIGIN'` is also applied so the dynamic linker finds the bundled `.so` relative to the binary regardless of system paths.

- **skill-skills submodule test**: Added `submodules: true` to the `rust-check` CI checkout step so the `skills/` git submodule is fully populated and `discover_real_skills_submodule` runs (and passes) in CI instead of seeing an empty directory.

## [0.0.63] — 2026-03-24

### Features

- **Editable bash commands in tool calling**: added a pluggable `BashEditHook` callback and `require_bash_edit` setting in `LlmToolConfig`. When enabled, every LLM-generated bash command is presented to the user for review/editing before execution. The user can modify the command or cancel it entirely. Register a UI callback via `set_bash_edit_hook()` at app startup. Backward compatible — when no hook is registered or the setting is off, commands execute as before.
- **Record average SNR in session sidecar JSON**: the session runner now accumulates SNR (dB) across all band-power snapshots during a recording and writes `avg_snr_db` to the session sidecar JSON file on session end. The `SessionEntry` struct in `skill-history` now includes `avg_snr_db: Option<f64>` for frontend filtering. Legacy sessions without the field gracefully default to `None`.

- **Bash edit hook registered at app startup**: the Tauri app now registers a native dialog-based bash edit hook during setup, so the `require_bash_edit` setting is functional end-to-end.

- **Backfill average SNR for legacy sessions**: when loading session history, sessions without `avg_snr_db` in their sidecar JSON now get it computed on the fly from the per-epoch metrics in the SQLite database. This is a lightweight `AVG()` query per session — no full data reload needed. Legacy recordings seamlessly show SNR values without re-recording.

- **Network retry with exponential backoff**: Web tools (`web_search`, `web_fetch`, `location`) now automatically retry on transient errors (HTTP 429/5xx, connection failures) with configurable exponential backoff. The retry policy is exposed in Settings > Tools with controls for max retries (0-3) and base delay (500-3000 ms). Default: 2 retries with 1 s base delay.

- **Per-round usage tracking**: Tool orchestration now emits `ToolEvent::RoundComplete` events with cumulative `prompt_tokens`, `completion_tokens`, and `tool_calls_count` after each inference + tool-execution round, enabling cost/usage observability in agentic loops.

- **Increased default max tool rounds**: Bumped the default maximum tool-calling rounds from 10 to 15, giving the agent more room for complex multi-step tasks. The Settings UI now includes a 15-round option.

- **Updater: no auto-restart, session-safe**: Updates are now downloaded automatically in the background, but the app no longer auto-restarts after downloading. The user must explicitly click "Restart Now" to apply. Restarting is **blocked** while an EEG session is being recorded — a warning is shown and the restart button is disabled until the session ends. Only the user can bypass this protection. When quitting the app (via menu, tray, or Cmd+Q), if a downloaded update is staged, the app relaunches to apply it instead of just exiting.

- **Persistent web cache for tool results**: Added a disk-backed cache (`skill_dir/web_cache/`) for `web_search` and `web_fetch` results. Avoids redundant network calls when the LLM re-fetches the same URL or repeats the same query within/across conversations. Entries are keyed by SHA-256 hash and expire via configurable TTLs. Expired entries are evicted on startup.

- **Configurable TTL and per-domain overrides**: `WebCacheConfig` in `LlmToolConfig` with `search_ttl_secs` (default 30 min), `fetch_ttl_secs` (default 2 hours), and `domain_ttl_overrides` map for fine-grained control (e.g. shorter TTL for news sites).

- **Frontend structured logger**: added `$lib/logger.ts` with `log.debug/info/warn/error` — `console.log`/`console.debug` are now stripped from production builds via esbuild `pure` config.

- **Interactive search logic extraction**: extracted `search-interactive-logic.ts` with pure functions for display graph building, screenshot enrichment, node serialisation, and closest-screenshot selection — all with unit tests (9 new tests).

- **Property-based tests**: added `proptest` to `skill-tools` (9 tests: JSON scanner invariants, tool-call extraction robustness, Llama XML parsing) and `skill-eeg` (4 tests: FFT/IFFT round-trip, PSD non-negativity, batch length).

- **Headless browser unit tests**: added `InterceptStore` tests (5 tests: empty state, push/snapshot, clear, snapshot-with-clear, thread safety).

- **skill-llm catalog type tests**: added 6 unit tests for `LlmModelEntry` — is_split, shard_count, all_filenames (single + sharded), DownloadState default/serde roundtrip.

### Performance

- **Fix sccache on Windows CI (57 min -> ~20 min builds)**: The `sccache-action` v0.0.9 on Windows fails to add the sccache directory to the Windows `Path` variable. Git Bash can find sccache (Unix PATH), but native processes like `cargo.exe` cannot, so `RUSTC_WRAPPER=sccache` silently fails and every build compiles 840+ crates from scratch with zero caching. Fixed by resolving sccache to its full Windows path and exporting `RUSTC_WRAPPER` as an absolute path via `GITHUB_ENV`. Also adds the sccache directory to `GITHUB_PATH` so cmake compiler launchers can find it.

- **criterion benchmark suite**: added `crates/skill-eeg/benches/dsp_bench.rs` with benchmarks for FFT (128–1024), IFFT, PSD, BandAnalyzer, and EegFilter. Run via `cargo bench -p skill-eeg`.

### Bugfixes

- **Fix 5 broken doctests across the workspace**: marked uncompilable doc examples as `ignore` in `skill-tools/src/log.rs`, `skill-llm/src/log.rs`, `skill-tts/src/log.rs`, `skill-headless/src/lib.rs`, `src-tauri/src/lib.rs`, and `src-tauri/src/skill_log.rs`. All workspace tests now pass with zero failures.

- **Fix dead_code CI error in skill-tools**: Changed `#[allow(dead_code)]` to `#[expect(dead_code)]` on `clear_bash_edit_hook()` so it correctly suppresses the lint under `-D warnings`.

- **LLM decode retry on transient Metal failure**: When the initial prompt decode fails (commonly seen on macOS as "decode error on prompt (batch at token 0)"), the engine now clears the KV cache, waits briefly, and retries the full prompt once before reporting an error. This handles transient Metal GPU failures (busy command buffer, timeout) that previously required restarting the LLM. The same retry logic applies to multimodal (mtmd) eval.

- **LLM server start never transitions from "loading"**: Fixed a bug where clicking "Start" in the chat window would show "loading" indefinitely. The root causes were: (1) the status poll timer self-cancelled on mount when status was "stopped" and was never restarted when the user clicked Start, (2) failed init paths (`init()` returning `None`, actor early-exit on model load or context creation failure) did not emit an `llm:status` event so the UI was never notified. Now the poll timer is restarted on every start attempt, and all failure paths emit a `"stopped"` status event with an error message.

- **Fix svelte-check errors**: Added missing `avg_snr_db` default in history-helpers test factory; added `aria-label` to dismiss-error button in chat page.

- **Fix Windows CI PowerShell parse error**: Replaced literal em dash characters in PowerShell `run:` blocks in `release-windows.yml`. Non-ASCII in CI scripts can cause encoding corruption and parser failures on Windows runners.

- **Add missing tests for under-tested crates**: added 28 new unit tests across three crates:
  - `skill-llm`: 6 tests for `estimate_memory_gb` and `recommend_ctx_size` (catalog/memory), 7 tests for `ThinkTracker` budget enforcement (engine/think_tracker).
  - `skill-commands`: 11 tests for `query_slug`, `file_ts`, `pca_2d`, and `pca_3d` pure utility functions.
  - `skill-screenshots`: 4 tests for `ScreenshotMetrics` atomics and `MetricsSnapshot` serialization.

### Refactor

- **`retry_with_backoff` helper**: Added a generic `retry_with_backoff(max_retries, base_delay, closure)` utility in `skill-tools::exec::helpers`, reusable across any blocking I/O path. Exported from the crate root.

- **`ToolRetryConfig` struct**: New `ToolRetryConfig { max_retries, base_delay_ms }` in `LlmToolConfig`, serialisable and configurable per-user. Wired through to `exec_location`, `exec_web_fetch_plain`, and `exec_web_search`.

- **Biome frontend linter/formatter**: added `biome.json` config, `npm run lint/format` scripts, and CI integration for consistent TS/Svelte code style. Formatted all 257 source files.

- **Structured error types**: added `thiserror` to workspace deps and created `error.rs` modules in `skill-data`, `skill-llm`, and `skill-tools` with typed error enums (`SessionError`, `DownloadError`, `ValidationError`, etc.) for pattern-matchable error handling.

- **Faster dev builds**: scoped the `incremental = false` workaround to just `candle-core` and increased `codegen-units` to 8 (was 1) for the dev profile.

- **Dead code cleanup**: removed blanket `#![allow(dead_code)]` from `skill-tools/parse`, replaced `HooksLog.path` with `_path` prefix, fixed broken syntax in `scripts/bump.js`.

- **Dashboard logic extraction**: extracted `dashboard-logic.ts` with pure functions for EEG score computation (`sigmoid100`, `computeRawScores`, `emaSmooth`), display formatting (`fmtUptime`, `fmtEeg`, `redact`), goal progress, and device classification — all with 22 unit tests.

- **Onboarding logic extraction**: extracted `onboarding-logic.ts` with model selection functions (`pickFamilyTarget`, `pickLlmTarget`) — 13 unit tests covering quantization preference, family matching, download-skip logic.

- **Doc comments**: added `///` documentation to 16 key public types across `skill-history` and `skill-commands` (SessionEntry, SessionMetrics, EpochRow, SearchResult, DayIndex, etc.).

- **SAFETY comment for unsafe block**: added missing `// SAFETY:` comment to `ManuallyDrop::new(unsafe { std::mem::zeroed() })` in `skill-llm/src/engine/actor.rs`.

- **Remove 12 unused `anyhow` deps**: removed `anyhow` from skill-commands, skill-devices, skill-eeg, skill-gpu, skill-headless, skill-health, skill-jobs, skill-label-index, skill-screenshots, skill-settings, skill-tray, skill-vision — none were importing it.

- **Wire thiserror into skill-tools API**: `validate_tool_arguments` now returns `Result<Value, ValidationError>` instead of `anyhow::Result<Value>`, making validation errors pattern-matchable by callers. Re-exported error types from all 3 crate roots.

- **Split skill-history**: extracted 748 lines of metrics/CSV loading into `skill-history/src/metrics.rs`. `lib.rs` reduced from 1766 to 1019 lines.

- **Zero clippy warnings**: fixed unsafe block SAFETY comment placement in skill-gpu, converted 2 `match` patterns to `let...else` in skill-tools web_cache, replaced redundant closure with method reference.

- **UmapViewer3D logic extraction**: extracted `umap-viewer-logic.ts` with pure functions for point-cloud normalization, random positions, Turbo colormap, RGB-to-hex conversion, and color array building — 14 unit tests.

- **DevicesTab logic extraction**: extracted `devices-logic.ts` with fuzzy matching, device image resolution (Muse, Emotiv, IDUN, OpenBCI, Hermes), OpenBCI channel labeling, and relative-time formatting — 21 unit tests.

- **LlmTab logic extraction**: extracted `llm-tab-logic.ts` with hardware-fit badge styling, icons, and labels — 11 unit tests.

- **GoalsTab logic extraction**: extracted `goals-logic.ts` with progress bar coloring and minute formatting — 8 unit tests.

- **HooksTab logic extraction**: extracted `hooks-logic.ts` with timestamp conversion (microsecond/millisecond/second auto-detection) and relative-age formatting — 10 unit tests.

- **ScreenshotsTab logic extraction**: extracted `screenshots-logic.ts` with rolling history buffer, microsecond/millisecond/ETA formatting, and SVG sparkline/area path generation — 16 unit tests.

- **InteractiveGraph3D logic extraction**: extracted `graph3d-logic.ts` with 3D vector math (`add3`, `scale3`, `normalize3`, `length3`), Fibonacci sphere layout, and Turbo colormap (`turbo`, `turboCss`, `turboHex`) — 16 unit tests.

- **Biome lint fixes**: removed unused variables and `as any` casts in test files, replaced `!` non-null assertions with type-narrowing guards.

- **Split `skill-llm/src/catalog.rs` into focused sub-modules**: decomposed the 1059-line monolithic file into 5 modules under `catalog/` — `mod.rs` (re-exports), `types.rs` (data types), `persistence.rs` (load/save/merge/queries), `memory.rs` (memory estimation, context-size recommendation), `download.rs` (resumable HuggingFace downloader). All public API re-exports preserved.

- **Split `skill-tools/src/exec.rs` into focused sub-modules**: decomposed the 1724-line monolithic file into 9 focused modules under `exec/` — `mod.rs` (dispatch, 97 lines), `tools_system.rs` (date/location/bash/skill), `tools_web.rs` (web_search/web_fetch), `tools_fs.rs` (read_file/write_file/edit_file/search_output), `status.rs` (status text formatter), `truncate.rs` (output truncation helpers), `safety.rs` (dangerous-operation detection and approval), `helpers.rs` (path resolution, UTC formatting), `tests.rs`. No file exceeds 345 lines.
- **Split `skill-commands/src/graph.rs` into focused sub-modules**: decomposed the 1482-line monolithic file into 5 modules under `graph/` — `mod.rs` (re-exports), `dot.rs` (Graphviz DOT generation), `svg.rs` (2-D layered SVG), `svg_3d.rs` (3-D perspective SVG), `tests.rs`. All public API re-exports preserved for backward compatibility.

### Build

- **Node 22 in CI**: Bumped all workflows from Node 20 to Node 22, satisfying `camera-controls@3.1.2` engine requirement and aligning with current LTS.
- **esbuild ETXTBSY retry**: Added `npm ci || npm ci` retry in ci.yml to handle transient ETXTBSY race condition during esbuild postinstall on Linux runners.

- **cargo-deny**: added `deny.toml` for license compliance, duplicate detection, and advisory checking. Integrated into CI audit job.

- **Enhanced pre-commit hook**: now checks Rust formatting (`cargo fmt`), frontend formatting (Biome), and i18n sync — not just i18n.

- **CI improvements**: added Biome format check, cargo-deny, and expanded Rust test coverage to include `skill-headless`, `skill-screenshots`, `skill-label-index`, `skill-skills`, `skill-jobs`, `skill-commands`, `skill-exg`.

- **Workspace dependency consolidation**: promoted `thiserror` and `base64` to `[workspace.dependencies]` to ensure single-version consistency.

- **rustfmt.toml**: added workspace-level rustfmt configuration (`edition = "2021"`, `max_width = 120`).

- **Leaner dependency graph**: 12 crates no longer transitively pull in `anyhow` when they don't use it.

### CLI

- **Tauri commands**: `web_cache_stats`, `web_cache_list`, `web_cache_clear`, `web_cache_remove_domain`, `web_cache_remove_entry` for frontend access to the cache.

### UI

- **Bash command review toggle in chat tools panel**: added a "Review bash commands" toggle (visible when bash tool is enabled) that activates the `require_bash_edit` setting. Every LLM-generated bash command is shown in a native dialog for user approval before execution. Includes i18n translations for all 5 languages.
- **SNR display in session history**: the expanded session detail view now shows the average signal-to-noise ratio (dB) with color coding — green (>= 10 dB good), amber (>= 0 dB fair), red (< 0 dB poor). Includes i18n translations for all 5 languages.

- **LLM loading progress steps in chat window**: The chat window now shows a step-by-step progress indicator while the LLM server is starting, replacing the generic bouncing dots. Each phase is shown with a spinner (active), checkmark (done), or dot (pending): Loading model weights → Creating context → Loading vision projector → Compiling GPU shaders. A progress bar tracks overall advancement. The current loading phase is also shown in the chat header (visible even when messages exist from a previous conversation). All steps are fully localised (en, de, fr, he, uk).

- **Retry settings in Tools tab**: New "Network retry" section in Settings > Tools with button-group selectors for max retries and base delay, matching the existing visual style.

- **Web cache management panel**: New section in Settings > Tools (under web search) with:
  - Enable/disable toggle
  - Search TTL and fetch TTL button-group selectors (5/15/30/60 min and 30/60/120/240 min)
  - Live stats (entry count, total size)
  - Scrollable entry list with kind badge, domain, label, age, and size
  - Per-entry remove button, per-domain remove button, and clear-all button

- **Show LLM start errors in chat window**: When the LLM server fails to start, a dismissible red error banner now appears below the chat header with the failure reason (e.g. "no model selected", "model file not found", "failed to create context"). Previously the error was only logged to the browser console.

### i18n

- **Retry setting strings**: Added `llm.tools.retry*` keys to all five locales (en, de, fr, he, uk).

- **Web cache strings**: Added `llm.tools.webCache*` keys to all five locales (en, de, fr, he, uk).

### Docs

- **Fix all 18 `cargo doc` warnings across the workspace**: resolved unresolved intra-doc links, fixed references to private items, escaped HTML tags in doc comments, and corrected macro link syntax in `skill-llm`, `skill-tools`, `skill-tts`, `skill-eeg`, `skill-data`, `skill-headless`, and `src-tauri`. The workspace now builds documentation with zero warnings.

- **CONTRIBUTING.md**: added comprehensive contributor guide covering prerequisites, project structure, dev workflow, code quality commands, commit checklist, and architecture notes.

- **TODO.md**: replaced completed items with remaining improvement opportunities (skill-history split, criterion benchmarks, Svelte component decomposition, i18n gaps).

- **Fix cargo doc warnings**: resolved unresolved link to `set_bash_edit_hook` in skill-tools and `gpu_fft::psd::psd` in skill-eeg.

- **Fix cargo doc warnings**: resolved unresolved doc links in skill-tools and skill-eeg.

### Dependencies

- **sha2**: Added `sha2 = "0.10"` to `skill-tools` for cache key hashing.

## [0.0.62] — 2026-03-24

### Performance

- **Batch clippy and test in bump**: `npm run bump` now runs `cargo clippy` and `cargo test` once each for all workspace crates instead of per-crate (38 → 7 steps), and no longer deletes `src-tauri/target/` after bumping, preserving the build cache for subsequent runs. Pass `--clean` to free disk space when needed (`npm run bump -- --clean`).

### Bugfixes

- **Fix useCargo ReferenceError in tauri-build.js**: The `useCargo` variable was referenced in the early-exit pass-through block (line 161) before its definition (line 637), causing a `ReferenceError` for non-build/dev subcommands. Inlined the check at the early-exit point.

### CLI

- **`npm run bump --dry-run`**: added `--dry-run` flag that runs all preflight checks (clippy, tests, i18n, etc.) without modifying any version files, changelog, or build artifacts.

- **Bump progress bar**: `npm run bump` now shows a live progress bar during preflight checks with step counter (N/total), elapsed time, and ETA. Clippy and test steps run per-crate for granular visibility instead of one monolithic command.

- **TUI progress bar for `npm run bump`**: Preflight checks now display a full TUI with scrolling log output above a pinned progress bar. The progress bar updates every second with elapsed time and ETA, and command output streams in real-time instead of being hidden.

## [0.0.61] — 2026-03-24

### CLI

- **Detect competing cargo processes in bump**: `npm run bump` now checks for other running `cargo` processes before starting clippy/test preflight checks and warns that they may cause hangs due to the global cargo package-cache lock. The user is prompted to continue or abort.

## [0.0.60] — 2026-03-24

### Performance

- **LLM E2E reuses clippy build cache**: The `llm-e2e` CI job now runs after `rust-check` (`needs: rust-check`) and shares its Cargo build cache via `Swatinem/rust-cache` (`save-if: false`). This turns the E2E compilation into a cheap incremental build instead of a full rebuild (~3–5 min saved). The linker (mold + clang), target triple, and `CMAKE_*_COMPILER_LAUNCHER: sccache` env vars are aligned with `rust-check` to maximise cache hits. The Vulkan SDK (~200 MB) now uses the same `actions/cache` key as `rust-check`, skipping download on cache hit. System deps (including mold + clang) are merged into one cached `awalsh128/cache-apt-pkgs-action` step, eliminating a separate `apt-get update` round-trip.
- **Halve LLM E2E test runtime (67s → 33s)**: Reduced `ctx_size` from 4096 to 2048 (test prompts are < 600 tokens). Reduced tool-chat `max_tokens` from 128 to 64 — the model was wasting inference echoing full JSON results (117-128 tokens) when it only needs ~25 tokens for the tool call and a short summary. Added `[profile.test.package.llama-cpp-sys-4]` and `llama-cpp-4` with `opt-level = 3` so the C++ inference engine runs optimized even in test builds, boosting tok/s by ~20%.
- **LLM E2E is now opt-in**: The `llm-e2e` job no longer runs automatically on push. It is triggered only via manual `workflow_dispatch` with the `run_llm_e2e` checkbox enabled. The Discord notification shows ⏭️ when skipped.

### Bugfixes

- **Regenerate Cargo.lock during bump**: `npm run bump` now runs `cargo generate-lockfile` after updating version in `src-tauri/Cargo.toml`, preventing CI `--locked` build failures due to stale lockfile.

- **Add missing safety comment for unsafe block**: Added `// SAFETY:` comment on the Linux `RLIMIT_STACK` unsafe block in `main.rs` to satisfy `clippy::undocumented_unsafe_blocks` (Rust 1.94).

- **Fix i18n test import shadowing**: The `extractKeysFromDir` import from `i18n-utils.ts` was shadowing the local test helper of the same name, causing 8 pre-existing key-sync test failures. Renamed to `extractKeysWithValues` in the test import.

### Build

- **CI/Release speed improvements**: Estimated 7-12 min savings per CI run via multiple optimizations:
  - Removed redundant `cargo check` step — `clippy` is a strict superset and already compiles everything.
  - Merged duplicate `cargo-audit` and `audit` jobs into a single security audit job.
  - Moved audit to run only on main/develop pushes (advisory, not PR-blocking).
  - Added concurrency groups to `ci.yml` to cancel superseded runs on the same branch/PR.
  - Replaced manual `--workspace -p crate1 -p crate2 ...` clippy invocations with `--workspace --exclude skill`.
  - Switched Linux CI from manual `actions/cache` to `Swatinem/rust-cache` for smarter per-crate invalidation (matching release workflows).
  - Cached Vulkan SDK on Linux CI (previously downloaded ~200 MB on every run).
  - Added `mold` fast linker + `clang` to Linux CI (previously only in release-linux).
  - Added `fetch-depth: 1` to CI jobs that don't need full git history.
  - Added `--locked` flag to all release cargo build commands for reproducible builds.
  - Added `--timings` flag and cargo-timings artifact upload to macOS and Linux release workflows (previously Windows only) for build profiling.
  - Removed redundant `cargo check` on Windows CI — `clippy` already covers it.
  - Fixed Discord notification to show "skipped" emoji for audit when it doesn't run on PRs.
  - Added `CMAKE_C_COMPILER_LAUNCHER` / `CMAKE_CXX_COMPILER_LAUNCHER` sccache integration to macOS, Linux CI, and preview builds (previously Windows release only) so cmake-based -sys crate compilations (llama-cpp-sys, espeak-ng) are cached across runs.

- **Commit Cargo.lock**: Removed `Cargo.lock` from `.gitignore` and committed it so that `cargo clippy --locked` and CI builds succeed without needing to regenerate the lockfile.

### i18n

- **Untranslated value detection**: Added a vitest check (`i18n untranslated value detection`) that fails when any non-English locale contains values identical to English that are not in the exemption list. The exemption logic (brand names, technical acronyms, formulas, academic citations, etc.) is now shared between `i18n-utils.ts`, the `audit-i18n.ts` script, and the test suite. A new `check:i18n` npm script runs the audit with `--check` for CI gating.

- **Translate all untranslated strings**: Translated 388 strings across 4 locales (de, fr, he, uk) that were still in English. Covers common UI (errors, dismiss, zoom reset, goal reached), supported devices and setup instructions, device API settings, API authentication, screenshot/OCR pipeline labels, screen recording permissions, history streaks, LLM tool settings, onboarding, and search screenshot tab. Added brand/product names and cross-language cognates to the exemption list.

## [0.0.59] — 2026-03-24

### Features

- **Windows NSIS installer: optional runtime components**: The installer now has a Components page with smart auto-detection. Optional sections for Vulkan Runtime (GPU acceleration), VC++ 2015-2022 Redistributable, WebView2 Runtime (required for UI on older Windows 10), and GPU TDR timeout increase (prevents driver resets during long LLM inference). Each component auto-selects only when its prerequisite is missing.
- **Windows NSIS installer: kill running instance**: The installer detects if the app is running before upgrading and offers to close it (WM_CLOSE then taskkill). The uninstaller also force-kills the app before removing files.
- **Windows NSIS installer: long path support**: Enables the `LongPathsEnabled` registry key so HuggingFace model cache paths exceeding 260 characters work correctly.
- **Windows NSIS installer: firewall rule**: Adds a Windows Firewall exception for the local LLM/WebSocket server, preventing the "allow access?" popup on first launch. Cleaned up on uninstall.
- **Windows NSIS installer: launch at login**: A "Launch at login" component is selected by default. Writes the same `HKCU\...\Run\skill` registry key used by the in-app autostart setting. Users can uncheck it during install or toggle it later in Settings.
- **Windows NSIS installer: clean uninstall**: The uninstaller now removes the autostart registry entry (`HKCU\...\Run\skill`), the firewall rule, and optionally offers to delete the user data folder (`%LOCALAPPDATA%\NeuroSkill`) including settings, sessions, and downloaded models.

### Refactor

- **Adopt `anyhow` across all workspace crates**: Replaced `Result<T, String>` error handling with `anyhow::Result<T>` throughout the crate layer. `map_err(|e| format!(...))` chains are now `.context()` / `.with_context()`, and `Err(format!(...))` becomes `anyhow::bail!(...)`. The Tauri command boundary in `src-tauri/` converts back to `String` via `.map_err(|e| e.to_string())`. `skill-jobs` retains `Result<Value, String>` for stored/cloned job results. `skill-headless` retains its `HeadlessError` enum. A migration script (`scripts/adopt_anyhow.py`) automates the bulk of the conversion.

## [0.0.58] — 2026-03-24

### Features

- **Auto-launch after install on Windows**: The NSIS finish page now shows a "Launch NeuroSkill™" checkbox (checked by default). The app is launched as the current user (not elevated) via the Explorer shell trick, so per-user paths, tray registration, and autostart work correctly.

- **Extracted `search-logic.ts` module**: Pulled pure business logic (mode normalization, UMAP label enrichment, time helpers, analysis chips) out of the 2,169-line search page into a testable TypeScript module with 9 unit tests.

- **Extracted `compare-logic.ts` module**: Pulled timeline helpers, session range selection, pointer-to-UTC conversion, and date formatting out of the 1,922-line compare page into a testable module with 14 unit tests.

- **Added DSP pipeline integration tests**: New `dsp_pipeline_test.rs` for `skill-eeg` covering band analysis at multiple sample rates, beta/alpha dominance detection, quality monitoring, and reset behaviour (5 tests).

- **Added `skill-headless` tests**: New `intercept_tests.rs` covering `InterceptStore` push/snapshot/clear, serialization round-trip, and default state (5 tests).

- **Added `skill-screenshots` tests**: New `context_tests.rs` covering mock context, fastembed model resolution, and `ActiveWindowInfo` defaults (5 tests).

- **Added `skill-jobs` integration tests**: New `queue_tests.rs` covering submit-and-poll, sequential execution ordering, error propagation, queue positioning, and unknown-job handling (5 tests).

- **Added `skill-data` tests**: New `hooks_log_tests.rs` (4 tests) and `screenshot_store_tests.rs` (4 tests) covering database creation, insert, count, and pagination.

### Bugfixes

- **Fixed 2 broken tests in `src-tauri/src/constants.rs`**: The `embedding_overlap_samples_correct` and `embedding_hop_samples_correct` tests were asserting stale values (640) after `EMBEDDING_OVERLAP_SECS` changed from 2.5 to 0.0. Tests now derive expected values from the actual constants.

- **Fixed compilation error in `skill-headless`**: Two `eval_fire()` calls passed `reply` by value instead of by reference.

- **Fix ™ symbol mangled in Windows NSIS installer**: The `.nsi` script was written as UTF-8 without BOM. NSIS with `Unicode True` requires a BOM to detect UTF-8; without it, NSIS falls back to the system ANSI codepage and corrupts non-ASCII characters like ™ in the product display name, version info, registry entries, and shortcuts. Changed to UTF-8 with BOM.

- **Fix Windows SxS "side-by-side configuration incorrect" error**: Added `CMAKE_MSVC_RUNTIME_LIBRARY = "MultiThreaded"` to the workspace `.cargo/config.toml` `[env]` section so that cmake-based C/C++ dependencies (llama-cpp-sys, espeak-ng) use static CRT (`/MT`) matching Rust's `+crt-static` target feature. Previously this env var was only set in CI, causing local Windows builds to produce a CRT mismatch that triggered the SxS error on machines without the Visual C++ Redistributable.

### Refactor

- **Eliminated all clippy warnings across the workspace**: Resolved every warning from `cargo clippy --workspace` — from 500+ warnings to zero.

  - Converted 21 `match` → `let...else` patterns across 10 files (api.rs, screenshot_cmds.rs, ws_commands, device_scanner, session_runner, session_connect, eeg_embeddings, label_cmds, skill-skills/sync, skill-llm/actor).
  - Replaced 17 `lock().expect("lock poisoned")` calls with `lock_or_recover()` across skill-llm (handlers, logging, state, tool_orchestration) and src-tauri/api.
  - Applied `cargo clippy --fix` auto-fixes: redundant closures, method call simplifications, let-else conversions.
  - Added `// SAFETY:` comments to remaining undocumented `unsafe` blocks (skill_log, quit, active_window, window_cmds).
  - Added `#![allow(clippy::panic, clippy::expect_used, clippy::unwrap_used)]` to `build.rs` (build-time panics are standard practice).
  - Added `#![allow(clippy::unwrap_used)]` to `image_encode_bench.rs` (benchmark binary).
  - Disabled `needless_pass_by_value` lint (280 false positives from Tauri `#[command]` handlers).
  - Disabled `expect_used` lint (50 legitimate uses in thread spawning, NonZero constructors, Tauri app builder — all unrecoverable).

- **Split `skill-tools/src/parse.rs` into modules**: The 2,441-line monolith is now split into 7 focused sub-modules (`types`, `coerce`, `validate`, `extract`, `strip`, `inject`, `json_scan`) while preserving the full public API. All 161 tests continue to pass.

- **Workspace-wide lint configuration**: Added `[workspace.lints]` in root `Cargo.toml` with consistent Clippy rules across all 22 crates — `unwrap_used`, `expect_used`, `undocumented_unsafe_blocks`, `needless_pass_by_value`, and more. All crates inherit via `[lints] workspace = true`.

- **Eliminated `unwrap()` in library code**: Replaced the sole remaining `unwrap()` in production code (`skill-commands/src/graph.rs`) with safe `let-else`. All other `unwrap()` calls were already in tests/examples/benchmarks.

- **Hardened `InterceptStore` lock handling**: Replaced `lock().expect("lock poisoned")` with graceful `if let Ok(guard)` pattern in `skill-headless` — poisoned locks now degrade gracefully instead of panicking.

### Build

- **Bust Windows CI caches**: Bumped Cargo and sccache cache keys (`windows-release-x86_64-msvc-v2`, `SCCACHE_GHA_VERSION=2`) to invalidate stale cmake artifacts that were compiled with dynamic CRT (`/MD`).
- **Verify static CRT in Windows CI**: Added a post-compile `dumpbin /dependents` check that fails the build if `vcruntime140.dll`, `ucrtbase.dll`, or `msvcp*.dll` appear as dependencies, catching dynamic CRT leaks before packaging.

- **Added `cargo audit` to CI**: New `cargo-audit` job in the CI pipeline scans for known dependency vulnerabilities on every push and PR.

### Docs

- **Added SAFETY comments to all `unsafe` blocks**: Documented invariants for every `unsafe` block in `skill-vision` (FFI OCR), `skill-gpu` (IOKit/sysctl), `skill-llm` (llama.cpp backend), and `skill-screenshots` (CoreFoundation/AppKit FFI).

## [0.0.57] — 2026-03-24

### Bugfixes

- **Fix skill-headless build with wry 0.54**: Enable `rwh_06` feature on `tao` so `Window` implements `HasWindowHandle` required by wry 0.54.

- **Fix mDNS discovery in smoke-test.sh**: Keep a single `dns-sd -B` process running continuously and poll its output every second, instead of spawning and killing a new process every 3 seconds (which could miss service announcements).

- **Fix smoke-test false-positive mDNS match**: The `dns-sd -B` header line already contains "skill", so the grep matched immediately before any service was actually discovered. Changed pattern to `Add.*_skill._tcp` which only matches a real service registration event.

- **Smoke test mDNS retry**: Moved mDNS discovery from `smoke-test.sh` (bash `dns-sd`) into `test.ts` (bonjour-service). Discovery now retries indefinitely with a 3-second backoff until the Skill server appears or the user presses Ctrl-C, fixing the "could not resolve port" failure on slow startups.

- **Smoke test port discovery**: Fixed `smoke-test.sh` failing because `test.ts` tried its own 5-second mDNS browse which raced and failed. The script now resolves the port via `dns-sd -L` after the browse succeeds and passes it explicitly to `test.ts`, eliminating the double-discovery race condition.

- **Fix smoke test on macOS**: Replace GNU `timeout` command (not available by default on macOS) with a portable Perl-based timeout in `smoke-test.sh`.

- **Fix smoke-test unbound variable**: `scripts/smoke-test.sh` failed with `unbound variable` when invoked without arguments due to `set -u` and bare `${*}`. Changed to `${*:-}` to default to empty string.

### CLI

- **Move smoke-test to scripts/ and add npm script**: Moved `smoke-test.sh` into `scripts/` and added `npm run test:smoke` shortcut.

## [0.0.56] — 2026-03-24

### Features

- **Smoke test script**: Added `smoke-test.sh` and `npm run test:smoke` to launch the app and run `test.ts` end-to-end inside a tmux session, waiting for mDNS service discovery before starting tests.

- **Full WS command test coverage**: Added smoke tests in `test.ts` for all previously untested WebSocket commands: `session_metrics`, calibration CRUD (`get_calibration`, `create_calibration`, `update_calibration`, `delete_calibration`), `sleep_schedule` / `sleep_schedule_set`, health commands (`health_summary`, `health_metric_types`, `health_query`, `health_sync`), and extended LLM management commands (`llm_downloads`, `llm_refresh_catalog`, `llm_hardware_fit`, `llm_select_model`, `llm_select_mmproj`, `llm_pause_download`, `llm_resume_download`, `llm_set_autoload_mmproj`, `llm_add_model`). All 58 dispatch table commands now have corresponding tests.

### Performance

- **Faster Windows release builds**: Added `CMAKE_C_COMPILER_LAUNCHER` and `CMAKE_CXX_COMPILER_LAUNCHER` env vars so sccache caches C/C++ compilations from cmake-based -sys crates (llama-cpp-sys-4, etc.), not just rustc. Added `[profile.release]` with `lto = "thin"` and `codegen-units = 8` for faster linking. Moved frontend build before SDK installs to reduce idle time. Combined NSIS and Vulkan SDK installs into a single parallel step. Added sccache GHA cache versioning and `--timings` cargo flag with artifact upload for build profiling.

### Bugfixes

- **Fix Windows CI zip assembly loading**: Load `System.IO.Compression` assembly explicitly before `System.IO.Compression.FileSystem` in the "Sign installer + create updater artifacts" step. On some PowerShell runtimes (notably PowerShell Core on GitHub Actions), loading only `FileSystem` does not implicitly load the base `System.IO.Compression` assembly, causing `Unable to find type [System.IO.Compression.ZipArchiveMode]` errors during NSIS updater zip creation.

## [0.0.55] — 2026-03-24

### Bugfixes

- **Fix Windows "side-by-side configuration" launch error**: Statically link the MSVC C/C++ runtime (`+crt-static`) into the Windows binary so the app no longer requires the Visual C++ Redistributable to be pre-installed. The espeak-ng static build and CI workflow also set `CMAKE_MSVC_RUNTIME_LIBRARY=MultiThreaded` to ensure all C/C++ dependencies use the matching static CRT.

- **Fix Windows auto-update "unsupported compression method"**: Replaced PowerShell `Compress-Archive` with .NET `ZipFile` using explicit Deflate compression (method 8) when creating updater `.nsis.zip` archives. `Compress-Archive` on newer Windows versions uses Deflate64 (method 9), which the Tauri updater's zip crate does not support.

## [0.0.54] — 2026-03-23

### Bugfixes

- **Fix clean:rust ENOTEMPTY on large target dirs**: Added retry options and `rm -rf` fallback to `scripts/clean-rust.js` so the Rust build artifact cleanup no longer fails with ENOTEMPTY on very large directory trees.

- **Windows app manifest for BLE access**: Added a custom Windows application manifest (`manifest.xml`) declaring Windows 10/11 compatibility via `supportedOS` and `maxversiontested`. Without this, Windows 11 may deny WinRT Bluetooth Low Energy API access to unpackaged desktop apps. Also includes Common Controls v6 and per-monitor DPI awareness v2.

- **Windows 11 Bluetooth permissions guidance**: Updated BLE error messages and the "Bluetooth is off" UI state to include Windows 11-specific instructions (Settings → Privacy & Security → Bluetooth). The "Open Settings" button now opens both the Bluetooth devices page and the Bluetooth privacy page on Windows. Added a Windows-specific adapter state check in `bluetooth_ok()` to detect powered-off adapters. Updated all locales (en, de, fr, he, uk).

### Build

- **Stop tracking Cargo.lock**: Removed all `Cargo.lock` files from version control and added `Cargo.lock` to `.gitignore`.

- **Add libgbm-dev to Linux CI**: Added missing `libgbm-dev` package to CI and release workflows to fix linker error (`library not found: gbm`).

## [0.0.53] — 2026-03-22

### Features

- **Auto-pair first discovered device**: When the paired devices list is empty (fresh install or all devices unpaired), the first discovered device is automatically added to paired devices and auto-connected. This provides a seamless first-run experience without requiring manual pairing. Only triggers when no paired devices exist.

- **Cortex WebSocket connection status indicator**: The Emotiv Cortex scanner backend now tracks and emits its WebSocket connection state (`disconnected`, `connecting`, `connected`) to the frontend in real time.

- **Llama XML tool-call parsing**: `extract_tool_calls()` now recognises the `<function=name><parameter=key>value</parameter></function>` XML format emitted by Llama-family models. Both `<parameter>` tag pairs and inline JSON bodies are supported. Stripping and streaming partial-tag handling are also included.

- **Tool-call self-healing**: When the LLM emits a garbled or malformed tool call that cannot be parsed, the orchestrator now detects the failed attempt, injects a corrective message containing the raw output, and asks the model to re-emit in the correct format. Up to 2 retry attempts are made before falling back to normal output. This significantly improves reliability with smaller local models.

### Bugfixes

- **Emotiv credentials not persisting across restarts**: `DeviceApiConfig` fields used `#[serde(skip_serializing)]` to keep secrets out of the JSON settings file, but this also caused `get_device_api_config` to return empty credentials to the frontend via Tauri IPC. The command now returns a `serde_json::Value` that bypasses the skip, so stored keychain credentials are correctly displayed after restart and are no longer accidentally overwritten with empty values on subsequent saves.

- **Keychain credentials not persisted on any platform**: The `keyring` crate v3.x requires explicit platform backend features. Without them, no credential store was compiled in and `set_password`/`get_password` silently failed, causing Emotiv (and IDUN / API token) credentials to be lost on restart. Enabled OS-specific backends via target dependencies (`apple-native` on macOS, `windows-native` on Windows, `linux-native-sync-persistent` + `crypto-rust` on Linux). Also improved error logging so keychain failures are no longer silently swallowed.

- **KittenTTS: re-open audio device on every utterance**: The KittenTTS backend opened the system audio output once at startup and reused the same stream for all subsequent speech. If the device was unplugged, switched, or became unavailable (e.g. Bluetooth disconnect, USB DAC removal), playback failed with "The requested device is no longer available." The worker now re-opens the default audio device before each utterance, matching the NeuTTS backend behaviour. This is cheap (~1 ms) relative to synthesis time and ensures the current default device is always used.

- **Screenshots not captured on Windows and Linux**: The `screenshots` feature (which enables the `xcap` screen-capture backend) was missing from the default Cargo features and from all CI/release build workflows. On macOS this was invisible because capture uses the `screencapture` CLI tool directly, but on Windows and Linux the capture function silently returned `None`. Added `screenshots` to the default feature set and to all build/CI workflows for Windows and Linux.

- **Fix WebView2 window class unregistration error on Windows**: Stopped dropping the WebView early during Close command handling. The `Chrome_WidgetWin_0` class unregistration race (Win32 error 1412) occurred because the WebView was destroyed while the event loop and parent window were still alive. The WebView is now dropped naturally when `run_return` finishes, allowing Chromium's internal child windows to clean up before class unregistration.

- **Fix Windows Vulkan STATUS_ACCESS_VIOLATION crash**: The crash was in the Vulkan validation layer (`VK_LAYER_KHRONOS_validation`) loaded by the Vulkan SDK during llama.cpp model loading on the `llm-actor` thread. In debug builds, the validation layer's `ErrorObject` constructor crashes inside `vkEnumeratePhysicalDevices` on certain Windows GPU drivers. Now disable the validation layer at app startup via `VK_LOADER_LAYERS_DISABLE` and `VK_INSTANCE_LAYERS` env vars (affects llama.cpp / ggml-vulkan) plus `WGPU_VALIDATION=0` (affects wgpu/cubecl). Added a Windows vectored exception handler that prints faulting address, thread name, and full backtrace for any future access violations.

- **Fix window flickering on Windows**: The screenshot capture worker iterated through all non-minimized windows calling `PrintWindow` (via xcap `capture_image()`) on each one until it got a result. `PrintWindow` sends a `WM_PRINT` message forcing each window to repaint, causing constant visible flickering across all open windows every few seconds. Now on Windows only the single foreground window is captured using xcap's `is_focused()` check, with a monitor-capture fallback if that fails.

### Build

- **`clean:rust` script fails on Windows**: Replaced Unix `rm -rf` with a cross-platform Node script (`scripts/clean-rust.js`) that works on Windows, macOS, and Linux. The script now reports the size of build artifacts and how much disk space was freed.

### UI

- **Colored status dots for Emotiv Cortex**: The Scanner Backends section now shows a green dot when connected to the Cortex WebSocket, a blinking yellow dot while connecting, and a red dot when disconnected — replacing the previous static text badges.

- **EMOTIV Launcher connection reminder**: Devices discovered via the Cortex transport (EMOTIV headsets) now show a hint reminding the user to confirm the connection in the EMOTIV Launcher app.

- **Native shortcut rendering in tray menu**: Keyboard shortcuts in the system tray context menu now use the native accelerator parameter on `MenuItem` instead of being appended to the label text. The OS renders them right-aligned in the platform-native style (e.g. ⌘⇧L on macOS, Ctrl+Shift+L on Linux/Windows).

### LLM

- **`detect_garbled_tool_call()`**: New public function that identifies malformed tool-call attempts in assistant output (broken `[TOOL_CALL]` blocks, incomplete XML `<function=` tags, or unbalanced JSON with tool-call keys).
- **`build_self_healing_message()`**: New public helper that constructs the corrective user message for the retry loop.

### i18n

- **New key `settings.scanner.cortexConnecting`**: Added "Connecting…" translations in EN, DE, FR, HE, and UK.

- **New key `settings.emotivLauncherHint`**: Added translations in EN, DE, FR, HE, and UK for the EMOTIV Launcher reminder.

- **Remove hardcoded Muse/OpenBCI two-device-only phrases**: Updated all i18n strings (en, de, fr, he, uk) that mentioned only "Muse or OpenBCI" to use device-agnostic wording like "your BCI device", "any supported BCI device", or "another BCI device". The app now supports Muse, MW75 Neuro, OpenBCI boards, Emotiv, IDUN Guardian, and more — user-facing text no longer implies only two brands are supported. Search keywords in the command palette were expanded to include all supported device names.

## [0.0.52] — 2026-03-21

### Features

- **System keychain for secrets**: API tokens and device credentials (`api_token`, Emotiv client ID/secret, IDUN API token) are now stored in the OS credential store (macOS Keychain, Linux Secret Service, Windows Credential Manager) instead of plaintext in `settings.json`. Secrets survive app reinstalls and build updates. Existing plaintext values are automatically migrated to the keychain on first launch and stripped from the JSON file.

- **Skills sync on launch**: Added a "Sync on launch" toggle in the Skills auto-refresh settings. When enabled, the app forces a community skills sync every time it starts, regardless of the periodic refresh interval. The periodic schedule continues to run normally afterwards.

- **Flamegraph profiling script**: Added `npm run tauri:flamegraph` to profile the Tauri app with `flamegraph` and produce an interactive SVG. Works on Linux (perf), macOS (dtrace), and Windows (dtrace/xperf). Supports optional duration argument (e.g. `npm run tauri:flamegraph -- 60`) and `--release` flag (default: dev profile to match `tauri dev`).

### Performance

- **Session listing 10x faster**: Replaced `serde_json::Value` (BTreeMap-backed) with a typed `SessionJsonMeta` struct for parsing session JSON sidecars, eliminating expensive BTreeMap construction and recursive drop overhead.
- **Metrics timestamp lookup O(1) instead of O(n)**: `read_metrics_csv_time_range` now reads only the first and last 4 KB of the file (via seek) instead of parsing every CSV record. For a 100 MB metrics file this reduces I/O from ~100 MB to ~8 KB.
- **Skip redundant timestamp patching**: `patch_session_timestamps` now skips sessions that already have valid start/end timestamps from their JSON sidecar, avoiding unnecessary metrics file reads on every session listing.
- **ZUNA encoder loads ~60% faster**: Encoder-only weight filter skips all decoder tensors during deserialization (halves bf16-to-f32 conversion work and memory). Weight data is moved instead of cloned via new `WeightMap::take()`. HashMap pre-sized from safetensors tensor count.
- **LUNA encoder loads faster**: Updated luna-rs to 0.0.3 with the same zero-copy weight loading, encoder-only filter, and HashMap pre-sizing optimizations.

### Bugfixes

- **Flamegraph permission errors on macOS**: Separated build (normal user) from profiling (`sudo flamegraph`) so dtrace runs as root and owns its trace files. Fixes "Trace file already exists" (exit 42) and "Permission denied" errors caused by root-owned artifacts from previous runs.
- **Flamegraph builds dev profile by default**: Now matches `tauri dev` behavior. Pass `--release` explicitly for optimized profiling.

- **Flamegraph script launching stale binaries**: Fixed `tauri:flamegraph` profiling old binaries instead of freshly-built ones. Added `-p skill` to target only the skill package, moved build cwd to workspace root, added sccache/mold detection to match `tauri-build.js` environment (prevents fingerprint mismatches), made `forceRemove` fail hard instead of silently continuing, and added post-build mtime verification to catch stale binaries before profiling.

- **Skills sync discovers all community skills**: The skill discovery algorithm stopped recursing into subdirectories when the repository root contained a valid `SKILL.md` with a `description` in frontmatter, causing only one skill (the index) to be loaded. Added support for an `index: true` frontmatter flag that marks a `SKILL.md` as an index file — the skill is loaded but the scanner continues recursing into child directories. The community skills repo root `SKILL.md` now uses this flag. Also fixed Phase 2 to skip re-processing `SKILL.md` files already handled in Phase 1.

### Build

- **Flamegraph script: full clean + WebView cache purge + user directory fixes**: `npm run tauri:flamegraph` now performs a complete clean before profiling: `cargo clean`, SvelteKit/Vite cache removal (`.svelte-kit`, `node_modules/.vite`, `build`), and — critically — **WebKitGTK/WebKit per-app cache clearing** (`~/.local/share/com.neuroskill.skill/` on Linux, `~/Library/WebKit/` and `~/Library/Caches/` on macOS). The stale WebView cache was causing the profiled binary to show old HTML/CSS and the wrong icon even after a fresh build. The sudo `--preserve-env` list is also expanded to include `CARGO_HOME`, `RUSTUP_HOME`, `DISPLAY`, `WAYLAND_DISPLAY`, `DBUS_SESSION_BUS_ADDRESS`, `XDG_RUNTIME_DIR`, and `LOGNAME`, ensuring the profiled app uses current-user directories instead of root's.

## [0.0.51] — 2026-03-21

### Features

- **Cross-modal screenshot ↔ EEG search in CLI**: Added new CLI commands and WS endpoints for bridging screenshots and EEG data:
  - `search-images --by-image <path>` — search screenshots by visual similarity using CLIP vision embeddings (base64 image sent over WS, server-side CLIP embedding + HNSW search).
  - `screenshots-for-eeg [--start --end] [--window N]` — find screenshots captured near EEG recording timestamps ("EEG → screen" bridge). Auto-selects the latest session when no range is given.
  - `eeg-for-screenshots "query" [--k N] [--window N]` — search screenshots by OCR text, then return EEG labels and session info near each match ("screen → EEG" bridge).
  - New WS commands: `search_screenshots_by_image_b64`, `search_screenshots_vision`, `screenshots_for_eeg`, `eeg_for_screenshots`.
  - All commands support `--json`, `--mode`, `--k`, `--window`, and `--limit` flags.

- **Implicit evidence collection for protocols**: Every protocol execution now produces structured measurement data using a standardised `px:start:`/`px:end:` label schema with pipe-separated metrics (bar, stress_index, relaxation, focus, hr, mood, faa, rmssd, deltas, outcome). The LLM captures before/after snapshots automatically and silently for every protocol, determining outcome as positive (≥10% target improvement), neutral, or negative. No user action required — evidence collection is invisible infrastructure.

- **Personal protocol effectiveness ranking**: After 5+ labeled protocol executions, the LLM aggregates outcomes via `search_labels "px:end"` to build a personal ranking by success rate and average metric delta. Surfaces time-of-day patterns, trigger-specific effectiveness, and modality preferences. Presents insights like "Cold water face splash is your most effective stress intervention — 92% success, average stress drop 31%."

- **Evidence-driven protocol selection**: New matching guidance rule — "Evidence first." Before suggesting any protocol, the LLM checks past effectiveness data. Leads with proven personal winners over generic recommendations. Retires consistent failures after 4+ negative outcomes. Explores new protocols occasionally even with strong data. Tracks modality preferences across interventions.

- **Implicit life-event labeling**: The LLM silently labels mentioned life events (caffeine intake, walks, meals, meetings, exercise, app switches, sleep quality) with EEG metric snapshots to build a complete personal effectiveness map beyond formal protocols. Privacy safeguards: full transparency if asked, user owns all local data, no inference of unmentioned events.

- **Interactive graph screenshots**: Added screenshot nodes to the interactive 3D cross-modal search graph. When the "Screenshots" checkbox is enabled, screenshots are discovered via three strategies: semantic text search on the query (attached to the query node), semantic search on each text-label (attached to label nodes), and timestamp proximity ±30 min around EEG points. Screenshots are injected as proper connected graph nodes (cyan spheres with async-loaded thumbnail sprites), connected via `screenshot_link` edges, and fully participate in selection highlighting. Duplicates are suppressed by filename.
- **SVG/DOT export includes screenshots**: When screenshots are visible in the interactive graph, the exported SVG and DOT files now include screenshot nodes. The frontend regenerates the SVG/DOT via new `regenerate_interactive_svg` / `regenerate_interactive_dot` Tauri commands, passing the full augmented graph (including client-side screenshot nodes) to the backend layout engine.

- **Interactive search: screenshot discovery and 3D visualization**: The interactive cross-modal search now discovers screenshots near EEG neighbor timestamps, ranking them by window-title and OCR-text proximity to the query. Screenshot nodes appear as a new layer in the graph with dedicated styling (pink `#ec4899`). Two new edge kinds (`screenshot_prox` for temporal proximity and `ocr_sim` for text-based matches) connect EEG points to relevant screenshots. A new 3D perspective-projected SVG (`svg_3d`) is generated alongside the existing 2D layouts using 3-component PCA across all text embeddings. The `InteractiveGraphNode` struct gains `proj_z`, `filename`, `app_name`, `window_title`, `ocr_text`, and `ocr_similarity` fields. The `InteractiveSearchResult` includes the new `svg_3d` field. Both the Tauri command and WebSocket `interactive_search` handler are updated. DOT and flat SVG exports also render screenshot nodes.

- **Modality Router**: Added a decision table mapping 12 EEG triggers to 7 intervention modalities (Breath, Tactile, Cognitive, Visual, Movement, Auditory, Passive Physiological) so the LLM selects the best modality for each person's circumstances before choosing a specific protocol. Breathing is presented as one equal option among many, never the default. Includes a modality selection guide by context and phrasing examples showing how to offer multi-modal choices.

- **Multi-modal protocol restructuring**: Restructured ~30 protocols that previously defaulted to breathing to present non-breathing alternatives at equal priority using ", OR " choice points. Affected protocols include Focus Reset, Pre-Performance Activation, Extended Exhale, Physiological Sigh, Kapalabhati, 4-Count Energising, Wim Hof, Anger Processing, Grief Holding, Emotion Surfing, Joy Amplification, Emotional Boundaries, Excitement Regulation, all morning routines, all workout protocols, social media protocols, Pre-Meal Pause, Sleep Wind-Down, Co-Regulation, One-Handed Calm, Cortical Quieting, Coherence Building, Break Reset, and Post-Scroll Reset.

- **Matching Guidance updated**: Added "Modality first, protocol second" as the top matching rule and "Always offer at least two modalities" as a new requirement, ensuring the LLM never prescribes breathing without presenting alternatives.

- **neuroskill-evidence standalone skill**: New skill defining the implicit evidence collection and personal effectiveness engine used across all intervention-delivering skills. Includes the standardised `px:` label schema (`px:start`, `px:end`, `px:note`, `px:skip`, `px:auto`) with pipe-separated key=value context format, required fields (8 core EEG metrics), outcome determination rules by trigger type, mandatory before/after measurement flow, life-event implicit labeling (caffeine, meals, walks, meetings, sleep, exercise, app switches), hook trigger tracking, evidence aggregation by data maturity, personal protocol ranking algorithm (success_rate × avg_delta_target), 10 evidence-driven selection rules, evidence surfacing and tone guidelines, privacy safeguards, complete protocol name reference (100+ snake_case names), and a quick-reference integration pattern for other skills.

- **Non-breathing protocol repertoire**: Added 30+ new protocols across 6 categories for users who find breathing exercises inconvenient — Cognitive Reset (micro-tasking, peripheral vision, mental arithmetic, internal narration silence, gratitude, mental time travel, worry parking, category switching), Tactile & Haptic Regulation (texture scanning, pressure points, temperature contrast, bilateral tapping, structured fidget, palm press, hand warming), Oculomotor & Visual (saccade reset, candle gaze, colour hunting, slow tracking, panoramic vision, near-far focus), Micro-Movement / Discreet (isometric squeeze-release, jaw release, micro-walk, seated spinal wave, wrist/ankle circles, shoulder blade pinch, single-leg floor press), Auditory / Non-Music (sound mapping, humming, toning, environmental sound bath, whisper reading, rhythmic tapping), and Passive Physiological (dive reflex, deliberate yawning, chewing, gargling, wrist cooling, ear massage, tongue press, postural gravity drop). Updated matching guidance to prioritise non-breathing protocols when user preference or context indicates.

- **Protocol API Integration Guide**: Added a comprehensive guide teaching the LLM how to use the NeuroSkill API at every phase of protocol delivery — a 4-phase lifecycle (Before: validate trigger and baseline with `status`, `search_labels`, `label`; During: voice-guide with `say`, protect with `dnd_set`, monitor with `status` polling, time with `timer`; After: measure deltas, `label` outcomes, `notify` summaries, `say` results; Longitudinal: `search_labels` to find past protocol instances, `compare` protocol vs non-protocol sessions, `hooks_set` for auto-triggers, `umap` for visual separation, `sleep` for sleep protocol impact). Includes a 16-tool quick reference table, concrete JSON payload examples for each phase, and 10 integration principles.

- **Protocol API annotations (🔧)**: Grounded 17 key protocols with specific API call sequences — Theta-Beta Neurofeedback (tbr live feedback), Box Breathing (full lifecycle with auto-hook setup after 5 sessions), Cardiac Coherence (rmssd as live biofeedback), Loving-Kindness (faa shift tracking), Alpha Induction (real-time alpha feedback loop), Sleep Wind-Down (sleep staging comparison next morning), NSDR/Yoga Nidra (full voice-guided delivery via `say`), Flow State (minimal interruption + neural signature hooks), Micro-Tasking Focus (quick measurement), Study Focus Sprint (Pomodoro + data-driven scheduling), Morning Clarity (sleep summary + streak), Caffeine Timing (personal sensitivity mapping via labels), Digital Sunset (sleep_schedule-driven auto-trigger), Sensory Overload (immediate DND + pattern recognition), Hyperfocus Exit (stillness-based proactive hook), Between-Patient Reset (shift stress trajectory), and Post-Doom-Scrolling (app correlation + proactive nudge hooks).

- **Protocol personalisation engine**: Added a top-level Personalisation Engine to the protocol repertoire with 12 adaptation dimensions (age, physical ability, neurodivergence, emotional state, cultural background, gender/identity, professional context, social context, location/time, language, need urgency, intimacy level), 10 adaptation rules, and 12 concrete example adaptations showing the same EEG trigger delivered differently for an executive, new mother, teenager, construction worker, elder, autistic teen, night-shift nurse, grieving person, person with ADHD, pregnant woman, wheelchair user, and ESL speaker.

- **Context-specific protocol collections**: Added 11 new protocol sections (77 protocols total) for underserved user groups — Parenting & Caregiving, Elderly & Aging, Teens & Students, Neurodivergent-Friendly (ADHD, autism, OCD, RSD), Commuters & Travellers, Manual & Physical Workers, Healthcare & Shift Workers, Intimate & Relational, Accessibility-Adapted (visual/hearing/mobility/cognitive impairment, chronic pain), Culturally Diverse Practices (Chinese Wuqinxi, Mesoamerican Temazcal, Japanese Shinrin-Yoku, Hawaiian Ho'oponopono, African Ubuntu, Indian Bhramari, Aboriginal Dadirri, Tai Chi, Islamic Dhikr, Christian Centering Prayer, Hindu/Buddhist Mala meditation), and Situational Micro-Protocols (11 ultra-short real-moment interventions).

- **Diverse contextual variants across existing protocols**: Enhanced Attention, Stress, Emotional Regulation, Grounding, Sleep, Music, and Dietary protocols with persona-specific variant examples (marked with ◈) covering children, elders, wheelchair users, breath-averse users, public settings, non-English speakers, introverts, athletes, and more. Expanded music and food suggestions to span Western, South Asian, East Asian, Latin, African, K-pop, Arabic, and vegan cultures.

- **Status command enriched with usage analytics**: The `status` command now returns most-used apps (all-time, 24h, 7d), most frequent label texts (all-time, 24h, 7d), screenshot/OCR summary counts (total, with embedding, with OCR, with OCR embedding), top screenshot apps (all-time, 24h), and label text-embedding count. New fields: `apps`, `screenshots`, and expanded `labels` section in the JSON response.

### Performance

- **Eliminate PNG encode/decode round-trip in screenshot embed pipeline**: The capture thread now sends the pre-decoded `DynamicImage` directly to the embed thread instead of encoding it to PNG first. The embed thread calls fastembed's `embed_images()` with the `DynamicImage` directly, avoiding the CPU-intensive PNG encode (capture thread) → PNG decode (fastembed) round-trip. For LLM vision and OCR paths that still need encoded bytes, JPEG is produced lazily (~10× faster than PNG). OCR via `ocrs` now also operates on the decoded image buffer directly (`run_ocr_from_image`) instead of encoding to PNG and decoding again.

### Bugfixes

- **LLM screenshot tool commands**: Added `search_screenshots`, `screenshots_around`, `screenshots_for_eeg`, and `eeg_for_screenshots` to the `skill` tool's command enum, description, and alias resolution. The LLM can now invoke screenshot search commands correctly instead of failing with validation errors. Also maps CLI names (`search-images`, `screenshots-around`, etc.) to their WebSocket API equivalents.

- **CPU FFT fallback for CI**: Added `rustfft`-based CPU fallback in `skill-eeg` behind a feature gate. The `gpu` feature (opt-in) uses `gpu-fft` with wgpu; without it, tests and headless CI environments use pure-Rust `rustfft`. Fixes 33 test failures on GPU-less CI runners.

- **Lazy embedder retry**: When the text embedding model fails to initialise at startup (e.g. missing download, network error), semantic label search, interactive search, and re-embed commands now retry initialisation on demand instead of permanently returning "embedder not initialized".

- **Fix text embedder model resolution**: `build_embedder` used `EmbeddingModel::from_str` which only matches debug variant names (e.g. `BGESmallENV15`), not the `model_code` strings persisted in settings (e.g. `Xenova/bge-small-en-v1.5`). Added `resolve_embedding_model()` that first looks up by `model_code` from the supported-models list, falling back to the variant-name parser. Applied to both model init and `set_embedding_model` validation.

- **LLM skill tool args coercion**: When an LLM flattens command arguments to the top level (e.g. `{"command":"search_screenshots","query":"today"}` instead of `{"command":"search_screenshots","args":{"query":"today"}}`), the coercion layer now automatically wraps stray properties into the `args` object before validation. Also handles the common `"arguments"` misspelling as an alias for `"args"`. This prevents validation errors for all `skill` tool commands.

- **Missing skill commands in tool enum**: Added `sleep_schedule`, `sleep_schedule_set`, `health_summary`, `health_query`, `health_metric_types`, `health_sync`, `search_screenshots_vision`, and `search_screenshots_by_image_b64` to the skill tool command enum, description, and `is_skill_api_command()`. These WS commands were functional but invisible to the LLM.

- **Coerced tool arguments now reach execution**: The tool orchestrator validated and coerced LLM arguments (e.g. wrapping flat `query` into `args`) but discarded the coerced value before calling `execute_builtin_tool_call`. The executor re-parsed the original un-coerced string, so flattened skill args like `{"command":"search_screenshots","query":"today"}` passed validation but still failed at runtime because `args.get("args")` found nothing. Fixed both sequential and parallel execution paths to write coerced arguments back to `tc.function.arguments` before execution.

- **Generic hyphenated CLI name resolution**: Added a generic fallback in `resolve_skill_alias` that converts any hyphenated tool name to underscored form and checks if it's a valid skill API command. This catches LLMs copying CLI names from docs (e.g. `search-labels`, `session-metrics`, `sleep-schedule`, `dnd-set`) without needing to enumerate every variant.

- **Block LLM download/management commands**: Added `llm_download`, `llm_cancel_download`, `llm_pause_download`, `llm_resume_download`, `llm_refresh_catalog`, and `llm_logs` to the BLOCKED list in skill tool execution. These LLM self-management commands should not be callable from the LLM itself.

- **Fix missing description warnings for root skill files**: Added YAML frontmatter with `name` and `description` to `skills/SKILL.md`, `skills/README.md`, and `skills/METRICS.md` so the skill discovery scanner no longer emits "description is required in frontmatter" warnings.

### Refactor

- **3D PCA utility**: Added `pca_3d()` function to `skill-commands` for 3-component power-iteration PCA, complementing the existing `pca_2d()`.
- **SVG 3D generator**: Added `generate_svg_3d()` to `skill-commands::graph` — renders a dark-themed perspective-projected SVG with depth cues (scale, opacity, drop shadows) and a grid floor.

- **Extracted evidence collection from neuroskill-protocols**: Removed the ~190-line inline Evidence Collection section from the protocols skill and replaced it with a reference to the standalone neuroskill-evidence skill. The protocols skill now depends on the evidence skill for all measurement, labeling, and ranking rules, allowing any other skill to also follow the same evidence framework.

- **Split neuroskill-protocols into 11 domain sub-skills**: The monolithic 1866-line protocols skill has been split into a slim 411-line hub (personalisation engine, API integration guide, modality router, matching guidance, sub-skill index) plus 11 contextually-loaded domain sub-skills: `neuroskill-protocols-focus` (136 lines), `neuroskill-protocols-stress` (101 lines), `neuroskill-protocols-emotions` (110 lines), `neuroskill-protocols-sleep` (54 lines), `neuroskill-protocols-body` (106 lines), `neuroskill-protocols-routines` (117 lines), `neuroskill-protocols-nutrition` (129 lines), `neuroskill-protocols-music` (79 lines), `neuroskill-protocols-digital` (71 lines), `neuroskill-protocols-breathfree` (279 lines), and `neuroskill-protocols-life` (537 lines). Each sub-skill loads independently based on the user's message domain, enabling efficient use with small-context-window LLMs. The hub always loads when protocol intent is detected, and sub-skills load contextually alongside the neuroskill-evidence skill.

### Build

- **Bump cleans Rust artifacts**: `npm run bump` now runs `npm run clean:rust` at the end to remove `src-tauri/target`, freeing disk space after the preflight checks.

### CLI

- **CLI v1.2.0**: Bumped version to reflect new cross-modal search capabilities.
- Fixed misplaced shebang line that prevented `npx tsx cli.ts` from running.

### LLM

- **Cross-modal query guide in tool description**: Added a decision guide to the `skill` tool description that maps question patterns to the correct command and direction (e.g. "What was on screen during EEG?" → `screenshots_for_eeg`, "How was my brain when I saw X?" → `eeg_for_screenshots`). This helps the LLM pick the right cross-modal bridging command instead of guessing.

- **Status tool results formatted as readable text**: When the internal LLM calls the `status` command, the result is now converted from raw JSON to a human-readable text block with clear section headers (Device, Session, EEG Embeddings, Labels, Most Used Apps, Screenshots, Signal Quality, Current Scores, Hooks, Sleep, Recording History). This makes status output in the Chat window much easier to read for both the user and the model.

### Docs

- **Cross-modal screenshot↔EEG documentation**: Updated SKILL.md, neuroskill-screenshots skill, and skills index with documentation for `search-images --by-image`, `screenshots-for-eeg`, `eeg-for-screenshots` commands, cross-modal workflows, and all new WS endpoints.

- **SKILL.md updated for new and updated functionality**: Documented the enriched `status` command response with `apps` (top apps by window switches, all-time/24h/7d), `labels.top_all_time`/`top_24h`/`top_7d` (most frequent label texts), `labels.embedded` count, and `screenshots` summary (total, with_embedding, with_ocr, with_ocr_embedding, top_apps). Updated `interactive` search documentation to cover the new 5th screenshot layer, `screenshot` node kind with `filename`, `app_name`, `window_title`, `ocr_text`, `ocr_similarity` fields, `proj_z` 3-D PCA coordinate, new `screenshot_prox` and `ocr_sim` edge kinds, and `svg_3d` pre-rendered 3-D perspective SVG output. Added the `skill` built-in LLM tool to the tool-calling table with supported and blocked command lists.

- **Skills synced with SKILL.md changes**: Updated `neuroskill-labels` skill to document the 5th screenshot layer in interactive search (screenshot nodes with `filename`, `app_name`, `window_title`, `ocr_text`, `ocr_similarity` fields), new `screenshot_prox` and `ocr_sim` edge kinds, 3-D PCA projection fields (`proj_x`/`proj_y`/`proj_z`), and three SVG outputs (`svg`, `svg_col`, `svg_3d`). Updated `neuroskill-llm` skill with full `skill` tool documentation including supported commands, argument coercion, blocked self-management commands, and status formatting. Updated skill index description from 4-layer to 5-layer graph search.

- **Cross-modal workflow examples in skills**: Replaced the bash-only Cross-Modal Workflows section in the screenshots SKILL.md with a direction-based query guide table and multi-step LLM tool-call examples (JSON). Added Cross-Modal Follow-Ups sections to the search and labels SKILL.md files showing how to chain commands across modalities.

- **Screenshot & skills search tests**: Added comprehensive smoke tests for all 6 screenshot WS commands (`search_screenshots`, `screenshots_around`, `search_screenshots_vision`, `search_screenshots_by_image_b64`, `screenshots_for_eeg`, `eeg_for_screenshots`) covering semantic/substring modes, cross-modal EEG↔screenshot bridging, CLIP vision search, base64 image upload, error handling for missing/invalid fields, and result structure validation. Also added skills command rejection tests confirming Tauri-only commands are correctly rejected over WS/HTTP.

- **LLM tool call examples in skills**: Added `## LLM Tool Calls` sections with concrete JSON examples to all skill SKILL.md files (screenshots, labels, search, sessions, sleep, hooks, DND, streaming). This helps the LLM use the correct `{"command": "...", "args": {...}}` format.

- **Improved skill tool description**: Updated the `skill` tool description and `args` field description with explicit examples showing the `command` + `args` nesting pattern. Added SLEEP SCHEDULE and HEALTH command groups to the description.

- **Status SKILL.md**: Updated to document the new `apps` (top apps by window switches), `labels.top_*` (most frequent label texts), and `screenshots` (OCR counts, top apps) fields in the status response. Added LLM Tool Calls section with guidance on using `status` for app usage queries. Fixed JSON response example to show correct field names (`switches`, `last_seen`, `last_used`).

## [0.0.50] — 2026-03-21

### Features

- **Chat model picker in titlebar**: Click the model name in the titlebar to switch between downloaded models without leaving the conversation. The dropdown shows all downloaded models grouped by family with size info, and seamlessly stops the current model and loads the selected one.

- **LFM2.5 1.2B Instruct model**: Added LiquidAI/LFM2.5-1.2B-Instruct-GGUF to the LLM catalog with 7 quant variants (Q4_0, Q4_K_M, Q5_K_M, Q6_K, Q8_0, F16, BF16). Ultra-compact 1.2B-parameter text model that fits in under 2 GB VRAM.

- **LFM2.5 1.2B Thinking model**: Added LiquidAI/LFM2.5-1.2B-Thinking-GGUF to the LLM catalog with 7 quant variants (Q4_0, Q4_K_M, Q5_K_M, Q6_K, Q8_0, F16, BF16). Ultra-compact 1.2B-parameter reasoning model with chain-of-thought capability.

### Performance

- **Screenshot capture pipeline ~3× faster**: Eliminated redundant image encode/decode round-trips in the capture thread. `resize_fit_pad` no longer encodes to PNG (deferred to embed-thread send); `encode_webp` operates on the already-decoded `DynamicImage` instead of re-decoding from bytes; Linux/Windows xcap capture skips the ~500ms PNG encoding entirely by passing the decoded RGBA image directly. Expected improvement: ~2.9s → ~0.8s per iteration on Linux.

### Bugfixes

- **Fix label_store tests failing with "readonly database"**: The `TempDir` was dropped at the end of the helper function, deleting the temporary directory before the SQLite connection could write to it. The `TempDir` handle is now kept alive for the duration of each test.

### Build

- **Bump aborts on warnings**: `npm run bump` now captures stdout and stderr from every preflight check step and scans for warning lines. If any warnings are detected, the bump is aborted before any files are modified. `cargo clippy` is invoked with `-D warnings` to promote Rust warnings to errors.
- **Bump mirrors CI checks**: preflight now runs all CI-equivalent steps — `npm test` (vitest), `cargo clippy` on all workspace crates (not just the app crate), and `cargo test --lib` on the same crate subset as CI — so issues are caught locally before pushing.
- **Bump checks libpipewire-0.3**: added `libpipewire-0.3` to the Linux system dependency preflight check so the missing `-dev` package is caught early with a clear install hint instead of a cryptic cargo build failure.

### UI

- **Language dropdown close on outside click**: The language picker dropdown now stays open after selecting a language (showing the updated checkmark), and reliably closes when clicking anywhere outside the dropdown. Switched to `pointerdown` capture for robust outside-click detection across all window areas including drag regions.

- **Language flag padding**: Added padding around the language flag button in the title bar for better spacing across all windows.

## [0.0.49] — 2026-03-21

### Bugfixes

- **Fix a11y warnings in HistoryCalendar**: Replaced clickable `<div>` with a `<button>` for session bars in the calendar view, resolving svelte-check warnings about missing keyboard handlers and ARIA roles.

### Refactor

- **Fix clippy warnings in main crate**: Replaced `match` with `if let` in `worker.rs`, derived `Default` for `DndRuntimeState`, and used struct initializer for `InputTrackingState` in `state.rs`.

## [0.0.47] — 2026-03-21

### Performance

- **LLM E2E test: reduce CI time from ~15 min to ~2-3 min**: Lowered `max_tokens` from 512 to 128 and `max_rounds` from 2 to 1 for tool-chat test steps. The 1.6B model on CPU was generating max tokens every round at ~1.6 tok/s, and hallucinating calls to disabled tools caused extra rounds. Also added "Do NOT call any other tool" to system prompts to reduce hallucinated tool calls.

### Bugfixes

- **Fix false-positive "kill" safety check on "skill" commands**: `check_bash_safety` now uses word-boundary detection so patterns like `"kill "` no longer match inside words like `"skill"`. Commands such as `skill --help` or `neuroskill-status` no longer trigger the dangerous-command approval dialog. Actual `kill`, `killall`, and `pkill` commands (including after pipes/semicolons) are still correctly flagged.

- **Fix CI compilation and clippy errors**: Added missing `std::io::Cursor` import for Linux in `skill-screenshots`, replaced `.err().expect()` with `.expect_err()` in `skill-llm` tool orchestration, used `.is_multiple_of()` instead of manual modulo check, replaced `.map_or(true, …)` with `.is_none_or(…)`, fixed `needless_range_loop` clippy warnings in generation and actor modules, added `Default` impl for `ScreenshotMetrics`, simplified boolean expression in capture backfill check, and replaced `len() > 0` with `!is_empty()` in HNSW guards.

- **Fix unused `Cursor` import in skill-screenshots**: Tightened the `cfg` gate on `std::io::Cursor` in `platform.rs` so it is only imported when actually used (macOS, or Linux/Windows with the `capture` feature). Fixes a compile error with `-D warnings` on Linux CI without the `capture` feature.

## [0.0.46] — 2026-03-20

### Features

- **Add tests for untested modules**: Added 61 new unit tests across 4 previously untested modules:
  - `skill-eeg/eeg_model_config` (13 tests): config defaults, JSON round-trip, persistence save/load, corrupt file handling.
  - `skill-tools/types` (18 tests): `LlmToolConfig` defaults, dangerous tool safety, serialization, `ToolContextCompression` levels.
  - `skill-tools/defs` (21 tests): builtin tool definitions integrity, `is_builtin_tool_enabled` toggle logic, `resolve_skill_alias` routing.
  - `skill-tools/context` (9 tests): token estimation, context trimming, compression levels, system/user message preservation.

- **API bearer token authentication**: Added optional bearer token authentication for the HTTP/WS API. When `api_token` is set in settings, all requests must include `Authorization: Bearer <token>`. Empty token (default) disables auth — suitable for localhost-only binds. Configurable via Settings UI and `get_api_token`/`set_api_token` Tauri commands. i18n keys added for all 5 languages.

- **Add tests for `skill-constants` and `skill-router`**: 18 tests for skill-constants (filter math, band continuity, channel counts, Emotiv sample rate derivation, MutexExt poison recovery) and 11 tests for skill-router (rounding precision, NaN/infinity handling, command list integrity).

- **In-app device switching**: Cancel connection and switch to a different paired device directly from the main dashboard without using the tray icon menu. When connected with multiple paired devices, a "Switch Device" button appears in the footer. During scanning, a "Cancel" button and alternate device list are shown inline.

- **GPU memory safety guard**: Added pre-decode GPU memory checks to prevent Metal/CUDA `abort()` crashes when GPU memory is exhausted. The LLM engine now verifies sufficient free GPU/unified memory before starting prompt decode, multimodal decode, warmup, and periodically during token generation. When memory is too low, requests are rejected with a recoverable error message instead of crashing the entire application.

- **LLM E2E integration test with benchmarking and mock EEG data**: Full end-to-end Rust integration test (`crates/skill-llm/tests/llm_e2e.rs`) that: downloads a capable model (>=1.5B), starts the LLM server, runs a plain chat, a date tool-calling chat, and a NeuroSkill status tool-calling chat with a mock EEG API server returning realistic brain-state data (device info, signal quality, meditation/focus scores, session history, labels). Every step is benchmarked with timing and throughput (tok/s). All responses and tool events are captured and displayed in a formatted report. Runnable via `npm run test:llm:e2e`.

- **LLM VLM screenshot backend**: Added a new `"llm-vlm"` screenshot embed backend that uses the LLM vision model for both image embeddings (mean-pooled vision tokens via mmproj) and OCR (VLM-based text extraction via chat completion). This allows benchmarking VLM-based OCR against traditional OCR engines (ocrs / Apple Vision). Selectable in Settings → Screenshots → Embed backend. Also added `ocr_via_llm()` to the `ScreenshotContext` trait.

- **VLM image embedding benchmark in E2E test**: Added step 9 to the LLM E2E test that benchmarks VLM image embedding via the `EmbedImage` request path. Reports embedding dimensions and timing. Skipped with a warning when no mmproj is loaded.

- **LUNA model backend for EXG embeddings**: Added support for the [luna-rs](https://crates.io/crates/luna-rs) EEG foundation model as an alternative embedding backend alongside ZUNA. Users can now select between ZUNA (default) and LUNA in the EEG Model settings tab, with LUNA offering `base`, `large`, and `huge` model size variants. The selected backend and model size are persisted in `model_config.json`.

- **Embedding speed tracking**: Every embedding inference now tracks wall-clock time in milliseconds. The last embedding speed and an exponential moving average are displayed in the EEG Model tab and published via the WebSocket status endpoint. SQLite stores per-row `embed_speed_ms` for post-hoc performance analysis.

- **Model provenance in SQLite and HNSW**: Each embedding row now records which model backend (`zuna` or `luna`) produced it via the new `model_backend` TEXT column in the daily `embeddings` table. Historical rows without the column are auto-migrated on open. This enables filtering, auditing, and future re-embedding by model.

- **Per-model HNSW indices**: Each model backend now gets its own HNSW index file per day (`eeg_embeddings.hnsw` for ZUNA, `eeg_embeddings_luna.hnsw` for LUNA) and globally (`eeg_global.hnsw` / `eeg_global_luna.hnsw`). This prevents dimension mismatches when switching backends and allows side-by-side nearest-neighbor search for each model. The daily SQLite remains shared with a `model_backend` column to differentiate rows. Search APIs accept an optional model backend parameter to load the correct index.

- **Re-embed from raw EXG data**: Added `estimate_reembed` and `trigger_reembed` Tauri commands. The re-embed worker reads raw EEG samples from session CSV files (`exg_*.csv` / `muse_*.csv`), reads channel names and sample rate from the JSON sidecar, chunks data into 5-second epochs with 50% overlap, resamples to model input size, runs the selected encoder (ZUNA or LUNA) on the GPU, and writes new embedding rows to SQLite. Per-model HNSW indices are rebuilt per day and globally. Progress is streamed to the frontend via the `reembed-progress` event.

- **No overlap by default**: Changed the default embedding epoch overlap from 2.5 s to 0.0 s, so consecutive epochs no longer overlap. Users can still configure overlap via settings. This doubles the effective epoch interval from 2.5 s to 5.0 s, reducing redundant GPU work.

- **Add tests for `skill-autostart`**: 3 tests covering Linux XDG autostart enable/disable lifecycle, non-existent app safety, and disable idempotency.
- **Add tests for `skill-tts`**: 9 tests across config (defaults, JSON round-trip, empty JSON deserialization) and logging (enable toggle, write without callback, disabled noop).

- **Epoch-aligned screenshot interval**: Screenshot capture interval is now aligned with EEG embedding epochs (5 s). The slider offers multipliers from 1× (every 5 s) to 12× (every 60 s) in 5-second steps, replacing the old 1–30 second free-form slider. Legacy config values are automatically snapped to the nearest epoch boundary.

- **Tool argument type coercion**: `validate_tool_arguments` now coerces argument types before JSON Schema validation, fixing multi-model compatibility. Handles string-to-boolean (`"true"` → `true`), string-to-number (`"3"` → `3`), number-to-string (`42` → `"42"`), integer-to-float rounding, string-encoded JSON objects/arrays, and `"yes"`/`"no"`/`"on"`/`"off"` boolean aliases. Added `coerce_tool_call_arguments` public API for pre-execution coercion. Different LLM backends (Llama, Qwen, Mistral, Gemma, DeepSeek) emit arguments in subtly different type formats — this coercion layer normalises them transparently.

- **LLM helpers integration tests**: Added 65 tests in `llm-helpers.test.ts` covering the full download-progress lifecycle — unit tests for each helper, plus end-to-end scenarios simulating download start, window reopen, completion, failure, pause, multi-family downloads, shard progress, and stale-state recovery.

- **Dynamic context window growth**: When a prompt exceeds the current context window, the LLM engine now automatically attempts to grow the context size (up to the model's max trained context length) if VRAM/memory permits, before falling back to trimming chat history. This avoids unnecessary message loss on systems with enough memory.

- **Add tests for `skill-gpu`**: 3 tests covering struct construction, JSON serialization, and `read()` no-panic guarantee.

- **Add tests for `skill-llm` chat store**: 9 tests covering session CRUD, message save/load, tool call persistence, archive/unarchive, session params roundtrip, and temp directory isolation.

- **Add tests for `skill-history` cache**: 14 tests (was 5) covering downsample edge cases (empty, single element, max=0, max=1, evenly spaced), metrics cache path generation, sleep summary defaults, and sleep stage analysis with epochs.

### Performance

- **Raise minimum auto context size from 2048 to 4096**: The `recommend_ctx_size` heuristic no longer returns 2048 as a minimum, which was too small for most conversations with system prompts and tool results. The new floor is 4096 tokens.

- **Debounce settings persistence**: `save_settings()` is now debounced with a 500ms window. Multiple rapid settings changes (e.g., toggling several options in quick succession) are collapsed into a single disk write, preventing I/O storms. A `save_settings_now()` function ensures settings are flushed during app shutdown.

- **Remove excessive .clone() in hot loops**: Eliminated redundant `String`, `PathBuf`, and `Vec<f32>` allocations in search, index rebuild, and downsampling hot paths. Key changes:
  - `skill-commands`: Search functions now store indices into `date_dirs`/`day_indices` instead of cloning `(String, PathBuf)` per embedding and per HNSW hit; owned copies are only materialized for the final top-k candidates after truncation.
  - `skill-history`: `list_embedding_sessions` interns day-name strings via an index vector instead of cloning per DB row; session-gap loop references interned names by index.
  - `skill-history`: `downsample_timeseries` uses in-place `swap` + `truncate` instead of cloning large `EpochRow` structs into a new `Vec`.
  - `skill-history`: `analyze_search_results` borrows `&str` date references instead of cloning into the frequency map.
  - `skill-label-index`: `rebuild_indices` iterates `rows` by value so `Vec<f32>` embeddings are moved into HNSW insert instead of cloned.
  - `skill-screenshots`: HNSW rebuild loop iterates rows by value to move embeddings instead of cloning.
  - `skill-commands/graph`: Parent-ID dedup uses `as_deref()` + `&str` set instead of cloning `Option<String>` twice.

### Bugfixes

- **LLM crash on Metal buffer allocation failure**: Fixed a crash (`SIGABRT` in `ggml_metal_synchronize` → `ggml_abort`) that occurred when the Metal GPU backend failed to allocate buffers during `llama_decode`. The ggml abort is unrecoverable in-process, so pre-flight memory checks now prevent reaching that code path.
- **Reduced dynamic context growth memory budget**: Lowered the memory headroom multiplier for dynamic context window resizing from 85% to 70% of available GPU memory, reducing the risk of Metal OOM during large context operations.

- **Screenshot LLM token receiver**: Fixed `blocking_recv()` return type mismatch (`Ok(tok)` → `Some(tok)`) in `screenshot.rs` that prevented compilation with newer tokio mpsc channel API.

- **Test coverage expansion**: Added 164 new tests across frontend and Rust crates. Frontend: 3 new test files — `format.test.ts` (47 tests covering 31 formatting functions), `history-helpers.test.ts` (53 tests covering day math, label colors, duration formatting), `umap-helpers.test.ts` (32 tests covering easing, color conversion, geometry normalisation, gradient ticks). Rust: `skill-tools/exec.rs` (24 tests for truncation, path resolution, bash/path safety, UTC offset formatting), `skill-label-index` (8 tests for HNSW insert, search, rebuild, empty-state handling). Total test count: frontend 280→412, Rust skill-tools 58→82, skill-label-index 0→8.

- **Fix all 21 TypeScript type errors**: Resolved every `svelte-check` error across the codebase (was 21, now 0):
  - `ChatToolCard`: Added typed `arg()` helper for `Record<string, unknown>` tool args, typed `SourceEntry` interface for web search sources, cast `tu.result` through `Record` instead of direct `unknown` access.
  - `compare/+page.svelte`: Fixed `UmapProgress` cast through `unknown`.
  - `history-helpers.test.ts`: Fixed `LabelRow` field names (`wall_start` → `label_start`).
  - `umap-helpers.test.ts`: Fixed `UmapPoint` construction to match current interface.

- **Calibration timer drift**: Replaced sequential `sleep(1000)` countdown with wall-clock-based timing (`Date.now()`) to prevent cumulative drift over long calibration phases.
- **Calibration TTS desynchronization**: The break-phase "Next: …" announcement was fire-and-forget, causing it to queue behind the next action's TTS cue and delay the countdown start. Both break announcements now await completion before the countdown begins, ensuring audio and visual phases stay in sync.

- **Fix cubecl GlobalConfig panic loop on embedder respawn**: When the EEG embedder worker was respawned (e.g. after switching models), `configure_cubecl_cache` called `GlobalConfig::set()` a second time, causing a panic. Replaced the `Once` guard (which itself gets poisoned after a panic, triggering an infinite respawn loop) with an `AtomicBool` compare-exchange that silently skips the already-configured case.

- **Fix LUNA model download using wrong weights filename**: `download_hf_weights` was hardcoded to download `model-00001-of-00001.safetensors` (ZUNA) even when the LUNA backend was selected, causing repeated download failures because the LUNA repo uses `LUNA_base.safetensors`. Parameterised the function to accept `weights_file` and `config_file`, and updated the worker and settings commands to pass the correct filenames per backend.

- **DSP: use actual device channel count for QualityMonitor**: `SessionDsp::new` and the Emotiv desc-may-change reset path passed the compile-time constant `EEG_CHANNELS` (12) instead of the actual device channel count, producing extra phantom `NoSignal` quality entries for devices with fewer channels.

- **DSP: periodic status emit scales with sample rate**: The `process_eeg` status emit interval was hardcoded to `count % 256`, firing every ~1 s only at 256 Hz. Now uses the device's actual sample rate so the interval is ~1 s at any rate (128 Hz, 250 Hz, 500 Hz, etc.).

- **DSP: AC-coupled clip detection in QualityMonitor**: The clip-count check used absolute sample values, causing DC-coupled devices (e.g. Emotiv with ~4200 µV baseline) to report every sample as a clip and mark all channels as `Poor`. Clip detection now subtracts the window mean first, matching the existing AC-coupled RMS logic.

- **Add keys to critical `{#each}` loops**: Added `(key)` expressions to Svelte `{#each}` loops for session lists, device lists, day lists, settings tabs, search presets, and score keys across dashboard, history, search, compare, and settings pages. Prevents DOM thrashing when lists update.

- **Remove debug console.log**: Replaced 7 `console.log` calls in `UmapScene.svelte` with `console.debug` (filtered in production DevTools by default).

- **EEG embeddings: epochs now fire for all device channel counts**: The `EegAccumulator` used a fixed 12-element buffer array (`EEG_CHANNELS`) but checked *all* 12 buffers to decide when an epoch was ready. Devices with fewer channels (Muse=4, Emotiv Insight=5, Hermes=8) had empty inactive buffers whose `len()==0` prevented the epoch trigger from ever firing — meaning **no EEG embeddings were produced**. Now only active device channels (`0..device_channels`) are checked; inactive channels are zero-filled in the model input tensor.

- **EEG embeddings: correct resampling for non-256 Hz devices**: The accumulator already had `resample_linear()` and `native_epoch_samples` logic, but it was unreachable due to the channel-count bug above. Verified that the full path now works: native-rate samples are accumulated, and when an epoch is complete, each channel is resampled from `native_epoch_samples` to `EMBEDDING_EPOCH_SAMPLES` (1280) for the ZUNA model. Devices at 256 Hz skip resampling (identity path).

- **EEG embeddings: channel name padding for ZUNA preprocessing**: The `load_from_named_tensor` function requires `channel_names.len() == data.nrows()`. With 12-row zero-padded tensors but only 4 device channel names, this assertion would fail. The worker now pads channel names with synthetic `_padN` labels for inactive rows, which don't match any 10-20 electrode position and get default spatial coordinates.

- **Add diagnostic logging for Emotiv Cortex session creation**: When connecting to an Emotiv headset, the session creation wait loop silently discarded all non-SessionCreated events. Now logs each event type (Connected, Authorized, Warning, HeadsetsQueried, etc.) so connection issues can be diagnosed from the log output.

- **Move disk I/O outside AppState lock in set_eeg_model_config**: `save_model_config` (disk write) was called while holding the AppState mutex, which could block other subsystems (including the async Cortex connection) from acquiring the lock. Now the config is persisted after the lock is released.

- **LLM download progress lost on window reopen**: When starting a model download in LLM settings, closing the window, and reopening it, the download progress bar was not shown. The poll timer only refreshed the catalog when it already knew about an active download, creating a chicken-and-egg problem on fresh mounts. Fixed by always polling the catalog (a cheap in-memory read) so in-flight downloads are detected regardless of initial component state. Also added the missing `"paused"` variant to the frontend `DownloadState` type.

- **Fix LUNA crash on channels with mixed-case names**: LUNA's channel vocabulary uses uppercase names (e.g. `PZ`) but some devices like EMOTIV INSIGHT send mixed-case (e.g. `Pz`), causing a panic in `channel_indices_unwrap`. The embed worker now normalises channel names to uppercase and filters out any channels not in the LUNA vocabulary instead of panicking.

- **Fix screenshot HNSW panic on vision model change**: Switching the screenshot embedding backend (e.g. fastembed 512-dim → mmproj 2048-dim) caused a panic when inserting into the existing HNSW index built with a different dimension. Both vision and OCR HNSW indices now detect dimension mismatches and reset to a fresh index, with a log message suggesting re-embed to backfill.

- **Fix metrics-only SQLite insert failing with NOT NULL constraint**: When the GPU device was poisoned and the embedder fell back to metrics-only mode, `insert_metrics_only` tried to insert NULL for `eeg_embedding` which has a NOT NULL constraint in existing databases. Now inserts an empty blob for backward compatibility, and new databases use a nullable column.

- **Fix LUNA huge/large variant crash (GroupNorm dimension mismatch)**: The HuggingFace `config.json` for LUNA lacks per-variant hyperparameters, so all variants loaded with the `base` config (embed_dim=64, depth=8). The `huge` variant (embed_dim=128, depth=24) crashed with "Expected 16, got 32" in GroupNorm. Added `LUNA_VARIANT_CONFIGS` constants with correct dimensions for each variant and a `luna_variant_config_path` helper that generates a variant-specific config file before loading the encoder.

- **Fix mmproj "Use" button deadlock**: `set_llm_active_mmproj` and `set_llm_active_model` called `save_catalog()` which re-acquires the LLM mutex while it was already held, causing a deadlock that froze the UI. Switched to `save_catalog_locked()` which operates on the already-held lock guard.

- **Fix "prompt too long" error when prompt exceeds n_ctx**: The LLM engine now automatically trims older chat history messages (keeping system prompt and latest user turn) when the tokenized prompt exceeds the context window budget and the context cannot be grown further. Previously, conversations would fail with "prompt too long (N >= n_ctx M)" once history grew past the context limit.

- **Remove SQL expect() panics in screenshot and health stores**: Replaced 26 `.expect("static SQL")` / `.expect("SQL query")` calls in `skill-data` crate (`screenshot_store.rs`, `health_store.rs`) with graceful fallbacks (`let-else` returning empty vectors, `if-let` skipping failed prepare). The app no longer panics on database corruption, disk-full, or other runtime SQLite errors in these code paths.

- **Replace silent catch blocks with console.warn logging**: Added descriptive `console.warn` messages to ~80 silent `catch {}` / `.catch(() => {})` blocks across all frontend source files. Each warning includes a bracketed module tag (e.g. `[chat]`, `[tts]`, `[goals]`) and the failed operation name, making previously invisible failures easy to diagnose in the browser console. Affected files span routes (+page.svelte for home, chat, search, history, calibration, onboarding, settings, session, labels, api, compare, focus-timer), layout, and shared lib components (TtsTab, LlmTab, GoalsTab, SleepTab, ToolsTab, ScreenshotsTab, UpdatesTab, ChatSidebar, ChatMessageList, MarkdownRenderer, TtsTestWidget, HelpElectrodes, theme-store, window-title, i18n).

- **Replace `any` types with proper TypeScript types**: Replaced `any` with `unknown` in error catch callbacks (`LlmTab`, `SettingsTab`), added `instanceof Error` checks for error message extraction, typed Tauri invoke results with `Record<string, unknown>` instead of `any[]`, and replaced `as any` event payload casts with typed `Record` casts.

- **Remove `.unwrap()` from production Rust code**: Replaced all 137 `.unwrap()` calls in production (non-test) code with safe alternatives — `expect("reason")` for provably-safe invariants, `unwrap_or`/`unwrap_or_default` for fallible values, `let Some(...) = ... else { ... }` for early returns, and `?` for propagation. Prevents potential panics in session CSV/Parquet writers, PPG analysis, screenshot store SQL queries, health store, GPU stats, history cache, tool call parsing, TTS engine, LLM init, headless browser, autostart, job queue, device session adapters, tray icons, settings commands, and all Mutex lock sites. Test code retains `.unwrap()` as is standard practice.

- **Replace `.lock().expect("lock poisoned")` with `.lock_or_recover()`**: Converted all 37 occurrences of panicking mutex locks across 6 files (`api.rs`, `lib.rs`, `settings_cmds`, `ws_commands`, `llm/cmds/server.rs`, `llm/cmds/streaming.rs`) to use the poison-recovering `MutexExt::lock_or_recover()` trait. The app will now gracefully recover from poisoned locks instead of crashing.

- **Fix missing `format` field in LlmModel**: Added the required `format: ModelFormat::Gguf` field to the `LlmModel` constructor in `hardware_fit.rs`, fixing a compilation error introduced by an upstream `llmfit-core` dependency update.

### Refactor

- **Extract DND/sleep from ws_commands**: Moved DND status/set and sleep schedule get/set into dedicated `dnd_sleep.rs` sub-module. `ws_commands/mod.rs` reduced from 873 to 695 lines.
- **Delete orphaned `bt_monitor.rs`**: Removed 71-line dead file (Bluetooth radio check) that was never imported by any module.

- **skill-headless is now optional in skill-tools**: The `skill-headless` dependency (wry/tao headless browser) is now behind an optional `"headless"` feature (on by default). `skill-llm` depends on `skill-tools` with `default-features = false`, so the LLM E2E test no longer pulls in wry/tao (fixes macOS compile error). When headless is disabled, web_fetch/web_search gracefully fall back to plain HTTP.

- **Deprecate Muse-defaulting DSP constructors**: `BandAnalyzer::new()`, `ArtifactDetector::new()`, and `QualityMonitor::new()` are now `#[deprecated]` with guidance to use the sample-rate-aware variants (`new_with_rate`, `with_channels`, `with_window`). Added `FilterConfig::passthrough_with_rate(sr)` for non-Muse passthrough configs. This prevents accidental use of 256 Hz defaults with non-Muse hardware.

- **`EegAccumulator` tracks `device_channels`**: New field set by `set_device_channels()`, used to scope buffer checks and epoch building to active channels only. Buffers are also cleared on device change to prevent stale data from a prior session.

- **Unit tests for `resample_linear`**: Added 5 tests covering identity, upsample, downsample, empty source, and zero target cases.

- **Extract LLM helpers into testable module**: Moved all pure functions and types from `LlmTab.svelte` into `$lib/llm-helpers.ts` (vendor labels, quant ranking, family grouping, entry sorting, entry group splitting, family auto-selection, download detection). The component now imports from the shared module.

- **Extract `load_and_apply_settings` from `setup_app`**: Moved the 80-line settings hydration block into its own `#[inline(never)]` function, reducing `setup_app` from 574 to ~490 lines.

- **Consolidate frontend stores into `src/lib/stores/`**: Moved 12 scattered `.svelte.ts` store files from `src/lib/` root into a dedicated `src/lib/stores/` directory. Merged 5 tiny titlebar-related files (`titlebar-state`, `chat-titlebar`, `history-titlebar`, `label-titlebar`, `help-search-state`) into a single `stores/titlebar.svelte.ts`. Renamed remaining stores to drop the `-store` suffix (e.g. `theme-store` → `theme`, `toast-store` → `toast`). Updated ~50 import paths across the codebase.

- **Extract `skill-gpu` crate**: Moved GPU utilisation/memory stats from `skill-data::gpu_stats` (698 lines, 19 `unsafe` blocks) into its own standalone `skill-gpu` crate with zero Tauri dependencies. `skill-data` re-exports `skill_gpu::*` for backward compatibility.

- **Extract history page canvas rendering into `history-canvas.ts`**: Moved 3 pure canvas rendering functions (`renderDayDots`, `renderDayGrid`, `renderSparkline`) and the `heatColor` utility from the 2,224-line `history/+page.svelte` into a dedicated `src/lib/history-canvas.ts` module (280 lines). The history page now delegates to these functions via thin wrappers, reducing it to 1,983 lines (-241). All rendering logic is now testable independently of Svelte reactive state.

- **Move hooks CRUD + keyword suggestions into `hook_cmds.rs`**: Extracted `sanitize_hook`, `get_hooks`, `set_hooks`, `get_hook_statuses`, `suggest_hook_keywords` (with helper functions `norm_keyword`, `fuzzy_score`, `merge_suggestion`) from `settings_cmds/mod.rs` into the existing `hook_cmds.rs` sub-module. `mod.rs` reduced from 959 to 761 lines.

- **Remove `any` types from core interfaces**: Replaced all `any` annotations in `chat-types.ts`, `search-types.ts`, and `chat-utils.ts` with proper types (`Record<string, unknown>`, `unknown`, discriminated union `ContentPart`). Added `typeof` guards in `detectToolDanger` for safe property access. Added explicit casts at `JobPollResult` consumption sites in search and compare pages.

- **Extract `skill-health` crate**: Moved HealthKit data store from `skill-data::health_store` into its own standalone `skill-health` crate with zero Tauri dependencies. The new crate has 9 unit tests covering sync idempotency, per-table queries, metric type listing, and aggregate summaries. `skill-data` re-exports `skill_health::*` for backward compatibility — no consumer changes needed.

- **Split AppState into domain sub-states**: Extracted `LlmState` and `DndRuntimeState` behind their own `Arc<Mutex<>>` inside `AppState`, eliminating lock contention between LLM operations (model loading, chat streaming, downloads) and the EEG/device hot path. LLM and DND code now acquires independent locks without blocking device status reads or UI commands. Reduced the `new_boxed()` stack allocation from 32 MB to 8 MB.

- **Split AppState into domain sub-states**: Extracted `ShortcutState`, `UiPrefsState`, `InputTrackingState`, and `EmbeddingModelState` from the monolithic 50+ field `AppState` struct. Fields are now accessed via `state.shortcuts.*`, `state.ui.*`, `state.input.*`, and `state.embedding.*`. This improves code organization and prepares for future independent locking to reduce mutex contention.

- **Split `lib.rs` and `settings_cmds/mod.rs`**: Extracted 5 new modules from the two largest files in `src-tauri`. `lib.rs` reduced from 1,580 to 1,413 lines by moving the macOS external renderer (172 lines) into `external_renderer.rs`. `settings_cmds/mod.rs` reduced from 1,560 to 938 lines by extracting device commands (`device_cmds.rs`), activity tracking (`activity_cmds.rs`), screenshot config/search (`screenshot_cmds.rs`), and skills management (`skills_cmds.rs`). All re-exports preserved — no changes to the public API or `generate_handler!` invocation.

- **Split ws_commands/mod.rs into sub-modules**: Extracted calibration (7 commands), health (4 commands), and screenshot (2 commands) handlers into dedicated `calibration.rs`, `health.rs`, and `screenshots.rs` sub-modules. Reduced `mod.rs` from 1168 to 873 lines while keeping the dispatch table as the single routing point.

- **Clean up TODO.md**: Removed 4 completed items, keeping only active/open tasks.

### Build

- **Centralize workspace dependencies**: Added `[workspace.dependencies]` for `rusqlite`, `serde`, and `serde_json` to the root `Cargo.toml`. Updated 7 crates (`skill-health`, `skill-data`, `skill-commands`, `skill-llm`, `skill-router`, `skill-label-index`, `skill-history`) plus `skill-gpu` to use `{ workspace = true }`, ensuring a single version and feature set across the workspace.

- **Cargo workspace consolidation**: All 21 Rust crates now share a single Cargo workspace defined in the project root `Cargo.toml`. A shared target directory (`src-tauri/target/`) eliminates the ~3.6 GB of duplicate build artifacts that previously accumulated in per-crate `target/` directories under `crates/`. The `.cargo/config.toml` was moved from `src-tauri/` to the project root, and `[patch.crates-io]` / `[profile.dev]` sections were moved to the workspace root as required by Cargo. All CI and release workflows updated accordingly (`Swatinem/rust-cache` workspace paths, `Cargo.lock` hash keys, `cargo build -p skill` from workspace root).

- **LLM E2E test in CI**: Added `llm-e2e` job to the CI workflow that runs the full LLM integration test on main pushes and manual triggers. The HuggingFace model cache is persisted across runs. The full report is saved as a build artifact (`llm-e2e-report`) with 30-day retention, and the formatted report table is rendered in the GitHub Actions Job Summary. The Discord notification now includes the LLM E2E result.

- **Clean all clippy warnings and enforce in CI**: Fixed ~25 clippy warnings across 7 crates (`skill-eeg`, `skill-data`, `skill-tools`, `skill-headless`, `skill-skills`, `skill-history`, `skill-label-index`, `skill-tts`). Fixes include adding `Default` impls, replacing manual modulo checks with `.is_multiple_of()`, fixing doc indentation, removing redundant closures, using `is_some_and`, and properly iterating with `enumerate`. Added a workspace-wide `cargo clippy -- -D warnings` step to CI so new warnings are caught before merge.

- **Update LUNA HuggingFace repo to PulpBio/LUNA**: Changed the default LUNA model repository from `thorir/LUNA` to `PulpBio/LUNA` (`https://huggingface.co/PulpBio/LUNA`). Existing configs with the old repo are auto-migrated on load.

### UI

- **Full context viewer popup**: Added a "View full context" button to the context breakdown chart. Clicking it opens a full-screen modal showing every message in the context window (system prompt, user messages, assistant responses, tool calls/results) with role-colored headers, per-message token estimates, and a copy-all button. Closeable via backdrop click or Escape key.

- **Company logos in Supported Devices**: Each company section in the Devices tab now displays the company's logo (fetched as favicon from the official website) instead of the first device image preview. Logos added for Interaxon (Muse), Neurable, OpenBCI, Emotiv, IDUN Technologies, and RE-AK.

- **Scanning cancel button**: Added an inline Cancel button during the scanning state so users no longer need the tray menu to abort a connection attempt.
- **Device switcher panel**: When connected, a collapsible panel lists other paired devices with one-click "Switch" buttons that disconnect and reconnect automatically.

- **GPU memory threshold settings**: Added configurable GPU memory safety thresholds in Settings → LLM → Inference Settings. Users can set the minimum free GPU memory required before decode (default: 0.5 GB) and during generation (default: 0.3 GB), or disable the checks entirely.

- **Smarter onboarding LLM model selection**: The onboarding model picker now uses a priority chain: already-downloaded model → Qwen3.5 4B Q4_K_M → LFM2.5-VL 1.6B Q8_0 (ultra-compact fallback) → any recommended model (smallest first). This ensures low-memory devices get a working LLM out of the box.

- **Extract `HistoryStatsBar` component**: Moved the 75-line recording streak / stats bar from `history/+page.svelte` into a reusable `$lib/HistoryStatsBar.svelte` component with proper i18n (7 new keys: streak messages, days/hours/sessions labels, week trend). History page reduced from 1985 to 1924 lines.

- **Screenshot interval slider**: Updated to show epoch-aligned steps (5 s, 10 s, …, 60 s) with multiplier badge (e.g. "10s (2× epoch)").

- **Company logo white background**: Company logos in the Devices tab now always have a white background instead of a translucent muted background, ensuring consistent logo visibility in both light and dark modes.

- **Proactive Hooks accent consistency**: Replaced all hardcoded `primary` (neutral black/white) and `ring-ring` references with accent-remapped `violet-500` family in the Hooks tab. Keyword suggestion focus highlight, scenario select focus ring, distance suggestion panel (border, background, threshold text, percentile bar zones and markers), relative-age text, and the enabled checkbox now all honor the user's Appearance accent setting.

- **LLM quick test snippet selectable & copyable**: The curl code snippet in LLM advanced settings is now text-selectable (overriding the global `user-select: none`) and includes a Copy button for one-click clipboard copy.

- **Show correct encoder name and loading state in EEG Model settings**: The encoder status section was hardcoded to show "ZUNA Encoder" regardless of the selected backend. Now dynamically shows the correct name (e.g. "LUNA Encoder (huge)") based on the selected backend and variant. The status indicator dot also pulses blue while the encoder is loading on the GPU.

- **Screenshot settings accent consistency**: Replaced all hardcoded `primary` (neutral black/white) references with accent-remapped `violet-500` family across the entire Screenshots tab. Toggles, range sliders, progress bar, badges, select focus rings, icons, and the privacy note now all honor the user's Appearance accent setting.

- **Add SvelteKit root error boundary**: New `+error.svelte` at the route root catches unhandled errors across all 16 routes with a friendly message, "Go to Dashboard" link, and "Reload Page" button — prevents blank white screens on `invoke()` failures.

- **Improve accessibility with aria-labels**: Added `aria-label` and `aria-expanded` attributes to icon-only and toggle buttons across high-traffic pages (search, history, chat sidebar). Covers interactive search expand/collapse, +/- stepper controls, analysis panel toggle, close/dismiss buttons, chat new/archive/restore/delete buttons.

- **OpenBCI config moved into Device API card**: The standalone OpenBCI configuration section has been folded into the Device API card as a collapsible sub-section, alongside Emotiv Cortex and IDUN Cloud, for a more consistent layout.

- **Remove 2K context size option from LLM settings**: The 2K (2048 tokens) context size option was removed from the LLM inference settings UI. The backend auto-recommend (`recommend_ctx_size`) already treats 2048 as "too small for practical use" and never selects it — the minimum auto-recommended context is 4K. The default remains "auto", which intelligently picks the largest context size that fits in available GPU/unified memory. Users can still manually select 4K, 8K, 16K, 32K, 64K, or 128K.

- **Extract `HistoryCalendar` component**: Moved the 208-line calendar heatmap (year/month/week views) from `history/+page.svelte` into a reusable `$lib/HistoryCalendar.svelte` component. History page reduced from 1924 to 1735 lines.

- **Extract `OnboardingChecklist` component**: Moved the 34-line onboarding checklist from the dashboard `+page.svelte` into `$lib/OnboardingChecklist.svelte`. Fixed hardcoded "Dismiss" button → `t("common.dismiss")`. Dashboard reduced from 1678 to 1644 lines.

### LLM

- **Qwen3 30B-A3B Instruct**: Added `byteshape/Qwen3-30B-A3B-Instruct-2507-GGUF` to the LLM catalog. MoE architecture with 30B total / 3B active parameters — fast inference with strong reasoning. Five quant variants: Q3_K_S 2.70bpw (10.3 GB), Q3_K_S 3.25bpw (12.4 GB), Q4_K_S 3.61bpw (13.8 GB, recommended), Q4_K_S 3.92bpw (15.0 GB), and IQ4_XS 4.67bpw (17.8 GB).

### i18n

- **Context viewer translations**: Added i18n keys for the context viewer (en, de, fr, uk, he).

- **Device switcher translations**: Added new i18n keys for device switching UI in English, German, French, Ukrainian, and Hebrew.

- **GPU memory safety strings**: Added translation keys for the new GPU memory threshold settings (`llm.inference.gpuMemThreshold`, `gpuMemThresholdDesc`, `gpuMemDecode`, `gpuMemGen`) in all languages (en, fr, uk, de, he).

- **Pre-commit hook enforces i18n key synchronisation**: Added `.githooks/pre-commit` that runs `npm run sync:i18n:check` when any file under `src/lib/i18n/` is staged. Blocks commits with missing translation keys and guides the developer to run `npm run sync:i18n:fix`. The hook is automatically activated via `postinstall` (`git config core.hooksPath .githooks`) and skips entirely when no i18n files are changed (~0 ms overhead on normal commits).

- **Screenshot interval strings**: Updated English, German, French, Hebrew, and Ukrainian translations for the epoch-aligned interval description and added `screenshots.intervalEpoch` key.

- **Migrate hardcoded English strings to i18n**: Replaced 12 raw English `title` attributes with `t()` calls across 8 components (ChatMessageList, CustomTitleBar, DevicesTab, TimeSeriesChart, dashboard, compare, api). Added 13 new i18n keys (`common.copy`, `common.copyToClipboard`, `common.newer`, `common.older`, `common.clickToHide`, `common.clickToReveal`, `common.goalReached`, `common.resetZoom`, `common.openComparison`, `common.error`, `error.description`, `error.goHome`, `error.reload`) for all 5 languages.

- **Add `common.dismiss` key**: New i18n key for all 5 languages, replacing the hardcoded "Dismiss" string in the onboarding checklist.

### Docs

- **Add doc comments to `skill-router` public API**: Added `///` documentation to all 6 rounding helper functions (`r1`, `r2`, `r3`, `r1d`, `r2d`, `r2f`).

### CI

- **Run Rust tests in CI**: Added `cargo test` step to the CI pipeline covering 13 testable crates (486 tests). Previously CI only ran `cargo check` and `clippy` — tests were never executed.
- **Add new crates to CI clippy**: Added `skill-health`, `skill-gpu`, `skill-screenshots`, and `skill-llm` to the workspace clippy check (were missing since extraction).

## [0.0.45] — 2026-03-19

### Features

- **Auto-connect to paired devices**: When the BLE scanner discovers a previously paired device while the app is idle (disconnected, no active session or pending reconnect), a session is automatically started. No cooldown is needed — `start_session()` immediately marks the app as connecting, preventing duplicate attempts, and the normal retry backoff handles failures.

- **Emotiv multi-headset selection**: when multiple Emotiv headsets are paired in the EMOTIV Launcher, the scanner now lists each one individually (e.g. `EPOCX-A1B2C3D4`, `INSIGHT-5AF2C39E`) in the discovered devices list instead of a single generic "Emotiv (Cortex)" entry. Users can pair and connect to the specific headset they want. The selected headset ID is passed to the Cortex API so the correct device is targeted.

- **Comprehensive GIF recording for all app UIs**: Added `scripts/screenshots/take-gifs.mjs` — a Playwright-based tool that records 58 animated GIFs covering every page, tab, toggle, expandable section, and hidden parameter panel in the app. Supports `--filter`, `--theme`, and `--list` CLI flags. Covers: dashboard (full scroll, electrode guide expand/collapse, collapsible sections), all 18 settings sub-tabs with toggle-revealed panels (DND automation with threshold/duration/lookback/SNR, OpenBCI config, calibration editor, LLM advanced inference, tool toggles with web search provider, screenshot OCR and metrics, skills), chat (sidebar, settings panel, tools panel), search (EEG/Text/Images modes), history (session expand), session detail, compare, help (all 11 tabs with scroll), calibration electrode tabs, onboarding wizard, labels search modes, focus timer config, downloads, API code examples, about, and what's new.

- **IMU data recording**: Devices with IMU sensors (Muse, Hermes, Emotiv, IDUN) now record accelerometer, gyroscope, and magnetometer data to `exg_<ts>_imu.csv` (or `.parquet`). Data includes `timestamp_s`, `accel_x/y/z`, `gyro_x/y/z`, `mag_x/y/z` columns.
- **Storage format selector in Settings**: Added a Recording Format picker (CSV / Parquet / Both) to the Settings tab. The "Both" option writes CSV and Parquet files simultaneously for every data stream (EEG, PPG, IMU, metrics).
- **GPU and memory stats moved to Settings**: The GPU / Unified Memory (RAM) panel is now shown in the Settings tab instead of the EXG tab, where it logically belongs alongside other system configuration.

- **Scanner backend settings**: New "Scanner Backends" section in the Devices tab with toggles for each transport (BLE, USB Serial, Emotiv Cortex). Changes are persisted in `settings.json` and take effect on next app restart.

- **Emotiv Cortex connection indicator**: The Cortex scanner toggle shows a live "Connected to Cortex" / "Not connected" badge based on whether Emotiv devices have been discovered via the Cortex WebSocket API.

- **Device log viewer**: New collapsible "Device Log" panel in the Devices tab showing a live, color-coded log of scanner and session events (BLE discovery, USB detection, Cortex polling, connect/disconnect, watchdog). Auto-refreshes every 3 seconds. Entries are kept in a 200-entry ring buffer.

- **Transport badge on devices**: Discovered devices now show a transport badge (USB, Cortex, WiFi) next to their name when they were found via a non-BLE transport.

- **Animated GIF capture for scrolling/animated windows**: The screenshot capture worker now detects motion between consecutive frames using pixel-difference scoring. When the change exceeds a configurable threshold (default 5%), a rapid burst of frames is captured and encoded as an animated GIF alongside the still WebP screenshot. New config fields: `gif_enabled`, `gif_frame_count`, `gif_frame_delay_ms`, `gif_motion_threshold`, `gif_max_size_kb`. The middle frame of the burst is used as the representative image for CLIP embedding and OCR. GIFs exceeding the size limit are automatically discarded.

- **Emotiv adapter test coverage**: Added tests for EEG translation, headset disconnect on stop-all-streams/close-session warnings, error-to-disconnect mapping, and warning filtering.

- **Tool thinking budget**: Added a configurable thinking budget override for tool-calling rounds in the Tools settings panel. Options: Chat (use chat-level setting), None (0), 256, 1K, 4K tokens. Lower values make the model respond faster after tool results. Stored in `LlmToolConfig.thinking_budget`.

### Bugfixes

- **Emotiv auto-connect no longer hijacks first headset**: Cortex devices are no longer blindly auto-connected as "trusted transport". Only explicitly paired headsets trigger auto-connect, preventing the first headset from being grabbed when multiple are available. Legacy `cortex:emotiv` paired entries are still honored for backward compatibility.
- **Emotiv scanner discovers headsets on first tick**: the Cortex scanner runs its first probe immediately at startup (before the 900 ms auto-connect fires) so all headsets are discovered and visible in the device list. Subsequent probes are skipped while a session is active to avoid invalidating the session's cortex token.

- **Data watchdog for silent BLE disconnects**: Added a 15-second data watchdog to the session event loop. If no device event arrives within the timeout, the connection is treated as silently lost and auto-reconnect is triggered. This catches scenarios where the BLE link stays alive but GATT notifications stop flowing (radio interference, device sleep, firmware hang).

- **Reconnect retry limit**: Auto-reconnect now gives up after 12 consecutive failed attempts (~51 seconds of total backoff) instead of retrying indefinitely. A toast notification informs the user to reconnect manually when the limit is reached. This prevents draining battery on a device that was intentionally turned off or moved out of range.

- **Disable GIF burst capture by default**: Changed `gif_enabled` default from `true` to `false` so the app only takes still screenshots during normal operation. GIF burst capture (motion detection + multi-frame capture) is intended for use in scripts only and can be explicitly enabled when needed.

- **Disable GIF capture in periodic screenshot loop**: The normal app screenshot worker no longer produces animated GIFs via motion detection. GIF burst capture is now reserved exclusively for scripts. The `gif_encode` module and config fields are preserved for the script-level API.

- **Emotiv disconnect cleanup**: when an Emotiv headset disconnects before any EEG data is recorded (e.g. powered off right after connecting, or Cortex session interrupted), the app now properly transitions to the disconnected state. Previously the UI would stay stuck on "connected" because `go_disconnected` was only called when the CSV recording file had been opened — which requires at least one EEG frame.
- **Disconnect events for all break paths**: the data watchdog timeout and event-channel-closed paths in the session runner now call `on_disconnected` to emit the `device-disconnected` event and toast, consistent with the explicit `DeviceEvent::Disconnected` path.
- **Emotiv headset disconnect/failure warnings**: the adapter now handles Cortex warning codes 102 (HEADSET_DISCONNECTED) and 103 (HEADSET_CONNECTION_FAILED) in addition to codes 0 and 1. This triggers an immediate disconnect instead of waiting for the 15-second data watchdog timeout.
- **Emotiv EEG subscribe error no longer causes infinite reconnect loop**: subscribe failures (e.g. Cortex error -32230 "stream not supported") are no longer treated as disconnects. The connect flow now fails with a clear user-facing error explaining the license requirement, and auto-reconnect is disabled for non-recoverable configuration errors.

- **Emotiv subscribe race condition**: `connect_emotiv` now waits for the `SessionCreated` event before calling `subscribe`. Previously it called `subscribe` immediately after `client.connect()`, but `connect()` only opens the WebSocket — the auth flow (hasAccessRight → authorize → queryHeadsets → createSession) runs asynchronously. The subscribe was sent with an empty cortexToken and session ID, causing an immediate `-32014 Cortex token is invalid` error.

- **Emotiv session stability**: Upgraded to emotiv crate v0.0.4 which prevents `ACCESS_RIGHT_GRANTED`, `HEADSET_CONNECTED`, and `HEADSET_SCANNING_FINISHED` warning handlers from re-authorizing or re-querying headsets when a session is already active.

- **Emotiv scanner is now side-effect-free**: The Cortex scanner probe only authorizes — it does NOT send `queryHeadsets` or `getCortexInfo`. The scanner also skips polling entirely when a session is active or a reconnect is pending, and waits 5 seconds at startup to avoid racing with the auto-connect flow.

- **Emotiv auto-connect without pairing**: Cortex-discovered and USB-discovered devices are now treated as trusted transports and auto-connect when the app is idle, without requiring manual pairing first. BLE devices still require pairing as before (since BLE advertisements can come from any nearby device).

- **Emotiv reconnect uses correct device ID**: `start_session` now pins the scanner-level device ID (e.g. `"cortex:emotiv"`) into `status.device_id` before the adapter runs. This ensures `on_connected` pairs the device with the correct ID (instead of the Cortex session ID), and reconnect retries route to `connect_emotiv` via the `cortex:` prefix.

- **Device kind routing by ID prefix**: `detect_device_kind` now checks the device ID prefix (`cortex:` → emotiv, `usb:` → ganglion) before falling back to name-based detection. This ensures Cortex-discovered devices route to `connect_emotiv` regardless of their headset ID format.

- **Emotiv EEG subscribe confirmation**: after subscribing to Cortex streams, the connect flow now waits up to 3 seconds for the EEG DataLabels response to confirm the subscription succeeded. If EEG subscription fails (e.g. due to a missing license), the error is logged and a toast is shown instead of silently streaming IMU-only data with empty EEG channels. Events consumed during the confirmation wait are replayed into the adapter so DataLabels are not lost.

- **Emotiv headset disconnect detection**: The Emotiv adapter now translates Cortex API warning codes `CORTEX_STOP_ALL_STREAMS` (0) and `CORTEX_CLOSE_SESSION` (1) into `DeviceEvent::Disconnected`, giving the session runner instant notification when a headset goes away instead of waiting up to 15 seconds for the data watchdog to fire. `CortexEvent::Error` is also surfaced as a disconnect to trigger immediate reconnection.

- **Emotiv sample rate per model**: EPOC X, EPOC+, EPOC Flex, Insight 2, MN8, and X-Trodes now correctly report 256 Hz instead of the hardcoded 128 Hz. The sample rate is derived from the headset ID prefix (e.g. `EPOCPLUS-*` → 256 Hz, `INSIGHT-*` → 128 Hz). This affects DSP filter configuration, band analysis, artifact detection, and CSV recording timestamps.

- **Emotiv device visible in UI with correct channel count**: The dashboard now uses dynamic `channel_names` from the connected device when available, instead of always using the hardcoded 14-channel EPOC layout. This fixes Emotiv Insight (5ch), MN8 (2ch), and Flex (32ch) devices showing wrong/missing EEG waveforms. Colors auto-extend by cycling the palette for high-channel-count devices.

- **Emotiv headset name in UI**: The Emotiv adapter now reports the actual headset ID (e.g. "INSIGHT-5AF2C39E") as the device name instead of the generic "Emotiv", so the dashboard and session metadata show which model is connected.

- **LLM skill sub-command auto-redirect**: When the LLM calls a Skill API sub-command (e.g. `status`, `say`) or a `neuroskill` alias (e.g. `neuroskill`, `neuroskill-status`, `neuroskill-hooks`) as a top-level tool, the call is silently rewritten to `skill` with the correct `{"command": "..."}` at three layers: extraction (parse.rs), validation (tool_orchestration.rs), and execution (exec.rs).
- **LLM dedup loop fix**: When the model re-emits the same tool call on round 2 (all calls deduped), instead of returning empty text, the orchestrator now injects a nudge message telling the model to summarize the existing results, then continues to a new inference round.
- **Tool result not misdetected as tool call**: JSON objects with `"ok"` or `"command"` keys (tool results) are no longer falsely extracted as tool calls when the model quotes them in its response.
- **Model not responding after tool call**: Improved tool result message prefix to prevent the model from interpreting tool results as new user questions. Assistant messages now include tool call details.

- **Devices are no longer auto-paired**: previously every device that connected was automatically added to the paired list, even without the user clicking "Pair". Now only explicitly paired devices are remembered. The single exception is first-time onboarding: if no devices are paired at all, the first successful connection auto-pairs as a convenience so new users can test immediately.
- **Auto-connect requires explicit pairing**: the scanner no longer auto-connects USB or other "trusted transport" devices without pairing. Only devices the user has explicitly paired (or first-time onboarding) trigger auto-connect.
- **Startup auto-connect skipped when no paired devices**: on first launch with no paired devices, the app no longer blindly scans and connects to the first device it finds. The user must discover and pair a device manually.

- **Paired Emotiv devices no longer appear in discovered list**: discovered Cortex devices (e.g. `cortex:EPOCPLUS-06F2DDBC`) now correctly match against the legacy `cortex:emotiv` paired entry, so paired headsets show in the "Paired" section instead of "Discovered". On first successful connection, the legacy ID is automatically migrated to the real headset ID.

- **Fix rustls CryptoProvider panic at startup**: Multiple transitive dependencies (`tauri-plugin-updater`, `emotiv`, `hf-hub`/`fastembed`) activated both the `ring` and `aws-lc-rs` features of rustls 0.23, preventing automatic provider selection. The app now explicitly installs the `ring` crypto provider at the start of `run()`, fixing the "Could not automatically determine the process-level CryptoProvider" panic.

- **Signal quality limited to actual electrodes**: The `status` WebSocket command now returns `signal_quality` entries only for electrodes that exist on the connected device, instead of always returning 12 entries (padded with `no_signal` for non-existent channels). The quality vector is also cleared on disconnect and at startup.

- **Unknown tool calls show misleading "disabled" error**: When the LLM hallucinates non-existent tool names (e.g. "status", "neuroskill-status"), the error now correctly says "unsupported tool" with guidance to use available tools, instead of the misleading "tool disabled in settings".

### Refactor

- **Extracted shared Tauri mock**: Moved `buildTauriMock()` from `take-screenshots.mjs` into a shared `scripts/screenshots/tauri-mock.mjs` module, imported by both the screenshot and GIF scripts.
- **Enhanced Tauri mocks**: Added full DND config (enabled by default with all sub-settings visible), complete LLM tools config (web search, execution mode, context compression, skills), expanded screenshot config (all sliders/pickers), and skills/license mocks.

- **Settings tab cleanup**: Removed device listings (Supported Devices, Paired/Discovered Devices, OpenBCI config, Device API) from the Settings tab. These are already available in the dedicated Devices tab.
- **Settings tab cleanup**: Removed Signal Processing and EEG Embedding sections from the Settings tab. These are already available in the dedicated EXG tab.
- **StorageFormat enum**: Extended `StorageFormat` with a `Both` variant and `as_str()` method. `SessionWriter` now supports `Both(CsvState, ParquetState)` for dual-format recording.
- **Session cleanup**: `delete_session` now removes IMU data files and Parquet variants alongside CSV files.

- **Scanner log tag**: Added `scanner` subsystem to the logging system (`LogConfig`) so scanner events can be toggled independently from `bluetooth` session events.

- **New `gif_encode` module in `skill-screenshots`**: Extracted GIF encoding and representative-frame extraction into a dedicated module (`gif_encode.rs`) with `encode_gif()` and `representative_frame_png()` helpers.
- **Motion detection and burst capture in `platform.rs`**: Added `motion_score()` for pixel-diff comparison and `capture_burst()` for rapid multi-frame capture.
- **`gif_filename` column in screenshot store**: Added SQLite migration and `update_gif_filename()` method to `ScreenshotStore`. All query result types (`ScreenshotResult`) now include the `gif_filename` field.

- **Renamed `bt_error` to `device_error`**: the error field in `DeviceStatus` was named `bt_error` (Bluetooth-specific) but is used for all device connection errors including Cortex WebSocket (Emotiv), USB serial (OpenBCI), and BLE. Renamed to `device_error` throughout the backend, frontend types, and dashboard UI to reflect the transport-agnostic nature of the field. Also renamed `classify_bt_error` → `classify_device_error`.

- **Generalized device scanner**: Replaced `ble_scanner.rs` with `device_scanner.rs` — a unified background scanner that runs multiple transport-specific backends in parallel:
  - **BLE** — discovers Muse, MW75, Hermes, Ganglion, IDUN devices (existing logic, unchanged)
  - **USB serial** — polls for OpenBCI Cyton/CytonDaisy dongles by detecting FTDI FT231X USB VID/PID and common port patterns (ttyUSB, usbserial), every 5 seconds
  - **Cortex WebSocket** — checks for Emotiv headsets via the local EMOTIV Launcher service (`wss://localhost:6868`), every 10 seconds; only polls when Emotiv credentials are configured
  - All backends share the same auto-connect logic: if a paired device is discovered while idle, connect automatically

- **Transport tag on discovered devices**: `DiscoveredDevice` now carries a `transport` field (`ble`, `usb_serial`, `wifi`, `cortex`) so the UI can display a transport badge. The transport is inferred from device ID prefixes (`usb:`, `cortex:`, or BLE by default).

- **Split EXG tab from Devices tab**: the EXG tab now has its own view with Signal Processing (notch/high-pass/low-pass filters), EEG Embedding (epoch overlap), and GPU/Memory stats. The Devices tab retains paired/discovered devices, supported devices, OpenBCI config, Device API credentials, Scanner Backends, and Device Log.

### Build

- **New npm scripts**: Added `npm run screenshots` and `npm run gifs` convenience commands.
- **New dev dependencies**: Added `gif-encoder-2` and `sharp` for GIF frame encoding and resizing.

### UI

- **Emotiv EEG license error page**: when the Cortex API rejects the EEG stream subscription (-32230), the dashboard shows a dedicated violet card explaining the issue with a "Manage Emotiv Account" button that opens the Cortex Apps settings page directly.

- **Device info badge shows actual sample rate**: the dashboard device badge (e.g. "EPOCPLUS-06F2DDBC · 14ch · 256 Hz") now reads the sample rate from the backend status instead of being hardcoded. All non-Muse device badges (Ganglion, Emotiv, IDUN, Hermes) were updated.

- **Tool card i18n fallback for unknown tools**: The chat tool card now shows the raw tool name instead of a raw i18n key (e.g. "status" instead of "chat.tools.status") when the LLM calls an unrecognized tool.

- **Compact supported devices list**: Collapsed supported devices into an accordion layout — each company shows as a single row with a thumbnail preview and device count. Device image grid and connection instructions are only revealed on expand. Reduces vertical space usage significantly when browsing the Devices tab.

### LLM

- **`neuroskill` tool alias**: `neuroskill` is recognized as an alias for `skill`. Hyphenated forms like `neuroskill-status`, `neuroskill-sessions`, `neuroskill-hooks` map to the corresponding API command.
- **Skill tool description restructured**: Compact grouped categories instead of a flat list of 30+ commands. The `command` parameter includes a JSON Schema `enum` constraint.
- **Skill tool call examples added**: Both full and compact tool prompts include an explicit `skill` calling example.

### i18n

- **Scanner & storage format translations**: Added 23 missing i18n keys to de, fr, he, and uk locales — scanner backends, device log, log scanner, and recording format settings.

### Docs

- **UI Guide**: Added `docs/UI.md` — comprehensive visual walkthrough of every screen, panel, and interaction in the app. Uses GitHub collapsible sections with animated GIFs showing real UI recordings and static light/dark PNGs. Covers: Dashboard (metrics, electrode guide, collapsible sections), Settings (all 18 tabs with toggle-revealed panels), Chat (sidebar, settings/tools panels), Search (EEG/Text/Images modes), History (day view, session expansion), Session Detail, Compare, Help (all 11 tabs), Calibration, Onboarding, Labels, Focus Timer, Downloads, API, About, and What's New.

- **Update all crate README files**: Synchronized README.md files across all workspace crates with current source code. Added missing modules (`health_store`, `session_parquet`, `session_writer` in skill-data), documented the full device adapter system with 6 hardware drivers in skill-devices, expanded skill-constants with all constant groups (device-specific, LLM, tool-calling, WebSocket, DND, calibration, app metadata, skills), updated skill-settings for `UserSettings` rename and new types (`DeviceApiConfig`, `SleepConfig`, `DoNotDisturbConfig`, `ScreenshotConfig`), documented skill-llm engine sub-modules (`actor`, `generation`, `protocol`, `state`, `think_tracker`, `tool_orchestration`), expanded skill-vision with full API docs, and added `skill-headless` and `skill-skills` to the AGENTS.md workspace crates table.

### Dependencies

- **emotiv**: bumped from 0.0.5 to 0.0.7 — adds `CortexEvent::HeadsetsQueried` and `CortexHandle::query_headsets()` for safe headset enumeration; guards `connect_headset`/`create_session` side effects behind `auto_create_session` flag.

- **emotiv**: bumped to 0.0.9 — adds `HEADSET_DISCONNECTED` (102) and `HEADSET_CONNECTION_FAILED` (103) warning constants; emits `CortexEvent::Disconnected` immediately when either is received; logs failed stream subscriptions.

- **emotiv**: bumped to 0.0.8 — failed stream subscriptions are now logged and emitted as `CortexEvent::Error` instead of being silently ignored.

## [0.0.44] — 2026-03-19

### Features

- **Skills auto-refresh**: Community skills are now periodically downloaded from GitHub to `~/.skill/skills/`. Users can configure the refresh interval (off / 12 h / 24 h / 7 d) or trigger a manual sync from the Tools settings tab. A background task checks freshness and downloads the latest tarball when stale. The new `sync` feature in `skill-skills` handles download, extraction, and metadata tracking via a `.skills_last_sync` sidecar file.
- **Skills download on onboarding**: Community skills are automatically downloaded when onboarding completes, so fresh installs have the latest skills available immediately.
- **Agent Skills settings card**: Separate card in the Tools tab listing all discovered skills with their descriptions (pulled from SKILL.md frontmatter). Individual skills can be toggled on/off, with bulk Enable All / Disable All actions. Disabled skills are excluded from the LLM system prompt. Changes are live-applied to the running LLM server without restart.

### Bugfixes

- **Missing `skill_api` in chat tool config**: Added the missing `skill_api` property when constructing `toolConfig` from loaded LLM config, fixing a TypeScript error.

- **Missing Cursor import on macOS**: Added `use std::io::Cursor` to `skill-screenshots/src/platform.rs` to fix compilation error on macOS.

### UI

- **Skill API tool toggle**: The NeuroSkill Skill API tool can now be enabled/disabled from both the Settings > Tools tab and the Chat tools panel, matching all other built-in tools. Previously the Skill API tool was always injected when available with no user-facing toggle.

### i18n

- **Skill API tool strings**: Added i18n labels and descriptions for the Skill API tool toggle in English, German, French, Hebrew, and Ukrainian.

## [0.0.43] — 2026-03-18

### Features

- **Auto context size recommendation**: When the user has not set an explicit context size, the LLM server now uses `llmfit`-style memory estimation to pick the largest power-of-two context (2K–128K) that fits the system's available GPU/unified memory with 15% headroom, instead of always defaulting to 4096. Each catalog entry now carries `params_b` and `max_context_length` metadata so the estimator knows the model's parameter count and trained context ceiling. User-set context values are capped at the model's maximum. The UI context-size picker dynamically hides options that exceed the active model's trained limit and now offers 64K and 128K choices for models that support them.

- **Parquet data consumption across the app**: All data-reading paths now check for both `.parquet` and `.csv` files, preferring Parquet when it exists. This ensures sessions recorded in Parquet format are fully visible in history, metrics analysis, session search, and the metrics cache.
  - `find_metrics_path` / `find_ppg_path` helpers try `.parquet` then `.csv`
  - `load_metrics_csv` dispatches to `load_metrics_from_parquet` for `.parquet` files
  - `read_metrics_time_range` handles both formats for timestamp patching
  - `is_session_data` matches both `.csv` and `.parquet` EEG data files
  - `extract_timestamp` strips `.csv`, `.parquet`, and `.json` suffixes
  - `skill-commands` session lookup resolves `.parquet` before `.csv`
  - Metrics disk cache validates mtime against whichever data file exists
  - File size reporting checks for `.parquet` data files

- **Parquet recording format**: EEG, PPG, and metrics data can now be stored in Apache Parquet format (Snappy compression) instead of CSV. Set `storage_format: "parquet"` in settings or use the `set_storage_format` Tauri command. Default remains CSV for backward compatibility.
  - `exg_<ts>.parquet` — raw EEG samples (timestamp + N channel columns)
  - `exg_<ts>_ppg.parquet` — PPG optical data with vitals
  - `exg_<ts>_metrics.parquet` — derived band-power metrics (~4 Hz)
  - New crate deps: `parquet`, `arrow-array`, `arrow-schema` (all v54)
  - `SessionWriter` enum wraps both `CsvState` and `ParquetState` with identical API
  - Tauri commands: `get_storage_format`, `set_storage_format`
  - Setting persisted in `settings.json` as `storage_format: "csv" | "parquet"`

- **Split/sharded GGUF support**: The LLM catalog and downloader now support multi-part (split) GGUF models. Added `shard_files` field to `LlmModelEntry` listing all shard filenames in order. The new `download_model()` function downloads shards sequentially with overall progress tracking, pause/resume per-shard, and cancellation between shards. Delete properly removes all shard files. The frontend shows shard count on download buttons and current shard progress during download.

- **MiniMax M2.5 full catalog**: Added 11 quant variants of MiniMax M2.5 (456B MoE, 46B active) to the LLM catalog via `unsloth/MiniMax-M2.5-GGUF` — from TQ1_0 (52 GB single file) through Q8_0 (226 GB, 6 shards). The Q4_K_M quant is marked as recommended.

- **Embedding pipeline resamples non-256 Hz devices**: The ZUNA model expects 1280 samples (5 s × 256 Hz). Non-256 Hz devices now accumulate 5 seconds at their native rate and linearly resample to 1280 samples before encoding. Previously, MW75 (500 Hz) fed 2.56 s of data and Emotiv (128 Hz) fed 10 s, producing wrong-duration epochs with mismatched frequency content.
- **EEG chart dynamically sized for device sample rate**: `EegChart` now accepts a `sampleRate` prop and sizes its waveform ring buffer and spectrogram columns to always show ≈15 seconds of history regardless of device. Added `bufSizeForRate()` and `specColsForRate()` helpers. Previously the buffer was hardcoded to 3840 samples (15 s at 256 Hz only).

### Bugfixes

- **Fix missing `storage_format` in `UserSettings::default()`**: Added the missing field initializer in the `Default` impl for `UserSettings` in `skill-settings`.

- **Fix duplicate `MUSE_SAMPLE_RATE` import**: Removed redundant import in `eeg_embeddings/mod.rs` and added missing `CHANNEL_NAMES` import.

- **Fix renamed IDUN field**: Updated `use_60hz` → `mains_freq_60hz` in `session_connect.rs` to match upstream `idun` crate API change.

- **MoE detection for hardware fit**: The hardware-fit analyzer now detects MoE models from the `"moe"` tag in addition to inferring from family name patterns, improving fit predictions for MiniMax M2.5 and similar models.

- **CI: add missing libpipewire-0.3-dev package**: `cargo check` on Linux CI failed because the `xcap` crate transitively depends on `pipewire-sys` / `libspa-sys`, which require the PipeWire development headers. Added `libpipewire-0.3-dev` to the apt package lists in `ci.yml` and `release-linux.yml` and bumped the cache version keys to force re-caching.

- **Fix clippy warnings**: Removed unused `std::io::Cursor` import in `skill-screenshots`, changed doc comment to plain comment in `session_runner.rs`, replaced `map_or(true, …)` with `is_none_or(…)` in LLM download/server commands, and used `matches!` macro in `session_connect.rs`.

- **Session history only loaded `muse_` files**: All session-file lookups across `skill-history`, `skill-commands`, and `settings_cmds` now accept both `exg_` and `muse_` prefixes. Previously recordings from non-Muse devices were invisible.
- **Orphaned CSV sessions hardcoded 256 Hz sample rate**: When a JSON sidecar was missing, `sample_rate_hz` was set to `Some(256)`. Now set to `None` (unknown) since the actual rate cannot be determined without metadata.
- **Emotiv electrode count in ElectrodeGuide**: Updated `EMOTIV_EPOC_LABELS` from 12 to all 14 electrodes, and tab count from "12" to "14".
- **Non-Muse electrode quality strip said "Muse signal"**: Changed label to generic "Signal".

- **Device routing missed "neurable" and "ige" prefixes**: `detect_device_kind` in `lifecycle.rs` had its own copy of device-name matching that was out of sync with `DeviceKind::from_name`. A Neurable headphone would be routed to the Muse connect path, and an IGE-prefixed IDUN Guardian would also fall through to Muse. Refactored to delegate to the canonical `DeviceKind::from_name` — single source of truth.
- **Emotiv auto-detects actual channel count**: `EmotivAdapter` now detects the real channel count from the first EEG packet. Previously it always assumed EPOC (14 channels); connecting an Insight (5-ch) or MN8 (2-ch) would produce misaligned EEG frames with wrong channel counts.

- **Emotiv TS channel count/names**: Fixed `EMOTIV_CAPS` in `device.ts` — `channelCount` was 12 instead of 14, and electrode names were missing `"F8"` and `"AF4"` (EPOC X/EPOC+ have 14 channels).
- **MW75 Rust detection missing "neurable"**: Added `n.contains("neurable")` to `DeviceKind::from_name` so devices advertising as "Neurable-XYZ" are correctly identified as MW75, matching the TypeScript detection logic.
- **Hermes TS electrode names**: Replaced generic `["Ch1",...,"Ch8"]` with proper 10-20 names `["Fp1","Fp2","AF3","AF4","F3","F4","FC1","FC2"]` to match the Rust constants in `skill-constants`.
- **IDUN adapter META capability**: Added `DeviceCaps::META` to `IdunAdapter` caps, since the adapter emits `DeviceEvent::Meta` for `GuardianEvent::DeviceInfo` but was not declaring the capability.

- **FAA uses name-based electrode lookup**: Frontal Alpha Asymmetry in `eeg_bands.rs` now resolves left/right frontal electrodes by 10-20 name instead of hardcoded indices [1]/[2]. Previously, non-Muse devices computed FAA from wrong electrodes (e.g. Emotiv used F7/F3 — both left hemisphere).
- **Cognitive load uses name-based electrode lookup**: `compute_cognitive_load` now finds frontal (theta) and parietal (alpha) electrodes by 10-20 name prefix instead of assuming Muse's 4-channel index layout. Falls back to index-based split for generic labels.
- **Laterality index uses name-based hemisphere detection**: `laterality_index_fn` now determines left/right hemisphere from 10-20 naming convention (odd=left, even=right) instead of hardcoded indices [0..1] vs [2..3]. Previously, MW75 computed laterality from 4 left-hemisphere channels only.
- **IDUN battery not shown in UI**: Added `isIdun` to the `hasBattery` derived flag in `+page.svelte` so the battery indicator renders for IDUN Guardian (which reports battery via BLE).
- **Ganglion showed Muse electrode labels**: Added `GANGLION_CH`/`GANGLION_COLOR` constants and wired them into the dashboard channel-label selector. Previously, connecting a Ganglion displayed Muse names (TP9/AF7/AF8/TP10) instead of generic Ch1–Ch4.
- **Hermes channel labels in constants.ts**: Updated `HERMES_CH` from generic `Ch1–Ch8` to proper 10-20 names matching Rust `HERMES_CHANNEL_NAMES`.
- **Emotiv constants.ts channel count/names**: Updated `EMOTIV_EEG_CHANNELS` from 12 to 14, added missing `F8`/`AF4` to `EMOTIV_CH` and corresponding colours.
- **Misleading Ganglion sample-rate comment**: Corrected doc comment on `MUSE_SAMPLE_RATE` — Ganglion uses 200 Hz, not 256 Hz.

- **Non-Muse devices had no device_id in status**: `on_connected` now stores `info.id` in `status.device_id` when it hasn't been set yet. Previously only Muse's connect path set this field, leaving it `None` for MW75, Hermes, Emotiv, IDUN, and Ganglion — breaking reconnection targeting and paired device tracking.
- **Device identity fields not populated from DeviceInfo**: `on_connected` now copies `serial_number`, `firmware_version`, `hardware_version`, `bootloader_version`, `mac_address`, and `headset_preset` from the adapter's `DeviceInfo` into `DeviceStatus`. Previously these were only populated from Muse Control JSON meta events.
- **Meta handler only parsed Muse short keys**: `process_meta` now accepts both Muse-style short keys (`sn`, `ma`, `fw`, `hw`, `bl`, `tp`) and long-form keys (`serial_number`, `mac_address`, `firmware_version`, `hardware_version`, `bootloader_version`, `headset_preset`). IDUN Guardian's `DeviceInfo` meta event (which uses long keys) was previously ignored.

- **Rust `DeviceKind` enum missing Ganglion, MW75, Hermes**: Added `Ganglion`, `Mw75`, and `Hermes` variants with correct capabilities (channel count, sample rate, IMU flags). Previously Ganglion was lumped into `OpenBci` and MW75/Hermes had no representation.
- **`DeviceKind::from_name` missing prefixes**: Added `simblee`, `mn8`, and `guardian` prefix detection; split Ganglion from OpenBCI; added MW75 (substring) and Hermes detection.
- **Frontend `deviceCapabilities()` incomplete**: Added `GANGLION_CAPS` (4ch/200Hz), `MW75_CAPS` (12ch/500Hz), and `HERMES_CAPS` (8ch/250Hz) with correct electrode names. Ganglion was previously detected as OpenBCI (8ch/250Hz).
- **Frontend Ganglion detection wrong**: `"simblee"` prefix now correctly returns Ganglion caps instead of falling through to unknown.

- **BandAnalyzer hardcoded 256 Hz sample rate**: Added `BandAnalyzer::new_with_rate(sample_rate)` and wired it through `SessionDsp`. Previously, PSD bin-frequency mapping and PAC computation used `MUSE_SAMPLE_RATE` (256 Hz) for all devices. On MW75 (500 Hz) this mapped bins at nearly 2x the correct frequency, placing "beta 13-30 Hz" at ~25-59 Hz. On Emotiv (128 Hz) the "beta band" actually integrated ~6.5-15 Hz (theta+alpha). All band powers, ratios (TAR, BAR, DTR), and derived scores (meditation, drowsiness, headache, consciousness indices) were wrong for non-Muse devices.
- **ArtifactDetector hardcoded 256 Hz and Muse electrode indices**: Rewrote blink detection to accept a sample rate and resolve frontal electrodes by 10-20 name via `ArtifactDetector::with_channels(sample_rate, channel_names)`. Previously, blink refractory timing was wrong for non-256 Hz devices, and only Muse's AF7/AF8 (indices 1/2) were checked — other devices detected blinks on wrong channels or not at all.
- **Session runner now sets device sample rate in filter config**: Before creating the DSP pipeline, the session runner writes the device's actual sample rate into `FilterConfig.sample_rate`, ensuring the `EegFilter` frequency-domain operations (low-pass, high-pass, notch) use correct bin frequencies for all devices.

- **Embedding worker used Muse channel names and 256 Hz for all devices**: `load_from_named_tensor` in the ZUNA embedding worker was called with hardcoded `CHANNEL_NAMES` (TP9/AF7/AF8/TP10) and `MUSE_SAMPLE_RATE` (256 Hz) regardless of the connected device. Now receives actual channel names and sample rate via `EpochMsg`, set from the device descriptor at session start.
- **Embedding overlap computation used 256 Hz**: `EegAccumulator::set_overlap_secs` converted seconds to samples using `MUSE_SAMPLE_RATE`. Now uses the device's actual sample rate.
- **Scanning message showed "Looking for Muse" for MW75, Hermes, and IDUN**: Added device-specific scanning messages using the `connectingTo` i18n key for non-Muse/non-Ganglion/non-Emotiv devices.

- **Emotiv dashboard showed 14 channel labels but only 12 had data**: The DSP pipeline caps at `EEG_CHANNELS` (12), so the last two EPOC electrodes (F8, AF4) were never forwarded to the frontend. Aligned `EMOTIV_CH`, `EMOTIV_COLOR`, `EMOTIV_CAPS`, ElectrodeGuide, and ElectrodePlacement to show 12 channels matching the pipeline output. Prevents undefined values in the signal quality grid and EEG expanded view.

- **Emotiv adapter uses DataLabels for channel detection**: When the Cortex API sends `DataLabels` for the "eeg" stream, the adapter now updates its channel count and names to match the actual headset (EPOC 14-ch, Insight 5-ch, MN8 2-ch, Flex 32-ch). Previously only the first-packet sample-count fallback was used.
- **IDUN Guardian notch filter now respects user setting**: `connect_idun` now reads the user's notch filter preference (50 Hz / 60 Hz) from the app settings and passes it to `GuardianClientConfig::use_60hz`. Previously the on-device notch always defaulted to 60 Hz, producing mains artifacts for users in 50 Hz countries (Europe, Asia, Africa, most of South America).

- **Emotiv & IDUN devices never connected**: `detect_device_kind` in the session lifecycle had no arms for Emotiv or IDUN device names, causing them to fall through to the Muse connect path. Added prefix matching for Emotiv (`emotiv`, `epoc`, `insight`, `flex`, `mn8`) and IDUN (`idun`, `guardian`) devices, and wired the `"emotiv"` / `"idun"` kinds to their respective `connect_emotiv` / `connect_idun` factory functions.

- **Emotiv & IDUN dashboard UI incomplete**: Added `isEmotiv` / `isIdun` capability flags to the dashboard, device images, alt text, battery visibility for Emotiv, and device-specific scanning message for Emotiv Cortex API connections.
- **ElectrodeGuide missing Emotiv & IDUN tabs**: Added Emotiv EPOC (14-ch) and IDUN Guardian (1-ch) tabs with correct electrode positions to the 3D electrode guide.
- **MN8 earbuds not detected in frontend**: `deviceCapabilities()` was missing the `mn8` prefix for Emotiv MN8 earbuds.

- **Emotiv/IDUN channel labels defaulted to Muse 4-ch**: Dashboard channel labels and colors now correctly show 14 Emotiv EPOC channels or 1 IDUN Guardian channel instead of falling back to the Muse TP9/AF7/AF8/TP10 labels.
- **ElectrodePlacement SVG missing Emotiv & IDUN**: Added 14-electrode Emotiv EPOC and single-electrode IDUN Guardian presets to the 2D electrode placement diagram with correct 10-20 positions.
- **Device info badge only shown for Ganglion**: Non-Muse device info badges (channel count, sample rate) now also appear for Emotiv, IDUN, and Hermes connections.
- **EEG expanded grid always 2-column**: The expanded EEG channel grid now adapts columns to the channel count (2 for ≤4ch, 3 for 5-8ch, 4 for >8ch) instead of being hardcoded to 2 columns except MW75.

- **Session runner re-reads pipeline channels after Emotiv auto-detection**: When Emotiv auto-detects the actual channel count (Insight 5-ch, MN8 2-ch vs assumed EPOC 14-ch), the session runner now picks up the updated `pipeline_channels` on the first EEG frame. Previously the snapshot was taken before any events arrived, so the DSP pipeline would process 14 channels even when only 5 were active.

- **MW75 reconnects to the correct device**: `connect_mw75` now uses `scan_all()` + `connect_to(device)` with `preferred_id` matching, so reconnection targets the previously-paired headphone instead of picking the first MW75 found.
- **Hermes reconnects to the correct device**: Same fix as MW75 — `connect_hermes` now uses `scan_all()` + `connect_to(device)` with `preferred_id` matching.
- **Emotiv CSV has correct channel header**: CSV creation is now deferred until the first EEG frame arrives, after Emotiv's channel auto-detection (DataLabels) has resolved the actual channel count. Previously the CSV was opened with 14-column EPOC headers even when an Insight (5-ch) was connected.

- **Fix screenshot capture on Linux (Wayland) and Windows**: Replaced shell-command-based screenshot capture (`xdotool`, `import`, `scrot`, `grim`, `swaymsg` on Linux; PowerShell `CopyFromScreen` on Windows) with the `xcap` crate. This provides native, dependency-free screen capture on both X11 and Wayland (via PipeWire) on Linux, and native Win32/WGC capture on Windows. Includes a dark-frame detection guard that skips all-black captures.

- **Downloaded LLM model auto-activates**: When the first model finishes downloading (or the current active model is missing/deleted), the newly downloaded model is automatically set as active. Previously only activated when `active_model` was empty — a stale reference to a deleted model prevented auto-selection.
- **"Start LLM Engine" auto-selects a model**: When the user clicks Start and no model is active (or the active model file is missing), the engine now automatically picks the first available downloaded model, activates it, and proceeds with loading. Previously it would fail with a generic error asking the user to manually click "Use".

- **Metrics CSV/Parquet header now uses actual channel names**: The metrics header was hardcoded to 4 Muse channels (TP9/AF7/AF8/TP10 × 12 bands = 48 columns). For MW75 (12-ch) or Emotiv (14-ch), the data row had more band-power columns than the header, corrupting CSV alignment. Added `build_metrics_header(channel_names)` that generates the correct per-channel band columns dynamically. Both CSV and Parquet metrics writers now use it.
- **Metrics reader detects channel count from header**: `load_metrics_csv` and `load_metrics_from_parquet` now find the "faa" column position to compute the cross-channel index offset dynamically. Previously hardcoded to column 49 (4 channels × 12 bands + 1), causing all index values to read from wrong columns for non-4-channel devices.

- **QualityMonitor window size now matches device sample rate**: Added `QualityMonitor::with_window(channels, window)` and wired it to use the device's EEG sample rate (≈1 second window). Previously the window was hardcoded to 256 samples — only 0.51 s at 500 Hz (MW75) or 2 s at 128 Hz (Emotiv), causing quality to be assessed over inconsistent time windows.
- **HeadPoseTracker IMU rate now configurable**: Added `HeadPoseTracker::with_imu_rate(hz)`. Gyro integration (`dt`), stillness EMA, gesture window, and refractory period all used the hardcoded Muse IMU rate (52 Hz). At different rates, `dt = 1/52` would produce wrong pitch/roll/yaw accumulation and incorrect stillness scores.

- **Fix screenshot service mock data**: Fixed GPU stats mock to use 0–1 fractions instead of raw integers (was showing 4000% instead of 30%). Fixed `get_csv_metrics` timeseries to use correct abbreviated `EpochRow` field names (`med`, `cog`, `drow`, `sc`, `mu`, `ha`, `hm`, `hc`, etc.) so session charts render properly. Fixed `get_sleep_stages` mock to return `{ epochs: [], summary: null }` instead of `null` to prevent crashes in the compare page.

- **Fix dashboard light screenshot empty**: Added a warm-up step before taking the first screenshot so Vite finishes compiling, and increased wait time for the dashboard page to ensure Svelte fully bootstraps before capture. Dashboard is now also captured as full-page.

- **Fix search EEG screenshot empty**: The `stream_search_embeddings` mock now properly sends streaming results through the Tauri Channel by extracting the Channel's callback ID and delivering `started`, `result`, and `done` messages with realistic neighbor data including labels and metrics.

- **Fix search images broken thumbnails**: Broken `<img>` elements (pointing to non-existent local screenshot server) are now replaced with coloured placeholder SVGs that mimic app windows (VS Code, Firefox, Terminal) after search results render. Removed duplicate search-images handler that was re-triggering search and overwriting the placeholder replacement.

- **Session metadata hardcoded Muse values**: `write_session_meta` wrote `"sample_rate_hz": 256`, `"channels": ["TP9","AF7","AF8","TP10"]`, and `"channel_count": 4` for all devices. Now uses actual device values from `DeviceStatus` (set at session start). Recordings from MW75, Emotiv, Hermes, IDUN, and Ganglion had incorrect metadata in the JSON sidecar file.
- **Ganglion connected at wrong sample rate**: `connect_ganglion` passed `EEG_SAMPLE_RATE` (256 Hz) to `OpenBciAdapter::make_descriptor`, but Ganglion runs at 200 Hz. Added `GANGLION_SAMPLE_RATE` (200 Hz) constant and use it instead. This caused the entire DSP pipeline (filter, band analyzer, artifact detector) to use 256 Hz for a 200 Hz device.
- **Missing constants in prelude**: Added `GANGLION_SAMPLE_RATE`, `GANGLION_CHANNEL_NAMES`, `HERMES_*`, and `MW75_*` constants to the `skill-constants` prelude for easier access across crates.

- **⚠️ BREAKING: `muse-status` event renamed to `status`**: The Tauri IPC event and WebSocket broadcast event `muse-status` has been renamed to `status` to reflect its device-agnostic nature. All frontend listeners, the WS server, and documentation have been updated. **External WS clients that subscribe to `muse-status` must update to `status`.**

### Refactor

- **Unified `exg_` session file convention**: New recordings use `exg_<timestamp>.csv` and `exg_<timestamp>.json` for all devices, replacing the Muse-only `muse_` prefix. Full backward compatibility: the history loader, session analysis, embedding search, and settings commands all accept both `exg_` and legacy `muse_` files.

- **`detect_device_kind` delegates to `DeviceKind::from_name`**: Eliminated duplicated device-name matching constants in `lifecycle.rs`. All device-name detection now flows through `skill-data::device::DeviceKind::from_name`.

- **DeviceKind type updated**: Added missing `"ganglion"`, `"mw75"`, and `"hermes"` variants to the TypeScript `DeviceKind` union so it matches all backend device kinds.
- **Stale JSDoc**: Updated `device_kind` field comment in `types.ts` to reference `DeviceKind` instead of listing an incomplete set.

- **Device config: Rust as single source of truth**: Eliminated duplicate device definitions between Rust and Svelte. `crates/skill-data/src/device.rs` is now the canonical source for device families, capabilities (channel count, PPG, IMU, central electrodes, full montage, sample rate, electrode names), and the supported-devices catalog (companies, models, images, instruction keys). The Svelte frontend receives capability flags via `DeviceStatus` fields (`has_ppg`, `has_imu`, `has_central_electrodes`, `has_full_montage`) and fetches the supported-devices catalog via the new `get_supported_companies` Tauri command. `src/lib/device.ts` is now a thin type-only wrapper; `src/lib/supported-devices.ts` loads data from Rust at startup. The old `detect_device_kind` in `lifecycle.rs` now delegates to `DeviceKind::as_str()`.

- **Split i18n into namespace files**: Replaced monolithic ~3000-line locale files (`en.ts`, `de.ts`, `fr.ts`, `he.ts`, `uk.ts`) with 15 namespace-based files per locale under `src/lib/i18n/<locale>/` (`common`, `dashboard`, `settings`, `search`, `calibration`, `history`, `hooks`, `llm`, `onboarding`, `screenshots`, `tts`, `perm`, `help`, `help-ref`, `ui`). Each locale folder has a barrel `index.ts` that merges all namespaces.
- **Added `TranslationKey` type safety**: Generated `keys.ts` with a union type of all 2731 valid translation keys. The `t()` function now accepts `TranslationKey` for compile-time checking on static keys, with a `string` overload for dynamic/computed keys.
- **Extracted shared `i18n-utils.ts`**: Moved duplicated `extractKeys()` logic from `sync-i18n.ts` and `audit-i18n.ts` into a shared `src/lib/i18n/i18n-utils.ts` module.
- **Updated i18n tests**: Test suite now validates per-namespace file consistency (74 tests, all passing).
- **Updated scripts**: `sync-i18n.ts` and `audit-i18n.ts` now operate on the new directory structure and use shared utilities.

### Build

- **Decouple skill-settings from skill-screenshots**: Moved `ScreenshotConfig` struct and its pure helper methods from `skill-screenshots` into `skill-settings`, breaking the transitive `skill-settings` → `skill-screenshots` → `xcap` → `pipewire` dependency chain. Crates like `skill-router`, `skill-history`, and `skill-settings` no longer require `libpipewire-0.3` to compile. The `fastembed_model_enum()` helper stays in `skill-screenshots` as a standalone function since it depends on the `fastembed` crate. `skill-screenshots` now imports `ScreenshotConfig` from `skill-settings` instead of owning it.

- **Feature-gate `xcap` in skill-screenshots**: Added a `capture` feature (default on) that gates the `xcap` dependency. The top-level binary sets `default-features = false` so `cargo clippy` works without `libpipewire-0.3` on dev machines. Screen capture gracefully returns `None` when the feature is disabled. CI and release builds work unchanged since pipewire is installed there.

### i18n

- **Emotiv scanning message**: Added `dashboard.connectingEmotiv` key in all five languages (en, de, fr, uk, he).

- **Replaced `\u` escape sequences with UTF-8 characters**: Converted 11 `\uXXXX` escapes in de, fr, he, uk i18n files to native UTF-8 per the project encoding rule.

### Dependencies

- **Added `xcap` 0.9 for Linux and Windows**: Cross-platform screen capture library replacing external CLI tool dependencies.

## [0.0.42] — 2026-03-18

### Features

- **Per-device API credentials**: added persistent device API configuration with a dedicated Emotiv Cortex client ID/secret editor in both Devices and Settings tabs.
- **Emotiv connection integration**: Emotiv connection now reads credentials from settings first, with environment variables as fallback.

- **Emotiv device support**: Added `EmotivAdapter` for Emotiv EPOC X, EPOC+, Insight, and Flex headsets via the Cortex WebSocket API (JSON-RPC 2.0). Streams EEG (up to 14 ch @ 128 Hz), motion/IMU, and battery data through the unified session runner.
- **IDUN Guardian device support**: Added `IdunAdapter` for the IDUN Guardian in-ear EEG earbud over BLE. Streams single-channel bipolar EEG (1 ch @ 250 Hz), 6-DOF IMU (accelerometer + gyroscope), and battery data.
- **Device constants**: Added hardware constants for Emotiv (EPOC 14-ch, Insight 5-ch, 128 Hz sample rate, channel labels) and IDUN Guardian (1-ch, 250 Hz, channel label) to `skill-constants`.
- **DeviceKind::Idun**: Extended `DeviceKind` enum in `skill-data` with the `Idun` variant, capability flags, and name-based detection (`idun`, `ige`, `guardian` prefixes).
- **TypeScript device layer**: Added `"idun"` to the `DeviceKind` union and `IDUN_CAPS` capability table in `device.ts` for UI-side device detection.

- **IDUN per-device API token**: extended device API configuration with an `IDUN` token field and UI editors in Devices and Settings tabs.
- **IDUN connector integration**: when configured, the IDUN token is applied to `IDUN_API_TOKEN` before connection, so cloud-decoding paths can consume it.

- **RE-AK device image**: `deviceImage()` in both DevicesTab and SettingsTab now maps Nucleus-Hermès BLE names (`hermes`, `nucleus`, `re-ak`, `reak`) to the correct device image.

### Performance

- **Improve Windows CI build caching**: stabilize Rust cache reuse with a shared `rust-cache` key and enable persisted GitHub Actions `sccache` backend to increase cache-hit rates and reduce repeated full recompiles.
- **Add compile cache diagnostics**: print `sccache --show-stats` before and after Rust compile in the Windows release workflow to make cache effectiveness visible in job logs.

- **Faster Windows builds with lld-link**: Auto-detect LLVM's `lld-link` linker on Windows (both CI and local dev via `tauri-build.js`), replacing the slower MSVC `link.exe`. Combined with the previously split compile/package CI steps, this should significantly reduce Windows release CI time.

### Bugfixes

- **Chart refs reactivity**: Declared `chartEl` and `bandChartEl` with `$state(...)` in `+page.svelte` to fix `non_reactive_update` warnings.

- **Device API save error handling**: `saveEmotivApi` and `saveIdunApi` now wrap the Tauri invoke in `try/catch`; failed saves surface an inline error message instead of silently flashing "Saved".
- **IDUN token env var safety**: replaced `std::env::set_var("IDUN_API_TOKEN", …)` (process-wide mutation, unsafe in multi-threaded Rust) with passing the token via `GuardianClientConfig::api_token`.

- **Restore ExgTab compatibility path**: added `src/lib/ExgTab.svelte` as a compatibility shim to prevent missing-module 404 requests.

- **Fix `app_log` ambiguity in lifecycle.rs**: Removed erroneous `app_log` item import from `crate` that conflicted with the `app_log!` macro defined in `lib.rs`. Replaced unused `Emitter` import with `Manager` to provide the `.state()` method needed by the macro.

- **Fix `app_log!` ambiguity in lifecycle module**: remove incorrect `app_log` item import from `src-tauri/src/lifecycle.rs` and clean unused Tauri imports in lifecycle/scanner modules so the Tauri crate compiles cleanly.

- **Enable GPU acceleration in Linux and macOS release builds**: Added missing `--features llm-vulkan` to the Linux CI cargo build and `--features llm-metal` to the macOS CI cargo build, ensuring GPU-accelerated LLM inference is included in release binaries (matching what `tauri-build.js` injects for local builds).

- **Windows CI: enable llm-vulkan feature**: Add missing `--features llm-vulkan` to the Windows release cargo build command, ensuring Vulkan GPU offloading for LLM inference is included in release builds (matching what `tauri-build.js` injects locally).

- **Fix `llama-cpp-4` version mismatch**: Non-macOS platforms used `0.2.10` while macOS used `0.2.12`; aligned to `0.2.12` everywhere.
- **Fix `package.json` dead script**: Removed duplicate `taur:build:win:nsis` key (typo missing `i`).
- **Replace silent `catch {}` in chat page**: All 18 empty catch blocks in `chat/+page.svelte` now log warnings/errors to the console instead of silently swallowing failures.

### Refactor

- **Extract lifecycle and quit modules from lib.rs**: Moved session lifecycle (start/cancel/disconnect/reconnect backoff) into `lifecycle.rs` and quit-confirmation dialogs into `quit.rs`, reducing `lib.rs` from 1,778 to 1,495 lines.
- **Add `DeviceStatus::reset_disconnected()` method**: Replaces 15+ manual field resets in `go_disconnected` with a single method call, preventing missed fields when new status fields are added.
- **Consolidate mutex lock acquisitions in `setup_app`**: Merged 4 separate lock/unlock cycles for LLM autostart, embedding model, model status, and HF repo into a single critical section.
- **Extract device-kind detection into constants**: Replaced inline string matching (`starts_with("ganglion")`, `contains("mw75")`) with named constants and a `detect_device_kind()` function with unit tests.

- **Remove duplicated sections from Settings tab**: Removed Muse Devices, OpenBCI, Signal Processing, EEG Embedding, and GPU/Memory sections from SettingsTab since they now live in the dedicated Devices and EXG tabs. The Settings tab now only contains Activity Tracking, Logging, Data Directory, and WebSocket Server settings.

- **Standardize device image naming**: renamed device assets to consistent vendor-prefixed names (`muse-*`, `openbci-*`, `emotiv-*`, `idun-*`) and updated all UI references accordingly.

- **Shared supported devices config**: extracted Supported Devices company/device/instruction definitions into a single shared module used by both Devices and Settings tabs.

### Build

- **Add `rustfmt.toml` and `clippy.toml`**: Added formatting and lint configuration for consistent Rust code style.

### UI

- **Devices settings tab**: Dedicated "Devices" tab (second in Settings sidebar) showing paired and discovered BCI devices with pair/forget/set-default actions.
- **EXG settings tab**: New "EXG" tab (third in Settings sidebar) for OpenBCI board configuration, signal processing filters (notch, high/low-pass), EEG embedding pipeline, and GPU/memory stats.

- **Fold Device API by company**: Device API settings are now grouped into collapsible company sections (`Emotiv Cortex`, `IDUN Cloud`) in both Devices and Settings tabs.

- **No visual behavior change**: existing device thumbnails continue to render, now via standardized local file names.

- **Device images**: Use `object-cover` so device thumbnails and previews fill their containers fully.

- **Add MN8 and X-Trodes device photos**: device cards now map `MN8` to `/devices/emotiv-mn8.webp` and `X-Trodes`/`Xtrodes` names to `/devices/emotiv-x-trodes.webp`.

- **Add model-specific Emotiv device photos**: device cards now map `EPOC`/`Emotiv` to `/devices/emotiv-epoc-x.webp`, `Insight` to `/devices/emotiv-insight.webp`, and `Flex` to `/devices/emotiv-flex-saline.webp`.

- **Use provided IDUN and Emotiv photos**: device rows now show the supplied IDUN Guardian and Emotiv Epoc images in Devices and Settings tabs instead of the generic fallback icon.

- **Add IDUN token helper link**: added a direct link to the IDUN dashboard in Device API settings to make token setup easier.

- **Localize IDUN and Emotiv device images**: switched device-card image sources to bundled local files (`/devices/idun-guardian.png` and `/devices/emotiv-epoc.png`) to avoid runtime dependence on external hosts.

- **RE-AK Nucleus-Hermès asset**: replaced the temporary Hermes placeholder with the official RE-AK image and bundled it as a local static asset.
- **Device naming correction**: updated Supported Devices labeling and instructions to use the proper spelling `Nucleus-Hermès`.

- **Save per API provider section**: Device API settings now have separate Save buttons inside each folded company section (`Emotiv Cortex`, `IDUN Cloud`) in both Devices and Settings tabs.

- **Company mapping correction**: moved `MW75 Neuro` from `Muse` to a dedicated `Neurable` section in Supported Devices.
- **RE-AK support card**: added `RE-AK` with `Hermes Nucleus` to Supported Devices and included a local bundled image asset.

- **Supported devices section**: added a compact Supported Devices section in Devices and Settings tabs that lists supported companies and device images.
- **Expandable connection guidance**: clicking a company name or any device image now expands brief connection instructions and opens the related settings panel (OpenBCI, Emotiv, IDUN).

- **Supported device card alignment**: refined Supported Devices grid spacing and card sizing for a cleaner, more consistent layout.
- **Uniform image framing**: all device images now render inside fixed-size white background frames with centered `object-contain` scaling to normalize mixed aspect ratios.

- **Supported Devices search**: Add fuzzy search to filter devices and companies by name in the Supported Devices section.
- **Supported Devices layout**: Reduce spacing and heights to make the grid more compact (smaller gaps, smaller thumbnails, smaller text).

### i18n

- **Device API section fully translated**: all hardcoded strings in the Device API card (section title, provider titles, descriptions, field labels, show/hide, save/saved) are now driven by `t()` keys under `settings.deviceApi.*`.

- **Supported Devices localization**: moved Supported Devices section title, company labels, device names, and connection instructions to i18n keys and replaced hardcoded strings in both tabs.

## [0.0.41] — 2026-03-17

### Features

- **Agent Skills discovery**: Added `skill-skills` crate that discovers `SKILL.md` files from `~/.skill/skills/` (user), `<cwd>/.skill/skills/` (project), and the bundled `skills/` git submodule. Discovered skills are injected into the LLM system prompt as an `<available_skills>` XML block so the model can load specialised instructions via `read_file` on demand. Skills require a `description` in YAML frontmatter; invalid index-style `SKILL.md` files (without description) allow recursion into subdirectories. Deduplication by name (first wins) and symlink real-path. Added `skills` git submodule from `https://github.com/NeuroSkill-com/skills.git` providing 10+ bundled EEG/protocol skills.

- **CLI: screenshot search commands**: Added `search-images` and `screenshots-around` CLI commands. `search-images "query"` searches screenshots by OCR text in semantic (embedding HNSW) or substring (SQL LIKE) mode. `screenshots-around --at <utc>` finds screenshots near a given timestamp within a configurable window. Both commands support `--json`, `--full`, and `--k` flags. Also added the corresponding `search_screenshots` and `screenshots_around` WebSocket/HTTP commands to the server dispatcher.

- **Configurable SNR exit threshold for DND**: The SNR level below which focus mode is forcibly deactivated is now a user setting (`snr_exit_db`) instead of a hardcoded constant. Default changed from 5 dB to 0 dB so DND only exits when the signal is completely lost. A new preset picker in the DND settings UI lets users choose 0 / 3 / 5 / 10 / 15 dB. Translations added for EN, DE, FR, UK, HE.

- **Context breakdown inspector**: Click the context usage ring in the chat header to open a popover showing the proportional breakdown of context window usage — system prompt, EEG context, tool definitions, user messages, assistant messages, thinking/reasoning, tool results, and current completion tokens. Includes a stacked bar visualization and detailed legend with token counts and percentages. Localized in all 5 languages.

- **skill-headless: Headless / Headful modes**: Replaced the `visible: bool` flag with a `Mode` enum (`Mode::Headless` and `Mode::Headful`). Headless mode positions the window off-screen so nothing is ever shown to the user while still giving the webview real pixel dimensions. Headful mode shows the window on-screen for debugging, demos, or interactive automation. In headless mode, `SetViewport` ensures the window stays off-screen after resize.

- **skill-headless: network interception**: Added request/response interception support. `EnableInterception` monkey-patches `fetch()` and `XMLHttpRequest` to capture all HTTP traffic. Navigation events are recorded via wry's navigation handler. `SetBlockedUrls` blocks navigations matching URL substring patterns. `GetInterceptedRequests` retrieves the full network log (requests, responses, navigations) with optional clear-on-read. Includes 11 tests covering fetch GET/POST, XHR with custom headers, navigation capture, URL blocking, and log clearing.

- **HealthKit data ingestion endpoints**: Added HTTP REST and WebSocket endpoints to receive Apple HealthKit data from a companion iOS app. New endpoints: `POST /v1/health/sync` (idempotent batch upsert of sleep, workouts, heart rate, steps, mindfulness, and generic metrics), `POST /v1/health/query` (query by type and time range), `GET /v1/health/summary` (aggregate counts), `GET /v1/health/metric_types` (list stored metric types). Data is stored in `~/.skill/health.sqlite` via a new `health_store` module in `skill-data`. All endpoints are also available as WS commands (`health_sync`, `health_query`, `health_summary`, `health_metric_types`).

- **Standalone LLM logger**: Added `skill_llm::log` module with pluggable callback sink (`set_log_callback`) and `llm_log!` macro. All `eprintln!("[llm] ...")` / `eprintln!("[chat_store] ...")` calls in `engine.rs` and `chat_store.rs` now route through the unified logger. On the Tauri side, `llm::init_llm_logger()` wires LLM output through `SkillLogger` so the `llm` and `chat_store` subsystem toggles in log config control visibility. Added `llm` and `chat_store` fields to `LogConfig`, new rows in the Settings logging grid, and i18n keys for all five locales (en, de, fr, uk, he). Also fixed the TtsTab `LogConfig` interface to include the missing `hooks` field.

- **LLM auto-start on launch**: Added `autostart` field to `LlmConfig`. When enabled + a model is downloaded and selected, the LLM server starts automatically during app setup with a 500ms delay to let the UI render first. Toggle added to the LLM settings tab.
- **Atomic model switch**: New `switch_llm_model` Tauri command that atomically stops the running server, waits for full shutdown, sets the new active model, and starts the new one — eliminating the fragile 150ms sleep race in the frontend.
- **Abort feedback in chat**: The stop/abort button now shows a spinner and "Aborting…" state while the abort is in flight, and is disabled to prevent double-clicks.
- **Context window warning**: A warning banner appears above the chat input when context usage exceeds 85% (amber) or 95% (red), showing the current usage percentage.
- **Per-session generation params**: Temperature, max tokens, top-k, top-p, and thinking level are now saved per chat session (new `params` column in `chat_sessions` table). Params auto-save on change (debounced 500ms) and restore when switching sessions.
- **Regenerate button**: Hover over the last assistant message to see a "Regenerate" button that removes the response and re-sends the last user message with current params.
- **Edit & resend on user messages**: Hover over any user message to see "Edit & resend" which populates the input with that message's text and removes all subsequent messages.
- **Live tok/s indicator**: During streaming, a live tokens-per-second counter is shown below the assistant message. After completion, the final tok/s is included in the timing line alongside TTFT and token counts.
- **Open LLM settings from chat**: The empty chat state (when server is stopped) now includes an "Open LLM settings" button alongside "Start server".
- **Reduced settings panel height**: Chat settings and tools panels reduced from 50vh to 40vh max to leave more room for the message list on small screens.

- **Best-result scoring for web search**: Rendered page content is now scored by text quality (word count, presence of numbers/data indicators like temperatures and percentages, uniqueness of words) with penalties for CSS/JS garbage. Only the best 1-2 results are included in the compact output instead of all 5, giving the LLM focused, high-quality content to summarize.

- **Configurable web search provider**: the `web_search` tool now supports three backends — **DuckDuckGo** (default, no API key), **Brave Search** (free tier: 2,000 queries/month with API key), and **SearXNG** (self-hosted instance URL). A new `WebSearchProvider` config struct holds the backend choice, Brave API key, and SearXNG URL. Each backend falls back to DuckDuckGo HTML scraping if it fails.
- **Search provider UI**: added a backend selector (DuckDuckGo / Brave / SearXNG) to the Tools settings tab, with conditional API key and URL inputs.

- **Skill API tool for LLM chat**: Added a built-in `skill` tool that gives the LLM direct access to the full NeuroSkill WebSocket API. The LLM can now query device status, list sessions, create labels, search EEG embeddings, manage hooks, control DND, run calibrations, and more — all without requiring the user to copy-paste data. The tool connects to the local HTTP API server and supports all commands from the CLI (status, sessions, session_metrics, label, search_labels, interactive_search, search, compare, sleep, say, notify, calibrate, timer, hooks, dnd, calibrations CRUD, umap, and read-only LLM management). Dangerous LLM self-management commands (start/stop/delete/select) are blocked for safety. Enabled by default via the `skill_api` toggle in Settings → LLM → Tools.

- **skill-headless crate**: New headless browser engine providing a CDP-like command API over wry/tao. Supports navigation, JS evaluation, DOM queries, CSS/script injection, cookie and storage management, viewport emulation, cache clearing, screenshots (canvas-based DOM walker), and element wait primitives — all driven from any thread via a channel-based command/response protocol over an off-screen system webview. Includes 29 advanced tests covering HTML rendering (background, font, flexbox, grid, positioning), PNG screenshots at multiple sizes, viewport resizing (320x240 to 1920x1080), WebGL/WebGL2 context creation and shader compilation, and custom user-agent injection.

- **Sleep schedule settings**: Added a new "Sleep" section in Settings with configurable bedtime and wake-up time. Includes five presets (Default 23:00–07:00, Early Bird 21:30–05:30, Night Owl 01:00–09:00, Short Sleeper 00:00–06:00, Long Sleeper 22:00–08:00), a 24-hour clock visualization, and duration summary. Sleep window is persisted and can be used for session classification and sleep staging analysis.

- **Tool-call logging toggle in Settings**: Added a `tools` flag to the logging configuration so users can enable/disable tool-call logging from Settings. The toggle controls the `skill-tools::log` subsystem which traces tool invocations, safety approvals, completion times, and errors. Wired into the central `SkillLogger` with `init_tool_logger` callback and `set_tool_logging` runtime toggle. Added i18n strings for en, de, fr, uk, he.

- **Separate tool-call logger**: Added a standalone pluggable logger (`skill-tools::log`) for tool-call tracing, following the same pattern as `skill-llm::log`. Logs tool invocations (name + args), completion times, safety approval events, and errors. Use `set_log_callback` to route output to the app logger and `set_log_enabled` to toggle at runtime. The `tool_log!` macro short-circuits formatting when logging is disabled.

- **Configurable tool context compression**: Added a new "Context compression" setting (Off / Normal / Aggressive) in Settings → LLM → Tools that controls how tool results are compressed before being injected into the conversation context. Normal mode caps web search results to 5, truncates long URLs, and compresses old tool results. Aggressive mode uses tighter limits for small context windows. Custom overrides for max search results and max result characters are available when compression is enabled.

- **Configurable tool hop limits**: Added UI controls for `max_rounds` (how many think→tool→think cycles per message) and `max_calls_per_round` (how many tools per cycle) in both the LLM Settings tools card and the Chat window tools panel. Preset buttons (1/3/5/10 for rounds, 1/2/4/8 for calls) with active highlight. i18n keys added for all 5 locales.

- **Tools master toggle**: Added an `enabled` master switch to the LLM tool configuration that disables all tools at once. When turned off, no built-in tools are available to the LLM regardless of individual tool toggles. The toggle appears in both the Settings LLM tab and the Chat tool panel, with i18n support for all languages (en, de, fr, uk, he).

- **Standalone TTS logger**: Added `skill_tts::log` module with pluggable callback sink (`set_log_callback`) and `tts_log!` macro. All `eprintln!("[tts] ...")` / `eprintln!("[neutts] ...")` calls in `kitten.rs`, `neutts.rs`, and `lib.rs` now route through the unified logger. On the Tauri side, `tts::init_tts_logger()` wires TTS output through `SkillLogger` so the `tts` subsystem toggle in log config controls TTS log visibility. Logging is enabled by default and can be toggled at runtime via `set_log_enabled`.

- **WebSocket & HTTP chat persistence**: Chats initiated via the WebSocket `llm_chat` command and `POST /llm/chat` endpoint are now persisted to the same SQLite chat store used by the Chat window. User and assistant messages (including tool calls) are saved to `chat_history.sqlite`, making them visible in the Chat window sidebar and recoverable across restarts. A `session_id` field is returned in both the WS `session` frame and the `done`/response payload, and callers can pass `session_id` in subsequent requests to continue an existing conversation. New sessions are auto-titled from the first user message.

- **Headless browser rendering in web_fetch tool**: Added `render` parameter to the `web_fetch` LLM tool. When `render=true`, pages are loaded in a headless browser (via `skill-headless`) that executes JavaScript, enabling content extraction from SPAs and dynamically rendered pages. Supports optional `wait_ms`, `selector` (CSS selector to wait for), and `eval_js` (custom JS to evaluate) parameters.

- **Headless browser rendering in web_search tool**: Added `render` and `render_count` parameters to the `web_search` LLM tool. When `render=true`, the top N search result URLs are visited in a headless browser and their rendered text content is included in the results under a `rendered_text` field, giving the LLM access to full page content including JS-rendered material.

### Performance

- **Parallel URL fetching for web search**: When `render=true`, all search result URLs are now fetched concurrently using scoped threads instead of sequentially. Total fetch time equals the slowest single URL rather than the sum of all. This typically reduces render=true latency from 10-15s to 3-5s for 3 URLs.

- **Drop full Linux portable build from CI**: Removed the `linux-portable-package` job from `ci.yml` so Linux CI only runs `cargo check` + `clippy` (matching Windows). The full release build and packaging are already covered by the dedicated `release-linux.yml` workflow. This dramatically reduces CI wall-clock time on every push and PR.

- **Screenshot duplicate detection**: when a new screenshot is identical to the previous one (same resized-PNG hash), the embed thread now copies the vision embedding, OCR text, and OCR text embedding from the previous row instead of re-running the vision encoder, OCR engine, and text embedder. This eliminates redundant GPU/CPU inference when the screen content hasn't changed (e.g. idle desktop, paused video).

### Bugfixes

- Fixed trailing garbage bytes in `crates/skill-tools/src/types.rs` that could cause a compilation failure.

- **Remove broken public SearXNG instance scraping**: public SearXNG instances universally block automated API access with 429 rate limits or anti-bot captchas. Removed the background instance list fetcher and random instance selection. SearXNG now requires a user-provided self-hosted instance URL.

- **Add tests to previously untested crates**: Added 35 new unit tests across 4 crates that had zero test coverage:
  - `skill-exg` (17 tests): cosine distance (identical, opposite, orthogonal, edge cases), fuzzy matching (exact, case-insensitive, substring, typo, empty), Levenshtein distance.
  - `skill-commands` (13 tests): DOT escaping, SVG escaping, text truncation, turbo colormap, graph generation.
  - `skill-eeg/band_metrics` (10 tests): spectral edge frequency, spectral centroid, Hjorth parameters, permutation entropy, sample entropy, DFA exponent, Higuchi fractal dimension.
  - `skill-history/cache` (5 tests): timeseries downsampling (noop, exact count, endpoint preservation, min-2), sleep stage analysis.

- **Web search no longer stalls after returning URLs**: Improved `web_search` tool description to instruct the LLM to use `render=true` for factual/current-data queries (weather, prices, scores, news). When `render=false`, the tool result now includes a follow-up hint telling the model to fetch page content. Added a weather example to the system prompt so the model learns the correct pattern.
- **Context window no longer fills up during multi-step tool chains**: The orchestration loop now condenses prior-round tool results to one-line summaries after each round (e.g. `[location: Boston, MA, US (America/New_York)]`, `[web_search: 5 results for "weather Boston"]`). The model already consumed those results and chose its next action, so the full content is no longer needed. This frees ~200-500 tokens per prior round, allowing 3-4 step chains (location → search → fetch → answer) to complete even on 4 K context models. Additionally, `web_search` returns compact text instead of verbose JSON, `web_fetch` is capped to configured limits, and headless-rendered page text is reduced from 4 K to 2 K chars per URL.

- **Accent color consistency**: All UI elements now honor the Appearance accent setting. Replaced hardcoded `oklch(0.58 0.24 293)` violet values in the markdown renderer CSS with `var(--color-violet-*)` tokens, converted every `purple-*` Tailwind class to the remapped `violet-*` family, and switched inline-style hex accent colors (`#8b5cf6`, `#a855f7`, `#c084fc`) to CSS custom properties across dashboard gauges, focus timer, compare page, and EEG indices.

- **Rotating browser User-Agents**: replaced bot-like User-Agent strings with a pool of 10 realistic browser UAs (Chrome, Firefox, Safari, Edge on Windows/macOS/Linux) rotated on each request to reduce fingerprinting.
- **Fix DuckDuckGo HTML search**: mimic real form submission by adding `Origin` header, correct `Referer`, and the `b=` submit-button field. Fixed HTML parser to split on `class="result results_links"` (the actual outer wrapper) instead of `class="result__body"` which is now a multi-class attribute and no longer matches.
- **Extracted `parse_ddg_html`**: separated HTML parsing from HTTP fetching for testability. Added offline unit tests for result parsing and DDG redirect URL unwrapping.

- **Fix macOS headless build**: Removed non-existent `EventLoopBuilderExtMacOS` import and `with_any_thread` call — tao 0.34 does not gate event loop thread affinity on macOS.
- **Fix skill-screenshots build**: Added missing `GenericImageView` import, made `CapturedImage` fields `pub(crate)`, cfg-gated `Path` import, removed unused `CapturedImage` re-import in capture.rs.
- **Fix skill-llm warnings**: Removed unused imports in engine.rs and handlers.rs (`SystemTime`, `UNIX_EPOCH`, axum types, `GenParams`, `unix_ts_ms`, `HeaderMap`, `Sse`).
- **Fix SleepTab Svelte error**: Wrapped `{@const}` tags in `{#if true}` blocks — `{@const}` must be an immediate child of a block tag, not a raw element like `<svg>`.

- **Headless browser no longer crashes the app on macOS**: On macOS, tao requires the event loop on the main thread, but Tauri already owns it. Previously, the headless browser launch panicked with "EventLoop must be created on the main thread!" on a spawned thread, aborting the process. Fix: on macOS, `Browser::launch` is disabled via `set_unavailable()` at startup and an external renderer is registered that reuses Tauri's existing webview infrastructure. A hidden `WebviewWindow` is created with `on_page_load` detection — the renderer waits for the actual page load event instead of a fixed delay, with a 30-second timeout. The user can cancel the fetch at any time via the tool cancellation button. This gives full JS-rendered page content (weather widgets, SPAs, etc.) without a second event loop. On Linux/Windows the standalone headless browser continues to work as before.
- **Web search render=true no longer returns empty/error results**: When the external renderer times out or returns empty content for a URL, the system now automatically falls back to plain HTTP fetch + HTML tag stripping for that specific URL instead of propagating the error. This ensures the LLM always gets usable content and doesn't waste rounds retrying.

- **Fix LLM crash when prompt exceeds n_batch**: Long prompts triggered a fatal `GGML_ASSERT(n_tokens_all <= cparams.n_batch)` abort in llama.cpp, killing the entire process. The prompt is now decoded in chunks of `n_batch` tokens, preventing the native assertion failure.

- **Screenshot "sessions only" gate never re-engages after disconnect**: `session_start_utc` was set when scanning began but never reset to `None` in `go_disconnected`, so `is_session_active()` permanently returned `true` after the first connection attempt. Screenshots continued capturing even with no device connected. Now `session_start_utc` is always cleared on disconnect, including during auto-reconnect retries (no data is streaming, so it is not an active session).

- **Fix unused import warnings in skill-router / src-tauri**: Removed ~30 unused `use` statements from `src-tauri/src/lib.rs`, gated `std::sync::Mutex` import in `state.rs` behind `#[cfg(not(feature = "llm"))]`, removed unused `CalibrationProfile` import from `helpers.rs`, and converted doc comment on macro invocation in `window_cmds.rs` to a regular comment to silence `unused_doc_comments` warning.

- **Remove broken DuckDuckGo JSON API search**: the DuckDuckGo Instant Answer API (`api.duckduckgo.com`) has been deprecated and returns empty results for most queries, adding unnecessary latency. Removed it and now use HTML scraping directly as the sole search strategy.

- **Fix all clippy warnings**: Resolved redundant field name, unnecessary cast, unneeded return statement, and unused imports across the workspace. Zero clippy warnings remain.

- **DND focus mode now works on all devices**: OpenBCI and Hermes sessions were missing the Do Not Disturb tick logic (only Muse and MW75 had it). The shared `session_runner` now runs DND for every device that produces EEG band snapshots.

- **Battery alerts use `BatteryEma` from `skill-devices`**: Replaced two inline EMA implementations (Muse, MW75) with the existing `BatteryEma` struct, ensuring consistent smoothing and alert thresholds across devices.

### Refactor

- **Shared calibration CRUD service**: Extracted `calibration_service.rs` with `create_profile`, `update_profile`, `delete_profile`, `list_profiles`, and `get_profile` functions. Both the Tauri IPC commands (`window_cmds`) and the WebSocket API (`ws_commands`) now delegate to this single service, eliminating duplicated state mutation logic.

- **Extract LLM HTTP handlers from `engine.rs` (2,502 → 2,079 lines)**: Moved all axum HTTP handlers, auth helpers, and the router builder into a new `handlers.rs` (449 lines) in the `skill-llm` crate. `engine.rs` retains the inference actor, tool orchestration, and shared state.

- **Extract web search backends from `skill-tools/exec.rs` (1,551 → 944 lines)**: Moved DuckDuckGo HTML, Brave API, SearXNG, and headless fetch code into `search.rs` (616 lines).

- **Extract platform capture from `skill-screenshots/capture.rs` (1,527 → 1,145 lines)**: Moved macOS/Linux/Windows window capture and image decoding into `platform.rs` (392 lines).

- **Extract EEG band metrics from `skill-eeg/eeg_bands.rs` (1,510 → 1,222 lines)**: Moved advanced metric functions (SEF, Hjorth, entropy, DFA, consciousness indices) into `band_metrics.rs` (298 lines).

- **Chat page component extraction**: Split the monolithic `chat/+page.svelte` (2720 lines) into 9 focused modules — `ChatHeader`, `ChatSettingsPanel`, `ChatToolsPanel`, `ChatMessageList`, `ChatInputBar`, `ChatToolCard` components plus `chat-types.ts` and `chat-eeg.ts` utility modules. The main page is now 897 lines (67% reduction) with each extracted component under 400 lines.
- **Compare page logic extraction**: Extracted types, constants, helpers, insight computation, UMAP analysis, and all canvas drawing functions from `compare/+page.svelte` (2582 lines) into `compare-types.ts` and `compare-canvas.ts`. The main page is now 1924 lines (25% reduction) with 663 lines of pure logic in reusable modules.
- **History page helper extraction**: Extracted types, constants, date helpers, format utilities, label analysis, and grid constants from `history/+page.svelte` (2332 lines) into `history-helpers.ts` (201 lines). The main page is now 2224 lines.
- **Search page logic extraction**: Extracted types, helpers, colormap functions, analysis computation, graph deduplication, and kNN visualization from `search/+page.svelte` (2192 lines) into `search-types.ts` (302 lines). The main page is now 1935 lines (12% reduction).
- **UMAP viewer helper extraction**: Extracted pure math, color, timestamp, and geometry helpers from `UmapViewer3D.svelte` (1806 lines) into `umap-helpers.ts` (167 lines). The main component is now 1765 lines.
- **Deduplicated shared code**: Consolidated `BandSnapshot` type (single canonical source in `BandChart.svelte`, re-exported via `chat-types.ts`), `SESSION_COLORS` (moved to `constants.ts`, re-exported from `compare-types.ts` and `history-helpers.ts`), and `fmtSize`/`fmtGB`/`fmtBytes` format helpers (added to `format.ts`, removed 3 inline duplicates).

- **Deduplicate metrics JSON serialization in `DayStore`**: Extracted the ~60-field metrics-to-JSON serialization into a shared `metrics_to_json()` function, eliminating the copy-pasted logic between `insert()` and `insert_metrics_only()`. Reduces `day_store.rs` by ~65 lines.

- **Replace raw `SQLITE_OPEN_READ_ONLY` flags with `open_readonly()` helper**: Replaced 11 inline `rusqlite::Connection::open_with_flags(…, READ_ONLY)` calls across 7 files with the existing `skill_data::util::open_readonly()` helper for consistency with the workspace crates.

- **Deduplicate `MutexExt` trait**: Moved the poison-recovering `MutexExt` trait to `skill-constants` (zero-dependency crate). `skill-data` and `skill-jobs` now re-export from the single canonical definition instead of maintaining independent copies.

- **Add `AppStateExt` helper trait**: Introduced a blanket `AppStateExt` trait on `Manager<Wry>` that replaces the verbose `app.state::<Mutex<Box<AppState>>>()` pattern (137 call sites) with `app.app_state()`. Cleaned up newly-unused `Mutex` and `AppState` imports across 13 files.

- **Remove 14 re-export shim modules**: Eliminated one-line facade modules in `lib.rs` (e.g. `mod eeg_bands { pub use skill_eeg::eeg_bands::*; }`) that only proxied upstream crate items. All 59 call sites now reference the source crates directly (`skill_eeg::`, `skill_data::`) making dependencies explicit and reducing indirection.

- **Unified device session via `DeviceAdapter` trait**: Replaced four copy-pasted session modules (`muse_session.rs`, `mw75_session.rs`, `openbci_session.rs`, `hermes_session.rs` — 2,070 lines) with a trait-based architecture (1,970 lines). Added `DeviceAdapter` async trait, unified event types (`DeviceEvent`, `EegFrame`, `PpgFrame`, `ImuFrame`, `BatteryFrame`), and capability flags (`DeviceCaps`) to `skill-devices::session`. Each device has a small adapter (107–223 lines) that translates vendor events into the common vocabulary. A single generic event loop in `session_runner.rs` handles DSP, CSV, DND, battery, and emit for all devices.

- **Split `llm.rs` into module directory**: Refactored the 1537-line `src-tauri/src/llm.rs` into `src-tauri/src/llm/` with focused sub-modules: `mod.rs` (re-exports, logger, emitter), `cmds/catalog.rs` (catalog queries), `cmds/downloads.rs` (download lifecycle), `cmds/selection.rs` (model selection), `cmds/server.rs` (server lifecycle), `cmds/chat.rs` (chat persistence), `cmds/streaming.rs` (IPC streaming), `cmds/hardware_fit.rs` (hardware prediction). All public API paths remain unchanged.

- **Split `skill-llm/engine.rs` into module directory**: Refactored the 2070-line `crates/skill-llm/src/engine.rs` into `engine/` with focused sub-modules: `mod.rs` (re-exports, macros), `logging.rs` (log buffer, file sink), `protocol.rs` (wire types), `state.rs` (server state, cell, status), `think_tracker.rs` (think-budget enforcement), `images.rs` (base64 decoding), `tool_orchestration.rs` (multi-round tool loop), `sampling.rs` (token sampling loop), `generation.rs` (text/multimodal generation), `actor.rs` (OS thread event loop), `init.rs` (public init). All public API paths remain unchanged.

- **Split `eeg_embeddings.rs` (2,434 → 3 files)**: Removed 624 lines of dead code (`#[cfg(any())]` blocks). Split the remaining 1,810 lines into `mod.rs` (414 lines, public API + EegAccumulator), `day_store.rs` (372 lines, per-day HNSW + SQLite), and `worker.rs` (1,051 lines, background embed worker + hook matcher + weight helpers).

- **Split `ws_commands.rs` (2,417 → 3 files)**: Extracted `hooks.rs` (315 lines, hooks_get/set/status/suggest/log handlers) and `llm_cmds.rs` (418 lines, all LLM WebSocket commands), reducing `mod.rs` to 1,714 lines. The central `dispatch()` function remains in `mod.rs`.

- **Split `settings_cmds.rs` (1,789 → 3 files)**: Extracted `dnd_cmds.rs` (207 lines, Do Not Disturb automation commands) and `hook_cmds.rs` (274 lines, hook distance suggestion and audit log), reducing `mod.rs` to 1,341 lines.

- **Split `skill-commands` crate `lib.rs` (1,710 → 2 files)**: Extracted `graph.rs` (803 lines, DOT and SVG generation for interactive search results), reducing `lib.rs` to 917 lines.

- **Extract `chat-utils.ts` from `chat/+page.svelte` (2,945 → 2,720 lines)**: Moved pure utility functions (tool-call fence stripping, danger detection, assistant output parsing) into a standalone TypeScript module for testability and reuse.

- **Split `lib.rs` into `state.rs` + `helpers.rs`**: Extracted `AppState`, `DeviceStatus`, `LlmState`, IPC packet structs, and all `impl` blocks into `state.rs` (~410 lines). Extracted time helpers, emit/toast helpers, settings persistence, device upsert, and state access shortcuts into `helpers.rs` (~250 lines). `lib.rs` dropped from 2,217 to 1,557 lines (–660).

- **Renamed `MuseStatus` → `DeviceStatus`**: The status struct is used for all devices (Muse, MW75, Hermes, OpenBCI) — the old name was misleading. Updated all Rust and TypeScript/Svelte references. The Tauri event name `"muse-status"` is kept for backward compatibility with existing WS clients.

- **`window_cmd!` / `window_tab_cmd!` macros for window commands**: Replaced 10 boilerplate `open_*_window` Tauri commands (each 7 lines) with 2–3 line macro invocations. New windows now require a single `window_cmd!` call instead of a full `#[tauri::command] pub async fn` definition.

- **Extract `ws_commands/search.rs` (525 lines)**: Moved search_labels, search, compare, session_metrics, and interactive_search handlers out of `ws_commands/mod.rs`, reducing it from 1,687 to 1,162 lines.

- **Extract `skill-history/cache.rs` (751 lines)**: Moved disk cache, metrics computation, sleep staging, and batch loading out of `skill-history/lib.rs`, reducing it from 1,384 to 654 lines.

- **Deduplicate `yyyymmdd_utc()`**: Replaced the 29-line hand-rolled calendar calculation in `helpers.rs` with a one-line delegation to the canonical `skill_data::util::yyyymmdd_utc()`.

- **Added 25 tests for `skill-devices::session` adapters**: Covers channel accumulation (Muse alignment of per-electrode delivery, partial/complete frames, out-of-range electrodes), event translation (EEG, PPG, IMU, battery, connected, disconnected, activation-skipped, packets-dropped-skipped), synthetic connected injection (MW75 RFCOMM, OpenBCI), capability flags, descriptor construction, and pipeline channel capping. Total: 34 tests in `skill-devices` (up from 9).

- **Made adapter handle fields `Option`-based for testability**: `Mw75Adapter` and `HermesAdapter` handle fields are now `Option<Handle>` with `new_for_test()` constructors that pass `None`, avoiding unsafe `MaybeUninit` hacks. `OpenBciAdapter` gained `from_receiver()` for direct async channel injection without needing private `StreamHandle` fields.

### Build

- **clean:rust npm script**: Added `npm run clean:rust` to remove all Rust `target` directories (under `crates/` and `src-tauri/`).

- **Added `tokio`, `async-trait`, `bitflags` deps to `skill-devices`**: Only `tokio/sync` (channel primitives) is used — no runtime. Added `tokio-util` to `src-tauri` for `CancellationToken`.

### CLI

- **`health` command**: New CLI command family for Apple HealthKit data. Subcommands: `health` / `health summary` (aggregate counts), `health sleep` / `health workouts` / `health hr` / `health steps` / `health metrics` (typed queries with `--start`, `--end`, `--limit`), `health metric-types` (list stored types), and `health sync` (push data from iOS companion). Human-readable formatting with color-coded output for each data type.

- **`sleep-schedule` command**: New CLI command to view and update the sleep schedule. Supports `sleep-schedule` (show current) and `sleep-schedule set --bedtime HH:MM --wake HH:MM --preset <id>` (update). Available presets: default, early_bird, night_owl, short_sleeper, long_sleeper.

- **LLM error diagnostics in CLI**: The `llm chat` command (both single-shot and REPL modes) now classifies known LLM error patterns (batch overflow, context window exceeded, decode failures, template errors, native panics, tokenization failures) and prints actionable hints alongside the error message.

### UI

- **Log viewer filtering**: The LLM server log in the settings tab now has level filter tabs (All / Info / Warn / Error) and a text search box. The line count shows filtered/total when a filter is active, and "No matching lines" is shown instead of "No log output yet" when filtering produces no results.

- **Tool context compression controls**: Added compression level selector and optional max-search-results / max-result-chars overrides to both the Settings → LLM → Tools tab and the inline chat tools panel.

- **Separate Tools settings tab**: Moved LLM chat tools configuration (per-tool toggles, SearXNG URL, execution mode, max rounds, max calls per round) out of the LLM tab into its own dedicated "Tools" settings tab with a wrench icon and i18n labels for all five languages.

- **Web search sources panel**: When expanding a `web_search` tool call in chat, a new "Sources" section shows each fetched domain with its quality score, content size, and a "best" badge for the highest-scoring result. Each source is expandable to reveal the URL and a content preview, allowing users to inspect exactly what data the LLM received from each site.

- **LLM tools section in own card**: Moved the built-in chat tools configuration out of the collapsible Advanced Inference Settings panel into its own top-level section card with a header showing the enabled count (e.g. "5/8"). The execution mode toggle is placed in a footer row within the same card. This makes tools discoverable without expanding the Advanced section.

- **Hide charts when no session is active**: Band Powers and EEG Waveforms charts are now hidden on the main dashboard when the device is not connected, reducing visual clutter in the disconnected/scanning states.
- **Show PPG/IMU only for capable devices**: PPG charts, PPG metrics, Head Pose card, and IMU chart are now gated by `deviceCaps.hasPpg` / `deviceCaps.hasImu` so they only appear for devices that actually have those sensors.

- **Open Folder button next to data dir input**: Moved the "Open Folder" button from below the default path label to inline next to the data directory input field in Settings, making it more discoverable and accessible.

### Server

- **`sleep_schedule` / `sleep_schedule_set` WS commands**: New WebSocket API commands for reading and writing the sleep schedule configuration. Partial updates supported — only fields present in the request are changed.

### i18n

- Added `llm.autostart`, `llm.autostartDesc` keys to all 5 locales (en, de, fr, uk, he).
- Added `chat.btn.aborting`, `chat.btn.regenerate`, `chat.btn.editResend`, `chat.tokSec`, `chat.ctxWarning`, `chat.noModelBtn`, `chat.logFilter.*` keys to en locale.

- **Search provider strings**: added search provider selector, Brave API key, and SearXNG URL translations in en, de, fr, uk, and he.

- **Context compression labels**: Added translations for context compression settings in English, German, French, Hebrew, and Ukrainian.

- **Sources label**: Added "Sources" / "Quellen" / "Sources" / "מקורות" / "Джерела" translations for the web search sources UI.

- **Convert all Unicode escapes to literal UTF-8**: Replaced all `\uXXXX` escape sequences in i18n locale files (en, de, fr, uk, he) with their literal UTF-8 characters. This makes the translation strings human-readable in source and avoids encoding issues across platforms.

- **Add 14 missing chat translation keys**: Added `chat.btn.aborting`, `chat.btn.editResend`, `chat.btn.regenerate`, `chat.ctxCompact`, `chat.ctxWarning`, `chat.logFilter.*`, `chat.noModel*`, and `chat.tokSec` to German, French, Ukrainian, and Hebrew. All 5 languages now have identical key sets (2,646 keys each).

### Docs

- **SKILL.md**: Added full `health` command reference with subcommand table, CLI examples, HTTP equivalents, JSON response shapes, sync payload format, and common metric types.

- **SKILL.md**: Added `sleep-schedule` command reference with examples, HTTP equivalents, JSON response shapes, and preset table.

- **Updated Discord invite link**: Changed Discord link to `https://discord.gg/Rcvb8Cx4cZ` across README, constants, and help dashboard.

### Dependencies

- **Upgrade skill-headless wry to 0.54**: Updated `skill-headless` crate from `wry` 0.49 to 0.54.3 to match the workspace's tauri-runtime-wry dependency, resolving a `kuchikiki` version conflict. The only API change was renaming `WebViewBuilder::with_web_context()` to `WebViewBuilder::new_with_web_context()`.

### CI

- **Discord commit message**: All CI workflow Discord notifications now include the last commit subject line, making it easier to identify which change triggered the build.

## [0.0.40] — 2026-03-16

### Bugfixes

- **CI: fix latest.json encoding and Python indentation**: Fixed IndentationError in macOS workflow (`if not notes:` was mis-indented inside `except` block) and inconsistent indentation in Linux workflow. Replaced literal `™` with `\u2122` escape in Python scripts and added `ensure_ascii=False` to all `json.dump` calls so `latest.json` is always coherent UTF-8 across all three platform CIs.

- **Fix LINUX.md path in packaging scripts**: Updated `package-linux-dist.sh` and `package-linux-system-bundles.sh` to reference `docs/LINUX.md` instead of the non-existent root-level `LINUX.md`, fixing a `cp: cannot stat` error on Linux CI.

- **Windows CI: fix trademark character encoding**: Replace literal `™` (U+2122) with PowerShell escape `$([char]0x2122)` in the Windows release workflow to prevent `Unexpected token` parse errors.

### Docs

- **AGENTS.md: add CI shared-artifact encoding rule**: New section documenting that all CI workflows must produce and consume `latest.json` as UTF-8 without BOM, with no literal non-ASCII characters in CI scripts.

## [0.0.39] — 2026-03-15

### Bugfixes

- **Fix syntax error in German locale**: Removed corrupted duplicate fragment (`ult de;`) at the end of `src/lib/i18n/de.ts` that caused a build failure.

- **Fix LLM catalog crash on Windows**: Replaced the `llm_catalog.json` symlink in `skill-llm` with a direct `include_str!` path to the source file. Git on Windows checks out symlinks as plain-text files (containing the target path), which caused an invalid-JSON panic at startup.

- **Fix mw75 RFCOMM Send bound violation on Windows**: Bumped `mw75` to 0.0.6 which wraps the RFCOMM `tokio::spawn` future with an `AssertSend` adapter. WinRT COM objects (`IInputStream`, `DataReader`, `StreamSocket`, `IVectorView`, etc.) are thread-safe under MTA but not marked `Send` by the `windows` crate.

- **Fix mw75 RFCOMM build on Windows**: Vendored `mw75` crate with fix for `READ_BUF_SIZE` constant that was gated behind `#[cfg(target_os = "linux")]` but used in the Windows RFCOMM code path, causing compilation failure.

- **Fix mw75 Windows RFCOMM build**: Updated to mw75 v0.0.4 which fixes compatibility with `windows` crate v0.62 by replacing removed `.get()` calls with async/await.

- **Fix 12 failing tests after EEG_CHANNELS bump to 12**: Updated `constants.test.ts` to expect `EEG_CHANNELS = 12` (matching Rust `skill-constants`), decoupled `EEG_CH`/`EEG_COLOR` length assertions from `EEG_CHANNELS` (they are Muse-specific with 4 entries), and updated `BAND_CANVAS_H` expected value from 290 to 642. Fixed over-escaped `\\\"` sequences in `helpTts.apiBody`, `helpFaq.a33`, and `helpSettings.calibrationTtsBody` across de/fr/he/uk locale files that caused the i18n key-extraction regex to detect spurious extra keys (`command\`, `text\`). Updated stale comment in `constants.ts`.

- **Fix Windows CI PowerShell parse errors**: Added UTF-8 BOM to `create-windows-nsis.ps1`, `release-windows.ps1`, and `setup-build-cache.ps1` so Windows PowerShell 5.1 correctly reads non-ASCII characters (™, —). Replaced `?.` null-conditional operator (PowerShell 7+ only) in `release-windows.ps1` with a PS 5.1-compatible alternative.

### Refactor

- **Remove vendored mw75 crate**: Migrated from a local vendored copy of `mw75` to the published crates.io version (0.0.3). The Windows RFCOMM `READ_BUF_SIZE` fix was upstreamed and published.

## [0.0.38] — 2026-03-15

### Features

- **Proactive Hooks**: background EEG monitoring that triggers actions when brain-state matches configured labels. Per-hook scenarios (cognitive/emotional/physical), keyword suggestions, distance threshold, fire history, WebSocket events, and full CLI CRUD.

- **Hook distance suggestion**: analyses HNSW data to recommend optimal thresholds with percentile visualization.

- **LLM coding-agent tools**: bash execution, read/write/edit file, web search (DuckDuckGo JSON + HTML fallback), web fetch, search_output for navigating large outputs. Safety approval dialogs for dangerous operations. Context-aware tool calling with automatic history trimming.

- **Tool-call cards with rich detail views**: expandable cards per tool type with cancel/stop, danger detection, and structured argument display.

- **Chat session archive (soft-delete)**: archive instead of permanent delete, with restore and permanent-delete from archive section.

- **Neurable MW75 Neuro headphone support**: Full 12-channel EEG session at 500 Hz. BLE activation + RFCOMM data streaming (behind `mw75-rfcomm` feature flag). Electrode placement guide shows MW75 ear-cup layout with 6 electrodes per ear (FT7/T7/TP7/CP5/P7/C5 left, FT8/T8/TP8/CP6/P8/C6 right). All 12 channels render in the dashboard: signal quality dots, EEG waveforms, spectrogram, and band powers. DSP pipeline processes all active channels. Device presets for Muse (4ch), Ganglion (4ch), and MW75 (12ch) in electrode guides.

- **Hermes 10-20 electrode labels**: Replaced generic `Ch1`–`Ch8` channel names with standard 10-20 positions (`Fp1`, `Fp2`, `AF3`, `AF4`, `F3`, `F4`, `FC1`, `FC2`) in `skill-constants` and the electrode placement SVG guide.

- **Hermes V1 EEG headset support**: Added full session support for the Hermes V1 headset (8-channel ADS1299 at 250 Hz, 9-DOF IMU). The `hermes-ble` crate is added to `skill-devices` and re-exported. All data streams over BLE GATT — no RFCOMM needed. BLE scanner recognises devices whose name starts with "Hermes". Session handles EEG (8 channels through DSP pipeline), IMU (accel + gyro → head pose), and packet-drop detection. Dashboard renders 8 channels dynamically with device-specific labels and colours. Electrode placement guide and 3D electrode guide include Hermes V1 tab. Constants added to `skill-constants` (`HERMES_EEG_CHANNELS`, `HERMES_SAMPLE_RATE`, `HERMES_CHANNEL_NAMES`).

- **Screenshot capture + vision embedding system**: periodic active-window capture with CLIP vision embedding (ONNX) and HNSW index. macOS CoreGraphics FFI, Linux X11/Wayland, Windows GDI. Configurable interval, size, quality, and embedding backend.

- **OCR text extraction + text embedding**: on-device OCR via `ocrs` crate. Dual HNSW architecture for visual and text similarity search.

- **Screenshots Settings UI tab**: full configuration with live re-embed progress.

- **Configurable OCR engine, GPU/CPU toggle**: `ScreenshotConfig` extended with OCR engine, model, and GPU settings.

### Performance

- **Unblock main & settings window startup by moving I/O-heavy Tauri commands to async threads**: Converted 12 synchronous `#[tauri::command]` handlers that performed directory scanning, JSON parsing, or SQLite queries on the Tauri IPC executor thread to `async` commands using `tokio::task::spawn_blocking`. This prevents those operations from stalling window rendering and other IPC calls during startup. Affected commands: `list_sessions`, `delete_session`, `list_embedding_sessions` (history_cmds); `get_daily_recording_mins`, `suggest_hook_distances`, `get_hook_log`, `get_hook_log_count`, `list_serial_ports` (settings_cmds); `query_annotations`, `get_recent_labels`, `delete_label`, `update_label`, `get_stale_label_count` (label_cmds).

- **History view: batch metrics loading with disk cache**: Replaced per-session IPC waterfall (4 concurrent `get_csv_metrics` calls) with a single `get_day_metrics_batch` call that loads all sessions' metrics in one roundtrip. Added persistent `_metrics_cache.json` disk cache next to each session CSV — subsequent loads skip CSV re-parsing entirely. Timeseries payloads are downsampled to ≤360 points on the backend, reducing transfer size for sparklines and heatmaps. Week view also uses a single batch call for all 7 days. Adjacent-day prefetching now batch-loads as well.

- **Shared text embedder — eliminate 3 redundant model copies (~390 MB RAM saved)**: consolidated four independent `fastembed::TextEmbedding` instances into a single app-wide `Arc<Mutex<TextEmbedding>>`.

- **GPU optimization across all model systems**: enabled flash attention and KQV offloading in LLM; GPU-first mmproj loading on Linux; DirectML on Windows and CUDA on Linux for screenshot embeddings.

- **Async history commands**: Converted `list_session_days`, `list_sessions_for_day`, `get_session_metrics`, `get_session_timeseries`, `get_csv_metrics`, `get_day_metrics_batch`, and `get_sleep_stages` from synchronous Tauri commands to async commands using `tokio::task::spawn_blocking`, preventing UI thread blocking during heavy file I/O and CSV/SQLite parsing.

### Bugfixes

- **Heatmap session counts**: `daySessionCounts` now reads from the localStorage day cache to return actual session counts instead of always 1, making month/year heatmap intensity meaningful.

- **Fix clippy warnings and a11y lint in history page**: Resolved 9 clippy warnings — `let_and_return` in ws_commands.rs, `double_ended_iterator_last` and `collapsible_str_replace` in llm.rs, `iter_cloned_collect` in hermes/mw75 sessions, `needless_borrow` in settings_cmds.rs. Converted label dot `<span>` elements to `<button>` with `aria-label`, `onfocus`/`onblur` handlers for keyboard accessibility.

- **History view hover fixes**: Fixed label hover not working on the day-grid heatmap canvas — hovering a cell containing a label now correctly highlights matching labels and shows the label tooltip. Fixed week view hover completely missing — added mouse event handlers to the day-dots canvas so both epoch dot and label circle hovers show tooltips. Moved tooltip rendering outside view-mode conditionals so tooltips are visible in all views.

- **History label dot tooltips no longer clipped**: Replaced `absolute`-positioned tooltips on label dots with a `fixed`-position portal element, preventing overflow clipping from parent containers with `overflow-hidden`. Tooltips now always render above all content regardless of DOM nesting.

- **Fix missing objc2 dependencies in skill-screenshots**: Added `objc2` and `objc2-app-kit` as macOS-only dependencies to resolve compilation errors when resolving the frontmost application PID via NSWorkspace.

- **Fix all svelte-check warnings and i18n gaps**: Resolved 5 Svelte compiler warnings — fixed `propColors` initial-value capture in `EegChart.svelte` by using `EEG_COLOR` fallback, fixed `device` prop capture in `ElectrodeGuide.svelte` by deferring initial tab selection to `$effect`, and fixed non-interactive element warning in `history/+page.svelte` by using `role="toolbar"` with `tabindex`. Added 215 missing English-fallback i18n keys across de, fr, he, and uk locales.

- **Fix `$state` invalid placement in titlebar-state**: Moved `$state(initial)` from a `return` statement to a variable declaration initializer, fixing the Svelte 5 `state_invalid_placement` error in `src/lib/titlebar-state.svelte.ts`.

### Refactor

- **Dynamic multi-channel DSP pipeline**: `EEG_CHANNELS` raised from 4 to 12 (max across all devices). `EegFilter` and `BandAnalyzer` track active channels and only wait for channels that have received data before firing GPU batches. Muse/Ganglion sessions use channels 0–3; MW75 uses all 12. Inactive channels have zero overhead.

- **Dynamic channel rendering**: EegChart, BandChart, signal quality, and EEG channel values all accept dynamic channel count/labels/colors via props. MW75 renders 12 channels in a 3-column grid; Muse/Ganglion render 4 in 2 columns.

- **Centralize hardcoded constants with prelude module**: added `skill_constants::prelude` re-exporting ~60 most-used constants; moved 50+ hardcoded values from `skill-eeg` (signal quality thresholds, artifact detection params, head pose params), `skill-data` (PPG IBI limits, DND identifiers), `skill-devices` (SNR thresholds), `skill-label-index` (index file names, HNSW params), `skill-llm` (catalog file, log dir/cap), `skill-tts` (event name, silence duration, KittenTTS config), `skill-tray` (menu rebuild debounce), `skill-tools` (tool-call delimiters, bash limits), and `src-tauri` (session gap, PPG channels, active window idle, WS request log cap) into `skill-constants` as the single source of truth; added `skill-constants` dependency to 5 crates that previously lacked it.

- **Cross-crate deduplication**: consolidated shared utilities into `skill-data::util` — `date_dirs`, `MutexExt`, UTC timestamp formatters (`yyyymmdd_utc`, `yyyymmddhhmmss_utc`, `unix_to_ts`, `ts_to_unix`, `fmt_unix_utc`, `civil_from_unix`), and `open_readonly` SQLite helper. Removed duplicate implementations from `skill-commands`, `skill-label-index`, `skill-exg`, `skill-screenshots`, and `skill-router`.
- **Screenshot HNSW dedup**: replaced near-identical vision/OCR HNSW load/save/rebuild function pairs in `skill-screenshots` with a generic `load_or_rebuild_hnsw_generic` + `save_hnsw_to` parametrized by path and fetch closure.
- **Band-snapshot enrichment**: extracted `enrich_band_snapshot` + `SnapshotContext` into `skill-devices`, eliminating ~90 lines of duplicated PPG/artifact/head-pose/composite-score/GPU enrichment from `muse_session.rs` and `openbci_session.rs`.
- **DND decision dedup**: replaced ~200-line inline DND decision block in `muse_session.rs` with the existing `skill_devices::dnd_tick()` pure function and proper state round-tripping.
- **Constants consistency**: fixed hardcoded `"labels.sqlite"` in `skill-data::label_store` to use `LABELS_FILE` from `skill-constants`.

- **Deduplicate frontend types and formatting helpers**: extracted shared TypeScript interfaces into `$lib/types.ts`; extracted 12 formatting functions into `$lib/format.ts`; extracted `SleepAnalysis` into `$lib/sleep-analysis.ts`; updated 15 consumer files; eliminates ~30 duplicate interface definitions and ~20 duplicate utility functions.

- **Canvas chart dedup**: migrated ImuChart, PpgChart, GpuChart, and BandChart to use the shared `animatedCanvas` Svelte action from `use-canvas.ts`, eliminating duplicated ResizeObserver + requestAnimationFrame + DPR scaling boilerplate. EegChart is intentionally left as-is due to its spectrogram tape + MutationObserver + frame-skip complexity.

- **HuggingFace cache path consolidation**: added `hf_cache_root()`, `hf_model_dir()`, and `hf_ensure_dirs()` helpers to `skill-data::util`. Replaced the manual env-var resolution in `skill-exg::resolve_hf_weights` and the duplicated `Cache::from_env().path()` + folder construction pattern in both `skill-exg` and `skill-llm::catalog`. Removed the now-unused `dirs` crate dependency from `skill-exg`.

- **Titlebar store factory**: added `createTitlebarState()` and `createTitlebarCallbacks()` in `titlebar-state.svelte.ts`. Refactored `chat-titlebar`, `history-titlebar`, and `label-titlebar` stores to use the shared factory instead of raw `$state()` calls.

- **Deduplicate Rust and frontend code**: replaced duplicated `MutexExt` trait in `src-tauri` with re-export from `skill-data::util`; removed duplicated `unix_secs_now()` in `eeg_embeddings.rs`; exported `rgba()` from `theme.ts` and removed duplicate in `GpuChart.svelte`; extracted `MuseStatus::reset_for_scanning()` to replace ~20-line identical status reset blocks in `muse_session.rs` and `openbci_session.rs`; converted 14 pure re-export shim files to inline module declarations in `lib.rs` and deleted the files; added `PPG_SAMPLE_RATE` to `skill-constants` and derived sample-rate constants in `session_csv` from the canonical source; eliminated `skill-eeg/src/constants.rs` shim file.

- **Deduplicate Tauri backend boilerplate**: Added `skill_dir()`, `read_state()`, `mutate_state()`, and `mutate_and_save()` helpers to `lib.rs`, replacing ~30 repetitive `state.lock_or_recover().skill_dir.clone()` call sites across `commands.rs`, `ws_commands.rs`, `settings_cmds.rs`, `label_cmds.rs`, `history_cmds.rs`, `session_analysis.rs`, and `global_eeg_index.rs`. Added `search_params()` helper in `commands.rs` to deduplicate the k/ef clamping pattern. Added generic `load_json_or_default()`, `save_json()`, and `init_wal_pragmas()` to `skill-data::util`, replacing duplicated JSON config I/O and SQLite PRAGMA patterns across `skill-settings`, `skill_log`, `activity_store`, `screenshot_store`, `hooks_log`, and `eeg_embeddings`. Applied `mutate_and_save()` to ~12 simple set-then-persist patterns in `window_cmds.rs` and `settings_cmds.rs`.

- **Extract `skill-commands` workspace crate**: moved EEG embedding search, timestamp helpers, SVG/DOT graph generation, PCA projection, and streaming search (2,321 lines) into `crates/skill-commands/`. Zero Tauri dependencies.

- **Extract `skill-devices` workspace crate**: moved DND focus-mode engine, composite EEG scores, and battery EMA from `muse_session.rs` (774 lines) into `crates/skill-devices/`. Zero Tauri dependencies. Includes 9 unit tests.

- **Extract `skill-exg` workspace crate**: moved cosine distance, fuzzy matching, HF weight management, GPU cache, and epoch metrics from `eeg_embeddings.rs` (2,613 lines) into `crates/skill-exg/`. Zero Tauri dependencies.

- **Extract `skill-jobs` workspace crate**: moved sequential job queue (384 lines) into `crates/skill-jobs/`. Zero Tauri dependencies. All 3 unit tests pass.

- **Extract `skill-router` workspace crate**: moved UMAP projection, embedding/label loaders, cluster analysis, and metric rounding from `ws_commands.rs` (2,408 lines) into `crates/skill-router/`. Zero Tauri dependencies.

- **Extract `skill-settings` workspace crate**: moved persistent configuration types and disk I/O (924 lines) into `crates/skill-settings/`. All 27 unit tests pass.

- **Extract `skill-tray` workspace crate**: moved progress-ring overlay, shortcut formatting, and dedup helpers from `tray.rs` (674 lines) into `crates/skill-tray/`. Pure `std`, zero dependencies. Includes 8 unit tests + 2 doc-tests.

- **Flatten `llm/` and `tts/` module directories into single files**: converted multi-file modules into single files. No API or import path changes.

- **Remove dead `llm/` duplicates**: deleted 1,277 lines of dead code (byte-identical copies of `catalog.rs` and `chat_store.rs`).

- **Deduplicate frontend formatting, canvas setup, navigation, and UI patterns**: replaced ~15 inline date formatting calls with shared `format.ts` helpers; added `dateToLocalKey()`, `dateToCompactKey()`, `fromUnix()`, `toUnix()`, `setupHiDpiCanvas()`, `getDpr()`, `fmtCountdown()`, `fmtDateTimeLocalInput()`, `parseDateTimeLocalInput()`, `localKeyToUnix()` to `format.ts`; replaced 23 DPR canvas setup blocks across 8 chart files; migrated `EegIndices`, `FaaGauge`, `HeadPoseCard` to `CollapsibleSection`; created `$lib/navigation.ts` with shared window-open helpers used by 8 files; created `ConfirmAction` UI component replacing inline confirm-delete in history and labels; cleaned up compare page internal duplication (5 DPR blocks, inline date helpers).

- **Move muse-rs and openbci dependencies to skill-devices crate**: Moved `muse-rs` and `openbci` dependency declarations from `src-tauri/Cargo.toml` to `crates/skill-devices/Cargo.toml` and re-exported them. Updated all imports in `muse_session.rs` and `openbci_session.rs` to use the re-exports via `skill_devices::`.

- **Extract `skill-history` crate**: Moved all session history, metrics, time-series, sleep staging, and analysis logic from `src-tauri/src/{history_cmds,session_analysis}.rs` into a new `crates/skill-history` workspace crate with zero Tauri dependencies. The Tauri files now contain only thin async IPC wrappers that delegate to `skill_history::*` and run on `spawn_blocking` threads. Types (`SessionEntry`, `SessionMetrics`, `EpochRow`, `CsvMetricsResult`, `SleepStages`, `HistoryStats`, `EmbeddingSession`) and all pure functions (`list_sessions_for_day`, `load_metrics_csv`, `get_session_metrics`, `get_sleep_stages`, `compute_compare_insights`, `analyze_sleep_stages`, `analyze_search_results`, `compute_status_history`, etc.) are now public API in the crate.

- **Shared canvas lifecycle action**: added `src/lib/use-canvas.ts` (`animatedCanvas` Svelte action) to DRY the ResizeObserver + requestAnimationFrame + DPR scaling boilerplate duplicated across EegChart, BandChart, PpgChart, GpuChart, and ImuChart. Existing charts are not yet migrated (tracked as a TODO).

### Build

- **Extract `skill-screenshots` workspace crate**: moved 1,533 lines of screenshot capture, vision embedding, HNSW search, and OCR into `crates/skill-screenshots/`.

- **Extract `skill-tools` workspace crate**: moved all LLM tool logic (definitions, execution, parsing, validation, safety) into `crates/skill-tools/`.

- **Create `skill-constants` crate**: single source of truth for all constants.

- **Extract `skill-data` workspace crate**: moved 2,984 lines of pure data/utility logic into `crates/skill-data/`.

- **Extract `skill-tts` workspace crate**: moved 1,307 lines of TTS logic into `crates/skill-tts/`.

- **Extract `skill-eeg` workspace crate**: moved 3,459 lines of EEG DSP into `crates/skill-eeg/`.

- **Extract `skill-llm` workspace crate**: moved 7,421 lines of LLM logic into `crates/skill-llm/`.

- **sccache + mold for faster builds**: auto-detected build caching. ~54% faster clean rebuilds.

- **Fragment-based changelog system**: replaced single-file `CHANGELOG.md` editing with `changes/unreleased/` fragments. Each change gets its own `.md` file; `npm run bump` compiles fragments into `changes/releases/<version>.md`, deletes consumed fragments, and rebuilds `CHANGELOG.md` from all release files. All 20 historical releases migrated. Supports `--rebuild` to regenerate from archives.

### CLI

- **Dynamic signal quality rendering**: `status` command now renders signal quality for any number of EEG channels (4/8/12) instead of hardcoding Muse's tp9/af7/af8/tp10 keys. Added device support note and tool-calling documentation to CLI header comment.

- **Full LLM management via CLI**: `llm select`, `llm mmproj`, `llm autoload-mmproj`, `llm pause`, `llm resume`, `llm downloads`, `llm refresh`, `llm fit`, `llm add`.

- **Calibration profile CRUD**: `calibrations create`, `calibrations update`, `calibrations delete`.

- **Hooks CRUD**: `hooks list`, `hooks add`, `hooks remove`, `hooks enable/disable`, `hooks update`, `hooks suggest`, `hooks log`.

### UI

- **History day view — 24×720 heatmap grid**: replaced the linear 24-hour timeline bar and epoch dot canvas with a dense heatmap grid (24 hour-columns × 720 five-second rows); cells colored by session color with opacity modulated by relaxation+engagement; hour headers, 15-minute grid lines, minute labels; scrollable canvas (max 420px); cursor-following tooltip with HH:MM:SS and data values.

- **History day view — rainbow label circles**: replaced text-based label display with tiny colored circles in session rows, expanded details, timeline legend, and canvas; rainbow hue distribution (0°–300° HSL) based on temporal order; hover highlights exact-match labels (glow ring + scale) and temporally close labels (within 5 min, brightness/glow); popover tooltips on hover; cross-session matching.

- **History day view — screenshot indicators on heatmap**: cells that have a corresponding screenshot show a small blue diamond indicator on the canvas grid; hovering a screenshot cell displays a floating image preview (loaded from the local API server); preview disappears immediately when the mouse moves to a different cell; screenshot data loaded per-day via `get_screenshots_around`.

- **History view UX improvements**: 14 enhancements to the history view:
  1. Week view shows total recording duration per day in sidebar (e.g. "2s · 1h 30m").
  2. Clicking on the day grid heatmap scrolls to and expands the corresponding session.
  3. Week view entire row is clickable to navigate to day view (not just sidebar).
  4. Month view calendar cells show mini duration bars proportional to recording hours.
  5. Day grid draws a red "now" marker hairline when viewing today.
  6. Keyboard shortcuts 1/2/3/4 switch between year/month/week/day views; arrow keys navigate in calendar views too.
  7. Week view today row has a visible left-edge accent bar instead of nearly invisible tint.
  8. Cross-highlighting between grid and session list — hovering a cell highlights the session row below; hovering grid sets a primary ring on the matching session card.
  9. Week↔Day view transitions use a subtle fade animation.
  10. Day view shows an aggregate daily summary card (total duration, avg relaxation/engagement, label count) above the session list when there are multiple sessions.
  11. Fixed `daySessionCounts` always returning 1 — heatmap intensity now reflects actual session count per day from cache.
  12. Month/year view tooltips show recording duration alongside session count.
  13. Expanding a session row scrolls it into view smoothly.
  14. Week view shows small session color legend dots in each day row.

- **MW75 Neuro device photo**: Added product image for the MW75 Neuro headphones, displayed on the dashboard when connected and in the device list in Settings.

- **History view accent consistency**: migrated all interactive/highlight colors in the history view from hardcoded `emerald-*` to accent-aware `violet-*` (remapped by the Appearance accent setting). Affected: year/month heatmap cells and legend swatches, month calendar day-count text, screenshot canvas diamond indicator (now reads `--color-violet-500` CSS property), and screenshot tooltip dot. Semantic status colors (`emerald-500` for positive trend, `red-*` for destructive actions) remain unchanged.

### i18n

- **Complete translation pass across de/fr/he/uk**: all untranslated keys now translated. `npm run audit:i18n` reports 0 untranslated keys.

- **Translation audit script** (`scripts/audit-i18n.ts`): detects untranslated keys with exemption system for technical terms.

### Docs

- **AI models reference**: Added `docs/AI.md` documenting all AI/ML models used across the codebase — LLM catalog, ZUNA EEG foundation model, CLIP/Nomic vision embeddings, OCR engines, TTS backends, HNSW search, and UMAP projection.

- **Add `README.md` to every workspace crate**: created README files for all 19 crates that were missing one; each contains heading and description from `Cargo.toml`.

- **Expand `crates/*/README.md`**: rewrote all 17 crate READMEs with comprehensive documentation — overview, public API tables, feature flags, and dependency lists.

- **Rewrite root `README.md`**: restructured to reference all workspace crate READMEs and `docs/` guides; added project layout section with crate table, vendored crates, and frontend structure.

- **Add workspace crate outline to `AGENTS.md`**: brief table of all 17 workspace crates and 2 vendored crates with one-line purpose descriptions.

- **Move `CHANGELOG.md` to project root**: merged the root-level entries into the comprehensive `docs/CHANGELOG.md` and moved to root.

- **Update SKILL.md for new devices, tools, and screenshots**: Added "Supported Devices" section documenting Muse, OpenBCI Ganglion, MW75 Neuro (12ch), and Hermes V1 (8ch) with channel counts and sample rates. Updated `status` JSON example to show device-agnostic fields (`eeg_channels`, multi-device name examples, dynamic signal quality keys). Added "Built-in Tool Calling" section to the LLM docs covering bash, read/write/edit, web search/fetch tools with safety info. Added "Screenshots (UI-only Feature)" section documenting capture, CLIP vision embedding, OCR, and dual HNSW search. Updated table of contents numbering.

- **Mermaid crate dependency graph**: Added an interactive Mermaid diagram to the README Architecture section showing how all workspace crates depend on each other and connect to the Tauri app and SvelteKit frontend.

- **Constants sync guard**: improved doc comment in `src/tests/constants.test.ts` explaining how the test file guards Rust↔TypeScript constant drift.

## [0.0.37] — 2026-03-14

### History — Calendar Heatmap View

- **Calendar heatmap** replaces the old single-day paginator as the default history view. Choose between **Year**, **Month** (default), **Week**, and **Day** granularity via a segmented control in the titlebar.
- **Year view**: GitHub-style contribution heatmap with weekly columns, month labels, and an intensity legend (emerald gradient). Click any day with recordings to drill into the day view.
- **Month view**: traditional calendar grid with per-day session counts and heat-colored backgrounds. Out-of-month days are dimmed; today is ring-highlighted.
- **Week view**: 24-hour timeline grid per day with canvas-rendered **epoch dots** — one dot per 5-second EEG embedding, color-coded by session. Each dot's Y position maps relaxation value within the session's band. Session bar fallback renders while timeseries loads. Day label sidebar shows weekday, date, and session count; click to navigate to day view. Empty days show a subtle placeholder.
- **Day view – epoch dot canvas**: new canvas timeline below the existing 24h bar renders all epoch dots for the day, with session color legend and label count summary. Label markers appear as amber triangles with text.
- **Always-rendered UI**: the calendar grid renders immediately, even while data is loading or when there are no recordings yet. A skeleton pulse animation shows in the day-label slot during loading; empty state shows a gentle hint with the clock icon and guidance text.
- **Reworked titlebar UX**: themed `history-viewmode-seg` segmented control (accent-aware via `--color-primary`); context-sensitive navigation — day mode shows prev/next arrows with position counter, calendar modes show period-appropriate prev/next with a formatted label (e.g. "March 2026"); skeleton loading animation for day labels.
- **i18n**: all new strings (`history.view.year/month/week/day`, `history.heatmap.less/more/none/hours/dayStreak`, `history.session`) translated across English, German, French, Hebrew, and Ukrainian.

### LLM — Hardware Fit Prediction

- **`get_model_hardware_fit` Tauri command**: uses `llmfit-core`'s `ModelFit::analyze` against the user's detected hardware (`SystemSpecs::detect()`, cached via `OnceLock`) to predict whether each catalog model will run. Returns fit level, run mode, estimated memory, tok/s, and composite score.
- **Per-quant fit badges** in the LLM settings tab: 🟢 Runs great / 🟡 Runs well / 🟠 Tight fit / 🔴 Won't fit. Hover tooltip shows run mode, memory breakdown, and estimated speed.
- **Hardware fit detail row** below each model entry showing run mode, memory utilization, estimated tok/s, and score.
- **Hardware summary line** above the family dropdown showing available memory from the detected GPU/RAM pool.
- **i18n**: `llm.fit.*` keys translated across all 5 locales.

### UI

- **Refactored `CustomTitleBar.svelte`**: collapsed the macOS and Windows/Linux duplicate branches into shared Svelte 5 snippets (`windowControls`, `centerContent`, `actionButtons`, `historyHead`, `tbBtn`, plus reusable icon snippets). Single unified template switches element order based on platform. Eliminated all duplicated HTML and CSS blocks. **975 → 533 lines (45% reduction).**
- Main-window titlebar now tints red when Bluetooth is unavailable (`bt_off` state), giving an immediate visual cue that the BLE adapter is off or missing. Uses the semantic `--color-error` token so the tint respects both light and dark themes.

### Chat — Tool Calling (pi-mono architecture)

- Implemented pi-mono style tool calling architecture with structured lifecycle events, argument validation, and configurable execution modes.
- Added **JSON Schema argument validation** for tool calls using the `jsonschema` crate — tool arguments are now validated against the tool's JSON Schema `parameters` definition before execution, with detailed error messages on validation failure (modelled after pi-mono's `validateToolArguments` with AJV).
- Added **configurable tool execution mode**: `parallel` (prepare sequentially, execute concurrently — default) and `sequential` (execute one-by-one in order). Persisted in `settings.json` under `llm.tools.execution_mode`.
- Added **configurable max tool rounds** (`max_rounds`, default 3) and **max tool calls per round** (`max_calls_per_round`, default 4) — both persisted in settings.
- Added **rich tool-execution lifecycle events** via IPC: `ToolExecutionStart` (with tool_call_id, tool_name, validated args) and `ToolExecutionEnd` (with result JSON and is_error flag), alongside the legacy `ToolUse` status events for backwards compatibility.
- Added **`BeforeToolCallFn` / `AfterToolCallFn` hook type definitions** for future extensibility — allows blocking tool execution or overriding results programmatically (modelled after pi-mono's `beforeToolCall`/`afterToolCall` hooks).
- Added execution mode toggle UI in both the Chat window settings panel and Settings → LLM tools section.
- Fully localised new strings in all five languages (EN, DE, FR, UK, HE).
- Added 4 new Rust unit tests for argument validation (valid args, missing required, no schema, wrong type) — all 15 tool tests pass.

### Chat — Tool Calling

- Added tool calling support to the LLM chat window with four built-in tools: **Date & Time**, **Location** (IP geolocation via ipwho.is), **Web Search** (DuckDuckGo Instant Answer API), and **Web Fetch** (fetch & read web pages).
- Added per-tool enable/disable toggles in the chat settings panel — persisted via `settings.json` under `llm.tools`.
- Added live tool-use indicators on assistant messages (calling → done/error) via a new `ToolUse` IPC chunk variant.
- Added a **Tools** badge in the chat header showing the number of enabled tools.
- Tool toggles are only shown when the model is running (`supports_tools` flag from the server status).
- Fully localised in all five languages (EN, DE, FR, UK, HE).

### Bugfixes

- Fixed quit confirmation dialog never receiving focus — set the parent window on the `rfd::MessageDialog` so the popup appears focused and modal on Linux/Windows instead of opening behind the main window.

- Fixed malformed thought traces that began with an unmatched opening `json` code fence and partial JSON fragments, which caused the rest of the thought bubble markdown to render incorrectly. The shared `normalizeMarkdown()` helper now strips that narrow orphaned preamble while preserving legitimate closed fenced code blocks.
- Fixed another chat tool-call transcript leak in the frontend parser. `stripToolCallFences()` now mirrors the Rust-side tool-call prefix heuristic instead of relying on narrow fence regexes, so incomplete or malformed fenced JSON blocks with blank lines or partial bodies are suppressed before they can appear in the lead-in bubble.
- Hardened chat markdown normalization for malformed model output. Emphasis repair now runs through a shared `normalizeMarkdown()` utility that protects fenced code blocks and inline code spans, trims stray spaces inside `*`/`**` delimiters, and falls back to raw `<strong>`/`<em>` tags when CommonMark flanking rules would still reject the emphasis. Added unit coverage for the repaired cases.
- Fixed expanded thought panels rendering raw markdown while final answers rendered correctly. The thought body now uses the shared `MarkdownRenderer`, so the same markdown normalization and parsing logic applies in both places. Added a muted renderer variant to preserve the thought-panel visual treatment.
- Fixed bold/italic not rendering when models emit `**word **` (space before closing delimiter) or `**Label:**value` (closing `**` preceded by punctuation followed by a non-whitespace character — CommonMark non-right-flanking edge case). Extended `normalizeMd()` in `MarkdownRenderer` with a trailing-space strip pass and a targeted conversion of punctuation-adjacent patterns to raw `<strong>`/`<em>` HTML so they always render as bold/italic regardless of CommonMark delimiter rules.

- Reworked chat-window assistant turn parsing/rendering so one streamed assistant turn can display as separate bubbles for lead-in text, tool activity, collapsed thinking, and final response, instead of merging tool chatter, `<think>` content, and the user-facing answer into one Markdown bubble.
- Fixed bold and italic text not rendering in the final answer bubble when a model emits `** Word:**` (space inside the `**` delimiters).  Added a `normalizeMd()` pre-pass in `MarkdownRenderer` that strips stray leading/trailing spaces inside `**…**` and `*…*` delimiters before handing the string to `marked`.
- Fixed multi-tool calls emitted as a single dict object: `{"date": {}, "location": {}}`.  Models like Qwen3 batch all tool calls as one JSON object where each key is a tool name and each value is the parameter object.  The previous extractor only recognised the OpenAI `{"name":"...","parameters":{}}` and `tool_calls:[...]` shapes.  Fix: added `KNOWN_TOOL_NAMES` constant (`date`, `location`, `web_search`, `web_fetch`), `is_dict_style_multi_tool()` helper, updated `extract_calls_from_value` (iterates over dict entries as calls), `is_tool_call_value` and `looks_like_tool_call_json_prefix` (early-exit when a known tool name appears as a JSON key), and frontend `stripToolCallFences` (same dict-style heuristic).  11/11 Rust unit tests pass.
- Fixed multi-tool and multi-round tool-calling rendering.  Root causes: (1) text emitted by the LLM before a tool call (e.g. "I'll use the date tool") appeared in the response bubble while the tool was running, then snapped to the lead-in position once the next inference round began; (2) with non-thinking models, consecutive rounds' text concatenated into one blob.  Fix: on the first `tool_use "calling"` event per round, the current `rawAcc` is parsed and frozen into `frozenLeadIn`/`frozenThinking`, `rawAcc` is reset to empty, and all subsequent delta/done/error handlers merge the frozen state back via `mergeWithFrozen()`.  Tools from multiple rounds accumulate correctly in the `toolUses` array.
- Replaced the blinking text cursor shown during LLM inference with a spinning SVG arc on the avatar column; the "AI" avatar is restored once generation completes.
- Reordered assistant turn sub-bubbles into strict chronological sequence: *thinking* (collapsed) → *lead-in text* → *tool-use indicators* → *response*.
- Fixed partial tool-call JSON fences and literal `</think>` tags appearing in the chat response bubble for tool-calling models (e.g. Qwen3).  Root causes: (1) the stream sanitizer emitted partial fence text before accumulating enough tokens to recognise the fence as a tool call; (2) tool-calling turns emit two separate `<think>` blocks (pre-tool and post-tool) which the single-pair extractor left unstripped in `content`.  Fix: added `stripToolCallFences()` to the frontend that removes both complete and incomplete fenced tool-call blocks, and rewrote `parseAssistantOutput()` to collect all `<think>…</think>` pairs across a multi-turn response, merging them into a single thinking block while routing the final segment to the answer bubble.
- Fixed chat message formatting when a model starts a JSON tool-call code fence and never closes it before `<think>`: the streaming sanitizer now suppresses incomplete trailing tool-call fences/JSON early enough that Markdown never treats the rest of the assistant reply as one giant code block.
- Fixed chat thinking-panel separation when `<think>` appears after other assistant text (for example after tool-call lead-in text): the chat UI now extracts think blocks from anywhere in the assistant message instead of only when `<think>` is the first visible token.
- Fixed chat tool-calling transcript leakage: assistant JSON tool payloads emitted in OpenAI-style inline/fenced blocks are now stripped from visible streamed output/history (not only `[TOOL_CALL]...[/TOOL_CALL]` markers), so users no longer see raw call JSON before the final natural-language answer.
- Updated the built-in `date` tool response to include explicit local-time metadata (`iso_local`, timezone abbreviation/name, and UTC offset) plus `iso_utc`, so assistant time answers can reliably default to the user's local timezone instead of guessing from epoch values.
- Fixed a follow-up chat tool-calling parser gap where some models output `{"tool":"date","parameters":{}}` instead of `{"name":"date",...}`; extractor now treats `tool` as a valid alias for function name so built-in tool execution triggers for that payload shape too.
- Fixed in-app chat tool-calling compatibility when models emit OpenAI-style function-call JSON directly in assistant text (including `{"name":"date","parameters":{}}`, fenced `json` blocks, and `{"tool_calls":[...]}` envelopes) instead of llama.cpp `[TOOL_CALL]...[/TOOL_CALL]` markers; the Rust extractor now detects these payload shapes and executes built-in tools correctly.
- Reduced title/menu redraw churn by deduplicating unchanged window-title writes (`setTitle`) and skipping no-op titlebar title-observer state updates.
- Reduced spacing between titlebar close/maximize/minimize controls across all windows by matching shared `CustomTitleBar` window-control button width to the other titlebar icon buttons (`30px`).
- Fixed Tailwind v4 `Invalid declaration: onMount` dev-server errors across `CustomTitleBar.svelte`, `+page.svelte`, `GpuChart.svelte`, `DisclaimerFooter.svelte`, and others — `@tailwindcss/vite` v4.2's `enforce:"pre"` transform matched `.svelte?svelte&type=style&lang.css` virtual modules before the Svelte compiler had extracted the `<style>` block, causing the CSS parser to choke on JavaScript imports. Patched `vite.config.js` with a shim that skips all `.svelte` style virtual module IDs in Tailwind's transform plugins. Also removed empty `<style></style>` blocks in `whats-new/+page.svelte` and `UmapViewer3D.svelte`.
- Fixed mmproj crash when the vision projector file is missing on disk — added an `exists()` guard before calling `mtmd_init_from_file` (which can abort/segfault on some platforms instead of returning null); switched from `active_mmproj_path()` to `resolve_mmproj_path(autoload)` so auto-detection works properly; stale paths where the file has been deleted are now filtered out with a warning instead of passed to the C library.
- Fixed app crash after mmproj fails to load — the clip/vision GPU warmup in `MtmdContextParams` (enabled by default) could corrupt Vulkan GPU state when the mmproj file was incompatible with the text model, causing the subsequent text-model warmup decode to abort the process. Disabled the clip warmup at init time (deferred to the first real multimodal request); wired up `no_mmproj_gpu` and `mmproj_n_threads` settings that were defined in `settings.rs` but never passed to the native library; added a file-size sanity check (files < 1 KB are rejected as truncated downloads); wrapped `init_from_file` in `catch_unwind` so a native panic cannot take down the application; improved error messages to include the file path and size for easier diagnostics.
- Fixed Linux mmproj startup crashes caused by unstable mtmd/Vulkan projector initialization paths on some driver stacks: mmproj now defaults to CPU projector init on Linux (while preserving normal text-model GPU offload), and advanced users can explicitly re-enable mmproj GPU init with `SKILL_FORCE_MMPROJ_GPU=1`.
- Fixed stale `mmproj` fallback selection on startup: when the active text model belongs to a known catalog repo, startup now rejects projector paths from a different repo (for example, a 27B projector with a 4B model), logs a clear mismatch warning, and continues in text-only mode without calling mtmd on the incompatible file.
- Fixed Linux WebKit startup abort on Wayland caused by `stacker::maybe_grow` swapping the main-thread stack before JavaScriptCoreGTK initialised. Linux now raises `RLIMIT_STACK` to 64 MiB and runs Tauri on the original main-thread stack; macOS and Windows keep the existing `stacker` path.
- Fixed Linux app auto-close after startup caused by implicit `RunEvent::ExitRequested` handling: implicit exits are now prevented consistently, main window is hidden instead, and only explicit quit paths run full shutdown.
- Fixed intermittent `npm run tauri dev` startup failure on Linux (`scripts/build-espeak-static.sh` exit `141`): replaced a SIGPIPE-prone `ar -t ... | head -1` cache-check pipeline (with `set -o pipefail`) by a safe `mapfile`-based first-object read, preventing false build-script aborts on valid archives.

### LLM Catalog

- Added Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled model family (`eugenehp/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF`) to the LLM catalog with 17 quant variants (Q2_K through BF16/F16).
- Added OmniCoder 9B model family (`Tesslate/OmniCoder-9B-GGUF`) to the LLM catalog with 13 quant variants (Q2_K through BF16) — a coding-focused 9B model.

## [0.0.36] — 2026-03-12
### CI Runtime

- Fixed macOS updater 404: renamed the macOS updater tarball from `NeuroSkill™.app.tar.gz` (non-ASCII URL) to `NeuroSkill_<version>_aarch64.app.tar.gz` so the URL stored in `latest.json` is pure ASCII and resolves correctly in the Tauri updater HTTP client.

## [0.0.35] — 2026-03-12
### CI Runtime

- Fixed cross-platform `latest.json` merge encoding in release workflows: Windows now writes `latest.json` as UTF-8 without BOM, and Linux/macOS loaders read with `utf-8-sig` to tolerate BOM-prefixed manifests and avoid `JSONDecodeError: Unexpected UTF-8 BOM`.

## [0.0.34] — 2026-03-12

### CI Runtime

- Fixed Windows release CI `Update latest.json` step crashing with "The property 'windows-x86_64' cannot be found" when `latest.json` already exists: `ConvertFrom-Json` returns a `PSCustomObject` whose properties cannot be set by dot-notation for new hyphenated names; the workflow now uses bracket-notation for hashtable/ordered-dict platforms and `Add-Member -Force` for PSCustomObject platforms.

## [0.0.33] — 2026-03-12
### CI Runtime

- Fixed Windows release CI PowerShell parser failures in `.github/workflows/release-windows.yml` by switching `latest.json` fallback `notes` text to ASCII-safe content and removing backtick-escaped tag/version string literals in the Discord webhook payload fields.

## [0.0.32] — 2026-03-12

### CI Runtime

- Windows release workflow now auto-detects NSIS artifacts across both valid output layouts (`src-tauri/target/x86_64-pc-windows-msvc/release/bundle/nsis` and `src-tauri/target/release/bundle/nsis`) in the build/sign step and artifact collection step, preventing false "bundle dir not found" / "installer not found" failures when Rust emits host-layout release outputs.

## [0.0.31] — 2026-03-12

### CI Runtime

- Release CI contributor attribution now comes only from git commit authors in the tagged release range (`previous_tag..current_tag`), and release workflows no longer append GitHub auto-generated release-note contributors.
- Windows release CI fallback for post-compile Tauri crash: `.github/workflows/release-windows.yml` now detects `npx tauri build` failure after the Rust binary is already produced (for example exit `-1073741571`/stack-overflow path), then recovers by running `npx tauri bundle --bundles nsis --no-sign`, recreating the updater zip (`*.nsis.zip`) from the generated installer, and signing that updater artifact via `npx tauri signer sign` so release publishing can continue.
- Windows release CI now uses the same primary packaging path that works locally (`npm run tauri:build:win:nsis`) instead of direct `npx tauri build` bundling, then signs the generated installer (when a cert is present) and creates/signs updater artifacts (`*.nsis.zip` + `*.nsis.zip.sig`) in workflow.
- Added npm script alias `taur:build:win:nsis` and switched Windows release CI to run `npm run taur:build:win:nsis` exactly.
- Windows release CI now installs NSIS explicitly before packaging (`choco install nsis` when `makensis` is missing), validates `makensis.exe` discovery, and exports `NSIS_DIR`/PATH so `scripts/create-windows-nsis.ps1` runs reliably on `windows-latest` runners.

## [0.0.30] — 2026-03-12

### Build / Tooling

- **Fix macOS release CI Pillow install**: added `--break-system-packages` to the `pip3 install Pillow` command in `.github/workflows/release-mac.yml` to resolve PEP 668 externally-managed-environment error on the `macos-26` runner.

## [0.0.29] — 2026-03-12

### Refactor

- macOS titlebar button order: switched the close and minimize button positions in the shared custom titlebar component so all macOS windows now use the requested control order.

- **macOS quit-time Metal teardown ordering**: added a one-time blocking shutdown helper in `src-tauri/src/lib.rs` and invoke it on explicit `RunEvent::ExitRequested` (`code = Some(_)`) before process exit continues. This now tears down LLM actor state and TTS backends earlier than `RunEvent::Exit`, reducing late `ggml-metal` static-destruction assertions (`GGML_ASSERT([rsets->data count] == 0)`) on macOS quit.

- **macOS shutdown abort in Metal teardown (`GGML_ASSERT([rsets->data count] == 0)`)**: added a blocking `Shutdown` command to the `tts/kitten.rs` worker and wired it into `tts_shutdown()` so `RunEvent::Exit` now waits for KittenTTS resources to drop before process exit/static destructor cleanup. This prevents late Metal/ggml teardown asserts when quitting after KittenTTS and LLM were active.

- **Heap-allocate AppState**: changed `Mutex<AppState>` → `Mutex<Box<AppState>>` across all Rust source files (`lib.rs`, `tray.rs`, `shortcut_cmds.rs`, `muse_session.rs`, `ws_commands.rs`, `openbci_session.rs`, `active_window.rs`, `label_cmds.rs`, `session_csv.rs`, `session_analysis.rs`, `llm/cmds.rs`, `session_dsp.rs`, `ble_scanner.rs`, `window_cmds.rs`, `history_cmds.rs`, `settings_cmds.rs`, `api.rs`, `commands.rs`, `global_eeg_index.rs`) to move the large `AppState` struct onto the heap, reducing main-thread stack frame size and mitigating stack overflow risk on platforms with smaller default stacks.
- **Extract LLM state into `Box<LlmState>`**: moved all LLM-related fields (`llm_config`, `llm_catalog`, `llm_downloads`, `llm_logs`, `llm_state_cell`, `llm_loading`, `llm_start_error`, `chat_store`) out of `AppState` into a dedicated `LlmState` sub-struct stored as `Box<LlmState>`, accessed via `s.llm.config`, `s.llm.catalog`, etc.  This further reduces `AppState`'s on-stack footprint and groups all LLM concerns behind a single heap-allocated pointer.
- **Construct AppState on a dedicated thread**: added `AppState::new_boxed()` that spawns a 32 MiB-stack thread to run `Box::new(AppState::default())`, avoiding the main-thread stack overflow that occurred on macOS when the large struct + `generate_handler!` frame exceeded the default stack limit.
- **Add macOS/Linux 32 MB main-thread stack size**: emit `-Wl,-stack_size,0x2000000` (macOS) and `-Wl,-z,stacksize=33554432` (Linux) via `cargo:rustc-link-arg-bins` in `build.rs`.  Using `rustc-link-arg-bins` instead of target-wide `rustflags` in `.cargo/config.toml` ensures the flag applies only to the final executable — ld64 rejects `-stack_size` when linking dylibs/cdylibs (the lib crate), which caused `ld: -stack_size option can only be used when linking a main executable`.
- **Extract `setup_app()` / `setup_background_tasks()`**: moved the ~650-line `.setup()` closure body and the updater/DND poll loops into separate `#[inline(never)]` top-level functions so LLVM cannot merge their stack frames with the already-huge `run()` frame produced by `generate_handler!` with ~150 commands.
- **Dynamic stack growth via `stacker`**: added `stacker = "0.1"` dependency and wrapped the `skill_lib::run()` call in `main()` with `stacker::maybe_grow(32 MiB, 64 MiB, ...)`.  This dynamically extends the main-thread stack using `mmap` + inline-asm stack-pointer swap (via `psm`) without changing the thread identity, which is required on macOS where Cocoa/AppKit mandates the event loop runs on the original main thread.  Linker flags (`-Wl,-stack_size`) were unreliable because macOS ld64 rejects them on dylibs and Tauri's mixed `crate-type = ["staticlib", "cdylib", "rlib"]` build triggers both lib and bin linking.

### Build / Tooling

- **Release notes now include changelog section**: all tagged release workflows (`.github/workflows/release-linux.yml`, `.github/workflows/release-mac.yml`, `.github/workflows/release-windows.yml`) now extract the matching `## [x.y.z]` block from `CHANGELOG.md` and pass it to `softprops/action-gh-release` via `body_path`, so GitHub Release information includes the version-specific changelog alongside generated release notes.

- **Preview artifacts now include changelog notes**: `.github/workflows/pr-build.yml` now generates `preview-notes.md` from the matching `CHANGELOG.md` `## [x.y.z]` section (based on `tauri.conf.json` version) and uploads it with the preview DMG/updater artifacts so pre-release testers get version-specific notes with each build.

- **Windows NSIS discovery false-negative fix**: corrected `scripts/create-windows-nsis.ps1` NSIS lookup when `makensis` is already on PATH. The script previously used `Split-Path` twice on `Get-Command makensis`, which resolved to the parent of the NSIS directory and could incorrectly fail `makensis.exe` checks. It now uses the direct parent folder and also accepts `NSIS_DIR` set to either the NSIS directory or a full `makensis.exe` path.

- **Windows NSIS PowerShell argument parsing fix**: corrected `scripts/create-windows-nsis.ps1` candidate-path construction to precompute `$TargetBinary`/`$HostBinary` and then build `$BinaryCandidates` from variables. This avoids a PowerShell parse/invocation edge case where comma-separated inline `Join-Path` calls inside `@(...)` could be interpreted as an array passed to `-ChildPath`, causing `Cannot convert 'System.Object[]' to the type 'System.String'`.

- **Windows NSIS standalone packaging path fallback**: `scripts/create-windows-nsis.ps1` now auto-detects the prebuilt release binary from either `src-tauri/target/x86_64-pc-windows-msvc/release/skill.exe` (explicit target build) or `src-tauri/target/release/skill.exe` (default host-target build). This fixes `npm run tauri:build:win:nsis` failing after a successful `tauri build --no-bundle` when Rust outputs the host-layout path; the script now also places NSIS output under the detected release directory's `bundle/nsis` folder.

- **macOS `.app` manual assembly fallback**: when the Tauri CLI bundler process itself stack-overflows (exit 134 SIGABRT or 139 SIGSEGV) during the `--bundles app` phase — which is a Tauri CLI issue, not the app binary — `scripts/tauri-build.js` now detects the crash, verifies the release binary was already built, and assembles the `.app` bundle manually using `ditto`, `codesign --force --deep --sign -`, the project's `Info.plist`, icons, entitlements, and resources from `tauri.conf.json`.  This makes `npm run tauri:build:mac -- --bundles app` reliable even when the Tauri CLI has stack issues.
- **Standalone macOS `.app` assembler**: added `scripts/assemble-macos-app.sh` that builds the `.app` directory structure from a pre-built release binary without invoking the Tauri CLI bundler at all.  New npm script `npm run tauri:build:mac:app` compiles with `--no-bundle` then runs the assembler.  Copies binary, merges `Info.plist` with required `CFBundle*` keys (including `CFBundleIconFile` and `NSHighResolutionCapable`), copies `icon.icns` + resources via `ditto`, and ad-hoc codesigns.
- **macOS DMG creator**: replaced the custom 800-line `scripts/create-macos-dmg.sh` with a single-pass [`appdmg`](https://github.com/LinusU/node-appdmg) approach.  Generates a branded background image (app icon + product name + version, 660×520 @1x + @2x Retina, dark/light mode adaptive) and a version-badged volume icon (`.icns` with "v0.0.28" pill overlay) via Pillow, then calls `appdmg` with a full spec: app + Applications symlink (top row), README.md + LICENSE + CHANGELOG.md (bottom row), icon positions, window size, ULFO+APFS format.  `appdmg` handles Finder view setup via AppleScript in one pass — no `hdiutil convert` round-trips that corrupt APFS volumes, no Python `ds_store`/`mac_alias` that crash Finder, no `hdiutil udifrez` SLA that corrupts DMGs on macOS 14+.  Both `release-mac.yml` and `pr-build.yml` CI workflows install `appdmg` + Pillow and use the shared script.  Also fixed `assemble-macos-app.sh` Info.plist generation: replaced regex string injection with `plistlib` so `NSHighResolutionCapable` is a proper boolean `<true/>` (not `<string>true</string>`) and added `CFBundleName`, `CFBundleDisplayName`, `LSMinimumSystemVersion`, `NSRequiresAquaSystemAppearance` keys.
- **Windows NSIS installer script**: added `scripts/create-windows-nsis.ps1` and `npm run tauri:build:win:nsis` for standalone Windows NSIS installer creation that bypasses the Tauri CLI bundler.  Generates branded installer images (header 150×57, welcome panel 164×314 with app icon + version via Pillow), bundles `README.md`, `CHANGELOG.md`, `LICENSE`, resources (espeak-ng-data, neutts-samples), creates Start Menu + Desktop shortcuts, registers in Add/Remove Programs, and optionally signs with `signtool.exe` via `CERTIFICATE_THUMBPRINT`.  The GPL-3.0 `LICENSE` is shown as a license agreement page during installation.

## [0.0.27]

### Bug Fixes

- **Feature-gated compilation for `--no-default-features` builds**: added `#[cfg(feature = "llm")]` / `#[cfg(not(feature = "llm"))]` guards in `lib.rs` (stub `llm_state_cell` field + Default impl), `api.rs` (conditional `Mutex` import), `settings_cmds.rs` (split `set_llm_config` into feature-gated paths), `tray.rs` (`ellipsize_middle` helper), and `llm/mod.rs` (`allowed_tools` field); added `#[cfg(any(feature = "tts-kitten", feature = "tts-neutts"))]` guards in `tts/mod.rs` for imports, helpers, constants, and `impl` blocks so the crate compiles cleanly when built with `--no-default-features` or any subset of feature flags.

### Dependencies

- **Auto-enable GPU backend for LLM by platform**: `llama-cpp-4` now receives `metal` on macOS and `vulkan` on Linux/Windows via target-specific dependency feature merging in `Cargo.toml`, so the `llm` feature automatically uses the correct GPU backend without requiring manual `--features llm-metal` / `--features llm-vulkan` flags.
- **Bump `llama-cpp-4` from 0.2.9 → 0.2.10** (and `llama-cpp-sys-4` 0.2.9 → 0.2.10).

### Build / Tooling

- **Fix CI release binaries missing embedded frontend**: release and preview workflows that bypass the Tauri CLI (`cargo build --release` directly) were producing dev-mode binaries that attempted to load the UI from `localhost:1420` instead of serving the embedded SvelteKit build output.  Root cause: the Tauri crate gates frontend embedding behind its `custom-protocol` Cargo feature, which `npx tauri build` activates automatically but raw `cargo build` does not.  Added a `custom-protocol` feature to `src-tauri/Cargo.toml` forwarding to `tauri/custom-protocol` and pass `--features custom-protocol` in `release-linux.yml`, `release-mac.yml`, `pr-build.yml`, and `release-windows.ps1`.
- `npm run bump` now also rotates the changelog release header automatically: it preserves a fresh `## [Unreleased]` section and inserts `## [x.y.z] — YYYY-MM-DD` for the newly bumped version.
- macOS local Tauri build stability: `scripts/tauri-build.js` now injects `--no-bundle` by default for `build` runs (while still forcing `--target aarch64-apple-darwin --no-sign`), unless the caller explicitly passes `--bundle`/`--bundles`/`--no-bundle`; this avoids post-compile bundle-phase crashes where `npx tauri build --target aarch64-apple-darwin --no-sign` fails but `--no-bundle` succeeds.

### Features

- **Hooks lifecycle completeness pass**: hook triggers now surface full runtime context (last-trigger time, matched label, and one-click session open), emit both in-app toast + native OS notification payloads, and persist immutable trigger snapshots into dedicated `hooks.sqlite` JSON audit rows; the trigger path runs in the background embedding worker with panic isolation and dedicated `hooks` logger toggles, while docs/tests/examples/locales were updated together (`SKILL.md`, `cli.ts`, `test.ts`, Help/FAQ + flow diagram, and `en`/`de`/`fr`/`he`/`uk` translations).
- **Proactive Hooks rename + scenarios**: renamed user-facing Hooks copy from “Automation Hooks” to **Proactive Hooks** and added per-hook scenario modes (`any`, `cognitive`, `emotional`, `physical`) so triggers can be gated by live state metrics in the background worker.
- **Hooks keyword picker keyboard UX**: keyword suggestions now support keyboard navigation (`↑` / `↓` / `Enter` / `Esc`) in addition to click-to-apply.
- **Hooks quick examples**: added one-click starter scenarios (cognitive deep-work guard, emotional calm recovery, physical body-break) to speed up hook creation.
- **Hooks keyword suggestions while typing**: Settings → Hooks now shows live keyword suggestions in the add-keyword flow by combining fuzzy matches from `labels.sqlite` with semantic nearest-label hits from the label text HNSW index; suggestion chips include source tags (`fuzzy`, `semantic`, or `fuzzy+semantic`) and can be clicked to add quickly.
- **Hooks button text-fit polish**: small action buttons in Hooks now use wrap-safe sizing (`h-auto` + multiline text) so localized labels fit without clipping.
- **Hooks scenario dropdown theming polish**: scenario selector now uses themed custom select styling (`appearance-none`, semantic border/ring tokens, custom chevron) for consistent dark/light appearance.
- **Hooks heading naming tweak**: Hooks tab card heading now uses the concise localized tab label ("Hooks") instead of longer variant text.
- **Settings sidebar resize**: Settings tab navigation sidebar is now mouse-resizable with a drag handle, bounded min/max width, and persisted width between opens.
- **Settings titlebar clarity**: settings window title now always includes localized “Settings” plus the active tab name (for example “Settings — Hooks”).

- **Hook distance suggestion**: new "Suggest threshold" button in Settings → Hooks that analyses real HNSW and SQLite data — finds labels matching the hook's keywords, computes cosine-distance distribution of recent EEG embeddings against those label references, and presents a percentile bar (min/p25/p50/p75/max) with a one-click "Apply" action to set the recommended threshold.
- **Hooks WS/CLI observability expansion**: added websocket commands `hooks_suggest` and `hooks_log`, plus CLI subcommands `hooks suggest` and `hooks log` (limit/offset pagination) for scriptable threshold recommendations and audit-log inspection over either WebSocket or HTTP tunnel transport.
- **Hook fire history viewer**: expandable "Hook fire history" section in Settings → Hooks with paginated (20/page) collapsible event rows showing timestamp, label, distance, command, and threshold-at-fire metadata.
- **Last-trigger relative age**: the last-trigger display in Settings → Hooks now shows a live relative-time label (e.g. "12s ago", "3m ago") that updates every second alongside the absolute timestamp.
- Added a new **Settings → Hooks** tab for user-defined automation hooks: each hook supports name, enabled flag, multiple keywords, command payload, custom text payload, configurable EEG distance threshold, and configurable recent-reference count (clamped to 10–20).
- Added backend hook persistence and runtime matching pipeline: hook rules are saved in `settings.json`, hook keyword queries use fuzzy matching plus text-embedding/HNSW nearest-label expansion, then map to recent label-window EEG references; incoming EEG embeddings now trigger websocket broadcasts when close enough, with payload `{ hook, context: "labels", command, text }`.

### Documentation

- **Proactive Hooks docs/examples refresh**: updated `SKILL.md` hooks scenarios and jq examples, refreshed CLI help/output text in `cli.ts` to include scenario metadata, and extended `test.ts` hook status smoke checks to validate `hook.scenario` when hooks exist.
- Added hooks explainers in Help/FAQ including a compact hook flow diagram and a dedicated trigger-mechanics FAQ entry.

### Bug Fixes

- **Single-instance runtime enforcement**: app startup now initializes `tauri-plugin-single-instance`, so opening NeuroSkill while it is already running no longer starts a second process; the existing `main` window is restored/focused instead.

- **Windows CI Rust warning cleanup (`dead_code`)**: removed the non-Linux `linux_has_appindicator_runtime()` stub from `src-tauri/src/lib.rs` so only the Linux implementation is compiled; this eliminates the Windows-only `function is never used` warning while preserving the Linux tray-runtime guard behavior.

### Documentation

- **README Linux packaging quickstart added**: added a concise Development-section command block in `README.md` for Linux release-style local packaging (`tauri:build:linux:x64:native` for AppImage, then `package:linux:system:x64:native -- --skip-build` for manual `.deb`/`.rpm`), including an explicit `ALLOW_LINUX_CROSS=1` cross-target example.
- **Linux setup docs now include tray runtime dependency guidance**: updated `LINUX.md` with a dedicated runtime prerequisite for `tauri dev` (`libayatana-appindicator3-1`, with `libappindicator3-1` fallback) and added troubleshooting steps for the startup error `Failed to load ayatana-appindicator3 or appindicator3 dynamic library`.
- **Linux docs cross-link clarity pass**: added a reciprocal pointer in `LINUX.md` back to `README.md` Development prerequisites and explicit wording that missing appindicator runtime packages can break `npm run tauri dev` at startup.
- **Linux packaging command docs aligned with workflows**: updated the `LINUX.md` build section to recommend the canonical local flow (`npm run tauri:build:linux:x64:native` for AppImage, then `npm run package:linux:system:x64:native -- --skip-build` for `.deb`/`.rpm` via `dpkg-deb`/`rpmbuild`), with cross-target examples when `ALLOW_LINUX_CROSS=1` is intentional.

### Bug Fixes

- **Rust clippy warning cleanup (embeddings/settings)**: marked argument-heavy constructor/spawn entry points in `src-tauri/src/eeg_embeddings.rs` with targeted `#[allow(clippy::too_many_arguments)]` (matching the existing worker rationale), and replaced the manual `Default` implementation for `HookStatus` with `#[derive(Default)]` in `src-tauri/src/settings.rs`.
- **Rust hooks settings compile fix (`E0596`)**: fixed `set_hooks` in `src-tauri/src/settings_cmds.rs` by binding the locked app state as mutable before assigning `s.hooks`, resolving `cannot borrow 's' as mutable, as it is not declared as mutable` during `cargo clippy`/build.
- **Linux tray is now mandatory with fail-fast startup guard**: before tray initialization, startup probes for loadable appindicator shared objects; when `libayatana-appindicator3` / `libappindicator3` is missing, startup aborts immediately with a clear prerequisite error instead of panicking inside `libappindicator-sys` or running without tray.
- **Linux `tauri dev` tray-runtime preflight**: `scripts/tauri-build.js` now checks for a loadable appindicator runtime (`libayatana-appindicator3.so*` or `libappindicator3.so*`) before launching `npx tauri dev`; when missing, it exits early with distro-aware install guidance (`apt`/`dnf`/`pacman`/`zypper`) instead of letting the app crash at startup with a `libappindicator-sys` panic.
- **`npm run bump` Linux preflight dependency clarity**: added an explicit `pkg-config` guard before `cargo clippy` in `scripts/bump.js` that checks `webkit2gtk-4.1`, `javascriptcoregtk-4.1`, and `libsoup-3.0`; when missing, bump now fails fast with actionable `apt install` guidance instead of surfacing a lower-level `webkit2gtk-sys` build-script crash.
- **Strictest non-status accent normalization (UMAP/Embeddings)**: removed remaining category-only orange/sky/emerald/violet highlight styling in UMAP and Embeddings controls (preset chips, pipeline badges, slider thumb/focus affordance, and dimension legend badges) in favor of semantic `primary` / `ring` tokens so generic interactive emphasis consistently follows Appearance accent settings.
- **Strict accent policy completion for generic selectors**: updated the remaining non-status selected controls in Calibration profile editing (break-duration and iterations chips) to use semantic `primary` tokens instead of hardcoded `amber`/`emerald`, and clarified `AGENTS.md` guidance that semantic status colors remain allowed only for true status signaling.
- **Follow-up accent normalization for non-status highlights**: converted remaining generic hardcoded `rose`/`emerald` selection and focus styles (UMAP timeout/cooldown controls, EEG overlap selector summary badges, and interactive search query focus ring) to semantic `primary` / `ring` tokens, while leaving semantic success/warning/error colors unchanged.
- **Broader accent-token consistency sweep**: replaced numerous hardcoded interactive blue states (selected chips/buttons, focus rings, and status badges) with semantic `primary` / `ring` tokens across Appearance, Settings, Focus Timer, History, Labels, Calibration, API, Search, and related tabs so accent-like UI feedback consistently follows the Appearance accent mapping.
- **Accent setting now applies to native form controls and remaining interactive toggles**: added a global `accent-color` base rule tied to the remapped accent palette so checkboxes/radios/ranges/progress controls follow the selected Appearance accent, and replaced remaining hardcoded non-remapped accent classes in interactive Search/UMAP controls.
- **Updater fallback on install failure**: when automatic update download/install fails in the Updates tab, the UI now gives an explicit "download online" fallback and automatically opens the latest GitHub releases page (`https://github.com/NeuroSkill-com/skill/releases/latest`) so users can immediately fetch the newest installer manually.
- **macOS white screen on first launch**: `win.show()` was called in Tauri's `setup` closure before WKWebView had loaded any content, producing a solid white frame until the next compositor cycle.  Fixed by removing the eager `setup` show and adding a new `show_main_window` Tauri command that is invoked from `+layout.svelte` `onMount`; the window now becomes visible only after the page has fully rendered.  Secondary windows (settings, help, calibration, etc.) and the new-user onboarding flow are unaffected — `show_main_window` is a no-op for any window whose label isn't `"main"` or whose onboarding flag is unset.
- **What's New version picker theme mismatch**: the navigation dropdown in `/whats-new` used transparent/native select styling that could ignore app theme colors in the standalone window. The picker now uses explicit themed control styles (`appearance-none`, theme-aware background/border/text) plus a custom caret so light/dark appearance matches the rest of the UI.
- **Appearance accent color not applied consistently across UI**: accent selection previously remapped only Tailwind `violet-*` variables, while many controls and gradients used `blue-*`, `indigo-*`, or `sky-*` classes and stayed on default hues. Accent application now remaps those accent-like families together so interactive highlights, rings, sliders, and accent gradients consistently follow the selected Appearance accent.

### CI Runtime

- Windows release workflow reliability fix: `.github/workflows/release-windows.yml` now uses ASCII-safe Discord title strings in the notify step to avoid Windows PowerShell parser/encoding failures, and the Tauri build step now runs with `--verbose` plus bundle-directory diagnostics when `npx tauri build` exits non-zero (so packaging failures surface actionable logs instead of a bare exit code).
- Linux release workflow now bypasses Tauri bundling entirely (macOS-style): it compiles frontend + Rust only, builds `.deb`/`.rpm` via `scripts/package-linux-system-bundles.sh`, builds the portable Linux tarball via `scripts/package-linux-dist.sh`, signs those outputs with `tauri signer`, and publishes updater metadata from the signed portable tarball instead of AppImage bundle artifacts.
- CI Linux packaging scope reduced to tarball-only in `.github/workflows/ci.yml`: removed the `linux-release` job that produced `.deb`/`.rpm`/`.AppImage`, so Linux CI now only runs the portable package flow and publishes `.tar.gz` artifacts.
- Tauri frontend bundling contract guard: added `scripts/verify-tauri-frontend-structure.js` and wired it into `npm run build` (`package.json`) so `tauri build` (via `beforeBuildCommand`) now fails fast unless the configured `src-tauri/tauri.conf.json` `build.frontendDist` path contains valid built assets (`index.html` + `_app/immutable` JS/CSS) rather than raw source files.
- Linux/macOS/Windows bundling workflows now run an explicit `npm run -s verify:tauri:frontend` step before packaging (`.github/workflows/ci.yml`, `.github/workflows/release-linux.yml`, `.github/workflows/release-mac.yml`, `.github/workflows/release-windows.yml`) to enforce the same Tauri asset layout contract in CI.
- Windows release Discord notifier fix: `.github/workflows/release-windows.yml` now sends the Discord payload from a PowerShell object serialized via `ConvertTo-Json` (instead of shell-escaped inline JSON), eliminating Discord API `50109` (`The request body contains invalid JSON`) failures after successful Windows builds.
- Windows release post-build hardening: `.github/workflows/release-windows.yml` now updates `latest.json` with native PowerShell (no `python3` dependency in Git Bash on `windows-latest`) and skips the Discord notification step when `DISCORD_WEBHOOK_URL` is unset, avoiding non-build-related exit failures after successful Windows artifact compilation.
- macOS release bundle frontend integrity: `.github/workflows/release-mac.yml` now copies the generated SvelteKit `build/` output into `Contents/Resources/app` with `ditto` during manual `.app` assembly and fails fast if `build/index.html`, copied `index.html`, copied `_app/immutable`, or copied JS/CSS assets are missing, preventing release artifacts that omit frontend HTML/JS/CSS/static files.
- Linux CI + release packaging now avoids Tauri for `.deb`/`.rpm`: both `.github/workflows/ci.yml` and `.github/workflows/release-linux.yml` build only AppImage via `tauri-build.js --bundles appimage`, then run `scripts/package-linux-system-bundles.sh` to generate `.deb` with `dpkg-deb` and `.rpm` with `rpmbuild`; this removes Tauri Linux deb/rpm bundler segfaults from automated Linux build paths while keeping artifact outputs unchanged.
- Linux workflow/script consistency pass: `package.json` Linux Tauri scripts (`tauri:build:linux:arm64`, `tauri:build:linux:x64:native`, `tauri:build:linux:x64`) now target AppImage-only bundling, and both Linux workflows call the npm script entrypoint for the AppImage build before running manual system-tool `.deb`/`.rpm` packaging.
- Linux CI/release workflow hardening: added native Linux x86_64 npm scripts (`tauri:build:linux:x64:native`, `package:linux:portable:x64:native`) and switched `.github/workflows/ci.yml` + `.github/workflows/release-linux.yml` to those scripts so hosted x86_64 runners no longer depend on `ALLOW_LINUX_CROSS` cross-mode execution paths.
- Linux CI execution policy refinement: in `.github/workflows/ci.yml`, heavy Linux bundling jobs (`linux-release` and `linux-portable-package`) now run by default on `push`, and can be explicitly enabled for manual `workflow_dispatch` runs via `run_linux_bundles=true`, keeping pull-request CI focused on faster validation.
- Updated GitHub Actions workflows to Node 24-ready action versions across CI and release workflows: `actions/checkout` → `v6`, `actions/setup-node` → `v6`, `actions/cache` → `v5`, and `Swatinem/rust-cache` → `v2.9.0`, removing the GitHub deprecation warnings about Node 20-based actions.
- Removed the Linux Rust job's apt archive cache from `.github/workflows/ci.yml`; that cache was low-value on hosted runners and was the most likely source of the `/usr/bin/tar` post-job save failure that was making the Rust CI job noisy or red despite successful build steps.
- Reintroduced Linux Tauri system dependency caching in CI and Linux release workflows via `awalsh128/cache-apt-pkgs-action` (`.github/workflows/ci.yml`, `.github/workflows/release-linux.yml`) so WebKit/GTK build dependencies are restored from cache instead of re-downloaded on every run.

### UI / Type Safety

- **Settings window width bump**: increased the default Settings window width from `680` to `760` (height unchanged) so tabs and controls have more horizontal room; applied consistently to Settings/Model/Updates entry paths that create the shared `settings` window.

### What's New window

- **Full changelog navigation**: the What's New window now parses the entire bundled `CHANGELOG.md` (via Vite `?raw` import) into individual version sections and renders each one with `MarkdownRenderer`; a compact navigation bar between the header and body provides "Newer ←" / "Older →" arrow buttons and a version-picker `<select>` dropdown so users can browse every release entry from a single window; scroll position resets to the top on each navigation step; a `1 / N` counter in the footer shows the current position; new i18n keys (`whatsNew.older`, `whatsNew.newer`, `whatsNew.unreleased`) added to all five locales (en, de, fr, he, uk)


- Reduced the untyped `any` surface in the Three.js-heavy UI components by introducing explicit typed scene/object wrappers in `src/lib/UmapViewer3D.svelte` and `src/lib/InteractiveGraph3D.svelte`; removed broad `any` refs and `@ts-ignore`, and kept behavior unchanged while making future refactors compile-time safer.

### i18n (0.0.4)

- Localized updater fallback messaging across all shipped locales (`en`, `de`, `fr`, `he`, `uk`) by adding translated keys for: (1) automatic-update install failure with online download guidance, and (2) failure to auto-open the download page; `UpdatesTab.svelte` now uses i18n keys instead of hardcoded English strings for both paths.
- Fixed a locale key-sync detection edge case for `de`, `fr`, `he`, and `uk`: normalized `llm.tools.locationDesc`, `llm.tools.webSearchDesc`, and `llm.tools.webFetchDesc` entries to standard `"key": "value"` formatting so `scripts/sync-i18n.ts --check` correctly counts them
- Ran `scripts/sync-i18n.ts --fix` to auto-backfill 138 missing keys in `src/lib/i18n/he.ts` with English fallbacks, restoring locale key-count parity (`2237` keys) so `npm run sync:i18n:check` passes.
- Completed German fallback translation coverage in [src/lib/i18n/de.ts](src/lib/i18n/de.ts) for the auto-synced OpenBCI/LLM/chat/help/downloads blocks and removed stale in-file TODO translation markers in that locale.
- Completed French/Hebrew/Ukrainian fallback translation coverage in [src/lib/i18n/fr.ts](src/lib/i18n/fr.ts), [src/lib/i18n/he.ts](src/lib/i18n/he.ts), and [src/lib/i18n/uk.ts](src/lib/i18n/uk.ts) for the same auto-synced OpenBCI/LLM/chat/help/downloads blocks, and removed stale in-file TODO translation markers.
- Fixed French placeholder consistency regression in [src/lib/i18n/fr.ts](src/lib/i18n/fr.ts) by restoring `llm.size` interpolation token to `{gb}` so runtime formatting and placeholder-consistency tests align.

### Focus / DND

- Linux Do Not Disturb automation support: implemented real Linux backend behavior in `src-tauri/src/dnd.rs` instead of non-macOS no-ops, with GNOME integration via `gsettings org.gnome.desktop.notifications show-banners` and KDE integration via `qdbus(6)` `org.kde.osdService.setDoNotDisturb`; OS-state polling now reports Linux DND state when detectable
- Linux DND fallback path: when GNOME and KDE DND APIs are unavailable, the backend now falls back to `xdg-desktop-portal` inhibit requests (`gdbus` to `org.freedesktop.portal.Inhibit`) with tracked request-handle lifecycle so disable calls close previously created portal requests
- Windows Do Not Disturb automation support: implemented a Windows backend in `src-tauri/src/dnd.rs` using the per-user notification banner toggle (`HKCU\\Software\\Microsoft\\Windows\\CurrentVersion\\PushNotifications\\ToastEnabled`) for enable/disable and OS-state query so focus automation works on Windows as well

### Repo hygiene

- Cleaned editor hygiene warnings by switching release workflows away from fragile cross-step `${{ env.* }}` references in expression contexts, and by normalizing historical changelog markdown structure so repo diagnostics stay quiet.
- Fixed a Rust docs lint warning (`clippy::doc_lazy_continuation`) in [src-tauri/src/dnd.rs](src-tauri/src/dnd.rs) by splitting the Linux bullet list and the Windows support note into separate rustdoc paragraphs.

### Build / CI (Unreleased)

- macOS `aarch64-apple-darwin` Tauri build fix: moved `MACOSX_DEPLOYMENT_TARGET` and `CMAKE_OSX_DEPLOYMENT_TARGET` into top-level Cargo `[env]` scope in `src-tauri/.cargo/config.toml` (they were accidentally nested under `[target.i686-pc-windows-gnu.env]`), so `llama-cpp-sys` now receives a 10.15 deployment target and avoids `std::filesystem` availability errors (`'path' is unavailable: introduced in macOS 10.15`) during CMake/C++ compilation.
- `npm run bump` now runs mandatory preflight gates before mutating versions: `npm run check`, `cargo clippy --manifest-path src-tauri/Cargo.toml`, then `npm run sync:i18n:check`; if any step fails, bump exits immediately and does not update version fields.
- Linux CI bundle stability: `scripts/tauri-build.js` now detects a Tauri CLI segfault (`exit 139`) during explicit multi-target bundle runs (for example `--bundles deb,appimage`) and automatically retries bundling sequentially per target so release jobs can still produce both `.deb` and `.AppImage` artifacts
- Linux CI single-target bundle stability: when an explicit Linux bundle run (for example `--bundles deb`) exits with `139`, `scripts/tauri-build.js` now verifies the expected bundle output for that target and treats the run as successful only if artifacts are present; the same artifact-aware tolerance is also applied per-target during sequential retry after a multi-target segfault.
- Linux CI per-target recovery hardened: when a Linux `tauri build --bundles <target>` run exits `139` before writing bundle artifacts, `scripts/tauri-build.js` now retries that target with `tauri bundle --bundles <target>` and only fails if expected artifacts are still missing after the fallback path.
- Linux CI release-bundle smoke test now fails if no `.deb` package is produced: `.github/workflows/ci.yml` verifies at least one `.deb` exists after bundling and checks both the explicit target-triple bundle path and fallback non-target path to catch segfault-recovery path regressions.
- Linux ARM64 build fallback (macOS-style crash isolation): for explicit bundle builds where Tauri crashes with `139`/`134` but the release binary already exists, `scripts/tauri-build.js` now exits successfully in compile-only mode and prints guidance; set `DISABLE_LINUX_CRASH_COMPILE_FALLBACK=1` to force hard failure.
- Added standalone Linux distribution packaging script `scripts/package-linux-dist.sh` to avoid Tauri bundling: it builds with `--no-bundle`, assembles `NeuroSkill/` (binary, bundled resources, launcher, icon, desktop entry, docs), and emits a portable `tar.gz` archive under `dist/linux/<target>/`.
- Added CI portable-package job in `.github/workflows/ci.yml`: `linux-portable-package` now runs `npm run package:linux:portable:x64`, verifies the generated `dist/linux/x86_64-unknown-linux-gnu/*.tar.gz`, and uploads it as a GitHub Actions artifact (`linux-portable-x86_64`).
- Added Linux `.deb` artifact upload in CI: the `linux-release` job in `.github/workflows/ci.yml` now resolves the generated package from the target/fallback bundle paths and uploads it as `linux-deb-x86_64` for direct download from Actions runs.
- Linux package matrix expanded to include `rpm`: Linux build scripts now request `--bundles deb,appimage,rpm`, and both CI/release workflows were updated to validate and publish `.rpm` alongside `.deb` and `.AppImage` artifacts.
- Added Linux integrity sidecars: workflows now generate `SHA256SUMS` files for Linux bundle outputs and portable tarball outputs, and `release-linux` now also generates detached `.sig` signatures for Linux release artifacts.
- Linux release stale-artifact guard: `.github/workflows/release-linux.yml` now removes cached `src-tauri/target/x86_64-unknown-linux-gnu/release/{bundle,skill}` and `dist/linux/x86_64-unknown-linux-gnu` before compile/package steps so rust-cache leftovers cannot be mistaken for fresh artifacts when assembling release outputs.
- Linux CI parity stale-artifact guard: `.github/workflows/ci.yml` now performs the same pre-build cleanup in the `linux-portable-package` job, clearing cached `target`/`dist` Linux output paths before packaging so uploaded CI tarballs always come from the current run.

## [0.0.24] — 2026-03-12

### UI

- Label window titlebar spacing + vertical fit: moved the add-label window title back to the side, rendered the EEG timer as a padded centered capsule in the shared titlebar, and changed `/label` from `h-screen` to `h-full min-h-0` so the bottom action row no longer gets clipped under the custom titlebar layout
- What's New window vertical fit fix: changed `/whats-new` root container from `h-screen` to `h-full min-h-0` and marked the changelog body as `min-h-0` so the shared custom titlebar no longer pushes the footer off-screen and the bottom `Got it` button remains visible
- Window vertical-fit sweep: switched the remaining titlebar-hosted route roots (`/`, `/about`, `/api`, `/calibration`, `/chat`, `/compare`, `/downloads`, `/focus-timer`, `/help`, `/history`, `/labels`, `/onboarding`, `/session`, `/settings`) from viewport height to parent-constrained height, adding `min-h-0` to the key scroll containers where needed so shared custom-titlebar layouts no longer clip bottom content or footers
- Search window titlebar center alignment: moved the mode segmented control to a true centered position in the shared titlebar (absolute center anchoring), increased control width budget, and tuned spacing/typography so all mode buttons render fully and stay visually aligned
- Label window titlebar timer: moved the live EEG-window elapsed timer from the add-label page header into the shared `CustomTitleBar` center area via a new `label-titlebar.svelte.ts` reactive store, removing the duplicate in-content strip while keeping the timer live
- Search window vertical fit fix: changed `/search` root container from `h-screen` to `h-full min-h-0` so it honors the `#main-content` constrained height under the custom 30px titlebar and no longer overflows/clips at the bottom
- Search window titlebar button rendering fix: updated the shared `CustomTitleBar` search layout to be shrink-safe (`search-window-head` + `search-mode-switch` now flex responsively, title truncates with ellipsis, and mode buttons use equal-width flex sizing) so all search mode buttons render reliably instead of clipping on narrower windows/locales
- History window titlebar consolidation: moved clock icon, title text, day pagination (prev/next + label + position indicator), compare toggle, labels toggle, and reload button from the in-page header into the shared custom titlebar via a new `history-titlebar.svelte.ts` reactive store and callbacks; the in-page header strip is removed and the history page retains only the labels browser panel and scroll content
- Help window titlebar consolidation: moved the search input, version badge, license label, ThemeToggle, and LanguagePicker from the in-page top bar into the shared custom titlebar via a new `help-search-state.svelte.ts` reactive store; the redundant in-page header strip is removed and the search state is shared between the help page and the titlebar seamlessly
- Fixed all windows being clipped at the bottom by exactly the custom titlebar height (30 px): `#main-content` now uses `box-sizing: border-box; height: 100vh` so the `padding-top: 30px` offset is contained within the viewport height rather than overflowing beneath the body's `overflow: hidden` boundary
- Settings window titlebar consolidation: moved the Settings title label, Help button, ThemeToggle, and LanguagePicker from the in-page top bar into the shared custom titlebar; the redundant in-page header strip is removed and the Help button is shown in the titlebar actions whenever the settings window is active
- API Status window: moved title and Refresh button from the in-page header into the shared custom titlebar; the title bar now shows the window title on all platforms, with a refresh icon button next to ThemeToggle and LanguagePicker; the in-page header section is removed
- Search window titlebar alignment: moved Search title and mode toggle buttons (EEG/Text/Interactive) from the in-content header into the shared custom titlebar, with mode switching synchronized between the titlebar and `/search` content
- Updated the shared custom titlebar to show each non-main window title in the titlebar itself and to scope main-only titlebar actions (label/history) to the main window; non-main windows now keep lightweight titlebar controls (theme/language + window controls)
- Removed duplicate in-content title bars from all secondary windows (about, compare, whats-new, focus-timer, session, labels, search, history, calibration, label, onboarding, chat); functional header controls (mode buttons, day pagination, compare toggle, recording badge, elapsed timer, TTS indicator) are preserved in-place while redundant title text, drag regions, and theme/language buttons are removed
- Added global themed scrollbar styling for app scroll containers so Windows windows no longer show default system scrollbars; includes light/dark variants and automatic fallback to system colors in forced-colors mode

### LLM

- Moved per-session LLM transcript files into a dedicated `~/.skill/llm_logs` directory (`skill_dir/llm_logs/llm_<unix-seconds>.txt`) so all LLM logs live in a standalone folder instead of the `skill_dir` root.
- Added i18n translations for all LLM built-in tool toggle labels and descriptions across all five supported locales (en, de, fr, he, uk); `TOOL_ROWS` in `LlmTab.svelte` is now a reactive `$derived` so labels update instantly on language change
- Added per-tool allow-list settings for LLM chat in Settings → LLM; `date`, `location`, `web_search`, and `web_fetch` can now be enabled or disabled individually, and running chat requests pick up the updated tool allow-list immediately
- Multimodal projector selection now stays attached to a compatible downloaded text model instead of behaving like a standalone model; selecting an `mmproj` can auto-pair to a matching downloaded LLM, incompatible projector selections are cleared when the base model changes, and startup now honors the resolved projector path when autoload is enabled
- Added simple built-in tool-calling support in `POST /v1/chat/completions` with a bounded execution loop for `date`, `location`, `web_search`, and `web_fetch`
- Wired Tauri IPC chat streaming (`chat_completions_ipc`) to the same tool-calling loop so the in-app chat window supports the same built-in tools
- IPC chat now emits incremental visible `delta` chunks while tool-calling runs, using a stream sanitizer that suppresses `[TOOL_CALL]...[/TOOL_CALL]` blocks from the UI
- Added tool schema injection and `[TOOL_CALL]...[/TOOL_CALL]` handling so models can call tools and continue generation with tool results
- Added basic external fetch/search integrations (`ipwho.is`, DuckDuckGo instant answer API, and HTTP(S) page fetch) with bounded payload truncation for safe prompt context

### Dependencies (0.0.17)

- `llama-cpp-4` `0.2.7` → `0.2.9` (with matching `llama-cpp-sys-4` lockfile update)

### Build / CI

- Windows release workflow stability fix: `.github/workflows/release-windows.yml` now generates the temporary Tauri `--config` JSON via PowerShell (`ConvertTo-Json`) instead of `bash` + `python3`, removing a fragile command-path dependency that could fail the post-compile build step with exit `127` on `windows-latest`
- Linux release artifact generation fixed: `scripts/tauri-build.js` now treats both `--bundle` and `--bundles` (including `--flag=value`) as explicit bundling requests, preventing accidental `--no-bundle` injection that skipped `.deb`/`.AppImage` outputs in CI
- Added explicit Linux bundle-flag guard steps in CI and release workflows to fail fast if `tauri:build:linux:x64` drops `--bundles deb,appimage` or if `scripts/tauri-build.js` stops recognizing `--bundles`
- Added post-build Linux bundle directory sanity checks in CI and release workflows to fail early when `bundle/deb` or `bundle/appimage` is missing

## [0.0.23] — 2026-03-12

### UI / Build (0.0.23)

- **Custom titlebar for all windows** — replaced native window decorations with a custom titlebar component (minimize, maximize, close buttons) for consistent cross-platform appearance on all windows including main, settings, help, search, history, calibration, chat, downloads, and more
- **Unified window close behavior across all platforms** — on all platforms including Linux, closing the main window now hides it instead of exiting. Users must select "Quit" from the tray menu to exit, which shows a confirmation dialog
- **Downloads window total size footer** — the standalone Downloads window now shows the combined size of all listed downloads in a bottom footer for quick storage visibility
- **Downloads footer visibility improved** — clarified the footer label to “Total download size”, added item count, and increased footer emphasis so the summary is easier to notice
- **Downloads status bar placement** — moved the total-size summary from the bottom footer to an always-visible status bar directly under the Downloads header
- **Custom titlebar controls centralized** — titlebar minimize/maximize/close now use a single shared Svelte handler path (no per-window DOM-id listener wiring), improving consistency across windows
- **All windows aligned to shared custom titlebar path** — added missing window-capability labels (`history`, `compare`, `downloads`, `whats-new`), routed shortcut-created Chat/History windows through shared open-window commands, and ensured recreated main window remains undecorated so custom drag/control behavior is uniform
- **Main window titlebar consolidation** — moved language picker, theme toggle, label, and history buttons from the main card header to the titlebar for a cleaner, more accessible layout; buttons remain icon-only and responsive
- **Titlebar spacing refinement** — action buttons (label, history, theme, language) now live on the left side with window controls (minimize, maximize, close) on the right side, utilizing flex layout for proper visual separation
- **Linux cross-target preflight guard** — `scripts/tauri-build.js` now fails fast when a Linux host attempts a non-native `*-unknown-linux-gnu` target (for example ARM host → x86_64) without explicit opt-in, and prints actionable guidance; this avoids long builds ending in `glib-sys` / `gobject-sys` `pkg-config` cross-compilation failures
- **Linux build docs updated for ARM hosts** — added `pkg-config` cross-compilation troubleshooting to `LINUX.md`, including native ARM build command guidance and recommended x86_64 release build strategy
- **Native ARM64 Linux build shortcut** — added `npm run tauri:build:linux:arm64` to run the correct local aarch64 target build (`deb` + `AppImage`, `llm-vulkan`) in one command
- **Explicit Linux x64 cross-build shortcut** — added `npm run tauri:build:linux:x64`, which sets `ALLOW_LINUX_CROSS=1` and then runs the x86_64 target build path; this keeps accidental cross-target builds blocked by default while allowing intentional ones
- **CI Linux build command aligned with npm scripts** — `.github/workflows/ci.yml` now runs `npm run tauri:build:linux:x64` for the Linux release bundle smoke test instead of an inline `npx tauri build ...` command, keeping CI and local build entrypoints consistent
- **Tagged Linux release workflow aligned with npm scripts** — `.github/workflows/release-linux.yml` now also runs `npm run tauri:build:linux:x64` (with existing signing/env vars), replacing the inline `npx tauri build ...` command so both CI and release workflows share the same build entrypoint
- **Workflow intent comments added** — both `.github/workflows/ci.yml` and `.github/workflows/release-linux.yml` now include inline comments noting that `tauri:build:linux:x64` intentionally sets `ALLOW_LINUX_CROSS=1`, reducing accidental regressions to implicit cross-build behavior

### Bug fixes (Linux)

- **Main window close/minimize/maximize buttons unresponsive** — on Linux
  (Wayland + GNOME/Mutter/KWin), window decoration buttons did nothing
  after the window was created with `visible(false)` and later shown;
  this is a known upstream issue (tauri-apps/tauri#11856); worked around
  by toggling fullscreen briefly after every `show()` call on the main
  window (`linux_fix_decorations()`), which forces the Wayland compositor
  to re-evaluate decoration state; applied in initial setup show,
  `show_and_recover_main()`, and `complete_onboarding()`
- **Window event diagnostic logging** — added `[window-event]` and
  `[run-event]` stderr logging for `CloseRequested`, `Destroyed`,
  `Focused`, `Moved`, `Resized`, `ScaleFactorChanged`, and
  `ExitRequested` events across all windows

### Onboarding (0.0.23)

- **Downloads complete success screen** — when all recommended models
  (Qwen3.5 4B, ZUNA encoder, NeuTTS, Kitten TTS) are downloaded, the
  onboarding done step now displays a prominent **green checkmark** with
  a success message and a clickable link to **settings** where users can
  download additional models or switch to alternatives
- **Downloads complete i18n** — added `onboarding.downloadsComplete`,
  `onboarding.downloadsCompleteBody`, and `onboarding.downloadMoreSettings`
  keys to all five locales (en, de, fr, he, uk)

## [0.0.17] — 2026-03-11

### UI / Build (0.0.17)

- **Tailwind Vite parser crash in MarkdownRenderer fixed** — resolved
  `[plugin:@tailwindcss/vite:generate:serve] Invalid declaration: Marked`
  by refactoring `src/lib/MarkdownRenderer.svelte` to use `marked.parse(...)`
  with a local renderer object and removing an empty local `<style>` block
- **MarkdownRenderer regression guard** — added
  `scripts/check-markdown-renderer.js` and wired it into `npm run check`
  so CI/local checks fail if `MarkdownRenderer.svelte` reintroduces
  `new Marked(...)` or a local `<style>` block
- **MarkdownRenderer guard now runs before dev startup** — `npm run dev`,
  `npm run build`, `npm run check:watch`, and `npm run tauri dev` now execute
  the MarkdownRenderer guard before Vite / SvelteKit startup so Tailwind
  parser regressions fail immediately instead of surfacing later from the
  Tailwind Vite pipeline

### Settings

- **Open `skill_dir` from Settings** — Data Directory now includes an
  **Open Folder** action that opens the fixed `~/.skill` directory in the
  system file manager

### Onboarding (0.0.17)

- **Recommended models quick setup** — onboarding now includes a one-click
  **Download Recommended Set** action that pulls the default local stack:
  **Qwen3.5 4B (Q4_K_M)**, **ZUNA encoder**, **NeuTTS**, and **Kitten TTS**
- **Qwen quant preference tightened** — when selecting the onboarding LLM
  target, the wizard now explicitly prefers **Q4_K_M** for Qwen3.5 4B
- **Staged background downloads** — onboarding now starts the recommended
  model downloads in sequence while the user continues setup: ZUNA →
  KittenTTS → NeuTTS → Qwen3.5 4B (`Q4_K_M` target)
- **Persistent footer model status** — all onboarding views now show a subtle
  footer line with staged model setup progress, and the onboarding window was
  enlarged slightly to keep spacing readable
- **Download order configured in Rust constants** — the onboarding queue no
  longer hardcodes download order in Svelte; it now reads the canonical
  sequence from `src-tauri/src/constants.rs`
- **Onboarding models i18n complete** — added the 16 missing
  `onboarding.step.models`, `onboarding.modelsHint`, `onboarding.modelsTitle`,
  `onboarding.modelsBody`, and `onboarding.models.*` keys to all four
  non-English locales (de, fr, he, uk)

### Tray / Downloads

- **LLM download progress in tray icon + menu** — while model files are
  downloading, the system tray now shows progress in the icon itself (a
  prominent circular ring around the tray icon) and in the tray menu
  (active download rows with filename, percent and live status text)
- **Standalone Downloads window** — added a dedicated downloads manager
  window (`/downloads`) that lists all model downloads at any time with
  per-item actions: pause, resume, cancel, and delete
- **Download initiated timestamp** — each download row now includes when it
  was started so long-running and resumed transfers are easier to track
- **Downloads i18n** — new downloads-window labels/status strings added to
  all shipped locales
- **Tray menu shortcut to Downloads** — added a direct **Downloads…** menu
  action in the tray, opening the standalone downloads window in one click

### Dependencies

- `llama-cpp-4` `0.2.6` → `0.2.7`

### CI / Build

- **Linux local `tauri build` segfault avoided** — `scripts/tauri-build.js`
  now injects `--no-bundle` by default for Linux `build` runs when the caller
  does not explicitly pass `--bundle` / `--no-bundle`; this avoids a native
  post-compile crash (status 139) in the Tauri CLI bundling/updater phase
  while still producing the release binary at
  `src-tauri/target/release/skill`

- **Windows release — wrong `link.exe`** — the GitHub-hosted `windows-latest`
  runner places `C:\Program Files\Git\usr\bin` (Git for Windows' Unix `link`
  utility) before the MSVC toolchain in `PATH`; Rust's MSVC backend resolved
  `link.exe` to that Unix binary, which rejected all MSVC linker flags with
  an "extra operand" error; fixed by adding a PowerShell step immediately
  after `ilammy/msvc-dev-cmd` in `release-windows.yml` that strips every
  `Git\usr\bin`-like entry from `PATH` via `$GITHUB_ENV`, ensuring the MSVC
  `link.exe` wins for all subsequent steps

---

## [0.0.16] — 2026-03-11

### EEG / Embeddings

- **Cross-day HNSW index** — similarity search is no longer scoped per-day;
  a persistent cross-day index (rolling 30-day merged index) is maintained
  under `~/.skill`; near-neighbours across months can be found in a single
  query
- **Label fuzzy semantic search** — label search now uses the vendored
  `fast-hnsw` label index for semantic matching in addition to plain-text
  filtering; queries like "find sessions where I felt anxious" surface
  nearest-neighbour label clusters rather than exact string hits

### LLM — Chat

- **Chat history persisted** — conversations are stored in SQLite at
  `~/.skill/chat_history.sqlite`; messages survive closing and reopening
  the chat window
- **Multi-conversation sidebar** — the chat window now has a sidebar listing
  named conversation threads; threads are persisted to disk and can be
  renamed or deleted
- **System prompt editor** — the system prompt is exposed as a text area in
  the chat settings panel so users can bias the model (e.g. "you are a
  neurofeedback coach") without recompiling
- **EEG context injection** — the current `eeg-bands` WebSocket event is
  automatically wired into the system prompt: "User's current focus: 72,
  relaxation: 58, SNR: 14 dB…" so the model can give contextualised advice
- **Prompt library** — a built-in set of neurofeedback prompt templates
  (e.g. "Summarise today's session", "Suggest a relaxation technique",
  "Explain what high theta means") is accessible from a `+` button in the
  chat input

### LLM — Downloads

- **Model download resumption** — interrupted downloads no longer restart
  from zero; the downloader uses `Content-Range` byte-range requests to
  resume from the last received byte

### UMAP Viewer

- **Export PNG / JSON** — "Export PNG" and "Export JSON" buttons added to
  the 3D scatter plot toolbar; PNG captures the current WebGL viewport,
  JSON exports the full point cloud with labels and timestamps

### Focus Timer

- **Session log** — a summary panel shows today's completed Pomodoro cycles,
  total focus time, and total break time; entries are labelled and persisted
  across restarts

### Onboarding (0.0.16)

- **Extended checklist** — onboarding now includes four additional steps:
  download an LLM model, run a similarity search, set a DND threshold, and
  try the REST API; previous four steps preserved

### UI / UX

- **Command Palette — fuzzy scoring** — the palette filter now uses an
  fzf-style scored fuzzy algorithm; partial matches are ranked by relevance
  instead of simple `includes()` containment
- **Theme — custom accent colour** — a primary-hue colour picker has been
  added to Settings → Appearance; the chosen accent is stored in
  `~/.skill/theme.json` and applied app-wide; dark / light / system mode
  selection is unaffected

---

## [0.0.15] — 2026-03-11

### Windows support

- **espeak-ng static build — Windows MSVC** — `scripts/build-espeak-static.ps1`
  builds `espeak-ng.lib` from source using CMake + MSVC on first run, then
  links it statically into the binary; subsequent builds are a no-op.
  Produces a single merged `.lib` (CMake + `lib.exe /OUT`) so that the linker
  sees no duplicate symbol conflicts
- **espeak-ng static build — Windows MinGW/GNU** — `scripts/build-espeak-static-mingw.sh`
  handles the `*-windows-gnu` target ABI (MSYS2 / cross-compile from
  Linux/macOS); output goes to `espeak-static-mingw/` to avoid collisions
  with the MSVC archive
- **`build.rs` — multi-platform espeak dispatch** — selects the correct
  build script and archive name based on `CARGO_CFG_TARGET_ENV`:
  `msvc` → PowerShell `.ps1`, `gnu` on Windows → MinGW `.sh`,
  macOS/Linux → Unix `.sh`; links `-lc++` on macOS, `-lstdc++` on Linux/MinGW,
  and omits the C++ flag on MSVC (runtime linked automatically)
- **`build.rs` — espeak data copy deferred to release** — the espeak-ng data
  directory is no longer copied during `cargo build` / `tauri dev`; the copy
  is skipped in debug builds to break the infinite rebuild loop where
  `build.rs` copies → Tauri watcher detects the change → `cargo run` → repeat.
  An empty placeholder directory is still created so Tauri's resource-path
  validation does not error at startup
- **`fast-hnsw` — vendored with Windows fix** — `memmap2::Mmap::advise()` and
  `memmap2::Advice` are `#[cfg(unix)]` and not available on Windows; the
  `fast_hnsw` crate's unconditional `advise(Advice::Random)` call caused a
  compile error on the MSVC target; patched locally via
  `src-tauri/vendor/fast-hnsw` with the `advise` call wrapped in
  `#[cfg(unix)]`; vendored until upstream releases a fix
- **`WINDOWS.md`** — updated prerequisites: Visual Studio Build Tools 2022
  (**Desktop development with C++** workload, provides `cl.exe`, `lib.exe`,
  Windows SDK) now listed as step 1; CMake doc note updated to cover
  espeak-ng's build system in addition to llama.cpp; added Git as a
  prerequisite for cloning the espeak-ng source; renumbered all steps

### Build / tooling

- **`scripts/tauri-build.js` refactored** — now a general Tauri wrapper that
  handles `dev`, `build`, and any other subcommand; non-`dev`/`build`
  subcommands (e.g. `tauri info`, `tauri signer`) pass straight through without
  triggering an espeak pre-build; platform detection now also covers
  `*-windows-gnu` (MinGW)
- **npm `tauri` script** — `"tauri": "node scripts/tauri-build.js"` routes all
  `npm run tauri …` invocations through the wrapper, so `npm run tauri info`,
  `npm run tauri dev`, `npm run tauri build -- --debug`, etc. all work
  consistently across platforms

### LLM — WebSocket / REST API

- **WebSocket commands** — `llm_status`, `llm_start`, `llm_stop`,
  `llm_catalog`, `llm_download`, `llm_cancel_download`, `llm_delete`,
  `llm_logs` added to the WebSocket command dispatcher (all behind the `llm`
  Cargo feature flag)
- **REST endpoints** — matching HTTP shortcuts added to the axum router in
  `api.rs`:
  - `GET  /llm/status` — running state, active model name, context size, vision flag
  - `POST /llm/start` — load the active model and start the inference server
  - `POST /llm/stop` — stop the inference server and free GPU/CPU resources
  - `GET  /llm/catalog` — model catalog with per-entry download states
  - `POST /llm/download` — start a background model download `{ "filename": "…" }`
  - `POST /llm/cancel_download` — cancel an in-progress download
  - `POST /llm/delete` — delete a locally-cached model file
  - `GET  /llm/logs` — last 500 LLM server log lines
  - `POST /llm/chat` — non-streaming chat completion; body: `{ message, images?, system?, temperature?, max_tokens? }`; returns `{ text, finish_reason, tokens }`
- **`LlmServerState::chat()`** — new method on the server-state actor handle;
  submits a generate request to the actor's channel and returns an
  `UnboundedReceiver<InferToken>` for streaming; returns `Err` immediately if
  the model is still loading or the actor has exited
- **`extract_images_from_messages()`** — helper that decodes all
  `data:<mime>;base64,…` data-URL image parts from an OpenAI-style messages
  array into raw `Vec<u8>` bytes; plain `https://…` URLs are silently skipped;
  call before passing messages to the actor so it receives pre-decoded bytes

### LLM — verbose logging

- **`LlmConfig.verbose`** (`bool`, default `false`) — when `false` (default),
  all internal llama.cpp / ggml and clip logs are silenced; set `true` to
  see raw tensor-load progress and other low-level detail
- **`mtmd_log_set` silence** — `clip_model_loader` uses a separate logger
  (`mtmd_log_set`) that is not affected by `llama_log_set`; the clip logger is
  now silenced via a no-op `extern "C"` callback when `verbose = false`,
  eliminating the tensor-load spam when loading a multimodal projector

### CLI (`cli.ts`)

- **`llm` subcommand group** added:
  - `llm status` — print LLM server status (stopped / loading / running)
  - `llm start` — load the active model and start the inference server
  - `llm stop` — stop the inference server, free GPU memory
  - `llm catalog` — list all catalog models with download states
  - `llm download <filename>` — start a background model download
  - `llm cancel <filename>` — cancel an in-progress download
  - `llm delete <filename>` — delete a cached model file
  - `llm logs` — print the last 500 LLM server log lines
  - `llm chat` — interactive multi-turn chat REPL (WebSocket streaming)
  - `llm chat "message"` — single-shot: send one message and stream the reply
  - `llm chat "message" --image a.jpg --image b.png` — vision: attach images
    (files are base64-encoded and embedded as `image_url` content parts;
    requires a vision-capable model with mmproj loaded)
- **`--image <path>`** — new flag (repeatable) for attaching image files to
  `llm chat` turns
- **`--system <prompt>`** — system prompt prepended as a `{ role: "system" }`
  message for `llm chat`
- **`--max-tokens <n>`** — maximum tokens to generate per turn
- **`--temperature <f>`** — sampling temperature (0 = deterministic, 1 = creative)

### Settings UI

- **Sidebar navigation** — the tab bar in Settings has been replaced with a
  persistent sidebar; each tab entry shows a 24 × 24 stroked SVG icon alongside
  the label and an active-indicator bar
- **Keyboard shortcuts** — `Cmd/Ctrl + 1–9` switch between the first nine
  settings tabs; tooltips on each sidebar item show the shortcut hint

### Help UI

- **Sidebar navigation + search** — the Help window now uses the same sidebar
  layout as Settings; a search box in the top bar filters across all help
  sections with keyboard-navigable results

### Internals

- **`SKILL_DIR` constant** — `src-tauri/src/constants.rs` now exports
  `pub const SKILL_DIR: &str = ".skill"` so the directory name is defined
  in one place; `default_skill_dir()` in `settings.rs` uses it
- **Data directory hardcoded** — the `data_dir` field has been removed from
  persisted settings; the skill directory is always `~/.skill` and is never
  configurable at runtime; `expand_tilde` helper and its tests removed

### Dependencies (0.0.15)

- `kittentts` `0.2.4` → `0.2.5`

---

## [0.0.13] — 2026-03-10

### Onboarding (0.0.13)

- **Recommended models quick setup** — onboarding now starts staged
  background downloads automatically while the user proceeds through steps,
  in this order: ZUNA → KittenTTS → NeuTTS → Qwen 3.5 4B (`Q4_K_M` target)
- **Persistent footer download status** — all onboarding views now show a
  subtle footer line with staged model setup progress (ZUNA, Kitten, NeuTTS,
  LLM), and the onboarding window size was increased to keep spacing readable
  with the always-visible footer indicator

### Dependencies (0.0.13)

- `llama-cpp-4` `0.2.3` → `0.2.5`
- `kittentts` `0.2.2` → `0.2.4`
- `neutts` `0.0.5` → `0.0.7`

### Bug fixes

- **Blank main window after long idle** — after a full day in the system
  tray with the window hidden, macOS can silently terminate WKWebView's
  web-content process under memory pressure, leaving a blank white page
  that only a full app restart could recover from
  - `+layout.svelte` sets `window.__skill_loaded = true` in `onMount` as
    a renderer-liveness sentinel
  - New `show_and_recover_main()` Rust helper checks the sentinel on every
    show via `eval()`; if the flag is absent it triggers `location.reload()`
    (renderer alive but content cleared), and falls back to `navigate()` if
    `eval()` itself returns `Err` (renderer process fully dead, WKWebView
    needs a fresh process spawned)
  - `RunEvent::Reopen` handler added — clicking the macOS Dock icon while
    all windows are hidden now shows the main window and runs the same
    two-layer recovery (previously a silent no-op)

- **Update loop — first check delayed by full interval** — the background
  updater slept `interval_secs` *before* the first check, so with the
  default 1-hour interval the first background check fired ~61 minutes after
  launch; pattern changed to check-then-sleep so the first check fires 30
  seconds after startup as intended

- **Update loop — update silently dropped on CDN race** — when the Rust
  background task emitted `update-available`, the frontend had to re-run
  `check()` to obtain a downloadable `Update` object; if `check()` returned
  `null` (latest.json not yet propagated to all CDN edge nodes), `available`
  was wiped and `phase` reverted to `"idle"` with no user feedback; fixed
  by threading the event payload as a `hint` through `checkAndDownload()` —
  the known version stays visible in the UI during the re-check, and a CDN
  race surfaces an actionable "Retry" error instead of a silent reset

- **What's New — dismiss race with uninitialised version** — `appVersion`
  started as the string `"…"` and was populated asynchronously via IPC;
  clicking "Got it" before the call resolved stored `"…"` in
  `last_seen_whats_new_version`, causing the window to reopen on every
  subsequent launch; fixed by seeding `appVersion` synchronously from the
  CHANGELOG version embedded at build time

- **What's New — markdown not rendered** — changelog entries containing
  `**bold**`, `` `code` `` spans, multi-line bullet continuations, and
  numbered sub-lists were all rendered as plain text; replaced the
  hand-rolled `parseChangelog` parser (which dropped any line not starting
  with `-` plus a trailing space) and the manual `{#each sections}` template with
  `MarkdownRenderer` (existing component backed by `marked` + GFM); scoped
  CSS overrides inside `.wn-body` preserve the compact window style without
  affecting the chat renderer

### Build / CI (0.0.13)

- **CI `cargo check --locked` failing on Linux** — `Cargo.lock` generated
  on macOS caused the Linux CI job to fail with "cannot update the lock file
  because --locked was passed"; added `cargo fetch --target
  x86_64-unknown-linux-gnu` before `cargo check --locked` to resolve
  platform-specific dependencies for Linux without touching the network
  during the check itself

- **Release — single notarization round trip** — the release workflow
  previously issued two separate `xcrun notarytool submit --wait` calls
  (one for the `.app` as a ZIP, one for the DMG), each waiting up to 20+
  minutes; consolidated to a single DMG submission — Apple's service
  registers notarization tickets for all signed content inside the container,
  so `xcrun stapler staple` succeeds on both the DMG and the `.app`
  afterward without a second submission; the updater tarball step is
  reordered to run after the DMG step so it always packages a stapled `.app`

---

## [0.0.11] — 2026-03-10

### LLM / Chat

- **LLM engine** — full on-device inference via `llama-cpp-4` (llama.cpp
  bindings). Runs text and multimodal (vision) models locally with no cloud
  dependency
- **Model catalog** (`llm_catalog.json`) — curated list of GGUF models
  (Qwen3.5 4B/27B, Llama-3.2-Vision, Gemma3, etc.) with per-entry metadata:
  repo, filename, quantisation, size, family description, tags, recommended
  flag. Bundled into the app at compile time
- **Tauri commands**: `get_llm_catalog`, `set_llm_active_model`,
  `set_llm_active_mmproj`, `download_llm_model`, `cancel_llm_download`,
  `delete_llm_model`, `refresh_llm_catalog`, `get_llm_logs`,
  `start_llm_server`, `stop_llm_server`, `get_llm_server_status`,
  `open_chat_window`
- **HTTP inference server** (`axum` router) — OpenAI-compatible endpoints
  (`/v1/chat/completions`, `/v1/completions`, `/v1/embeddings`) served
  locally so third-party tools can connect to the on-device model
- **Vision / multimodal** — image inputs decoded from data-URL or base64
  and fed through a clip mmproj; `autoload_mmproj` setting automatically
  selects the best downloaded projector for the active model
- **Thinking-model support** — forced `</think>` injection after a budget
  cap; orphaned tail tokens are discarded (decoded into KV cache for
  coherence, suppressed from output) until the next clean line boundary
- **File upload** in chat — images attachable to messages; previewed in
  the UI before sending
- **Markdown renderer** (`MarkdownRenderer.svelte`) — renders streamed
  assistant output with code blocks, tables, and inline formatting
- **Chat window** (`src/routes/chat/+page.svelte`) — full chat UI with
  message history, streaming tokens, stop button, model/mmproj selectors,
  generation parameter controls
- **Global chat shortcut** — configurable keyboard shortcut (stored in
  settings) focuses the existing chat window or opens a new one
- **i18n** — `llm.*` keys added to all five language files (en, de, fr,
  he, uk)

### Build / CI (0.0.11)

- **Bypass Tauri's built-in signing pipeline** in both `release.yml` and
  `pr-build.yml` — Tauri's `create-dmg` subprocess crashes with `SIGILL`
  on macOS 26 (hdiutil API change); replaced with explicit steps:
  1. `npx tauri build --bundles app --no-sign` — compile only
  2. `codesign` — deep-sign with `--options runtime` + `--entitlements`
  3. `xcrun notarytool submit … --wait` — notarize
  4. `xcrun stapler staple` — staple ticket to bundle
  5. Recreate `.app.tar.gz` from the signed bundle, then
     `npx tauri signer sign` — re-sign the updater artifact with Ed25519
- `release.sh` — minor fix to `TAURI_TARGET` default propagation

---

## [0.0.9] — 2026-03-10

### Dependencies (0.0.9)

- Migrated `llama-cpp-4` and `llama-cpp-sys-4` to local path via
  `[patch.crates-io]` (`../../../llama-cpp-rs/llama-cpp-4` and
  `../../../llama-cpp-rs/llama-cpp-sys-4`) — ensures the SIGILL fix
  (correct `CMAKE_OSX_ARCHITECTURES` / `CMAKE_CROSSCOMPILING` for Apple
  cross-arch builds) is always active; both the `llm` feature and neutts's
  backbone resolve to the same local crate, preserving the `links = "llama"`
  deduplication

### Build / CI (0.0.9)

- macOS builds now target `aarch64-apple-darwin` (arm64) only — x86_64
  is no longer compiled
  - `tauri:build:mac` npm script passes `--target aarch64-apple-darwin`
  - `release.sh` defaults `TAURI_TARGET` to `aarch64-apple-darwin` (still
    overridable via env var for universal or x86_64 builds)
  - `build-espeak-static.sh` defaults `CMAKE_OSX_ARCHITECTURES` to `arm64`
    instead of the host architecture (still overridable via `ESPEAK_ARCHS`)
  - `.cargo/config.toml` sets `[build] target = "aarch64-apple-darwin"` so
    plain `cargo build` / `cargo check` / `npx tauri build` all default to
    arm64 without requiring an explicit `--target` flag
  - `ci.yml` Linux `cargo check` / `cargo clippy` steps now pass
    `--target x86_64-unknown-linux-gnu` to override the config.toml default;
    espeak build step passes `ESPEAK_ARCHS=x86_64` explicitly
  - `pr-build.yml` and `release.yml` were already correct (`--target
    aarch64-apple-darwin`, `ESPEAK_ARCHS=arm64`)
- Fixed SIGILL crash after successful compile on macOS 26.3 in both local
  and CI builds; root cause traced via lldb + macOS crash report:
  - Tauri's bundled `create-dmg` script spawns `bundle_dmg.sh` as a child
    process which fails on macOS 26 (hdiutil API change); Node.js propagates
    the child's fatal exit as `process.kill(pid, SIGILL)` via
    `ProcessWrap::OnExit` → promise rejection chain
  - Local dev (`tauri:build:mac`): added `--no-sign` — no certificate on dev
    machines, codesign would have failed at the same stage
  - CI (`release.yml`, `pr-build.yml`): replaced `--bundles app,dmg` with
    `--bundles app`; added an explicit "Create DMG" step that uses `hdiutil`
    directly, stamps the version badge, then signs and notarizes — identical
    end result with no dependency on Tauri's create-dmg script
- Fixed pre-commit hook failing on macOS when CUDA Toolkit is absent
  - `cargo clippy --all-features` activated `llm-cuda` and `llm-vulkan`,
    causing `llama-cpp-sys` to pass `-DGGML_CUDA=ON -DGGML_VULKAN=ON` to
    CMake, which hard-errors if no CUDA Toolkit is found
  - Hook now selects platform-appropriate features: `--features llm-metal`
    on macOS, default features on Linux/Windows — CUDA/Vulkan features are
    never activated where their native toolkits are unavailable

---

## [0.0.6] — 2026-03-06

### Do Not Disturb / Focus Mode

- Replaced hand-rolled ObjC FFI + XPC DND implementation with the
  [`macos-focus`](https://crates.io/crates/macos-focus) crate — pure Rust,
  no private frameworks, no special entitlements
- DND now works reliably on macOS 12–15; the XPC path that consistently
  returned "operation failed" errors has been removed
- Added **Focus mode picker** in Settings → Goals: choose any Focus mode
  configured in System Settings (Do Not Disturb, Work, Personal, Sleep,
  Driving, …) rather than always activating Do Not Disturb
- `focus_mode_identifier` persisted in settings; defaults to Do Not Disturb
  for backwards compatibility with existing configs
- Added `list_focus_modes` Tauri command backed by
  `FocusManager::available_modes()`; falls back to the full first-party mode
  list if `ModeConfigurations.json` is unreadable
- Added TODO stubs for Linux (D-Bus / xdg-portal) and Windows
  (WinRT / IQuietHoursSettings) DND support

### Quit Dialog

- macOS quit confirmation dialog now uses `NSAlert` via `objc2-app-kit`
  dispatched through `dispatch2::DispatchQueue::main().exec_sync()`,
  eliminating the `CFUserNotificationDisplayAlert: called from main
  application thread` log warning that `rfd` triggered

### Bug fixes and warnings

- Fixed `CFStringCreateWithCString` / `CFRelease` clashing `extern "C"`
  signatures between `dnd.rs` and `gpu_stats.rs`
- Removed three unnecessary `unsafe {}` blocks around safe `iimp()` closure
  calls in the (now-deleted) ObjC FFI path
- Removed unused `vm_deallocate` extern declaration in `gpu_stats.rs`
- Removed unnecessary `unsafe {}` block wrapping safe `NSAlert` method calls
- Fixed unescaped ASCII `"` inside German DND strings in `de.ts` that caused
  587 cascading TypeScript parse errors
- Replaced `map_or(false, |v| v == 1)` with `== Some(1)` in `gpu_stats.rs`
- Replaced manual `div_ceil` closure with `u64::div_ceil` in `job_queue.rs`
- Replaced `&&` range assertions with `.contains()` in `ppg_analysis.rs`
- Replaced `vec![…]` with array literals in test push calls
- Replaced `for ch in 0..N` index loops with `enumerate()` iterators in
  `eeg_bands.rs` and `eeg_filter.rs`
- Moved constant-value `assert!` calls into `const { }` blocks in
  `constants.rs`
- Fixed doc comment continuation indent warnings in `gpu_stats.rs`

### i18n

- Added `dnd.focusMode`, `dnd.focusModeDesc`, `dnd.focusModeLoading` keys
  to all five language files (en, de, fr, he, uk)

---

## [0.0.3] — 2026-03-06

- Added NeuTTS engine support alongside KittenTTS, with seamless switching between engines
- TTS engine switching now works reliably in both directions
- Graceful shutdown for NeuTTS on engine change or app exit
- TTS caching and quality improvements
- UI updates for TTS tab including progress/error state display
- Fixed TypeScript type for TTS error phase
- Added translations
- Better updater configuration

---