Adds battery-state awareness to the thread-count helper: when
probe_power_state() returns OnBattery, chosen is halved before the
[MIN, MAX] clamp. Also removes the leading underscore from the
workload/gpu_offloaded parameters (they are silenced via let _ = ...
until the GPU-clamp task).
New test battery_halves_thread_count verifies on_battery <= on_ac and
>= MIN when the host has more than MIN physical cores. Restructured
to sequential with_override calls (not nested) to avoid re-entrant
deadlock on power::TEST_LOCK.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Without a lock, env_var_bypasses_clamps (set_var + remove_var) and
matches_existing_clamp_when_no_clamps_apply (remove_var) race under
cargo's parallel test runner. Task 2.2 will add another remove_var
call, compounding the race.
Adds a #[cfg(test)] static THREAD_ENV_LOCK and a closure-style
with_thread_env_lock helper, mirroring power::with_override.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds crates/core/src/tuning.rs with MIN/MAX_INFERENCE_THREADS consts,
Workload enum (Llm/Whisper), and inference_thread_count() helper matching
the existing constants::inference_thread_count clamp behaviour. Three unit
tests pass. tracing dep added to Cargo.toml (used by future tasks).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Cache sits between override resolution and platform_probe; override
paths (test + env-var) bypass it entirely so they always take effect
immediately. OnceLock<Mutex<Option<CachedState>>> initialised lazily on
first probe. Two test-only helpers (force_clear_cache, force_set_cache)
expose the cache slot for unit tests. Mutex import ungated — now used in
production cache code as well as test override.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
TEST_OVERRIDE and test_get_override are now #[cfg(test)]-gated and
the read in probe_power_state is wrapped in a #[cfg(test)] block,
so test scaffolding doesn't compile into production builds.
with_override now uses a drop-guard pattern so a panic inside body
still resets TEST_OVERRIDE, preventing stale state from leaking
into subsequent tests.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds the public probe entry point with two override mechanisms:
- In-process TEST_OVERRIDE slot serialised by TEST_LOCK mutex for
unit tests (avoids races with cargo's parallel runner).
- MAGNOTIA_POWER_STATE_OVERRIDE env var for integration tests set
externally.
Platform dispatch: Linux reads /sys/class/power_supply; all others
return Unknown. Five new override tests pass alongside the six sysfs
parser tests from Task 1.2 (12 total green).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Pure function takes a directory path and returns OnAc, OnBattery, or
Unknown by reading `type` and `online` files in each supply entry.
Matches the kernel ABI documented at
Documentation/ABI/testing/sysfs-class-power. Failure modes (missing
dir, unreadable files, malformed contents) all fall through to Unknown
rather than panicking.
Seven tests cover: Mains online, battery-only, USB-PD, empty dir,
missing dir, and malformed entries. All pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Introduces crates/core/src/power.rs with the PowerState enum
(OnAc / OnBattery / Unknown) and a unit test confirming the three
variants are distinct. Registers the module in lib.rs and adds
tempfile = "3" to dev-dependencies for use in later tasks.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bite-sized TDD plan implementing the design at
docs/superpowers/specs/2026-05-09-battery-gpu-aware-thread-tuning-design.md.
8 phases, ~13 commits:
- Phase 1: power.rs (PowerState enum, sysfs parser, probe + overrides,
10s TTL cache)
- Phase 2: tuning.rs (Workload enum, helper, battery clamp, GPU clamp,
per-process logging)
- Phase 3: vulkan_loader_available moves from src-tauri to magnotia-core
- Phase 4: whisper backend wires through new helper
- Phase 5: LLM call site wires through new helper after model load
- Phase 6: remove old constants::inference_thread_count facade
- Phase 7: thread_sweep.rs prints 4-panel power-aware RTF table
- Phase 8: manual battery validation smoke
Each task is TDD: write failing test, run, implement, run, commit.
Tests live in unit-test modules inside power.rs and tuning.rs.
Override design uses an in-process with_override(state, |closure|)
helper for unit tests (cargo runs them in parallel; env-var path
would race) and MAGNOTIA_POWER_STATE_OVERRIDE env var for
integration tests like thread_sweep.rs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Spec for two new clamps on the existing inference_thread_count helper:
- Battery clamp: drop to physical/2 when on battery (Linux sysfs probe;
macOS/Windows return Unknown, treated as OnAc).
- GPU-offload clamp: 2 threads for fully-offloaded LLM, 4 for Whisper
(Whisper keeps mel spectrogram, decoder bookkeeping, and beam search
on CPU even with full Vulkan offload).
Codex + online-research consult: no prior art in llama.cpp, Ollama,
Jan, or mistral.rs. Apple's Low Power Mode is OS-level (P-core
frequency cap), not a published software API. Lumenote would be first
in the open Rust inference-wrapper space; instrumentation matters,
hence the thread_sweep.rs extension + per-process INFO log.
Architecture (Approach B):
- New crates/core/src/power.rs (PowerState, sysfs probe, 10s TTL).
- New crates/core/src/tuning.rs (Workload enum, helper).
- Move vulkan_loader_available() from src-tauri to magnotia-core
so crates/transcription can call it without a Tauri dep.
- Existing constants::inference_thread_count() becomes deprecated
facade for one commit, then removed.
GPU-offload detection is intent-based (use_gpu && requested_layers
>= model.n_layer() for LLM; cfg(whisper-vulkan) && libvulkan
resolvable for Whisper). Residency-based detection deferred:
llama-cpp-2 0.1.145 doesn't expose post-load offload count, so a
true outcome flag would need llama_log_set callback parsing or an
upstream PR. Caveat documented in the spec.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bundled work from a low-end laptop (Ryzen 5 4650U / Vega 6 / Linux Mint
22.2 / X11) profiling pass.
perf: WEBKIT_DISABLE_DMABUF_RENDERER=1 default on all Linux
Previously only set on Wayland sessions. Empirically it's a
significant idle-cost win on integrated GPUs in either session type:
env-var matrix (release binary, 75s settle, 10s jiffies CPU sample)
showed magnotia idle CPU 12.30% → 2.80% of one core and idle GPU
17% → 10% on this hardware. Users can opt back in by exporting
WEBKIT_DISABLE_DMABUF_RENDERER=0. The Wayland-only XWayland
fallback (GDK_BACKEND=x11, WINIT_UNIX_BACKEND=x11) is unchanged.
fix: secondary-windows ACL — allow set-always-on-top
The float window's pin toggle calls setAlwaysOnTop() but the
secondary-windows capability didn't permit it, so the popout was
stuck always-on-top regardless of the pin state. Adds the
core:window:allow-set-always-on-top permission. Narrow scope.
fix: guard registerGlobalHotkey against non-main webviews
Cross-window settings sync via localStorage can re-fire the
$effect(() => settings.globalHotkey) callback inside popout webviews
where the main layout's registerGlobalHotkey is reachable. Adds an
early-return when the current window label is not "main", so the
popout doesn't trigger an ACL-denied register/unregister and the
user no longer sees a spurious "Hotkey not registered" toast when
popouts are open. Keeps the global-shortcut perm scoped to main.
build: pin @tauri-apps/api 2.10.1 + @tauri-apps/plugin-dialog 2.7.1
Match the Rust crate versions tauri-cli's version-mismatch check
enforces during release builds. Without this, `npm run tauri build`
exits 0 silently while emitting an Error and never producing
binaries.
test: add crates/transcription/tests/jfk_bench.rs
Reproducible RTF regression fixture. Env-gated on
MAGNOTIA_WHISPER_TEST_MODEL + MAGNOTIA_WHISPER_TEST_AUDIO so it
never runs in CI without setup. Loads the JFK WAV inline (no hound
dep), times model load + cold + warm transcribe, prints SUMMARY.
Baselines on this hardware:
--release --features whisper: cold RTF 0.054, warm 0.050, RSS 125 MB
--release --features whisper,whisper-vulkan: cold RTF 0.029, warm 0.028, RSS 125 MB
Vulkan on RADV/Vega 6 nearly halves transcription latency for
Whisper Tiny — useful baseline for Phase 10 hardware-recommendation
scoring.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When an async onChange handler rejects, the toggle snapped back to its
previous value silently — screen reader users had no signal that the
action failed, only that the toggle suddenly returned to its original
state. Now aria-invalid="true" exposes the failure on the role=switch
button for 1.5s after rejection, so assistive tech announces the error.
Self-clears via setTimeout; no API change for callers.
Hard-coded rgba(214,132,80,X) shadows were tied to the dark-theme accent
and stayed off-hue in light mode (where the accent is a darker #a3683a).
Six tokens added to app.css @theme and mirrored in the light-theme block
plus design-system/colors_and_type.css:
--shadow-accent-sm 0 0 8px 0.25 alpha Toggle on-state
--shadow-accent-md 0 4px 16px 0.3 ModelDownloader card
--shadow-accent-lg 0 4px 20px 0.3 DictationPage record button
--shadow-accent-glow 0 0 8px 0.4 progress-bar glow
--shadow-accent-pill 0 1px 4px 0.3 SegmentedButton pill
--shadow-accent-raised 0 2px 8px 0.2 FilesPage browse tile
Migrated seven sites: Toggle, SegmentedButton, ModelDownloader (x2),
FilesPage (x2), DictationPage record button, SettingsPage locale pill.
Focus-ring shadows (0_0_0_3px_rgba(...,0.1)) left alone — handled in
parallel via the --accent-shadow-focus token.
@font-face: app.css and colors_and_type.css both declare the same five
fonts. Duplication is unavoidable because preview/*.html loads the
design-system file directly without Vite, so it cannot share the
runtime declarations. font-weight ranges and font-style reconciled to
match exactly (Atkinson 200-800 not 400-700, JetBrains gains font-style:
normal); src URLs intentionally differ (root-absolute vs relative).
Comments added to both copies explaining the contract.
- Extend isInputFocused to also bail when document.activeElement sits
inside a custom widget with role combobox/listbox/radio/switch/menuitem/tab
so single-letter shortcuts like '[' do not collapse the sidebar while
focused on SegmentedButton, ZonePicker, etc.
- Move the transcript font-size CSS variable write from documentElement
to document.body. Consumers inherit body styles, and scoping the write
there avoids invalidating root-level styles every time fontSize changes.
Record button bumped from 40x40 to 48x48 (WCAG 2.5.5 AA requires 44x44; 48 chosen as forgivable oversize for the most-touched control). Inner stop-square and record-circle bumped 14x14 to 16x16, Loader2 16 to 18, to keep the visual ratio.
Recording-status sr-only live region now announces both transitions: "Recording started" on start and "Recording stopped" on stop, with the stop message clearing after 1200ms so the live region does not hold stale text. Transition tracked via a plain let prevRecording (not $state) since it is bookkeeping and must not retrigger the effect. aria-atomic="true" added so the whole region re-reads on update.
Brand guidelines (docs/brand/magnotia-brand-guidelines.md §4) say "Never go
below 12px for any text" and "tertiary text must be ≥18px bold or ≥24px
regular". Audit found 246 sub-12px className sites and 166 cases of
text-text-tertiary paired with body sizes.
Bumped text-[9/10/11px] → text-[12px] across 26 .svelte files in src/lib
and src/routes. Promoted text-text-tertiary → text-text-secondary at body
sizes (12-14px) where context allowed; left placeholder pseudo-states,
ternary-conditional inactive states, and line-through "done" states alone
(8 residual pairings, all decorative).
Affects neurodivergent users on 1366×768 budget hardware most directly.
svelte-check: 0 errors, 0 warnings.
Reduces first-run cognitive load. Save / Copy / Clear / Export now appear only
once the user has a transcript or is recording, sliding in via animate-fade-in
(brand --duration-ui + ease-out-quart). Template and Extract Tasks remain
always-visible because they shape transcription rather than acting on completed
content.
Empty state now leads with the catchphrase as a brand statement
(font-display italic 28px) with the hotkey hint demoted to body-font tertiary.
EmptyState gains an optional headline prop; callers without it render unchanged.
Phase 10b colour push: theme reads as Magnotia, not "subdued generic dark".
Same lightness band as Phase 10a (AA preserved), more chroma across accent,
status tokens, and zone surfaces.
Dark accent: #c98555 -> #d68450 (subtle/hover/glow recomputed).
Light accent: #a06a3e -> #a3683a.
Dark status: success #7ec89a -> #5fc28a, danger #e87171 -> #e85f5f,
warning #e8c86e -> #e8be4a.
Light status: success #2f7549 -> #1f7344, danger #a83838 -> #b32626,
warning #b89a3e -> #a08a1f.
Sensory zones pushed further apart so each zone reads as its own
environment: cave cools toward blue-grey, energy warms toward red-orange,
reset cools toward green. AA on body text verified for every Card surface
in both themes.
Stale accent rgba (201,133,85) replaced with new (214,132,80) across nine
files. --shadow-accent in colors_and_type.css mirrored.
No secondary tint token added: Magnotia's identity is amber-as-the-only-
meaningful-colour. Cheaper to add a second meaningful colour later if a
real need appears than to dilute the discipline now.
Default surfaces, text tokens, and borders untouched per scope.
Six changes, one coherent polish pass on SettingsPage:
1. Autostart toggle dedupe. The 27-line custom button that mirrored
Toggle's markup to dodge a bind:checked race is gone. Toggle's new
async onChange + snap-back covers the OS round-trip; setLaunchAtLogin
re-throws on failure so Toggle reverts checked. autostartSyncing
state retired.
2. AI Assistant Default/Advanced split. Cleanup preset and GPU
concurrency moved into a nested SettingsGroup title="Advanced",
closed by default. The Assistant page now leads with four
first-decision controls (tier, model picker, status, prewarm) and
parks the tuning knobs behind a disclosure. Search keywords cascade
from the Advanced group up to AI Assistant up to AI & Processing.
3. Vocabulary disclosure consistency. The hand-rolled Profiles and
Templates sub-panels (with their own ChevronRight buttons +
showProfiles/showTemplates state) are now nested SettingsGroups.
The count badges live in the description slot ("3 profiles ·
custom vocabulary..."). showProfiles, showTemplates, and the
ChevronRight import retired.
4. Model-download ETA in Settings. Both whisper and LLM download
chips now show "{percent}% · {bytes} / {total} · {duration} left",
reusing formatDuration from utils/time.ts. New formatBytes +
etaSecondsFromPercent helpers in SettingsPage; bytes pulled from
the existing emit payload (DownloadProgress snake_case for whisper,
done/total for the LLM event). FirstRunPage already does the
timestamp-based ETA; we mirror its shape.
5. Privacy badge in sticky header. Single chip "100% local · no
telemetry · no cloud" added next to the Settings heading using
bg-accent-subtle + text-text-tertiary. The longer About list
stays put as the deep-dive.
6. SettingsGroup icons on the seven top-level groups. Mic, BookOpen,
Type, Sparkles, SquareCheck, Clipboard, Sliders — paired with the
existing literal text titles per the icon-with-label brand rule.
Inner nested groups stay icon-free.
Verification: npm run check returns 0 errors, 0 warnings.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Both props default to current behaviour, so all existing call sites render
unchanged. Subtle uses bg-bg-elevated for quieter empty-state contexts;
danger uses bg-danger/5 + border-danger/30 for calm caution panels (not
red-alert loud, matching brand). Raised layers shadow-md on top of the
existing low-contrast shadow for floating or detached surfaces.
Title bumped to font-semibold so it scans first; description gets tracking-wide
plus mt-1 for cleaner rhythm. Chevron alignment shifted from mt-0.5 to mt-1 to
match the new spacing. Optional `icon` prop renders a Lucide component between
the chevron and the title block at 16x16, text-tertiary, no colour change on
open. Layout is preserved exactly when no icon is passed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extends the shared Toggle with three optional props while keeping the
existing bind:checked path untouched for current call sites.
- disabled: dims the control, sets aria-disabled, and early-returns on click.
- loading: forces a Loader2 spinner inside the thumb, applies cursor-wait
and aria-busy, and blocks interaction. Cursor-wait wins over disabled's
not-allowed when both are set, matching loading-precedence guidance.
- onChange(next): when supplied, replaces the immediate flip. Promise
returns drive an internal pending flag so the toggle renders loading
while the handler resolves; on resolve checked flips, on reject the
toggle snaps back. Synchronous handlers flip immediately after the call.
Brand motion preserved: 150ms duration, cubic-bezier(0.2, 0.8, 0.2, 1),
no transition-all.
- Add --color-overlay-dim token to app.css @theme (rgba(15,14,12,0.7) dark, rgba(26,24,22,0.55) light); mirror as --overlay-dim in design-system/colors_and_type.css.
- MorningTriageModal: replace inlined rgba scrim with var(--color-overlay-dim), drop backdrop-blur-sm (brand opposes glassmorphism by default), bump container to z-[60] so it sits above the .grain z-50 overlay.
- FilesPage: collapse two near-identical drop-zone blocks into one bordered region with conditional inner content; preserves isDragOver hover behaviour and aria region.
Two scoped fixes to HotkeyRecorder:
- Append "Esc to cancel." to the recording prompt so the existing Escape
handler is discoverable. srStatus already mentions Escape; no harmonisation
needed.
- Replace animate-pulse-warm on the captured state with a static
bg-success/10 + border-success/40 flash. pulse-warm is the
recording-active keyframe; reusing it for the saved state muddled the
visual vocabulary. The global * transition rule already animates
background-color and border-color over --duration-ui (150ms) with the
brand cubic-bezier ease, so no new keyframe is needed and prefers-reduced-motion
is honoured automatically. Also swapped transition-all for transition-colors
to comply with the no-transition-all brand-motion rule.
Native window.confirm() is OS-styled and breaks the warm-amber UI
tone with an abrupt brutalist modal. The brand voice is "calm,
informative, solution-first; never blame the user", and confirm() is
none of those.
Replaced both destructive paths (Clear All, Delete N selected) with
a two-click inline pattern: first click arms the trigger and morphs
it into "Delete X? [Confirm] [Cancel]" inline pills; auto-cancels
after 4s so a user who walked away doesn't return to a primed delete
button. No new modal, no new component, no new dependency.
Both timers are cleared in onDestroy so unmounting the page mid-arm
doesn't leak.
Side-stripe borders (border-left: 2px or 3px in an accent colour, with
the rest of the element having no border) are the textbook AI-coded-IDE
"selected-state" pattern. The brand vocabulary is colour-tint and
chevron, not stripe-and-fill.
Replaced eight sites:
- VirtualSegmentList active/match/default segment rows: bg-accent/10,
bg-warning/10, hover:bg-hover (drop the stripes; bg-warning bumped
/5 → /10 so it's visible without the stripe doing the work).
- viewer/+page.svelte: same pattern as VirtualSegmentList.
- TasksPage profile-list tabs: bg-accent/10 + font-medium for active,
hover:bg-hover for inactive. rounded-md added so the tint follows
the row outline.
- ToastViewport: replaced border-left + variant colour with full 1px
border + bg tint + border-color tint (color-mix() with the matching
semantic token at 35% / 10%). Also removed the orphan --moss /
--signal / --ember tokens — they were not defined in app.css and
fell back to hex literals.
- preview/+page.svelte: dropped the phase-coloured left stripe and
the borderColorClass derivation; the header already has a pulsing
dot / animated bars / spinner for each phase, so the stripe was
redundant.
Per the project's no-dashes feedback rule (see CORBEL-Main memory
feedback_no_dashes.md), em-dashes in user-visible copy are the single
biggest "AI wrote this" surface tell. Forty-two occurrences across ten
files: ten in SettingsPage's group descriptions and option labels
(audio device picker, vocabulary help, prompt placeholders, model
descriptions), four in FirstRunPage's setup steps, three in
DictationPage / FilesPage / Energy chip tooltips, plus the privacy
bullets ("100% offline, no Python required, no cloud, no accounts,
no telemetry"), the rejection toast, MicroSteps thumb labels, and the
shutdown ritual prompts.
Replacement rule: subordinate-clause em-dashes followed by lowercase
become commas; sentence-break em-dashes followed by capitals become
full stops. Code comments left intact.
One non-em-dash regression caught while reviewing: TasksPage's energy
filter "no selection" placeholder showed "—" as a literal label; the
script changed it to a comma which read as a UI bug. Replaced with
"Any" instead.
The settingsSearch state was declared in Phase 9c with a comment
describing the intended behaviour but no <input> bound to it. With 21
SettingsGroup instances and a 2,261-line page, finding any specific
control required scrolling and opening every group. For the
neurodivergent audience this is exactly the cognitive-load failure
the brand commits to avoiding.
Added a sticky search input at the top of the page (paired with the
"Settings" title in a single sticky header that bg-bg-matches the
page so groups don't bleed under it). Each SettingsGroup now passes
through searchMatches against its title, description, and a curated
keyword string. Parent groups inherit their children's keywords so a
search for "GPU" expands AI & Processing and the AI Assistant inner
group together.
Updated SettingsGroup to push prop changes to the DOM via $effect:
the native <details> element's open attribute is mutated by user
clicks outside Svelte's reactive graph, so a parent prop change
needs to force-sync. User-driven toggles where the prop value didn't
change keep their local state because $effect only re-runs on diff.
Verified live: baseline shows 3 groups open (Transcription, Terms &
profiles, Capture & export), "GPU" opens AI & Processing + AI
Assistant only (other 19 closed), "ritual" opens Tasks & Rituals +
Rituals only, X-button clear restores baseline.
Eighteen sites across SegmentedButton, Toggle, HotkeyRecorder,
ModelDownloader, FilesPage, and SettingsPage hardcoded the old amber
rgba(232,168,124,...) for focus rings, glows, and progress bars. After
the Phase 10a a11y darkening of --accent to #c98555, the inline shadows
no longer matched the accent fill they were paired with, so every
focused input rendered a glow at a noticeably different hue from its
border.
Replaced all of them with the current accent rgba(201,133,85,...).
Toggle component also had two motion violations: an active:scale-95
press-shrink and a cubic-bezier(0.16,1,0.3,1) (ease-out-expo) thumb
transition. Both replaced with the brand's --ease curve
(0.2,0.8,0.2,1) and the scale dropped, in line with the record-button
fix in the previous commit.
Removed redundant inline transition-duration overrides on FilesPage
Browse button (the global * rule already sets --duration-ui) and on
ModelDownloader's primary CTA (also dropped its active:scale-[0.97]).
Record button had active:scale-[0.93] transition-all, which contradicts
the brand motion guideline ("never bounce, slow, calm, deliberate"). The
record button is the most-touched control in the app; a 7% press-shrink
on it sets the wrong tone for everything else.
Removed the scale transform, removed the inline transition-duration
override (the global * rule in app.css already animates the named
properties at --duration-ui). Also updated the inline shadow rgba from
(232,168,124,0.3) to (201,133,85,0.3) so the glow tracks the new
a11y-darkened accent fill instead of the pre-Phase-10a amber.
Body had user-select:none globally so users could not copy error
strings, model paths, profile names, transcripts, status messages,
or any text into search or another tool. The brand essence is
"clarity without friction" and this was friction.
Body now defaults to user-select:text (the standard for text content),
and the chrome elements that should never be drag-selected (button,
summary, role=button|switch|tab|menuitem, data-no-select) opt out.
Verified live: body=text, button=none, paragraph=text (inherited).
The runtime app reads tokens from app.css @theme; this file ships values
to the buildless preview pages under design-system/preview/*.html.
Several values had drifted from the Phase 10a a11y darkening done in
app.css, so the previews showed brighter accents and lighter tertiary
text than what users actually see.
Aligned: dark --text-tertiary, dark --accent (+ hover/subtle/glow/
border-focus/shadow-accent), light --text-tertiary, light --accent
(+ hover/subtle/glow), light --success, light --danger.
Header banner now states the file's role explicitly so future edits
don't recreate the drift.
The Tailwind v4 @theme block defines --color-danger as the semantic error
token, and 15+ files across the codebase use text-danger consistently.
SettingsPage was the lone outlier with five text-error / border-error
sites (audio devices, profile delete, vocabulary, term delete, diagnostic
report). Without a --color-error token, those utilities resolved to the
default text colour, so error messages rendered in warm cream instead of
red. Verified live via Playwright probe before and after.
README LLM-formatting section now states four Qwen tiers (Qwen3.5 2B /
4B / 9B + Qwen3.6 27B) and the magnotia-llm crate row reflects the
four-tier registry. The whisper-ecosystem context doc gets the same
refresh and cites unsloth as the GGUF source.
Older roadmap and Phase-0 audit docs left untouched — they are dated
historical artefacts and rewriting them would muddy the audit trail.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Picks up the registry rename in the front-end and Tauri command layer:
- src/lib/types/app.ts: LlmModelIdStr now lists the four new ids
(qwen3_5_2b / qwen3_5_4b / qwen3_5_9b / qwen3_6_27b).
- src/lib/pages/SettingsPage.svelte: LLM_MODELS table rebuilt with
four tiers (Minimal / Standard / High / Maximum), matching subtitles
and download-size copy. selectedLlmModelId fallback, hardware-warning
thresholds, tier-availability check, and ensureRecommendedLlmTier
fallback all retargeted at the new ids. The Maximum tier surfaces a
64 GB / 24 GB warning so users with mid-range hardware see honest
expectations.
- src-tauri/src/commands/llm.rs and commands/tasks.rs: doc-comment
examples refreshed (Qwen3 4B → Qwen3.5 4B, Qwen3's tokenizer →
Qwen's tokenizer — the BPE family is shared).
- src/lib/stores/llmStatus.svelte.ts: chip-detail example updated.
cargo build --workspace clean. cargo test --workspace clean.
npx svelte-check reports one pre-existing error in vite.config.js
(unused @ts-expect-error directive, dates back to the original
scaffold commit 9926a42); not introduced here, out of scope to fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the three older Qwen3 variants with a four-tier ladder spanning
a wider hardware range:
- Qwen3_5_2B_Q4 (Minimal, 8 GB RAM, ~1.3 GB download)
- Qwen3_5_4B_Q4 (Standard, 16 GB RAM / 6 GB VRAM, ~2.7 GB) — DEFAULT
- Qwen3_5_9B_Q4 (High, 32 GB RAM / 12 GB VRAM, ~5.7 GB)
- Qwen3_6_27B_Q4 (Maximum, 64 GB RAM / 24 GB VRAM, ~17 GB)
All four GGUFs sourced from unsloth's HF org with pinned commit SHAs.
Sizes and SHA256 hashes verified against the live X-Linked-Etag /
X-Linked-Size headers on the LFS CDN. Q4_K_M quantisation throughout
(common sweet-spot for cleanup + task extraction).
recommend_tier rewritten to span four bands; default_tier moves from
the old 4B-Instruct-2507 to Qwen3.5 4B. The 27B Maximum tier honestly
needs 64 GB RAM to run without partial offload — surfaced in the
description string so the Settings UI can warn realistically.
In-tree smoke tests (smoke.rs, content_tags_smoke.rs) updated to
reference the new smallest tier so a developer's MAGNOTIA_LLM_TEST_MODEL
points at the cheapest GGUF to download. Crate description in
crates/llm/Cargo.toml refreshed to mention the new family.
NOTE (out of scope; not fixed): the size_bytes / sha256 / hf_url
methods could collapse into a single LlmModelMetadata table to remove
four parallel match arms. Layer 2 cleanup, separate session.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
