# Memory — whispassist > Generated: 2026-07-07 21:30:45 > Total memories: **75** > Breakdown: fact: 3, decision: 10, goal: 1, preference: 1, context: 3, event: 1, learning: 24, observation: 2, artifact: 25, error: 5 --- ## Instructions *Standing rules, constraints, and guidelines to always follow.* *No memories of this type.* --- ## Facts *Verified information, project status, and established truths.* ### parakeet-rs (altunenes) DOES have a genuine increm... parakeet-rs (altunenes) DOES have a genuine incremental streaming decode API for Parakeet-family models: ParakeetEOU and Nemotron structs thread real recurrent state between calls internally (EncoderCache: cache_last_channel/cache_last_time/cache_last_channel_len; decoder LSTM state_h/state_c; last_token) plus a 4s rolling audio ring buffer, so callers just feed sequential small chunks (160ms for EOU, 560ms for Nemotron) and get incremental partial text -- it is not naive re-chunking of a batch decoder. Source: github.com/altunenes/parakeet-rs src/parakeet_eou.rs and model_eou.rs, examples/streaming.rs (checked 2026-07-02). *Confidence: 0.95 | Status: active | Created: 2026-07-02T20:09:27* ### WhispAssist app icon/branding: real logo provided ... WhispAssist app icon/branding: real logo provided by user (paperclip mascot + purple speech-bubble-with-sparkles mark, WhispAssist wordmark). The app icon (title bar/taskbar/tray/installer) is cropped from just the small speech-bubble-with-sparkles mark, not the full marketing graphic or the paperclip mascot (too much fine detail to read at 16-32px). Updated twice: first a transparent-background crop, then a refined version on its own gradient purple background from a cleaner logo revision the user provided. Regenerated via ; that command generates iOS/Android/Appx/macOS outputs by default which must be deleted since WhispAssist is Windows-only (not referenced by tauri.conf.json). tray.png is a manual 32x32 export, not one of tauri icon's own output names. *Confidence: 1.0 | Status: active | Created: 2026-07-02T20:13:49* ### The EOU streaming variant used by parakeet-rs is a... The EOU streaming variant used by parakeet-rs is a genuinely separate ONNX export, not a mode of the batch Parakeet TDT model: it's NVIDIA's own nvidia/parakeet_realtime_eou_120m-v1 (120M params, cache-aware FastConformer encoder + LSTM decoder, 80-160ms chunks, English-only, no punctuation/casing, emits token). This is distinct from istupakov/parakeet-tdt-0.6b-v3-onnx (600M, the community ONNX conversion used for WhispAssist's originally-considered batch/full-file transcription path, which has the ~4-5min length limit). Both are downloaded separately; author's own code comment on reset_on_eou says 'I must admit that this is not work very well on my real world tests'. *Confidence: 0.9 | Status: active | Created: 2026-07-02T20:09:30* --- ## Decisions *Architectural choices, approach selections, and their rationale.* ### M4.4 (MS Graph calendar source, T8.9, FR-CAL-6) SH... M4.4 (MS Graph calendar source, T8.9, FR-CAL-6) SHIPPED 2026-07-07 on branch feature_chore_bug_005 (commits 83cb30e..7df76df + 1cf0be1). This LIFTS the 2026-07-02 'T8.9 on hold' decision — user explicitly asked to build it in this session, overriding the earlier hold. Implementation: calendar::GraphSource implementing CalendarSource, reusing sync::oauth's provider-agnostic PKCE/loopback machinery (added a 'graph-calendar' OAuth provider entry, made sync::resolve_access_token pub(crate) for cross-module reuse) rather than building new OAuth plumbing. This completes all of milestone M4 (M4.1 chunked upload, M4.2 multi-language, M4.3 Dropbox/Box, M4.4 Graph calendar). *Confidence: 0.95 | Status: active | Created: 2026-07-08T02:22:48* ### WhispAssist roadmap: T8.8 (at-rest encryption / va... WhispAssist roadmap: T8.8 (at-rest encryption / vault) and T8.9 (Microsoft Graph calendar) are ON HOLD per user decision (2026-07-02). Consequence: Phase 9c (T9.12 client-side encryption before upload) is blocked since it depends on the T8.8 vault - skip 9c for now. Building Phase 9 (remote sync) starting with 9a (WebDAV). *Confidence: 1.0 | Status: active | Created: 2026-07-03T01:01:49* ### WhispAssist GPU acceleration decision (2026-07): w... WhispAssist GPU acceleration decision (2026-07): whisper.cpp GPU transcription via VULKAN as primary cross-vendor path — ONE binary covers NVIDIA+AMD+Intel (whisper-rs 'vulkan' feature). whisper-rs GPU features: cuda (NVIDIA-only, CUDA toolkit at build), vulkan (cross-vendor, Vulkan SDK at build + vulkan-1.dll loader shipping with every GPU driver), hipblas (AMD/ROCm LINUX-ONLY, unusable on Windows), metal (Apple). KEY: whisper.cpp GPU backends are COMPILE-TIME/static (cannot download runtime on-demand like the ort load-dynamic NPU path); binary must be built with the feature. Vulkan is default GPU build; CUDA is optional NVIDIA-only turbo variant LATER (user: 'we will do cuda later'). Vulkan degrades to CPU if no GPU present. *Confidence: 1.0 | Status: active | Created: 2026-07-05T22:03:34* ### WhispAssist roadmap: T8.8 (at-rest encryption/vaul... WhispAssist roadmap: T8.8 (at-rest encryption/vault) and T8.9 (MS Graph calendar) ON HOLD per user (2026-07-02). Consequence: Phase 9c (T9.12 client-side encryption) blocked (needs T8.8 vault) - skip for now. Building Phase 9 remote sync starting with 9a WebDAV. *Confidence: 1.0 | Status: active | Created: 2026-07-03T01:02:11* ### In WhispAssist's calendar module, CalendarSource::... In WhispAssist's calendar module, CalendarSource::import() is a synchronous trait method (matches PstSource's blocking subprocess call). GraphSource (MS Graph, async reqwest HTTP) bridges into that sync signature via tauri::async_runtime::block_on inside fetch_events, and callers run the whole import() call inside tauri::async_runtime::spawn_blocking (same pattern commands.rs already used for PstSource's readpst subprocess) — avoids making CalendarSource async just for one source. Kept a separate WA_GRAPH_CALENDAR_BASE_URL env var (distinct from sync's WA_GRAPH_BASE_URL used by OneDriveTarget) so calendar and sync tests never race on the same process-global env var in cargo test. *Confidence: 0.9 | Status: active | Created: 2026-07-08T02:22:55* ### WhispAssist PST usability fixes (2026-07-06, commi... WhispAssist PST usability fixes (2026-07-06, commits 00de9d3/9d757be/75ec24d): (1) calendar path is now remembered in Settings (pst_last_path field) so the user doesn't re-browse every launch; (2) added an opt-in 'auto-sync on launch' checkbox (pst_auto_sync) that re-imports the remembered path once at startup - deliberately a ONE-SHOT pass mirroring the existing sync-job-resume pattern in lib.rs, NOT a periodic timer, because NFR-RES-1 ('no polling timers when idle') is enforced consistently everywhere else in this codebase (every startup task has a comment noting this). User asked about a periodic 'read frequency' option too; I declined to build that specific piece and explained the NFR-RES-1 conflict rather than silently building or silently dropping it. *Confidence: 0.95 | Status: active | Created: 2026-07-06T20:40:52* ### Completed a full UI/UX redesign of WhispAssist: se... Completed a full UI/UX redesign of WhispAssist: semantic CSS design-token system (light/dark, WCAG AA verified), all emoji/Unicode icons replaced with @lucide/svelte SVG icons, segmented Monitor/Sun/Moon theme toggle defaulting to system preference, custom theme-aware scrollbars. Informed by researching Granola and Meetily's UIs; kept WhispAssist's 3-pane layout since diarization/speaker-naming already beats both competitors. *Confidence: 0.95 | Status: active | Created: 2026-07-02T20:13:34* ### WhispAssist M1 feature-briefs decomposition (2026-... WhispAssist M1 feature-briefs decomposition (2026-07-06, docs-only, branch feature_chore_bug_004). KEY INSIGHT: M1 is heavily scaffolded already, do NOT re-create — IPC types FeatureBrief/FeatureBriefInfo/ContextExcerpt (models.rs), api.ts bindings createFeatureBrief/listFeatureBriefs/getFeatureBrief/setBriefExposed, the 4 commands registered in lib.rs, DB tables feature_briefs + mcp_access_log (migrations/0003_ai_mcp.sql), the briefs/.json schema (docs/03-data-model.md), and the FeatureBriefBuilder trait (docs/04-api-contracts.md) ALL EXIST; only the 4 command bodies return not_implemented(). REMAINING to build: (1) Store methods insert_feature_brief/list_feature_briefs/get_feature_brief_row/set_brief_exposed; (2) a new non-streaming LlmProvider::complete(system,user)->String primitive (mirrors suggest_tags; impl Ollama+OpenAiCompat, Anthropic later); (3) FeatureBriefBuilder in a new briefs module (strict '## Title/## Problem/## Desired Outcome/## Acceptance Criteria' prompt like RESPONSE_FORMAT_INSTRUCTIONS + parse_brief mirroring parse_summary); (4) the 4 command bodies (add state: State, Tauri injects it, api.ts unchanged); (5) UI in SummaryPanel; (6) golden-transcript tests. KEY DESIGN DECISIONS: context_excerpts are VERBATIM transcript substrings (grounding invariant enforced by the golden test), selected by keyword overlap, NOT model paraphrase; the on-disk file is a sealed BriefFile envelope {schema,generated_at,provider,model,...fields,source} indexed by the feature_briefs table (file=truth, row=index); write file+row only AFTER a successful distill (no partial artifacts on LLM failure). Full implementation-ready checklist in docs/05-roadmap.md M1; JSON schema in docs/03-data-model.md; tests in docs/06-test-strategy.md P10. *Confidence: 1.0 | Status: active | Created: 2026-07-07T04:28:36* ### Decision: NOT pursuing NVIDIA Parakeet or DirectML... Decision: NOT pursuing NVIDIA Parakeet or DirectML/NPU acceleration for WhispAssist transcription near-term, as of 2026-07-02. Researched achetronic/parakeet (Go, Linux-only, dead end) and altunenes/parakeet-rs (Rust, built on ort, has genuine streaming via a separate EOU 120M model with real internally-threaded encoder cache and LSTM decoder state). Reasons not to pursue now: (1) DirectML support for this model family is unproven - zero reports of anyone running it, and there is a live unresolved ONNX Runtime bug (microsoft/onnxruntime issue 19837) producing wrong output on DirectML for the exact LSTM+Einsum op combination these models use; (2) DirectML itself is now in Microsoft maintenance mode, with new NPU/GPU work moving to Windows ML instead, which calls WhispAssist's existing ADR-0004 (ort + DirectML for NPU) into question independent of Parakeet; (3) Parakeet streaming needs a continuous per-meeting state machine fed small sequential chunks, fundamentally incompatible with WhispAssist's current stateless independent 4-second-window architecture - a real rearchitecture, not a swap, and it would lose whisper.cpp's crash-recoverable-per-window property; (4) competitor Meetily does not actually do live Parakeet+hardware-acceleration either - they only use it for offline batch re-transcription. Recommended future path if revisited: CPU-only Parakeet-EOU streaming spike first to validate the rearchitecture and quality tradeoffs (EOU is English-only, no punctuation/capitalization), treat NPU acceleration as a separate track that should probably target Windows ML rather than raw DirectML. *Confidence: 0.9 | Status: active | Created: 2026-07-02T20:14:37* ### WhispAssist CUDA + AMD-non-Vulkan acceleration PLA... WhispAssist CUDA + AMD-non-Vulkan acceleration PLAN (2026-07-05): whisper.cpp bakes in exactly ONE GPU backend at compile time. Universal build=Vulkan (all vendors). NEW: NVIDIA-turbo installer variant = whisper.cpp CUDA (feature already declared Cargo.toml:94) — but a CUDA build has NO Vulkan, so AMD/Intel GPUs in that variant need a non-Vulkan accel path. DECISION (recommended, user was away/didn't confirm): AMD-non-Vulkan = DirectML via the EXISTING ort/ONNX NPU path (OnnxNpuTranscriber generalized to accept EP: OpenVINO-NPU vs DirectML-GPU+device_id). DirectML works on any DX12 GPU, load-dynamic (no new build toolchain), one binary, degrades to CPU. hipBLAS/ROCm DEFERRED (narrow gfx coverage, fragile on Windows, needs 3rd installer). CORE CODE CHANGE: add AccelPath enum {WhisperCuda,WhisperVulkan,WhisperCpu,OnnxOpenVino,OnnxDirectML} + resolver in hardware/mod.rs from (BackendId+cfg!(feature)+runtime readiness); transcriber factory switches on it. This also CLOSES the known detection-honesty gap (GPU marked available from DXGI regardless of compiled feature -> no-op GPU routing). PHASES: P0 detection-honesty+AccelPath (small, no hw, do first), P2 AMD/Intel DirectML (validate on Intel Arc iGPU locally), P1 CUDA (needs NVIDIA hw/CI). ort needs 'directml' feature + an onnxruntime.dll built with DML EP in the runtime bundle; DirectML.dll ships with Win10 1903+. *Confidence: 0.85 | Status: active | Created: 2026-07-05T23:28:58* --- ## Goals *Objectives, targets, and milestones to track progress.* ### WhispAssist NEXT STEPS — CPU transcription slownes... WhispAssist NEXT STEPS — CPU transcription slowness testing round (open bug): the full transcribe_file path takes ~90s for a 6.3s clip on a 14-thread CPU with base.en (should be ~2-3s). This is PRE-EXISTING (predates Vulkan) and independent of the GPU work. To root-cause next: (a) verify whisper set_n_threads actually applies available_threads()=14 (log n_threads in run_full); (b) check whether the encoder pays full 1500 audio_ctx per 30s window even for a 6s clip in the non-streaming path (the streaming path fix in commit 8530a22 scaled audio_ctx by window length, but transcribe_file/run_full may not); (c) run stock whisper-cli directly on the same model+wav to isolate whether it's OUR run_full config vs whisper.cpp itself; (d) test greedy vs beam params and q5_1 vs f16 model; (e) confirm it's not thermal/throttle. The non-vulkan CPU baseline re-run (task, ~90s expected) was in progress to formally confirm parity with the vulkan build's CPU number. *Confidence: 1.0 | Status: active | Created: 2026-07-05T22:04:11* --- ## Commitments *Promises, obligations, and TODOs that need follow-through.* *No memories of this type.* --- ## Preferences *User and entity preferences for personalization.* ### WhispAssist user working style (observed 2026-07-0... WhispAssist user working style (observed 2026-07-06): (1) Demands EMPIRICAL PROOF over theory. When I attributed silent new recordings to the mic-not-in-WAV design, they pushed back ('the old file is also vault-sealed and plays fine, so it must be the 32->16bit change'). Resolving it required an actual real-hardware loopback capture test (play a known sound, read peak i16 from the WAV) to prove the 16-bit path records real audio — only then accept the diagnosis. HOW TO APPLY: when diagnosing a bug, verify the cause with a runnable test/measurement and show the evidence; don't just assert a root cause. (2) Highly protective of encryption-at-rest. They independently spotted that the decrypted audio.play.wav on disk and the browser 'download' button undermined the vault, and asked to switch playback to in-memory on-the-fly decryption. HOW TO APPLY: proactively avoid writing plaintext of vault-sealed data to disk and close off easy exfiltration paths (downloads, temp files). *Confidence: 0.9 | Status: active | Created: 2026-07-07T01:05:30* --- ## Relationships *Entity connections, team context, and collaboration patterns.* *No memories of this type.* --- ## Context *Session summaries, status updates, and conversation state.* ### WhispAssist release/versioning process (as of v0.2... WhispAssist release/versioning process (as of v0.2.0, 2026-07-06): the version string lives in THREE files that must be bumped together — package.json, src-tauri/tauri.conf.json, src-tauri/Cargo.toml (Cargo.lock updates on build). The shipped UNIVERSAL installer is built with: npm run tauri build -- --features vulkan --config src-tauri/tauri.vulkan.conf.json, with env VULKAN_SDK=C:\VulkanSDK\1.4.350.0, CMAKE_GENERATOR=Ninja, CARGO_TARGET_DIR=C:\wt, and vcvars64.bat loaded from 'C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Auxiliary\Build\vcvars64.bat'. Outputs land at C:\wt\release\bundle\msi\WhispAssist__x64_en-US.msi and \nsis\WhispAssist__x64-setup.exe (MSI ~44MB, NSIS ~9MB). The build does NOT create/push git tags — after merging to main, tag manually: git tag v; git push origin v. Last release before 0.2.0 was v0.1.6 (PR #15). *Confidence: 1.0 | Status: active | Created: 2026-07-07T01:05:16* ### Project status as of 2026-07-02: Phase 8 tasks T8.... Project status as of 2026-07-02: Phase 8 tasks T8.7 (multi-language transcription + i18n scaffold) and T8.8 (at-rest encryption/vault, needs an ADR decision on SQLCipher vs file-level encryption first) remain deferred/pending. User paused them to do a full UI/UX redesign, then a transcription-latency investigation and fix, then Parakeet/NPU research. Both T8.7 and T8.8 are still the next planned work whenever the user returns to the Phase 8 roadmap. *Confidence: 0.9 | Status: active | Created: 2026-07-02T20:14:43* ### WhispAssist remaining-work assessment + plan (2026... WhispAssist remaining-work assessment + plan (2026-07-06, docs-only planning on branch feature_chore_bug_004, NOT built). As of v0.2.0, Phases 1-9 ship. The only real gap is Phase 10 (external AI + agent handoff, ADR-0011) plus a few polish items. STUBBED (commands return Err(not_implemented)): create/list/get_feature_brief, set_brief_exposed, mcp_status, set_mcp_enabled, set_mcp_scope, mcp_access_log, run_agent, create_issue_from_brief. SKELETON/ABSENT: mcp/mod.rs (trait + todo!() only), AnthropicProvider (returns 'not built yet'), MS Graph CalendarSource, multi-language transcription, DropboxTarget/BoxTarget (OAuth wired in sync/oauth.rs but no upload impl), chunked/resumable upload. BUILT already: OpenAiCompatProvider. The phased plan lives in docs/05-roadmap.md section 'Remaining work — post-v0.2.0 execution plan': M1 feature briefs (first; LLM-only, no egress) -> M2 MCP server (the differentiator; inbound loopback, zero added egress, FR-MCP-7 egress-unchanged is the merge gate) ; M3 hosted AI (parallel; finish Anthropic + hosted-key/banner/allowlist) ; M4 reliability/breadth (chunked upload is TOP item because recordings are now ~50-100MB and put() buffers whole file, OneDrive/Graph caps PUT at 250MB; then multi-language, Dropbox/Box) ; M5 push handoff (agent runner + issue tracker, Could-priority, last/optional). MS Graph calendar deferred (Could). *Confidence: 1.0 | Status: active | Created: 2026-07-07T04:28:28* --- ## Events *Important conversations, milestones, and temporal occurrences.* ### WhispAssist reached v0.2.0 (2026-07-06), supersedi... WhispAssist reached v0.2.0 (2026-07-06), superseding v0.1.6. Additions since 0.1.6: microphone capture (records+transcribes the user's voice, mixed into transcript AND the saved recording at native quality via MicBridge), in-app recording playback with in-memory on-the-fly decryption (waaudio:// custom protocol, no plaintext on disk, download disabled), AI-generated tags + chip tag editor + tag filtering, meeting rename (editable title), notes single-pane Editor/Preview toggle, cancel-a-recording, 16-bit half-size recordings, live per-item sync upload progress, audio output + microphone device pickers, Outlook .pst recurring-event import/filtering + auto-sync-on-launch, and a SINGLE universal Vulkan installer (bundled vulkan-1.dll) replacing the separate CPU/NPU vs Vulkan builds. Built on branch feature_chore_bug_003 — still needs merge to main + tag v0.2.0. Installers already produced at C:\wt\release\bundle. *Confidence: 1.0 | Status: active | Created: 2026-07-07T01:05:36* --- ## Learnings *Knowledge acquired from experience, corrections, and insights.* ### Real libpst quirk found via a real 7.2GB mailbox t... Real libpst quirk found via a real 7.2GB mailbox test (2026-07-06): this readpst/libpst build joins a multi-value RRULE BYDAY with semicolons instead of RFC 5545's commas, e.g. 'RRULE:FREQ=WEEKLY;COUNT=10;BYDAY=MO;TU;WE;TH;FR' - so TU/WE/TH/FR appear as bare semicolon-separated tokens with no '='. A naive RRULE parser using '?' on split_once('=') silently bails out (returns None) for any event with more than one weekday, which is why the first version of expand_rrule worked for single-BYDAY series but silently dropped recurrence for multi-weekday ones. Fix: track the last-seen key and attribute a bare (no '=') token to it as a continuation value. This joins other already-documented libpst 0.6.63 quirks in ADR-0008 (no ORGANIZER/ATTENDEE emitted, no -8 flag support, wrong -t usage string). *Confidence: 1.0 | Status: active | Created: 2026-07-06T20:40:50* ### WhispAssist WASAPI microphone finding (2026-07-06)... WhispAssist WASAPI microphone finding (2026-07-06): unlike get_default_device(Render) which FAILS in cargo test on this dev machine, capturing the DEFAULT MIC (Direction::Capture) DOES work under cargo test AND delivers real frames (~3840 16kHz frames in 0.6s). Caveat: the FIRST COM activation of the mic can deliver 0 frames within the first ~600ms (cold start); a second run delivers normally. So a hardware mic smoke test should assert open+stop succeed (summary.sample_rate>0), not frames>0. Test: audio::tests::microphone_capture_opens_and_stops_cleanly (#[ignore], run with --ignored). *Confidence: 0.9 | Status: active | Created: 2026-07-06T21:34:39* ### CONFIRMED FIX for WhispAssist release build crashe... CONFIRMED FIX for WhispAssist release build crashes: Windows Defender real-time scanning of src-tauri/target was corrupting rustc.exe's compilation (random STATUS_STACK_BUFFER_OVERRUN crashes on different crates each run). Adding Defender exclusions (Add-MpPreference -ExclusionPath for target/, ~/.cargo, ~/.rustup, and -ExclusionProcess for rustc.exe/cargo.exe) fixed it completely -- full release build (whisper-rs, sherpa-rs native deps, MSI+NSIS bundling) now succeeds cleanly with the project's normal aggressive release profile (opt-level=z, codegen-units=1, lto=true). No toolchain reinstall or profile change was needed after all; those were red herrings from earlier in the debugging session. Also: a leftover running whispassist.exe instance can block cargo from overwriting the binary with 'Access is denied' -- close it before rebuilding. *Confidence: 0.95 | Status: active | Created: 2026-07-01T21:09:24* ### Dark-mode bug pattern learned: CSS custom properti... Dark-mode bug pattern learned: CSS custom properties don't cascade upward to ancestor elements. If data-theme (or similar theme attribute) is only set on an inner .app div, html/body keep the browser's default white background + 8px UA-stylesheet margin, invisible in light mode but a bright white border around the whole window in dark mode. Fix: mirror the theme attribute onto document.documentElement via an effect, and add a global html/body margin:0 + background:var(--bg) rule. *Confidence: 0.9 | Status: active | Created: 2026-07-02T20:13:41* ### WhispAssist NPU spike (T3.4): Intel AI Boost NPU (... WhispAssist NPU spike (T3.4): Intel AI Boost NPU (Core Ultra 5 135U, Meteor Lake, PCI VEN_8086&DEV_7D1D) runs the Whisper base.en ONNX encoder via ONNX Runtime OpenVINO EP (device_type=NPU) at ~66 ms/window vs ~236 ms/window on CPU = 3.58x faster, correct output shape (1,1500,512), full op coverage. Validates the plan: offload the fixed-shape Whisper encoder to the NPU, keep the dynamic decoder on CPU. *Confidence: 0.95 | Status: active | Created: 2026-07-02T20:31:08* ### WhispAssist: an OAuth-linked calendar source (MS G... WhispAssist: an OAuth-linked calendar source (MS Graph) is stored in Settings (settings.json: graph_calendar_enabled + graph_calendar_credential_ref) rather than as a sync_targets DB row, because it is not a SyncTarget/upload destination — reusing add_sync_target's OAuth flow would have wrongly surfaced it in the Sync UI and risked the upload pump trying to build a SyncTarget for it. begin_graph_calendar_link is a separate command from begin_oauth_link for this reason, duplicating ~60 lines of PKCE handshake rather than sharing it, since the two flows diverge in storage/eventing. *Confidence: 0.85 | Status: active | Created: 2026-07-08T02:23:03* ### cargo fmt -- does not scope forma... cargo fmt -- does not scope formatting to those files in the WhispAssist repo (src-tauri) — it reformats the entire crate regardless of file args passed after --, pulling in unrelated pre-existing drift in untouched files (observed: src/audio/mod.rs). After running cargo fmt scoped to touched files, always git status/diff to catch and git checkout -- any unrelated files it touched before committing. Also: memanto's on-prem backend (localhost:8080) needs Ollama (localhost:11434, embedding model nomic-embed-text) running for recall/export/sync to work, and the active agent session (memanto agent activate whispassist) can expire/drop mid-session — 'remember' can report success even when the write doesn't actually persist/index, so verify with 'memanto recall --recent' after a batch of remember calls rather than trusting the success message alone. *Confidence: 0.9 | Status: active | Created: 2026-07-08T02:23:10* ### WhispAssist runtime bundle URLs CORRECTED to gitea... WhispAssist runtime bundle URLs CORRECTED to gitea /media/ path (2026-07-06, commit 224aaf9): the .7z runtime bundles are hosted via Git LFS, and gitea's /raw/ endpoint returns the LFS POINTER (text) not the file, so the download URLs were changed from /raw/branch/main/ to /media/branch/main/ (gitea's media endpoint resolves LFS objects). Final URLs: NPU=https://git.dou.bet/iamdoubz/WhispAssist/media/branch/main/runtime/openvino.7z, DirectML=https://git.dou.bet/iamdoubz/WhispAssist/media/branch/main/runtime/directml.7z. SHAs unchanged (openvino ca0be9fc..., directml 34369222...). Files tracked via Git LFS (.gitattributes: runtime/*.7z filter=lfs). Still on branch chore_debug (no PR yet); URLs point at main so they resolve after merge. GOTCHA for future: GitHub raw.githubusercontent AND gitea /raw/ both serve LFS pointers not content — always use gitea /media/ (or GitHub media/LFS URL) for LFS-backed download targets. *Confidence: 1.0 | Status: active | Created: 2026-07-06T14:45:20* ### cargo test in WhispAssist src-tauri reliably crash... cargo test in WhispAssist src-tauri reliably crashes at LINK time (rustc.exe exit 0xc0000409 STATUS_STACK_BUFFER_OVERRUN) when linking the full debug test binary against whisper.cpp+sherpa-onnx native libs with full debug info (debuginfo=2 default for test profile). This reproduces even from a clean target/debug, with reduced --jobs, regardless of Defender exclusions (which fixed the earlier release-build crashes but not this). FIX: set env var CARGO_PROFILE_TEST_DEBUG=0 (drop debug info) before cargo test -- this shrinks the PDB/link footprint enough to avoid the crash. Confirmed working: 'cargo test privacy_self_check' passed cleanly with CARGO_PROFILE_TEST_DEBUG=0 --jobs 4. Also noted: commands.rs is NOT feature-gated for cpu-transcription/diarization (unconditionally imports SherpaDiarizer/WhisperTranscriber/run_streaming_worker), so cargo test --no-default-features fails to compile -- can't lighten the native-link footprint that way, must use the debug-info trick instead. *Confidence: 0.95 | Status: active | Created: 2026-07-01T22:01:31* ### WhispAssist dark-theme native-control gotcha (2026... WhispAssist dark-theme native-control gotcha (2026-07-06, commits 4442f80/8038918): the app theme is a MANUAL toggle (data-theme on html + .app), independent of the OS prefers-color-scheme. Native form controls default to LIGHT and render bright-white in dark mode. Fixes applied: (1) added CSS 'color-scheme: light' on :global(:root) and 'color-scheme: dark' on :global([data-theme="dark"]) in App.svelte — this alone fixed unstyled controls. (2) A