544 lines
87 KiB
Markdown
544 lines
87 KiB
Markdown
# 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 <EOU> 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/<id>.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<AppState>, 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_<ver>_x64_en-US.msi and \nsis\WhispAssist_<ver>_x64-setup.exe (MSI ~44MB, NSIS ~9MB). The build does NOT create/push git tags — after merging to main, tag manually: git tag v<ver>; git push origin v<ver>. 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 -- <specific files> does not scope forma...
|
|
|
|
cargo fmt -- <specific files> 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 <textarea> was white because Settings.svelte's 'input, select { background:var(--bg); color:var(--fg) }' rule EXCLUDED textarea; fix = add textarea to that selector. (3) The header template <select> (.theme-select) options popup stayed WHITE even with color-scheme:dark because it had 'background: transparent' — an AUTHOR-styled select makes Chromium/WebView2 render its options popup in light unless the options carry their own colors. Fix = give .theme-select an explicit 'background: var(--bg-elevated)' AND style '.theme-select option { background: var(--bg-elevated); color: var(--fg) }'. LESSON: for dark-mode selects, set color-scheme on the root AND author the <option> background/color with theme tokens; don't rely on transparent backgrounds.
|
|
|
|
*Confidence: 0.95 | Status: active | Created: 2026-07-06T13:02:41*
|
|
|
|
### WhispAssist test-environment quirk found 2026-07-0...
|
|
|
|
WhispAssist test-environment quirk found 2026-07-06: wasapi::get_default_device(&Direction::Render) (and thus find_render_device(None)) FAILS when called from within 'cargo test' on this dev machine, even though the real desktop app (running as a normal foreground GUI process) resolves it fine. Enumeration itself (DeviceCollection) works fine in both contexts -- it's specifically the 'default device' role query that needs a real interactive audio session. Tests for this were written to assert consistency (find_render_device(None) vs a raw wasapi::get_default_device call, or unknown-id fallback vs None) rather than assuming a default device is resolvable, so they pass in both environments. Implication: don't trust 'cargo test' alone to validate anything touching wasapi's default-device APIs -- verify audio-device-selection behavior via the actual running app instead.
|
|
|
|
*Confidence: 0.9 | Status: active | Created: 2026-07-06T21:02:54*
|
|
|
|
### WhispAssist NPU Rust recipe (T3.4, validated): ort...
|
|
|
|
WhispAssist NPU Rust recipe (T3.4, validated): ort crate v2.0.0-rc.10 with features [load-dynamic, openvino] links against Intel's pip onnxruntime-openvino DLLs and runs the Whisper base.en encoder on the Intel NPU at 65 ms/window (matches Python; CPU is 236ms). Recipe: (1) ORT_DYLIB_PATH -> site-packages/onnxruntime/capi/onnxruntime.dll (the OpenVINO-enabled ORT build); (2) prepend BOTH openvino/libs AND onnxruntime/capi to PATH so dependent DLLs (openvino.dll, onnxruntime_providers_openvino.dll, onnxruntime_providers_shared.dll) resolve; (3) OpenVINOExecutionProvider::default().with_device_type("NPU").build().error_on_failure() to make a failed NPU registration LOUD instead of silently falling back to CPU; (4) Session::run needs &mut session. For shipping, bundle these DLLs with the Tauri app (resource/sidecar) instead of relying on pip. load-dynamic means no C++/OpenVINO build step in cargo.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-02T20:42:31*
|
|
|
|
### DirectML support in parakeet-rs for NVIDIA Parakee...
|
|
|
|
DirectML support in parakeet-rs for NVIDIA Parakeet/EOU/Nemotron models is purely theoretical/unvalidated, not a proven working combination. Evidence: (1) parakeet-rs's own Cargo.toml/execution.rs just forwards to ort's generic directml feature with zero model-specific notes -- contrast with its explicit CoreML warning ('CoreML EP currently runs slower than CPU for Sortformer/Parakeet models because the ONNX graphs have dynamic input shapes'); no equivalent DirectML note exists. (2) Searched all 113 issues in altunenes/parakeet-rs GitHub repo: zero mention DirectML. (3) microsoft/onnxruntime issue #19837 (opened 2024, still unresolved as of check) reports DirectML EP producing wrong numeric results on a model containing LSTM+Einsum ops -- root cause never found. (4) Microsoft's own microsoft/DirectML GitHub repo now carries a banner: DirectML is in maintenance/sustained-engineering mode, with new feature development moved to Windows ML (WinML); relevant since WhispAssist ADR-0004 specifies ort+DirectML for NPU accel. Recommend flagging ADR-0004 for review given this shift.
|
|
|
|
*Confidence: 0.85 | Status: active | Created: 2026-07-02T20:09:34*
|
|
|
|
### Rebuilding WhispAssist release binary after Phase ...
|
|
|
|
Rebuilding WhispAssist release binary after Phase 6: initial 'tauri build' failed with 'only metadata stub found for rlib dependency core' / cannot find crate for std,num_traits (whisper-rs-sys build script, atoi). Root cause: stale/corrupted 19GB target/ dir from a prior interrupted build. Fix: cargo clean in src-tauri, then rebuild clean. Always load vcvars64.bat (VS2022 BuildTools) before cargo/tauri build.
|
|
|
|
*Confidence: 0.9 | Status: active | Created: 2026-07-01T20:27:51*
|
|
|
|
### Root cause of WhispAssist release build crashes fo...
|
|
|
|
Root cause of WhispAssist release build crashes found: NOT toolchain corruption. rustc.exe crashes with STATUS_STACK_BUFFER_OVERRUN (0xc0000409) specifically compiling the windows-rs crate (v0.61.3) and pxfm crate, reproducible under both rustc 1.94.1 and 1.96.1. Root cause is the project's aggressive release profile in src-tauri/Cargo.toml: opt-level='z' + codegen-units=1 + lto=true triggers an LLVM/rustc codegen crash on these large generated crates. Confirmed fix: overriding just opt-level=2, codegen-units=16 via CARGO_PROFILE_RELEASE_OPT_LEVEL/CARGO_PROFILE_RELEASE_CODEGEN_UNITS env vars lets the windows crate compile cleanly in isolation. Earlier 'toolchain corruption' and 'cargo clean' theories were red herrings -- the missing-std/core-prelude errors seen on other crates were a cascade effect of cargo continuing after the crashed crate's .rlib was never written.
|
|
|
|
*Confidence: 0.9 | Status: active | Created: 2026-07-01T20:40:50*
|
|
|
|
### WhispAssist recording fix + diagnosis (2026-07-06,...
|
|
|
|
WhispAssist recording fix + diagnosis (2026-07-06, branch feature_chore_bug_003): User reported new recordings had 'no sound' while old ones played. DIAGNOSED: NOT the 32->16bit change. Proved via ignored hardware test loopback_16bit_wav_captures_played_audio (played Windows Alarm01.wav through default device, captured peak i16=11670) that 16-bit loopback capture records real audio fine. Root cause: the recorded WAV was loopback-only, so a mic-only moment (user talking, nothing playing through speakers) recorded as silence while their voice still reached the transcript. FIX (user chose native-quality): added audio::MicBridge (AtomicU32 rate + Mutex<VecDeque<f32>>, cap ~0.5s for clock-drift): loopback thread publishes its rate + pulls mic samples per-frame and mixes into every channel in write_wav_bytes(mic:&[f32]); mic thread resamples its audio to loopback rate (Resampler::new_to) and pushes to the bridge. New WasapiCapture::start_loopback_recording / start_microphone_recording; commands.rs uses them with a shared bridge when mic enabled. Recording is now native rate/stereo 16-bit WITH the user's voice. Verified by ignored test loopback_recording_with_mic_bridge_captures_played_audio (peak i16=22358). This SUPERSEDES the earlier 'mic transcript-only, not in WAV' limitation.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T23:27:12*
|
|
|
|
### WhispAssist T3.4 steps 1-3 DONE & validated end-to...
|
|
|
|
WhispAssist T3.4 steps 1-3 DONE & validated end-to-end: OnnxNpuTranscriber transcribes speech correctly on the Intel NPU. Pipeline: hand-rolled Whisper log-mel (transcription/mel.rs, matches HF WhisperFeatureExtractor) -> NPU encoder (OpenVINO EP) -> CPU greedy decoder (no KV cache, re-feeds prefix) -> hand-rolled byte-BPE detok from tokenizer.json (no tokenizers crate). Behind cargo feature 'npu' = [dep:ort, dep:rustfft]; renamed from the old empty 'directml' feature. Decode config from generation_config.json: decoder_start=50257, eos=50256, forced_decoder_ids=[[1,50362]] (notimestamps); mask token ids >=50257 (except eot) in argmax. TTS test: spoke 'testing one two three four, the quick brown fox...' got 'testing 1234 the quick brown fox jumps over the lazy dog.' All gates green first try (clippy -D, fmt, 69 default tests, 77 npu tests). NOT YET wired into commands.rs dispatch (that is step 6) — nothing routes to it in-app yet.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-02T21:14:31*
|
|
|
|
### WhispAssist NPU build gotcha: ort's default downlo...
|
|
|
|
WhispAssist NPU build gotcha: ort's default download-binaries do NOT include the OpenVINO execution provider. Must supply an ONNX Runtime built with OpenVINO (Intel's prebuilt onnxruntime-openvino) PLUS the OpenVINO runtime DLLs on the DLL path. HARD VERSION PIN: onnxruntime-openvino 1.24.1 requires openvino runtime 2025.4.1 EXACTLY. Version mismatch (e.g. openvino 2026.2) does NOT error loudly — it silently falls back to CPUExecutionProvider (Win Error 127 'procedure could not be found'). Pin the ort<->onnxruntime<->openvino version triple and assert the active provider is OpenVINOExecutionProvider at load, else the NPU is silently unused.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-02T20:31:10*
|
|
|
|
### CONFIRMED ROOT CAUSE (2026-07-02) of the WhispAssi...
|
|
|
|
CONFIRMED ROOT CAUSE (2026-07-02) of the WhispAssist whisper.cpp model-load crash: it is Trend Micro Security Agent (Worry-Free Business Security, corporate-managed -- services ntrtscan/TMBMServer/TmCCSF/tmlisten all running) injecting into whispassist.exe and hooking file I/O / MessageBox APIs. Live cdb attach to the process while the 'Debug Assertion Failed: _osfile(fh) & FOPEN' dialog was showing revealed 'tmmon64' (Trend Micro's monitoring module) sitting directly in the call stack between USER32!MessageBoxW and ucrtbased!__acrt_MessageBoxW, and the read path (whisper.cpp's std::ifstream -> xsgetn -> fread) resolves into ucrtbased.dll (debug CRT) even though whisper-rs-sys's CMakeCache.txt confirms /MD (release CRT) was used to build it -- i.e. Trend Micro's hook is corrupting the CRT call path, not a real build misconfiguration. Ruled out first: NOT a stack-size issue (tried 16MiB worker thread stack, crash identical), NOT stale/corrupted build artifacts (crash reproduces identically from a fully clean cargo clean --profile dev rebuild), NOT Windows Defender (already excluded target/ and C:\Users\dadous\AppData\Local\WhispAssist, crash persisted). Fix requires excluding whispassist.exe / the WhispAssist install and model directories from Trend Micro's real-time scan and behavior monitoring -- likely needs corporate IT/policy admin involvement since TMBMServer implies tamper-protected central management, not a self-service local exclusion like Defender. Tooling note: installed WinDbg Preview via 'winget install --id Microsoft.WinDbg' -- ships cdbX64.exe (classic command-line debugger) alongside the modern WinDbgX.exe GUI, usable for live process attach analysis without needing the full Visual Studio IDE debugger.
|
|
|
|
*Confidence: 0.95 | Status: active | Created: 2026-07-02T13:30:42*
|
|
|
|
### Fixed a real bug behind reported 30-65 second live...
|
|
|
|
Fixed a real bug behind reported 30-65 second live-transcription lag: whisper.cpp's encoder always runs over a full padded 30-second mel window (1500 encoder positions) unless audio_ctx is explicitly reduced via set_audio_ctx. WhispAssist's live-transcription streaming windows are only about 4 seconds each but were never setting audio_ctx, so every window paid the full 30-second-equivalent encode cost, serially, on one worker thread. Fixed in src-tauri/src/transcription/mod.rs (function audio_ctx_for_window) by scaling audio_ctx proportionally to the real window length (1500 positions = 30s, so a 4s window gets about 201). Committed as 8530a22. Verified about 30 percent faster in a controlled A/B benchmark, though that specific test ran under heavy CPU contention from Docker Desktop and other concurrent Claude Code sessions on this machine, which likely masks a larger real-world improvement since the fix targets the encoder O(n^2)-ish attention cost specifically.
|
|
|
|
*Confidence: 0.9 | Status: active | Created: 2026-07-02T20:14:23*
|
|
|
|
### WhispAssist dev/test workflow gotchas (2026-07-06)...
|
|
|
|
WhispAssist dev/test workflow gotchas (2026-07-06): (1) A running 'npm run tauri dev' / whispassist.exe holds the cargo build lock on its target dir — STOP it (Stop-Process -Name whispassist, plus kill the node/cargo/vite procs whose CommandLine matches tauri|whispassist|vite) BEFORE running cargo build/clippy/test or the build blocks on the lock. (2) 'npm run tauri dev' writes output to the Windows console handle, NOT the redirected background-task log file (which stays empty) — confirm the app actually launched via Get-Process whispassist, not by reading the log; it typically appears ~30s after launch. (3) The audio-device tests find_render_device_none_matches_get_default_device and find_render_device_unknown_id_falls_back_exactly_like_none are FLAKY under cargo test (the wasapi get_default_device(Render) COM quirk) — a rerun passes; do NOT chase them as regressions. Loopback/mic hardware smoke tests (loopback_16bit_wav_captures_played_audio, etc.) are #[ignore]'d and run with --ignored.
|
|
|
|
*Confidence: 0.95 | Status: active | Created: 2026-07-07T01:05:20*
|
|
|
|
### WhispAssist tooling FOOTGUN (2026-07-06): 'npm run...
|
|
|
|
WhispAssist tooling FOOTGUN (2026-07-06): 'npm run format' = 'prettier --write .' — it reformats the ENTIRE repo, not just changed files. Running it during a small change reformatted 120+ files (all docs/ADRs/.claude skills/CLAUDE.md/stores/etc) because the repo isn't uniformly prettier-clean, burying the real diff. RECOVERY that worked: git diff --name-only | grep out the intended KEEP files | xargs git checkout -- , then verify only intended files remain; the reformats were content+EOL noise (git diff -w showed EOL-only for many). LESSON: to format/verify only your touched files use 'npx prettier --write <files>' or just 'npm run check' (svelte-check) + 'npx eslint <files>' which don't write. ALSO: 'npm run lint' currently reports ~143 PRE-EXISTING errors (mostly 'console'/'process' is not defined no-undef in node-context files) unrelated to app code — don't be alarmed, they predate any given change. Repo has mixed LF/CRLF (git warns 'LF will be replaced by CRLF'); harmless line-ending churn.
|
|
|
|
*Confidence: 0.95 | Status: active | Created: 2026-07-06T12:48:15*
|
|
|
|
### WhispAssist dev-run finding (2026-07-06): 'npm run...
|
|
|
|
WhispAssist dev-run finding (2026-07-06): 'npm run tauri dev' builds the Rust with features audio,cpu-transcription,diarization,pst,sync,npu = the DEFAULT set MINUS vulkan and cuda. Consequence: the everyday dev/default build has NEITHER a whisper.cpp GPU backend NOR CUDA, so on the Intel Core Ultra + Arc iGPU dev machine the ONLY GPU accel path is DirectML (via the ort/npu feature) — and directml_would_help() returns TRUE there, so the DirectML Settings card is visible. To exercise the Vulkan path you must build with --features vulkan explicitly (see whispassist-vulkan-build-recipe). Incremental dev rebuild ~42s once whisper.cpp/deps are cached; debug binary at src-tauri/target/debug/whispassist.exe. Launch recipe: load vcvars64.bat (VS2022 BuildTools) then 'npm run tauri dev'; the WebView2 window opens on the user's desktop. Opening Settings does NOT trigger the debug-CRT Abort/Retry/Ignore assertion dialog (that only fires on the transcription file-read path), so a UI-only visual check on the debug build is safe.
|
|
|
|
*Confidence: 0.9 | Status: active | Created: 2026-07-06T12:27:54*
|
|
|
|
### Windows dev-loop gotchas confirmed again this sess...
|
|
|
|
Windows dev-loop gotchas confirmed again this session (2026-07-06): (1) Git Bash's quoting of 'cmd.exe /c "...vcvars64.bat" && cargo ...' silently no-ops (just opens/closes an interactive cmd shell) - must run that exact vcvars64.bat wrapper via the PowerShell tool instead, never Bash; (2) both 'cargo fmt' (no path args) and 'npm run format' (prettier --write .) reformat the ENTIRE repo/workspace, not just touched files - this touched unrelated pre-existing files (commands.rs WebDavTarget chain, all of docs/, .claude/skills/, package.json, CLAUDE.md, package-lock.json) and had to be reverted via targeted git checkout, keeping only the intended diff. Going forward: use 'rustfmt --edition 2021 <specific files>' and 'npx prettier --write <specific files>' instead of the whole-repo commands.
|
|
|
|
*Confidence: 0.95 | Status: active | Created: 2026-07-06T20:40:55*
|
|
|
|
---
|
|
|
|
## Observations
|
|
|
|
*Patterns noticed, behavioral notes, and recurring themes.*
|
|
|
|
### WhispAssist transcription benchmark (2026-07, rele...
|
|
|
|
WhispAssist transcription benchmark (2026-07, release 0.1.4, SAME 6.3s TTS clip 'Testing 1234 the quick brown fox...', base.en q5_1 model, full transcribe_file path, machine=Intel Core Ultra 14-thread + Arc iGPU): VULKAN on Intel Arc iGPU = load 274ms, infer ~4.5s (fastest, edges out NPU). NPU OpenVINO = load 1355ms, infer ~5.7s. CPU whisper.cpp = load ~190ms, infer 92-99s (PATHOLOGICAL, ~15x SLOWER than real-time). All three produce the correct transcript. CRITICAL FINDING (user-confirmed): CPU was ALREADY ~90s BEFORE Vulkan was added — so the Vulkan build did NOT regress the CPU path; the 0.1.4 Vulkan build is SAFE to ship (CPU fallback unchanged). The ~90s CPU is a PRE-EXISTING bug in the full transcribe_file path, NOT caused by Vulkan. Vulkan and NPU are ~16-20x faster than the broken CPU path. Note the STREAMING path was already fixed earlier (audio_ctx scaling, commit 8530a22); transcribe_file (single_segment=false, 30s-padded seek loop) is the still-slow one.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-05T22:04:09*
|
|
|
|
### Meetily (Zackriya-Solutions/meetily), the competit...
|
|
|
|
Meetily (Zackriya-Solutions/meetily), the competitor app cited as prior art for Parakeet integration, uses Parakeet only for BATCH/offline transcription via the transcribe-rs crate (built on istupakov's ONNX conversion, per meetily's own README credits) -- not live streaming. Its 'Import & Enhance' feature (post-hoc re-transcription) is the actual use case; no DirectML-specific hardware acceleration for Parakeet is documented in its backend README. A true streaming fork (Nemotron streaming ASR engine) exists only as a community fork (Amitsurya2000/transcribe-rs), not upstream. sherpa-onnx (k2-fsa) also lacks true streaming Parakeet TDT support as of its open issues #2918 and #3573. Checked 2026-07-02, informs WhispAssist NPU/Parakeet research.
|
|
|
|
*Confidence: 0.85 | Status: active | Created: 2026-07-02T20:09:37*
|
|
|
|
---
|
|
|
|
## Artifacts
|
|
|
|
*Tool outputs, files, reports, and external references.*
|
|
|
|
### M4.1 (chunked/resumable upload, T9.2 refinement, F...
|
|
|
|
M4.1 (chunked/resumable upload, T9.2 refinement, FR-SYNC-2/5) SHIPPED 2026-07-07 on branch feature_chore_bug_005 (commits be025bb, 2cc48c9), same day as M4.2/M4.3/M4.4 but a separate prior session. Recordings are now native-quality (~50-100 MB) and the old put() buffered the whole file into memory in one PUT; OneDrive/Graph also caps a single PUT at 250 MB. Fixed by streaming disk-to-network in fixed-size chunks on both sync backends: WebDavTarget::put and OneDriveTarget::put now use tokio::fs::File + BufReader (O(chunk) memory, not O(file size)) for every upload. Files at/below LARGE_FILE_THRESHOLD (8 MiB) still take a single streamed PUT — only large artifacts (.wav recordings) take the chunked path. WebDAV (Nextcloud/ownCloud) implements the chunking-v2 protocol; OneDrive uses Graph's upload-session API. Needed enabling tokio's io-util/net features for AsyncReadExt/AsyncSeekExt/BufReader (separate chore commit be025bb).
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-08T02:27:25*
|
|
|
|
### WhispAssist calendar/PST recurrence expansion (202...
|
|
|
|
WhispAssist calendar/PST recurrence expansion (2026-07-06, commits fa8a7f2/6f07f28): expand_rrule in calendar/mod.rs now parses RRULE (DAILY/WEEKLY/MONTHLY/YEARLY, INTERVAL/COUNT/UNTIL/BYDAY/BYMONTHDAY/BYMONTH) and expands each recurring PST event into its own stored row keyed by '{uid}@{ymd}' for dedup, instead of only storing the first occurrence. Rewritten on chrono::Local (promoted from transitive to direct dependency) instead of hand-rolled epoch-day math, because the first version had a real DST bug: it kept a fixed UTC time-of-day per occurrence, so a meeting created in winter (CST) drifted an hour once its weekly recurrence crossed into summer (CDT) - e.g. 16:30 UTC showed correctly as 10:30 in January but wrongly as 11:30 in July. Fix: convert dtstart to local wall-clock once, keep hour/min/sec fixed, re-resolve the UTC offset per occurrence date. Tests assert local wall-clock time is identical across all occurrences (would fail under the old code).
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T20:40:48*
|
|
|
|
### WhispAssist audio-device-selection feature (2026-0...
|
|
|
|
WhispAssist audio-device-selection feature (2026-07-06, commits bb1d173/8e3e22c/2759a0e/5b28153/1376059/38b9cb2/b8dc5fc): added Settings > Hardware > 'Audio Devices' picker overriding the default 'Default system audio' WASAPI loopback render device. Backend: AudioCapture::start now takes device_id: Option<&str> (Device::get_id() string) instead of always resolving wasapi::get_default_device(&Direction::Render); new find_render_device() enumerates via wasapi::DeviceCollection and falls back to system default if the configured device is gone (same degrade-gracefully spirit as the existing mid-recording reconnect, which now retries the SAME selection first instead of switching to whatever's currently default). New list_render_devices()/list_audio_devices command (wasapi crate already supported enumeration, just was unused until now). Settings.audio_output_device: Option<String>, #[serde(default)] for backward compat. This is playback/render-device-only (loopback capture) -- WhispAssist has no microphone capture path at all (FR-CAP-1), so there is no separate 'input device' setting.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T21:02:52*
|
|
|
|
### WhispAssist M3 (hosted AI providers) SHIPPED (2026...
|
|
|
|
WhispAssist M3 (hosted AI providers) SHIPPED (2026-07-07, branch feature_chore_bug_005, merge commit e966181, on top of M1+M2 merge 2582d89). Built by an agent in an isolated git worktree, but that worktree's base snapshot was stale (pre-dated M1/M2 - based on 3038b9d, not the then-current feature_chore_bug_005 HEAD) - a real limitation of this session's worktree isolation mechanism worth remembering: it can silently reuse an earlier repo snapshot rather than the live current branch when a new worktree agent is launched later in the same session. Concretely this meant the M3 agent never saw M1's LlmProvider::complete() trait method addition, so AnthropicProvider::complete() was left as M1's not-yet-implemented stub even though M3 finished summarize/suggest_tags/status for real. Caught this by grepping the worktree for 'fn complete' before merging (found nothing) rather than trusting the agent's done report, then implemented AnthropicProvider::complete_with_key (non-streaming POST to /v1/messages, mirrors suggest_tags_with_key) myself as part of merge reconciliation. Also resolved three merge conflicts: llm/mod.rs (the complete() gap above), and two lucide-icon-import-list conflicts in Settings.svelte/SummaryPanel.svelte (M2 and M3 each added their own icon imports at the same insertion point) - unioned both, verified every icon is actually used before committing. What M3 built: AnthropicProvider real implementation (x-api-key + anthropic-version headers, SSE streaming for summarize, non-streaming for tags/complete), set_llm_provider storing the Anthropic key only in the OS credential store (never settings.json/DB, tested) and adding api.anthropic.com to the egress allowlist only when actually configured, HostedAiBanner.svelte one-time third-party-egress acknowledgment component, active-provider indicator + per-use quick-switch in SummaryPanel. Verified independently end to end on the fully merged M1+M2+M3 tree: cargo test --features mcp = 165 passed/0 failed, clippy clean (default + --features mcp), cargo fmt clean (only the same pre-existing unrelated audio/mod.rs drift as before), svelte-check 0 errors. Lesson for future milestone builds in this repo: after any worktree-isolated agent finishes, grep for the specific trait methods/functions the previous milestone added before trusting 'this builds on M<n-1>' claims - isolation snapshots can silently drift stale mid-session.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-07T13:16:21*
|
|
|
|
### WhispAssist meeting title + tags features (2026-07...
|
|
|
|
WhispAssist meeting title + tags features (2026-07-06, commits 2efe295/e68e888 and 68c32ff/16fcb15): (1) meetings previously had NO way to rename from 'Untitled meeting' anywhere in the UI - added an inline-editable title header above the transcript/notes pane (TranscriptNotes.svelte) backed by a new rename_meeting command; (2) attach_meeting_to_event now also mirrors the linked calendar event's subject onto the meeting's title when it has one, so linking 'Jerry / Daniel - Weekly 1:1' auto-renames the meeting; (3) added a 'Generate tags' feature mirroring 'Generate summary' - new LlmProvider::suggest_tags trait method (implemented for Ollama/OpenAI-compatible/Anthropic-stub), reads the transcript via the same build_prompt assembly, non-streamed, returns 1-8 tags merged into (not replacing) the existing tag list; (4) replaced the old comma-separated text-input tags UI with GitHub-topics-style removable/clickable chips (new shared TagChip.svelte component using existing --accent/--accent-soft tokens) - clicking a chip's label calls meetings.filterByTag() which filters the sidebar meeting list, an 'x' removes it.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T20:40:53*
|
|
|
|
### WhispAssist Phase 9a increment 1 DONE (branch feat...
|
|
|
|
WhispAssist Phase 9a increment 1 DONE (branch feature_another_one): WebDAV target management + connection test, validated end-to-end against a live wsgidav server (file physically uploaded via PROPFIND->MKCOL->PUT->HEAD). Added: storage SyncTargetRow CRUD (sqlx FromRow); sync::WebDavTarget real impl; sync::credentials keyring wrapper (service 'WhispAssist-sync', secret keyed by credential_ref, never in DB); TLS enforcement (enforce_transport: https always, http only for LAN+opt-in); commands list/add/update/remove/test_sync_target + set_sync_enabled (take State now); privacy_self_check wired to real targets. Enabled 'sync' in default cargo features. Anonymous targets supported (basic_auth only sent when a secret exists). All gates green: clippy -D, fmt, 74 tests. Local test server: python -m wsgidav.server.server_cli --host 127.0.0.1 --port 8899 --root <dir> --auth anonymous; test env WA_WEBDAV_URL/USER/PASS, run 'cargo test webdav_round_trip -- --ignored'. STILL TODO increment 2: durable queue + pump/backoff + sync_meeting + finalize-hook enqueue + sync://job events + sync_status/retry_sync_job (currently still not_implemented) + Settings sync UI. Increment 3: 9b OAuth (OneDrive/Dropbox/Box). 9c encryption on hold (needs T8.8 vault).
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-03T01:17:19*
|
|
|
|
### M4.2 (multi-language transcription, T8.7, FR-TRX-4...
|
|
|
|
M4.2 (multi-language transcription, T8.7, FR-TRX-4) SHIPPED 2026-07-07 on branch feature_chore_bug_005 (large commit chain: 6bc0949 add multilingual/language fields to ModelInfo+Settings, 37a6dd1 multilingual model catalog entries, 7f88101 whisper language catalog for Settings dropdown, 496a8aa wire whisper language param through Transcriber trait, 66d9eff persist requested language at meeting creation, 75ca3df track resolved per-meeting language on RecordingSession, 7490ff5 wire language selection through recording/reprocess/recovery, 34fff3f register list_whisper_languages command, plus a full UI chain (4c00496/4892719/dc969df/e6e995e/7bff392/353d582) for a Settings language picker + per-meeting badge + reprocess override, and a same-day bugfix 40709b2 making resolve_language normalize case-insensitive 'en' matches so the English-only-model guard fires correctly). Delivers: multilingual model option, whisper language param (select/auto-detect), per-meeting language persisted, Settings dropdown + reprocess picker in the UI.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-08T02:27:34*
|
|
|
|
### WhispAssist Phase 9a increment 2 DONE: durable upl...
|
|
|
|
WhispAssist Phase 9a increment 2 DONE: durable upload queue (sync_jobs CRUD, pump w/ backoff, finalize hook, startup pump, sync://job events). 78 tests green. TODO: Upload-now UI. Increment 3: 9b OAuth; 9c on hold.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-03T02:57:46*
|
|
|
|
### WhispAssist in-memory recording playback (2026-07-...
|
|
|
|
WhispAssist in-memory recording playback (2026-07-06, branch feature_chore_bug_003): Replaced the file-based player (which decrypted vault-sealed audio.wav to a plaintext audio.play.wav on disk, weakening encryption-at-rest) with an in-memory custom Tauri protocol. commands::serve_recording backs a registered 'waaudio' uri scheme (URL http://waaudio.localhost/<meeting_id> on Windows): reads audio.wav, vault::open decrypts in RAM, streams audio/wav with Range support (parse_byte_range) for seeking; 404 no file, 403 sealed+locked, path-traversal guarded (id must be alnum/hyphen). recording_playback_path now returns that URL after a cheap sealed-prefix + is_unlocked precheck and deletes stale audio.play.wav. commands::cleanup_playback_temp() sweeps all meetings/*/audio.play.wav at startup (called in lib.rs setup). Removed assetProtocol config + tauri protocol-asset feature + convertFileSrc; CSP media-src now 'self' http://waaudio.localhost. UI <audio> has controlsList=nodownload noplaybackrate + oncontextmenu preventDefault so the decrypted audio can't be saved to disk. Verified: clippy clean, svelte-check clean, full build, startup sweep removed leftover play-temp files.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T23:41:52*
|
|
|
|
### WhispAssist runtime bundles switched to 7z from re...
|
|
|
|
WhispAssist runtime bundles switched to 7z from repo raw URLs (2026-07-06, branch chore_debug, commits 27ba4f9/e6f1916). The NPU (OpenVINO) and DirectML runtimes are now downloaded as .7z (LZMA2+BCJ) from gitea RAW paths: NPU=https://git.dou.bet/iamdoubz/WhispAssist/raw/branch/main/runtime/openvino.7z (SHA ca0be9fc52c78ee623b152f790450b3d4020c5a7ebe99d27736455b308782191), DirectML=https://git.dou.bet/iamdoubz/WhispAssist/raw/branch/main/runtime/directml.7z (SHA 34369222fcc1be2e72a957b868b1976a90150ba704a06c9e8992c34ee368926b). REPLACED the zip crate with sevenz-rust2 (pinned 0.7.0 — newer needs rustc>1.77 MSRV; optional + npu-gated). extract_zip_flat -> extract_7z_flat (uses decompress_file_with_extract_fn, flattens basenames — archives nest DLLs under directml/ and openvino/ folders). stage_directml_runtime now defaults to DIRECTML_RUNTIME_URL const (no longer requires WA_DIRECTML_RUNTIME_URL env). Env overrides WA_NPU_RUNTIME_URL / WA_DIRECTML_RUNTIME_URL still honored. sevenz-rust2 0.7.0 confirmed to decode LZMA2+BCJ (has src/bcj/x86.rs); validated by test extract_7z_flat_unpacks_the_directml_bundle (extracts ../runtime/directml.7z, asserts onnxruntime.dll == 17253408 bytes, flattened) — PASSES. clippy -D warnings clean on default/shipped build; 104+ tests green. NOTE: raw URLs point to branch/main, so they only resolve once runtime/openvino.7z + runtime/directml.7z are committed to MAIN. As of now those .7z files are UNTRACKED (left for the user to commit — plain git add vs Git LFS decision; ~27MB total). The old .zip package-registry URL (git.dou.bet/api/packages/.../npu-runtime/...) is retired. archives were created with 7z LZMA2:24m BCJ.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T13:26:14*
|
|
|
|
### WhispAssist M1 (feature briefs) + M2 (MCP server) ...
|
|
|
|
WhispAssist M1 (feature briefs) + M2 (MCP server) SHIPPED (2026-07-07, branch feature_chore_bug_005, merge commit 2582d89). Built in parallel by two agents in isolated git worktrees, then merged sequentially (M1 first, ff-merge; M2 second, conflict-merged). M1: Store methods for feature_briefs (insert/list/get_row/set_exposed), LlmProvider::complete() primitive (Ollama+OpenAI-compat), FeatureBriefBuilder distiller (briefs/mod.rs, keyword-overlap grounding), the 4 create/list/get_feature_brief+set_brief_exposed commands, SummaryPanel.svelte UI, golden-transcript tests. M2: mcp module on rmcp (loopback bind + token gate, Streamable HTTP + stdio adapter), 4 tools (list_recent_meetings/get_transcript/get_action_items/get_feature_brief), scope control (none|selected|all), mcp_status/set_mcp_enabled lifecycle (token in OS credential store), disclosure UI + mcp_access_log audit, privacy panel integration. Gated behind an optional mcp Cargo feature. Merge required resolving one real conflict (storage/mod.rs trait+impl interleaving) plus one real cross-branch integration bug: M2's handler.rs called commands::get_feature_brief(id) with the pre-M1 stub arity; fixed by extracting get_feature_brief_core(store, id) shared between the Tauri command and the MCP handler, and wired the previously-stubbed selected-scope exposed-flag check (scope::brief_visible) that M2 had left as a documented KNOWN GAP. Verified independently (not just trusting agent reports): cargo test --features mcp = 154 passed/0 failed, clippy clean (default + --features mcp), cargo fmt clean (only pre-existing unrelated audio/mod.rs drift), svelte-check 0 errors. Known environment gotcha hit repeatedly during verification: whisper-rs-sys native build intermittently fails with MSVC error C1056 'cannot update time date stamp' - Trend Micro AV interference, same class as the previously-recorded model-loading hang; resolved by retrying the build, not a code defect. Also: building in a deeply-nested git worktree path (.claude/worktrees/agent-id/...) can overflow Windows MAX_PATH during whisper.cpp's CMake TryCompile scratch dirs - work around with a short CARGO_TARGET_DIR (e.g. C:/wa-build-x) when testing worktrees directly. M2's one incomplete acceptance item: no compiled end-to-end MCP wire-protocol client test (blocked on an rmcp reqwest-0.13-vs-0.12 version conflict pulling in aws-lc-rs; agent reverted the attempt cleanly rather than leave it unverified) - unit-level loopback-bind-refusal and scope-enforcement tests substitute for now. Next: M3 (hosted AI providers, Anthropic) per user pre-authorization to skip the usage checkpoint and go straight to a scheduled resume.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-07T06:17:52*
|
|
|
|
### M4.3 (Dropbox/Box upload targets, T9.10, FR-SYNC-9...
|
|
|
|
M4.3 (Dropbox/Box upload targets, T9.10, FR-SYNC-9) SHIPPED 2026-07-07 on branch feature_chore_bug_005 (commit 6991c1d), same day as M4.1/M4.2/M4.4 but a separate prior session. Implements DropboxTarget and BoxTarget SyncTarget impls (OAuth PKCE was already wired for both providers). DropboxTarget: path-addressed like OneDrive/WebDAV; create_folder_v2 creates the whole intermediate path in one call; upload-session chunking above LARGE_FILE_THRESHOLD (start/append_v2/finish). BoxTarget: Box addresses items by numeric ID not path, so ensure_dir/exists/put all walk (and lazily create) the folder chain from root ('0') by listing each level's children; always uploads via Box's session API regardless of file size (its session API takes plain PUT bodies, consistent with every other target, needs no new reqwest feature, and Box computes/returns each part's digest so no local hashing needed). Both follow the same documented ceiling as OneDriveTarget's put_chunked (M4.1): the upload-session id lives only for one put() call, not persisted across process restarts — a crash mid-chunked-upload restarts that file's upload from scratch rather than resuming (unlike WebDAV's chunking-v2 which is genuinely resumable).
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-08T02:27:46*
|
|
|
|
### WhispAssist T3.4 NPU acceleration COMPLETE (steps ...
|
|
|
|
WhispAssist T3.4 NPU acceleration COMPLETE (steps 1-6, feature_npu branch). Dispatch: commands.rs load_transcriber() routes BackendId::Npu -> OnnxNpuTranscriber (onnx model dir), else whisper.cpp, with graceful CPU fall-through; used by streaming worker and batch reprocess. run_streaming_worker now takes &dyn Transcriber (T: ?Sized). hardware_status returns npu:{present,runtimeReady,modelInstalled}. download_npu_package command + npu://download progress events; startup auto-fetches ONNX model in background if NPU present && model missing. Settings>Hardware shows NPU detected + download indicator. All gates green first try: fmt, clippy default+npu -D warnings, 69 default tests, 77 npu tests, svelte-check 0 errors, eslint. Real NPU inference re-verified post-refactor. KNOWN GAP: OpenVINO runtime staging (stage_npu_runtime) copies DLLs from local dirs in env WA_NPU_RUNTIME_SRC (';'-separated) not a hosted download - no hosted runtime bundle URL yet. Upgrade path: host versioned ORT+OpenVINO bundle, download+unzip into paths::npu_runtime_dir(). Repo frontend NOT prettier-clean (117 files pre-existing); only touched files formatted.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-02T22:17:13*
|
|
|
|
### WhispAssist merged single installer (2026-07-06, b...
|
|
|
|
WhispAssist merged single installer (2026-07-06, branch feature_chore_bug_003): Combined the two installers (13MB default/NPU-DirectML at src-tauri/target vs 67MB Vulkan at C:\wt) into ONE universal Vulkan installer. Verified facts: default exe 13MB imports no vulkan-1.dll (runs anywhere, GPU via runtime DirectML); Vulkan exe 67MB imports vulkan-1.dll at LOAD time (dumpbin) so it won't launch without the loader. Fix (option A = bundle the loader): build.rs stage_vulkan_loader() runs when CARGO_FEATURE_VULKAN set — copies vulkan-1.dll from %VULKAN_SDK%\Bin (fallback C:\Windows\System32) next to the exe (OUT_DIR ancestors nth(3) = target/<profile>, correct under CARGO_TARGET_DIR=C:\wt) AND into src-tauri/ (gitignored) for bundling. New src-tauri/tauri.vulkan.conf.json overlay adds bundle.resources ['vulkan-1.dll']. Release build cmd: npm run tauri build -- --features vulkan --config src-tauri/tauri.vulkan.conf.json (needs VULKAN_SDK, CMAKE_GENERATOR=Ninja, CARGO_TARGET_DIR=C:\wt). VERIFIED end-to-end: built MSI+NSIS 0.1.6; WiX main.wxs shows vulkan-1.dll as a Component in the same install dir as whispassist.exe + sibling DLLs (onnxruntime/sherpa/whispassist_lib), so the loader finds it. DirectML stays hidden when vulkan compiled (directml_would_help returns false). STILL TO TEST BY USER: launch the merged installer on a clean VM with NO vulkan-1.dll / no GPU driver to confirm graceful CPU fallback before retiring the 13MB build.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-07T00:06:25*
|
|
|
|
### WhispAssist 5-feature batch (2026-07-06, branch fe...
|
|
|
|
WhispAssist 5-feature batch (2026-07-06, branch feature_chore_bug_003): (1) Notes single-pane Editor/Preview toggle in TranscriptNotes.svelte (one button flips label Preview<->Editor). (2) Recording playback FR-REC-5: command recording_playback_path decrypts vault-sealed wav to audio.play.wav; enabled tauri protocol-asset feature + assetProtocol scope [$LOCALDATA/WhispAssist/meetings/**] + media-src CSP; SummaryPanel <audio controls> via convertFileSrc. (3) Smaller wav FR-CAP-8: wav_spec_for now forces 16-bit Int at native rate; write_wav_bytes quantizes f32->i16 via f32_to_i16 (clamp*i16::MAX) — halves 205MB->~103MB. Kept native rate (48k), did NOT resample to 44.1k (marginal, needs multichannel resampler in hot path). (4) Cancel recording FR-CAP-9: cancel_recording command stops loopback+mic, joins worker, store.delete_meeting (row+folder), emits recording://state state:cancelled; App.svelte Cancel button with confirm. (5) Live sync progress FR-SYNC-11: sync put() streams body via futures_util::stream::unfold + reqwest wrap_stream + Content-Length, sends incremental (sent,total); upload_job drains on a std thread (throttled 250ms) calling on_progress; pump_sync emits live sync://job; SummaryPanel <progress> bar. All clippy-clean, svelte-check clean, vite+full cargo build pass.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T22:57:29*
|
|
|
|
### WhispAssist 0.1.6 release build (2026-07-06): Vulk...
|
|
|
|
WhispAssist 0.1.6 release build (2026-07-06): Vulkan build (npm run tauri build -- --features vulkan; VULKAN_SDK=C:\VulkanSDK\1.4.350.0, CMAKE_GENERATOR=Ninja, CARGO_TARGET_DIR=C:\wt, use QUOTED 'set "VAR=val"' to avoid trailing-space bug). Headline fix vs 0.1.5: keyring windows-native (OS credential store was a no-op mock; sync/AI creds never persisted). Artifacts in C:\wt\release\bundle\: msi\WhispAssist_0.1.6_x64_en-US.msi (27MB, SHA256 ffbe32d9b53f2feaaa4b8a6a858b2f283cc5520b7e77814bc5cf7a41e04b5301), nsis\WhispAssist_0.1.6_x64-setup.exe (8.8MB, SHA256 669147d9578b2644b0838de346dda9ce7edd2b614a18f63ca5f61ffd64c6b526). SHA256SUMS.txt written to C:\wt\release\bundle\. Upload the .msi + -setup.exe + SHA256SUMS.txt; NOT the .7z runtime bundles (hosted in-repo via Git LFS, pulled from /media/branch/main/runtime/). 27MB MSI confirms Vulkan (CPU-only=11MB). Commits since 0.1.5: keyring fix 33c0397, version bump 6a4b832, plus sync-target-edit, Nextcloud server-URL auto-build, dark-theme fixes, About page, 7z runtime download.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T18:14:30*
|
|
|
|
### WhispAssist 0.1.5 shipped (2026-07-06, branch chor...
|
|
|
|
WhispAssist 0.1.5 shipped (2026-07-06, branch chore_debug): two UI fixes + version bump, 6 commits (035abb9..0d62416). (1) Settings panel horizontal-scroll/close-button-cutoff FIXED: the tab <nav> in Settings.svelte didn't wrap inside the fixed-width panel (width:min(720px,92vw)), overflowing and pushing the .close (X) button off-edge + adding a horizontal scrollbar. Fix: nav { flex:1 1 auto; min-width:0; flex-wrap:wrap }, header align-items:flex-start, .close flex:0 0 auto. (2) New ABOUT page (Settings ▸ About, Info icon tab): shows version (env!(CARGO_PKG_VERSION)) + build commit hash + a source link to https://git.dou.bet/iamdoubz/WhispAssist. Commit hash baked at build time via build.rs (git rev-parse --short HEAD -> cargo:rustc-env=WA_GIT_HASH, rerun-if-changed=../.git/logs/HEAD). Two new Tauri commands: app_info()->{version,commit}, open_url(url) (validates http(s), Windows-only via 'explorer <url>' — no shell injection, reuses installed toolchain instead of adding tauri-plugin-opener). api.ts got AppInfo type + appInfo()/openUrl() bindings. NOTE the section {#if}/{:else if} chain in Settings.svelte: privacy was the catch-all {:else} — adding an About branch required converting privacy to {:else if section==="privacy"} because {:else if} can't follow {:else}. Version bumped 0.1.4->0.1.5 in package.json, src-tauri/Cargo.toml, src-tauri/tauri.conf.json, Cargo.lock (package-lock.json tracks app version as 0.0.0, untouched). clippy -D warnings + svelte-check + eslint(my files) all clean.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T12:48:09*
|
|
|
|
### WhispAssist P2 DirectML GREEN end-to-end (2026-07-...
|
|
|
|
WhispAssist P2 DirectML GREEN end-to-end (2026-07-05): staged Microsoft.ML.OnnxRuntime.DirectML v1.24.1 (EXACT match to the OpenVINO bundle's ORT 1.24.1, from nuget flat-container api.nuget.org/v3-flatcontainer/microsoft.ml.onnxruntime.directml/1.24.1/...nupkg) into %LOCALAPPDATA%/WhispAssist/runtime/directml/ = onnxruntime.dll (17MB, DML EP baked in) + onnxruntime_providers_shared.dll. DirectML.dll NOT in the nuget — the Win11 System32 DirectML.dll (v1.15.5) satisfied it. Ran directml_transcribes_speech spike (no ORT_DYLIB_PATH; ensure_runtime_env pointed ort at runtime/directml/onnxruntime.dll + prepended its dir to PATH): backend=Intel (Arc iGPU via DirectMLExecutionProvider device_id 0), load=1222ms, infer=453ms, transcript='(gentle music)' (test wav was music; non-empty => PASS). So the full P0+P2 DirectML path is proven working on real GPU hardware, infer time on par with Vulkan/NPU. Repro: nuget version MUST be >= the ORT the OpenVINO bundle ships (1.24.x) for ABI/symbol match with ort rc.10. Older DirectML nugets (1.20-1.23) would risk GetProcAddress misses.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T02:30:59*
|
|
|
|
### WhispAssist sync-target EDIT feature added (2026-0...
|
|
|
|
WhispAssist sync-target EDIT feature added (2026-07-06, commits 149f2c9/45b2acb/496a337/fb79009). Root cause of the user's Nextcloud auth failure: the saved target's base_url was missing the /remote.php/dav/files/<user>/ path (they entered just the host) -> PROPFIND hit a non-DAV path -> 401. Credentials + app password were CORRECT (verified via curl PROPFIND returning 207 + X-User-Id). The app had NO way to edit a saved target — only add/delete/toggle-enabled — so they couldn't fix the URL. FIX: the backend update_sync_target command + store + api.updateSyncTarget ALREADY existed (was only used by the enable toggle); added the missing UI. Changes: (1) SyncTargetInfo (models.rs + api.ts) gained upload_transcript/notes/summary/recording, trigger_on_finalize, allow_plaintext_lan, encrypt_before_upload + row_to_info populates them (secret still never exposed, FR-SYNC-6); (2) settings.svelte.ts store.updateTarget(); (3) Settings.svelte: per-webdav-target Edit button -> startEdit() loads it into the add-form (secret blank), heading/submit become 'Edit target'/'Save changes', kind tabs hidden during edit, Cancel button. Password left blank on save = keep stored (update only rotates the credential when a non-empty secret is provided). Test-connection in edit mode tests the TYPED values (form has no id) so it needs the password re-entered; the simpler fix-and-save path preserves the secret. Correct Nextcloud WebDAV base_url = https://HOST/remote.php/dav/files/USERNAME/ . clippy + svelte-check + eslint all clean.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T15:41:28*
|
|
|
|
### WhispAssist 0.1.5 release build (2026-07-06, first...
|
|
|
|
WhispAssist 0.1.5 release build (2026-07-06, first public release): Vulkan build via 'npm run tauri build -- --features vulkan' with env VULKAN_SDK=C:\VulkanSDK\1.4.350.0, CMAKE_GENERATOR=Ninja (ninja 1.10.2 at C:\Tools\Standalone), CARGO_TARGET_DIR=C:\wt, vcvars64 loaded. GOTCHA that failed the first attempt: cmd 'set CARGO_TARGET_DIR=C:\wt && ...' captured a TRAILING SPACE (C:\wt ) -> 'failed to create directory C:\wt \release'; fix = quoted set: set "CARGO_TARGET_DIR=C:\wt". Artifacts in C:\wt\release\bundle\: msi\WhispAssist_0.1.5_x64_en-US.msi (27MB, SHA256 ed83ad0001c654221f3e5787088d922a1211d2722acbd5c2b4523f4f98c745e6), nsis\WhispAssist_0.1.5_x64-setup.exe (8.8MB, SHA256 dfa3a3acf7e9fdfe6d104527d55f660f0b9c0c1364da0cf665113aad60c61422). SHA256SUMS.txt written to C:\wt\release\bundle\. 27MB MSI size confirms Vulkan (CPU-only was 11MB). Release page should upload: the .msi, the -setup.exe, and SHA256SUMS.txt — NOT the .7z runtime bundles (those are hosted in-repo via Git LFS, pulled on-demand from /media/branch/main/runtime/).
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T15:03:01*
|
|
|
|
### WhispAssist microphone-capture feature FR-CAP-7 (2...
|
|
|
|
WhispAssist microphone-capture feature FR-CAP-7 (2026-07-06, branch feature_chore_bug_002): WA now optionally captures the user's mic alongside loopback and mixes both 16kHz-mono streams into the single transcription worker via audio::spawn_mixer (Mixer struct sums+clamps aligned samples, forwards survivor when one source stalls/ends). New: WasapiCapture::start_microphone (Direction::Capture, no WAV), audio::list_capture_devices, list_input_devices command, Settings.microphone_enabled(default true)+audio_input_device. capture_loop generalized: wav_path Option, direction param, emit_level only for loopback. RecordingSession.mic_capture Option; stop/pause/resume handle both. Settings>Hardware>Audio Devices got a Microphone picker (Off/Default/devices). Loopback WAV stays byte-accurate native; mic is transcript-only (not in WAV/diarization) - tracked ponytail limitation.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T21:34:29*
|
|
|
|
### WhispAssist P0+P2 GPU acceleration SHIPPED (2026-0...
|
|
|
|
WhispAssist P0+P2 GPU acceleration SHIPPED (2026-07-05, branch chore_debug). P0 (2 commits e9567b8,7fe0d87): AccelPath enum {WhisperCpu,WhisperVulkan,WhisperCuda,OnnxOpenVino,OnnxDirectML} + resolve_accel(_with) in hardware/mod.rs = single source of truth; BackendInfo.available now derived from it (DXGI + NPU), closing the no-op-GPU detection gap; load_transcriber drives engine choice from it. 3 new pure resolver unit tests, all green. P2 (DirectML for AMD/Intel non-Vulkan): ort gains 'directml' feature; OnnxNpuTranscriber RENAMED to OnnxTranscriber (file still npu.rs) — load() now picks OpenVINO(NPU) vs DirectML(Amd/Intel, device_id 0) EP by BackendId over the SAME onnx artifacts; ensure_runtime_env takes the runtime dll path; paths::directml_runtime_dir/dll/ready added (C:\Users\dadous\AppData\Local/WhispAssist/runtime/directml/onnxruntime.dll); download_and_extract_runtime parameterized (sha+ready) and reused by new stage_directml_runtime + download_directml_package command (registered in lib.rs); hardware_status reports a directml field. 104 lib tests + clippy -D warnings all green across feature sets. TESTED end-to-end: directml_transcribes_speech spike (ignored test in npu.rs, env WA_DML_MODEL_DIR/WA_DML_TEST_WAV/WA_DML_BACKEND) run with ORT_DYLIB_PATH=the existing OpenVINO onnxruntime.dll -> FAILED as expected with 'GetProcAddress OrtSessionOptionsAppendExecutionProvider_DML failed' = the OpenVINO ORT build has NO DirectML EP. This PROVES the DML dispatch path selects the EP and fails LOUDLY (error_on_failure) instead of silently. REMAINING for a GREEN GPU run: stage a real DirectML-EP onnxruntime.dll (Microsoft.ML.OnnxRuntime.DirectML) into runtime/directml/, VERSION-MATCHED to ort rc.10 / ORT ~1.24.x (OpenVINO bundle is ORT 1.24.1). No hosted DirectML bundle published yet; WA_DIRECTML_RUNTIME_URL env drives download, empty SHA. CUDA (P1) NOT done yet (needs NVIDIA hw/CI).
|
|
|
|
*Confidence: 0.95 | Status: active | Created: 2026-07-06T00:45:47*
|
|
|
|
### WhispAssist DirectML Settings toggle WIRED (2026-0...
|
|
|
|
WhispAssist DirectML Settings toggle WIRED (2026-07-06, branch chore_debug, 4 commits 281ccf0/ceb7e59/1b743da/1b74e0b). Backend: hardware::directml_would_help() (pub, in hardware/mod.rs) = cfg!(feature=npu) && !cfg!(feature=vulkan) && a present AMD/Intel GPU (or NVIDIA when !cfg!(cuda)) via dxgi::enumerate_gpus() presence — gates the card so it stays HIDDEN on the shipping Vulkan build (Vulkan already covers all GPUs) and only shows on CUDA/non-Vulkan builds where a GPU lacks coverage. hardware_status now returns directml:{applicable,runtimeReady,modelInstalled} (mirrors the npu:{} field). Frontend: api.ts HardwareStatus gained directml?:{applicable,runtimeReady,modelInstalled} + api.downloadDirectmlPackage(); Settings.svelte has a 'GPU acceleration (DirectML)' card mirroring the NPU package card (reuses .npu-package CSS), shown when directml.applicable, Ready badge when runtimeReady&&modelInstalled else a Download button -> downloadDirectmlPackage() -> loadHardware(). Progress is await-driven (busy flag), NO % listener — DirectML runtime/model progress still emits on the cosmetic npu://download channel (only 'done' on directml://download). It's a package-DOWNLOAD action, not a persistent on/off toggle; the real 'use this GPU' switch is the existing Preferred-backend dropdown (which now enables AMD/Intel once staged, via the P0 availability fix). clippy -D warnings + svelte-check both clean (fixed a needless_return in directml_would_help by using the cfg-block tail-expression pattern like npu_hardware_present).
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T12:27:50*
|
|
|
|
### WhispAssist release 0.1.4 = first Vulkan-enabled b...
|
|
|
|
WhispAssist release 0.1.4 = first Vulkan-enabled build. Artifacts: C:\wt\release\bundle\msi\WhispAssist_0.1.4_x64_en-US.msi (27MB), C:\wt\release\bundle\nsis\WhispAssist_0.1.4_x64-setup.exe (8.7MB), C:\wt\release\whispassist.exe (64MB) — sizes jumped from 11MB/4.1MB/13MB (0.1.3 CPU-only) because the Vulkan backend + embedded SPIR-V shaders are statically compiled in. Built via 'npm run tauri build -- --features vulkan' with the [[whispassist-vulkan-build-recipe]] env. REMAINING WIRING GAPS (not yet done): (1) detection-honesty gating — mark GPU backends 'available' only when a vulkan/cuda feature is compiled (cfg!(feature=...)), else best() routes to a no-op GPU; (2) make Vulkan the standing release build flag instead of manual --features vulkan; (3) preferred_backend UI so the user can force Intel GPU (currently best() picks NPU rank-0 over Intel rank-3 on this machine, so the app uses NPU not Vulkan unless overridden).
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-05T22:04:13*
|
|
|
|
### WhispAssist Nextcloud sync UX overhaul (2026-07-06...
|
|
|
|
WhispAssist Nextcloud sync UX overhaul (2026-07-06, commits 8f6e804/9c670f9). ROOT CAUSE of user's confusion (dwdoubet@box.dou.bet): they were editing with the password field BLANK. WebDavTarget.test() does PROPFIND on base_url+remote_base_path. Bare host https://box.dou.bet + blank pw -> PROPFIND https://box.dou.bet/WhispAssist = NON-DAV path -> 404 (no auth needed) -> test treats 404 as success -> FALSE 'Connected'. Full DAV URL + blank pw -> real DAV endpoint -> 401 -> 'auth failed'. So bare-host 'Connected' was a false positive. FIX 1: WebDavTarget gained provider_hint field; for nextcloud/owncloud, dav_root() derives origin (scheme+host+port) from base_url and builds {origin}/remote.php/dav/files/{username}/ — so users enter ONLY the server URL (https://box.dou.bet) and the app builds the canonical DAV path; a pasted full path is normalized via origin(). Other providers (seafile /seafdav, synology /dav, cloudreve, generic) still use base_url verbatim. url_for uses dav_root(). FIX 2: test_sync_target — for an EXISTING webdav target (id present), it now builds WebDavTarget from the stored row's credential_ref (STORED password) + applies form overrides (base_url/username/remote_base_path/provider_hint/allow_plaintext_lan), unless the user typed a NEW password (temp cred). So Test works after a URL fix WITHOUT re-entering the password. Frontend: davAutoPath derived (provider is nextcloud/owncloud) drives the Server URL placeholder/hint; testConnection passes id in edit mode. Test nextcloud_builds_dav_path_from_server_url added. Correct Nextcloud username here = dwdoubet, host box.dou.bet. NOTE: their existing saved target may just start working after this (url_for rebuilds path from origin+username) if provider_hint=nextcloud. clippy --all-targets + svelte-check + eslint clean; 13 sync tests pass.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T16:11:14*
|
|
|
|
---
|
|
|
|
## Errors
|
|
|
|
*Failure records, bugs, and lessons learned from mistakes.*
|
|
|
|
### WhispAssist crash to INVESTIGATE LATER (2026-07-06...
|
|
|
|
WhispAssist crash to INVESTIGATE LATER (2026-07-06, dev build, branch feature_chore_bug_002): app crashed with exit code 0x80000003 (STATUS_BREAKPOINT) around DirectML use on an Intel GPU. Repro sequence: NPU recording worked fine (mic capture confirmed working); user then switched to test DirectML, downloaded the DirectML components, then hit Record and it crashed — crash may have occurred DURING the component download or immediately after starting recording. Log evidence at crash: WARN 'ONNX engine load failed (model load failed: Error attempting to load symbol OrtSessionOptionsAppendExecutionProvider_DML from dynamic library: GetProcAddress failed); falling back to CPU', then whisper_model_load loading ggml-small.en-q5_1.bin, then process exited 0x80000003. CONFIRMED unrelated to the microphone/mixer feature (FR-CAP-7) — it's in the DirectML/ONNX + whisper model-load path. Suspect: DirectML runtime download/activation or DML EP symbol-load failure interacting with recording start. Next: reproduce by enabling DirectML on Intel GPU + start recording; check GetProcAddress DML symbol load and whether crash is during download vs whisper load.
|
|
|
|
*Confidence: 0.9 | Status: active | Created: 2026-07-06T21:59:37*
|
|
|
|
### WhispAssist build gotcha (cost ~1hr this session, ...
|
|
|
|
WhispAssist build gotcha (cost ~1hr this session, masqueraded as a Node 26 incompatibility): NEVER use PowerShell 'Set-Content -Encoding utf8' on package.json / tauri.conf.json / any JSON or TOML — Windows PowerShell 5.1 writes UTF-8 WITH a BOM. The BOM in package.json breaks vite (fails 'type:module' detection -> 'This package is ESM only but was loaded by require' for @sveltejs/vite-plugin-svelte) AND vitefu (JSON.parse chokes: 'Unexpected token, not valid JSON' -> 'Unable to read package.json'), which fails 'npm run build' / the whole tauri build. Fix: use the Edit tool, or sed, or [System.IO.File]::WriteAllText. Strip an existing BOM with: sed -i '1s/^\xef\xbb\xbf//' file. Node was v26.3.0 at C:\Tools\node but Node was NOT the cause.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-05T22:04:14*
|
|
|
|
### Critical recurring issue: WhispAssist debug (dev) ...
|
|
|
|
Critical recurring issue: WhispAssist debug (dev) builds hang/stall during whisper.cpp model loading on this machine, apparently due to Trend Micro AV behavior-monitoring interfering with ZwWriteVirtualMemory calls made during model load. Confirmed reproducible even in a bare standalone Rust example binary with zero Tauri/webview involvement, so it is specific to whisper.cpp model loading in a debug-profile binary, not the app shell. Release (optimized+stripped) builds do NOT hit this - confirmed by the user and by direct testing. Escalated to IT, unresolved as of 2026-07-02. Workaround: use release builds for real testing/spot-checking; expect dev-mode launches to sometimes hang at model load and need force-killing.
|
|
|
|
*Confidence: 0.9 | Status: active | Created: 2026-07-02T20:14:14*
|
|
|
|
### WhispAssist CRITICAL BUG FOUND + FIXED (2026-07-06...
|
|
|
|
WhispAssist CRITICAL BUG FOUND + FIXED (2026-07-06): the OS credential store was a NO-OP the entire time. keyring 3.x feature-gates its platform backends and they are OFF by default; 'keyring = { version = "3", optional = true }' had NO backend feature, so on Windows keyring silently used its MOCK keystore. The mock does NOT persist across keyring::Entry instances, so credentials::set (Entry A) appeared to succeed while credentials::get (a fresh Entry B, same service+name) always returned NoEntry -> has_secret=false -> no HTTP Basic auth sent -> 401. This broke ALL sync WebDAV auth, and would break MCP/hosted-AI keys + OAuth tokens too (anything via sync::credentials, SERVICE='WhispAssist-sync'). FIX: keyring = { version = "3", optional = true, features = ["windows-native"] } (pulls dep:windows-sys = real Windows Credential Manager). Confirmed: WA_SYNC_TEST_DIAG went from status=401 has_secret=false to status=207 has_secret=true. DIAGNOSIS JOURNEY (Nextcloud box.dou.bet user dwdoubet): symptom 'auth failed'; the user's credentials + full DAV URL were valid (curl PROPFIND 207). Red herrings: (a) earlier the Nextcloud base_url was missing /remote.php/dav/files/<user>/; (b) a bare-host test gave a FALSE 'Connected' because a non-DAV path 404 is treated as success; (c) blank password in edit test. Real root cause was keyring. TOOLING NOTE: diagnosed via temporary tracing::info! logs (WA_SYNC_CMD_DIAG in test_sync_target = has_id+secret_len; WA_SYNC_SET_DIAG in add branch; WA_SYNC_TEST_DIAG in WebDavTarget::test() = url+status+has_secret) read live from the 'npm run tauri dev' output (tracing filter is 'info' in lib.rs). Also learned keyring feature name = windows-native, and the app's tracing default level is info. These temp diagnostics MUST be removed before release.
|
|
|
|
*Confidence: 1.0 | Status: active | Created: 2026-07-06T18:06:13*
|
|
|
|
### WhispAssist debug-CRT assertion (found 2026-07 whi...
|
|
|
|
WhispAssist debug-CRT assertion (found 2026-07 while running the non-vulkan CPU baseline in DEBUG): the whispassist_lib debug test binary throws MSVC Debug Assertion 'Expression: _osfile(fh) & FOPEN' at ucrt read.cpp:381 = a read() on a CLOSED/invalid file handle. It pops a MODAL Abort/Retry/Ignore dialog that HANGS the test (this is what stalled the overnight non-vulkan baseline run for 8 hours). Only fires under the debug CRT (-MDd); the RELEASE 0.1.4 build does NOT assert (release CRT skips the check), so shipping is unaffected — but the underlying 'read on a closed handle' is latent UB worth root-causing. Likely in the transcription file-read path (whisper.cpp model load or our audio read_wav_mono_16k / vault::open passthrough). ADD to the CPU testing round: investigate this handle bug alongside the ~90s transcribe_file slowness (may or may not be related). Workaround to get a clean non-vulkan CPU number: run 'cargo test --release' (no debug CRT dialog), not plain 'cargo test'.
|
|
|
|
*Confidence: 0.95 | Status: active | Created: 2026-07-05T22:07:51*
|
|
|
|
---
|
|
|
|
*End of memory export.*
|