v3.3.1 - Ask Docs toolbar tab + retrieval overhaul + 3rd Groq key

Plug-and-play zip: extract anywhere, run Hotkeys.exe. No installer, no admin.

Highlights

• Ask Docs toolbar tab: Shift+F11 opens Ask Docs, now with a proper 📚 tab in the main toolbar (was showing as a placeholder). Same NotebookLM-style Q&A workspace.
• 3 Groq keys in rotation (was 2). Cerebras still 2 free-tier keys. All bundled, works out of the box.
• Retrieval overhaul in Ask Docs: IDF-weighted lexical search with log-saturation, location anchors (Surah N / verse S:V), hybrid semantic+lex round-robin merge, char_budget 20k→50k for exhaustive queries, ultra-compact one-line-per-verse output format, CJK bracket normalisation. 8+ additional retrieval fixes.
• Cerebras max_tokens 4096→16384 so reasoning models have headroom for long enumerated answers.
• Prewarm 2000ms → 1000ms — first Shift+F11 press opens Ask Docs faster.

Hotkey reminder (unchanged; safeguarded)

• refine: Alt+Shift+W
• library: Alt+Shift+E
• undo refine: Alt+Shift+Z
• whisper (STT): Ctrl+Enter (primary), Alt+Space (secondary)
• Quick notes: Shift+F7 or Home
• Ask Docs: Shift+F11

Bundled + verified
18/18 critical assets present including all 5 API keys, ask_docs subpackage, MiniLM embedding model, Magika file-detector models, WebView2 runtime, Whisper models, Tenacity audio editor.

Size: 1.8 GB zip → ~3.3 GB extracted.

v3.3.0 - Ask Docs (NotebookLM-style Q&A, Shift+F11)

v3.3.0 — Ask Docs (NotebookLM-style document Q&A)

New feature: Shift+F11 → Ask Docs

A full NotebookLM-inspired document Q&A workspace, bundled directly into Hotkeys.
Press Shift+F11 anywhere to open it.

What you can do:

• Load PDFs, Word docs (.docx), PowerPoint (.pptx), Excel (.xlsx), HTML, EPUB, images (with OCR), audio (transcribed via bundled Whisper), YouTube URLs, and pasted text
• Ask questions about your sources; answers arrive with inline numbered citation chips that link back to the exact passage
• Auto-generated summary + suggested follow-up questions on first source added
• Per-notebook artifacts: Summary, FAQ, Study Guide, Timeline
• Chat history persisted per notebook

Under the hood:

• Local all-MiniLM-L6-v2 ONNX embedding model + per-notebook SQLite vector index (fully offline retrieval)
• Chat inference via Groq (openai/gpt-oss-120b), automatic fallback to local Qwen if Groq unreachable
• MarkItDown for universal document parsing (dozens of formats)
• OCR fallback via Groq vision (qwen/qwen3.6-27b) for scanned PDFs and images
• Audio transcription via Hotkeys' bundled faster-whisper small model

Startup:

• Prewarmed at app launch — first Shift+F11 press opens the window instantly
• markitdown import ~2s + MiniLM ONNX load ~250 ms happens in the background thread while you use other Hotkeys features

Data location:
%APPDATA%\Hotkeys\ask_docs\<notebook-uuid>\ — sources, chats, vectors, artifacts

Build integrity

• verify_dist.py grew 3 new checks (ask_docs subpackage, MiniLM ONNX ≥80MB, tokenizer.json). Build fails non-zero if any missing.
• hotkeys.spec bundles the full ask_docs subpackage + MiniLM (86 MB) + tokenizer + collect_all('markitdown') for its dynamic converter backends

Size

Zip: 1.76 GB (+~150 MB over v3.2.9). Extracted dist: 3.3 GB.

Ship as Hotkeys-Windows.zip — extract anywhere, run Hotkeys.exe.

v3.2.9 - Cerebras: gpt-oss-120b primary (fixes 404 on llama-3.3-70b + preps for Aug 17 glm-4.7 deprecation)

What's new in v3.2.9

Cerebras model migration ahead of their 2026-08-17 deprecation.

Cerebras (chat/refine primary)

• Retired llama-3.3-70b (has been 404-ing for weeks). Replaced with:
  • Primary: gpt-oss-120b — same 120B model as our Groq fallback, on Cerebras' faster infrastructure. Verified 200 OK end-to-end from the built dist.
  • Fallback: gemma-4-31b (Gemma 4)
• Retired zai-glm-4.7 (deprecating Aug 17). Config migration silently upgrades any existing user config naming these models to gpt-oss-120b on next launch.

Everything else unchanged

• Groq continues to serve: whisper-large-v3-turbo (STT), qwen/qwen3.6-27b (vision/OCR), openai/gpt-oss-120b (chat fallback)
• All features remain free-tier (Cerebras Developer + Groq Developer)
• 4 bundled API keys (2 Groq + 2 Cerebras) rotate on 429 rate limits

End-to-end verification of this dist

• verify_dist.py: 14/14 critical assets present + valid
• _bundled_keys.py shipped inside _internal/ (341 bytes, contains real gsk_ and csk- keys)
• Fresh launch: no "BUNDLED KEYS MISSING" warning
• Fresh launch: no Cerebras 404 (was constant before this release)
• Fresh launch: live POST api.cerebras.ai/v1/chat/completions HTTP/1.1 200 OK

Ship as Hotkeys-Windows.zip — extract anywhere, run Hotkeys.exe.

v3.2.8 - Paste-verify rewrite + RTL Arabic/Urdu + Home key + STT cold-start

What's new in v3.2.8

Paste flow, RTL rendering, hotkeys, robustness — a big cleanup pass.

Paste-then-open-notepad fallback (the big one)

Complete rewrite of the "did the paste actually land?" logic around a universal signal:

• Editable detection: UIA TextPattern is the discriminator for Chromium/Electron apps (Claude, VS Code, Discord, Slack) — editable inputs expose it, read-only content areas don't. Was falsely trusting message-history areas as editable before.
• Caret movement: for Win32-native apps (Notepad, Word, Explorer rename), we watch the caret rect via GetGUIThreadInfo and confirm it moved after Ctrl+V.
• UIA text-in-after: strong-proof check that the pasted text appears in the focused element's snapshot.
• Chrome omnibox / autocomplete popup case handled without special-casing: if focus vanishes to a no-caret popup after paste, we trust it.

Result: dictate → paste lands in real input → no MiniNotepad. Dictate with cursor on desktop / read-only page / message history → paste doesn't land → MiniNotepad opens with the text.

Quick Notes: bare Home key as secondary hotkey

SHIFT+F7 still works; HOME also toggles Quick Notes (open on first press, save+close on second press). Notes tab description updated to show both.

Arabic / Hebrew / Urdu in MiniNotepad

Tk's Text widget doesn't apply the Unicode BiDi Algorithm, so RTL scripts came out with words reversed and characters scrambled. Fixed via arabic-reshaper + python-bidi preprocessing:

• Display: right-to-left reading order, right-aligned at the widget edge (native RTL layout)
• Ctrl+C / Ctrl+X copies the ORIGINAL logical-order Unicode, so pasting into any BiDi-aware app (Word, browsers, Notepad) renders correctly on their end
• Paste and typing INTO MiniNotepad also get transformed + realigned automatically (not just the initial content)
• Save-to-file writes original logical bytes
• Mixed content (English + Arabic in same note): each line aligned per its script

Cloud STT for slow-internet users

• Cloud transcription timeout raised from 3s to 15s (only affects the failure/offline case — successful calls unchanged)
• Reachability probe timeout 0.5s -> 3s
• Every cloud-fallback case now shows a diagnostic pill: Cloud too slow, Cloud unreachable (DNS), Antivirus blocked cloud, Cloud failed (Type). No more silent falls to local model.
• Mic pre-warm at app startup: first Ctrl+Enter / Alt+Space captures your speech from the very first syllable instead of eating 200-500ms of WASAPI cold-start latency
• Mic-open retry: escalating backoff (up to 2.6s total) for Bluetooth headset HFP switches, WASAPI handle release, transient PortAudio busy

Bulletproof bundled resources

• _bundled_keys.py verified present in every dist via verify_dist.py (post-build check that fails the build if anything critical is missing). Was silently missing from every dist since v3.1 — all users had been running with NO API keys, silent fall to local models for every cloud feature. Fixed for good.
• python-bidi + arabic-reshaper bundled for RTL rendering.

Ship as Hotkeys-Windows.zip — extract anywhere, run Hotkeys.exe.

v3.2.7 - CRITICAL: fixes bundled API keys (was silently missing since v3.1)

What's new in v3.2.7

CRITICAL FIX: bundled API keys are actually shipped now. Since v3.1 the dist has been shipping WITHOUT _bundled_keys.py, causing every user (including your dad) to silently fall back to local-only for STT/refine/vision. Fixed the packaging bug and added a build-time verifier so this class of drift can't happen again.

Voice-to-text

• Mic pre-warm at app startup: first Ctrl+Enter / Alt+Space captures your speech from the very first syllable instead of eating 200-500ms of WASAPI cold-start latency
• Escalating-backoff retry: transient PortAudio busy conditions (Bluetooth HFP switch, WASAPI handle release, another app briefly holding the mic) retry silently across 4 attempts / 2.6s before showing an error
• Every cloud-failure case now shows a diagnostic pill: Cloud too slow (internet slow?), Cloud unreachable (DNS), Antivirus blocked cloud, Cloud failed (Type). No more silent fallback to local.
• Alt+Space works as a second STT hotkey alongside Ctrl+Enter

Quick Notes

• Bare Home key as a second toggle hotkey alongside Shift+F7 (1st press opens, 2nd press saves + closes)
• No more spurious MiniNotepad: pasting into Quick Notes (or any Hotkeys-owned window) no longer triggers the fallback popup

Audio Editor (Shift+F10)

• Suppresses every known first-launch dialog: update-check, welcome, what's-new, analytics, crash-survivor, file-association prompt
• Hidden-dialog rescue: any first-launch dialog we haven't catalogued gets force-brought-to-front so it can't hide behind the main window and silently steal every click

Whiteboard (Shift+F8)

• Bundled Microsoft WebView2 Runtime Fixed Version 150.0.4078.65 — whiteboard opens instantly on any Windows 10/11 PC without needing separate installers
• Friendly error dialog with fix links if underlying .NET runtime is missing

Downloads (Ctrl+Alt+D)

• curl_cffi bundled: Dailymotion, Reddit, X (Twitter) now work
• Playlist detection dialog: "Just this video" / "Entire playlist" / "Cancel"
• Post-error rescue: if merged file exists on disk despite yt-dlp error, we report success
• Friendlier failure pills with plain-language causes
• yt-dlp's own logger routed into app.log for future diagnosis

Reliability

• Root close guard: hidden root window can no longer take the app down via stray WM_DELETE_WINDOW events
• Deep 40-frame stack trace when _quit() fires, so any future close cause is traceable
• Hourly heartbeat log

Build system

• verify_dist.py: post-build script that checks every critical asset (bundled API keys, diarize worker, whiteboard, audio editor, WebView2 runtime, whisper models with min-size validation, tessdata, etc.). Build now FAILS if any asset is missing — preventing the class of silent-drift bug that caused _bundled_keys.py to hide from every release for months.

Ship as Hotkeys-Windows.zip — extract anywhere, run Hotkeys.exe.

v3.2.6 - Whiteboard works on any PC (WebView2 Fixed bundled)

What's new in v3.2.6

Whiteboard just works on every PC now

• Bundled Microsoft WebView2 Runtime Fixed Version 150.0.4078.65 (x64) inside the app. Whiteboard uses the bundled runtime always, never relies on system-installed WebView2. Result: Shift+F8 opens instantly on any Windows 10 / 11 PC, no separate installer needed.
• whiteboard.py sets WEBVIEW2_BROWSER_EXECUTABLE_FOLDER env var before webview.start() so pywebview picks up the bundled binaries.

Size trade-off

• Zip grew from 1.32 GB (v3.2.5) to 1.61 GB (+290 MB). Extracted dist ~3.12 GB.
• One-time bloat, paid for by zero-config whiteboard on every user's PC.

Note

• WebView2 Runtime version 150.0.4078.65 is Microsoft's stable release (support through 2027). Future Hotkeys releases will update to newer WebView2 versions as Microsoft ships them.
• .NET Framework 4.7.2+ is still required (pywebview uses pythonnet). Windows 10 1903+ and all Windows 11 machines have it pre-installed. On the rare PC without it, the whiteboard is silently disabled (no crash, no popup) - every other feature works.

Ship as Hotkeys-Windows.zip — extract anywhere, run Hotkeys.exe.

v3.2.5 - Audio editor: no first-launch popups, no hidden dialogs

What's new in v3.2.5

Fixes reported by real users on v3.2.4.

Audio Editor (Shift+F10)

• Suppress every known first-launch dialog via cfg: update-check prompt, "check for updates" itself (audio editor is fully offline, never phones home), analytics / crash-report / usage-info consent, Welcome, What's New, crash-survivor scan.
• Hidden-dialog rescue: if a first-launch dialog we haven't catalogued still fires, the walker now force-brings it to the front (AttachThreadInput trick) so it can't hide behind the main window and silently steal every click.
• Previous suppressions still active: Splash, FirstProjectSave, MissingExtension, MixMono/Stereo/Unknown, .aup3 file-associations popup.
• Net: first launch goes straight to a functional editor OR a visible dialog you can dismiss with one click. No more "clicks do nothing."

Ship as Hotkeys-Windows.zip — extract anywhere, run Hotkeys.exe.

v3.2.4 - Alt+Space + slow-internet STT + reliability

What's new in v3.2.4

Voice-to-text

• New: Alt+Space works as a second STT hotkey alongside Ctrl+Enter. Easier one-handed. Both always active.
• Slow-internet fix: cloud transcription now waits up to 15s (was 3s) so upload can complete on weak connections. Fast connections unaffected.
• Reachability probe bumped 0.5s to 3s (helps first-attempt on slow WiFi).
• Every cloud-fallback case now shows a diagnostic pill: Cloud too slow (internet slow?), Cloud unreachable (DNS), Antivirus blocked cloud - add api.groq.com to AV exceptions, or Cloud failed (Type). No more silent local fallbacks.

Reliability

• Root close guard: hidden root window can no longer take the app down via stray WM_DELETE_WINDOW events.
• Hook self-heal no longer thrashes on mouse activity (5-min silence threshold).
• Flood guard: no longer punishes users spamming the same hotkey.
• Deep 40-frame stack trace logged when _quit() fires, so any future close cause is traceable.
• Heartbeat log every hour.

Downloads (Ctrl+Alt+D)

• curl_cffi bundled: Dailymotion, Reddit, X (Twitter) now work.
• Playlist detection dialog: pick "Just this video", "Entire playlist", or "Cancel".
• Post-error rescue: if the merged file exists on disk despite a yt-dlp error, we report success instead of failure.
• Friendlier failure pills (plain-language cause, not raw yt-dlp text).
• Progress-pill coverage for every stage: Reading URL, Downloading, Merging, Saved.
• yt-dlp's own logger routed into app.log for future diagnosis.

Whiteboard / Audio editor

• Whiteboard preflight: if .NET Framework is missing/broken on the user's PC, silently skip pre-warm (no scary Python traceback popup at launch).
• If Shift+F8 is pressed on such a PC, friendly dialog with direct download URLs for the two Microsoft installers.
• Audio editor no longer prompts to associate .aup3 files on first launch.

Distribution

• Every download link on the sprawf/hotkeys profile page now uses /releases/latest/download/ pattern, always resolves to newest release automatically.

v3.2.3 - Hotkey self-heal + flood guard fix

Fixes

• Hotkey self-heal: WH_KEYBOARD_LL hook can be silently unhooked by Windows (LowLevelHooksTimeout, session lock/unlock, sleep). The watchdog now cross-checks GetLastInputInfo against our last hook callback tick and reinstalls the hook within 2s if Windows saw input we did not. No more spam-then-dead hotkeys.
• Flood guard: no longer trips on same-hotkey spam (user impatience). Only pauses on 3+ DIFFERENT hotkey events in 2s.

Also carries the v3.2.2 Groq model migration (gpt-oss-120b text, qwen/qwen3.6-27b vision).

v3.2.2 - Groq model migration

What's new since v3.2.1

Migrated off Groq models being decommissioned:

• llama-4-scout-17b-16e-instruct (decommission: 2026-07-17)
• llama-3.3-70b-versatile (decommission: 2026-08-16)

New Groq model chain

• Text default: `openai/gpt-oss-120b` (was llama-3.3-70b-versatile)
• Vision: `qwen/qwen3.6-27b` (was llama-4-scout-17b-16e-instruct)
• Fallback: llama-3.1-8b-instant kept for speed cases

Also fixed

• `vision.py` now strips `...` reasoning tags from Qwen's OCR responses so extracted text doesn't get polluted with the model's chain-of-thought.
• `settings.py` Test-API-Key button now pulls the model from `GROQ_MODELS[0]` instead of hardcoding, so future model rotations don't break the test.

Cerebras untouched

Their `llama-3.3-70b` is a separate model, not affected by Groq's deprecation.

Installation

Same as v3.2.1: download the zip, extract anywhere, run Hotkeys.exe. If upgrading, replace the folder.