Bidet AI Desktop — Overhaul Plan

Drafted 2026-05-23 after today's 22-minute brain dump crashed during transcription. Plan only — no code changed yet. Review and redline before any touch.

(a) What's there now

Stack

• Tkinter UI (app.py, 888 lines), always-on-top dark window, title Bidet AI — Honest Answers
• recorder.py — sounddevice 16 kHz mono, RMS silence detection, beep ladder at 30/45/60s (NOT the 10/15/20 in HANDOFF — already drifted)
• transcriber.py — openai-whisper library, model medium, GPU fp16 on RTX 4070, one big blocking model.transcribe() call, regex de-loop pass after
• processor.py — Ollama local-first (gemma3:4b) with Gemini cloud fallback. Glossary + speaker-context prepended to every prompt
• distributor.py — webhook POST + Google Drive upload (Drive currently skipped, no DRIVE_FOLDER_ID in .env)
• tp3_ingest.py — MinIO audio archive + Postgres ingest with Gemini embedding (works when Ollama is up)
• Prompts library at ~/.bidet/prompts.json — user-overridable per-tab, plus custom tab support

Tabs today

1. Clean Raw (builtin clean)
2. Clean Analysis (builtin analysis)
3. Clean for AI (builtin forai)
4. Clean for Judges (builtin judges) + a Generate Judges Pitch button below the notebook

Icon

• bidet_ai.ico exists in repo root (256×256 RGBA, looks old per Mark)
• NEVER WIRED INTO THE APP. app.py has no iconbitmap / iconphoto call. So the running window shows Tk's default feather. The .ico is just sitting there.

Today's working sessions in bidet_ai.log — 30+ prior runs, all 30s–3min audio → transcribe in 30s–2min, "Whisper done: X chars" always logs. One outlier today.

(b) The 22-min crash — root cause

Log shows exactly one line for today:

12:16:01 INFO Starting Whisper transcription of …audio_2026-05-23_12-16-01.wav

Then nothing. No Whisper done, no Pipeline failed, no exception. App went black.

The pipeline() thread in app.py:659 is wrapped in try/except Exception and would have logged anything Python-level. So model.transcribe() either:

1. Hung in a decode loop. Even with compression_ratio_threshold=2.0 and condition_on_previous_text defaulted on, openai-whisper has a known long-audio failure mode: 30s chunks where it gets stuck repeating, retries internally, never returns. The dedup regex runs after transcribe; it cannot help if transcribe never finishes.
2. CUDA died / OOM with no Python exception. 22 min × 16 kHz of fp16 medium-model decoding can blow the VRAM-host-memory pinned buffers. CUDA OOM in some configurations only shows up in stderr (which we swallowed with io.StringIO() at the top of app.py and transcriber.py — see lines 17–20 of both). If the Whisper C extension faulted, the GUI process would die silently.
3. Tk freeze masquerading as crash. pipeline() runs in a daemon thread so Tk should still pump, but very long GPU work with the Python GIL released and re-acquired in chunks can stall Tk's event loop on Windows. Result: the window stops repainting, Mark sees "black," kills it.

We cannot know which without re-running with proper stderr capture. Mitigation regardless of which one it was:

• Switch transcribe path to faster-whisper (CTranslate2 backend). Already installed in system Python (Mark used it to recover the 22-min audio today — succeeded on CPU int8 in reasonable time, would be much faster on GPU). It:
  • Streams segments back as an iterator → we can write to disk and update the UI progressively, so a crash at minute 18 still saves minutes 0–18
  • Uses ~30–40% less VRAM at the same model size
  • Has battle-tested VAD filtering (vad_filter=True) that solves the loop-on-silence problem at the source instead of regex'ing it later
• Restore stderr capture to a log file instead of StringIO(). The current if sys.stderr is None: sys.stderr = io.StringIO() block silently eats every CUDA / cuBLAS / ctranslate2 message. Route to a rotating file handler so the next mystery crash has a tail.
• Add a transcribe-progress watchdog. If no segment lands for 90s, surface "transcribe stalled" in the status bar and let Mark either wait or kill cleanly with the partial transcript already saved.

(c) UI changes Mark asked for (verbatim)

1. Window title Bidet AI — Honest Answers → Bidet AI. One line in app.py:146.
2. Wire the icon. Add self.iconbitmap(default=str(PROJECT_ROOT / "bidet_ai.ico")) in _build_ui. NEW icon asset required from Mark — current .ico is the old logo he wants replaced. Need him to drop the new PNG/SVG in repo root; I'll convert to multi-resolution .ico (16/32/48/64/128/256) for taskbar + alt-tab + title bar.
3. Tab renames + reshuffle.

4. REMOVE Clean for Judges tab + Generate Judges Pitch button. Contest is shipped, no longer needed. Per the sister rule from today (feedback_never_delete_work_archive_2026-05-23), archive — don't delete:
   • Move JUDGES_PROMPT, format_for_judges, _run_judges, _fill_judges into archive/judges_mode.py so the strings survive
   • Strip the builtin judges tab from the default tab list + remove the button from _build_ui
   • ~/.bidet/prompts.json migration: drop the judges entry on first run after upgrade so Mark's existing state doesn't re-add it
5. Migration of existing user state. ~/.bidet/prompts.json on Mark's machine has the old four-tab structure. On startup, detect old schema and rewrite to new four-tab structure preserving any overrides Mark made on clean / analysis / forai. Back up the old file as prompts.json.bak.20260523 first.

(d) "Premiere level" — what's worth adding, what's gold-plating

Worth doing now   low risk, big lift

Worth considering   ask Mark

• CustomTkinter swap. Modern dark theme out of the box, better widget styling, drop-in replacement for tk.Tk in 90% of cases. Risk: minor layout drift, +1 dep. Payoff: looks more 2026.
• PyWebView / Tauri rewrite. Browser-based UI rendered in a desktop chrome. Bigger lift, but unlocks the same look as Bidet AI Web. Probably premature — Mark's pain points are in the pipeline, not the Tk look-and-feel.
• System tray companion so the window can hide-to-tray instead of being always-on-top. Mark's old gripe was the topmost window blocking other work. Optional.
• Hotkey to start/stop recording the same way Bidet Quick uses Ctrl+Shift+; — would let Mark brain-dump without alt-tabbing.

Not worth it

• Rewriting in Electron / Tauri / Qt. The Tk stack works, the bugs are in the audio path and the UI surface.
• Adding speaker diarization. Solo voice, no need.
• Cloud Whisper. Mark wants local; he has the GPU; it costs nothing to keep local.

(e) Recommended sequence