Personal/community inventory tool. A timestamped, append-only event log (data/log/log.md) fed by:

• Hand-written notes dropped into data/inbox/openclaw/ (or any inbox subdir) as .md/.txt.
• Phone audio synced via Syncthing into data/inbox/audio/ → transcribed by Whisper.
• Phone video synced via Syncthing into data/inbox/video/ → frames every 5s + Whisper on the audio track.
• One-time Google Doc import of the existing inventory_final.md (TODO: komventory import-gdoc <path>).
• Later: a sandboxed reasoning VM ("openclaw") that reads data/log/log.md and drops new notes into data/inbox/openclaw/.

Entries look like:

## 2026-05-16T20:42:11+02:00 — source: whisper@audio/20260516-2041.m4a — loc: "Nářadí přepravka 1"

Štětec, šroubováky, řezák malý…

![[media/audio/20260516-2041.m4a]]

loc: is optional. source: tags the origin so a note can be regenerated (e.g. re-run Whisper with a better model) without losing track of which entry replaced which.

data/
  log/log.md              source of truth, append-only
  log/media/              attachments referenced by entries
  inbox/audio/            Syncthing target for phone audio
  inbox/video/            Syncthing target for phone video
  inbox/openclaw/         (phase 2) VM drops *.note.md here
  inbox/imports/          drop the Google Doc export here
  cache/whisper/          downloaded Whisper models (persists across rebuilds)
src/komventory/           the package
docker/Dockerfile         uv + ffmpeg + faster-whisper
compose.yml               default service: `komventory watch`

# Build once (uv sync inside the container; first run pulls the Whisper model).
docker compose build

# Run the watcher in the foreground.
docker compose up

# One-shot sweep of inbox/ (no watcher).
docker compose run --rm komventory komventory ingest

# Process one specific file.
docker compose run --rm komventory komventory ingest /data/inbox/audio/foo.m4a

# Print resolved paths (useful when debugging bind mounts).
docker compose run --rm komventory komventory paths

uv sync
uv run komventory paths

komventory-api (compose service) serves a single-page web app on port 3411:

• Type a note or hit 🎙 záznam to start continuous VAD-chunked recording (Silero VAD in the browser).
• Each detected utterance posts to /api/notes/audio → Whisper → entry in log.md.
• "mluvit zpět" reads transcripts back in Czech (Piper, thomcles-high — Thomcles/Piper-TTS-Czech, a jirka fine-tune — auto-downloaded on first use; override with KOMVENTORY_TTS_VOICE, cs_CZ-jirka-medium still works per-request via voice).
• "odpovídat na otázky" routes utterances that look like questions through /api/ask (LLM is stubbed for v1 — wire a real backend in src/komventory/qa.py:_call_llm).

Mobile mic requires HTTPS. Reverse-proxy from your existing Caddy (or any HTTPS front), passing the SSE endpoint through unbuffered:

inv.example.org {
    reverse_proxy <docker-host>:3411 {
        # SSE: flush each event immediately.
        flush_interval -1
    }
}

The PWA and the API share one origin (Caddy fronts both / and /api/* at the same hostname) — no CORS, no mixed content.

Default large-v3, multilingual, pinned to Czech (KOMVENTORY_WHISPER_LANG=cs). On CPU with int8 expect ~1–3× realtime; a backlog of phone videos can take hours. Override in compose.yml to medium/small/base for faster (lower quality) runs, or set KOMVENTORY_WHISPER_DEVICE=cuda if you have a GPU.