S.H.O.N.A.R._Desktop_Companion/STATUS.md

# SHONAR Desktop — Status

Checkpoint for the standalone desktop app. Prepend a new dated section per
material change; keep history below.

## 2026-09-22 — open-item audit closed; Large v3 quiet benchmark; in-app update check

- **pump-error.log (open item 1, Sep 16): CLOSED — already gone.** Nothing
  named pump-error.log exists under `~/.config/shonar-desktop/` anymore
  (only `logs/engine.log`). The code path that writes it
  (`DesktopState.kt` pump failure logger) is current and legitimate.
- **Large v3 full-length timing (open item 2, Sep 16): MEASURED.** Quiet
  machine (app + engine stopped), CPU int8, same args as the engine
  (beam_size=5), clips cut from a real 2 h 09 m recording:
  - 300 s audio → 1107 s wall (**0.27× real-time**), first segment 37 s,
    model load 8.7 s.
  - 1200 s audio → 2768 s wall (**0.43× real-time**), first segment 52 s.
  The earlier note ("300 s clip ran well under 7 min") was optimistic —
  under real quiet-CPU conditions Large v3 is ~4–8× slower than that.
  Practical read: a 2-hour recording ≈ 4–8 h of CPU transcription;
  Base (~7× real-time) stays the sensible default, Large v3 is for
  shorter, quality-critical clips.
- **In-app update check** (`7195aa1`, pushed): Settings gains
  "About & updates" — build commit baked into the jar
  (`generateBuildInfo` → `shonar-build.properties`), a Check-for-updates
  button against the Forgejo commits API (up to date / update available
  with Open-repo link / newer-than-remote for dev builds), verified live
  on-device.

## 2026-09-16 — open-item audit; Large v3 provisioning verified

- **Large v3 first-use download verified end-to-end** in an isolated
  `HF_HUB_CACHE`: `snapshot_download` fetched 2.9 GB in ~2.5 min,
  `is_model_downloaded('large-v3')` → True, model loaded in 4.6 s
  (CPU int8) and transcribed a real clip. Note: `setup-standalone.sh`
  vendors only **Base**; Large v3 is deliberately a first-use auto-download
  into the `models/hf-cache` the app exports — the registry's
  `download_instructions()` already covers the manual path. No defect.
- Skipped backend test resolved: intentional (`test_ai_adapters.py:55`
  skips when faster-whisper IS installed — the missing-dep path can't be
  exercised then). Not a Postgres test. Current count: 80 passed, 1 skipped.
- Earlier open items 1–2 (startup job adoption, upload-mapping reuse) were
  closed by `ba54cb8` + the `.shonar.json` sidecar.

## 2026-09-15 (evening) — summarize progress bar, job adoption

**Committed:** `7fa99e8` streaming summarize progress, `ba54cb8` job adoption.

- Summarize now shows a real determinate bar + `Summarizing… N%` (same look
  as the main screen). Backend: both LLM adapters stream (SSE / NDJSON) and
  report 0–99% as tokens arrive; 100 only on success. No fake movement —
  before the first token the label reads "Summarizing… waiting for model"
  (cold LLM / server queueing was indistinguishable from a dead app).
- Adoption of engine jobs across app restarts: startup scan no longer skips
  files that have a report (re-summarize was orphaned), a reprocess that
  hits 409 "already running" attaches to the running job instead of
  erroring, and live work outranks a saved report in library status.
- Verified live: screencap 19:41 — library row shows the bar + waiting
  label mid re-summarize; job settled to succeeded 100 with the % racing
  48→99→100 once tokens flowed (H200 queueing caused the earlier "nothing
  happens"). pytest 79 passed/1 skipped · ruff clean · gradle green.
- Note: the dev loop died once (~19:35, cause unclear — gradle `run` exit
  143 chain); relaunched via background `./desktop-dev.sh`.

**Done earlier (Sep 15 morning):** see below.

## 2026-09-15 — standalone desktop, Android fully removed

**Verified state (live this session):**
- HEAD `6599935`, clean tree, single branch `master`, no stashes
- Backend tests: `78 passed, 1 skipped` (pytest, SQLite) · `ruff check` clean
- Frontend tests: `39 passed` across 7 gradle suites
- Engine runs locally on `127.0.0.1:8000` (healthz ok); runtime state in
  `~/.config/shonar-desktop/` (`engine.db`, `settings.json`)

**What the app is:** Compose Desktop (Kotlin) UI + bundled FastAPI engine
(SQLite + in-process queue, self-migrating on boot). No Postgres/Redis/Docker,
no server, no phone component. faster-whisper (Base vendored under
`models/hf-cache/`, Large v3 downloadable) + an LLM for summaries
(`openai_compat` or `ollama`).

**Architecture decisions worth keeping:**
- All Android/phone code deleted (`71b5571`) — desktop is the product. The
  old server-era source lives on the main repo's `deferred/desktop-server`
  branch for reference only.
- Shared code sits at `shared/com/shonar` (no `src/`), pulled in via
  `srcDir("../shared")` in `app/build.gradle.kts`.
- Search picks its path by SQL dialect inside `services/search.py` (SQLite
  substring scan / Postgres tsvector) — intentionally **no** `search_backend`
  config knob (removed `6599935`; nothing ever read it).
- Engine endpoint root is `/api/v1` (`/healthz`, `/system/status`, …).
- Dev loop: `./desktop-dev.sh` (inotifywait restarts app on `.kt`; uvicorn
  `--reload` for backend). Debian path: `INSTALL-debian.md`.

**Done this cycle (Sep 14–15):** batch transcribe-all with overall +
per-file stage progress; library multi-select/bulk delete (single delete
path); queued-file delete/rename; folder picker in-app + path in Settings;
zoom keys; search-within-recording; click-hit-to-playback; model choice
narrowed to Base + Large v3; Debian packaging + install guide; standalone
provisioning with vendored base model.

**Open items / candidates:**
1. Stale-noise cleanup: `~/.config/shonar-desktop/pump-error.log` holds old
   Android→server upload 404s (pre-desktop pivot); no action taken.
2. Optional: full-length Large v3 transcription timing (300 s clip ran well
   under 7 min on CPU int8 while the app + LLM were also active; a quiet
   benchmark would pin the number).

**Known debt:** none blocking. No dead stubs; one former TODO (search
backend) resolved by deletion.