app.py

"""JamAI long-track rating app -- Hugging Face Space entrypoint.

Wraps s04_longtrack_rating_app.py (copied verbatim into the Space repo,
next to this file, with longtrack_pool/ beside it -- s04 resolves both
relative to its own location, so no path surgery is needed).

What this wrapper adds, and why:
1. RATINGS PERSISTENCE. Space containers have EPHEMERAL filesystems --
   anything written locally is wiped on restart. A background
   CommitScheduler pushes logs/ to a PRIVATE HF dataset repo every 2
   minutes. Each server boot writes its own uniquely-named CSV
   (longtrack_ratings_<bootid>.csv): CommitScheduler mirrors folder
   state, so reusing one filename would let a fresh boot's empty file
   OVERWRITE history in the dataset -- unique-per-boot files only ever
   ADD. Analysis concatenates logs/*.csv from the dataset repo.
2. LOCKED WRITES. CSV appends take scheduler.lock so a commit never
   snapshots a half-written row.
3. Plain launch() -- Spaces serves the app directly; no share tunnel,
   no relay, stable URL.

Space setup (one-time, ~10 min -- see README.md for details):
  - Create the Space (SDK: Gradio), upload this folder's files +
    s04_longtrack_rating_app.py + longtrack_pool/.
  - Space settings -> Secrets: HF_TOKEN = a WRITE-scoped token.
  - Optionally set RATINGS_DATASET (else edit the default below).
Without HF_TOKEN the app still runs but ratings are NOT persisted --
it prints a loud warning (fine for a quick look, never for real raters).
"""

import os
import sys
import uuid
from pathlib import Path

try:
    import spaces
    @spaces.GPU
    def _dummy_gpu():
        pass
except ImportError:
    pass

HERE = Path(__file__).resolve().parent
# On the Space, s04_longtrack_rating_app.py sits NEXT TO this file (repo
# root); for a local smoke test it is one level up (02_Conductor/).
for _p in (HERE, HERE.parent):
    if str(_p) not in sys.path:
        sys.path.insert(0, str(_p))

import s04_longtrack_rating_app as core

# ---- persistence: local logs/ -> private HF dataset repo ----------------
# Dataset namespace is derived from the TOKEN OWNER (whoami) unless
# RATINGS_DATASET overrides it -- a token can only create repos in its
# own namespace (a hardcoded guess 403s, as the first deploy showed).
BOOT_ID = uuid.uuid4().hex[:8]
core.RATINGS_CSV = core.LOG_DIR / f"longtrack_ratings_{BOOT_ID}.csv"
core.LOG_DIR.mkdir(parents=True, exist_ok=True)

# Persistence engages only when actually running on a Space (SPACE_ID is
# set by HF) AND a token secret exists; any failure degrades to
# no-persistence with a loud warning rather than killing the app.
scheduler = None
if os.environ.get("SPACE_ID") and os.environ.get("HF_TOKEN"):
    try:
        from huggingface_hub import CommitScheduler, HfApi
        DATASET_ID = os.environ.get("RATINGS_DATASET") or (
            HfApi(token=os.environ["HF_TOKEN"]).whoami()["name"]
            + "/jamai-longtrack-ratings")
        scheduler = CommitScheduler(
            repo_id=DATASET_ID, repo_type="dataset", private=True,
            folder_path=str(core.LOG_DIR), path_in_repo="logs",
            every=2, token=os.environ["HF_TOKEN"],
        )
        _orig_append = core.append_log

        def _locked_append(row, csv_path=None):
            with scheduler.lock:
                _orig_append(row, csv_path)

        core.append_log = _locked_append
        print(f"ratings persist to dataset {DATASET_ID} "
              f"(file logs/longtrack_ratings_{BOOT_ID}.csv, every 2 min)")
    except Exception as e:  # bad token / wrong namespace / network
        scheduler = None
        print(f"persistence setup FAILED ({e}) -- running WITHOUT it")
if scheduler is None:
    print("=" * 70)
    print("WARNING: ratings will NOT survive a restart (no persistence "
          "active). On the Space, set the HF_TOKEN secret (write scope) "
          "and check RATINGS_DATASET before real raters use this.")
    print("=" * 70)

tracks = core.load_tracks()
print(f"loaded {len(tracks)} tracks")
demo = core.build_demo(tracks)
demo.launch(share=True, ssr_mode=False)