Prune stems whose source song was removed or replaced

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

server/stemd.py

@@ -21,6 +21,9 @@ Settings (environment):
   DEVICE      cuda             cuda, cuda:N or cpu
   RESCAN_MIN  30               minutes between scans once caught up
   AUTO_SCAN   0                1 = scan the library on a timer; 0 = only separate Stem Now requests
+  PRUNE       1                remove stems whose song left the library; 0 = off
+  PRUNE_CHANGED 0              1 = also remove stems of songs replaced in place (redone by the next scan)
+  PRUNE_MAX   0.2              skip a prune that would remove more than this share of the stems
   FP16        1                half precision on the GPU
   OVERLAP / SEGMENT / SHIFTS   as in the desktop app's settings
 """
@@ -40,8 +43,8 @@ sys.path.insert(0, os.path.join(_ROOT, "desktop"))  # in the image the package s
 sys.path.insert(0, _ROOT)
 
 from stemify.catalog import MODELS_BY_ID  # noqa: E402
-from stemify.library_core import (MARKER, MAX_SECONDS, SOURCE_RANK, duration, is_done, is_share_error,  # noqa: E402
-                                     iter_tracks, signature, write_track)
+from stemify.library_core import (MARKER, MAX_SECONDS, SOURCE_RANK, duration, find_stale, is_done,  # noqa: E402
+                                     is_share_error, iter_tracks, remove_stems, signature, write_track)
 
 MUSIC_DIR = os.environ.get("MUSIC_DIR", "/music")
 STEMS_DIR = os.environ.get("STEMS_DIR", "/stems")
@@ -51,6 +54,9 @@ MODEL = os.environ.get("MODEL", "roformer_bs_sw")
 DEVICE = os.environ.get("DEVICE", "cuda")
 RESCAN_S = float(os.environ.get("RESCAN_MIN", "30")) * 60
 AUTO_SCAN = os.environ.get("AUTO_SCAN", "0") in ("1", "true", "yes")
+PRUNE = os.environ.get("PRUNE", "1") not in ("0", "false", "no", "")
+PRUNE_CHANGED = os.environ.get("PRUNE_CHANGED", "0") in ("1", "true", "yes")
+PRUNE_MAX = float(os.environ.get("PRUNE_MAX", "0.2"))
 FP16 = os.environ.get("FP16", "1") not in ("0", "false", "no", "")
 OVERLAP = float(os.environ.get("OVERLAP", "0.25"))
 SEGMENT = float(os.environ.get("SEGMENT", "0"))
@@ -331,6 +337,26 @@ def serve_queue(sep, failed):
     return used
 
 
+def prune():
+    """Remove stems whose song is gone from MUSIC_DIR (and optionally replaced ones)."""
+    if not PRUNE:
+        return
+    orphans, changed = find_stale(MUSIC_DIR, STEMS_DIR)
+    drop = orphans + (changed if PRUNE_CHANGED else [])
+    if changed and not PRUNE_CHANGED:
+        log(f"prune: {len(changed):,} stems are of a replaced song (PRUNE_CHANGED=1 removes them)")
+    if not drop:
+        return
+    total = sum(1 for _ in iter_tracks(MUSIC_DIR, STEMS_DIR))
+    kept = total - len(orphans)
+    if kept <= 0 or len(drop) > PRUNE_MAX * (kept + len(drop)):
+        log(f"prune: refusing to remove {len(drop):,} of {kept + len(drop):,} stems "
+            f"(over PRUNE_MAX {PRUNE_MAX:.0%}, or the music share looks empty); run server/prune.py --apply to force")
+        return
+    remove_stems(STEMS_DIR, drop)
+    log(f"prune: removed {len(drop):,} stems folders no longer in the library")
+
+
 def run_pass(sep, failed):
     """Separate everything that needs it. Returns (tracks done, whether the GPU was used)."""
     t0 = time.monotonic()
@@ -404,6 +430,7 @@ def main():
         log("AUTO_SCAN off: only Stem Now requests will be separated")
     while not stop.is_set():
         try:
+            prune()
             if not AUTO_SCAN:
                 idle(sep, failed, RESCAN_S)
                 continue