feat: ml-worker self-heals models; restore modal tag-input auto-focus

ml-worker now fetches WD14 + SigLIP on container start via an entrypoint
script that calls huggingface_hub with a populated local_dir. HF does a
content-based integrity check, so subsequent starts are a no-op. Transient
HF outages warn and continue when a local copy is already present. This
drops the baked-in model layer (~4 GB) from the image so it pushes cleanly
through registry proxies and no longer bloats every tagged release.

view-modal.js restores the desktop-only auto-focus of the tag input on
openModal (with touch-device guard). The block existed in an earlier
main commit (0bf1961) but was absent from the post-corruption
consolidation commit.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

.gitignore

@@ -14,3 +14,4 @@ venv/
 
 # Claude files
 .claude/
+.remember/

Dockerfile.ml

@@ -22,14 +22,9 @@ RUN pip install --no-cache-dir \
         --index-url https://download.pytorch.org/whl/cpu \
         "torch>=2.5"
 
-# Pre-install only what the model downloader needs, then fetch models. By keeping the model-download
-# layer above the full requirements install, a change to requirements-ml.txt won't invalidate the
-# (slow, ~2 GB) model download layer.
-RUN pip install --no-cache-dir "huggingface_hub>=0.25"
-COPY scripts/download_models.py /app/scripts/download_models.py
-RUN mkdir -p ${ML_MODEL_DIR} && python /app/scripts/download_models.py
-
-# Install the rest of the ML requirements. Model files are already on disk from the step above.
+# Install ML requirements. Models are NOT baked in — the entrypoint fetches them
+# at container start into the ${ML_MODEL_DIR} volume, and huggingface_hub's
+# content-based integrity check makes subsequent starts a no-op.
 COPY requirements-ml.txt /app/requirements-ml.txt
 RUN pip install --no-cache-dir -r requirements-ml.txt
 
@@ -37,6 +32,9 @@ RUN pip install --no-cache-dir -r requirements-ml.txt
 # config.py lives at the repo root and is imported by app/__init__.py via 'config.Config'.
 COPY app /app/app
 COPY config.py /app/config.py
+COPY scripts/ensure_models.py /app/scripts/ensure_models.py
+COPY scripts/entrypoint-ml.sh /app/scripts/entrypoint-ml.sh
+RUN chmod +x /app/scripts/entrypoint-ml.sh && mkdir -p ${ML_MODEL_DIR}
 
-# Celery needs these env vars set at runtime via docker-compose.
+ENTRYPOINT ["/app/scripts/entrypoint-ml.sh"]
 CMD ["celery", "-A", "app.celery_app:celery", "worker", "--loglevel=info", "-Q", "ml", "--concurrency=1"]

app/static/js/view-modal.js

@@ -811,6 +811,14 @@ document.addEventListener('DOMContentLoaded', () => {
     modal.classList.add('active');
     updateImage(index);
     history.replaceState({ modalIndex: index }, '', window.location.href);
+    // Desktop only — skip on touch devices to avoid popping the on-screen keyboard
+    if (tagInput && !isTouchDevice()) {
+      setTimeout(() => tagInput.focus(), 100);
+    }
+  }
+
+  function isTouchDevice() {
+    return ('ontouchstart' in window) || (navigator.maxTouchPoints > 0);
   }
 
   function closeModal() {

scripts/ensure_models.py

@@ -0,0 +1,84 @@
+"""Ensure WD14 and SigLIP model files exist under $ML_MODEL_DIR.
+
+Run at container start by entrypoint-ml.sh. huggingface_hub performs content-based
+integrity checks: matching files are skipped, changed/missing files are fetched.
+If the remote is unreachable but a local copy is already present, log a warning
+and continue so a transient HF outage does not take a healthy worker down.
+"""
+from __future__ import annotations
+
+import logging
+import os
+import sys
+
+from huggingface_hub import hf_hub_download, snapshot_download
+from huggingface_hub.utils import HfHubHTTPError
+
+logging.basicConfig(level=logging.INFO, format='%(asctime)s %(levelname)s %(name)s: %(message)s')
+log = logging.getLogger('ensure_models')
+
+MODEL_DIR = os.environ.get('ML_MODEL_DIR', '/models')
+WD14_REPO = os.environ.get('WD14_REPO', 'SmilingWolf/wd-eva02-large-tagger-v3')
+SIGLIP_REPO = os.environ.get('SIGLIP_REPO', 'google/siglip-so400m-patch14-384')
+WD14_REVISION = os.environ.get('WD14_REVISION') or None
+SIGLIP_REVISION = os.environ.get('SIGLIP_REVISION') or None
+
+WD14_FILES = ('model.onnx', 'selected_tags.csv')
+SIGLIP_CORE_FILES = (
+    'config.json',
+    'preprocessor_config.json',
+    'tokenizer_config.json',
+    'tokenizer.json',
+    'special_tokens_map.json',
+    'spiece.model',
+    'model.safetensors',
+)
+
+
+def _all_present(dir_path: str, files: tuple[str, ...]) -> bool:
+    return all(os.path.exists(os.path.join(dir_path, f)) for f in files)
+
+
+def ensure_wd14() -> None:
+    target = os.path.join(MODEL_DIR, 'wd14')
+    os.makedirs(target, exist_ok=True)
+    try:
+        for fname in WD14_FILES:
+            hf_hub_download(
+                repo_id=WD14_REPO,
+                filename=fname,
+                revision=WD14_REVISION,
+                local_dir=target,
+            )
+        log.info('WD14 ready at %s', target)
+    except (HfHubHTTPError, OSError) as exc:
+        if _all_present(target, WD14_FILES):
+            log.warning('WD14 refresh failed (%s); continuing with local copy', exc)
+        else:
+            log.error('WD14 missing and cannot download: %s', exc)
+            raise
+
+
+def ensure_siglip() -> None:
+    target = os.path.join(MODEL_DIR, 'siglip')
+    os.makedirs(target, exist_ok=True)
+    try:
+        snapshot_download(
+            repo_id=SIGLIP_REPO,
+            revision=SIGLIP_REVISION,
+            local_dir=target,
+            allow_patterns=list(SIGLIP_CORE_FILES),
+        )
+        log.info('SigLIP ready at %s', target)
+    except (HfHubHTTPError, OSError) as exc:
+        if _all_present(target, SIGLIP_CORE_FILES):
+            log.warning('SigLIP refresh failed (%s); continuing with local copy', exc)
+        else:
+            log.error('SigLIP missing and cannot download: %s', exc)
+            raise
+
+
+if __name__ == '__main__':
+    ensure_wd14()
+    ensure_siglip()
+    sys.exit(0)

scripts/entrypoint-ml.sh

@@ -0,0 +1,4 @@
+#!/bin/sh
+set -e
+python /app/scripts/ensure_models.py
+exec "$@"