test: import-integrity and packaging guards + 8 latent bug fixes

.github/workflows/release.yml

@@ -57,3 +57,21 @@ jobs:
         uses: python-semantic-release/publish-action@v9.15.2
         with:
           github_token: ${{ secrets.ADMIN_TOKEN }}
+
+      # Releases failed silently for 3 months (Mar-Jun 2026) while PyPI
+      # went stale; see OpenAdaptAI/OpenAdapt#999. Fail loudly instead.
+      - name: File issue on release failure
+        if: failure()
+        env:
+          GH_TOKEN: ${{ secrets.ADMIN_TOKEN }}
+        run: |
+          TITLE="Release workflow failed on main"
+          BODY="The release workflow failed: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+
+          Until this is fixed, merged fix/feat commits are NOT being published to PyPI, and users install stale versions."
+          EXISTING=$(gh issue list --repo "${{ github.repository }}" --state open --search "in:title \"$TITLE\"" --json number --jq '.[0].number // empty')
+          if [ -n "$EXISTING" ]; then
+            gh issue comment "$EXISTING" --repo "${{ github.repository }}" --body "$BODY"
+          else
+            gh issue create --repo "${{ github.repository }}" --title "$TITLE" --body "$BODY"
+          fi

openadapt_ml/cloud/azure_inference.py

@@ -424,10 +424,10 @@ def main():
 
     elif args.command == "inference-worker":
         # Start inference worker
-        from openadapt_ml.adapters.qwen import QwenVLAdapter
+        from openadapt_ml.models.qwen_vl import QwenVLAdapter
 
         print(f"Starting inference worker with model: {args.model}")
-        adapter = QwenVLAdapter(model_name=args.model)
+        adapter = QwenVLAdapter.from_pretrained(args.model)
         queue.poll_and_process(adapter)
 
     elif args.command == "inference-watch":

openadapt_ml/cloud/local.py

@@ -499,9 +499,7 @@ def cmd_serve(args: argparse.Namespace) -> int:
         if not args.no_regenerate:
             print("Regenerating dashboard and viewer...")
             try:
-                # Use keep_polling=True so JavaScript fetches live data from training_log.json
-                # This ensures the dashboard shows current data instead of stale embedded data
-                regenerate_local_dashboard(str(serve_dir), keep_polling=True)
+                regenerate_local_dashboard(str(serve_dir))
                 # Also regenerate viewer if comparison data exists
                 _regenerate_viewer_if_possible(serve_dir)
             except Exception as e:

openadapt_ml/evals/grounding.py

@@ -19,8 +19,8 @@
 if TYPE_CHECKING:
     from PIL import Image
 
-    from openadapt_ml.data.types import Episode
     from openadapt_ml.grounding.base import GroundingModule, RegionCandidate
+    from openadapt_ml.schema import Episode
 
 
 @dataclass

openadapt_ml/experiments/demo_prompt/run_experiment.py

@@ -423,7 +423,7 @@ def run_experiment(
     from openadapt_ml.ingest.capture import capture_to_episode
 
     print(f"Loading demo from: {demo_capture_path}")
-    episode = capture_to_episode(demo_capture_path, goal=goal)
+    episode = capture_to_episode(demo_capture_path, instruction=goal)
     print(f"  Loaded {len(episode.steps)} steps, goal: {episode.goal}")
 
     print(f"\nTest task: {test_task}")

openadapt_ml/ingest/__init__.py

@@ -11,8 +11,8 @@
     - load_episodes(): Load Episodes from JSON files (primary entry point)
     - save_episodes(): Save Episodes to JSON file
     - capture_to_episode(): Converts one openadapt-capture recording → one Episode
-    - capture_to_session(): Converts one recording → Session containing one Episode
-    - load_captures_as_sessions(): Loads multiple recordings → list of Sessions
+    - capture_to_episodes(): Converts one recording → list of Episodes
+    - load_captures_as_episodes(): Loads multiple recordings → list of Episodes
     - generate_synthetic_episodes(): Creates synthetic training data
 """
 
@@ -29,15 +29,15 @@
 try:
     from openadapt_ml.ingest.capture import (  # noqa: F401
         capture_to_episode,
-        capture_to_session,
-        load_captures_as_sessions,
+        capture_to_episodes,
+        load_captures_as_episodes,
     )
 
     __all__.extend(
         [
             "capture_to_episode",
-            "capture_to_session",
-            "load_captures_as_sessions",
+            "capture_to_episodes",
+            "load_captures_as_episodes",
         ]
     )
 except ImportError:

openadapt_ml/scripts/compare.py

@@ -202,6 +202,29 @@ def generate_comparison_data(
     return comparison_data
 
 
+def generate_comparison(
+    capture_path: str | Path,
+    adapter=None,
+    output_path: str | Path | None = None,
+) -> Path:
+    """Generate a comparison HTML for a capture using a loaded adapter.
+
+    Convenience wrapper over generate_comparison_data and
+    generate_comparison_html for callers that hold a capture path and a
+    model adapter (e.g. openadapt_ml.cloud.azure_inference).
+    """
+    from openadapt_ml.ingest.capture import capture_to_episode
+
+    capture_path = Path(capture_path)
+    episode = capture_to_episode(capture_path)
+    comparison_data = generate_comparison_data(episode, model=adapter)
+    if output_path is None:
+        output_path = capture_path / "comparison.html"
+    output_path = Path(output_path)
+    generate_comparison_html(capture_path, episode, comparison_data, output_path)
+    return output_path
+
+
 def generate_comparison_html(
     capture_path: Path,
     episode: Episode,
@@ -804,7 +827,7 @@ def main():
 
     # Convert capture to episode
     print(f"Loading capture from: {capture_path}")
-    episode = capture_to_episode(capture_path, goal=args.goal)
+    episode = capture_to_episode(capture_path, instruction=args.goal)
     print(f"Loaded {len(episode.steps)} steps")
 
     # Load model if checkpoint provided

pyproject.toml

@@ -66,6 +66,7 @@ training = [
 # See https://docs.unsloth.ai/get-started/installation
 
 dev = [
+    "build>=1.0.0",  # tests/test_packaging.py builds the wheel
     "pytest>=9.0.0",
     "ruff>=0.1.0",
 ]