fix(medcat-trainer): Fix issue with meta-annotation tasks disappearing (#558)

* Add test for signals

* Add fix for signals to allow keeping manually selected meta tasks

* Add comment regarding multiple sources of meta-tasks

---------

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

Author: mart-r

medcat-trainer/webapp/api/api/signals.py

@@ -107,8 +107,31 @@ def project_tasks_changed(sender, instance, action, **kwargs):
     # post_remove or post_add actions, overwrite to model_pack supplied MetaCAT tasks.
     if (action.startswith('post') and isinstance(instance, ProjectAnnotateEntitiesFields) and
             instance.model_pack is not None):
-        instance.tasks.set([MetaTask.objects.filter(prediction_model_id=meta_cat.id).first() for meta_cat in
-                            instance.model_pack.meta_cats.all()])
+        # NOTE: This part deals with two different sources of information:
+        #       1. sometimes the model pack associated with the project can have meta-cats for meta-annotations
+        #       2. sometimes the project itself defines meta-tasks for the annotator to use
+        #
+        #       Currently the proccess here defaults to useing model-pack defined meta-tasks (if present),
+        #       while allowing for the project-defined ones otherwise.
+
+        # Find automated tasks from the model pack
+        db_tasks = [
+            MetaTask.objects.filter(prediction_model_id=meta_cat.id).first()
+            for meta_cat in instance.model_pack.meta_cats.all()
+        ]
+        # Filter out None values
+        automated_tasks = [t for t in db_tasks if t is not None]
+
+        # Only overwrite if the model pack actually brought automated tasks to the table.
+        # This preserves manual workflows when training from scratch.
+        if automated_tasks:
+            # Disconnect the signal temporarily to prevent infinite recursion loops
+            m2m_changed.disconnect(project_tasks_changed, sender=ProjectAnnotateEntitiesFields.tasks.through)
+            try:
+                instance.tasks.set(automated_tasks)
+            finally:
+                # Always reconnect the signal
+                m2m_changed.connect(project_tasks_changed, sender=ProjectAnnotateEntitiesFields.tasks.through)
 
 
 m2m_changed.connect(project_tasks_changed, sender=ProjectAnnotateEntitiesFields.tasks.through)

medcat-trainer/webapp/api/api/tests/test_signals.py

@@ -170,3 +170,44 @@ def test_syncs_tasks_from_model_pack_meta_cats(self):
             action='post_add',
         )
         self.assertIn(task, project.tasks.all())
+
+    def test_manual_tasks_persist_when_model_pack_has_no_meta_cats(self):
+        # 1. Create a manual task not tied to any automated prediction model
+        manual_task = MetaTask.objects.create(name='ManualPresenceTask')
+
+        # 2. Build a lightweight ModelPack that does NOT contain any meta_cats
+        cdb = ConceptDB(name='manual-cdb', cdb_file='manual-cdb.dat')
+        cdb.save(skip_load=True)
+        vocab = Vocabulary(name='manual-vocab', vocab_file='manual-vocab.dat')
+        vocab.save(skip_load=True)
+
+        mp = ModelPack(name='base-ner-pack', concept_db=cdb, vocab=vocab)
+        mp.save(skip_load=True)
+
+        # Write a dummy zip file onto the filesystem so model_pack file validation passes
+        pack_path = os.path.join(settings.MEDIA_ROOT, 'manual_base.zip')
+        with open(pack_path, 'wb') as fh:
+            fh.write(b'fake-zip')
+        ModelPack.objects.filter(pk=mp.pk).update(model_pack='manual_base.zip')
+        mp.refresh_from_db()
+
+        # 3. Create a project and attach this empty ModelPack
+        project = create_basic_project(name='manual-tasks-proj')
+        project.model_pack = mp
+        project.concept_db = None
+        project.vocab = None
+        project.save()
+
+        # 4. Add the manual task to the project's many-to-many field
+        project.tasks.add(manual_task)
+
+        # 5. Manually trigger the signal matching a Django admin 'post_add' save operation
+        api_signals.project_tasks_changed(
+            sender=ProjectAnnotateEntitiesFields.tasks.through,
+            instance=project,
+            action='post_add',
+        )
+
+        # 6. Assert that our manual task survived the signal handler execution
+        self.assertIn(manual_task, project.tasks.all())
+        self.assertEqual(project.tasks.count(), 1)