Add comprehensive tests for provider plugin system

.github/workflows/ci.yaml

@@ -133,6 +133,48 @@ jobs:
           fi
           tox -e live-api
 
+  plugin-integration-test:
+    needs: test
+    runs-on: ubuntu-latest
+    if: github.event_name == 'pull_request'
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Detect provider-related changes
+        id: provider-changes
+        uses: tj-actions/changed-files@v46
+        with:
+          files: |
+            langextract/providers/**
+            langextract/factory.py
+            langextract/inference.py
+            tests/provider_plugin_test.py
+            pyproject.toml
+            .github/workflows/ci.yaml
+
+      - name: Skip if no provider changes
+        if: steps.provider-changes.outputs.any_changed == 'false'
+        run: |
+          echo "No provider-related changes detected – skipping plugin integration test."
+          exit 0
+
+      - name: Set up Python 3.11
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.11"
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install tox
+
+      - name: Run plugin smoke test
+        run: tox -e plugin-smoke
+
+      - name: Run plugin integration test
+        run: tox -e plugin-integration
+
   ollama-integration-test:
     needs: test
     runs-on: ubuntu-latest

langextract/providers/__init__.py

@@ -69,27 +69,42 @@ def load_plugins_once() -> None:
     _PLUGINS_LOADED = True
     return
 
-  _PLUGINS_LOADED = True
-
   try:
     entry_points_group = metadata.entry_points(group="langextract.providers")
   except Exception as exc:
     logging.debug("No third-party provider entry points found: %s", exc)
     return
 
+  # Set flag after successful entry point query to avoid disabling discovery
+  # on transient failures during enumeration.
+  _PLUGINS_LOADED = True
+
   for entry_point in entry_points_group:
     try:
       provider = entry_point.load()
-
+      # Validate provider subclasses but don't auto-register - plugins must
+      # use their own @register decorators to control patterns.
       if isinstance(provider, type):
-        registry.register(entry_point.name)(provider)
-        logging.info(
-            "Registered third-party provider from entry point: %s",
-            entry_point.name,
-        )
+        # pylint: disable=import-outside-toplevel
+        # Late import to avoid circular dependency
+        from langextract import inference
+
+        if issubclass(provider, inference.BaseLanguageModel):
+          logging.info(
+              "Loaded third-party provider class from entry point: %s",
+              entry_point.name,
+          )
+        else:
+          logging.warning(
+              "Entry point %s returned non-provider class %r; ignoring",
+              entry_point.name,
+              provider,
+          )
       else:
+        # Module import triggers decorator execution
         logging.debug(
-            "Loaded provider module from entry point: %s", entry_point.name
+            "Loaded provider module/object from entry point: %s",
+            entry_point.name,
         )
     except Exception as exc:
       logging.warning(

pyproject.toml

@@ -92,6 +92,8 @@ python_functions = "test_*"
 addopts = "-ra"
 markers = [
     "live_api: marks tests as requiring live API access",
+    "requires_pip: marks tests that perform pip install/uninstall operations",
+    "integration: marks integration tests that test multiple components together",
 ]
 
 [tool.pyink]