From b91b0b3c3a018dfd5c8306a3a0380326f4b76f0b Mon Sep 17 00:00:00 2001
From: RaresKeY <158580472+RaresKeY@users.noreply.github.com>
Date: Sat, 18 Jul 2026 13:41:07 +0000
Subject: [PATCH] fix(models): preserve Hugging Face identity-only cards

---
 src/model_capability_readers/huggingface.py |  2 +-
 src/provider_capability_schemas.py          |  8 ++--
 tests/test_provider_capability_schemas.py   | 42 +++++++++++++++++++++
 3 files changed, 47 insertions(+), 5 deletions(-)

diff --git a/src/model_capability_readers/huggingface.py b/src/model_capability_readers/huggingface.py
index f6f5ec6fc..bdbfb45ad 100644
--- a/src/model_capability_readers/huggingface.py
+++ b/src/model_capability_readers/huggingface.py
@@ -136,7 +136,7 @@ def records_from_payload(
     endpoint_id: Any = "",
     base_url: Any = "",
 ) -> tuple[ModelCapabilityRecord, ...]:
-    if isinstance(payload, Mapping) and "pipeline_tag" in payload:
+    if isinstance(payload, Mapping):
         record = record_from_model(payload, endpoint_id=endpoint_id, base_url=base_url)
         return (record,) if record else ()
     if not isinstance(payload, (list, tuple)):
diff --git a/src/provider_capability_schemas.py b/src/provider_capability_schemas.py
index 71b86162c..faffb917e 100644
--- a/src/provider_capability_schemas.py
+++ b/src/provider_capability_schemas.py
@@ -394,8 +394,10 @@ HUGGINGFACE_MODEL_SHAPE = ProviderCatalogShape(
     provider_id="huggingface",
     envelope=ENVELOPE_SINGLE,
     identity_paths=("modelId", "id"),
-    required_item_paths=("pipeline_tag",),
-    item_types=(("pipeline_tag", (str,)),),
+    # Hub ModelInfo exposes pipeline_tag as optional metadata.  Provider/host
+    # context is still required because this shape has priority zero, so an
+    # identity-only card can stay native without making generic ``id`` payloads
+    # look like Hugging Face catalogs.
     detection_priority=0,
 )
 HUGGINGFACE_MODELS_LIST_SHAPE = ProviderCatalogShape(
@@ -403,8 +405,6 @@ HUGGINGFACE_MODELS_LIST_SHAPE = ProviderCatalogShape(
     provider_id="huggingface",
     envelope=ENVELOPE_BARE_LIST,
     identity_paths=("modelId", "id"),
-    required_item_paths=("pipeline_tag",),
-    item_types=(("pipeline_tag", (str,)),),
     detection_priority=0,
 )
 COHERE_MODELS_SHAPE = ProviderCatalogShape(
diff --git a/tests/test_provider_capability_schemas.py b/tests/test_provider_capability_schemas.py
index bf814d678..228f5120f 100644
--- a/tests/test_provider_capability_schemas.py
+++ b/tests/test_provider_capability_schemas.py
@@ -1001,6 +1001,48 @@ def test_huggingface_reader_maps_provider_specific_pipeline_metadata():
     assert record.capability.confidence == mc.CONFIDENCE_REGISTRY
 
 
+def test_huggingface_optional_pipeline_tag_preserves_identity_only_records():
+    cases = (
+        (
+            {"modelId": "org/no-pipeline-tag"},
+            "huggingface.hub.model-info.v1",
+        ),
+        (
+            {"modelId": "org/null-pipeline-tag", "pipeline_tag": None},
+            "huggingface.hub.model-info.v1",
+        ),
+        (
+            [{"modelId": "org/list-no-pipeline-tag"}],
+            "huggingface.hub.model-info-list.v1",
+        ),
+        (
+            [{"modelId": "org/list-null-pipeline-tag", "pipeline_tag": None}],
+            "huggingface.hub.model-info-list.v1",
+        ),
+    )
+
+    for payload, shape_id in cases:
+        resolution = pcs.resolve_provider(payload, provider="huggingface")
+        direct = huggingface.records_from_payload(payload)
+        wrapped = records_from_payload(payload, vendor="huggingface")
+
+        assert resolution.shape_id == shape_id
+        assert resolution.fallback is False
+        assert len(direct) == 1
+        assert len(wrapped) == 1
+        assert wrapped[0].model_id == direct[0].model_id
+        assert wrapped[0].capability.family == mc.FAMILY_UNKNOWN
+        assert wrapped[0].capability.capabilities == ()
+        assert wrapped[0].catalog_shape_id == shape_id
+        assert wrapped[0].fallback is False
+
+    # An identity-only singleton remains insufficient to infer Hugging Face
+    # without configured provider or host context.
+    unscoped = {"modelId": "org/unscoped"}
+    assert pcs.resolve_provider(unscoped).provider_id == pcs.PROVIDER_UNKNOWN
+    assert records_from_payload(unscoped) == ()
+
+
 def test_cohere_reader_maps_only_native_endpoint_and_limit_fields():
     chat, ambiguous = cohere.records_from_payload(
         {