fix(llm): AttributeError instead of BadRequestError on list-typed instruction content

_validate_messages in openai.py, anthropic.py, and google.py all check
`(m.content or "").strip()` to detect empty system/developer content.
But `SystemMessage`/`DeveloperMessage.content` is typed
`str | Sequence[TextContent]` -- a plain-dict payload (e.g. content as
a list of `{"type": "text", "text": ...}` blocks, matching real
provider API shapes) validates fine against the Pydantic model, but a
non-empty list is truthy, so `.strip()` is called on a `list` and
raises AttributeError instead of the intended graceful
BadRequestError -- or, for genuinely non-empty content, crashes a
request that should have succeeded.

Fix: use the existing `.text` property (already defined on these
message models for exactly this normalization) instead of raw
`.content`. Same one-line fix duplicated across all three providers.

Added one regression test per provider constructing a developer
message with list-typed content and asserting the completion
succeeds instead of raising AttributeError. TDD red->green verified:
reverting only the three provider files reproduces `AttributeError:
'list' object has no attribute 'strip'` for all three; reapplying
passes. Full libs/giskard-llm unit suite (211 passed, 7 skipped);
ruff/ruff-format clean; basedpyright 0 errors.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: chuenchen309

libs/giskard-llm/src/giskard/llm/providers/anthropic.py

@@ -224,7 +224,7 @@ def _validate_messages(self, messages: Sequence[ChatMessage]) -> None:
                 raise BadRequestError(
                     400, "Tool messages must have a tool_call_id.", PROVIDER
                 )
-            if m.role in _ANTHROPIC_INSTRUCTION_ROLES and not (m.content or "").strip():
+            if m.role in _ANTHROPIC_INSTRUCTION_ROLES and not (m.text or "").strip():
                 raise BadRequestError(
                     400,
                     "System and developer messages must have non-empty content.",

libs/giskard-llm/src/giskard/llm/providers/google.py

@@ -331,7 +331,7 @@ def _validate_messages(self, messages: Sequence[ChatMessage]) -> None:
                 raise BadRequestError(
                     400, "Tool messages must have a tool_call_id.", PROVIDER
                 )
-            if m.role in _INSTRUCTION_ROLES and not (m.content or "").strip():
+            if m.role in _INSTRUCTION_ROLES and not (m.text or "").strip():
                 raise BadRequestError(
                     400, "System messages must have non-empty content.", PROVIDER
                 )

libs/giskard-llm/src/giskard/llm/providers/openai.py

@@ -222,7 +222,7 @@ def _validate_messages(self, messages: Sequence[ChatMessage]) -> None:
                 raise BadRequestError(
                     400, "Tool messages must have a tool_call_id.", self._PROVIDER
                 )
-            if m.role in _INSTRUCTION_ROLES and not (m.content or "").strip():
+            if m.role in _INSTRUCTION_ROLES and not (m.text or "").strip():
                 raise BadRequestError(
                     400, "System messages must have non-empty content.", self._PROVIDER
                 )

libs/giskard-llm/tests/test_providers.py

@@ -596,6 +596,27 @@ async def test_openai_validate_empty_system_content(mock_import):
         )
 
 
+@patch("giskard.llm.providers.openai._import_openai")
+@pytest.mark.openai
+async def test_openai_validate_developer_content_as_content_blocks(mock_import):
+    """Developer ``content`` typed as a list of ``TextContent`` blocks (the SDK-native
+    shape, reachable via plain-dict validation) must be normalized through ``.text``,
+    not crash with ``AttributeError`` from calling ``.strip()`` on a list."""
+    mock_import.return_value = MagicMock()
+    provider = _make_openai_provider()
+    provider._client.chat.completions.create = AsyncMock(
+        return_value=_make_openai_response("Hello")
+    )
+    resp = await provider.complete(
+        "gpt-4o",
+        [
+            {"role": "developer", "content": [{"type": "text", "text": "Be helpful"}]},
+            {"role": "user", "content": "Hi"},
+        ],
+    )
+    assert resp.choices[0].message.content == "Hello"
+
+
 @patch("giskard.llm.providers.openai._import_openai")
 @pytest.mark.openai
 async def test_openai_multiple_system_works(mock_import):
@@ -719,6 +740,31 @@ async def test_anthropic_validate_empty_developer_content(mock_import):
         )
 
 
+@patch("giskard.llm.providers.anthropic._import_anthropic")
+async def test_anthropic_validate_developer_content_as_content_blocks(mock_import):
+    """Developer ``content`` typed as a list of ``TextContent`` blocks must be
+    normalized through ``.text``, not crash with ``AttributeError`` from calling
+    ``.strip()`` on a list."""
+    mock_import.return_value = MagicMock()
+    provider = _make_anthropic_provider()
+
+    mock_raw = MagicMock()
+    mock_raw.content = [SimpleNamespace(type="text", text="Hello")]
+    mock_raw.stop_reason = "end_turn"
+    mock_raw.model = "claude-3"
+    mock_raw.usage = SimpleNamespace(input_tokens=10, output_tokens=5)
+    provider._client.messages.create = AsyncMock(return_value=mock_raw)
+
+    resp = await provider.complete(
+        "claude-3",
+        [
+            {"role": "developer", "content": [{"type": "text", "text": "Be helpful"}]},
+            {"role": "user", "content": "Hi"},
+        ],
+    )
+    assert resp.choices[0].message.content == [TextContent(text="Hello")]
+
+
 @patch("giskard.llm.providers.anthropic._import_anthropic")
 async def test_anthropic_validate_alternation(mock_import):
     mock_import.return_value = MagicMock()
@@ -734,6 +780,47 @@ async def test_anthropic_validate_alternation(mock_import):
         )
 
 
+# -- Google message validation --------------------------------------------------
+
+
+@patch("giskard.llm.providers.google._import_genai_errors")
+@pytest.mark.google
+async def test_google_validate_developer_content_as_content_blocks(mock_errors):
+    """Developer ``content`` typed as a list of ``TextContent`` blocks must be
+    normalized through ``.text``, not crash with ``AttributeError`` from calling
+    ``.strip()`` on a list."""
+    genai_types = pytest.importorskip("google.genai.types")
+    mock_errors.return_value = MagicMock()
+    provider = _make_google_provider()
+    provider._client.aio = MagicMock()
+    provider._client.aio.models = MagicMock()
+    raw = genai_types.GenerateContentResponse.model_validate(
+        {
+            "candidates": [
+                {
+                    "content": {"parts": [{"text": "Hello"}]},
+                    "finish_reason": "STOP",
+                }
+            ],
+            "usage_metadata": {
+                "prompt_token_count": 3,
+                "candidates_token_count": 4,
+                "total_token_count": 7,
+            },
+        }
+    )
+    provider._client.aio.models.generate_content = AsyncMock(return_value=raw)
+
+    resp = await provider.complete(
+        "gemini-3.5-flash",
+        [
+            {"role": "developer", "content": [{"type": "text", "text": "Be helpful"}]},
+            {"role": "user", "content": "Hi"},
+        ],
+    )
+    assert resp.choices[0].message.content == [TextContent(text="Hello")]
+
+
 # -- OpenAI Responses API (respond) -------------------------------------------