Merge branch 'main' into feat/l1-workspace

# Conflicts:
#	frontend/src/router.tsx

backend/tests/test_ai_provider.py

@@ -96,7 +96,8 @@ class TestAnthropicProvider:
         )
 
         mock_response = MagicMock()
-        mock_response.content = [MagicMock(text='{"result": "ok"}')]
+        mock_response.content = [MagicMock(type="text", text='{"result": "ok"}')]
+        mock_response.stop_reason = "end_turn"
         mock_response.usage = MagicMock(input_tokens=100, output_tokens=50)
 
         mock_client = AsyncMock()
@@ -120,6 +121,170 @@ class TestAnthropicProvider:
             messages=[{"role": "user", "content": "Hello"}],
         )
 
+    @pytest.mark.asyncio
+    async def test_generate_json_skips_non_text_blocks(self):
+        """A leading non-text block (e.g. thinking) is skipped; the first
+        text block's text is returned instead of content[0].text."""
+        from app.core import ai_provider
+
+        ai_provider._anthropic_clients.clear()
+
+        provider = AnthropicProvider(
+            api_key="skip-key", model="claude-sonnet-4-6", timeout=31
+        )
+
+        thinking_block = MagicMock(type="thinking", thinking="hmm...")
+        text_block = MagicMock(type="text", text='{"ok": 1}')
+        mock_response = MagicMock()
+        mock_response.content = [thinking_block, text_block]
+        mock_response.stop_reason = "end_turn"
+        mock_response.usage = MagicMock(input_tokens=10, output_tokens=5)
+
+        mock_client = AsyncMock()
+        mock_client.messages.create = AsyncMock(return_value=mock_response)
+
+        with patch("anthropic.AsyncAnthropic", return_value=mock_client):
+            text, _, _ = await provider.generate_json(
+                system_prompt="You are a helper.",
+                messages=[{"role": "user", "content": "Hi"}],
+            )
+
+        assert text == '{"ok": 1}'
+
+    @pytest.mark.asyncio
+    async def test_generate_json_raises_when_no_text_block(self):
+        """A response with no text block (e.g. a bare refusal) raises a clear
+        error instead of returning a non-text block's attributes."""
+        from app.core import ai_provider
+
+        ai_provider._anthropic_clients.clear()
+
+        provider = AnthropicProvider(
+            api_key="empty-key", model="claude-sonnet-4-6", timeout=32
+        )
+
+        mock_response = MagicMock()
+        mock_response.content = [MagicMock(type="thinking", thinking="...")]
+        mock_response.stop_reason = "refusal"
+        mock_response.usage = MagicMock(input_tokens=10, output_tokens=0)
+
+        mock_client = AsyncMock()
+        mock_client.messages.create = AsyncMock(return_value=mock_response)
+
+        with patch("anthropic.AsyncAnthropic", return_value=mock_client):
+            with pytest.raises(ValueError, match="no text block"):
+                await provider.generate_json(
+                    system_prompt="You are a helper.",
+                    messages=[{"role": "user", "content": "Hi"}],
+                )
+
+    @pytest.mark.asyncio
+    async def test_generate_json_logs_warning_on_truncation(self, caplog):
+        """When stop_reason is max_tokens, a warning is logged (truncation
+        signal) and the partial text is still returned."""
+        import logging
+
+        from app.core import ai_provider
+
+        ai_provider._anthropic_clients.clear()
+
+        provider = AnthropicProvider(
+            api_key="trunc-key", model="claude-sonnet-4-6", timeout=33
+        )
+
+        text_block = MagicMock(type="text", text='{"partial": tr')
+        mock_response = MagicMock()
+        mock_response.content = [text_block]
+        mock_response.stop_reason = "max_tokens"
+        mock_response.usage = MagicMock(input_tokens=10, output_tokens=4096)
+
+        mock_client = AsyncMock()
+        mock_client.messages.create = AsyncMock(return_value=mock_response)
+
+        with patch("anthropic.AsyncAnthropic", return_value=mock_client):
+            with caplog.at_level(logging.WARNING, logger="app.core.ai_provider"):
+                text, _, _ = await provider.generate_json(
+                    system_prompt="You are a helper.",
+                    messages=[{"role": "user", "content": "Hi"}],
+                )
+
+        assert text == '{"partial": tr'
+        truncation_records = [
+            r for r in caplog.records if getattr(r, "stop_reason", None) == "max_tokens"
+        ]
+        assert truncation_records, "expected a warning record for max_tokens truncation"
+
+    @pytest.mark.asyncio
+    async def test_generate_json_passes_output_config_when_schema_given(self):
+        """When a JSON schema is supplied, it is forwarded as
+        output_config.format so the API constrains the response shape."""
+        from app.core import ai_provider
+
+        ai_provider._anthropic_clients.clear()
+
+        provider = AnthropicProvider(
+            api_key="schema-key", model="claude-sonnet-4-6", timeout=34
+        )
+
+        mock_response = MagicMock()
+        mock_response.content = [MagicMock(type="text", text='{"title": "x"}')]
+        mock_response.stop_reason = "end_turn"
+        mock_response.usage = MagicMock(input_tokens=10, output_tokens=5)
+
+        mock_client = AsyncMock()
+        mock_client.messages.create = AsyncMock(return_value=mock_response)
+
+        schema = {
+            "type": "object",
+            "properties": {"title": {"type": "string"}},
+            "required": ["title"],
+            "additionalProperties": False,
+        }
+
+        with patch("anthropic.AsyncAnthropic", return_value=mock_client):
+            await provider.generate_json(
+                system_prompt="You are a helper.",
+                messages=[{"role": "user", "content": "Hi"}],
+                max_tokens=512,
+                schema=schema,
+            )
+
+        mock_client.messages.create.assert_called_once_with(
+            model="claude-sonnet-4-6",
+            max_tokens=512,
+            system="You are a helper.",
+            messages=[{"role": "user", "content": "Hi"}],
+            output_config={"format": {"type": "json_schema", "schema": schema}},
+        )
+
+    @pytest.mark.asyncio
+    async def test_generate_json_no_output_config_when_schema_none(self):
+        """With no schema, output_config is not sent (backward compatible)."""
+        from app.core import ai_provider
+
+        ai_provider._anthropic_clients.clear()
+
+        provider = AnthropicProvider(
+            api_key="noschema-key", model="claude-sonnet-4-6", timeout=35
+        )
+
+        mock_response = MagicMock()
+        mock_response.content = [MagicMock(type="text", text="{}")]
+        mock_response.stop_reason = "end_turn"
+        mock_response.usage = MagicMock(input_tokens=1, output_tokens=1)
+
+        mock_client = AsyncMock()
+        mock_client.messages.create = AsyncMock(return_value=mock_response)
+
+        with patch("anthropic.AsyncAnthropic", return_value=mock_client):
+            await provider.generate_json(
+                system_prompt="You are a helper.",
+                messages=[{"role": "user", "content": "Hi"}],
+            )
+
+        _, call_kwargs = mock_client.messages.create.call_args
+        assert "output_config" not in call_kwargs
+
 
 class TestGeminiProvider:
     """Tests for GeminiProvider.generate_json."""
@@ -174,6 +339,48 @@ class TestGeminiProvider:
 
         mock_client.aio.models.generate_content.assert_called_once()
 
+    @pytest.mark.asyncio
+    async def test_generate_json_accepts_and_ignores_schema(self):
+        """Gemini accepts the schema kwarg (interface parity) and still
+        returns JSON; it does not error on the param."""
+        provider = GeminiProvider(api_key="test-key", model="gemini-2.5-flash")
+
+        mock_usage = MagicMock()
+        mock_usage.prompt_token_count = 5
+        mock_usage.candidates_token_count = 3
+
+        mock_response = MagicMock()
+        mock_response.text = '{"answer": 1}'
+        mock_response.usage_metadata = mock_usage
+
+        mock_client = MagicMock()
+        mock_client.aio.models.generate_content = AsyncMock(return_value=mock_response)
+
+        mock_genai_module = MagicMock()
+        mock_genai_module.Client.return_value = mock_client
+
+        mock_types = MagicMock()
+        mock_types.Content.side_effect = lambda **kw: kw
+        mock_types.Part.side_effect = lambda **kw: kw
+        mock_types.GenerateContentConfig.side_effect = lambda **kw: kw
+
+        mock_google = MagicMock()
+        mock_google.genai = mock_genai_module
+        mock_genai_module.types = mock_types
+
+        with patch.dict(sys.modules, {
+            "google": mock_google,
+            "google.genai": mock_genai_module,
+            "google.genai.types": mock_types,
+        }):
+            text, _, _ = await provider.generate_json(
+                system_prompt="Generate JSON.",
+                messages=[{"role": "user", "content": "data"}],
+                schema={"type": "object"},
+            )
+
+        assert text == '{"answer": 1}'
+
     @pytest.mark.asyncio
     async def test_generate_json_handles_none_usage(self):
         """Token counts default to 0 when usage_metadata attributes are None."""

backend/tests/test_kb_conversion_schema.py

@@ -0,0 +1,104 @@
+"""Tests for the structured-output JSON schemas used by KB conversion.
+
+These validate that the schemas are well-formed against the Anthropic
+structured-output limits (every object carries additionalProperties: false,
+`required` is a subset of declared properties, no numeric/length constraints)
+and that the target_type -> schema selector returns the right shape. They do
+NOT exercise the live API — constrained decoding must be smoke-tested against
+a real model before AI_KB_CONVERT_STRUCTURED_OUTPUT is enabled in production.
+"""
+
+from app.core.kb_conversion_service import (
+    PROCEDURAL_SCHEMA,
+    TROUBLESHOOTING_SCHEMA,
+    _schema_for_target_type,
+)
+
+# Constraints disallowed by Anthropic structured outputs (must be absent so the
+# API does not reject the schema or silently strip them).
+_DISALLOWED_KEYS = {
+    "minimum",
+    "maximum",
+    "multipleOf",
+    "minLength",
+    "maxLength",
+    "minItems",
+    "maxItems",
+}
+
+
+def _assert_well_formed(schema: dict) -> None:
+    """Recursively assert a JSON schema obeys the structured-output limits."""
+    if schema.get("type") == "object":
+        assert schema.get("additionalProperties") is False, (
+            f"object schema missing additionalProperties: false: {schema}"
+        )
+        props = schema.get("properties", {})
+        required = set(schema.get("required", []))
+        assert required <= set(props), (
+            f"required keys not all declared as properties: {required - set(props)}"
+        )
+        for sub in props.values():
+            _assert_well_formed(sub)
+    elif schema.get("type") == "array":
+        _assert_well_formed(schema["items"])
+
+    assert not (_DISALLOWED_KEYS & set(schema)), (
+        f"schema uses unsupported constraint(s): {_DISALLOWED_KEYS & set(schema)}"
+    )
+
+
+class TestStructuredOutputSchemas:
+    def test_troubleshooting_schema_is_well_formed(self):
+        _assert_well_formed(TROUBLESHOOTING_SCHEMA)
+
+    def test_procedural_schema_is_well_formed(self):
+        _assert_well_formed(PROCEDURAL_SCHEMA)
+
+    def test_troubleshooting_schema_top_level_shape(self):
+        props = TROUBLESHOOTING_SCHEMA["properties"]
+        assert set(props) >= {"title", "description", "nodes"}
+        node = props["nodes"]["items"]
+        # Every field the troubleshooting prompt may emit must be modelled,
+        # else additionalProperties: false makes them impossible to produce.
+        assert set(node["properties"]) >= {
+            "id",
+            "type",
+            "question",
+            "options",
+            "next_node_id",
+            "confidence",
+            "source_excerpt",
+        }
+
+    def test_procedural_schema_top_level_shape(self):
+        props = PROCEDURAL_SCHEMA["properties"]
+        assert set(props) >= {"title", "description", "steps", "intake_form"}
+        step = props["steps"]["items"]
+        assert set(step["properties"]) >= {
+            "id",
+            "type",
+            "content",
+            "confidence",
+            "source_excerpt",
+        }
+        intake = props["intake_form"]["items"]
+        assert set(intake["properties"]) >= {
+            "variable_name",
+            "label",
+            "field_type",
+            "required",
+            "display_order",
+        }
+
+
+class TestSchemaSelector:
+    def test_returns_troubleshooting_schema(self):
+        assert _schema_for_target_type("troubleshooting") is TROUBLESHOOTING_SCHEMA
+
+    def test_returns_procedural_schema_for_procedural(self):
+        assert _schema_for_target_type("procedural") is PROCEDURAL_SCHEMA
+
+    def test_defaults_to_procedural_for_unknown(self):
+        # convert_document treats any non-"troubleshooting" target as procedural.
+        assert _schema_for_target_type("something-else") is PROCEDURAL_SCHEMA