unique_toolkit 2026.28.0__tar.gz → 2026.28.0.dev1__tar.gz

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-toolkit
-Version: 2026.28.0
+Version: 2026.28.0.dev1
 Summary: 
 Author: Cedric Klinkert, Martin Fadler, Sadique Sheik, Fabian Schläpfer, Pascal Hauri, Ejiroghene Oki-Peter, Rami Azouz, Ricardo Visini, Sanzio Monti
 Author-email: Cedric Klinkert <cedric.klinkert@unique.ch>, Martin Fadler <martin.fadler@unique.ch>, Sadique Sheik <sadique@unique.ch>, Fabian Schläpfer <fabian@unique.ch>, Pascal Hauri <pascal@unique.ch>, Ejiroghene Oki-Peter <oki-peter.ejiroghene@unique.ai>, Rami Azouz <rami.ext@unique.ch>, Ricardo Visini <ricardo.visini@unique.ch>, Sanzio Monti <sanzio.monti@unique.ai>
@@ -16,7 +16,7 @@ Requires-Dist: sseclient>=0.0.27,<0.0.28
 Requires-Dist: openai>=1.105.0,<3
 Requires-Dist: platformdirs>=4.0.0,<5
 Requires-Dist: pillow>=12.1.1
-Requires-Dist: unique-sdk>=2026.28.0
+Requires-Dist: unique-sdk>=2026.28.0.dev2,<2026.28.0rc0
 Requires-Dist: jambo>=0.1.2,<0.2
 Requires-Dist: jinja2>=3.1.6,<4
 Requires-Dist: docxtpl>=0.20.1,<0.21

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "unique_toolkit"
-version = "2026.28.0"
+version = "2026.28.0.dev1"
 description = ""
 readme = "README.md"
 requires-python = ">=3.12"
@@ -28,7 +28,7 @@ dependencies = [
     "openai>=1.105.0,<3",
     "platformdirs>=4.0.0,<5",
     "pillow>=12.1.1",
-    "unique-sdk>=2026.28.0",
+    "unique-sdk>=2026.28.0.dev2,<2026.28.0rc0",
     "jambo>=0.1.2,<0.2",
     "jinja2>=3.1.6,<4",
     "docxtpl>=0.20.1,<0.21",

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/_common/chunk_relevancy_sorter/service.py

@@ -71,7 +71,7 @@ class ChunkRelevancySorter:
         company_id: str | None = None,
         user_id: str | None = None,
     ):
-        if isinstance(event, BaseEvent):
+        if isinstance(event, (ChatEvent, BaseEvent)):
             self.chunk_relevancy_evaluator = ContextRelevancyEvaluator.from_event(event)
         else:
             [company_id, user_id] = validate_required_values([company_id, user_id])

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/_common/pydantic/rjsf_tags.py

@@ -928,7 +928,6 @@ def ui_schema_for_model(
     ui: dict[str, Any] = {}
 
     for fname in model_cls.model_fields:
-        field_info = model_cls.model_fields[fname]
         ann = hints[fname]
         node: dict[str, Any] = {}
 
@@ -1020,21 +1019,13 @@ def ui_schema_for_model(
                 _maybe_set_string_empty_value(branch, alt_b, default_string_empty_value)
                 branches.append(branch)
             if branches:
-                # RJSF v6 selects the per-branch uiSchema using the SAME keyword
-                # the JSON schema uses for the union: ``schema.oneOf`` reads
-                # ``uiSchema.oneOf`` and ``schema.anyOf`` reads ``uiSchema.anyOf``
-                # (see @rjsf/core ``AnyOfField``). Pydantic emits ``oneOf`` for a
-                # *discriminated* union and ``anyOf`` otherwise, so the uiSchema
-                # branch key MUST match the schema keyword — a ``oneOf`` schema
-                # paired with an ``anyOf`` uiSchema makes RJSF silently drop every
-                # per-branch widget and fall back to default widgets.
-                branch_key = (
-                    "oneOf" if field_info.discriminator is not None else "anyOf"
-                )
-                # Keep any branch structure already supplied by a composer
-                # (Optional/Union write their own ``anyOf`` into ``meta``).
-                if "anyOf" not in node and "oneOf" not in node:
-                    node[branch_key] = branches
+                # Check if the metadata already has an anyOf structure (from Union composer)
+                if "anyOf" in node:
+                    # Use the anyOf from the metadata (from Union composer)
+                    pass  # Keep the existing anyOf from the Union composer
+                else:
+                    # Create anyOf structure for Union types
+                    node["anyOf"] = branches
 
         # Scalars: node already has metadata if any
         _maybe_set_string_empty_value(node, base, default_string_empty_value)

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/agentic/evaluation/context_relevancy/service.py

@@ -81,7 +81,7 @@ class ContextRelevancyEvaluator:
         company_id: str | None = None,
         user_id: str | None = None,
     ):
-        if isinstance(event, BaseEvent):
+        if isinstance(event, (ChatEvent, BaseEvent)):
             self.language_model_service = LanguageModelService.from_event(event)
         else:
             [company_id, user_id] = validate_required_values([company_id, user_id])

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/agentic/evaluation/hallucination/constants.py

@@ -52,18 +52,6 @@ class HallucinationConfig(EvaluationMetricConfig):
         default={},
         description="Additional options to pass to the language model.",
     )
-    max_external_context_chars_per_source: SkipJsonSchema[int] = Field(
-        default=20_000,
-        ge=0,
-        description=(
-            "Per-source character cap applied to external (non-chunk) "
-            "context texts, e.g. MCP tool output. Raw/large tool results "
-            "inflate the judge prompt and the eval-LLM call; each external "
-            "text is truncated to this many characters before being added "
-            "to the context. 0 disables truncation. Does not affect "
-            "chunk-derived context."
-        ),
-    )
     score_to_label: dict[str, str] = {
         "LOW": "GREEN",
         "MEDIUM": "YELLOW",

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/agentic/evaluation/hallucination/hallucination_evaluation.py

@@ -54,17 +54,6 @@ class HallucinationEvaluation(Evaluation):
             reference_pattern=self.config.reference_pattern,
         )
 
-        # Append non-chunk groundedness context (e.g. MCP tool output). These
-        # are added in full ("all retrieved output as context") rather than via
-        # source-marker selection, since they carry no [sourceN] markers.
-        external_context_texts = self._reference_manager.get_external_context_texts()
-        max_chars = self.config.max_external_context_chars_per_source
-        if max_chars > 0:
-            external_context_texts = [
-                text[:max_chars] for text in external_context_texts
-            ]
-        context_texts = context_texts + external_context_texts
-
         evaluation_result: EvaluationMetricResult = await check_hallucination(
             company_id=self._company_id,
             user_id=self._user_id,

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/agentic/evaluation/hallucination/utils.py

@@ -16,8 +16,6 @@ from unique_toolkit.agentic.evaluation.schemas import (
 )
 from unique_toolkit.content import ContentReference
 from unique_toolkit.content.schemas import ContentChunk
-from unique_toolkit.language_model.builder import MessagesBuilder
-from unique_toolkit.language_model.infos import ModelCapabilities
 from unique_toolkit.language_model.schemas import (
     LanguageModelMessages,
     LanguageModelStreamResponse,
@@ -106,28 +104,19 @@ def _get_msgs(
     """
     Composes the messages for hallucination analysis based on the provided input and configuration.
 
-    This method composes messages with or without context based on the availability of context texts,
-    history message texts, and (for vision-capable models) context images in the input.
+    This method composes messages with or without context based on the availability of context texts
+    and history message texts in the input.
 
     Args:
-        input (EvaluationMetricInput): The input data that includes context texts, history message texts,
-                                      and context images for the analysis.
+        input (EvaluationMetricInput): The input data that includes context texts and history message texts
+                                      for the analysis.
         config (EvaluationMetricConfig): The configuration settings for composing messages.
         logger (Optional[logging.Logger], optional): The logger used for logging debug information.
 
     Returns:
         The composed messages as per the provided input and configuration.
     """
-    # Context images only count as grounding when the model can actually consume
-    # them; otherwise _compose_msgs drops them and the message would be text-only.
-    model_supports_vision = (
-        ModelCapabilities.VISION in config.language_model.capabilities
-    )
-    has_context = bool(
-        input.context_texts
-        or input.history_messages
-        or (input.context_images and model_supports_vision)
-    )
+    has_context = bool(input.context_texts or input.history_messages)
 
     if has_context:
         _LOGGER.debug("Using context / history for hallucination evaluation.")
@@ -176,36 +165,6 @@ def _compose_msgs(
         else None,
         output_text=input.output_text,
     )
-
-    # When images are provided AND the configured model is vision-capable, build
-    # a multimodal user message (rendered user text + one image part per data
-    # URL) so the model can ground the check against the rendered pages. If the
-    # model lacks VISION we drop the images and fall back to the text-only
-    # message: attaching images to a non-vision model would make the LLM call
-    # fail. Without images the message stays the exact text-only form, so
-    # existing callers are unaffected.
-    context_images = input.context_images or []
-    model_supports_vision = (
-        ModelCapabilities.VISION in config.language_model.capabilities
-    )
-    # When has_context is True for an images-only input, the caller already
-    # confirmed the model is vision-capable, so this branch stays unreachable in
-    # that case and only logs for a genuine non-vision misconfiguration.
-    if context_images and not model_supports_vision:
-        _LOGGER.warning(
-            "Hallucination model %s lacks VISION capability; ignoring "
-            "%d context image(s) and grounding on text only.",
-            config.language_model.name,
-            len(context_images),
-        )
-    if context_images and model_supports_vision:
-        return (
-            MessagesBuilder()
-            .append(system_msg)
-            .image_message_append(content=user_msg_content, images=context_images)
-            .build()
-        )
-
     user_msg = LanguageModelUserMessage(content=user_msg_content)
 
     return LanguageModelMessages([system_msg, user_msg])

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/agentic/evaluation/schemas.py

@@ -34,16 +34,6 @@ class EvaluationMetricInput(BaseModel):
     context_texts: Optional[list[str]] = None
     history_messages: Optional[list[ChatMessage]] = None
     output_text: Optional[str] = None
-    context_images: Optional[list[str]] = Field(
-        default=None,
-        description=(
-            "Optional base64 data URLs (e.g. 'data:image/png;base64,...') of "
-            "context the metric should evaluate against visually, in addition "
-            "to context_texts. Used by metrics that support a multimodal "
-            "(vision) check; ignored by metrics that do not. Defaults to None "
-            "so existing text-only callers are unaffected."
-        ),
-    )
 
     def get_joined_context_texts(self, tag_name: str = "reference") -> str:
         """

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/agentic/evaluation/tests/test_hallucination_utils.py

@@ -20,10 +20,6 @@ from unique_toolkit.agentic.evaluation.schemas import EvaluationMetricInput
 from unique_toolkit.chat.schemas import ChatMessage
 from unique_toolkit.content import ContentReference
 from unique_toolkit.content.schemas import ContentChunk
-from unique_toolkit.language_model.infos import (
-    LanguageModelName,
-    ModelCapabilities,
-)
 from unique_toolkit.language_model.schemas import (
     LanguageModelMessageRole,
     LanguageModelMessages,
@@ -1017,162 +1013,3 @@ def test_compose_msgs__generates_non_empty_messages__from_config_prompts(
     # Assert
     assert len(result.root[0].content) > 0
     assert len(result.root[1].content) > 0
-
-
-_IMAGE_DATA_URL = "data:image/png;base64,QQ=="
-
-
-@pytest.mark.ai
-def test_compose_msgs__builds_multimodal_user_message__with_context_images(
-    hallucination_config: HallucinationConfig,
-) -> None:
-    """
-    Purpose: Verify that context_images produce a multimodal user message.
-    Why this matters: Vision faithfulness needs the rendered pages as image parts.
-    Setup summary: Provide context_images, assert the user message carries text +
-    one image part per data URL.
-    """
-    # Arrange
-    input_data: EvaluationMetricInput = EvaluationMetricInput(
-        input_text="Question",
-        output_text="Output",
-        context_images=[_IMAGE_DATA_URL],
-    )
-
-    # Act
-    result: LanguageModelMessages = _compose_msgs(
-        input_data, hallucination_config, has_context=True
-    )
-
-    # Assert
-    assert len(result.root) == 2
-    user_content = result.root[1].content
-    assert isinstance(user_content, list)
-    text_parts = [p for p in user_content if p.get("type") == "text"]
-    image_parts = [p for p in user_content if p.get("type") == "image_url"]
-    assert len(text_parts) == 1
-    assert len(image_parts) == 1
-    assert image_parts[0]["imageUrl"]["url"] == _IMAGE_DATA_URL
-
-
-@pytest.mark.ai
-def test_compose_msgs__stays_text_only__without_context_images(
-    hallucination_config: HallucinationConfig,
-) -> None:
-    """
-    Purpose: Verify that omitting context_images keeps the exact text-only message.
-    Why this matters: Backward compatibility for all existing (non-vision) callers.
-    Setup summary: No context_images -> user message content is a plain string.
-    """
-    # Arrange
-    input_data: EvaluationMetricInput = EvaluationMetricInput(
-        input_text="Question",
-        context_texts=["Context 1"],
-        output_text="Output",
-    )
-
-    # Act
-    result: LanguageModelMessages = _compose_msgs(
-        input_data, hallucination_config, has_context=True
-    )
-
-    # Assert
-    assert isinstance(result.root[1].content, str)
-
-
-@pytest.mark.ai
-def test_compose_msgs__drops_images__when_model_lacks_vision() -> None:
-    """
-    Purpose: Verify images are dropped when the configured model is not vision-capable.
-    Why this matters: Attaching images to a non-vision model makes the LLM call fail;
-    we must degrade gracefully to a text-only message instead.
-    Setup summary: Configure a text-only model (no VISION capability) with
-    context_images, assert the user message content stays a plain string.
-    """
-    # Arrange
-    text_only_config = HallucinationConfig(
-        enabled=True,
-        language_model=LanguageModelName.AZURE_GPT_4_0613,
-    )
-    assert ModelCapabilities.VISION not in (
-        text_only_config.language_model.capabilities
-    )
-    input_data: EvaluationMetricInput = EvaluationMetricInput(
-        input_text="Question",
-        output_text="Output",
-        context_images=[_IMAGE_DATA_URL],
-    )
-
-    # Act
-    result: LanguageModelMessages = _compose_msgs(
-        input_data, text_only_config, has_context=True
-    )
-
-    # Assert
-    assert isinstance(result.root[1].content, str)
-
-
-@pytest.mark.ai
-def test_get_msgs__treats_context_images_as_context(
-    hallucination_config: HallucinationConfig,
-) -> None:
-    """
-    Purpose: Verify that context_images alone count as context in _get_msgs.
-    Why this matters: A vision-only turn must use the grounded (has_context) prompt
-    and build the multimodal message.
-    Setup summary: Provide only context_images; assert the system prompt matches the
-    has_context variant and the user message is multimodal.
-    """
-    # Arrange
-    image_only: EvaluationMetricInput = EvaluationMetricInput(
-        input_text="Question",
-        output_text="Output",
-        context_images=[_IMAGE_DATA_URL],
-    )
-    grounded: EvaluationMetricInput = EvaluationMetricInput(
-        input_text="Question",
-        context_texts=["Context 1"],
-        output_text="Output",
-    )
-
-    # Act
-    result: LanguageModelMessages = _get_msgs(image_only, hallucination_config)
-    grounded_result: LanguageModelMessages = _get_msgs(grounded, hallucination_config)
-
-    # Assert
-    assert result.root[0].content == grounded_result.root[0].content
-    assert isinstance(result.root[1].content, list)
-
-
-@pytest.mark.ai
-def test_get_msgs__ignores_context_images__when_model_lacks_vision() -> None:
-    """
-    Purpose: Verify image-only input is not treated as context for a non-vision model.
-    Why this matters: _compose_msgs drops the images for a non-vision model, so the
-    grounded prompt would otherwise claim references that are never sent. has_context
-    must reflect the message actually delivered.
-    Setup summary: Configure a text-only model with only context_images; assert the
-    system prompt matches the ungrounded variant and the user message is text-only.
-    """
-    # Arrange
-    text_only_config = HallucinationConfig(
-        enabled=True,
-        language_model=LanguageModelName.AZURE_GPT_4_0613,
-    )
-    image_only: EvaluationMetricInput = EvaluationMetricInput(
-        input_text="Question",
-        output_text="Output",
-        context_images=[_IMAGE_DATA_URL],
-    )
-    no_context: EvaluationMetricInput = EvaluationMetricInput(
-        input_text="Question",
-        output_text="Output",
-    )
-
-    # Act
-    result: LanguageModelMessages = _get_msgs(image_only, text_only_config)
-    no_context_result: LanguageModelMessages = _get_msgs(no_context, text_only_config)
-
-    # Assert
-    assert result.root[0].content == no_context_result.root[0].content
-    assert isinstance(result.root[1].content, str)

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/agentic/reference_manager/reference_manager.py

@@ -32,7 +32,6 @@ class ReferenceManager:
         self._tool_chunks: dict[str, tool_chunks] = {}
         self._chunks: list[ContentChunk] = []
         self._references: list[list[ContentReference]] = []
-        self._external_context_texts: list[str] = []
 
     def extract_referenceable_chunks(
         self, tool_responses: list[ToolCallResponse]
@@ -48,21 +47,6 @@ class ReferenceManager:
     def get_chunks(self) -> list[ContentChunk]:
         return self._chunks
 
-    def add_external_context_texts(self, texts: list[str]) -> None:
-        """Register groundedness context that is not backed by ContentChunks.
-
-        Some retrieval sources (e.g. MCP tool output) provide text the answer
-        is built on but are deliberately *not* modelled as ContentChunks — they
-        are a different entity from ingested document chunks and must not flow
-        through the chunk/reference pipeline. They still need to ground the
-        hallucination check, so they are stored here and appended to the
-        evaluation context alongside the chunk-derived texts.
-        """
-        self._external_context_texts.extend(text for text in texts if text)
-
-    def get_external_context_texts(self) -> list[str]:
-        return self._external_context_texts
-
     def get_tool_chunks(self) -> dict[str, tool_chunks]:
         return self._tool_chunks
 

unique_toolkit-2026.28.0.dev1/unique_toolkit/agentic/tools/a2a/manager.py

@@ -0,0 +1,55 @@
+from logging import Logger
+
+from unique_toolkit.agentic.tools.a2a.response_watcher import SubAgentResponseWatcher
+from unique_toolkit.agentic.tools.a2a.tool import SubAgentTool, SubAgentToolConfig
+from unique_toolkit.agentic.tools.config import ToolBuildConfig
+from unique_toolkit.agentic.tools.tool_progress_reporter import ToolProgressReporter
+from unique_toolkit.app.schemas import ChatEvent
+
+
+class A2AManager:
+    def __init__(
+        self,
+        logger: Logger,
+        tool_progress_reporter: ToolProgressReporter,
+        response_watcher: SubAgentResponseWatcher,
+    ):
+        self._logger = logger
+        self._tool_progress_reporter = tool_progress_reporter
+        self._response_watcher = response_watcher
+
+    def get_all_sub_agents(
+        self,
+        tool_configs: list[ToolBuildConfig],
+        event: ChatEvent,
+    ) -> tuple[list[ToolBuildConfig], list[SubAgentTool]]:
+        sub_agents = []
+
+        for tool_config in tool_configs:
+            if not tool_config.is_sub_agent:
+                continue
+
+            if not isinstance(tool_config.configuration, SubAgentToolConfig):
+                self._logger.error(
+                    "tool_config.configuration must be of type SubAgentToolConfig"
+                )
+                continue
+
+            sub_agent_tool_config = tool_config.configuration
+
+            sub_agents.append(
+                SubAgentTool(
+                    configuration=sub_agent_tool_config,
+                    event=event,
+                    tool_progress_reporter=self._tool_progress_reporter,
+                    name=tool_config.name,
+                    display_name=tool_config.display_name,
+                    response_watcher=self._response_watcher,
+                )
+            )
+
+        filtered_tool_config = [
+            tool_config for tool_config in tool_configs if not tool_config.is_sub_agent
+        ]
+
+        return filtered_tool_config, sub_agents

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/agentic/tools/a2a/tool/service.py

@@ -20,6 +20,7 @@ from unique_toolkit._common.referencing import (
 from unique_toolkit._common.utils.jinja.render import render_template
 from unique_toolkit.agentic.evaluation.schemas import EvaluationMetricName
 from unique_toolkit.agentic.feature_flags import feature_flags
+from unique_toolkit.agentic.message_log_manager.service import MessageStepLogger
 from unique_toolkit.agentic.tools.a2a.response_watcher import SubAgentResponseWatcher
 from unique_toolkit.agentic.tools.a2a.tool._memory import (
     get_sub_agent_short_term_memory_manager,
@@ -52,7 +53,6 @@ from unique_toolkit.chat.service import ChatService
 from unique_toolkit.content import ContentChunk, ContentReference
 from unique_toolkit.language_model import (
     LanguageModelFunction,
-    LanguageModelService,
     LanguageModelToolDescription,
 )
 
@@ -72,20 +72,12 @@ class SubAgentTool(Tool[SubAgentToolConfig]):
         name: str = "SubAgentTool",
         display_name: str = "SubAgentTool",
         response_watcher: SubAgentResponseWatcher | None = None,
-        *,
-        chat_service: ChatService | None = None,
-        language_model_service: LanguageModelService | None = None,
     ) -> None:
-        if chat_service is not None and language_model_service is not None:
-            super().__init__(
-                configuration,
-                event,
-                tool_progress_reporter=tool_progress_reporter,
-                chat_service=chat_service,
-                language_model_service=language_model_service,
-            )
-        else:
-            super().__init__(configuration, event, tool_progress_reporter)
+        super().__init__(configuration)
+        self._event = event
+        self._tool_progress_reporter = tool_progress_reporter
+        self._chat_service = ChatService(event)
+        self._message_step_logger = MessageStepLogger(chat_service=self._chat_service)
         self._user_id = event.user_id
         self._company_id = event.company_id
 

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/agentic/tools/config.py

@@ -1,4 +1,3 @@
-import logging
 from enum import StrEnum
 from typing import Annotated, Any, Generic
 
@@ -15,36 +14,9 @@ from typing_extensions import TypeVar
 from unique_toolkit._common.pydantic_helpers import get_configuration_dict
 from unique_toolkit.agentic.tools.schemas import BaseToolConfig
 
-logger = logging.getLogger(__name__)
-
 T = TypeVar("T", bound=BaseToolConfig, default=BaseToolConfig)
 
 
-def _ensure_base_tool_config(value: dict[str, Any]) -> None:
-    """Make configuration parseable; disabled/demoted tools are never built at runtime."""
-    configuration = value.get("configuration")
-    if not isinstance(configuration, BaseToolConfig):
-        value["configuration"] = BaseToolConfig()
-
-
-def _non_empty_str(value: dict[str, Any], snake_key: str, camel_key: str) -> str:
-    raw = value.get(snake_key, value.get(camel_key, ""))
-    if raw is None:
-        return ""
-    return str(raw)
-
-
-def _is_mcp_tool_payload(value: dict[str, Any]) -> bool:
-    if _non_empty_str(value, "mcp_source_id", "mcpSourceId"):
-        return True
-
-    configuration = value.get("configuration", {})
-    if isinstance(configuration, dict):
-        if _non_empty_str(configuration, "mcp_source_id", "mcpSourceId"):
-            return True
-    return False
-
-
 class ToolIcon(StrEnum):
     ANALYTICS = "IconAnalytics"
     BOOK = "IconBook"
@@ -104,18 +76,11 @@ class ToolBuildConfig(BaseModel, Generic[T]):
         value: Any,
         info: ValidationInfo,
     ) -> Any:
-        """Validate and resolve tool configuration based on tool type and name.
-
-        Both enabled and disabled tools resolve to their concrete config type when
-        the stored configuration is valid, so downstream consumers that key off the
-        tool name keep seeing the expected config type regardless of the enabled
-        flag. A tool whose configuration is invalid is demoted to disabled with a
-        BaseToolConfig fallback instead of raising.
-        """
+        """Check the given values for."""
         if not isinstance(value, dict):
             return value
 
-        is_mcp_tool = _is_mcp_tool_payload(value)
+        is_mcp_tool = value.get("mcp_source_id", "") != ""
         mcp_configuration = value.get("configuration", {})
 
         # Import at runtime to avoid circular imports
@@ -137,39 +102,26 @@ class ToolBuildConfig(BaseModel, Generic[T]):
 
         configuration = value.get("configuration", {})
 
-        try:
-            if is_sub_agent_tool:
-                from unique_toolkit.agentic.tools.a2a import ExtendedSubAgentToolConfig
-
-                config = ExtendedSubAgentToolConfig.model_validate(configuration)
-            elif isinstance(configuration, dict):
-                # Local import to avoid circular import at module import time
-                from unique_toolkit.agentic.tools.factory import ToolFactory
-
-                config = ToolFactory.build_tool_config(
-                    value["name"],
-                    **configuration,
-                )
-            else:
-                # Check that the type of config matches the tool name
-                from unique_toolkit.agentic.tools.factory import ToolFactory
-
-                assert isinstance(
-                    configuration,
-                    ToolFactory.tool_config_map[value["name"]],  # pyright: ignore[reportArgumentType]
-                )
-                config = configuration
-        except Exception:
-            tool_name = value.get("name", "<unknown>")
-            logger.warning(
-                "Tool '%s' has invalid configuration and will be disabled.",
-                tool_name,
-                exc_info=True,
+        if is_sub_agent_tool:
+            from unique_toolkit.agentic.tools.a2a import ExtendedSubAgentToolConfig
+
+            config = ExtendedSubAgentToolConfig.model_validate(configuration)
+        elif isinstance(configuration, dict):
+            # Local import to avoid circular import at module import time
+            from unique_toolkit.agentic.tools.factory import ToolFactory
+
+            config = ToolFactory.build_tool_config(
+                value["name"],
+                **configuration,
             )
-            value["is_enabled"] = False
-            value["isEnabled"] = False
-            _ensure_base_tool_config(value)
-            return value
+        else:
+            # Check that the type of config matches the tool name
+            from unique_toolkit.agentic.tools.factory import ToolFactory
 
+            assert isinstance(
+                configuration,
+                ToolFactory.tool_config_map[value["name"]],  # pyright: ignore[reportArgumentType]
+            )
+            config = configuration
         value["configuration"] = config
         return value

{unique_toolkit-2026.28.0 → unique_toolkit-2026.28.0.dev1}/unique_toolkit/agentic/tools/experimental/__init__.py

@@ -1,7 +1,3 @@
-from unique_toolkit.agentic.tools.experimental.ask_user_tool import (
-    AskUserTool,
-    AskUserToolConfig,
-)
 from unique_toolkit.agentic.tools.experimental.open_file_tool import (
     OpenFileTool,
     OpenFileToolRuntime,
@@ -23,8 +19,6 @@ from unique_toolkit.agentic.tools.experimental.todo import (
 )
 
 __all__ = [
-    "AskUserTool",
-    "AskUserToolConfig",
     "DisplayMode",
     "OpenFileTool",
     "OpenFileToolRuntime",