unique-toolkit 2026.40.0.dev2__tar.gz → 2026.40.0.dev4__tar.gz

{unique_toolkit-2026.40.0.dev2 → unique_toolkit-2026.40.0.dev4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-toolkit
-Version: 2026.40.0.dev2
+Version: 2026.40.0.dev4
 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>=2026.40.0.dev0,<2026.40.0rc0
+Requires-Dist: unique-sdk>=2026.40.0.dev3,<2026.40.0rc0
 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-2026.40.0.dev2 → unique_toolkit-2026.40.0.dev4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "unique_toolkit"
-version = "2026.40.0.dev2"
+version = "2026.40.0.dev4"
 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>=2026.40.0.dev0,<2026.40.0rc0",
+    "unique-sdk>=2026.40.0.dev3,<2026.40.0rc0",
     "jambo>=0.1.2,<0.2",
     "jinja2>=3.1.6,<4",
     "docxtpl>=0.20.1,<0.21",

{unique_toolkit-2026.40.0.dev2 → unique_toolkit-2026.40.0.dev4}/unique_toolkit/agentic/feature_flags/feature_flags.py

@@ -23,9 +23,6 @@ class FeatureFlagNames(StrEnum):
         "FEATURE_FLAG_ENABLE_MCP_METADATA_FALLBACK_UN_19145"
     )
     enable_html_rendering_un_15131 = "FEATURE_FLAG_ENABLE_HTML_RENDERING_UN_15131"
-    enable_code_execution_fence_un_17972 = (
-        "FEATURE_FLAG_ENABLE_CODE_EXECUTION_FENCE_UN_17972"
-    )
     enable_html_with_fence_un_17927 = "FEATURE_FLAG_ENABLE_HTML_WITH_FENCE_UN_17927"
     enable_web_search_argument_screening_un_18741 = (
         "FEATURE_FLAG_ENABLE_WEB_SEARCH_ARGUMENT_SCREENING_UN_18741"
@@ -110,12 +107,18 @@ class FeatureFlags(BaseSettings):
 
     enable_code_execution_fence_un_17972: FeatureFlag = Field(
         default=FeatureFlag(False),
-        description="Emit codeExecution fences in message.text for code interpreter outputs (UN-17972). When disabled, inline file refs are kept as-is. Can be 'true' or comma-separated company IDs.",
+        description=(
+            "Retired: replaced by the `enable_code_execution_fence` config field "
+            "on ShowExecutedCodePostprocessorConfig (default true). "
+            "No code reads this field anymore. Kept only so removing it doesn't "
+            "trip the config-checker's breaking-change gate on this settings "
+            "model; delete once that checker allows field removal."
+        ),
     )
 
     enable_html_with_fence_un_17927: FeatureFlag = Field(
         default=FeatureFlag(False),
-        description="Render HTML code interpreter files as htmlWithSource fences instead of HtmlRendering blocks (UN-17927). Requires enable_code_execution_fence_un_17972 to also be enabled. Can be 'true' or comma-separated company IDs.",
+        description="Render HTML code interpreter files as htmlWithSource fences instead of HtmlRendering blocks (UN-17927). Requires the `enable_code_execution_fence` config field (ShowExecutedCodePostprocessorConfig) to also be enabled. Can be 'true' or comma-separated company IDs.",
     )
 
     enable_web_search_argument_screening_un_18741: FeatureFlag = Field(

{unique_toolkit-2026.40.0.dev2 → unique_toolkit-2026.40.0.dev4}/unique_toolkit/agentic/tools/openai_builtin/code_interpreter/postprocessors/code_display.py

@@ -6,15 +6,10 @@ from typing import override
 from pydantic import BaseModel, Field
 from pydantic.json_schema import SkipJsonSchema
 
-from unique_toolkit.agentic.feature_flags import FeatureFlagNames
 from unique_toolkit.agentic.postprocessor.postprocessor_manager import (
     ResponsesApiPostprocessor,
 )
 from unique_toolkit.agentic.tools.config import get_configuration_dict
-from unique_toolkit.experimental.resources.feature_flags import (
-    COMPANY_ID_PLACEHOLDER,
-    is_flag_enabled,
-)
 from unique_toolkit.language_model.schemas import ResponsesLanguageModelStreamResponse
 
 _TEMPLATE = """
@@ -39,6 +34,13 @@ class ShowExecutedCodePostprocessorConfig(BaseModel):
         default=True,
         description="Show the executed source code to the user",
     )
+    enable_code_execution_fence: bool = Field(
+        default=True,
+        description=(
+            "Show generated files as interactive cards with a built-in code view "
+            "and hide the legacy collapsible code block below."
+        ),
+    )
     remove_from_history: SkipJsonSchema[bool] = (
         Field(  # At the moment, it's not possible to keep executed code in the history
             default=True,
@@ -55,20 +57,15 @@ class ShowExecutedCodePostprocessor(ResponsesApiPostprocessor):
     def __init__(
         self,
         config: ShowExecutedCodePostprocessorConfig,
-        company_id: str | None = None,
     ):
         super().__init__(self.__class__.__name__)
         self._config = config
-        self._company_id = company_id
-        # Resolved in run() (before apply_postprocessing_to_response) since flag evaluation is async.
-        self._is_enabled = False
+        # When fences are enabled, the fence itself shows the executed code,
+        # so this legacy <details> display must stay off.
+        self._is_enabled = config.enable and not config.enable_code_execution_fence
 
     @override
     async def run(self, loop_response: ResponsesLanguageModelStreamResponse) -> None:
-        self._is_enabled = self._config.enable and not await is_flag_enabled(
-            FeatureFlagNames.enable_code_execution_fence_un_17972,
-            company_id=self._company_id or COMPANY_ID_PLACEHOLDER,
-        )
         if self._is_enabled:
             await asyncio.sleep(self._config.sleep_time_before_display)
 

{unique_toolkit-2026.40.0.dev2 → unique_toolkit-2026.40.0.dev4}/unique_toolkit/agentic/tools/openai_builtin/code_interpreter/postprocessors/generated_files.py

@@ -376,6 +376,7 @@ class DisplayCodeInterpreterFilesPostProcessor(
         company_id: str | None = None,
         user_id: str | None = None,
         chat_id: str | None = None,
+        code_execution_fence_enabled: bool = True,
     ) -> None:
         super().__init__(self.__class__.__name__)
 
@@ -404,8 +405,11 @@ class DisplayCodeInterpreterFilesPostProcessor(
         self._file_size_map: dict[str, int] = {}
         self._container_files: list[CodeInterpreterContainerFile] = []
 
+        # Fence rendering is a plain config switch now (was FF UN-17972). The
+        # value lives on ShowExecutedCodePostprocessorConfig ("Code display" on
+        # the config page) and is passed in by the orchestrator.
+        self._fence_enabled = code_execution_fence_enabled
         # Resolved in run() (before apply_postprocessing_to_response) since flag evaluation is async.
-        self._fence_ff_on = False
         self._html_fence_ff_on = False
 
     def _build_retry(self) -> AsyncRetrying:
@@ -511,12 +515,8 @@ class DisplayCodeInterpreterFilesPostProcessor(
         run_t0 = time.monotonic()
         self._log.info("run() started — fetching and uploading code interpreter files")
 
-        self._fence_ff_on = await is_flag_enabled(
-            FeatureFlagNames.enable_code_execution_fence_un_17972,
-            company_id=self._company_id or COMPANY_ID_PLACEHOLDER,
-        )
-        # htmlWithSource requires BOTH the fence FF and the HTML-fence FF on; default
-        # (FF off) keeps HtmlRendering, so existing deployments are unaffected.
+        # htmlWithSource requires BOTH the fence config and the HTML-fence FF on;
+        # default (FF off) keeps HtmlRendering, so existing deployments are unaffected.
         self._html_fence_ff_on = await is_flag_enabled(
             FeatureFlagNames.enable_html_with_fence_un_17927,
             company_id=self._company_id or COMPANY_ID_PLACEHOLDER,
@@ -668,9 +668,9 @@ class DisplayCodeInterpreterFilesPostProcessor(
     ) -> bool:
         apply_t0 = time.monotonic()
         self._log.info(
-            "apply_postprocessing started — %d file(s) in content_map, fence_ff=%s",
+            "apply_postprocessing started — %d file(s) in content_map, fence_enabled=%s",
             len(self._content_map),
-            self._fence_ff_on,
+            self._fence_enabled,
         )
 
         if loop_response.message.references is None:
@@ -711,9 +711,9 @@ class DisplayCodeInterpreterFilesPostProcessor(
                 )
                 changed |= replaced
 
-            # HTML rendered as HtmlRendering block unless the dedicated html-fence FF
-            # is on (enable_html_with_fence_un_17927). Default keeps HtmlRendering.
-            elif is_html and not self._html_fence_ff_on:
+            # HTML uses HtmlRendering unless BOTH the fence config and the
+            # html-fence FF are on (htmlWithSource needs both).
+            elif is_html and not (self._html_fence_ff_on and self._fence_enabled):
                 loop_response.message.text, replaced = _replace_container_html_citation(
                     text=loop_response.message.text or "",
                     filename=filename,
@@ -721,15 +721,15 @@ class DisplayCodeInterpreterFilesPostProcessor(
                 )
                 changed |= replaced
 
-            # Non-HTML files, or HTML when self._html_fence_ff_on → inline content link for
-            # subsequent fence injection (htmlWithSource / imgWithSource / fileWithSource)
+            # Non-HTML files, or HTML when both fence switches are on → inline
+            # content link for subsequent fence injection
             else:
                 loop_response.message.text, replaced = _replace_container_file_citation(
                     text=loop_response.message.text or "",
                     filename=filename,
                     content_id=content_id,
                     ref_number=ref_number,
-                    use_content_link=self._fence_ff_on,
+                    use_content_link=self._fence_enabled,
                 )
                 changed |= replaced
 
@@ -746,7 +746,7 @@ class DisplayCodeInterpreterFilesPostProcessor(
 
             # HtmlRendering and htmlWithSource both embed contentId directly — no ContentReference needed
             is_html_rendered = is_html
-            if replaced and not (is_image or is_html_rendered or self._fence_ff_on):
+            if replaced and not (is_image or is_html_rendered or self._fence_enabled):
                 loop_response.message.references.append(
                     ContentReference(
                         sequence_number=ref_number,
@@ -765,7 +765,7 @@ class DisplayCodeInterpreterFilesPostProcessor(
             error_files,
         )
 
-        if self._fence_ff_on:
+        if self._fence_enabled:
             code_blocks = _build_code_blocks(
                 loop_response,
                 self._content_map,
@@ -1503,7 +1503,7 @@ def _warn_unmatched_code_blocks(
 ) -> None:
     """Warn for files that were uploaded but could not be matched to any code block.
 
-    When the fence feature flag is on, every uploaded file should map to a code block
+    When fence rendering is enabled, every uploaded file should map to a code block
     via its /mnt/data/<filename> path so it can receive a fence.  If a file is not
     matched (e.g. the LLM used a variable for the output path rather than a literal
     string) it falls back to a plain unique://content/ link with no code context.
@@ -1634,10 +1634,10 @@ def _replace_container_file_citation(
 ) -> tuple[str, bool]:
     """Replace a sandbox file link with either an inline content link or a superscript ref.
 
-    When the fence feature flag is on (use_content_link=True), the sandbox link is
-    replaced with [filename](unique://content/{id}) so the subsequent fence injection
-    step can locate and wrap it. When the flag is off (use_content_link=False), the
-    original pre-fence behaviour is restored: the link is replaced with <sup>N</sup>
+    When fence rendering is enabled in the config (use_content_link=True), the sandbox
+    link is replaced with [filename](unique://content/{id}) so the subsequent fence
+    injection step can locate and wrap it. When it is disabled (use_content_link=False),
+    the original pre-fence behaviour is restored: the link is replaced with <sup>N</sup>
     and the file remains accessible via the references panel.
     """
     file_markdown = _sandbox_link_pattern(filename)

{unique_toolkit-2026.40.0.dev2 → unique_toolkit-2026.40.0.dev4}/unique_toolkit/agentic/tools/openai_builtin/code_interpreter/tool/config.py

@@ -96,9 +96,9 @@ DEFAULT_TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT = (
     _TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT_BASE + _TOOL_DESCRIPTION_FILES_SUFFIX
 ).strip()
 
-# Used when the code-execution fence feature flag (UN-17972) is enabled.
-# The frontend derives the artifact title from the filename itself, so the
-# LLM no longer needs to produce a markdown heading before the sandbox link.
+# Default system prompt. The frontend derives the artifact title from the
+# filename itself, so the LLM does not need to produce a markdown heading
+# before the sandbox link.
 #
 # UN-19711 / Horizon follow-up: The HTML rules in the fence prompt reduce iframe height
 # collapse when chat auto-measures the embedded document. Remaining gaps belong in the

{unique_toolkit-2026.40.0.dev2 → unique_toolkit-2026.40.0.dev4}/unique_toolkit/chat/functions.py

@@ -188,6 +188,46 @@ async def modify_message_async(
         raise e
 
 
+async def mark_turn_system_interrupted_async(
+    user_id: str,
+    company_id: str,
+    chat_id: str,
+    user_message_id: str,
+) -> ChatMessage:
+    """Marks the turn anchored on *user_message_id* as system-interrupted.
+
+    Persists ``turnInterruptionReason = SYSTEM_INTERRUPTED`` on the originating
+    USER message via a dedicated service-authenticated node-chat endpoint. The
+    server guarantees idempotency and never overwrites ``userAbortedAt`` (a
+    user-initiated stop wins if both race).
+
+    Args:
+        user_id (str): The user ID.
+        company_id (str): The company ID.
+        chat_id (str): The chat ID.
+        user_message_id (str): The originating USER message of the turn.
+
+    Returns:
+        ChatMessage: The USER message after the update (unchanged if the call
+            was a no-op due to idempotency or a racing user stop).
+
+    Raises:
+        Exception: If the request fails.
+
+    """
+    try:
+        message = await unique_sdk.Message.mark_turn_system_interrupted_async(
+            user_id=user_id,
+            company_id=company_id,
+            id=user_message_id,
+            chatId=chat_id,
+        )
+        return ChatMessage(**message)
+    except Exception as e:
+        _LOGGER.error(f"Failed to mark turn as system-interrupted: {e}")
+        raise e
+
+
 def map_references(references: list[ContentReference]) -> list[dict[str, Any]]:
     return [
         {

{unique_toolkit-2026.40.0.dev2 → unique_toolkit-2026.40.0.dev4}/unique_toolkit/chat/schemas.py

@@ -192,6 +192,12 @@ class MessageLogStatus(StrEnum):
     RUNNING = "RUNNING"
     COMPLETED = "COMPLETED"
     FAILED = "FAILED"
+    CANCELLED = "CANCELLED"
+    """The user pressed Stop while this thinking/tool step was in flight.
+
+    Use ``FAILED`` for system-side interruption (timeout, truncation, terminal
+    backend failure), not for user-initiated cancellation.
+    """
 
 
 class MessageExecutionStatus(StrEnum):

{unique_toolkit-2026.40.0.dev2 → unique_toolkit-2026.40.0.dev4}/unique_toolkit/services/chat_service.py

@@ -49,6 +49,7 @@ from unique_toolkit.chat.functions import (
     get_message_tools,
     get_message_tools_async,
     get_selection_from_history,
+    mark_turn_system_interrupted_async,
     modify_message,
     modify_message_assessment,
     modify_message_assessment_async,
@@ -383,6 +384,36 @@ class ChatService(ChatServiceDeprecated):
             set_completed_at=set_completed_at or False,
         )
 
+    async def mark_turn_system_interrupted_async(self) -> ChatMessage:
+        """Marks the current turn as system-interrupted.
+
+        Persists ``turnInterruptionReason = SYSTEM_INTERRUPTED`` on the current
+        chat context's originating USER message via a dedicated
+        service-authenticated node-chat endpoint.
+
+        Call this only for system-side unexpected endings: timeout, truncation,
+        terminal backend failure, and equivalents. Do not call it for
+        user-initiated cancellation or normal completion.
+
+        The operation is idempotent and never overwrites ``userAbortedAt``: if
+        the user pressed Stop and the system interruption race, the user stop
+        wins and this call is a server-side no-op.
+
+        Returns:
+            ChatMessage: The USER message anchoring the turn.
+
+        Raises:
+            Exception: If the request fails. Callers invoking this from
+                failure-handling paths should treat it as best-effort.
+
+        """
+        return await mark_turn_system_interrupted_async(
+            user_id=self._user_id,
+            company_id=self._company_id,
+            chat_id=self._chat_id,
+            user_message_id=self._user_message_id,
+        )
+
     def modify_assistant_message(
         self,
         content: str | None = None,