unique-toolkit 2026.38.0.dev16__tar.gz → 2026.38.0.dev17__tar.gz

{unique_toolkit-2026.38.0.dev16 → unique_toolkit-2026.38.0.dev17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-toolkit
-Version: 2026.38.0.dev16
+Version: 2026.38.0.dev17
 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>

{unique_toolkit-2026.38.0.dev16 → unique_toolkit-2026.38.0.dev17}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "unique_toolkit"
-version = "2026.38.0.dev16"
+version = "2026.38.0.dev17"
 description = ""
 readme = "README.md"
 requires-python = ">=3.12"

{unique_toolkit-2026.38.0.dev16 → unique_toolkit-2026.38.0.dev17}/unique_toolkit/agentic/feature_flags/feature_flags.py

@@ -23,6 +23,9 @@ 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"
@@ -107,17 +110,12 @@ class FeatureFlags(BaseSettings):
 
     enable_code_execution_fence_un_17972: FeatureFlag = Field(
         default=FeatureFlag(False),
-        description=(
-            "Retired (UN-25450): code interpreter output is always rendered as "
-            "codeExecution fences now, no code reads this field. 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."
-        ),
+        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.",
     )
 
     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). Can be 'true' or comma-separated company IDs.",
+        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.",
     )
 
     enable_web_search_argument_screening_un_18741: FeatureFlag = Field(

unique_toolkit-2026.38.0.dev17/unique_toolkit/agentic/tools/openai_builtin/code_interpreter/postprocessors/code_display.py

@@ -0,0 +1,95 @@
+import asyncio
+import logging
+import re
+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 = """
+<details><summary>Code Interpreter Call</summary>    
+
+```python
+{code}
+```  
+
+</details>    
+</br>
+
+""".lstrip()
+
+
+logger = logging.getLogger(__name__)
+
+
+class ShowExecutedCodePostprocessorConfig(BaseModel):
+    model_config = get_configuration_dict()
+    enable: bool = Field(
+        default=True,
+        description="Enable display of executed code before the assistant message",
+    )
+    remove_from_history: SkipJsonSchema[bool] = (
+        Field(  # At the moment, it's not possible to keep executed code in the history
+            default=True,
+            description="If set, the code interpreter call will be removed from the history on subsequent calls to the assistant.",
+        )
+    )
+    sleep_time_before_display: float = Field(
+        default=0.2,
+        description="Time to sleep before displaying the executed code. Please increase this value if you experience rendering issues.",
+    )
+
+
+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
+
+    @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)
+
+    @override
+    def apply_postprocessing_to_response(
+        self, loop_response: ResponsesLanguageModelStreamResponse
+    ) -> bool:
+        if not self._is_enabled:
+            return False
+
+        prepended_text = ""
+        for output in loop_response.code_interpreter_calls:
+            prepended_text += _TEMPLATE.format(code=output.code)
+
+        loop_response.message.text = prepended_text + (loop_response.message.text or "")
+
+        return prepended_text != ""
+
+    @override
+    async def remove_from_text(self, text) -> str:
+        if not self._config.remove_from_history:
+            return text
+        pattern = r"<details><summary>Code Interpreter Call</summary>.*?</details>"
+        return re.sub(pattern, "", text, flags=re.DOTALL)

{unique_toolkit-2026.38.0.dev16 → unique_toolkit-2026.38.0.dev17}/unique_toolkit/agentic/tools/openai_builtin/code_interpreter/postprocessors/generated_files.py

@@ -38,6 +38,7 @@ from unique_toolkit.agentic.tools.openai_builtin.code_interpreter.schemas import
     CodeInterpreterFile,
     CodeInterpreterFileType,
 )
+from unique_toolkit.content.schemas import ContentReference
 from unique_toolkit.content.service import ContentService
 from unique_toolkit.experimental.resources.feature_flags import (
     COMPANY_ID_PLACEHOLDER,
@@ -404,6 +405,7 @@ class DisplayCodeInterpreterFilesPostProcessor(
         self._container_files: list[CodeInterpreterContainerFile] = []
 
         # 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:
@@ -509,8 +511,12 @@ class DisplayCodeInterpreterFilesPostProcessor(
         run_t0 = time.monotonic()
         self._log.info("run() started — fetching and uploading code interpreter files")
 
-        # HTML files stay in an HtmlRendering block unless this flag is on; the
-        # default (FF off) keeps HtmlRendering, so existing deployments are unaffected.
+        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.
         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,
@@ -662,8 +668,9 @@ class DisplayCodeInterpreterFilesPostProcessor(
     ) -> bool:
         apply_t0 = time.monotonic()
         self._log.info(
-            "apply_postprocessing started — %d file(s) in content_map",
+            "apply_postprocessing started — %d file(s) in content_map, fence_ff=%s",
             len(self._content_map),
+            self._fence_ff_on,
         )
 
         if loop_response.message.references is None:
@@ -671,6 +678,7 @@ class DisplayCodeInterpreterFilesPostProcessor(
         if loop_response.message.text is None:
             loop_response.message.text = ""
 
+        ref_number = _get_next_ref_number(loop_response.message.references)
         changed = False
 
         replaced_files: list[str] = []
@@ -720,6 +728,8 @@ class DisplayCodeInterpreterFilesPostProcessor(
                     text=loop_response.message.text or "",
                     filename=filename,
                     content_id=content_id,
+                    ref_number=ref_number,
+                    use_content_link=self._fence_ff_on,
                 )
                 changed |= replaced
 
@@ -734,6 +744,20 @@ class DisplayCodeInterpreterFilesPostProcessor(
             else:
                 missed_files.append(filename)
 
+            # 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):
+                loop_response.message.references.append(
+                    ContentReference(
+                        sequence_number=ref_number,
+                        source_id=content_id,
+                        source="node-ingestion-chunks",
+                        url=f"unique://content/{content_id}",
+                        name=filename,
+                    )
+                )
+                ref_number += 1
+
         self._log.info(
             "Stage-1 replacement summary — replaced=%s, missed=%s, error=%s",
             replaced_files,
@@ -741,29 +765,30 @@ class DisplayCodeInterpreterFilesPostProcessor(
             error_files,
         )
 
-        code_blocks = _build_code_blocks(
-            loop_response,
-            self._content_map,
-            self._container_files,
-            include_html=self._html_fence_ff_on,
-        )
-        self._log.info(
-            "Fence injection — %d code block(s), files: %s",
-            len(code_blocks),
-            [f.filename for b in code_blocks for f in b.files],
-        )
-        _warn_unmatched_code_blocks(
-            self._content_map, code_blocks, include_html=self._html_fence_ff_on
-        )
-        text_before = loop_response.message.text
-        loop_response.message.text = _inject_code_execution_fences(
-            loop_response.message.text or "",
-            code_blocks,
-        )
-        fences_changed = loop_response.message.text != text_before
-        changed |= fences_changed
-        if fences_changed:
-            self._log.info("Fence injection modified the message text")
+        if self._fence_ff_on:
+            code_blocks = _build_code_blocks(
+                loop_response,
+                self._content_map,
+                self._container_files,
+                include_html=self._html_fence_ff_on,
+            )
+            self._log.info(
+                "Fence injection — %d code block(s), files: %s",
+                len(code_blocks),
+                [f.filename for b in code_blocks for f in b.files],
+            )
+            _warn_unmatched_code_blocks(
+                self._content_map, code_blocks, include_html=self._html_fence_ff_on
+            )
+            text_before = loop_response.message.text
+            loop_response.message.text = _inject_code_execution_fences(
+                loop_response.message.text or "",
+                code_blocks,
+            )
+            fences_changed = loop_response.message.text != text_before
+            changed |= fences_changed
+            if fences_changed:
+                self._log.info("Fence injection modified the message text")
 
         _warn_missing_content_ids(loop_response.message.text or "", self._content_map)
         ci_ran = bool(loop_response.code_interpreter_calls)
@@ -1502,6 +1527,15 @@ def _warn_unmatched_code_blocks(
             )
 
 
+def _get_next_ref_number(references: list[ContentReference] | None) -> int:
+    if not references:
+        return 1
+    max_ref_number = 0
+    for ref in references:
+        max_ref_number = max(max_ref_number, ref.sequence_number)
+    return max_ref_number + 1
+
+
 def _replace_container_file_error(
     text: str, filename: str, error_message: str
 ) -> tuple[str, bool]:
@@ -1595,11 +1629,16 @@ def _replace_container_file_citation(
     text: str,
     filename: str,
     content_id: str,
+    ref_number: int,
+    use_content_link: bool,
 ) -> tuple[str, bool]:
-    """Replace a sandbox file link with an inline content link.
+    """Replace a sandbox file link with either an inline content link or a superscript ref.
 
-    The sandbox link becomes [filename](unique://content/{id}) so the fence
-    injection step that runs afterwards can find it and wrap it in a fence.
+    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>
+    and the file remains accessible via the references panel.
     """
     file_markdown = _sandbox_link_pattern(filename)
 
@@ -1613,6 +1652,9 @@ def _replace_container_file_citation(
         return text, False
 
     logger.info("Displaying file %s", filename)
-    return re.sub(
-        file_markdown, f"[{filename}](unique://content/{content_id})", text
-    ), True
+    replacement = (
+        f"[{filename}](unique://content/{content_id})"
+        if use_content_link
+        else f"<sup>{ref_number}</sup>"
+    )
+    return re.sub(file_markdown, replacement, text), True

{unique_toolkit-2026.38.0.dev16 → unique_toolkit-2026.38.0.dev17}/unique_toolkit/agentic/tools/openai_builtin/code_interpreter/tool/config.py

@@ -96,7 +96,7 @@ DEFAULT_TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT = (
     _TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT_BASE + _TOOL_DESCRIPTION_FILES_SUFFIX
 ).strip()
 
-# Default prompt since code-execution fences went GA (UN-25450).
+# 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.
 #

unique_toolkit-2026.38.0.dev16/unique_toolkit/agentic/tools/openai_builtin/code_interpreter/postprocessors/code_display.py

@@ -1,59 +0,0 @@
-import logging
-import re
-from typing import override
-
-from pydantic import BaseModel, Field
-from pydantic.json_schema import SkipJsonSchema
-
-from unique_toolkit.agentic.postprocessor.postprocessor_manager import (
-    ResponsesApiPostprocessor,
-)
-from unique_toolkit.agentic.tools.config import get_configuration_dict
-from unique_toolkit.language_model.schemas import ResponsesLanguageModelStreamResponse
-
-logger = logging.getLogger(__name__)
-
-
-class ShowExecutedCodePostprocessorConfig(BaseModel):
-    model_config = get_configuration_dict()
-    remove_from_history: SkipJsonSchema[bool] = (
-        Field(  # At the moment, it's not possible to keep executed code in the history
-            default=True,
-            description="If set, the code interpreter call will be removed from the history on subsequent calls to the assistant.",
-        )
-    )
-
-
-class ShowExecutedCodePostprocessor(ResponsesApiPostprocessor):
-    """Strips legacy `<details>` code blocks from the stored conversation history.
-
-    Executed code used to be prepended to the assistant message as a
-    `<details><summary>Code Interpreter Call</summary>` block. Since UN-25450 the
-    code travels inside each `codeExecution` fence instead, so nothing is
-    prepended any more. Messages written before that change still hold the old
-    block, so it is removed here before the history goes back to the model.
-    """
-
-    def __init__(self, config: ShowExecutedCodePostprocessorConfig):
-        super().__init__(self.__class__.__name__)
-        self._config = config
-
-    # `remove_from_text` does all the work, but the manager still calls `run` and
-    # `apply_postprocessing_to_response` on every turn and the base class raises
-    # NotImplementedError. So both must stay, even though they do nothing.
-    @override
-    async def run(self, loop_response: ResponsesLanguageModelStreamResponse) -> None:
-        return None
-
-    @override
-    def apply_postprocessing_to_response(
-        self, loop_response: ResponsesLanguageModelStreamResponse
-    ) -> bool:
-        return False
-
-    @override
-    async def remove_from_text(self, text) -> str:
-        if not self._config.remove_from_history:
-            return text
-        pattern = r"<details><summary>Code Interpreter Call</summary>.*?</details>"
-        return re.sub(pattern, "", text, flags=re.DOTALL)