unique-toolkit 2026.40.0.dev10__tar.gz → 2026.40.0.dev12__tar.gz

{unique_toolkit-2026.40.0.dev10 → unique_toolkit-2026.40.0.dev12}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-toolkit
-Version: 2026.40.0.dev10
+Version: 2026.40.0.dev12
 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.dev6,<2026.40.0rc0
+Requires-Dist: unique-sdk>=2026.40.0.dev7,<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.dev10 → unique_toolkit-2026.40.0.dev12}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "unique_toolkit"
-version = "2026.40.0.dev10"
+version = "2026.40.0.dev12"
 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.dev6,<2026.40.0rc0",
+    "unique-sdk>=2026.40.0.dev7,<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.dev10 → unique_toolkit-2026.40.0.dev12}/unique_toolkit/agentic/loop_runner/__init__.py

@@ -21,6 +21,7 @@ from unique_toolkit.agentic.loop_runner.runners import (
     BasicLoopIterationRunner,
     BasicLoopIterationRunnerConfig,
     MistralLoopIterationRunner,
+    PromptForcedToolLoopIterationRunner,
     QwenLoopIterationRunner,
     ResponsesBasicLoopIterationRunner,
 )
@@ -34,6 +35,7 @@ __all__ = [
     "BasicLoopIterationRunnerConfig",
     "BasicLoopIterationRunner",
     "MistralLoopIterationRunner",
+    "PromptForcedToolLoopIterationRunner",
     "QwenLoopIterationRunner",
     "handle_forced_tools_iteration",
     "handle_last_iteration",

{unique_toolkit-2026.40.0.dev10 → unique_toolkit-2026.40.0.dev12}/unique_toolkit/agentic/loop_runner/runners/__init__.py

@@ -6,6 +6,10 @@ from unique_toolkit.agentic.loop_runner.runners.basic import (
 from unique_toolkit.agentic.loop_runner.runners.mistral import (
     MistralLoopIterationRunner,
 )
+from unique_toolkit.agentic.loop_runner.runners.prompt_forced_tool import (
+    PROMPT_FORCED_TOOL_CALL_INSTRUCTION,
+    PromptForcedToolLoopIterationRunner,
+)
 from unique_toolkit.agentic.loop_runner.runners.qwen import (
     QWEN_FORCED_TOOL_CALL_INSTRUCTION,
     QWEN_LAST_ITERATION_INSTRUCTION,
@@ -18,6 +22,8 @@ __all__ = [
     "BasicLoopIterationRunner",
     "ResponsesBasicLoopIterationRunner",
     "MistralLoopIterationRunner",
+    "PROMPT_FORCED_TOOL_CALL_INSTRUCTION",
+    "PromptForcedToolLoopIterationRunner",
     "QwenLoopIterationRunner",
     "QWEN_FORCED_TOOL_CALL_INSTRUCTION",
     "QWEN_LAST_ITERATION_INSTRUCTION",

unique_toolkit-2026.40.0.dev12/unique_toolkit/agentic/loop_runner/runners/prompt_forced_tool/__init__.py

@@ -0,0 +1,9 @@
+from unique_toolkit.agentic.loop_runner.runners.prompt_forced_tool.runner import (
+    PROMPT_FORCED_TOOL_CALL_INSTRUCTION,
+    PromptForcedToolLoopIterationRunner,
+)
+
+__all__ = [
+    "PROMPT_FORCED_TOOL_CALL_INSTRUCTION",
+    "PromptForcedToolLoopIterationRunner",
+]

unique_toolkit-2026.40.0.dev12/unique_toolkit/agentic/loop_runner/runners/prompt_forced_tool/helpers.py

@@ -0,0 +1,32 @@
+"""Message helpers for models that reject a forced tool_choice."""
+
+from unique_toolkit.language_model.schemas import (
+    LanguageModelMessageRole,
+    LanguageModelMessages,
+)
+
+
+def append_prompt_forced_tool_instruction(
+    *,
+    messages: LanguageModelMessages,
+    instruction: str,
+) -> LanguageModelMessages:
+    """Append a must-call instruction to the last user message.
+
+    Args:
+        messages (LanguageModelMessages): Conversation copied for this tool.
+        instruction (str): Instruction that names the tool to call.
+
+    Returns:
+        LanguageModelMessages: A new message list with the instruction appended
+        when the last user message has string content.
+    """
+    messages_list = list(messages)
+    for i in range(len(messages_list) - 1, -1, -1):
+        msg = messages_list[i]
+        if msg.role == LanguageModelMessageRole.USER and isinstance(msg.content, str):
+            messages_list[i] = msg.model_copy(
+                update={"content": msg.content + "\n" + instruction}
+            )
+            break
+    return LanguageModelMessages(root=messages_list)

unique_toolkit-2026.40.0.dev12/unique_toolkit/agentic/loop_runner/runners/prompt_forced_tool/runner.py

@@ -0,0 +1,53 @@
+"""Loop runner for models that reject forced tool_choice values."""
+
+from typing import Unpack
+
+from unique_toolkit.agentic.loop_runner._iteration_handler_utils import (
+    run_forced_tools_iteration,
+)
+from unique_toolkit.agentic.loop_runner.base import _LoopIterationRunnerKwargs
+from unique_toolkit.agentic.loop_runner.runners.basic import (
+    BasicLoopIterationRunner,
+)
+from unique_toolkit.agentic.loop_runner.runners.prompt_forced_tool.helpers import (
+    append_prompt_forced_tool_instruction,
+)
+from unique_toolkit.chat.functions import LanguageModelStreamResponse
+
+PROMPT_FORCED_TOOL_CALL_INSTRUCTION = (
+    "You must call the tool {TOOL_NAME}. Call it instead of replying with text."
+)
+
+
+class PromptForcedToolLoopIterationRunner(BasicLoopIterationRunner):
+    """Runner for models that reject named and ``any`` tool choices.
+
+    Claude Opus 5.5 and Fable 5.1 return HTTP 400 for those ``tool_choice``
+    types. This runner keeps one request per forced tool, limits ``tools`` to
+    that tool, sends ``tool_choice="auto"``, and asks for the call in the prompt.
+    """
+
+    async def _handle_forced_tools(
+        self,
+        **kwargs: Unpack[_LoopIterationRunnerKwargs],
+    ) -> LanguageModelStreamResponse:
+        original_messages = kwargs["messages"].model_copy(deep=True)
+
+        def _prepare(
+            func_name: str | None,
+            per_choice_kwargs: _LoopIterationRunnerKwargs,
+        ) -> _LoopIterationRunnerKwargs:
+            instruction = PROMPT_FORCED_TOOL_CALL_INSTRUCTION.format(
+                TOOL_NAME=func_name or ""
+            )
+            per_choice_kwargs["messages"] = append_prompt_forced_tool_instruction(
+                messages=original_messages,
+                instruction=instruction,
+            )
+            return per_choice_kwargs
+
+        return await run_forced_tools_iteration(
+            loop_runner_kwargs=kwargs,
+            prepare_loop_runner_kwargs=_prepare,
+            tool_choice_override="auto",
+        )

{unique_toolkit-2026.40.0.dev10 → unique_toolkit-2026.40.0.dev12}/unique_toolkit/language_model/infos.py

@@ -110,6 +110,7 @@ class LanguageModelName(StrEnum):
     GEMINI_2_5_PRO_PREVIEW_0605 = "litellm:gemini-2-5-pro-preview-06-05"
     GEMINI_3_1_PRO_PREVIEW = "litellm:gemini-3-1-pro-preview"
     GEMINI_3_6_FLASH = "litellm:gemini-3-6-flash"
+    GEMINI_3_8_FLASH = "litellm:gemini-3-8-flash"
     GEMINI_3_FLASH_PREVIEW = "litellm:gemini-3-flash-preview"
     GEMINI_3_PRO_PREVIEW = "litellm:gemini-3-pro-preview"
     GROK_4_1_FAST_NON_REASONING = "litellm:grok-4-1-fast-non-reasoning"
@@ -372,6 +373,10 @@ class LanguageModelInfo(BaseModel):
 
     supported_reasoning_efforts: list[ReasoningEffort] | None = None
 
+    # False for models that return HTTP 400 when tool_choice is a named tool or
+    # "any". Only "auto" and "none" are accepted there.
+    supports_forced_tool_choice: bool = True
+
     _ENV_VAR: ClassVar[str] = "LANGUAGE_MODEL_INFOS"
 
     def get_encoder(self) -> TypeEncoder:
@@ -2028,6 +2033,7 @@ class LanguageModelInfo(BaseModel):
                     info_cutoff_at=date(2026, 6, 1),
                     published_at=date(2026, 9, 22),
                     supported_reasoning_efforts=[],
+                    supports_forced_tool_choice=False,
                 )
             case (
                 LanguageModelName.ANTHROPIC_CLAUDE_FABLE_5
@@ -2265,6 +2271,33 @@ class LanguageModelInfo(BaseModel):
                     published_at=date(2026, 7, 21),
                     supported_reasoning_efforts=["minimal", "low", "medium", "high"],
                 )
+            case LanguageModelName.GEMINI_3_8_FLASH:
+                return cls(
+                    name=model_name,
+                    capabilities=[
+                        ModelCapabilities.FUNCTION_CALLING,
+                        ModelCapabilities.STREAMING,
+                        ModelCapabilities.VISION,
+                        ModelCapabilities.STRUCTURED_OUTPUT,
+                        ModelCapabilities.REASONING,
+                    ],
+                    provider=LanguageModelProvider.LITELLM,
+                    family=ModelFamily.GOOGLE,
+                    version="gemini-3-8-flash",
+                    encoder_name=EncoderName.O200K_BASE,  # TODO: Update encoder with litellm
+                    token_limits=LanguageModelTokenLimits(
+                        # https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash
+                        token_limit_input=1_048_576,
+                        token_limit_output=65_536,
+                    ),
+                    # https://deepmind.google/models/model-cards/gemini-3-8-flash/
+                    info_cutoff_at=date(2026, 3, 1),
+                    # https://ai.google.dev/gemini-api/docs/changelog
+                    published_at=date(2026, 9, 2),
+                    # minimal is rejected; default thinking level is medium when unset.
+                    # https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash
+                    supported_reasoning_efforts=["low", "medium", "high"],
+                )
             case LanguageModelName.GEMINI_3_FLASH_PREVIEW:
                 return cls(
                     name=model_name,