PyPI - unique_toolkit - Versions diffs - 2026.44.0.dev5__tar.gz → 2026.44.0.dev7__tar.gz - Mend

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-toolkit
-Version: 2026.44.0.dev5
+Version: 2026.44.0.dev7
 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>=2026.44.0.dev8,<2026.44.0rc0
+Requires-Dist: unique-sdk>=2026.44.0.dev9,<2026.44.0rc0
 Requires-Dist: jambo>=0.1.2,<0.2
 Requires-Dist: jinja2>=3.1.6,<4
 Requires-Dist: docxtpl>=0.20.1,<0.21

@@ -1,6 +1,6 @@
 [project]
 name = "unique_toolkit"
-version = "2026.44.0.dev5"
+version = "2026.44.0.dev7"
 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>=2026.44.0.dev8,<2026.44.0rc0",
+    "unique-sdk>=2026.44.0.dev9,<2026.44.0rc0",
     "jambo>=0.1.2,<0.2",
     "jinja2>=3.1.6,<4",
     "docxtpl>=0.20.1,<0.21",

@@ -1,10 +1,4 @@
-"""Keep agent-loop prompts inside the provider's context window.
-The local reducer estimates tokens with an OpenAI tokenizer, which can undercount
-other providers by 2x or more. ``ContextOverflowGuard`` verifies near-limit
-payloads with the provider's own counter before the LLM call and recovers when the
-provider still rejects a prompt as too long.
-"""
+"""Keep agent-loop prompts inside the provider's context window, verified by the provider's own token counter."""
 from __future__ import annotations
@@ -225,7 +219,8 @@ class ContextOverflowGuard:
                 raise
             self._logger.warning(
                 "LLM call still exceeded the context window. Retrying with the "
-                "latest tool call results withheld."
+                "latest tool call results withheld.",
+                exc_info=retry_exc,
             )
             await recompose()
             return await run_iteration()

@@ -237,7 +237,8 @@ class HistoryManager:
         The withheld results' chunks are dropped from the reference manager. They
         are the newest chunks, so source numbers of earlier results stay valid.
-        Returns the number of withheld results.
+        Returns the number of newly withheld results; results that already hold
+        *notice* are not counted.
         """
         batch_start: int | None = None
         for index in range(len(self._loop_history) - 1, -1, -1):
@@ -256,12 +257,16 @@ class HistoryManager:
             message = self._loop_history[index]
             if not isinstance(message, LanguageModelToolMessage):
                 continue
+            if message.content == notice:
+                continue
             self._loop_history[index] = LanguageModelToolMessage(
                 content=notice,
                 tool_call_id=message.tool_call_id,
                 name=message.name,
             )
             withheld_tool_call_ids.append(message.tool_call_id)
+        if not withheld_tool_call_ids:
+            return 0
         withheld_chunks = {
             id(chunk)

@@ -140,11 +140,7 @@ class LoopTokenReducer:
         return self._last_token_count
     def tighten_effective_token_limit(self, factor: float) -> None:
-        """Shrink the input-token budget for the rest of the run.
-        Used after the provider rejected a payload this reducer counted as within
-        budget, i.e. the encoder underestimates the model's real token count.
-        """
+        """Shrink the input-token budget for the rest of the run after the provider counted more tokens than the local encoder."""
         previous_limit = self._effective_token_limit
         self._effective_token_limit = int(previous_limit * factor)
         self._logger.warning(

@@ -1,9 +1,4 @@
-"""Build request bodies for LiteLLM's Anthropic-compatible ``count_tokens`` endpoint.
-Agent-loop messages are in OpenAI chat format. The endpoint expects the Anthropic
-Messages format, so system messages, tool calls, tool results and images are
-converted before counting.
-"""
+"""Convert OpenAI chat payloads to request bodies for LiteLLM's Anthropic-compatible ``count_tokens`` endpoint."""
 from __future__ import annotations

unique_toolkit 2026.44.0.dev5__tar.gz → 2026.44.0.dev7__tar.gz