language-model-common 2.0.90__tar.gz → 2.0.92__tar.gz

{language_model_common-2.0.90/language_model_common.egg-info → language_model_common-2.0.92}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: language-model-common
-Version: 2.0.90
+Version: 2.0.92
 Summary: Provides the underlying framework to enhance langchain and add loading configurations
 Author-email: Imran Qureshi <imran.qureshi@bwell.com>
 License: Apache-2.0

language_model_common-2.0.92/VERSION

@@ -0,0 +1 @@
+2.0.92

{language_model_common-2.0.90 → language_model_common-2.0.92/language_model_common.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: language-model-common
-Version: 2.0.90
+Version: 2.0.92
 Summary: Provides the underlying framework to enhance langchain and add loading configurations
 Author-email: Imran Qureshi <imran.qureshi@bwell.com>
 License: Apache-2.0

{language_model_common-2.0.90 → language_model_common-2.0.92}/languagemodelcommon/mcp/mcp_client/tool_invocation.py

@@ -269,11 +269,14 @@ async def _execute_tool_call_with_heartbeat(
                         e,
                         name,
                     )
-    except asyncio.CancelledError:
-        call_task.cancel()
-        with contextlib.suppress(asyncio.CancelledError):
-            await call_task
-        raise
+    finally:
+        # Ensure call_task is never left orphaned, whether this coroutine
+        # exits via CancelledError, a return, or any other exception raised
+        # from the loop body (e.g. adispatch_custom_event).
+        if not call_task.done():
+            call_task.cancel()
+            with contextlib.suppress(asyncio.CancelledError):
+                await call_task
 
 
 def _make_execute_tool(

{language_model_common-2.0.90 → language_model_common-2.0.92}/languagemodelcommon/models/model_factory.py

@@ -151,6 +151,39 @@ class ModelFactory:
     def _is_anthropic_model(model_name: str) -> bool:
         return model_name.startswith(("anthropic.", "us.anthropic."))
 
+    @staticmethod
+    def _resolve_anthropic_bedrock_max_tokens(*, model_name: str) -> int | None:
+        """Look up the real max_output_tokens for a Bedrock Anthropic model ID.
+
+        ``ChatAnthropicBedrock`` inherits ``ChatAnthropic.set_default_max_tokens``,
+        which resolves an unset ``max_tokens`` from ``langchain_anthropic``'s own
+        model-profile registry — keyed by bare Anthropic API names (e.g.
+        ``claude-sonnet-4-5-20250929``), not Bedrock cross-region inference
+        profile IDs (e.g. ``us.anthropic.claude-sonnet-4-5-20250929-v1:0``). For
+        every Bedrock-prefixed ID that lookup misses, silently capping output at
+        ``_FALLBACK_MAX_OUTPUT_TOKENS`` (4096) even when the model supports far
+        more (64000 for Claude Sonnet 4.5) — which truncates long tool-call
+        arguments (e.g. a full skill/document body) mid-generation.
+        ``langchain_aws``'s own Bedrock-aware profile registry (used elsewhere,
+        e.g. by ``ChatAnthropicBedrock._set_model_profile``, but too late to
+        affect ``max_tokens``) does have entries keyed by the full Bedrock ID, so
+        look the real limit up there and pass it through explicitly.
+        """
+        try:
+            from langchain_aws.chat_models.anthropic import (
+                _get_default_model_profile,
+            )
+        except ImportError:
+            logger.warning(
+                "langchain_aws._get_default_model_profile unavailable; "
+                "falling back to unset max_tokens for %s",
+                model_name,
+            )
+            return None
+        profile = _get_default_model_profile(model_name)
+        max_output_tokens = profile.get("max_output_tokens")
+        return int(max_output_tokens) if max_output_tokens else None
+
     def _create_anthropic_bedrock_model(
         self,
         *,
@@ -164,6 +197,21 @@ class ModelFactory:
         from pydantic import SecretStr
         import boto3
 
+        model_parameters_dict = dict(model_parameters_dict)
+
+        if "max_tokens" not in model_parameters_dict:
+            resolved_max_tokens = self._resolve_anthropic_bedrock_max_tokens(
+                model_name=model_name
+            )
+            if resolved_max_tokens is not None:
+                model_parameters_dict["max_tokens"] = resolved_max_tokens
+                logger.info(
+                    "Resolved max_tokens=%d for Bedrock model %s from profile "
+                    "(overriding langchain_anthropic's unprefixed-lookup default)",
+                    resolved_max_tokens,
+                    model_name,
+                )
+
         if thinking_budget and thinking_budget > 0:
             model_parameters_dict["thinking"] = {
                 "type": "enabled",

{language_model_common-2.0.90 → language_model_common-2.0.92}/languagemodelcommon/utilities/environment/language_model_common_environment_variables.py

@@ -17,6 +17,7 @@ logger = logging.getLogger(__name__)
 DEFAULT_STREAMING_BUFFER_FLUSH_INTERVAL_SECONDS = 0.05
 DEFAULT_LANGGRAPH_MAX_CONCURRENCY = 4
 DEFAULT_LANGGRAPH_RECURSION_LIMIT = 100
+DEFAULT_MCP_TOOL_HEARTBEAT_INTERVAL_SECONDS = 15.0
 
 # Default generic error message when not exposing technical details
 DEFAULT_GENERIC_ERROR_MESSAGE = (
@@ -293,7 +294,19 @@ class LanguageModelCommonEnvironmentVariables(
     def mcp_tool_heartbeat_interval_seconds(self) -> float:
         """Interval in seconds between synthetic heartbeat events emitted
         while an MCP tool call is in flight without reporting progress."""
-        return float(os.environ.get("MCP_TOOL_HEARTBEAT_INTERVAL_SECONDS", "15"))
+        value = os.environ.get("MCP_TOOL_HEARTBEAT_INTERVAL_SECONDS")
+        if value is None:
+            return DEFAULT_MCP_TOOL_HEARTBEAT_INTERVAL_SECONDS
+        try:
+            parsed = float(value)
+            return max(1.0, parsed)
+        except ValueError:
+            logger.warning(
+                "Invalid MCP_TOOL_HEARTBEAT_INTERVAL_SECONDS value '%s'; using default=%s",
+                value,
+                DEFAULT_MCP_TOOL_HEARTBEAT_INTERVAL_SECONDS,
+            )
+            return DEFAULT_MCP_TOOL_HEARTBEAT_INTERVAL_SECONDS
 
     @property
     def emit_tool_heartbeat_in_chat_completions(self) -> bool:

language_model_common-2.0.90/VERSION

@@ -1 +0,0 @@
-2.0.90