letta-nightly 0.8.4.dev20250614104137__py3-none-any.whl → 0.8.4.dev20250615221417__py3-none-any.whl

letta/functions/helpers.py

@@ -14,6 +14,7 @@ from letta.orm.errors import NoResultFound
 from letta.schemas.enums import MessageRole
 from letta.schemas.letta_message import AssistantMessage
 from letta.schemas.letta_response import LettaResponse
+from letta.schemas.letta_stop_reason import LettaStopReason, StopReasonType
 from letta.schemas.message import Message, MessageCreate
 from letta.schemas.user import User
 from letta.server.rest_api.utils import get_letta_server
@@ -292,7 +293,11 @@ async def _send_message_to_agent_no_stream(
     )
 
     final_messages = interface.get_captured_send_messages()
-    return LettaResponse(messages=final_messages, usage=usage_stats)
+    return LettaResponse(
+        messages=final_messages,
+        stop_reason=LettaStopReason(stop_reason=StopReasonType.end_turn.value),
+        usage=usage_stats,
+    )
 
 
 async def _async_send_message_with_retries(

letta/functions/mcp_client/types.py

@@ -4,6 +4,10 @@ from typing import List, Optional
 from mcp import Tool
 from pydantic import BaseModel, Field
 
+# MCP Authentication Constants
+MCP_AUTH_HEADER_AUTHORIZATION = "Authorization"
+MCP_AUTH_TOKEN_BEARER_PREFIX = "Bearer"
+
 
 class MCPTool(Tool):
     """A simple wrapper around MCP's tool definition (to avoid conflict with our own)"""
@@ -12,6 +16,7 @@ class MCPTool(Tool):
 class MCPServerType(str, Enum):
     SSE = "sse"
     STDIO = "stdio"
+    STREAMABLE_HTTP = "streamable_http"
 
 
 class BaseServerConfig(BaseModel):
@@ -20,14 +25,44 @@ class BaseServerConfig(BaseModel):
 
 
 class SSEServerConfig(BaseServerConfig):
+    """
+    Configuration for an MCP server using SSE
+
+    Authentication can be provided in multiple ways:
+    1. Using auth_header + auth_token: Will add a specific header with the token
+       Example: auth_header="Authorization", auth_token="Bearer abc123"
+
+    2. Using the custom_headers dict: For more complex authentication scenarios
+       Example: custom_headers={"X-API-Key": "abc123", "X-Custom-Header": "value"}
+    """
+
     type: MCPServerType = MCPServerType.SSE
     server_url: str = Field(..., description="The URL of the server (MCP SSE client will connect to this URL)")
+    auth_header: Optional[str] = Field(None, description="The name of the authentication header (e.g., 'Authorization')")
+    auth_token: Optional[str] = Field(None, description="The authentication token or API key value")
+    custom_headers: Optional[dict[str, str]] = Field(None, description="Custom HTTP headers to include with SSE requests")
+
+    def resolve_token(self) -> Optional[str]:
+        if self.auth_token and self.auth_token.startswith(f"{MCP_AUTH_TOKEN_BEARER_PREFIX} "):
+            return self.auth_token[len(f"{MCP_AUTH_TOKEN_BEARER_PREFIX} ") :]
+        return self.auth_token
 
     def to_dict(self) -> dict:
         values = {
             "transport": "sse",
             "url": self.server_url,
         }
+
+        # TODO: handle custom headers
+        if self.custom_headers is not None or (self.auth_header is not None and self.auth_token is not None):
+            headers = self.custom_headers.copy() if self.custom_headers else {}
+
+            # Add auth header if specified
+            if self.auth_header is not None and self.auth_token is not None:
+                headers[self.auth_header] = self.auth_token
+
+            values["headers"] = headers
+
         return values
 
 
@@ -46,3 +81,63 @@ class StdioServerConfig(BaseServerConfig):
         if self.env is not None:
             values["env"] = self.env
         return values
+
+
+class StreamableHTTPServerConfig(BaseServerConfig):
+    """
+    Configuration for an MCP server using Streamable HTTP
+
+    Authentication can be provided in multiple ways:
+    1. Using auth_header + auth_token: Will add a specific header with the token
+       Example: auth_header="Authorization", auth_token="Bearer abc123"
+
+    2. Using the custom_headers dict: For more complex authentication scenarios
+       Example: custom_headers={"X-API-Key": "abc123", "X-Custom-Header": "value"}
+    """
+
+    type: MCPServerType = MCPServerType.STREAMABLE_HTTP
+    server_url: str = Field(..., description="The URL path for the streamable HTTP server (e.g., 'example/mcp')")
+    auth_header: Optional[str] = Field(None, description="The name of the authentication header (e.g., 'Authorization')")
+    auth_token: Optional[str] = Field(None, description="The authentication token or API key value")
+    custom_headers: Optional[dict[str, str]] = Field(None, description="Custom HTTP headers to include with streamable HTTP requests")
+
+    def resolve_token(self) -> Optional[str]:
+        if self.auth_token and self.auth_token.startswith(f"{MCP_AUTH_TOKEN_BEARER_PREFIX} "):
+            return self.auth_token[len(f"{MCP_AUTH_TOKEN_BEARER_PREFIX} ") :]
+        return self.auth_token
+
+    def model_post_init(self, __context) -> None:
+        """Validate the server URL format."""
+        # Basic validation for streamable HTTP URLs
+        if not self.server_url:
+            raise ValueError("server_url cannot be empty")
+
+        # For streamable HTTP, the URL should typically be a path or full URL
+        # We'll be lenient and allow both formats
+        if self.server_url.startswith("http://") or self.server_url.startswith("https://"):
+            # Full URL format - this is what the user is trying
+            pass
+        elif "/" in self.server_url:
+            # Path format like "example/mcp" - this is the typical format
+            pass
+        else:
+            # Single word - might be valid but warn in logs
+            pass
+
+    def to_dict(self) -> dict:
+        values = {
+            "transport": "streamable_http",
+            "url": self.server_url,
+        }
+
+        # Handle custom headers
+        if self.custom_headers is not None or (self.auth_header is not None and self.auth_token is not None):
+            headers = self.custom_headers.copy() if self.custom_headers else {}
+
+            # Add auth header if specified
+            if self.auth_header is not None and self.auth_token is not None:
+                headers[self.auth_header] = self.auth_token
+
+            values["headers"] = headers
+
+        return values

letta/groups/sleeptime_multi_agent_v2.py

@@ -144,7 +144,8 @@ class SleeptimeMultiAgentV2(BaseAgent):
         for message in response.messages:
             yield f"data: {message.model_dump_json()}\n\n"
 
-        yield f"data: {response.usage.model_dump_json()}\n\n"
+        for finish_chunk in self.get_finish_chunks_for_stream(response.usage):
+            yield f"data: {finish_chunk}\n\n"
 
     @trace_method
     async def step_stream(

letta/helpers/decorators.py

@@ -1,7 +1,13 @@
 import inspect
+import json
+from dataclasses import dataclass
 from functools import wraps
 from typing import Callable
 
+from pydantic import BaseModel
+
+from letta.constants import REDIS_DEFAULT_CACHE_PREFIX
+from letta.data_sources.redis_client import NoopAsyncRedisClient, get_redis_client
 from letta.log import get_logger
 from letta.plugins.plugins import get_experimental_checker
 from letta.settings import settings
@@ -67,3 +73,88 @@ def deprecated(message: str):
         return wrapper
 
     return decorator
+
+
+@dataclass
+class CacheStats:
+    """Note: this will be approximate to not add overhead of locking on counters.
+    For exact measurements, use redis or track in other places.
+    """
+
+    hits: int = 0
+    misses: int = 0
+    invalidations: int = 0
+
+
+def async_redis_cache(
+    key_func: Callable, prefix: str = REDIS_DEFAULT_CACHE_PREFIX, ttl_s: int = 300, model_class: type[BaseModel] | None = None
+):
+    """
+    Decorator for caching async function results in Redis. May be a Noop if redis is not available.
+    Will handle pydantic objects and raw values.
+
+    Attempts to write to and retrieve from cache, but does not fail on those cases
+
+    Args:
+        key_func: function to generate cache key (preferably lowercase strings to follow redis convention)
+        prefix: cache key prefix
+        ttl_s: time to live (s)
+        model_class: custom pydantic model class for serialization/deserialization
+
+    TODO (cliandy): move to class with generics for type hints
+    """
+
+    def decorator(func):
+        stats = CacheStats()
+
+        @wraps(func)
+        async def async_wrapper(*args, **kwargs):
+            redis_client = await get_redis_client()
+
+            # Don't bother going through other operations for no reason.
+            if isinstance(redis_client, NoopAsyncRedisClient):
+                return await func(*args, **kwargs)
+            cache_key = get_cache_key(*args, **kwargs)
+            cached_value = await redis_client.get(cache_key)
+
+            try:
+                if cached_value is not None:
+                    stats.hits += 1
+                    if model_class:
+                        return model_class.model_validate_json(cached_value)
+                    return json.loads(cached_value)
+            except Exception as e:
+                logger.warning(f"Failed to retrieve value from cache: {e}")
+
+            stats.misses += 1
+            result = await func(*args, **kwargs)
+            try:
+                if model_class:
+                    await redis_client.set(cache_key, result.model_dump_json(), ex=ttl_s)
+                elif isinstance(result, (dict, list, str, int, float, bool)):
+                    await redis_client.set(cache_key, json.dumps(result), ex=ttl_s)
+                else:
+                    logger.warning(f"Cannot cache result of type {type(result).__name__} for {func.__name__}")
+            except Exception as e:
+                logger.warning(f"Redis cache set failed: {e}")
+            return result
+
+        async def invalidate(*args, **kwargs) -> bool:
+            stats.invalidations += 1
+            try:
+                redis_client = await get_redis_client()
+                cache_key = get_cache_key(*args, **kwargs)
+                return (await redis_client.delete(cache_key)) > 0
+            except Exception as e:
+                logger.error(f"Failed to invalidate cache: {e}")
+                return False
+
+        def get_cache_key(*args, **kwargs):
+            return f"{prefix}:{key_func(*args, **kwargs)}"
+
+        # async_wrapper.cache_invalidate = invalidate
+        async_wrapper.cache_key_func = get_cache_key
+        async_wrapper.cache_stats = stats
+        return async_wrapper
+
+    return decorator

letta/interfaces/anthropic_streaming_interface.py

@@ -26,6 +26,8 @@ from letta.constants import DEFAULT_MESSAGE_TOOL, DEFAULT_MESSAGE_TOOL_KWARG
 from letta.helpers.datetime_helpers import get_utc_timestamp_ns, ns_to_ms
 from letta.local_llm.constants import INNER_THOUGHTS_KWARG
 from letta.log import get_logger
+from letta.otel.context import get_ctx_attributes
+from letta.otel.metric_registry import MetricRegistry
 from letta.schemas.letta_message import (
     AssistantMessage,
     HiddenReasoningMessage,
@@ -35,6 +37,7 @@ from letta.schemas.letta_message import (
     ToolCallMessage,
 )
 from letta.schemas.letta_message_content import ReasoningContent, RedactedReasoningContent, TextContent
+from letta.schemas.letta_stop_reason import LettaStopReason, StopReasonType
 from letta.schemas.message import Message
 from letta.schemas.openai.chat_completion_response import FunctionCall, ToolCall
 from letta.server.rest_api.json_parser import JSONParser, PydanticJSONParser
@@ -90,6 +93,8 @@ class AnthropicStreamingInterface:
 
     def get_tool_call_object(self) -> ToolCall:
         """Useful for agent loop"""
+        if not self.tool_call_name:
+            raise ValueError("No tool call returned")
         # hack for tool rules
         try:
             tool_input = json.loads(self.accumulated_tool_call_args)
@@ -140,6 +145,10 @@ class AnthropicStreamingInterface:
                         ttft_span.add_event(
                             name="anthropic_time_to_first_token_ms", attributes={"anthropic_time_to_first_token_ms": ns_to_ms(ttft_ns)}
                         )
+                        metric_attributes = get_ctx_attributes()
+                        if isinstance(event, BetaRawMessageStartEvent):
+                            metric_attributes["model.name"] = event.message.model
+                        MetricRegistry().ttft_ms_histogram.record(ns_to_ms(ttft_ns), metric_attributes)
                         first_chunk = False
 
                     # TODO: Support BetaThinkingBlock, BetaRedactedThinkingBlock
@@ -377,6 +386,8 @@ class AnthropicStreamingInterface:
                         self.anthropic_mode = None
         except Exception as e:
             logger.error("Error processing stream: %s", e)
+            stop_reason = LettaStopReason(stop_reason=StopReasonType.error.value)
+            yield stop_reason
             raise
         finally:
             logger.info("AnthropicStreamingInterface: Stream processing complete.")