fast-agent-mcp 0.2.57__py3-none-any.whl → 0.3.0__py3-none-any.whl

fast_agent/llm/internal/passthrough.py

@@ -0,0 +1,137 @@
+import json  # Import at the module level
+from typing import Any, Dict, List, Optional
+
+from mcp import CallToolRequest, Tool
+from mcp.types import CallToolRequestParams, PromptMessage
+
+from fast_agent.core.logging.logger import get_logger
+from fast_agent.core.prompt import Prompt
+from fast_agent.llm.fastagent_llm import (
+    FastAgentLLM,
+    RequestParams,
+)
+from fast_agent.llm.provider_types import Provider
+from fast_agent.llm.usage_tracking import create_turn_usage_from_messages
+from fast_agent.mcp.helpers.content_helpers import get_text
+from fast_agent.types import PromptMessageExtended
+from fast_agent.types.llm_stop_reason import LlmStopReason
+
+CALL_TOOL_INDICATOR = "***CALL_TOOL"
+FIXED_RESPONSE_INDICATOR = "***FIXED_RESPONSE"
+
+
+class PassthroughLLM(FastAgentLLM):
+    """
+    A specialized LLM implementation that simply passes through input messages without modification.
+
+    This is useful for cases where you need an object with the AugmentedLLM interface
+    but want to preserve the original message without any processing, such as in a
+    parallel workflow where no fan-in aggregation is needed.
+    """
+
+    def __init__(
+        self, provider=Provider.FAST_AGENT, name: str = "Passthrough", **kwargs: dict[str, Any]
+    ) -> None:
+        super().__init__(name=name, provider=provider, **kwargs)
+        self.logger = get_logger(__name__)
+        self._messages = [PromptMessage]
+        self._fixed_response: str | None = None
+        self._correlation_id: int = 0
+
+    async def initialize(self) -> None:
+        pass
+
+    def _parse_tool_command(self, command: str) -> tuple[str, Optional[dict]]:
+        """
+        Parse a tool command string into tool name and arguments.
+
+        Args:
+            command: The command string in format "***CALL_TOOL <tool_name> [arguments_json]"
+
+        Returns:
+            Tuple of (tool_name, arguments_dict)
+
+        Raises:
+            ValueError: If command format is invalid
+        """
+        parts = command.split(" ", 2)
+        if len(parts) < 2:
+            raise ValueError("Invalid format. Expected '***CALL_TOOL <tool_name> [arguments_json]'")
+
+        tool_name = parts[1].strip()
+        arguments = None
+
+        if len(parts) > 2:
+            try:
+                arguments = json.loads(parts[2])
+            except json.JSONDecodeError:
+                raise ValueError(f"Invalid JSON arguments: {parts[2]}")
+
+        self.logger.info(f"Calling tool {tool_name} with arguments {arguments}")
+        return tool_name, arguments
+
+    async def _apply_prompt_provider_specific(
+        self,
+        multipart_messages: List["PromptMessageExtended"],
+        request_params: RequestParams | None = None,
+        tools: list[Tool] | None = None,
+        is_template: bool = False,
+    ) -> PromptMessageExtended:
+        # Add messages to history with proper is_prompt flag
+        self.history.extend(multipart_messages, is_prompt=is_template)
+
+        last_message = multipart_messages[-1]
+        tool_calls: Dict[str, CallToolRequest] = {}
+        stop_reason: LlmStopReason = LlmStopReason.END_TURN
+        if self.is_tool_call(last_message):
+            tool_name, arguments = self._parse_tool_command(last_message.first_text())
+            tool_calls["correlationId" + str(self._correlation_id)] = CallToolRequest(
+                method="tools/call",
+                params=CallToolRequestParams(name=tool_name, arguments=arguments),
+            )
+            self._correlation_id += 1
+            stop_reason = LlmStopReason.TOOL_USE
+
+        if last_message.first_text().startswith(FIXED_RESPONSE_INDICATOR):
+            self._fixed_response = (
+                last_message.first_text().split(FIXED_RESPONSE_INDICATOR, 1)[1].strip()
+            )
+
+        if len(last_message.tool_results or {}) > 0:
+            assert last_message.tool_results
+            concatenated_content = " ".join(
+                [
+                    (get_text(tool_result.content[0]) or "<empty>")
+                    for tool_result in last_message.tool_results.values()
+                ]
+            )
+            result = Prompt.assistant(concatenated_content, stop_reason=stop_reason)
+
+        elif self._fixed_response:
+            result = Prompt.assistant(
+                self._fixed_response, tool_calls=tool_calls, stop_reason=stop_reason
+            )
+        else:
+            concatenated_content = "\n".join(
+                [message.all_text() for message in multipart_messages if "user" == message.role]
+            )
+            result = Prompt.assistant(
+                concatenated_content,
+                tool_calls=tool_calls,
+                stop_reason=stop_reason,
+            )
+
+        turn_usage = create_turn_usage_from_messages(
+            input_content=multipart_messages[-1].all_text(),
+            output_content=result.all_text(),
+            model="passthrough",
+            model_type="passthrough",
+            tool_calls=len(tool_calls),
+            delay_seconds=0.0,
+        )
+        self.usage_accumulator.add_turn(turn_usage)
+
+        return result
+
+    def is_tool_call(self, message: PromptMessageExtended) -> bool:
+        return message.first_text().startswith(CALL_TOOL_INDICATOR)

mcp_agent/llm/augmented_llm_playback.py → fast_agent/llm/internal/playback.py

@@ -1,16 +1,19 @@
 from typing import Any, List, Type, Union
 
+from mcp import Tool
 from mcp.types import PromptMessage
 
-from mcp_agent.core.exceptions import ModelConfigError
-from mcp_agent.core.prompt import Prompt
-from mcp_agent.llm.augmented_llm import RequestParams
-from mcp_agent.llm.augmented_llm_passthrough import PassthroughLLM
-from mcp_agent.llm.provider_types import Provider
-from mcp_agent.llm.usage_tracking import create_turn_usage_from_messages
-from mcp_agent.mcp.interfaces import ModelT
-from mcp_agent.mcp.prompt_message_multipart import PromptMessageMultipart
-from mcp_agent.mcp.prompts.prompt_helpers import MessageContent
+from fast_agent.core.exceptions import ModelConfigError
+from fast_agent.core.prompt import Prompt
+from fast_agent.interfaces import ModelT
+from fast_agent.llm.internal.passthrough import PassthroughLLM
+from fast_agent.llm.provider_types import Provider
+from fast_agent.llm.usage_tracking import create_turn_usage_from_messages
+from fast_agent.mcp.helpers.content_helpers import normalize_to_extended_list
+from fast_agent.mcp.prompts.prompt_helpers import MessageContent
+from fast_agent.types import PromptMessageExtended, RequestParams
+
+# TODO -- support tool usage/replay
 
 
 class PlaybackLLM(PassthroughLLM):
@@ -28,11 +31,11 @@ class PlaybackLLM(PassthroughLLM):
 
     def __init__(self, name: str = "Playback", **kwargs: dict[str, Any]) -> None:
         super().__init__(name=name, provider=Provider.FAST_AGENT, **kwargs)
-        self._messages: List[PromptMessageMultipart] = []
+        self._messages: List[PromptMessageExtended] = []
         self._current_index = -1
         self._overage = -1
 
-    def _get_next_assistant_message(self) -> PromptMessageMultipart:
+    def _get_next_assistant_message(self) -> PromptMessageExtended:
         """
         Get the next assistant message from the loaded messages.
         Increments the current message index and skips user messages.
@@ -53,14 +56,23 @@ class PlaybackLLM(PassthroughLLM):
 
     async def generate(
         self,
-        multipart_messages: List[Union[PromptMessageMultipart, PromptMessage]],
+        messages: Union[
+            str,
+            PromptMessage,
+            PromptMessageExtended,
+            List[Union[str, PromptMessage, PromptMessageExtended]],
+        ],
         request_params: RequestParams | None = None,
-    ) -> PromptMessageMultipart:
+        tools: List[Tool] | None = None,
+    ) -> PromptMessageExtended:
         """
         Handle playback of messages in two modes:
         1. First call: store messages for playback and return "HISTORY LOADED"
         2. Subsequent calls: return the next assistant message
         """
+        # Normalize all input types to a list of PromptMessageExtended
+        multipart_messages = normalize_to_extended_list(messages)
+
         # If this is the first call (initialization) or we're loading a prompt template
         # with multiple messages (comes from apply_prompt)
         if -1 == self._current_index:
@@ -72,24 +84,16 @@ class PlaybackLLM(PassthroughLLM):
             # Reset the index to the beginning for proper playback
             self._current_index = 0
 
-            await self.show_assistant_message(
-                message_text=f"HISTORY LOADED ({len(self._messages)} messages)",
-                title="ASSISTANT/PLAYBACK",
-            )
-
             # In PlaybackLLM, we always return "HISTORY LOADED" on initialization,
             # regardless of the prompt content. The next call will return messages.
-            return Prompt.assistant("HISTORY LOADED")
+            return Prompt.assistant(f"HISTORY LOADED ({len(self._messages)}) messages")
 
         response = self._get_next_assistant_message()
-        await self.show_assistant_message(
-            message_text=MessageContent.get_first_text(response), title="ASSISTANT/PLAYBACK"
-        )
 
         # Track usage for this playback "turn"
         try:
             input_content = str(multipart_messages) if multipart_messages else ""
-            output_content = MessageContent.get_first_text(response)
+            output_content = MessageContent.get_first_text(response) or ""
 
             turn_usage = create_turn_usage_from_messages(
                 input_content=input_content,
@@ -108,10 +112,10 @@ class PlaybackLLM(PassthroughLLM):
 
     async def structured(
         self,
-        multipart_messages: List[Union[PromptMessageMultipart, PromptMessage]],
+        messages: List[PromptMessageExtended],
         model: Type[ModelT],
         request_params: RequestParams | None = None,
-    ) -> tuple[ModelT | None, PromptMessageMultipart]:
+    ) -> tuple[ModelT | None, PromptMessageExtended]:
         """
         Handle structured requests by returning the next assistant message.
         """

mcp_agent/llm/augmented_llm_silent.py → fast_agent/llm/internal/silent.py

@@ -2,47 +2,40 @@
 
 from typing import Any
 
-from mcp_agent.llm.augmented_llm_passthrough import PassthroughLLM
-from mcp_agent.llm.provider_types import Provider
-from mcp_agent.llm.usage_tracking import TurnUsage, UsageAccumulator
+from fast_agent.llm.internal.passthrough import PassthroughLLM
+from fast_agent.llm.provider_types import Provider
+from fast_agent.llm.usage_tracking import TurnUsage, UsageAccumulator
 
 
 class ZeroUsageAccumulator(UsageAccumulator):
     """Usage accumulator that always reports zero usage."""
-    
+
     def add_turn(self, turn: TurnUsage) -> None:
         """Override to do nothing - no usage accumulation."""
         pass
 
 
+# TODO -- this won't work anymore
 class SilentLLM(PassthroughLLM):
     """
     A specialized LLM that processes messages like PassthroughLLM but suppresses all display output.
-    
+
     This is particularly useful for parallel agent workflows where the fan-in agent
     should aggregate results without polluting the console with intermediate output.
     Token counting is disabled - the model always reports zero usage.
     """
-    
+
     def __init__(
         self, provider=Provider.FAST_AGENT, name: str = "Silent", **kwargs: dict[str, Any]
     ) -> None:
         super().__init__(name=name, provider=provider, **kwargs)
         # Override with zero usage accumulator - silent model reports no usage
         self.usage_accumulator = ZeroUsageAccumulator()
-    
-    def show_user_message(self, message: Any, **kwargs) -> None:
-        """Override to suppress user message display."""
-        pass
-    
-    async def show_assistant_message(self, message: Any, **kwargs) -> None:
-        """Override to suppress assistant message display."""
-        pass
-    
+
     def show_tool_calls(self, tool_calls: Any, **kwargs) -> None:
         """Override to suppress tool call display."""
         pass
-    
+
     def show_tool_results(self, tool_results: Any, **kwargs) -> None:
         """Override to suppress tool result display."""
-        pass
+        pass

fast_agent/llm/internal/slow.py

@@ -0,0 +1,38 @@
+import asyncio
+from typing import Any, List
+
+from mcp import Tool
+
+from fast_agent.llm.fastagent_llm import (
+    RequestParams,
+)
+from fast_agent.llm.internal.passthrough import PassthroughLLM
+from fast_agent.llm.provider_types import Provider
+from fast_agent.types import PromptMessageExtended
+
+
+class SlowLLM(PassthroughLLM):
+    """
+    A specialized LLM implementation that sleeps for 3 seconds before responding like PassthroughLLM.
+
+    This is useful for testing scenarios where you want to simulate slow responses
+    or for debugging timing-related issues in parallel workflows.
+    """
+
+    def __init__(
+        self, provider=Provider.FAST_AGENT, name: str = "Slow", **kwargs: dict[str, Any]
+    ) -> None:
+        super().__init__(name=name, provider=provider, **kwargs)
+
+    async def _apply_prompt_provider_specific(
+        self,
+        multipart_messages: List["PromptMessageExtended"],
+        request_params: RequestParams | None = None,
+        tools: list[Tool] | None = None,
+        is_template: bool = False,
+    ) -> PromptMessageExtended:
+        """Sleep for 3 seconds then apply prompt like PassthroughLLM."""
+        await asyncio.sleep(3)
+        return await super()._apply_prompt_provider_specific(
+            multipart_messages, request_params, tools, is_template
+        )

{mcp_agent → fast_agent}/llm/memory.py

@@ -35,7 +35,9 @@ class SimpleMemory(Memory, Generic[MessageParamT]):
     def __init__(self) -> None:
         self.history: List[MessageParamT] = []
         self.prompt_messages: List[MessageParamT] = []  # Always included
-        self.conversation_cache_positions: List[int] = []  # Track active conversation cache positions
+        self.conversation_cache_positions: List[
+            int
+        ] = []  # Track active conversation cache positions
         self.cache_walk_distance: int = 6  # Messages between cache blocks
         self.max_conversation_cache_blocks: int = 2  # Maximum conversation cache blocks
 
@@ -109,79 +111,85 @@ class SimpleMemory(Memory, Generic[MessageParamT]):
     def should_apply_conversation_cache(self) -> bool:
         """
         Determine if conversation caching should be applied based on walking algorithm.
-        
+
         Returns:
             True if we should add or update cache blocks
         """
         total_messages = len(self.history)
-        
+
         # Need at least cache_walk_distance messages to start caching
         if total_messages < self.cache_walk_distance:
             return False
-            
+
         # Check if we need to add a new cache block
-        return len(self._calculate_cache_positions(total_messages)) != len(self.conversation_cache_positions)
-    
+        return len(self._calculate_cache_positions(total_messages)) != len(
+            self.conversation_cache_positions
+        )
+
     def _calculate_cache_positions(self, total_conversation_messages: int) -> List[int]:
         """
         Calculate where cache blocks should be placed using walking algorithm.
-        
+
         Args:
             total_conversation_messages: Number of conversation messages (not including prompts)
-            
+
         Returns:
             List of positions (relative to conversation start) where cache should be placed
         """
         positions = []
-        
+
         # Place cache blocks every cache_walk_distance messages
-        for i in range(self.cache_walk_distance - 1, total_conversation_messages, self.cache_walk_distance):
+        for i in range(
+            self.cache_walk_distance - 1, total_conversation_messages, self.cache_walk_distance
+        ):
             positions.append(i)
             if len(positions) >= self.max_conversation_cache_blocks:
                 break
-        
+
         # Keep only the most recent cache blocks (walking behavior)
         if len(positions) > self.max_conversation_cache_blocks:
-            positions = positions[-self.max_conversation_cache_blocks:]
-            
+            positions = positions[-self.max_conversation_cache_blocks :]
+
         return positions
-    
+
     def get_conversation_cache_updates(self) -> dict:
         """
         Get cache position updates needed for the walking algorithm.
-        
+
         Returns:
             Dict with 'add', 'remove', and 'active' position lists (relative to full message array)
         """
         total_conversation_messages = len(self.history)
         new_positions = self._calculate_cache_positions(total_conversation_messages)
-        
+
         # Convert to absolute positions (including prompt messages)
         prompt_offset = len(self.prompt_messages)
         new_absolute_positions = [pos + prompt_offset for pos in new_positions]
-        
+
         old_positions_set = set(self.conversation_cache_positions)
         new_positions_set = set(new_absolute_positions)
-        
+
         return {
-            'add': sorted(new_positions_set - old_positions_set),
-            'remove': sorted(old_positions_set - new_positions_set), 
-            'active': sorted(new_absolute_positions)
+            "add": sorted(new_positions_set - old_positions_set),
+            "remove": sorted(old_positions_set - new_positions_set),
+            "active": sorted(new_absolute_positions),
         }
-    
+
     def apply_conversation_cache_updates(self, updates: dict) -> None:
         """
         Apply cache position updates.
-        
+
         Args:
             updates: Dict from get_conversation_cache_updates()
         """
-        self.conversation_cache_positions = updates['active'].copy()
+        self.conversation_cache_positions = updates["active"].copy()
 
-    def remove_cache_control_from_messages(self, messages: List[MessageParamT], positions: List[int]) -> None:
+    def remove_cache_control_from_messages(
+        self, messages: List[MessageParamT], positions: List[int]
+    ) -> None:
         """
         Remove cache control from specified message positions.
-        
+
         Args:
             messages: The message array to modify
             positions: List of positions to remove cache control from
@@ -196,14 +204,16 @@ class SimpleMemory(Memory, Generic[MessageParamT]):
                             if isinstance(content_block, dict) and "cache_control" in content_block:
                                 del content_block["cache_control"]
 
-    def add_cache_control_to_messages(self, messages: List[MessageParamT], positions: List[int]) -> int:
+    def add_cache_control_to_messages(
+        self, messages: List[MessageParamT], positions: List[int]
+    ) -> int:
         """
         Add cache control to specified message positions.
-        
+
         Args:
-            messages: The message array to modify  
+            messages: The message array to modify
             positions: List of positions to add cache control to
-            
+
         Returns:
             Number of cache blocks successfully applied
         """

{mcp_agent → fast_agent}/llm/model_database.py

@@ -158,8 +158,9 @@ class ModelDatabase:
         context_window=2097152, max_output_tokens=8192, tokenizes=GOOGLE_MULTIMODAL
     )
 
+    # 31/08/25 switched to object mode (even though groq says schema supported and used to work..)
     KIMI_MOONSHOT = ModelParameters(
-        context_window=131072, max_output_tokens=16384, tokenizes=TEXT_ONLY
+        context_window=262144, max_output_tokens=16384, tokenizes=TEXT_ONLY, json_mode="object"
     )
 
     # FIXME: xAI has not documented the max output tokens for Grok 4. Using Grok 3 as a placeholder. Will need to update when available (if ever)
@@ -245,7 +246,7 @@ class ModelDatabase:
         "grok-3-mini": GROK_3,
         "grok-3-fast": GROK_3,
         "grok-3-mini-fast": GROK_3,
-        "moonshotai/kimi-k2-instruct": KIMI_MOONSHOT,
+        "moonshotai/kimi-k2-instruct-0905": KIMI_MOONSHOT,
         "qwen/qwen3-32b": QWEN3_REASONER,
         "deepseek-r1-distill-llama-70b": DEEPSEEK_DISTILL,
         "openai/gpt-oss-120b": OPENAI_GPT_OSS_SERIES,
@@ -275,6 +276,38 @@ class ModelDatabase:
         params = cls.get_model_params(model)
         return params.tokenizes if params else None
 
+    @classmethod
+    def supports_mime(cls, model: str, mime_type: str) -> bool:
+        """
+        Return True if the given model supports the provided MIME type.
+
+        Normalizes common aliases (e.g., image/jpg->image/jpeg, document/pdf->application/pdf)
+        and also accepts bare extensions like "pdf" or "png".
+        """
+        from fast_agent.mcp.mime_utils import normalize_mime_type
+
+        tokenizes = cls.get_tokenizes(model) or []
+
+        # Normalize the candidate and the database entries to lowercase
+        normalized_supported = [t.lower() for t in tokenizes]
+
+        # Handle wildcard inputs like "image/*" quickly
+        mt = (mime_type or "").strip().lower()
+        if mt.endswith("/*") and "/" in mt:
+            prefix = mt.split("/", 1)[0] + "/"
+            return any(s.startswith(prefix) for s in normalized_supported)
+
+        normalized = normalize_mime_type(mime_type)
+        if not normalized:
+            return False
+
+        return normalized.lower() in normalized_supported
+
+    @classmethod
+    def supports_any_mime(cls, model: str, mime_types: List[str]) -> bool:
+        """Return True if the model supports any of the provided MIME types."""
+        return any(cls.supports_mime(model, m) for m in mime_types)
+
     @classmethod
     def get_json_mode(cls, model: str) -> str | None:
         """Get supported json mode (structured output) for a model"""