Feature/compaction

pyproject.toml

@@ -17,7 +17,7 @@ dependencies = [
     "fastapi>=0.115.6",
     "mcp==1.12.4",
     "opentelemetry-distro>=0.55b0",
-    "opentelemetry-exporter-otlp-proto-http>=1.7.0",
+    "opentelemetry-exporter-otlp-proto-http>=1.34.1",
     "pydantic-settings>=2.7.0",
     "pydantic>=2.10.4",
     "pyyaml>=6.0.2",
@@ -38,7 +38,7 @@ dependencies = [
     "deprecated>=1.2.18",
     "a2a-sdk>=0.3.0",
     "email-validator>=2.2.0",
-    "pyperclip>=1.9.0",
+    "pyperclip>=1.9.0"
 ]
 
 # For Azure OpenAI with DefaultAzureCredential support, install with: pip install fast-agent-mcp[azure]
@@ -53,6 +53,7 @@ dev = [
     "ruff>=0.8.4",
     "tomli>=2.2.1",
     "pytest>=7.4.0",
+    "pytest-mock",
     "pytest-asyncio>=0.21.1",
     "pytest-cov",
 ]
@@ -103,7 +104,9 @@ dev = [
     "pytest>=7.4.0",
     "pytest-asyncio>=0.21.1",
     "pytest-cov>=6.1.1",
+    "pytest-mock",
     "ipdb>=0.13.13",
+    "trio>=0.30.0",
 ]
 
 [project.scripts]

scripts/event_viewer.py

@@ -173,7 +173,7 @@ def load_events(path: Path) -> List[Event]:
                         events.append(event)
                     except Exception as e:
                         print(f"Error on line {line_num}: {e}")
-                        print(f"Line content: {line.strip()}")
+                        print(f"Line content: {line.strip()[:500]}")
                         raise
     except Exception as e:
         print(f"Error loading file: {e}")

src/mcp_agent/core/agent_types.py

@@ -25,6 +25,14 @@ class AgentType(Enum):
     ITERATIVE_PLANNER = "iterative_planner"
 
 
+class ContextTruncationMode(str, Enum):
+    """Enumeration of supported context truncation strategies."""
+
+    NONE = None
+    SUMMARIZE = "summarize"
+    REMOVE = "remove"
+
+
 @dataclass
 class AgentConfig:
     """Configuration for an Agent instance"""
@@ -37,6 +45,10 @@ class AgentConfig:
     prompts: Optional[Dict[str, List[str]]] = None
     model: str | None = None
     use_history: bool = True
+
+    context_truncation_mode: Optional[ContextTruncationMode] = None
+    context_truncation_length_limit: Optional[int] = None
+
     default_request_params: RequestParams | None = None
     human_input: bool = False
     agent_type: AgentType = AgentType.BASIC

src/mcp_agent/core/enhanced_prompt.py

@@ -683,6 +683,8 @@ def pre_process_input(text):
                 return "SHOW_USAGE"
             elif cmd == "markdown":
                 return "MARKDOWN"
+            elif cmd == "compact":
+                return "COMPACT_HISTORY"
             elif cmd == "prompt":
                 # Handle /prompt with no arguments as interactive mode
                 if len(cmd_parts) > 1:
@@ -858,6 +860,7 @@ async def handle_special_commands(command, agent_app=None):
         rich_print("  /agents        - List available agents")
         rich_print("  /prompt <name> - Apply a specific prompt by name")
         rich_print("  /usage         - Show current usage statistics")
+        rich_print("  /compact       - Compact conversation history using summarization")
         rich_print("  /markdown      - Show last assistant message without markdown formatting")
         rich_print("  @agent_name    - Switch to agent")
         rich_print("  STOP           - Return control back to the workflow")
@@ -897,6 +900,10 @@ async def handle_special_commands(command, agent_app=None):
         # Return a dictionary to signal that markdown display should be shown
         return {"show_markdown": True}
 
+    elif command == "COMPACT_HISTORY":
+        # Return a dictionary to signal that history compaction should be performed
+        return {"compact_history": True}
+
     elif command == "SELECT_PROMPT" or (
         isinstance(command, str) and command.startswith("SELECT_PROMPT:")
     ):

src/mcp_agent/core/interactive_prompt.py

@@ -193,6 +193,10 @@ async def prompt_loop(
                         # Handle markdown display
                         await self._show_markdown(prompt_provider, agent)
                         continue
+                    elif "compact_history" in command_result:
+                        # Handle history compaction
+                        await self._compact_history(prompt_provider, agent)
+                        continue
 
                 # Skip further processing if:
                 # 1. The command was handled (command_result is truthy)
@@ -772,3 +776,103 @@ async def _show_markdown(self, prompt_provider: PromptProvider, agent_name: str)
 
         except Exception as e:
             rich_print(f"[red]Error showing markdown: {e}[/red]")
+
+    async def _compact_history(self, prompt_provider: PromptProvider, agent_name: str) -> None:
+        """
+        Compact the conversation history using summarization.
+
+        Args:
+            prompt_provider: Provider that has access to agents
+            agent_name: Name of the current agent
+        """
+        try:
+            # Get agent to compact history for
+            if hasattr(prompt_provider, "_agent"):
+                # This is an AgentApp - get the specific agent
+                agent = prompt_provider._agent(agent_name)
+            else:
+                # This is a single agent
+                agent = prompt_provider
+
+            # Check if agent has message history
+            if not hasattr(agent, "_llm") or not agent._llm:
+                rich_print("[yellow]No message history available to compact[/yellow]")
+                return
+
+            message_history = agent._llm.message_history
+            if not message_history or len(message_history) <= 2:
+                rich_print("[yellow]Insufficient message history to compact (need more than 2 messages)[/yellow]")
+                return
+
+            rich_print(f"[bold]Compacting conversation history for agent [cyan]{agent_name}[/cyan]...[/bold]")
+            rich_print(f"Current history: {len(message_history)} messages")
+
+            # Import the ContextTruncation class
+            from mcp_agent.core.agent_types import ContextTruncationMode
+            from mcp_agent.llm.context_truncation_and_summarization import ContextTruncation
+
+            # Message history is already in PromptMessageMultipart format
+            multipart_messages = message_history
+
+            # Apply summarization-based compaction
+            # Use a reasonable token limit - this could be made configurable
+            token_limit = 4000  # Conservative limit to trigger compaction
+            model_name = agent._llm.default_request_params.model or 'gpt-4'
+            system_prompt = ""  # System prompt is part of message history, not separate
+
+            compacted_messages = await ContextTruncation.truncate_if_required(
+                messages=multipart_messages,
+                truncation_mode=ContextTruncationMode.SUMMARIZE,
+                limit=token_limit,
+                model_name=model_name,
+                system_prompt=system_prompt,
+                provider=agent._llm
+            )
+
+            # Update both histories to keep them in sync
+
+            # 1. Update universal history (PromptMessageMultipart format)
+            agent._llm._message_history = compacted_messages
+
+            # 2. Update provider-specific history
+            # Need to convert PromptMessageMultipart to provider format and update provider history
+            # This ensures next API calls work correctly
+            try:
+                # Check if this is Anthropic provider
+                if hasattr(agent._llm, 'history') and 'anthropic' in str(type(agent._llm)).lower():
+                    from mcp_agent.llm.providers.multipart_converter_anthropic import (
+                        AnthropicConverter,
+                    )
+                    provider_messages = AnthropicConverter.convert_from_multipart_to_anthropic_list(compacted_messages)
+                    agent._llm.history.set(provider_messages)
+
+                # Check if this is OpenAI provider  
+                elif hasattr(agent._llm, 'history') and 'openai' in str(type(agent._llm)).lower():
+                    from mcp_agent.llm.providers.multipart_converter_openai import OpenAIConverter
+                    # Convert each message individually since there's no bulk converter
+                    provider_messages = []
+                    for msg in compacted_messages:
+                        openai_msg = OpenAIConverter.convert_to_openai(msg)
+                        provider_messages.append(openai_msg)
+                    agent._llm.history.set(provider_messages)
+
+                # Check if this is Google provider
+                elif hasattr(agent._llm, 'history') and 'google' in str(type(agent._llm)).lower():
+                    from mcp_agent.llm.providers.google_converter import GoogleConverter
+                    provider_messages = GoogleConverter.convert_to_google_content(compacted_messages)
+                    agent._llm.history.set(provider_messages)
+
+                # Add other providers as needed
+                else:
+                    rich_print("[yellow]Warning: Unknown provider type, only universal history updated[/yellow]")
+
+            except Exception as provider_error:
+                rich_print(f"[yellow]Warning: Could not update provider-specific history: {provider_error}[/yellow]")
+                rich_print("[yellow]Universal history was updated, but next API calls might have issues[/yellow]")
+
+            rich_print(f"[green]✓ History compacted from {len(message_history)} to {len(compacted_messages)} messages[/green]")
+
+        except Exception as e:
+            import traceback
+            rich_print(f"[red]Error compacting history: {e}[/red]")
+            rich_print(f"[dim]{traceback.format_exc()}[/dim]")

src/mcp_agent/core/request_params.py

@@ -2,7 +2,7 @@
 Request parameters definitions for LLM interactions.
 """
 
-from typing import Any, Dict, List
+from typing import Any, Dict, List, Optional
 
 from mcp import SamplingMessage
 from mcp.types import CreateMessageRequestParams
@@ -53,7 +53,22 @@ class RequestParams(CreateMessageRequestParams):
     Optional dictionary of template variables for dynamic templates. Currently only works for TensorZero inference backend
     """
 
+    context_truncation_mode: Optional[str] = None
+    """
+    The mode to use for context truncation: 'summarize', 'remove', or None
+    """
+
+    context_truncation_length_limit: Optional[int] = None
+    """
+    The token limit for context truncation. When exceeded, truncation will be applied.
+    """
+
+    request_delay_seconds: Optional[float] = None
+    """
+    Optional delay in seconds between requests to prevent rate limiting.
+    """
+
     mcp_metadata: Dict[str, Any] | None = None
     """
     Metadata to pass through to MCP tool calls via the _meta field.
-    """
+    """