feature: Memory Truncation (CG-10994) (#838)

Truncates conversation history + summarizes in two instances

1. Number of messages passed in to the LLM exceeds `max_messages` which
is default set to 100
2. The context limit for the LLM is approaching shortly (measured by
counting tokens and measuring if the current message input tokens >
`max_input_tokens_for_model - 10000`)

Algorithm -> similar to [OpenHands
summarizer](https://github.com/All-Hands-AI/OpenHands/blob/8043612420f0ef7d0c82ab5ab1c584385a8c57f3/openhands/memory/condenser/impl/llm_summarizing_condenser.py)
1. Keep the first two messages (System Prompt + Human Message) 
2. Truncate the number of messages to almost half
`(len(state[messages])) // 2)` and leave the last N messages untouched
3. Summarize the conversation in between the first two
(`keep_first_messages`) until the `tail`.

Misc
- We are able to pass in `agent_config` param to `CodeAgent` if we want
to play around with the max_messages, or `keep_first_messages` property
<img width="697" alt="image"
src="https://github.com/user-attachments/assets/d47450cf-f478-4bcd-b7cc-f8b397656870"
/>


Sample Trace (setting max_messages = 10) ->
https://smith.langchain.com/public/60c0eef2-4e46-4506-86d7-9acd48b01cbe/r

---------

Co-authored-by: jemeza-codegen <jmeza@codegen.com>

Author: tawsifkamal

src/codegen/agents/code_agent.py

@@ -5,6 +5,7 @@
 from langchain.tools import BaseTool
 from langchain_core.messages import AIMessage, HumanMessage
 from langchain_core.runnables.config import RunnableConfig
+from langgraph.graph.graph import CompiledGraph
 from langsmith import Client
 
 from codegen.extensions.langchain.agent import create_codebase_agent
@@ -15,16 +16,20 @@
 if TYPE_CHECKING:
     from codegen import Codebase
 
+from codegen.agents.utils import AgentConfig
+
 
 class CodeAgent:
     """Agent for interacting with a codebase."""
 
     codebase: "Codebase"
-    agent: any
+    agent: CompiledGraph
     langsmith_client: Client
     project_name: str
     thread_id: str | None = None
-    config: dict = {}
+    run_id: str | None = None
+    instance_id: str | None = None
+    difficulty: int | None = None
 
     def __init__(
         self,
@@ -35,6 +40,8 @@ def __init__(
         tools: Optional[list[BaseTool]] = None,
         tags: Optional[list[str]] = [],
         metadata: Optional[dict] = {},
+        agent_config: Optional[AgentConfig] = None,
+        thread_id: Optional[str] = None,
         **kwargs,
     ):
         """Initialize a CodeAgent.
@@ -60,15 +67,28 @@ def __init__(
             model_name=model_name,
             memory=memory,
             additional_tools=tools,
+            config=agent_config,
             **kwargs,
         )
         self.model_name = model_name
         self.langsmith_client = Client()
 
+        if thread_id is None:
+            self.thread_id = str(uuid4())
+        else:
+            self.thread_id = thread_id
+
         # Get project name from environment variable or use a default
         self.project_name = os.environ.get("LANGCHAIN_PROJECT", "RELACE")
         print(f"Using LangSmith project: {self.project_name}")
 
+        # Store SWEBench metadata if provided
+        self.run_id = metadata.get("run_id")
+        self.instance_id = metadata.get("instance_id")
+        # Extract difficulty value from "difficulty_X" format
+        difficulty_str = metadata.get("difficulty", "")
+        self.difficulty = int(difficulty_str.split("_")[1]) if difficulty_str and "_" in difficulty_str else None
+
         # Initialize tags for agent trace
         self.tags = [*tags, self.model_name]
 
@@ -79,7 +99,7 @@ def __init__(
             **metadata,
         }
 
-    def run(self, prompt: str, thread_id: Optional[str] = None) -> str:
+    def run(self, prompt: str) -> str:
         """Run the agent with a prompt.
 
         Args:
@@ -89,12 +109,9 @@ def run(self, prompt: str, thread_id: Optional[str] = None) -> str:
         Returns:
             The agent's response
         """
-        if thread_id is None:
-            thread_id = str(uuid4())
-        self.thread_id = thread_id
         self.config = {
             "configurable": {
-                "thread_id": thread_id,
+                "thread_id": self.thread_id,
                 "metadata": {"project": self.project_name},
             },
             "recursion_limit": 100,
@@ -104,15 +121,15 @@ def run(self, prompt: str, thread_id: Optional[str] = None) -> str:
         # see more https://langchain-ai.github.io/langgraph/concepts/low_level/#reducers
         input = {"query": prompt}
 
-        config = RunnableConfig(configurable={"thread_id": thread_id}, tags=self.tags, metadata=self.metadata, recursion_limit=100)
+        config = RunnableConfig(configurable={"thread_id": self.thread_id}, tags=self.tags, metadata=self.metadata, recursion_limit=200)
         # we stream the steps instead of invoke because it allows us to access intermediate nodes
         stream = self.agent.stream(input, config=config, stream_mode="values")
 
         # Keep track of run IDs from the stream
         run_ids = []
 
         for s in stream:
-            if len(s["messages"]) == 0:
+            if len(s["messages"]) == 0 or isinstance(s["messages"][-1], HumanMessage):
                 message = HumanMessage(content=prompt)
             else:
                 message = s["messages"][-1]

src/codegen/agents/utils.py

@@ -0,0 +1,8 @@
+from typing import TypedDict
+
+
+class AgentConfig(TypedDict, total=False):
+    """Configuration options for the CodeAgent."""
+
+    keep_first_messages: int  # Number of initial messages to keep during summarization
+    max_messages: int  # Maximum number of messages before triggering summarization

src/codegen/extensions/langchain/agent.py

@@ -1,12 +1,13 @@
 """Demo implementation of an agent with Codegen tools."""
 
-from typing import TYPE_CHECKING, Optional
+from typing import TYPE_CHECKING, Any, Optional
 
 from langchain.tools import BaseTool
 from langchain_core.messages import SystemMessage
 from langgraph.checkpoint.memory import MemorySaver
 from langgraph.graph.graph import CompiledGraph
 
+from codegen.agents.utils import AgentConfig
 from codegen.extensions.langchain.llm import LLM
 from codegen.extensions.langchain.prompts import REASONER_SYSTEM_MESSAGE
 from codegen.extensions.langchain.tools import (
@@ -38,6 +39,7 @@ def create_codebase_agent(
     memory: bool = True,
     debug: bool = False,
     additional_tools: Optional[list[BaseTool]] = None,
+    config: Optional[AgentConfig] = None,
     **kwargs,
 ) -> CompiledGraph:
     """Create an agent with all codebase tools.
@@ -57,7 +59,7 @@ def create_codebase_agent(
     Returns:
         Initialized agent with message history
     """
-    llm = LLM(model_provider=model_provider, model_name=model_name, max_tokens=8192, **kwargs)
+    llm = LLM(model_provider=model_provider, model_name=model_name, **kwargs)
 
     # Get all codebase tools
     tools = [
@@ -89,7 +91,7 @@ def create_codebase_agent(
 
     memory = MemorySaver() if memory else None
 
-    return create_react_agent(model=llm, tools=tools, system_message=system_message, checkpointer=memory, debug=debug)
+    return create_react_agent(model=llm, tools=tools, system_message=system_message, checkpointer=memory, debug=debug, config=config)
 
 
 def create_chat_agent(
@@ -100,6 +102,7 @@ def create_chat_agent(
     memory: bool = True,
     debug: bool = False,
     additional_tools: Optional[list[BaseTool]] = None,
+    config: Optional[dict[str, Any]] = None,  # over here you can pass in the max length of the number of messages
     **kwargs,
 ) -> CompiledGraph:
     """Create an agent with all codebase tools.
@@ -138,7 +141,7 @@ def create_chat_agent(
 
     memory = MemorySaver() if memory else None
 
-    return create_react_agent(model=llm, tools=tools, system_message=system_message, checkpointer=memory, debug=debug)
+    return create_react_agent(model=llm, tools=tools, system_message=system_message, checkpointer=memory, debug=debug, config=config)
 
 
 def create_codebase_inspector_agent(
@@ -148,6 +151,7 @@ def create_codebase_inspector_agent(
     system_message: SystemMessage = SystemMessage(REASONER_SYSTEM_MESSAGE),
     memory: bool = True,
     debug: bool = True,
+    config: Optional[dict[str, Any]] = None,
     **kwargs,
 ) -> CompiledGraph:
     """Create an inspector agent with read-only codebase tools.
@@ -175,7 +179,7 @@ def create_codebase_inspector_agent(
     ]
 
     memory = MemorySaver() if memory else None
-    return create_react_agent(model=llm, tools=tools, system_message=system_message, checkpointer=memory, debug=debug)
+    return create_react_agent(model=llm, tools=tools, system_message=system_message, checkpointer=memory, debug=debug, config=config)
 
 
 def create_agent_with_tools(
@@ -185,6 +189,7 @@ def create_agent_with_tools(
     system_message: SystemMessage = SystemMessage(REASONER_SYSTEM_MESSAGE),
     memory: bool = True,
     debug: bool = True,
+    config: Optional[dict[str, Any]] = None,
     **kwargs,
 ) -> CompiledGraph:
     """Create an agent with a specific set of tools.
@@ -209,4 +214,4 @@ def create_agent_with_tools(
 
     memory = MemorySaver() if memory else None
 
-    return create_react_agent(model=llm, tools=tools, system_message=system_message, checkpointer=memory, debug=debug)
+    return create_react_agent(model=llm, tools=tools, system_message=system_message, checkpointer=memory, debug=debug, config=config)