From 33258ce2ad2712d47c89aaac0cc6e319e74818be Mon Sep 17 00:00:00 2001
From: opencode <opencode@home.ms>
Date: Sun, 5 Jul 2026 07:03:20 +0000
Subject: [PATCH] fix: add sliding window to conversation history (issue #8)

- Add _trim_conversation_history() with message count + token limits
- Configurable via CLOVER_MAX_HISTORY_MESSAGES (default: 20)
- Configurable via CLOVER_MAX_HISTORY_TOKENS (default: 8000)
- Trims oldest messages first, preserves last 6 minimum
- Prevents OOM and degraded LLM quality on long sessions
---
 ai_agent.py        | 46 ++++++++++++++++++++++++++++++++++++++++++++++
 config/settings.py |  4 ++++
 2 files changed, 50 insertions(+)

diff --git a/ai_agent.py b/ai_agent.py
index 0cbda1c..05f7630 100644
--- a/ai_agent.py
+++ b/ai_agent.py
@@ -34,6 +34,8 @@ class AIAgent:
         self.model_manager = ModelManager()
         self.conversation_history = []
         self.available_tools = self._setup_tools()
+        self.max_history_messages = self.config.get("max_history_messages", 20)
+        self.max_history_tokens = self.config.get("max_history_tokens", 8000)
 
     def _setup_tools(self):
         """Setup available tools for the AI agent"""
@@ -267,6 +269,47 @@ When using tools, be methodical and explain each step. Always test your creation
         except Exception as e:
             return {"error": f"Tool execution failed: {str(e)}"}
 
+    def _trim_conversation_history(self):
+        """
+        Trim conversation history to prevent unbounded memory growth.
+        Uses a sliding window approach keeping the most recent messages.
+
+        Preserves at least the last N messages (max_history_messages)
+        and stays within token budget (max_history_tokens).
+        """
+        if len(self.conversation_history) <= self.max_history_messages:
+            return
+
+        # Rough token estimation: ~4 chars per token
+        def estimate_tokens(msg):
+            return len(msg.get("content", "")) // 4
+
+        # Calculate total tokens in history
+        total_tokens = sum(estimate_tokens(msg) for msg in self.conversation_history)
+
+        if total_tokens <= self.max_history_tokens:
+            # Within token budget, just apply message count limit
+            if len(self.conversation_history) > self.max_history_messages:
+                self.conversation_history = self.conversation_history[
+                    -self.max_history_messages :
+                ]
+            return
+
+        # Over token budget - trim from the front, keeping recent messages
+        while (
+            len(self.conversation_history) > 6
+            and sum(estimate_tokens(msg) for msg in self.conversation_history)
+            > self.max_history_tokens
+        ):
+            # Remove pairs of messages (user + assistant) from the front
+            self.conversation_history = self.conversation_history[2:]
+
+        # Also enforce message count limit
+        if len(self.conversation_history) > self.max_history_messages:
+            self.conversation_history = self.conversation_history[
+                -self.max_history_messages :
+            ]
+
     def chat(self, user_message: str) -> str:
         """
         Have a conversation with the user, using tools as needed
@@ -288,6 +331,9 @@ When using tools, be methodical and explain each step. Always test your creation
 
             # Prepare messages for AI
             messages = [{"role": "system", "content": self._create_system_prompt()}]
+
+            # Trim history to prevent unbounded memory growth
+            self._trim_conversation_history()
             messages.extend(self.conversation_history)
 
             # Get AI response
diff --git a/config/settings.py b/config/settings.py
index 2a14913..20077c1 100644
--- a/config/settings.py
+++ b/config/settings.py
@@ -43,6 +43,10 @@ def load_config():
         "project_root": os.getenv("CLOVER_PROJECT_ROOT", "."),
         "summary_file": os.getenv("CLOVER_SUMMARY_FILE", "clover.md"),
         "structure_file": os.getenv("CLOVER_STRUCTURE_FILE", "structure.md"),
+        "max_history_messages": int(
+            os.getenv("CLOVER_MAX_HISTORY_MESSAGES", "20")
+        ),
+        "max_history_tokens": int(os.getenv("CLOVER_MAX_HISTORY_TOKENS", "8000")),
     }
 
     # Debug output if enabled