deepcode-hku 1.0.1__py3-none-any.whl

workflows/agents/memory_agent_concise.py

@@ -0,0 +1,923 @@
+"""
+Concise Memory Agent for Code Implementation Workflow
+简洁的代码实现工作流内存代理
+
+This memory agent implements a focused approach:
+1. Before first file: Normal conversation flow
+2. After first file: Keep only system_prompt + initial_plan + current round tool results
+3. Clean slate for each new code file generation
+
+Key Features:
+- Preserves system prompt and initial plan always
+- After first file generation, discards previous conversation history
+- Keeps only current round tool results from essential tools:
+  * read_code_mem, read_file, write_file
+  * execute_python, execute_bash
+  * search_code, search_reference_code, get_file_structure
+- Provides clean, focused input for next write_file operation
+"""
+
+import json
+import logging
+import os
+import time
+from datetime import datetime
+from typing import Dict, Any, List, Optional
+
+
+class ConciseMemoryAgent:
+    """
+    Concise Memory Agent - Focused Information Retention
+
+    Core Philosophy:
+    - Preserve essential context (system prompt + initial plan)
+    - After first file generation, use clean slate approach
+    - Keep only current round tool results from all essential MCP tools
+    - Remove conversational clutter and previous tool calls
+
+    Essential Tools Tracked:
+    - File Operations: read_code_mem, read_file, write_file
+    - Code Analysis: search_code, search_reference_code, get_file_structure
+    - Execution: execute_python, execute_bash
+    """
+
+    def __init__(
+        self,
+        initial_plan_content: str,
+        logger: Optional[logging.Logger] = None,
+        target_directory: Optional[str] = None,
+        default_models: Optional[Dict[str, str]] = None,
+    ):
+        """
+        Initialize Concise Memory Agent
+
+        Args:
+            initial_plan_content: Content of initial_plan.txt
+            logger: Logger instance
+            target_directory: Target directory for saving summaries
+            default_models: Default models configuration from workflow
+        """
+        self.logger = logger or self._create_default_logger()
+        self.initial_plan = initial_plan_content
+
+        # Store default models configuration
+        self.default_models = default_models or {
+            "anthropic": "claude-sonnet-4-20250514",
+            "openai": "gpt-4o",
+        }
+
+        # Memory state tracking - new logic: trigger after each write_file
+        self.last_write_file_detected = (
+            False  # Track if write_file was called in current iteration
+        )
+        self.should_clear_memory_next = False  # Flag to clear memory in next round
+        self.current_round = 0
+
+        # Parse phase structure from initial plan
+        self.phase_structure = self._parse_phase_structure()
+
+        # Memory configuration
+        if target_directory:
+            self.save_path = target_directory
+        else:
+            self.save_path = "./deepcode_lab/papers/1/"
+
+        # Code summary file path
+        self.code_summary_path = os.path.join(
+            self.save_path, "implement_code_summary.md"
+        )
+
+        # Current round tool results storage
+        self.current_round_tool_results = []
+
+        # Track all implemented files
+        self.implemented_files = []
+
+        self.logger.info(
+            f"Concise Memory Agent initialized with target directory: {self.save_path}"
+        )
+        self.logger.info(f"Code summary will be saved to: {self.code_summary_path}")
+        # self.logger.info(f"🤖 Using models - Anthropic: {self.default_models['anthropic']}, OpenAI: {self.default_models['openai']}")
+        self.logger.info(
+            "📝 NEW LOGIC: Memory clearing triggered after each write_file call"
+        )
+
+    def _create_default_logger(self) -> logging.Logger:
+        """Create default logger"""
+        logger = logging.getLogger(f"{__name__}.ConciseMemoryAgent")
+        logger.setLevel(logging.INFO)
+        return logger
+
+    def _parse_phase_structure(self) -> Dict[str, List[str]]:
+        """Parse implementation phases from initial plan"""
+        try:
+            phases = {}
+            lines = self.initial_plan.split("\n")
+            current_phase = None
+
+            for line in lines:
+                if "Phase" in line and ":" in line:
+                    # Extract phase name
+                    phase_parts = line.split(":")
+                    if len(phase_parts) >= 2:
+                        current_phase = phase_parts[0].strip()
+                        phases[current_phase] = []
+                elif current_phase and line.strip().startswith("-"):
+                    # This is a file in the current phase
+                    file_line = line.strip()[1:].strip()
+                    if file_line.startswith("`") and file_line.endswith("`"):
+                        file_name = file_line[1:-1]
+                        phases[current_phase].append(file_name)
+                elif current_phase and not line.strip():
+                    # Empty line might indicate end of phase
+                    continue
+                elif current_phase and line.strip().startswith("###"):
+                    # New section, end current phase
+                    current_phase = None
+
+            return phases
+
+        except Exception as e:
+            self.logger.warning(f"Failed to parse phase structure: {e}")
+            return {}
+
+    def record_file_implementation(
+        self, file_path: str, implementation_content: str = ""
+    ):
+        """
+        Record a newly implemented file (simplified version)
+        NEW LOGIC: File implementation is tracked via write_file tool detection
+
+        Args:
+            file_path: Path of the implemented file
+            implementation_content: Content of the implemented file
+        """
+        # Add file to implemented files list if not already present
+        if file_path not in self.implemented_files:
+            self.implemented_files.append(file_path)
+
+        self.logger.info(f"📝 File implementation recorded: {file_path}")
+
+    async def create_code_implementation_summary(
+        self,
+        client,
+        client_type: str,
+        file_path: str,
+        implementation_content: str,
+        files_implemented: int,
+    ) -> str:
+        """
+        Create LLM-based code implementation summary after writing a file
+        Uses LLM to analyze and summarize the implemented code
+
+        Args:
+            client: LLM client instance
+            client_type: Type of LLM client ("anthropic" or "openai")
+            file_path: Path of the implemented file
+            implementation_content: Content of the implemented file
+            files_implemented: Number of files implemented so far
+
+        Returns:
+            LLM-generated formatted code implementation summary
+        """
+        try:
+            # Record the file implementation first
+            self.record_file_implementation(file_path, implementation_content)
+
+            # Create prompt for LLM summary
+            summary_prompt = self._create_code_summary_prompt(
+                file_path, implementation_content, files_implemented
+            )
+            summary_messages = [{"role": "user", "content": summary_prompt}]
+
+            # Get LLM-generated summary
+            llm_response = await self._call_llm_for_summary(
+                client, client_type, summary_messages
+            )
+            llm_summary = llm_response.get("content", "")
+
+            # Format the summary in the requested structure
+            formatted_summary = self._format_code_implementation_summary(
+                file_path, llm_summary, files_implemented
+            )
+
+            # Save to implement_code_summary.md (append mode)
+            await self._save_code_summary_to_file(formatted_summary, file_path)
+
+            self.logger.info(f"Created and saved code summary for: {file_path}")
+            return formatted_summary
+
+        except Exception as e:
+            self.logger.error(
+                f"Failed to create LLM-based code implementation summary: {e}"
+            )
+            # Fallback to simple summary
+            return self._create_fallback_code_summary(
+                file_path, implementation_content, files_implemented
+            )
+
+    def _create_code_summary_prompt(
+        self, file_path: str, implementation_content: str, files_implemented: int
+    ) -> str:
+        """
+        Create prompt for LLM to generate code implementation summary
+
+        Args:
+            file_path: Path of the implemented file
+            implementation_content: Content of the implemented file
+            files_implemented: Number of files implemented so far
+
+        Returns:
+            Prompt for LLM summarization
+        """
+        current_round = self.current_round
+
+        # Create formatted list of implemented files
+        implemented_files_list = (
+            "\n".join([f"- {file}" for file in self.implemented_files])
+            if self.implemented_files
+            else "- None yet"
+        )
+
+        prompt = f"""You are an expert code implementation summarizer. Analyze the implemented code file and create a structured summary.
+
+**🚨 CRITICAL: The files listed below are ALREADY IMPLEMENTED - DO NOT suggest them in Next Steps! 🚨**
+
+**All Previously Implemented Files:**
+{implemented_files_list}
+
+**Current Implementation Context:**
+- **File Implemented**: {file_path}
+- **Current Round**: {current_round}
+- **Total Files Implemented**: {files_implemented}
+
+
+**Initial Plan Reference:**
+{self.initial_plan[:]}
+
+**Implemented Code Content:**
+```
+{implementation_content[:]}
+```
+
+**Required Summary Format:**
+
+1. **Status Marker**: Mark the phase and round corresponding to this code file
+   Format: Phase {{phase_name}}, Round {{round_number}}
+
+2. **Implementation Progress**: List the code file completed in current round and core implementation ideas
+   Format: {{file_path}}: {{core implementation ideas}}
+
+3. **Dependencies**: According to the File Structure and initial plan, list functions that may be called by other files
+   Format: {{file_path}}: Function {{function_name}}: core ideas--{{ideas}}; Required parameters--{{params}}; Return parameters--{{returns}}
+   Required packages: {{packages}}
+
+4. **Next Steps**: List code files that will be implemented in the next round (EXCLUDE already implemented files)
+   Format: Code will be implemented: {{file_path}}; will stay on Phase {{phase}}/ will go to Phase {{next_phase}}
+   **WARNING: Do NOT suggest any file from the "All Previously Implemented Files" list above!**
+
+**Instructions:**
+- Be precise and concise
+- Focus on function interfaces that other files will need
+- Extract actual function signatures from the code
+- **CRITICAL: For Next Steps, ONLY suggest files that are NOT in the "All Previously Implemented Files" list above**
+- **NEVER suggest implementing a file that is already in the implemented files list**
+- Predict next implementation steps based on the initial plan but exclude already completed files
+- Use the exact format specified above
+
+**Summary:**"""
+
+        return prompt
+
+    def _format_code_implementation_summary(
+        self, file_path: str, llm_summary: str, files_implemented: int
+    ) -> str:
+        """
+        Format the LLM-generated summary into the final structure
+
+        Args:
+            file_path: Path of the implemented file
+            llm_summary: LLM-generated summary content
+            files_implemented: Number of files implemented so far
+
+        Returns:
+            Formatted summary
+        """
+        timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+
+        # Create formatted list of implemented files
+        implemented_files_list = (
+            "\n".join([f"- {file}" for file in self.implemented_files])
+            if self.implemented_files
+            else "- None yet"
+        )
+
+        formatted_summary = f"""# Code Implementation Summary
+**All Previously Implemented Files:**
+{implemented_files_list}
+**Generated**: {timestamp}
+**File Implemented**: {file_path}
+**Total Files Implemented**: {files_implemented}
+
+{llm_summary}
+
+---
+*Auto-generated by Memory Agent*
+"""
+        return formatted_summary
+
+    def _create_fallback_code_summary(
+        self, file_path: str, implementation_content: str, files_implemented: int
+    ) -> str:
+        """
+        Create fallback summary when LLM is unavailable
+
+        Args:
+            file_path: Path of the implemented file
+            implementation_content: Content of the implemented file
+            files_implemented: Number of files implemented so far
+
+        Returns:
+            Fallback summary
+        """
+        # Create formatted list of implemented files
+        implemented_files_list = (
+            "\n".join([f"- {file}" for file in self.implemented_files])
+            if self.implemented_files
+            else "- None yet"
+        )
+
+        summary = f"""# Code Implementation Summary
+**All Previously Implemented Files:**
+{implemented_files_list}
+**Generated**: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}
+**File Implemented**: {file_path}
+**Total Files Implemented**: {files_implemented}
+**Summary failed to generate.**
+
+---
+*Auto-generated by Concise Memory Agent (Fallback Mode)*
+"""
+        return summary
+
+    async def _save_code_summary_to_file(self, new_summary: str, file_path: str):
+        """
+        Append code implementation summary to implement_code_summary.md
+        Accumulates all implementations with clear separators
+
+        Args:
+            new_summary: New summary content to append
+            file_path: Path of the file for which the summary was generated
+        """
+        try:
+            # Create directory if it doesn't exist
+            os.makedirs(os.path.dirname(self.code_summary_path), exist_ok=True)
+
+            # Check if file exists to determine if we need header
+            file_exists = os.path.exists(self.code_summary_path)
+
+            # Open in append mode to accumulate all implementations
+            with open(self.code_summary_path, "a", encoding="utf-8") as f:
+                if not file_exists:
+                    # Write header for new file
+                    f.write("# Code Implementation Progress Summary\n")
+                    f.write("*Accumulated implementation progress for all files*\n\n")
+
+                # Add clear separator between implementations
+                f.write("\n" + "=" * 80 + "\n")
+                f.write(
+                    f"## IMPLEMENTATION File {file_path}; ROUND {self.current_round} \n"
+                )
+                f.write("=" * 80 + "\n\n")
+
+                # Write the new summary
+                f.write(new_summary)
+                f.write("\n\n")
+
+            self.logger.info(
+                f"Appended LLM-based code implementation summary to: {self.code_summary_path}"
+            )
+
+        except Exception as e:
+            self.logger.error(f"Failed to save code implementation summary: {e}")
+
+    async def _call_llm_for_summary(
+        self, client, client_type: str, summary_messages: List[Dict]
+    ) -> Dict[str, Any]:
+        """
+        Call LLM for code implementation summary generation ONLY
+        调用LLM生成代码实现总结（仅用于代码总结）
+
+        This method is used only for creating code implementation summaries,
+        NOT for conversation summarization which has been removed.
+        """
+        if client_type == "anthropic":
+            response = await client.messages.create(
+                model=self.default_models["anthropic"],
+                system="You are an expert code implementation summarizer. Create structured summaries of implemented code files that preserve essential information about functions, dependencies, and implementation approaches.",
+                messages=summary_messages,
+                max_tokens=5000,
+                temperature=0.2,
+            )
+
+            content = ""
+            for block in response.content:
+                if block.type == "text":
+                    content += block.text
+
+            return {"content": content}
+
+        elif client_type == "openai":
+            openai_messages = [
+                {
+                    "role": "system",
+                    "content": "You are an expert code implementation summarizer. Create structured summaries of implemented code files that preserve essential information about functions, dependencies, and implementation approaches.",
+                }
+            ]
+            openai_messages.extend(summary_messages)
+
+            response = await client.chat.completions.create(
+                model=self.default_models["openai"],
+                messages=openai_messages,
+                max_tokens=5000,
+                temperature=0.2,
+            )
+
+            return {"content": response.choices[0].message.content or ""}
+
+        else:
+            raise ValueError(f"Unsupported client type: {client_type}")
+
+    def start_new_round(self, iteration: Optional[int] = None):
+        """Start a new dialogue round and reset tool results
+
+        Args:
+            iteration: Optional iteration number from workflow to sync with current_round
+        """
+        if iteration is not None:
+            # Sync with workflow iteration
+            self.current_round = iteration
+            # self.logger.info(f"🔄 Synced round with workflow iteration {iteration}")
+        else:
+            # Default behavior: increment round counter
+            self.current_round += 1
+            self.logger.info(f"🔄 Started new round {self.current_round}")
+
+        self.current_round_tool_results = []  # Clear previous round results
+        # Note: Don't reset last_write_file_detected and should_clear_memory_next here
+        # These flags persist across rounds until memory optimization is applied
+        # self.logger.info(f"🔄 Round {self.current_round} - Tool results cleared, memory flags preserved")
+
+    def record_tool_result(
+        self, tool_name: str, tool_input: Dict[str, Any], tool_result: Any
+    ):
+        """
+        Record tool result for current round and detect write_file calls
+
+        Args:
+            tool_name: Name of the tool called
+            tool_input: Input parameters for the tool
+            tool_result: Result returned by the tool
+        """
+        # Detect write_file calls to trigger memory clearing
+        if tool_name == "write_file":
+            self.last_write_file_detected = True
+            self.should_clear_memory_next = True
+
+            # self.logger.info(f"🔄 WRITE_FILE DETECTED: {file_path} - Memory will be cleared in next round")
+
+        # Only record specific tools that provide essential information
+        essential_tools = [
+            "read_code_mem",  # Read code summary from implement_code_summary.md
+            "read_file",  # Read file contents
+            "write_file",  # Write file contents (important for tracking implementations)
+            "execute_python",  # Execute Python code (for testing/validation)
+            "execute_bash",  # Execute bash commands (for build/execution)
+            "search_code",  # Search code patterns
+            "search_reference_code",  # Search reference code (if available)
+            "get_file_structure",  # Get file structure (for understanding project layout)
+        ]
+
+        if tool_name in essential_tools:
+            tool_record = {
+                "tool_name": tool_name,
+                "tool_input": tool_input,
+                "tool_result": tool_result,
+                "timestamp": time.time(),
+            }
+            self.current_round_tool_results.append(tool_record)
+            # self.logger.info(f"📊 Essential tool result recorded: {tool_name} ({len(self.current_round_tool_results)} total)")
+
+    def should_use_concise_mode(self) -> bool:
+        """
+        Check if concise memory mode should be used
+
+        Returns:
+            True if first file has been generated and concise mode should be active
+        """
+        return self.last_write_file_detected
+
+    def create_concise_messages(
+        self, system_prompt: str, messages: List[Dict[str, Any]], files_implemented: int
+    ) -> List[Dict[str, Any]]:
+        """
+        Create concise message list for LLM input
+        NEW LOGIC: Always clear after write_file, keep system_prompt + initial_plan + current round tools
+
+        Args:
+            system_prompt: Current system prompt
+            messages: Original message list
+            files_implemented: Number of files implemented so far
+
+        Returns:
+            Concise message list containing only essential information
+        """
+        if not self.last_write_file_detected:
+            # Before any write_file, use normal flow
+            self.logger.info(
+                "🔄 Using normal conversation flow (before any write_file)"
+            )
+            return messages
+
+        # After write_file detection, use concise approach with clean slate
+        self.logger.info(
+            f"🎯 Using CONCISE memory mode - Clear slate after write_file, Round {self.current_round}"
+        )
+
+        concise_messages = []
+
+        # 1. Add initial plan message (always preserved)
+        initial_plan_message = {
+            "role": "user",
+            "content": f"""**Task: Implement code based on the following reproduction plan**
+
+**Code Reproduction Plan:**
+{self.initial_plan}
+
+**Working Directory:** Current workspace
+
+**Current Status:** {files_implemented} files implemented
+
+**Objective:** Continue implementation by analyzing dependencies and implementing the next required file according to the plan's priority order.""",
+        }
+        concise_messages.append(initial_plan_message)
+
+        # 2. Add Knowledge Base
+        knowledge_base_message = {
+            "role": "user",
+            "content": f"""**Below is the Knowledge Base of the LATEST implemented code file:**
+{self._read_code_knowledge_base()}
+
+**Development Cycle - START HERE:**
+
+**For NEW file implementation:**
+1. **You need to call read_code_mem(already_implemented_file_path)** to understand existing implementations and dependencies - agent should choose relevant ALREADY IMPLEMENTED file paths for reference, NOT the new file you want to create
+2. Write_file can be used to implement the new component
+3. Finally: Use execute_python or execute_bash for testing (if needed)
+
+**When all files implemented:**
+**Use execute_python or execute_bash** to test the complete implementation""",
+        }
+        concise_messages.append(knowledge_base_message)
+
+        # 3. Add current tool results (essential information for next file generation)
+        if self.current_round_tool_results:
+            tool_results_content = self._format_tool_results()
+            tool_results_message = {
+                "role": "user",
+                "content": f"""**Current Tool Results:**
+{tool_results_content}""",
+            }
+            concise_messages.append(tool_results_message)
+        else:
+            # If no tool results yet, add guidance for next steps
+            guidance_message = {
+                "role": "user",
+                "content": f"""**Current Round:** {self.current_round}
+
+**Development Cycle - START HERE:**
+
+**For NEW file implementation:**
+1. **You need to call read_code_mem(already_implemented_file_path)** to understand existing implementations and dependencies - agent should choose relevant ALREADY IMPLEMENTED file paths for reference, NOT the new file you want to create
+2. Write_file can be used to implement the new component
+3. Finally: Use execute_python or execute_bash for testing (if needed)
+
+**When all files implemented:**
+1. **Use execute_python or execute_bash** to test the complete implementation""",
+            }
+            concise_messages.append(guidance_message)
+        # **Available Essential Tools:** read_code_mem, write_file, execute_python, execute_bash
+        # **Remember:** Start with read_code_mem when implementing NEW files to understand existing code. When all files are implemented, focus on testing and completion. Implement according to the original paper's specifications - any reference code is for inspiration only."""
+        # self.logger.info(f"✅ Concise messages created: {len(concise_messages)} messages (original: {len(messages)})")
+        return concise_messages
+
+    def _read_code_knowledge_base(self) -> Optional[str]:
+        """
+        Read the implement_code_summary.md file as code knowledge base
+        Returns only the final/latest implementation entry, not all historical entries
+
+        Returns:
+            Content of the latest implementation entry if it exists, None otherwise
+        """
+        try:
+            if os.path.exists(self.code_summary_path):
+                with open(self.code_summary_path, "r", encoding="utf-8") as f:
+                    content = f.read().strip()
+
+                if content:
+                    # Extract only the final/latest implementation entry
+                    return self._extract_latest_implementation_entry(content)
+                else:
+                    return None
+            else:
+                return None
+
+        except Exception as e:
+            self.logger.error(f"Failed to read code knowledge base: {e}")
+            return None
+
+    def _extract_latest_implementation_entry(self, content: str) -> Optional[str]:
+        """
+        Extract the latest/final implementation entry from the implement_code_summary.md content
+        Uses a simpler approach to find the last implementation section
+
+        Args:
+            content: Full content of implement_code_summary.md
+
+        Returns:
+            Latest implementation entry content, or None if not found
+        """
+        try:
+            import re
+
+            # Pattern to match the start of implementation sections
+            section_pattern = (
+                r"={80}\s*\n## IMPLEMENTATION File .+?; ROUND \d+\s*\n={80}"
+            )
+
+            # Find all implementation section starts
+            matches = list(re.finditer(section_pattern, content))
+
+            if not matches:
+                # No implementation sections found
+                lines = content.split("\n")
+                fallback_content = (
+                    "\n".join(lines[:10]) + "\n... (truncated for brevity)"
+                    if len(lines) > 10
+                    else content
+                )
+                self.logger.info(
+                    "📖 No implementation sections found, using fallback content"
+                )
+                return fallback_content
+
+            # Get the start position of the last implementation section
+            last_match = matches[-1]
+            start_pos = last_match.start()
+
+            # Take everything from the last section start to the end of content
+            latest_entry = content[start_pos:].strip()
+
+            # self.logger.info(f"📖 Extracted latest implementation entry from knowledge base")
+            # print(f"DEBUG: Extracted content length: {len(latest_entry)}")
+            # print(f"DEBUG: First 200 chars: {latest_entry[:]}")
+
+            return latest_entry
+
+        except Exception as e:
+            self.logger.error(f"Failed to extract latest implementation entry: {e}")
+            # Return last 1000 characters as fallback
+            return content[-500:] if len(content) > 500 else content
+
+    def _format_tool_results(self) -> str:
+        """
+        Format current round tool results for LLM input
+
+        Returns:
+            Formatted string of tool results
+        """
+        if not self.current_round_tool_results:
+            return "No tool results in current round."
+
+        formatted_results = []
+
+        for result in self.current_round_tool_results:
+            tool_name = result["tool_name"]
+            tool_input = result["tool_input"]
+            tool_result = result["tool_result"]
+
+            # Format based on tool type
+            if tool_name == "read_code_mem":
+                file_path = tool_input.get("file_path", "unknown")
+                formatted_results.append(f"""
+**read_code_mem Result for {file_path}:**
+{self._format_tool_result_content(tool_result)}
+""")
+            elif tool_name == "read_file":
+                file_path = tool_input.get("file_path", "unknown")
+                formatted_results.append(f"""
+**read_file Result for {file_path}:**
+{self._format_tool_result_content(tool_result)}
+""")
+            elif tool_name == "write_file":
+                file_path = tool_input.get("file_path", "unknown")
+                formatted_results.append(f"""
+**write_file Result for {file_path}:**
+{self._format_tool_result_content(tool_result)}
+""")
+            elif tool_name == "execute_python":
+                code_snippet = (
+                    tool_input.get("code", "")[:50] + "..."
+                    if len(tool_input.get("code", "")) > 50
+                    else tool_input.get("code", "")
+                )
+                formatted_results.append(f"""
+**execute_python Result (code: {code_snippet}):**
+{self._format_tool_result_content(tool_result)}
+""")
+            elif tool_name == "execute_bash":
+                command = tool_input.get("command", "unknown")
+                formatted_results.append(f"""
+**execute_bash Result (command: {command}):**
+{self._format_tool_result_content(tool_result)}
+""")
+            elif tool_name == "search_code":
+                pattern = tool_input.get("pattern", "unknown")
+                file_pattern = tool_input.get("file_pattern", "")
+                formatted_results.append(f"""
+**search_code Result (pattern: {pattern}, files: {file_pattern}):**
+{self._format_tool_result_content(tool_result)}
+""")
+            elif tool_name == "search_reference_code":
+                target_file = tool_input.get("target_file", "unknown")
+                keywords = tool_input.get("keywords", "")
+                formatted_results.append(f"""
+**search_reference_code Result for {target_file} (keywords: {keywords}):**
+{self._format_tool_result_content(tool_result)}
+""")
+            elif tool_name == "get_file_structure":
+                directory = tool_input.get(
+                    "directory_path", tool_input.get("path", "current")
+                )
+                formatted_results.append(f"""
+**get_file_structure Result for {directory}:**
+{self._format_tool_result_content(tool_result)}
+""")
+
+        return "\n".join(formatted_results)
+
+    def _format_tool_result_content(self, tool_result: Any) -> str:
+        """
+        Format tool result content for display
+
+        Args:
+            tool_result: Tool result to format
+
+        Returns:
+            Formatted string representation
+        """
+        if isinstance(tool_result, str):
+            # Try to parse as JSON for better formatting
+            try:
+                result_data = json.loads(tool_result)
+                if isinstance(result_data, dict):
+                    # Format key information
+                    if result_data.get("status") == "summary_found":
+                        return (
+                            f"Summary found:\n{result_data.get('summary_content', '')}"
+                        )
+                    elif result_data.get("status") == "no_summary":
+                        return "No summary available"
+                    else:
+                        return json.dumps(result_data, indent=2)
+                else:
+                    return str(result_data)
+            except json.JSONDecodeError:
+                return tool_result
+        else:
+            return str(tool_result)
+
+    def get_memory_statistics(self, files_implemented: int = 0) -> Dict[str, Any]:
+        """Get memory agent statistics"""
+        return {
+            "last_write_file_detected": self.last_write_file_detected,
+            "should_clear_memory_next": self.should_clear_memory_next,
+            "current_round": self.current_round,
+            "concise_mode_active": self.should_use_concise_mode(),
+            "current_round_tool_results": len(self.current_round_tool_results),
+            "essential_tools_recorded": [
+                r["tool_name"] for r in self.current_round_tool_results
+            ],
+            "implemented_files_tracked": files_implemented,
+            "implemented_files_list": self.implemented_files.copy(),
+            "phases_parsed": len(self.phase_structure),
+        }
+
+    def get_implemented_files(self) -> List[str]:
+        """Get list of all implemented files"""
+        return self.implemented_files.copy()
+
+    def should_trigger_memory_optimization(
+        self, messages: List[Dict[str, Any]], files_implemented: int = 0
+    ) -> bool:
+        """
+        Check if memory optimization should be triggered
+        NEW LOGIC: Trigger after write_file has been detected
+
+        Args:
+            messages: Current message list
+            files_implemented: Number of files implemented so far
+
+        Returns:
+            True if concise mode should be applied
+        """
+        # Trigger if we detected write_file and should clear memory
+        if self.should_clear_memory_next:
+            # self.logger.info(f"🎯 Triggering CONCISE memory optimization (write_file detected, files: {files_implemented})")
+            return True
+
+        # No optimization before any write_file
+        return False
+
+    def apply_memory_optimization(
+        self, system_prompt: str, messages: List[Dict[str, Any]], files_implemented: int
+    ) -> List[Dict[str, Any]]:
+        """
+        Apply memory optimization using concise approach
+        NEW LOGIC: Clear all history after write_file, keep only system_prompt + initial_plan + current tools
+
+        Args:
+            system_prompt: Current system prompt
+            messages: Original message list
+            files_implemented: Number of files implemented so far
+
+        Returns:
+            Optimized message list
+        """
+        if not self.should_clear_memory_next:
+            # Before any write_file, return original messages
+            return messages
+
+        # Apply concise memory optimization after write_file detection
+        # self.logger.info(f"🧹 CLEARING MEMORY after write_file - creating clean slate")
+        optimized_messages = self.create_concise_messages(
+            system_prompt, messages, files_implemented
+        )
+
+        # Clear the flag after applying optimization
+        self.should_clear_memory_next = False
+
+        compression_ratio = (
+            ((len(messages) - len(optimized_messages)) / len(messages) * 100)
+            if messages
+            else 0
+        )
+        self.logger.info(
+            f"🎯 CONCISE optimization applied: {len(messages)} → {len(optimized_messages)} messages ({compression_ratio:.1f}% compression)"
+        )
+
+        return optimized_messages
+
+    def clear_current_round_tool_results(self):
+        """Clear current round tool results (called when starting new round)"""
+        self.current_round_tool_results = []
+        self.logger.info("🧹 Current round tool results cleared")
+
+    def debug_concise_state(self, files_implemented: int = 0):
+        """Debug method to show current concise memory state"""
+        stats = self.get_memory_statistics(files_implemented)
+
+        print("=" * 60)
+        print("🎯 CONCISE MEMORY AGENT STATE (Write-File-Based)")
+        print("=" * 60)
+        print(f"Last write_file detected: {stats['last_write_file_detected']}")
+        print(f"Should clear memory next: {stats['should_clear_memory_next']}")
+        print(f"Files implemented: {stats['implemented_files_tracked']}")
+        print(f"Current round: {stats['current_round']}")
+        print(f"Concise mode active: {stats['concise_mode_active']}")
+        print(f"Current round tool results: {stats['current_round_tool_results']}")
+        print(f"Essential tools recorded: {stats['essential_tools_recorded']}")
+        print(f"Implemented files tracked: {len(self.implemented_files)}")
+        print(f"Implemented files list: {self.implemented_files}")
+        print(f"Code summary file exists: {os.path.exists(self.code_summary_path)}")
+        print("")
+        print(
+            "📊 NEW LOGIC: write_file → clear memory → accumulate tools → next write_file"
+        )
+        print("📊 Essential Tools Tracked:")
+        essential_tools = [
+            "read_code_mem",
+            "read_file",
+            "write_file",
+            "execute_python",
+            "execute_bash",
+            "search_code",
+            "search_reference_code",
+            "get_file_structure",
+        ]
+        for tool in essential_tools:
+            tool_count = sum(
+                1 for r in self.current_round_tool_results if r["tool_name"] == tool
+            )
+            print(f"  - {tool}: {tool_count} calls")
+        print("=" * 60)