feat: improve memory cleanup when plan approve

Author: dori

src/mcp_as_a_judge/db/interface.py

@@ -78,3 +78,19 @@ async def get_recent_sessions(self, limit: int = 10) -> list[tuple[str, int]]:
             List of tuples: (session_id, last_activity_timestamp), ordered by most recent first
         """
         pass
+
+    @abstractmethod
+    async def delete_previous_plan(self, session_id: str) -> int:
+        """
+        Delete all previous judge_coding_plan records except the most recent one.
+
+        This method removes all but the last conversation record with source='judge_coding_plan'
+        for the given session to avoid keeping multiple failed plan attempts.
+
+        Args:
+            session_id: Session identifier
+
+        Returns:
+            Number of records deleted
+        """
+        pass

src/mcp_as_a_judge/db/providers/sqlite_provider.py

@@ -8,7 +8,7 @@
 import time
 import uuid
 
-from sqlalchemy import create_engine, func
+from sqlalchemy import create_engine, delete, func
 from sqlmodel import Session, SQLModel, asc, desc, select
 
 from mcp_as_a_judge.core.constants import MAX_CONTEXT_TOKENS
@@ -303,3 +303,46 @@ async def get_recent_sessions(self, limit: int = 10) -> list[tuple[str, int]]:
             results = session.exec(stmt).all()
             # results are tuples (session_id, last_activity)
             return [(row[0], int(row[1])) for row in results]
+
+    async def delete_previous_plan(self, session_id: str) -> int:
+        """
+        Delete all previous judge_coding_plan records except the most recent one.
+
+        Uses SQL ORM to find all judge_coding_plan records for the session,
+        keeps only the most recent one, and deletes the rest.
+        """
+        with Session(self.engine) as session:
+            # Find all judge_coding_plan records for this session, ordered by timestamp DESC
+            stmt = (
+                select(ConversationRecord)
+                .where(ConversationRecord.session_id == session_id)
+                .where(ConversationRecord.source == "judge_coding_plan")
+                .order_by(
+                    desc(ConversationRecord.timestamp),
+                    desc(ConversationRecord.id),
+                )
+            )
+            plan_records = list(session.exec(stmt).all())
+
+            if len(plan_records) <= 1:
+                # No previous plans to delete
+                logger.info(f"No previous judge_coding_plan records to delete for session {session_id}")
+                return 0
+
+            # Keep the first record (most recent), delete the rest
+            records_to_delete = plan_records[1:]  # Skip the first (most recent)
+            record_ids_to_delete = [record.id for record in records_to_delete]
+
+            logger.info(f"Deleting {len(records_to_delete)} previous judge_coding_plan records for session {session_id}")
+
+            # Delete records using SQL IN clause
+            delete_stmt = delete(ConversationRecord).where(
+                ConversationRecord.id.in_(record_ids_to_delete)
+            )
+            result = session.exec(delete_stmt)
+            session.commit()
+
+            deleted_count = result.rowcount if hasattr(result, 'rowcount') else len(records_to_delete)
+            logger.info(f"Successfully deleted {deleted_count} previous judge_coding_plan records")
+
+            return deleted_count

src/mcp_as_a_judge/server.py

@@ -1686,6 +1686,11 @@ async def judge_coding_plan(
             # Mark plan as approved for completion validation and update state
             updated_task_metadata.mark_plan_approved()
             updated_task_metadata.update_state(TaskState.PLAN_APPROVED)
+
+            # Delete previous failed plan attempts, keeping only the most recent approved one
+            deleted_count = await conversation_service.db.delete_previous_plan(updated_task_metadata.task_id)
+            logger.info(f"Plan approved - deleted {deleted_count} previous plan attempts")
+
             # Force next step to code review implementation gate
             workflow_guidance.next_tool = "judge_code_change"
             if not workflow_guidance.reasoning: