Phase 7: Self-Optimizing Agents (Feedback analysis on rejection, lesson injection on retry)

backend/routers/agent_runner.py

@@ -363,8 +363,16 @@ async def approve_task(task_id: str, background_tasks: BackgroundTasks):
     return {"message": "Task approved", "task": task}
 
 @router.post("/{task_id}/reject")
-async def reject_task(task_id: str):
+async def reject_task(task_id: str, background_tasks: BackgroundTasks, feedback: str | None = None):
     task = update_task_status(task_id, "todo")
+    
+    # Trigger Self-Optimization Loop
+    background_tasks.add_task(
+        memory_service.analyze_rejection, 
+        task_id=task_id, 
+        feedback=feedback
+    )
+    
     await audit_service.log_action(
         user_id=None,
         action="task_rejected",

backend/services/agent_runner_service.py

@@ -104,6 +104,19 @@ class AgentRunnerService:
                     for m in memories:
                         memory_blocks.append(f"- Memory: {m['content']}")
                     extra_context += memory_header + "\n".join(memory_blocks)
+                
+                # Fetch Self-Optimization Lessons for this specific task
+                lessons_res = supabase.table("project_memory") \
+                    .select("content") \
+                    .eq("task_id", task_id) \
+                    .eq("memory_type", "self_optimization_lesson") \
+                    .order("created_at", desc=True) \
+                    .limit(1) \
+                    .execute()
+                
+                if lessons_res.data:
+                    lesson = lessons_res.data[0]["content"]
+                    extra_context += f"\n\n### CRITICAL LESSON FROM PREVIOUS ATTEMPT\n{lesson}\n"
 
             import time
             import hashlib

backend/services/memory_service.py

@@ -109,4 +109,66 @@ class MemoryService:
             }
         )
 
+    async def analyze_rejection(self, task_id: str, feedback: Optional[str] = None):
+        """
+        Analyzes a task rejection to generate a 'Self-Optimization Lesson'.
+        Triggered when a human rejects an agent's output.
+        """
+        try:
+            # 1. Fetch task and its failed output
+            task_res = supabase.table("tasks").select("*, projects(name, description)").eq("id", task_id).single().execute()
+            if not task_res.data:
+                return
+            
+            task = task_res.data
+            output = task.get("output_data") or {}
+            
+            # 2. Get an analyst agent
+            from agents.agent_factory import AgentFactory
+            from services.llm_config import getDefaultProvider, getDefaultModel
+            
+            provider = getDefaultProvider()
+            model = getDefaultModel(provider)
+            
+            analyst = AgentFactory.get_agent(
+                provider=provider,
+                name="Optimization Analyst",
+                role="Self-Optimization Expert",
+                model=model,
+                system_prompt="You analyze task rejections. Your goal is to produce a single, concise 'Lesson Learned' that the next agent should follow to avoid repeating the mistake. Focus on the core reason for rejection."
+            )
+            
+            # 3. Construct prompt for analysis
+            analysis_prompt = f"""
+            TASK: {task.get('title')}
+            DESCRIPTION: {task.get('description')}
+            
+            REJECTED OUTPUT:
+            {str(output)[:2000]}
+            
+            HUMAN FEEDBACK: {feedback or "No explicit feedback provided, but the output did not meet quality standards."}
+            
+            Provide a concise 'LESSON LEARNED' for the next agent. Start with 'Next time, you must...'
+            """
+            
+            result = await analyst.run(analysis_prompt, [])
+            lesson_text = result.get("raw_output") or result.get("data")
+            
+            if lesson_text:
+                await self.save_memory(
+                    project_id=task.get("project_id"),
+                    task_id=task_id,
+                    content=f"Optimization Lesson for '{task.get('title')}': {lesson_text}",
+                    memory_type="self_optimization_lesson",
+                    metadata={
+                        "original_task_id": task_id,
+                        "was_rejected": True,
+                        "feedback": feedback
+                    }
+                )
+                logger.info(f"Saved self-optimization lesson for task {task_id}")
+                
+        except Exception as e:
+            logger.error(f"Failed to analyze rejection for task {task_id}: {e}")
+
 memory_service = MemoryService()