support: flush support/summarizer model after summarization (cleanup + call)

src/cai/repl/commands/memory.py

@@ -4,6 +4,7 @@ Manages memory storage in .cai/memory for persistent context.
 """
 
 from typing import List, Optional, Dict, Any
+import inspect
 import os
 import asyncio
 import json
@@ -1256,15 +1257,31 @@ This session is being continued from a previous conversation that ran out of con
                 input=f"Please summarize the following conversation:\n\n{conversation_text}",
                 max_turns=1
             )
-            
+
             if result.final_output:
                 return str(result.final_output)
             else:
                 return None
-                
+
         except Exception as e:
             console.print(f"[red]Error generating summary: {e}[/red]")
             return None
+        finally:
+            # Best-effort: explicitly cleanup the temporary summary/support model
+            try:
+                model_inst = getattr(summary_agent, "model", None)
+                # Some Agent constructions put the Model object directly on `agent.model`
+                # and some providers expose a cleanup coroutine.
+                if model_inst is not None and hasattr(model_inst, "cleanup"):
+                    try:
+                        coro = model_inst.cleanup()
+                        if inspect.isawaitable(coro):
+                            await coro
+                    except Exception:
+                        # best-effort cleanup — swallow any errors
+                        pass
+            except Exception:
+                pass
     
     def _format_history_for_summary(self, history: List[Dict[str, Any]]) -> str:
         """Format message history for summarization.

src/cai/sdk/agents/models/openai_chatcompletions.py

@@ -470,6 +470,49 @@ class OpenAIChatCompletionsModel(Model):
             # Ignore any errors during cleanup
             pass
 
+    async def cleanup(self) -> None:
+        """Explicitly cleanup underlying clients and free instance registry.
+
+        This is intended to be called when a temporary model instance (for
+        example the summary/support model) is no longer needed. It will try
+        to close the HTTP/async client if available, remove the instance
+        from the legacy `ACTIVE_MODEL_INSTANCES` registry and clear the
+        in-memory message history so any backing LLM server can free slots
+        or context.
+        """
+        try:
+            client = getattr(self, "_client", None)
+            if client is not None:
+                aclose = getattr(client, "aclose", None)
+                if aclose:
+                    try:
+                        res = aclose()
+                        # Await if it's awaitable
+                        if inspect.isawaitable(res):
+                            await res
+                    except Exception:
+                        # Best-effort close
+                        pass
+                try:
+                    delattr(self, "_client")
+                except Exception:
+                    pass
+        except Exception:
+            pass
+
+        try:
+            key = (getattr(self, '_display_name', None), getattr(self, 'agent_id', None))
+            if key in ACTIVE_MODEL_INSTANCES:
+                del ACTIVE_MODEL_INSTANCES[key]
+        except Exception:
+            pass
+
+        try:
+            if hasattr(self, 'message_history') and isinstance(self.message_history, list):
+                self.message_history.clear()
+        except Exception:
+            pass
+
     def add_to_message_history(self, msg):
         """Add a message to this instance's history if it's not a duplicate.