fix(agent): persist _last_summary across restarts with used sentinel

The previous implementation popped _last_summary from session.metadata
after injecting it into the prompt, then saved the session. This caused
the summary to be permanently lost after a process restart, making the
AI forget archived context and appear to ignore memory or reference
non-existent previous messages.

Replace the destructive pop with a _last_summary_used sentinel:
- _last_summary stays in metadata for restart survival
- _last_summary_used prevents duplicate injection within the same turn
- Clear the sentinel whenever a new summary is generated

Updates tests to match the new persistence behavior.

nanobot/agent/autocompact.py

@@ -89,6 +89,7 @@ class AutoCompact:
             if summary and summary != "(nothing)":
                 self._summaries[key] = (summary, last_active)
                 session.metadata["_last_summary"] = {"text": summary, "last_active": last_active.isoformat()}
+                session.metadata.pop("_last_summary_used", None)
             session.messages = kept_msgs
             session.last_consolidated = 0
             session.updated_at = datetime.now()
@@ -111,13 +112,15 @@ class AutoCompact:
             logger.info("Auto-compact: reloading session {} (archiving={})", key, key in self._archiving)
             session = self.sessions.get_or_create(key)
         # Hot path: summary from in-memory dict (process hasn't restarted).
-        # Also clean metadata copy so stale _last_summary never leaks to disk.
         entry = self._summaries.pop(key, None)
         if entry:
-            session.metadata.pop("_last_summary", None)
+            session.metadata["_last_summary_used"] = True
             return session, self._format_summary(entry[0], entry[1])
-        if "_last_summary" in session.metadata:
-            meta = session.metadata.pop("_last_summary")
-            self.sessions.save(session)
+        # Cold path: summary persisted in session metadata (process restarted).
+        # Keep it in metadata so it survives restarts; only inject once per
+        # turn via the _last_summary_used sentinel.
+        meta = session.metadata.get("_last_summary")
+        if meta and not session.metadata.get("_last_summary_used"):
+            session.metadata["_last_summary_used"] = True
             return session, self._format_summary(meta["text"], datetime.fromisoformat(meta["last_active"]))
         return session, None

nanobot/agent/memory.py

@@ -585,6 +585,7 @@ class Consolidator:
                 "text": summary,
                 "last_active": session.updated_at.isoformat(),
             }
+            session.metadata.pop("_last_summary_used", None)
             self.sessions.save(session)
 
     def estimate_session_prompt_tokens(

tests/agent/test_auto_compact.py

@@ -1021,13 +1021,15 @@ class TestSummaryPersistence:
         assert summary is not None
         assert "User said hello." in summary
         assert "Inactive for" in summary
-        # Metadata should be cleaned up after consumption
-        assert "_last_summary" not in reloaded.metadata
+        # Metadata persists so the summary survives restarts; _last_summary_used
+        # sentinel prevents duplicate injection within the same turn.
+        assert "_last_summary" in reloaded.metadata
+        assert reloaded.metadata.get("_last_summary_used") is True
         await loop.close_mcp()
 
     @pytest.mark.asyncio
     async def test_metadata_cleanup_no_leak(self, tmp_path):
-        """_last_summary should be removed from metadata after being consumed."""
+        """_last_summary persists in metadata for restart survival; _last_summary_used sentinel prevents duplicate injection."""
         loop = _make_loop(tmp_path, session_ttl_minutes=15)
         session = loop.sessions.get_or_create("cli:test")
         _add_turns(session, 6, prefix="hello")
@@ -1050,10 +1052,13 @@ class TestSummaryPersistence:
         _, summary = loop.auto_compact.prepare_session(reloaded, "cli:test")
         assert summary is not None
 
-        # Second call: no summary (already consumed)
+        # Second call: no summary (already consumed this turn)
         _, summary2 = loop.auto_compact.prepare_session(reloaded, "cli:test")
         assert summary2 is None
-        assert "_last_summary" not in reloaded.metadata
+        # _last_summary stays in metadata for restart survival;
+        # _last_summary_used sentinel prevents duplicate injection.
+        assert "_last_summary" in reloaded.metadata
+        assert reloaded.metadata.get("_last_summary_used") is True
         await loop.close_mcp()
 
     @pytest.mark.asyncio
@@ -1081,6 +1086,8 @@ class TestSummaryPersistence:
         # In-memory path is taken (no restart)
         _, summary = loop.auto_compact.prepare_session(reloaded, "cli:test")
         assert summary is not None
-        # Metadata should also be cleaned up
-        assert "_last_summary" not in reloaded.metadata
+        # _last_summary stays in metadata for restart survival;
+        # _last_summary_used sentinel prevents duplicate injection.
+        assert "_last_summary" in reloaded.metadata
+        assert reloaded.metadata.get("_last_summary_used") is True
         await loop.close_mcp()

tests/agent/test_loop_consolidation_tokens.py

@@ -190,7 +190,10 @@ async def test_consolidation_persists_summary_for_next_prepare_session(tmp_path,
     reloaded, pending = loop.auto_compact.prepare_session(reloaded, "cli:test")
     assert pending is not None
     assert "User discussed project status." in pending
-    assert "_last_summary" not in reloaded.metadata
+    # _last_summary persists for restart survival; _last_summary_used prevents
+    # duplicate injection within the same turn.
+    assert "_last_summary" in reloaded.metadata
+    assert reloaded.metadata.get("_last_summary_used") is True
 
 
 @pytest.mark.asyncio