diff --git a/.claude/hooks/.diagnostics_state.json b/.claude/hooks/.diagnostics_state.json
deleted file mode 100644
index 16a76768..00000000
--- a/.claude/hooks/.diagnostics_state.json
+++ /dev/null
@@ -1 +0,0 @@
-{"file": "/home/patrick/Projects/AIPass/src/aipass/drone/apps/modules/git_module.py", "errors": [{"line": 671, "message": "E501: Line too long (193 > 120)"}]}
\ No newline at end of file
diff --git a/src/aipass/prax/apps/handlers/monitoring/filesystem_handler.py b/src/aipass/prax/apps/handlers/monitoring/filesystem_handler.py
index 807d0ed7..6e91faeb 100644
--- a/src/aipass/prax/apps/handlers/monitoring/filesystem_handler.py
+++ b/src/aipass/prax/apps/handlers/monitoring/filesystem_handler.py
@@ -535,11 +535,15 @@ class MonitoringFileHandler(FileSystemEventHandler):
     # CLAUDE CODE AGENT ACTIVITY PARSING (JSONL sessions)
     # =========================================================================
 
+    _MAX_AGENT_EVENTS_PER_BATCH = 3
+
     def _parse_agent_activity(self, file_path, branch):
         """Parse Claude Code session JSONL to show agent actions.
 
         Returns True if an event was emitted (or deduped), False on failure.
-        Iterates forward through new lines, emitting all distinct actions.
+        Iterates forward through all new lines (for model extraction), but only
+        emits the last few distinct actions to prevent doom-scrolling when large
+        batches accumulate between polls or after compaction resets.
         """
         try:
             path_key = str(file_path)
@@ -560,6 +564,9 @@ class MonitoringFileHandler(FileSystemEventHandler):
             if not lines:
                 return True
 
+            pending_actions = []
+            last_seen = self._last_agent_action.get(path_key)
+
             for line in lines:
                 try:
                     entry = _json.loads(line)
@@ -575,10 +582,12 @@ class MonitoringFileHandler(FileSystemEventHandler):
                 if not action_text:
                     continue
 
-                if self._last_agent_action.get(path_key) == action_text:
+                if action_text == last_seen:
                     continue
-                self._last_agent_action[path_key] = action_text
+                last_seen = action_text
+                pending_actions.append(action_text)
 
+            for action_text in pending_actions[-self._MAX_AGENT_EVENTS_PER_BATCH :]:
                 tagged_branch = self._tag_branch_with_model(path_key, branch)
                 evt = MonitoringEvent(
                     priority=1,
@@ -591,6 +600,9 @@ class MonitoringFileHandler(FileSystemEventHandler):
                 if self._event_queue:
                     self._event_queue.enqueue(evt)
 
+            if pending_actions:
+                self._last_agent_action[path_key] = pending_actions[-1]
+
             return True
 
         except Exception as e:
diff --git a/src/aipass/prax/tests/test_filesystem_handler.py b/src/aipass/prax/tests/test_filesystem_handler.py
index 0268aef4..32060233 100644
--- a/src/aipass/prax/tests/test_filesystem_handler.py
+++ b/src/aipass/prax/tests/test_filesystem_handler.py
@@ -1763,6 +1763,56 @@ class TestParseAgentActivity:
             result = handler._parse_agent_activity(mock_path, "PRAX")
         assert result is True
 
+    def test_batch_cap_limits_emitted_events(self):
+        """Should only emit last N events when batch exceeds cap."""
+        mod, _, _, _, queue = _import_filesystem_handler()
+        handler = _make_handler(mod, queue=queue)
+        mock_path = MagicMock()
+        mock_path.__str__ = MagicMock(return_value="/fake/claude.jsonl")
+        mock_path.name = "claude.jsonl"
+        entries = []
+        tools = ["Read", "Edit", "Write", "Grep", "Glob", "Bash"]
+        for i, tool in enumerate(tools):
+            entries.append(
+                {
+                    "type": "assistant",
+                    "message": {"content": [{"type": "tool_use", "name": tool, "input": {"file_path": f"f{i}.py"}}]},
+                }
+            )
+        data = "\n".join(json.dumps(e) for e in entries) + "\n"
+        mock_path.stat.return_value.st_size = len(data)
+        with patch("builtins.open", _mopen(read_data=data)):
+            result = handler._parse_agent_activity(mock_path, "PRAX")
+        assert result is True
+        cap = mod.MonitoringFileHandler._MAX_AGENT_EVENTS_PER_BATCH
+        assert queue.enqueue.call_count == cap
+
+    def test_batch_cap_emits_tail_not_head(self):
+        """Should emit the last actions in a large batch, not the first."""
+        mod, mock_eq, _, _, queue = _import_filesystem_handler()
+        handler = _make_handler(mod, queue=queue)
+        mock_path = MagicMock()
+        mock_path.__str__ = MagicMock(return_value="/fake/claude.jsonl")
+        mock_path.name = "claude.jsonl"
+        entries = []
+        for i in range(6):
+            entries.append(
+                {
+                    "type": "assistant",
+                    "message": {"content": [{"type": "tool_use", "name": f"Tool{i}", "input": {}}]},
+                }
+            )
+        data = "\n".join(json.dumps(e) for e in entries) + "\n"
+        mock_path.stat.return_value.st_size = len(data)
+        with patch("builtins.open", _mopen(read_data=data)):
+            handler._parse_agent_activity(mock_path, "PRAX")
+        cap = mod.MonitoringFileHandler._MAX_AGENT_EVENTS_PER_BATCH
+        assert queue.enqueue.call_count == cap
+        last_call_kwargs = mock_eq.MonitoringEvent.call_args_list[-1][1]
+        assert "Tool5" in last_call_kwargs["message"]
+        first_call_kwargs = mock_eq.MonitoringEvent.call_args_list[-cap][1]
+        assert "Tool3" in first_call_kwargs["message"]
+
 
 # =============================================
 # CHECK COMMAND INDICATOR TESTS