fix(compression): skip empty post-handoff summary windows

agent/context_compressor.py

@@ -2777,6 +2777,19 @@ def compress(self, messages: List[Dict[str, Any]], current_tokens: int = None, f
             if summary_body and not self._previous_summary:
                 self._previous_summary = summary_body
             turns_to_summarize = messages[max(compress_start, summary_idx + 1):compress_end]
+            if not turns_to_summarize:
+                self._ineffective_compression_count += 1
+                self._last_compression_savings_pct = 0.0
+                if not self.quiet_mode:
+                    logger.warning(
+                        "Compression skipped: latest context summary leaves no "
+                        "new turns to summarize in window %d-%d. "
+                        "ineffective_compression_count=%d",
+                        compress_start,
+                        compress_end,
+                        self._ineffective_compression_count,
+                    )
+                return messages
         elif self._previous_summary:
             # No handoff summary found in the current messages, but
             # _previous_summary is non-empty — it was set by a different

tests/agent/test_context_compressor_summary_continuity.py

@@ -85,3 +85,34 @@ def fake_generate_summary(turns_to_summarize, focus_topic=None):
     assert compressor._previous_summary == old_summary
     assert seen_turns
     assert all(old_summary not in str(msg.get("content", "")) for msg in seen_turns)
+
+
+def test_empty_post_handoff_window_noops_without_summary_call():
+    """A latest handoff that consumes the window must not trigger an empty summary."""
+    compressor = _compressor()
+    old_summary = "WINDOW-END-SUMMARY durable facts already captured"
+    messages = [
+        {"role": "system", "content": "system prompt"},
+        {"role": "user", "content": "protected first user turn"},
+        {"role": "assistant", "content": "old assistant turn already captured"},
+        {"role": "user", "content": f"{SUMMARY_PREFIX}\n{old_summary}"},
+        {"role": "assistant", "content": "recent tail response"},
+        {"role": "user", "content": "latest tail request"},
+        {"role": "assistant", "content": "latest tail answer"},
+    ]
+
+    with (
+        patch.object(compressor, "_find_tail_cut_by_tokens", return_value=4),
+        patch.object(compressor, "_generate_summary") as mock_generate_summary,
+    ):
+        result = compressor.compress(messages, current_tokens=90_000)
+
+    mock_generate_summary.assert_not_called()
+    assert result == messages
+    assert compressor._previous_summary == old_summary
+    assert compressor.compression_count == 0
+    assert compressor._ineffective_compression_count == 1
+    assert compressor._last_compression_savings_pct == 0.0
+    assert compressor._last_summary_dropped_count == 0
+    assert compressor._last_summary_fallback_used is False
+    assert compressor._last_compress_aborted is False