diff --git a/agent/conversation_compression.py b/agent/conversation_compression.py
index 22decb591a2..0cf9d364c24 100644
--- a/agent/conversation_compression.py
+++ b/agent/conversation_compression.py
@@ -1098,7 +1098,11 @@ def _compress_context_via_codex_app_server(
             approx_tokens=approx_tokens,
             force=True,
         )
-        if getattr(result, "token_usage_last", None):
+        # An empty usage report must consume the pending post-compaction verdict
+        # rather than leaving preflight deferral armed until some unrelated later
+        # Codex turn supplies usage. Minimal external test engines may not expose
+        # the ContextEngine update hook; preserve their existing bookkeeping.
+        if hasattr(agent.context_compressor, "update_from_response"):
             _record_codex_app_server_usage(agent, result)
     except Exception:
         logger.debug("codex compaction bookkeeping failed", exc_info=True)
diff --git a/tests/run_agent/test_codex_app_server_compaction.py b/tests/run_agent/test_codex_app_server_compaction.py
index 4bd5e8431dd..e5e7d668e74 100644
--- a/tests/run_agent/test_codex_app_server_compaction.py
+++ b/tests/run_agent/test_codex_app_server_compaction.py
@@ -102,6 +102,8 @@ def test_codex_app_server_manual_compression_routes_to_codex_thread():
     assert agent._codex_session.calls == 1
     assert agent.context_compressor.compression_count == 1
     assert agent.context_compressor.last_compression_rough_tokens == 100000
+    # This minimal fake compressor does not implement update_from_response(),
+    # so the runtime preserves its existing pending-usage bookkeeping here.
     assert agent.context_compressor.last_prompt_tokens == -1
     assert agent.context_compressor.last_completion_tokens == 0
     assert agent.context_compressor.awaiting_real_usage_after_compression is True