diff --git a/agent/codex_runtime.py b/agent/codex_runtime.py
index 214ed013f57..664e4fbe9b1 100644
--- a/agent/codex_runtime.py
+++ b/agent/codex_runtime.py
@@ -279,6 +279,11 @@ def _record_codex_app_server_compaction(
             compressor, "compression_count", 0
         ) + 1
         compressor.last_compression_rough_tokens = approx_tokens or 0
+        # The app server has already completed a real compaction boundary. Its
+        # usage update (when supplied) is therefore the same real-vs-real
+        # effectiveness verdict used by the normal compression path.
+        if hasattr(compressor, "_verify_compaction_cleared_threshold"):
+            compressor._verify_compaction_cleared_threshold = True
         if not getattr(turn, "token_usage_last", None):
             compressor.last_prompt_tokens = -1
             compressor.last_completion_tokens = 0
diff --git a/tests/run_agent/test_codex_app_server_integration.py b/tests/run_agent/test_codex_app_server_integration.py
index 802096fd0c8..c88d45c483a 100644
--- a/tests/run_agent/test_codex_app_server_integration.py
+++ b/tests/run_agent/test_codex_app_server_integration.py
@@ -143,6 +143,13 @@ class TestRunConversationCodexPath:
                 turn_id="turn-compact-1",
                 thread_id="thread-compact-1",
                 compacted=True,
+                token_usage_last={
+                    "totalTokens": 300_000,
+                    "inputTokens": 300_000,
+                    "cachedInputTokens": 0,
+                    "outputTokens": 0,
+                    "reasoningOutputTokens": 0,
+                },
             )
 
         monkeypatch.setattr(CodexAppServerSession, "run_turn", fake_run_turn)
@@ -157,10 +164,11 @@ class TestRunConversationCodexPath:
 
         assert result["completed"] is True
         assert agent.context_compressor.compression_count == 1
-        # The app-server omitted usage, so the runtime consumes the pending
-        # post-compaction verdict instead of deferring preflight indefinitely.
-        assert agent.context_compressor.last_prompt_tokens == 0
+        # A compacted turn with real usage is judged against that same real
+        # prompt count, exactly like a normal completed compression boundary.
+        assert agent.context_compressor.last_prompt_tokens == 300_000
         assert agent.context_compressor.awaiting_real_usage_after_compression is False
+        assert agent.context_compressor._ineffective_compression_count == 1
         assert events == [
             (
                 "session:compress",