diff --git a/agent/codex_responses_adapter.py b/agent/codex_responses_adapter.py
index ffac5e20fa..de6305a6fa 100644
--- a/agent/codex_responses_adapter.py
+++ b/agent/codex_responses_adapter.py
@@ -350,9 +350,8 @@ def _replay_reasoning_items(
     """Replay persisted encrypted reasoning/compaction items for one assistant turn. Skips duplicate
     ids, ``compaction`` checkpoints unless THIS request carries ``context_management`` (else a persisted
     checkpoint erases pre-checkpoint history on a model that cannot decrypt it), and items stamped by
-    another issuer or model (HTTP 400). Endpoint-stamped legacy items without model provenance drop
-    (fail closed) when the current model is known; fully unstamped items pass. ``id`` (store=False
-    lookups 404) and the Hermes provenance fields are stripped."""
+    another issuer or model (HTTP 400). Items without a model stamp (legacy or unstamped) replay on a
+    matching issuer. ``id`` (store=False lookups 404) and the Hermes provenance fields are stripped."""
     global _CROSS_ISSUER_WARN_EMITTED
     replayed: List[Dict[str, Any]] = []
     for ri in _as_list(msg.get("codex_reasoning_items")):
@@ -364,9 +363,11 @@ def _replay_reasoning_items(
         item_issuer = ri.get("_issuer_kind")
         item_model = ri.get("_issuer_model")
         foreign_issuer = current_issuer_kind is not None and item_issuer is not None and item_issuer != current_issuer_kind
-        foreign_model = current_issuer_model is not None and (
-            (item_model is not None and item_model != current_issuer_model)
-            or (item_model is None and item_issuer is not None)
+        # No model stamp → trust the endpoint stamp. Native compaction checkpoints and reasoning persisted
+        # before model stamping exist carry none; dropping them would erase every existing session's
+        # context once. A wrong guess is caught by the invalid_encrypted_content 400 classifier.
+        foreign_model = (
+            current_issuer_model is not None and item_model is not None and item_model != current_issuer_model
         )
         if foreign_issuer or foreign_model:
             if not _CROSS_ISSUER_WARN_EMITTED:
diff --git a/tests/agent/test_codex_responses_adapter.py b/tests/agent/test_codex_responses_adapter.py
index 61894dc517..ec382875af 100644
--- a/tests/agent/test_codex_responses_adapter.py
+++ b/tests/agent/test_codex_responses_adapter.py
@@ -593,16 +593,21 @@ def test_reasoning_replay_requires_matching_issuer_model_on_same_endpoint():
     assert not any(i.get("type") == "reasoning" for i in other)
 
 
-def test_reasoning_replay_drops_endpoint_stamped_legacy_item_without_model():
-    # Fail closed: an endpoint-only stamp cannot prove the minting model once the request model is known.
+def test_legacy_endpoint_stamped_item_without_model_replays_on_same_issuer():
+    # WHY: native compaction checkpoints and reasoning persisted before model stamping carry only the
+    # endpoint stamp; dropping them would erase every existing session's context once after upgrade.
     issuer = "other:https://responses.example.com/v1"
     legacy = {"type": "reasoning", "encrypted_content": "legacy-blob", "_issuer_kind": issuer}
     items = _chat_messages_to_responses_input(
         _reasoning_history(legacy), current_issuer_kind=issuer, current_issuer_model="gpt-5.6-sol"
     )
-    assert not any(i.get("type") == "reasoning" for i in items)
-    # Ordinary assistant text stays replayable; only the sealed blob is withheld.
-    assert any(i.get("role") == "assistant" for i in items)
+    replayed = [i for i in items if i.get("type") == "reasoning"]
+    assert [i["encrypted_content"] for i in replayed] == ["legacy-blob"]
+    # A different endpoint stamp still drops.
+    foreign = _chat_messages_to_responses_input(
+        _reasoning_history(legacy), current_issuer_kind="codex_backend", current_issuer_model="gpt-5.6-sol"
+    )
+    assert not any(i.get("type") == "reasoning" for i in foreign)
 
 
 def test_preflight_codex_api_kwargs_drops_oversized_message_id_end_to_end():