diff --git a/src/tokenops/control/integration.py b/src/tokenops/control/integration.py
index 4a28e5d..a228c21 100644
--- a/src/tokenops/control/integration.py
+++ b/src/tokenops/control/integration.py
@@ -183,25 +183,41 @@ def consume_carry(
 
 
 def _compact_messages(messages):
-    """Deep context_compaction MUTATE: rewrite the outgoing messages — pin every system
-    message, drop duplicate non-system messages (deduped tool outputs / repeated context)."""
+    """Deep context_compaction MUTATE: rewrite the outgoing messages into a cache-friendly
+    shape and drop redundant context.
+
+    Two moves, both aimed at the prompt-cache discount (cached prefix tokens bill far cheaper
+    than fresh ones):
+
+    * **Hoist system messages into a stable leading prefix.** The static instructions
+      (system prompt, schema, constraints) form the cacheable prefix; the volatile
+      conversation follows. Relative order within the system block, and within the tail, is
+      preserved, so this is a no-op for the common case where system is already first.
+    * **Drop duplicate non-system messages** (deduped tool outputs / repeated context),
+      keeping the first occurrence so the prefix stays stable across calls.
+
+    Trajectory hints are steer context and stay in the tail (recency favors the correction).
+    Note: the first reorder after an interleaved system message busts the old cache once, then
+    the new stable order caches; net win only when the system block is stable across calls.
+    """
     seen: set = set()
-    out: list = []
+    system_msgs: list = []
+    tail: list = []
     for msg in messages:
         role = msg.get("role") if isinstance(msg, dict) else None
         content = msg.get("content", "") if isinstance(msg, dict) else str(msg)
         if role == "system":
-            out.append(msg)
+            system_msgs.append(msg)
             continue
         if isinstance(content, str) and content.startswith("[TokenOps trajectory hint"):
-            out.append(msg)
+            tail.append(msg)
             continue
         key = (role, content)
         if key in seen:
             continue
         seen.add(key)
-        out.append(msg)
-    return out
+        tail.append(msg)
+    return system_msgs + tail
 
 
 def wrap_complete(
diff --git a/tests/test_context_compaction.py b/tests/test_context_compaction.py
index 7b56193..53f9f23 100644
--- a/tests/test_context_compaction.py
+++ b/tests/test_context_compaction.py
@@ -36,3 +36,32 @@ def test_no_hook_is_telemetry_only():
     det, pol = context_compaction.build(ctx_max=10_000, has_hook=False)
     sig = det.pre_call(_req(10_000), FakeView())
     assert pol.decide(sig, FakeView()).kind is ActionKind.ALLOW  # never HALT, never mutate
+
+
+def test_compact_hoists_system_into_stable_prefix():
+    from tokenops.control.integration import _compact_messages
+
+    msgs = [
+        {"role": "user", "content": "u1"},
+        {"role": "system", "content": "sys-a"},
+        {"role": "assistant", "content": "a1"},
+        {"role": "user", "content": "u1"},  # duplicate of first user turn
+        {"role": "system", "content": "sys-b"},
+    ]
+    out = _compact_messages(msgs)
+    # system messages hoisted to the front, in their original relative order
+    assert [m["role"] for m in out[:2]] == ["system", "system"]
+    assert [m["content"] for m in out[:2]] == ["sys-a", "sys-b"]
+    # tail keeps non-system order and drops the duplicate user turn
+    assert [m["content"] for m in out[2:]] == ["u1", "a1"]
+
+
+def test_compact_is_noop_when_system_already_first():
+    from tokenops.control.integration import _compact_messages
+
+    msgs = [
+        {"role": "system", "content": "sys"},
+        {"role": "user", "content": "u1"},
+        {"role": "assistant", "content": "a1"},
+    ]
+    assert _compact_messages(msgs) == msgs