diff --git a/cueapi/cli.py b/cueapi/cli.py
index e81ebaf..d5f394b 100644
--- a/cueapi/cli.py
+++ b/cueapi/cli.py
@@ -2174,6 +2174,21 @@ def _resolve_recipient(client, recipient: str) -> str:
     help="Decoupled reply target (defaults to the sender).",
 )
 @click.option("--metadata", default=None, help="JSON metadata blob")
+@click.option(
+    "--mode",
+    "mode",
+    default="auto",
+    type=click.Choice(["live", "bg", "inbox", "webhook", "auto"]),
+    show_default=True,
+    help=(
+        "Delivery mode hint. live = recipient's attached Live session, bg = "
+        "spawn a fresh background session, inbox = leave in inbox for pull, "
+        "webhook = POST to recipient's configured webhook, auto = server "
+        "picks the best supported mode based on recipient capabilities. "
+        "Server may downgrade if the requested mode isn't supported — see "
+        "`Sent via X` in the response."
+    ),
+)
 @click.option(
     "--idempotency-key",
     "idempotency_key",
@@ -2195,6 +2210,7 @@ def message_to(
     expects_reply: bool,
     reply_to_agent: Optional[str],
     metadata: Optional[str],
+    mode: str,
     idempotency_key: Optional[str],
 ) -> None:
     """Send a message to a recipient by name, slug, or agent ID.
@@ -2219,6 +2235,12 @@ def message_to(
             body["metadata"] = json.loads(metadata)
         except json.JSONDecodeError:
             raise click.UsageError("--metadata must be valid JSON")
+    # Default-omit discipline: only send delivery_mode when the user opted
+    # away from `auto`. Server treats absent == auto, so this avoids payload
+    # noise on the common path and keeps wire-format identical to pre-Surface-6
+    # senders. `auto` is also redundant to send.
+    if mode != "auto":
+        body["delivery_mode"] = mode
 
     headers: dict = {"X-Cueapi-From-Agent": from_agent}
     if idempotency_key:
@@ -2246,6 +2268,20 @@ def message_to(
                 if m.get("thread_id"):
                     echo_info("Thread:", m["thread_id"])
                 echo_info("Delivery state:", m.get("delivery_state", "?"))
+                # Surface the server's chosen delivery mode. The response's
+                # `effective_delivery_mode` is the mode the server actually
+                # used, which may differ from the requested `mode` if the
+                # recipient doesn't support it (e.g. requested live, recipient
+                # has no live session, downgraded to inbox).
+                effective = m.get("effective_delivery_mode")
+                if effective:
+                    if mode != "auto" and effective != mode:
+                        echo_info(
+                            "Sent via:",
+                            f"{effective} (requested {mode}, recipient does not support it)",
+                        )
+                    else:
+                        echo_info("Sent via:", effective)
                 downgraded_header = None
                 try:
                     downgraded_header = resp.headers.get("X-CueAPI-Priority-Downgraded")
diff --git a/tests/test_cli.py b/tests/test_cli.py
index 90ca5ff..58fa8f6 100644
--- a/tests/test_cli.py
+++ b/tests/test_cli.py
@@ -5,6 +5,7 @@
 
 from typing import Any, Optional
 
+import pytest
 from click.testing import CliRunner
 
 from cueapi.cli import main
@@ -2599,3 +2600,149 @@ def test_top_level_help_lists_message_to():
     result = runner.invoke(main, ["--help"])
     assert result.exit_code == 0
     assert "message-to" in result.output
+
+
+# --- message-to --mode flag (Surface 6 delivery_mode) ---
+
+
+def test_message_to_mode_default_auto_omits_field(monkeypatch):
+    # Default is `auto` and the server treats absent == auto, so we don't
+    # send the field on the common path — keeps wire-format identical to
+    # pre-Surface-6 senders and avoids payload noise.
+    holder: dict = {}
+    _patch_messages_client(
+        monkeypatch,
+        holder,
+        responses={
+            ("POST", "/messages"): lambda: _FakeResp(
+                201, {"id": "msg_z", "delivery_state": "queued"}
+            )
+        },
+    )
+    result = runner.invoke(
+        main,
+        ["message-to", "agt_x", "--from", "y@z", "--body", "hi"],
+    )
+    assert result.exit_code == 0, result.output
+    body = holder["client"].calls[-1][2]
+    assert "delivery_mode" not in body
+
+
+def test_message_to_mode_explicit_auto_still_omits_field(monkeypatch):
+    # User explicitly typing --mode auto is the same wire-format as omitting
+    # it. No reason to send the redundant field.
+    holder: dict = {}
+    _patch_messages_client(
+        monkeypatch,
+        holder,
+        responses={
+            ("POST", "/messages"): lambda: _FakeResp(
+                201, {"id": "msg_z", "delivery_state": "queued"}
+            )
+        },
+    )
+    result = runner.invoke(
+        main,
+        ["message-to", "agt_x", "--from", "y@z", "--body", "hi", "--mode", "auto"],
+    )
+    assert result.exit_code == 0, result.output
+    body = holder["client"].calls[-1][2]
+    assert "delivery_mode" not in body
+
+
+@pytest.mark.parametrize("mode", ["live", "bg", "inbox", "webhook"])
+def test_message_to_mode_non_auto_passed_through(monkeypatch, mode):
+    holder: dict = {}
+    _patch_messages_client(
+        monkeypatch,
+        holder,
+        responses={
+            ("POST", "/messages"): lambda: _FakeResp(
+                201,
+                {
+                    "id": "msg_z",
+                    "delivery_state": "queued",
+                    "effective_delivery_mode": mode,
+                },
+            )
+        },
+    )
+    result = runner.invoke(
+        main,
+        ["message-to", "agt_x", "--from", "y@z", "--body", "hi", "--mode", mode],
+    )
+    assert result.exit_code == 0, result.output
+    body = holder["client"].calls[-1][2]
+    assert body["delivery_mode"] == mode
+    # echo_info pads labels — match the components, not the raw concat.
+    assert "Sent via:" in result.output
+    assert mode in result.output
+
+
+def test_message_to_mode_invalid_value_rejected_by_click():
+    # Click.Choice covers validation; we just confirm the gate is in place
+    # so a typo doesn't silently sail past as "auto".
+    result = runner.invoke(
+        main,
+        ["message-to", "agt_x", "--from", "y@z", "--body", "hi", "--mode", "bogus"],
+    )
+    assert result.exit_code != 0
+    assert "bogus" in result.output or "invalid choice" in result.output.lower()
+
+
+def test_message_to_surfaces_downgraded_delivery_mode(monkeypatch):
+    # Server downgrade case: requested `live` but recipient has no live
+    # session, server delivered via inbox. CLI surfaces both the chosen
+    # mode and the "you asked for X" hint.
+    holder: dict = {}
+    _patch_messages_client(
+        monkeypatch,
+        holder,
+        responses={
+            ("POST", "/messages"): lambda: _FakeResp(
+                201,
+                {
+                    "id": "msg_z",
+                    "delivery_state": "queued",
+                    "effective_delivery_mode": "inbox",
+                },
+            )
+        },
+    )
+    result = runner.invoke(
+        main,
+        ["message-to", "agt_x", "--from", "y@z", "--body", "hi", "--mode", "live"],
+    )
+    assert result.exit_code == 0, result.output
+    assert "Sent via:" in result.output
+    assert "inbox" in result.output
+    assert "requested live" in result.output
+
+
+def test_message_to_omits_sent_via_when_server_does_not_return_it(monkeypatch):
+    # Pre-Surface-6 server (or auto + no field) returns no
+    # effective_delivery_mode. The CLI should not emit a "Sent via:" line.
+    holder: dict = {}
+    _patch_messages_client(
+        monkeypatch,
+        holder,
+        responses={
+            ("POST", "/messages"): lambda: _FakeResp(
+                201, {"id": "msg_z", "delivery_state": "queued"}
+            )
+        },
+    )
+    result = runner.invoke(
+        main,
+        ["message-to", "agt_x", "--from", "y@z", "--body", "hi"],
+    )
+    assert result.exit_code == 0, result.output
+    assert "Sent via:" not in result.output
+
+
+def test_message_to_help_lists_mode_flag():
+    result = runner.invoke(main, ["message-to", "--help"])
+    assert result.exit_code == 0
+    assert "--mode" in result.output
+    for choice in ("live", "bg", "inbox", "webhook", "auto"):
+        assert choice in result.output