diff --git a/benchmarks/harbor-buzz-orchestra/src/harbor_buzz_orchestra/container_runtime.py b/benchmarks/harbor-buzz-orchestra/src/harbor_buzz_orchestra/container_runtime.py
index 149a5295a7..ed883a820a 100644
--- a/benchmarks/harbor-buzz-orchestra/src/harbor_buzz_orchestra/container_runtime.py
+++ b/benchmarks/harbor-buzz-orchestra/src/harbor_buzz_orchestra/container_runtime.py
@@ -157,9 +157,17 @@ async def run(
             )
             # The task arrives exactly as it would in production Buzz: a
             # user prompt @mentioning the orchestrator. The harness never
-            # speaks as any agent.
+            # speaks as any agent. The orchestrator is mentioned by pubkey,
+            # not by name resolution: task text is untrusted payload, and any
+            # @-token inside it (e.g. Vim's `:%normal! @a`) would otherwise
+            # fail member resolution and kill the trial before the agent
+            # ever saw the task. An explicit --mention demotes unresolved
+            # @-tokens in the text to presentation-only.
             await self._send(
-                trial.user, trial, f"@{orchestrator.agent_id} {instruction}"
+                trial.user,
+                trial,
+                f"@{orchestrator.agent_id} {instruction}",
+                mention=orchestrator.nostr_pubkey,
             )
             final_message = await asyncio.wait_for(
                 self._wait_for_done(environment, orchestrator, trial, agents + infra),
@@ -519,18 +527,24 @@ async def _verify_m1_output(
             )
 
     async def _send(
-        self, credential: AgentCredential, trial: TrialHandle, content: str
+        self,
+        credential: AgentCredential,
+        trial: TrialHandle,
+        content: str,
+        *,
+        mention: str | None = None,
     ) -> None:
-        await self._buzz_json(
-            credential,
-            trial,
+        args = [
             "messages",
             "send",
             "--channel",
             trial.channel_id,
             "--content",
             content,
-        )
+        ]
+        if mention is not None:
+            args += ["--mention", mention]
+        await self._buzz_json(credential, trial, *args)
 
     async def _buzz_json(
         self, credential: AgentCredential, trial: TrialHandle, *args: str
diff --git a/benchmarks/harbor-buzz-orchestra/tests/test_container_runtime.py b/benchmarks/harbor-buzz-orchestra/tests/test_container_runtime.py
index ebf0eb4b5d..5fc0e63e54 100644
--- a/benchmarks/harbor-buzz-orchestra/tests/test_container_runtime.py
+++ b/benchmarks/harbor-buzz-orchestra/tests/test_container_runtime.py
@@ -372,6 +372,37 @@ async def test_m1_output_probe_matches_grader_and_is_condition_scoped(
     assert bool(probed) == (condition == "M1-hello-world")
 
 
+async def test_send_mentions_by_pubkey_so_task_text_stays_inert(
+    tmp_path, monkeypatch
+):
+    """Task text is untrusted payload: `:%normal! @a` in a task statement must
+    not be fed to member-name resolution (it would fail and kill the trial).
+    An explicit --mention pins delivery to the orchestrator's pubkey."""
+    rt = runtime(tmp_path)
+    orch = credential("orch-1", "orchestrator", "orch-model")
+    trial = trial_handle((orch,))
+    calls = []
+
+    async def buzz_json(credential, trial, *args):
+        calls.append(args)
+        return {}
+
+    monkeypatch.setattr(rt, "_buzz_json", buzz_json)
+
+    await rt._send(
+        trial.user,
+        trial,
+        "@orch-1 run `:%normal! @a` on the file",
+        mention=orch.nostr_pubkey,
+    )
+    assert calls[-1][-2:] == ("--mention", "pubkey-orch-1")
+
+    # Without an explicit mention the send is unchanged (name resolution).
+    await rt._send(trial.user, trial, "plain content")
+    assert "--mention" not in calls[-1]
+    assert calls[-1][-2:] == ("--content", "plain content")
+
+
 async def test_wait_for_done_requires_orchestrator_authorship(tmp_path, monkeypatch):
     rt = runtime(tmp_path, poll_seconds=0)
     orch = credential("orch-1", "orchestrator", "orch-model")