emrg: the batch-call reminder is a system turn, not the host's voice

emrg/server/daemon.py

@@ -714,6 +714,16 @@ def _with_batch_call_nudge(
     states it at the tail of the next request, where the next round's calls are
     actually decided.
 
+    It is a **system** message, not a user one. The content is what the model
+    reads; the role is what every *other* reader reads. A user turn is the
+    host's voice: it is the role a client renders as the host's bubble, the role
+    a transcript attributes to the host, and the role a compaction summarises as
+    something the host said. This reminder is the daemon's instruction to the
+    model, so a user role tells every one of those readers that the host said
+    something the host never typed (reported 2026-11-03). A trailing system
+    message is accepted where it is needed: measured against the configured
+    provider, `{"role": "system"}` after the user turn answers normally.
+
     A **new list** is returned and `messages` is never touched, and that is the
     whole of "the reminder exists only in this request": the loop's own list is
     what `_log_llm_exchange` writes and what a rebuild is rebuilt from, so a
@@ -727,7 +737,7 @@ def _with_batch_call_nudge(
     """
     request = [dict(m) for m in messages]
     if previous_round_tool_calls == 1:
-        request.append({"role": "user", "content": BATCH_CALL_NUDGE})
+        request.append({"role": "system", "content": BATCH_CALL_NUDGE})
     return request
 
 

tests/test_batch_call_nudge.py

@@ -6,8 +6,11 @@
 requirement stated in the prompt sits too far from the point where the next
 round's calls are decided.
 
-So the daemon appends `{"role": "user", "content": "batch call"}` to the
-**request** that follows a round dispatching exactly one call. The readings here
+So the daemon appends `{"role": "system", "content": "batch call"}` to the
+**request** that follows a round dispatching exactly one call. A **system** turn
+and not a user one: the reminder is the daemon's instruction, while a user turn
+is the host's voice — the role a client renders as the host's bubble and a
+transcript attributes to the host. The readings here
 are the `messages` lists the provider was actually handed — captured where the
 stream is consumed — so nothing asserts the shape of the code that built them:
 removing the append, inverting its condition, or moving it ahead of the tool
@@ -36,7 +39,7 @@
 from emrg.session import Session
 
 SESSION_ID = "batch-call-nudge"
-NUDGE = {"role": "user", "content": BATCH_CALL_NUDGE}
+NUDGE = {"role": "system", "content": BATCH_CALL_NUDGE}
 
 
 class _StubTool:
@@ -124,7 +127,7 @@ async def fake_summary(session, records, source=None):
 
 
 def _nudges(messages: list[dict]) -> int:
-    return sum(1 for m in messages if m.get("role") == "user"
+    return sum(1 for m in messages if m.get("role") == "system"
                and m.get("content") == BATCH_CALL_NUDGE)