fix: retain Responses assistant replies in RPC session history

Carve the history-only implementation and regression from #104754
(b4bfa76facb43111a863b1256c89875e3f01fde0) by PLASMA-FR; omit
the unrelated locale-picker changes. Complement merged #104523.

Real serve + WebSocket reconnect: SQLite retains two rows; before,
resume/activate/history return only the user; after, both survive.
Plain-content control retains both rows in both arms. Native macOS
sleep and the full desktop symptom are not established by this probe.

Refs #68321

tests/tui_gateway/test_session_history_codex_sidecar.py

@@ -0,0 +1,54 @@
+"""RPC history must retain visible Responses-API assistant sidecars (#68321)."""
+
+from __future__ import annotations
+
+import threading
+
+from hermes_state import SessionDB
+import tui_gateway.server as server
+
+
+def test_session_history_preserves_codex_message_items(tmp_path):
+    message_items = [
+        {
+            "type": "message",
+            "role": "assistant",
+            "phase": "final_answer",
+            "content": [{"type": "output_text", "text": "Persisted answer"}],
+        }
+    ]
+    db = SessionDB(db_path=tmp_path / "state.db")
+    db.create_session("stored-session", source="desktop")
+    db.append_message(
+        "stored-session",
+        "assistant",
+        "",
+        codex_message_items=message_items,
+    )
+    previous_db = server._db
+    setattr(server, "_db", db)
+    server._sessions["runtime-session"] = {
+        "session_key": "stored-session",
+        "history": [],
+        "history_lock": threading.Lock(),
+        "running": False,
+        "agent": None,
+    }
+
+    try:
+        response = server.handle_request(
+            {"id": "1", "method": "session.history", "params": {"session_id": "runtime-session"}}
+        )
+    finally:
+        server._sessions.pop("runtime-session", None)
+        setattr(server, "_db", previous_db)
+        db.close()
+
+    assert isinstance(response, dict)
+    assert "error" not in response
+    assert response["result"]["count"] == 1
+    messages = response["result"]["messages"]
+    assert len(messages) == 1
+    assert messages[0]["role"] == "assistant"
+    assert messages[0]["text"] == ""
+    assert messages[0]["codex_message_items"] == message_items

tui_gateway/session_history.py

@@ -167,7 +167,13 @@ def _legacy_display_kind(role: str, text: str) -> str | None:
     return "auto_continue" if role == "user" and text.lstrip().startswith(_AUTO_CONTINUE_NOTE_PREFIX) else None
 
 
-_HISTORY_REASONING_KEYS = ("reasoning", "reasoning_content", "reasoning_details", "codex_reasoning_items")
+_HISTORY_ASSISTANT_DETAIL_KEYS = (
+    "reasoning",
+    "reasoning_content",
+    "reasoning_details",
+    "codex_reasoning_items",
+    "codex_message_items",
+)
 _HISTORY_ROLES = frozenset({"user", "assistant", "tool", "system"})
 
 
@@ -205,9 +211,9 @@ def _history_to_messages(history: list[dict]) -> list[dict]:
             # `context` is an 80-char preview; ship args so a full-call renderer isn't truncated.
             messages.append({"role": "tool", "name": name, "context": _tool_ctx(name, args), **({"args": args} if args else {})})
             continue
-        # A reasoning-only assistant turn is kept so "Thinking…" still shows after resume/reload.
-        has_reasoning = role == "assistant" and any(m.get(key) for key in _HISTORY_REASONING_KEYS)
-        if not content_text.strip() and not has_reasoning:
+        # Assistant detail sidecars can carry the only visible reply or reasoning after resume/reload.
+        has_assistant_detail = role == "assistant" and any(m.get(key) for key in _HISTORY_ASSISTANT_DETAIL_KEYS)
+        if not content_text.strip() and not has_assistant_detail:
             continue
         msg = {"role": role, "text": content_text}
         # Authoring time (Unix seconds) for display.timestamps; display-only.
@@ -223,7 +229,7 @@ def _history_to_messages(history: list[dict]) -> list[dict]:
         if invocation:
             msg.update(text=invocation, display_kind="skill_invocation")
         if role == "assistant":
-            msg.update((key, m[key]) for key in _HISTORY_REASONING_KEYS if m.get(key) is not None)
+            msg.update((key, m[key]) for key in _HISTORY_ASSISTANT_DETAIL_KEYS if m.get(key) is not None)
         # Display-only timeline metadata (model switches, delegation events).
         display_kind = m.get("display_kind") or _legacy_display_kind(role, content_text)
         if display_kind:

website/docs/developer-guide/session-storage.md

@@ -138,6 +138,7 @@ CREATE INDEX IF NOT EXISTS idx_messages_session_id ON messages(session_id, id);
 Notes:
 - `tool_calls` is stored as a JSON string (serialized list of tool call objects)
 - `reasoning_details`, `codex_reasoning_items`, and `codex_message_items` are stored as JSON strings
+- Desktop history hydration retains assistant sidecars in both REST and JSON-RPC (`session.resume`, `session.activate`, `session.history`) projections. A Responses reply may live only in `codex_message_items` while `content` is empty; that is not an empty transcript row.
 - `reasoning` stores the raw reasoning text for providers that expose it
 - `api_content` is a byte-fidelity sidecar: the exact content string sent to the API for this message when it differs from `content` (ephemeral memory/plugin injections, persist overrides). It preserves the wire bytes for prompt-cache-stable replay — stored as sent, except lone surrogates, which sqlite3 cannot bind and which the conversation loop scrubs from every outgoing payload anyway. `NULL` means `content` was sent verbatim.
 - Timestamps are Unix epoch floats (`time.time()`)