fix: carry WebUI reply polish into gateway chat

CHANGELOG.md

@@ -5,6 +5,7 @@
 
 ### Fixed
 - `ctl.sh` now honors `HERMES_WEBUI_NO_DOTENV=1`, letting tests and scripted launches opt out of repo-local `.env` loading so host-specific `HERMES_WEBUI_STATE_DIR` values do not make the ctl test suite flaky (#3246).
+- Gateway-backed chat now carries the same WebUI final-answer polish guidance as the in-process chat paths, so terse scratchpad fragments such as "Need script" are not encouraged as visible assistant replies.
 
 ## [v0.51.192] — 2026-05-31 — Release FL (stage-batch4 — per-model context_length default-only guard)
 

api/gateway_chat.py

@@ -247,13 +247,17 @@ def _run_gateway_chat_streaming(
         cfg = get_config()
         try:
             from api.streaming import (
+                _WEBUI_PROGRESS_PROMPT,
                 _load_webui_prefill_context,
                 _prefill_messages_with_webui_context,
                 _public_prefill_context_status,
             )
 
             prefill_context = _load_webui_prefill_context(cfg)
-            prefill_messages = _prefill_messages_with_webui_context(prefill_context, cfg)
+            prefill_messages = [
+                {"role": "system", "content": _WEBUI_PROGRESS_PROMPT},
+                *_prefill_messages_with_webui_context(prefill_context, cfg),
+            ]
             put_gateway_event("context_status", {
                 "session_id": session_id,
                 "prefill": _public_prefill_context_status(prefill_context),

api/streaming.py

@@ -203,7 +203,8 @@ WebUI progress guidance:
 - Each update should say what you are about to check, what you just confirmed, or why the next tool call is needed.
 - Keep updates concise, factual, and in the user's language. One or two short sentences are enough.
 - Do not reveal hidden reasoning, chain-of-thought, private scratchpads, secrets, raw logs, or long tool output.
-- Do not include terse planning fragments or scratchpad shorthand in visible assistant text. Avoid fragments like "Need check logs", "Need inspect email", or "maybe invite"; either omit them or rewrite them as clear user-facing progress.
+- Final visible assistant replies must be clear user-facing English, not private planning notes.
+- Do not include terse planning fragments or scratchpad shorthand in visible assistant text. Avoid fragments like "Need script", "Need check logs", "Need inspect email", or "maybe invite"; either omit them or rewrite them as clear user-facing progress.
 - For direct answers or very short tasks, skip progress updates and answer normally.
 """.strip()
 

tests/test_webui_gateway_chat_backend.py

@@ -286,10 +286,14 @@ def test_gateway_chat_worker_translates_sse_and_persists_session(tmp_path, monke
     assert '"stream": true' in captured["body"]
     payload = json.loads(captured["body"])
     assert [m["content"] for m in payload["messages"]] == [
+        streaming._WEBUI_PROGRESS_PROMPT,
         "prefill",
         "webui session context",
         "Say hello",
     ]
+    assert payload["messages"][0]["role"] == "system"
+    assert "Final visible assistant replies" in payload["messages"][0]["content"]
+    assert "Need script" in payload["messages"][0]["content"]
     events = []
     while not subscriber.empty():
         events.append(subscriber.get_nowait())
@@ -361,7 +365,9 @@ def test_gateway_chat_worker_forwards_image_attachments_as_multimodal_parts(tmp_
     )
 
     content = captured["body"]["messages"][-1]["content"]
-    assert captured["body"]["messages"][0] == {"role": "user", "content": "webui session context"}
+    assert captured["body"]["messages"][0]["role"] == "system"
+    assert "Final visible assistant replies" in captured["body"]["messages"][0]["content"]
+    assert captured["body"]["messages"][1] == {"role": "user", "content": "webui session context"}
     assert content[0] == {"type": "text", "text": "What is in this image?"}
     assert content[1]["type"] == "image_url"
     assert content[1]["image_url"]["url"].startswith("data:image/png;base64,")

tests/test_webui_surface_context.py

@@ -26,6 +26,8 @@ def test_webui_ephemeral_prompt_includes_browser_surface_context():
     assert "explicit captures" in prompt
     assert "durable user preferences" in prompt
     assert "Do not include terse planning fragments" in prompt
+    assert "Final visible assistant replies" in prompt
+    assert "Need script" in prompt
     assert "Need inspect email" in prompt
     assert "clear user-facing progress" in prompt