From 752c3a663cdf83aca33cd7021ed7f99cbfb448ae Mon Sep 17 00:00:00 2001
From: jasmine889966 <31839662+jasmine889966@users.noreply.github.com>
Date: Tue, 8 Sep 2026 09:10:31 +0800
Subject: [PATCH] fix(server): auto-title in the session's language, not the
 prompt's
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The titling instruction asked for a '4-5 word title' in English without
pinning an output language, so non-English sessions (zh/ja/ko/…) got
English titles. Tell the model to write the title in the same language
as the user's opening message, and treat the word count as an
English-scale hint only — the natural equivalent brevity elsewhere
(CJK titles are a few characters, not 4-5 words).

The small-talk sentinel stays the fixed English token it always was
(normalization already tolerates casing/punctuation riffs).

Tests: the title call's system prompt pins the same-language rule; a
CJK opener keeps a CJK title end to end (sanitizer whitespace/quote
handling and the 80-char cap are script-agnostic); the sentinel still
fires for a non-English small-talk opener and earns its single retry.
---
 coworker/server/manager.py | 17 ++++++++++++-----
 tests/test_autotitle.py    | 35 +++++++++++++++++++++++++++++++++++
 2 files changed, 47 insertions(+), 5 deletions(-)

diff --git a/coworker/server/manager.py b/coworker/server/manager.py
index cf9b740fa..266b0f8c3 100644
--- a/coworker/server/manager.py
+++ b/coworker/server/manager.py
@@ -5211,13 +5211,20 @@ def _extra_roots_of(
         ]
 
     # -- LLM auto-titles (FB-010) -------------------------------------------------
+    # The title must speak the user's language: with no language pinned, an English
+    # instruction ("4-5 word title") yields English titles for Chinese/Japanese/…
+    # sessions. Same-language also relaxes the word count to a length that is natural
+    # for the opener's script (CJK titles are a few characters, not 4-5 "words").
     _AUTOTITLE_PROMPT = (
         "You title chat sessions. Given the user's opening message(s) — and, when "
-        "present, the assistant's first reply for context — reply with ONLY a 4-5 word "
-        "title for the session, named after what the session is actually about — no "
-        "quotes or punctuation wrapping it. If there is no topic at all ("
-        '"hey", "how are you", "hi there" and a generic reply), reply with exactly: '
-        "small-talk"
+        "present, the assistant's first reply for context — reply with ONLY a short "
+        "title for the session, named after what the session is actually about. "
+        "Write the title in the same language as the user's opening message "
+        "(4-5 words for English; the natural equivalent brevity for other "
+        "languages) — no quotes or punctuation wrapping it. If there is no topic "
+        "at all ("
+        '"hey", "how are you", "hi there" and a generic reply), reply with '
+        "exactly: small-talk"
     )
 
     def _maybe_autotitle(self, session_id: str) -> None:
diff --git a/tests/test_autotitle.py b/tests/test_autotitle.py
index a6929ae6a..f722c2b71 100644
--- a/tests/test_autotitle.py
+++ b/tests/test_autotitle.py
@@ -173,3 +173,38 @@ async def go():
         assert mgr._autotitle_attempts[sid] == 1
 
     asyncio.run(go())
+
+
+async def test_title_call_pins_the_opener_language(tmp_path):
+    # With no language pinned in the titling instruction, an English system prompt
+    # ("4-5 word title") biases the model toward English titles even for non-English
+    # sessions. The prompt must tell the model to match the opener's language.
+    mgr, provider = _mgr(tmp_path, [_text("好的")], ["日本行程规划"])
+    await _turn(mgr, "lang1", "帮我规划一下日本的旅行")
+
+    system = provider.title_calls[0][0]["content"]
+    assert "same language as the user's opening message" in system
+
+
+async def test_non_english_opener_keeps_non_english_title(tmp_path):
+    # End to end: a CJK title survives the sanitizer (whitespace collapse and quote
+    # strip don't touch CJK) and the 80-char absurdity cap, and is stored verbatim.
+    mgr, _ = _mgr(tmp_path, [_text("好的，我来帮你安排")], ["日本行程规划"])
+    await _turn(mgr, "lang2", "帮我规划一下日本的旅行")
+    assert mgr.session_store.load("lang2").title == "日本行程规划"
+
+
+async def test_non_english_small_talk_still_hits_the_sentinel(tmp_path):
+    # The sentinel is a fixed English token regardless of the session's language;
+    # normalization must keep catching it after the prompt rewording.
+    mgr, provider = _mgr(
+        tmp_path,
+        [_text("你好呀"), _text("在的")],
+        ["small-talk", "季度汇报整理"],
+    )
+    await _turn(mgr, "lang3", "你好")
+    assert mgr.session_store.title_state("lang3")["auto_title"] is None
+
+    await _turn(mgr, "lang3", "帮我整理季度汇报")
+    assert len(provider.title_calls) == 2
+    assert mgr.session_store.load("lang3").title == "季度汇报整理"