fix(chat): harden session auto-title and keep a single no-tool LLM call

- Add session_auto_title prompts, finalize_auto_title (reject prose, truncate, fallback).
- Third-round naming: one model.chat(system, user) with empty tools; reuse executor.model.
- Extend worker title tests for garbage model output.

Co-authored-by: Cursor <cursoragent@cursor.com>

tests/test_session_auto_title.py

@@ -0,0 +1,27 @@
+from __future__ import annotations
+
+from oclaw.runtime.session_auto_title import AUTO_TITLE_CHAR_MAX, finalize_auto_title, should_reject_auto_title
+
+
+def test_should_reject_long_prose() -> None:
+    blob = "您好！我是 Claude Code，一个软件工程助手。" + "x" * 80
+    assert should_reject_auto_title(blob) is True
+
+
+def test_should_reject_claude_code_intro() -> None:
+    assert should_reject_auto_title("Hello! I am Claude Code, a software engineering assistant.") is True
+
+
+def test_should_accept_short_title() -> None:
+    assert should_reject_auto_title("告警查询配置") is False
+
+
+def test_finalize_truncates_fallback() -> None:
+    long_fb = "一二三四五六七八九十11121314151617181920"
+    out = finalize_auto_title(raw="x" * 100, fallback=long_fb)
+    assert len(out) == AUTO_TITLE_CHAR_MAX
+    assert out == long_fb[:AUTO_TITLE_CHAR_MAX]
+
+
+def test_finalize_uses_model_when_sane() -> None:
+    assert finalize_auto_title(raw="  网管同步  ", fallback="fallback") == "网管同步"