ct-public/isparkclaw-webui
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: remove hardcoded chinese title heuristics (#887)

Browse Source 
* fix: remove hardcoded chinese title heuristics

* fix: use english placeholder for non-latin fallback titles
This commit is contained in:
Pavol Biely
2026-04-23 20:45:34 +04:00
committed by GitHub
parent ae7be6deba
commit 96c97c5e0e
3 changed files with 74 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
api/streaming.py
View File

@@ -106,7 +106,7 @@ def _sanitize_generated_title(text: str) -> str:
"""Sanitize LLM-generated title text before persisting to session."""
s = _strip_thinking_markup(text or '')
s = re.sub(
r'^\s*(?:[*_`~]+\s*)?(?:session\s+title|title)\s*[:：]\s*(?:[*_`~]+\s*)?',
r'^\s*(?:[*_`~]+\s*)?(?:session\s+title|title)\s*:\s*(?:[*_`~]+\s*)?',
'',
s,
flags=re.IGNORECASE,
@@ -132,10 +132,7 @@ def _looks_invalid_generated_title(text: str) -> bool:
or re.search(r'^\s*(i|we)\s+(should|need to|will|can)\b', s, flags=re.IGNORECASE)
or re.search(r'^\s*let me\b', s, flags=re.IGNORECASE)
or re.search(r"^\s*here(?:'s| is) (?:a |my )?(?:thinking|thought)", s, flags=re.IGNORECASE)
or re.search(r'用户(要求|希望|想让|让我)', s)
or re.search(r'请只?回复', s)
or re.search(r'^\s*(ok|okay|done|all set|complete|completed|finished)\b[\s.!?]*$', s, flags=re.IGNORECASE)
or re.search(r'^\s*(好的|好啦|完成了|已完成|测试完成|测试已完成|可以了|没问题)\s*[！!。\.\s]*$', s)
)
@@ -209,10 +206,10 @@ def _title_prompts(user_text: str, assistant_text: str) -> tuple[str, list[str]]
"Return only the title text, 3-8 words, as a topic label.\n"
"Do not use markdown, bullets, labels, or prefixes like Session Title:.\n"
"Do not output a full sentence.\n"
"Do not output acknowledgements or completion phrases like OK, done, all set, 测试完成.\n"
"Do not output acknowledgements or completion phrases like OK, done, or all set.\n"
"Do not describe internal reasoning.\n"
"Bad: The user is asking..., OK, 好的，测试完成！\n"
"Good: 自动标题生成测试, Clarify Dialog Layout, GitHub Issue Triage"
"Bad: The user is asking..., OK, all set.\n"
"Good: Title Generation Test, Clarify Dialog Layout, GitHub Issue Triage"
),
(
"Rewrite this conversation start as a concise noun-phrase title.\n"
@@ -437,10 +434,10 @@ def _fallback_title_from_exchange(user_text: str, assistant_text: str) -> Option
combined = f"{user_text} {assistant_text}".strip().lower()
combined_raw = f"{user_text} {assistant_text}".strip()
def _contains_latin(text: str) -> bool:
return bool(re.search(r'[A-Za-z]', text or ''))
def _extract_named_topic(text: str) -> str:
m = re.search(r'《([^》]{2,24})》', text)
if m:
return (m.group(1) or '').strip()
m = re.search(r'"([^"\n]{2,24})"', text)
if m:
return (m.group(1) or '').strip()
@@ -451,57 +448,53 @@ def _fallback_title_from_exchange(user_text: str, assistant_text: str) -> Option
topic_name = _extract_named_topic(combined_raw)
if topic_name:
if any(k in combined for k in ('时间', 'time', '安排', '效率', '怎么办', '健身', '唱歌', '写毛笔', '不够用了')):
return f'{topic_name}与时间管理'
if not _contains_latin(topic_name):
if any(k in combined for k in ('time', 'schedule', 'efficiency', 'manage', 'fitness', 'singing', 'calligraphy')):
return 'Time management discussion'
if any(k in combined for k in ('hermes', 'codex', 'ai')):
return 'AI productivity discussion'
return 'Conversation topic'
if any(k in combined for k in ('time', 'schedule', 'efficiency', 'manage', 'fitness', 'singing', 'calligraphy')):
return f'{topic_name} time management'
if any(k in combined for k in ('hermes', 'codex', 'ai')):
return f'{topic_name}与AI效率'
return f'{topic_name}讨论'
return f'{topic_name} AI productivity'
return f'{topic_name} discussion'
if any(k in combined for k in ('title', '标题')) and any(k in combined for k in ('summary', 'summar', '摘要', '短标题')):
if any(k in combined for k in ('test', '测试', 'ok', '回复ok')):
return '会话标题自动摘要测试'
return '会话标题自动摘要'
if any(k in combined for k in ('clarify', '澄清')) and any(k in combined for k in ('dialog', 'card', '对话', '卡片')):
return 'Clarify 对话卡片'
if any(k in combined for k in ('issue', 'github', 'pr')) and any(k in combined for k in ('triage', 'bug', 'review', '问题')):
if any(k in combined for k in ('title', 'session title')) and any(k in combined for k in ('summary', 'summar', 'short title')):
if any(k in combined for k in ('test', 'ok', 'reply ok')):
return 'Session title auto-summary test'
return 'Session title auto-summary'
if any(k in combined for k in ('clarify', 'clarification')) and any(k in combined for k in ('dialog', 'card')):
return 'Clarify dialog card'
if any(k in combined for k in ('issue', 'github', 'pr')) and any(k in combined for k in ('triage', 'bug', 'review')):
return 'GitHub Issue Triage'
head = re.split(r'[。！？.!?\n]', user_text)[0].strip()
head = re.split(r'[.!?\n]', user_text)[0].strip()
if not head:
return None
stop_cjk = {
'我们', '看看', '一下', '这个', '标题', '是否', '可以', '用户', '理解', '这里', '测试', '一下',
'你只', '需要', '回复', '就可', '可以', '不需', '需要做', '什么', '自动', '成用户', '短标题',
}
stop_en = {
'the', 'this', 'that', 'with', 'from', 'into', 'just', 'reply', 'please',
'need', 'needs', 'want', 'wants', 'user', 'assistant', 'could', 'would',
'should', 'about', 'there', 'here', 'test', 'testing', 'title', 'summary',
}
tokens = re.findall(r'[\u4e00-\u9fff]{2,6}|[A-Za-z0-9][A-Za-z0-9_./+-]*', head)
tokens = re.findall(r'[A-Za-z0-9][A-Za-z0-9_./+-]*', head)
if not tokens:
return head[:64]
return 'Conversation topic'
picked = []
for tok in tokens:
lower_tok = tok.lower()
if re.search(r'[\u4e00-\u9fff]', tok):
if tok in stop_cjk:
continue
else:
if lower_tok in stop_en or len(lower_tok) < 3:
continue
if lower_tok in stop_en or len(lower_tok) < 3:
continue
if tok not in picked:
picked.append(tok)
if len(picked) >= 4:
break
if picked:
if any(re.search(r'[\u4e00-\u9fff]', t) for t in picked):
return ''.join(picked)[:20]
return ' '.join(picked)[:60]
return head[:24]
return 'Conversation topic'
def _run_background_title_update(session_id: str, user_text: str, assistant_text: str, placeholder_title: str, put_event, agent=None):

15
tests/test_sprint41.py
View File

@@ -137,16 +137,21 @@ class TestIssue495TitleStreaming(unittest.TestCase):
)
def test_streaming_rejects_generic_completion_titles(self):
self.assertIn(
"测试完成",
STREAMING_PY,
"streaming.py should reject generic completion phrases as session titles",
)
self.assertIn(
"all set",
STREAMING_PY,
"streaming.py should reject generic English completion phrases as session titles",
)
self.assertIn(
"completed",
STREAMING_PY,
"streaming.py should reject completion-status titles as session titles",
)
self.assertNotIn(
"测试完成",
STREAMING_PY,
"streaming.py title generation should stay English-only",
)
def test_streaming_uses_reasoning_split_for_minimax_titles(self):
self.assertIn(

35
tests/test_title_sanitization.py
View File

@@ -1,6 +1,11 @@
import unittest
from pathlib import Path
from api.streaming import _first_exchange_snippets, _sanitize_generated_title
from api.streaming import (
_fallback_title_from_exchange,
_first_exchange_snippets,
_sanitize_generated_title,
)
class TestGeneratedTitleSanitization(unittest.TestCase):
@@ -33,3 +38,31 @@ class TestGeneratedTitleSanitization(unittest.TestCase):
_first_exchange_snippets(messages),
("What time is it in San Francisco?", "It is 6:16 PM in San Francisco."),
)
def test_fallback_title_uses_english_discussion_suffix(self):
self.assertEqual(
_fallback_title_from_exchange('Please review "random cancel"', ""),
"random cancel discussion",
)
def test_fallback_title_summary_label_is_english(self):
self.assertEqual(
_fallback_title_from_exchange("Generate a short title summary test", ""),
"Session title auto-summary test",
)
def test_fallback_title_non_latin_input_uses_english_placeholder(self):
self.assertEqual(
_fallback_title_from_exchange("讨论一下这个问题", ""),
"Conversation topic",
)
def test_fallback_title_non_latin_quoted_topic_uses_english_placeholder(self):
self.assertEqual(
_fallback_title_from_exchange('Please review "讨论主题"', ""),
"Conversation topic",
)
def test_title_generation_source_has_no_cjk_literals(self):
src = Path("api/streaming.py").read_text(encoding="utf-8")
self.assertNotRegex(src, r"[\u4e00-\u9fff]", "title generation code should stay English-only")