zulip/zerver/lib/mention.py

42 lines
1.4 KiB
Python
Raw Normal View History

from typing import Optional, Set, Tuple
import re
# Match multi-word string between @** ** or match any one-word
# sequences after @
find_mentions = r'(?<![^\s\'\"\(,:<])(?P<silent>_?)@(?P<match>\*\*[^\*]+\*\*|all|everyone|stream)'
user_group_mentions = r'(?<![^\s\'\"\(,:<])@(\*[^\*]+\*)'
wildcards = ['all', 'everyone', 'stream']
def user_mention_matches_wildcard(mention: str) -> bool:
return mention in wildcards
def extract_mention_text(m: Tuple[str, str]) -> Optional[str]:
# re.findall provides tuples of match elements; we want the second
# to get the main mention content.
s = m[1]
if s.startswith("**") and s.endswith("**"):
text = s[2:-2]
if text in wildcards:
return None
return text
# We don't care about @all, @everyone or @stream
return None
def possible_mentions(content: str) -> Set[str]:
matches = re.findall(find_mentions, content)
# mention texts can either be names, or an extended name|id syntax.
texts_with_none = (extract_mention_text(match) for match in matches)
texts = {text for text in texts_with_none if text}
return texts
def extract_user_group(matched_text: str) -> str:
return matched_text[1:-1]
def possible_user_group_mentions(content: str) -> Set[str]:
matches = re.findall(user_group_mentions, content)
return {extract_user_group(match) for match in matches}