Group of 3 You'll notice
Several heuristics that judge a subagent reply's meaning now understand Chinese, Japanese, and Korean text, not just English
What
A set of internal text-classification checks that Claude Code uses to interpret short or automated replies from subagents (helper AI processes that do sub-tasks) were extended to handle East Asian languages:
- A new script-detection helper recognizes Han (Chinese/Japanese kanji), Hiragana/Katakana, Thai/Lao/Khmer/Myanmar, and Hangul (Korean) characters, and uses this to compute a more accurate word count for text in these scripts. The "is this reply just saying done" check now also matches Japanese/Chinese/Korean completion phrases, not only the literal word "done".
- The check that decides a subagent reply means "nothing to report" (silence, or no suggestion) now also matches Chinese, Japanese, and Korean equivalents (for example 沈黙, 提案なし, 제안 없음), and it now recognizes fullwidth and Japanese-style brackets (（）［］【】〔〕) as well as ASCII ones.
- The check that decides whether a very short reply counts as a real yes/no-style answer now also counts Han, phonetic, and Hangul characters, treating a single such character specially, alongside the existing English yes/no word list.
Why
Without these changes, a subagent replying in Chinese, Japanese, or Korean could have its "done", "nothing to report", or short-answer replies misclassified simply because the matching logic only understood English words and ASCII punctuation, leading to incorrect handling of otherwise normal responses.