Nice fix! A couple of suggestions from our implementation (#25991, closed in favor of this one) that proved important in testing:

1. 500-character minimum gate

Most false positives come from short conversational replies ("Based on the results, the best next" — 39 chars), not just emoji. These short responses can't possibly hit token limits, so they're almost certainly complete. Adding a length gate eliminates the vast majority of false positives before we even check the ending character.

In _should_treat_stop_as_truncated(), after the existing 20-char guard:

if len(visible_text) < 500:
    return False

2. Config opt-out

Some users may want to disable the heuristic entirely. A config flag with default true (current behavior) lets them opt out. Read in __init__ from agent.glm_truncation_heuristic, checked at the top of _should_treat_stop_as_truncated().

Your _get_terminal_response_char helper is solid — we used unicodedata.category() to strip Mn/Me/Cf which covers the same ZWJ/VS16/skin-tone cases plus any future combining marks. Functional equivalent though.

Happy to help with either addition if useful.

Superseded by #28324 (merged), which is a salvage of #28168 and covers emoji + caret as natural response endings.