diff --git a/CHANGELOG.md b/CHANGELOG.md
index 534142f5a..2fe67b9be 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -14,6 +14,8 @@ metadata and the backend fallback mirror it.
 
 ### Fixed
 
+- Call replies use sentence boundaries after their first short utterance (#2628) — thanks @rudycelekli!
+
 - MCP speech tools wait through model loading and progress-extended CPU renders instead of timing out before the backend (#2609)
 
 ## [0.5.7] — 2026-10-05
diff --git a/backend/services/sentence_chunker.py b/backend/services/sentence_chunker.py
index 10e9c1181..cd2325d2b 100644
--- a/backend/services/sentence_chunker.py
+++ b/backend/services/sentence_chunker.py
@@ -478,6 +478,7 @@ def _maybe_short_flush(self) -> list[str]:
                 return []
 
         self._buffer = ""
+        self._is_first_flush = False
         return [stripped]
 
     def _maybe_aggressive_first_flush(self) -> str | None:
diff --git a/docs/integrations/calls.md b/docs/integrations/calls.md
index e593ae868..f227a72b9 100644
--- a/docs/integrations/calls.md
+++ b/docs/integrations/calls.md
@@ -196,3 +196,5 @@ turn the other person talked over.
   resampled to 8 kHz μ-law and sent as 20 ms frames. On barge-in (about 200 ms
   of the other person's speech while the agent talks), VoiceStudio sends Twilio
   a `clear` to drop buffered audio.
+
+Call speech may start at a clause boundary for the first utterance of each turn. After a complete short reply has played, later speech waits for normal sentence boundaries.
diff --git a/tests/test_sentence_chunker.py b/tests/test_sentence_chunker.py
index b1a92e70e..34f482df9 100644
--- a/tests/test_sentence_chunker.py
+++ b/tests/test_sentence_chunker.py
@@ -69,3 +69,19 @@ def test_single_sentence_request_stays_whole():
 def test_italian_language_disables_aggressive_flush():
     chunker = SentenceChunker(language="it", aggressive_first_flush=True)
     assert chunker._aggressive_first_flush is False
+
+
+def test_short_flush_retires_first_clause_window_until_next_turn():
+    chunker = SentenceChunker(aggressive_first_flush=True)
+    assert chunker.push("Hello!") == ["Hello!"]
+    second = "Here is a longer follow-up with more information, "
+    assert chunker.push(second) == []
+    assert chunker.flush() == [second.strip()]
+    assert chunker.push(second) == [second.strip()]
+
+
+def test_short_flush_disabled_threshold_keeps_first_clause_window():
+    chunker = SentenceChunker(aggressive_first_flush=True, min_words_for_short_flush=2)
+    assert chunker.push("Hello!") == []
+    rest = " Here is a longer follow-up with more information, "
+    assert chunker.push(rest) == [("Hello!" + rest).strip()]