diff --git a/crates/kernel/src/lib.rs b/crates/kernel/src/lib.rs
index 23dc4cee..c4cc9797 100644
--- a/crates/kernel/src/lib.rs
+++ b/crates/kernel/src/lib.rs
@@ -50,6 +50,7 @@ pub mod proactive;
 pub mod queue;
 pub mod schedule;
 pub mod security;
+pub mod sentence_buffer;
 pub mod session;
 pub(crate) mod syscall;
 pub mod task_report;
diff --git a/crates/kernel/src/sentence_buffer.rs b/crates/kernel/src/sentence_buffer.rs
new file mode 100644
index 00000000..a09b367f
--- /dev/null
+++ b/crates/kernel/src/sentence_buffer.rs
@@ -0,0 +1,154 @@
+//! Accumulates streaming text deltas and emits complete sentences.
+//!
+//! Used by channel adapters that need sentence-level chunking for TTS
+//! synthesis — feeding one sentence at a time produces natural prosody
+//! while keeping latency low.
+
+/// Accumulates text and emits complete sentences split on sentence-ending
+/// punctuation (`。！？.!?\n`).
+///
+/// No async, no I/O — pure text segmentation. Designed to sit between an
+/// LLM `TextDelta` stream and a TTS synthesizer.
+#[derive(Debug, Default)]
+pub struct SentenceBuffer {
+    buf: String,
+}
+
+/// Characters that terminate a sentence.
+const SENTENCE_ENDS: &[char] = &['。', '！', '？', '.', '!', '?', '\n'];
+
+impl SentenceBuffer {
+    /// Create an empty buffer.
+    pub fn new() -> Self { Self::default() }
+
+    /// Push a text delta. Returns any complete sentences that were formed.
+    ///
+    /// A sentence is delimited by any sentence-ending character
+    /// (`。！？.!?\n`). The
+    /// delimiter is included in the returned sentence. Consecutive delimiters
+    /// (e.g. `"?!"`) are collapsed into one sentence.
+    pub fn push(&mut self, delta: &str) -> Vec<String> {
+        self.buf.push_str(delta);
+        self.drain_sentences()
+    }
+
+    /// Drain any remaining text that hasn't been terminated by a sentence
+    /// delimiter. Call this when the LLM turn ends.
+    pub fn flush(&mut self) -> Option<String> {
+        let rest = std::mem::take(&mut self.buf);
+        let trimmed = rest.trim().to_owned();
+        if trimmed.is_empty() {
+            None
+        } else {
+            Some(trimmed)
+        }
+    }
+
+    /// Returns `true` if the buffer has no accumulated text.
+    pub fn is_empty(&self) -> bool { self.buf.trim().is_empty() }
+
+    fn drain_sentences(&mut self) -> Vec<String> {
+        let mut sentences = Vec::new();
+
+        loop {
+            let pos = self.buf.find(SENTENCE_ENDS);
+            let Some(byte_pos) = pos else { break };
+
+            // Include the delimiter character.
+            let end = byte_pos + self.buf[byte_pos..].chars().next().unwrap().len_utf8();
+
+            // Skip consecutive delimiters (e.g. "?!" or "。\n").
+            let mut scan = end;
+            for ch in self.buf[end..].chars() {
+                if SENTENCE_ENDS.contains(&ch) {
+                    scan += ch.len_utf8();
+                } else {
+                    break;
+                }
+            }
+
+            let sentence = self.buf[..scan].trim().to_owned();
+            self.buf = self.buf[scan..].to_owned();
+
+            if !sentence.is_empty() {
+                sentences.push(sentence);
+            }
+        }
+
+        sentences
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn single_sentence() {
+        let mut buf = SentenceBuffer::new();
+        assert!(buf.push("こんに").is_empty());
+        assert!(buf.push("ちは").is_empty());
+        let out = buf.push("。");
+        assert_eq!(out, vec!["こんにちは。"]);
+        assert!(buf.is_empty());
+    }
+
+    #[test]
+    fn multiple_sentences_in_one_push() {
+        let mut buf = SentenceBuffer::new();
+        let out = buf.push("Hello. How are you? Fine!");
+        assert_eq!(out, vec!["Hello.", "How are you?", "Fine!"]);
+    }
+
+    #[test]
+    fn flush_remaining() {
+        let mut buf = SentenceBuffer::new();
+        buf.push("no terminator");
+        assert_eq!(buf.flush(), Some("no terminator".to_owned()));
+        assert!(buf.is_empty());
+    }
+
+    #[test]
+    fn flush_empty() {
+        let mut buf = SentenceBuffer::new();
+        assert_eq!(buf.flush(), None);
+    }
+
+    #[test]
+    fn consecutive_delimiters_collapsed() {
+        let mut buf = SentenceBuffer::new();
+        let out = buf.push("Really?! Yes.");
+        assert_eq!(out, vec!["Really?!", "Yes."]);
+    }
+
+    #[test]
+    fn newline_as_delimiter() {
+        let mut buf = SentenceBuffer::new();
+        let out = buf.push("Line one\nLine two\n");
+        assert_eq!(out, vec!["Line one", "Line two"]);
+    }
+
+    #[test]
+    fn chinese_mixed_punctuation() {
+        let mut buf = SentenceBuffer::new();
+        let out = buf.push("今天天气真好。明天呢？不知道！");
+        assert_eq!(out, vec!["今天天气真好。", "明天呢？", "不知道！"]);
+    }
+
+    #[test]
+    fn incremental_deltas() {
+        let mut buf = SentenceBuffer::new();
+        assert!(buf.push("I am ").is_empty());
+        assert!(buf.push("fine").is_empty());
+        let out = buf.push(". Thank you!");
+        assert_eq!(out, vec!["I am fine.", "Thank you!"]);
+    }
+
+    #[test]
+    fn trailing_text_after_sentence() {
+        let mut buf = SentenceBuffer::new();
+        let out = buf.push("Done. Now");
+        assert_eq!(out, vec!["Done."]);
+        assert_eq!(buf.flush(), Some("Now".to_owned()));
+    }
+}