diff --git a/src/server/terminal_service.rs b/src/server/terminal_service.rs
index fb6b4fd29..3029d095f 100644
--- a/src/server/terminal_service.rs
+++ b/src/server/terminal_service.rs
@@ -531,6 +531,91 @@ impl OutputBuffer {
     }
 }
 
+/// Find the largest prefix of `buf` that does not end in the middle of a UTF-8
+/// code point. Invalid bytes are treated as complete so they can continue
+/// downstream and be rendered with replacement characters if needed.
+fn find_utf8_split_point(buf: &[u8]) -> usize {
+    if buf.is_empty() {
+        return 0;
+    }
+
+    let start = buf.len().saturating_sub(3);
+    for i in (start..buf.len()).rev() {
+        let b = buf[i];
+        if b & 0x80 == 0 {
+            return buf.len();
+        }
+        if b & 0xC0 == 0x80 {
+            continue;
+        }
+
+        let seq_len = if b & 0xE0 == 0xC0 {
+            2
+        } else if b & 0xF0 == 0xE0 {
+            3
+        } else if b & 0xF8 == 0xF0 {
+            4
+        } else {
+            return buf.len();
+        };
+
+        return if buf.len() - i >= seq_len {
+            buf.len()
+        } else {
+            i
+        };
+    }
+
+    buf.len()
+}
+
+#[derive(Default)]
+struct Utf8ChunkAccumulator {
+    remainder: Vec<u8>,
+}
+
+impl Utf8ChunkAccumulator {
+    fn push_chunk(&mut self, mut data: Vec<u8>) -> Option<Vec<u8>> {
+        if data.is_empty() {
+            return None;
+        }
+
+        let had_remainder = !self.remainder.is_empty();
+        if had_remainder {
+            let mut combined = std::mem::take(&mut self.remainder);
+            combined.extend_from_slice(&data);
+            data = combined;
+        }
+
+        let split = find_utf8_split_point(&data);
+        if split == data.len() {
+            return Some(data);
+        }
+
+        // Only hold back a candidate incomplete suffix when we have evidence that
+        // the bytes before it are already UTF-8 text. If split is 0, the whole
+        // read may be the start of a UTF-8 character, so keep it for the next read.
+        if !had_remainder && split > 0 && std::str::from_utf8(&data[..split]).is_err() {
+            return Some(data);
+        }
+
+        self.remainder = data.split_off(split);
+        if data.is_empty() {
+            None
+        } else {
+            Some(data)
+        }
+    }
+
+    fn finish(&mut self) -> Option<Vec<u8>> {
+        if self.remainder.is_empty() {
+            None
+        } else {
+            Some(std::mem::take(&mut self.remainder))
+        }
+    }
+}
+
 /// Try to send data through the output channel with rate-limited drop logging.
 /// Returns `true` if the caller should break out of the read loop (channel disconnected).
 fn try_send_output(
@@ -1086,6 +1171,7 @@ impl TerminalServiceProxy {
         let reader_thread = thread::spawn(move || {
             let mut reader = reader;
             let mut buf = vec![0u8; 4096];
+            let mut utf8_chunks = Utf8ChunkAccumulator::default();
             let mut drop_count: u64 = 0;
             // Initialize to > 5s ago so the first drop triggers a warning immediately.
             let mut last_drop_warn = Instant::now() - Duration::from_secs(6);
@@ -1095,13 +1181,25 @@ impl TerminalServiceProxy {
                         // EOF
                         // This branch can be reached when the child process exits on macOS.
                         // But not on Linux and Windows in my tests.
+                        if let Some(data) = utf8_chunks.finish() {
+                            let _ = try_send_output(
+                                &output_tx,
+                                data,
+                                terminal_id,
+                                "",
+                                &mut drop_count,
+                                &mut last_drop_warn,
+                            );
+                        }
                         break;
                     }
                     Ok(n) => {
                         if exiting.load(Ordering::SeqCst) {
                             break;
                         }
-                        let data = buf[..n].to_vec();
+                        let Some(data) = utf8_chunks.push_chunk(buf[..n].to_vec()) else {
+                            continue;
+                        };
                         // Use try_send to avoid blocking the reader thread when channel is full.
                         // During disconnect, the run loop (sp.ok()) stops and read_outputs() is
                         // no longer called, so the channel won't be drained. Blocking send would
@@ -1308,12 +1406,23 @@ impl TerminalServiceProxy {
         let terminal_id = open.terminal_id;
         let reader_thread = thread::spawn(move || {
             let mut buf = vec![0u8; 4096];
+            let mut utf8_chunks = Utf8ChunkAccumulator::default();
             let mut drop_count: u64 = 0;
             // Initialize to > 5s ago so the first drop triggers a warning immediately.
             let mut last_drop_warn = Instant::now() - Duration::from_secs(6);
             loop {
                 match output_pipe.read(&mut buf) {
                     Ok(0) => {
+                        if let Some(data) = utf8_chunks.finish() {
+                            let _ = try_send_output(
+                                &output_tx,
+                                data,
+                                terminal_id,
+                                " (helper)",
+                                &mut drop_count,
+                                &mut last_drop_warn,
+                            );
+                        }
                         // EOF - helper process exited
                         log::debug!("Terminal {} helper output EOF", terminal_id);
                         break;
@@ -1322,7 +1431,9 @@ impl TerminalServiceProxy {
                         if exiting.load(Ordering::SeqCst) {
                             break;
                         }
-                        let data = buf[..n].to_vec();
+                        let Some(data) = utf8_chunks.push_chunk(buf[..n].to_vec()) else {
+                            continue;
+                        };
                         // Use try_send to avoid blocking the reader thread (same as direct PTY mode)
                         if try_send_output(
                             &output_tx,
@@ -1845,3 +1956,94 @@ impl TerminalServiceProxy {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::{find_utf8_split_point, Utf8ChunkAccumulator};
+
+    #[test]
+    fn utf8_split_point_returns_full_len_for_complete_input() {
+        assert_eq!(find_utf8_split_point(b"hello"), 5);
+        assert_eq!(find_utf8_split_point("中文".as_bytes()), "中文".len());
+        assert_eq!(find_utf8_split_point("😀".as_bytes()), "😀".len());
+    }
+
+    #[test]
+    fn utf8_split_point_detects_incomplete_trailing_sequence() {
+        let data = [b'a', 0xE4, 0xB8];
+        assert_eq!(find_utf8_split_point(&data), 1);
+    }
+
+    #[test]
+    fn utf8_split_point_keeps_malformed_prefix_but_buffers_trailing_lead_byte() {
+        let data = [0xFF, 0xE4];
+        assert_eq!(find_utf8_split_point(&data), 1);
+    }
+
+    #[test]
+    fn utf8_split_point_treats_orphan_continuations_as_complete() {
+        let data = [0x80, 0x81, 0x82];
+        assert_eq!(find_utf8_split_point(&data), data.len());
+    }
+
+    #[test]
+    fn utf8_chunk_accumulator_reassembles_split_multibyte_output() {
+        let full = "你好世界".as_bytes();
+        let mut chunker = Utf8ChunkAccumulator::default();
+        let mut output = Vec::new();
+
+        for chunk in full.chunks(5) {
+            if let Some(data) = chunker.push_chunk(chunk.to_vec()) {
+                output.extend_from_slice(&data);
+            }
+        }
+
+        if let Some(data) = chunker.finish() {
+            output.extend_from_slice(&data);
+        }
+
+        assert_eq!(output, full);
+    }
+
+    #[test]
+    fn utf8_chunk_accumulator_buffers_leading_split_multibyte_output() {
+        let mut chunker = Utf8ChunkAccumulator::default();
+
+        assert!(chunker.push_chunk(vec![0xE4]).is_none());
+        assert!(chunker.push_chunk(vec![0xB8]).is_none());
+        assert_eq!(
+            chunker.push_chunk(vec![0xAD]),
+            Some("中".as_bytes().to_vec())
+        );
+        assert!(chunker.finish().is_none());
+    }
+
+    #[test]
+    fn utf8_chunk_accumulator_flushes_incomplete_tail_on_finish() {
+        let mut chunker = Utf8ChunkAccumulator::default();
+        assert_eq!(chunker.push_chunk(vec![b'a', 0xE4]), Some(vec![b'a']));
+        assert_eq!(chunker.finish(), Some(vec![0xE4]));
+        assert!(chunker.finish().is_none());
+    }
+
+    #[test]
+    fn utf8_chunk_accumulator_does_not_stall_on_malformed_bytes() {
+        let mut chunker = Utf8ChunkAccumulator::default();
+        assert_eq!(chunker.push_chunk(vec![0xFF]), Some(vec![0xFF]));
+        assert!(chunker.finish().is_none());
+    }
+
+    #[test]
+    fn utf8_chunk_accumulator_buffers_lone_utf8_lead_bytes() {
+        let mut chunker = Utf8ChunkAccumulator::default();
+        assert!(chunker.push_chunk(vec![0xE4]).is_none());
+        assert_eq!(chunker.finish(), Some(vec![0xE4]));
+    }
+
+    #[test]
+    fn utf8_chunk_accumulator_does_not_hold_back_non_utf8_prefixes() {
+        let mut chunker = Utf8ChunkAccumulator::default();
+        assert_eq!(chunker.push_chunk(vec![0xFF, 0xE4]), Some(vec![0xFF, 0xE4]));
+        assert!(chunker.finish().is_none());
+    }
+}