diff --git a/crates/edit/src/buffer/mod.rs b/crates/edit/src/buffer/mod.rs
index 2019c9513a1..1ff899b095b 100644
--- a/crates/edit/src/buffer/mod.rs
+++ b/crates/edit/src/buffer/mod.rs
@@ -4190,4 +4190,45 @@ mod tests {
 
         assert_eq!(buffer_stored_bytes(&buf), b"ax\nbx\nx\n");
     }
+
+    // Replace All must not search the text again from the start after
+    // replacing a match that ends at the end of the text.
+    #[test]
+    fn test_replace_all_match_at_end_of_text() {
+        for (text, pattern, replacement, expected) in [
+            ("x foo", "foo", "Foo", "x Foo"),
+            ("x foo", "foo", "foo", "x foo"),
+            ("foo foo", "foo", "Foo", "Foo Foo"),
+            ("x a", "a", "aa", "x aa"),
+        ] {
+            let mut buf = prep_buffer(text);
+            buf.find_and_replace_all(pattern, Default::default(), replacement.as_bytes()).unwrap();
+            assert_eq!(buffer_stored_bytes(&buf), expected.as_bytes(), "{text:?}");
+        }
+    }
+
+    // Same as above, but with "insert final newline" adding a newline after the last match.
+    #[test]
+    fn test_replace_all_match_at_end_of_text_final_newline() {
+        let mut buf = prep_buffer("x a");
+        buf.set_insert_final_newline(true);
+
+        buf.find_and_replace_all("a", Default::default(), b"aa").unwrap();
+
+        assert_eq!(buffer_stored_bytes(&buf), b"x aa\n");
+    }
+
+    // After Replace hits a match that ends at the end of the text,
+    // the search must wrap around instead of finding the replaced text again.
+    #[test]
+    fn test_replace_match_at_end_of_text_wraps_around() {
+        let mut buf = prep_buffer("foo x foo");
+        buf.find_and_select("foo", Default::default()).unwrap();
+        buf.find_and_select("foo", Default::default()).unwrap();
+        buf.find_and_replace("foo", Default::default(), b"Foo").unwrap();
+
+        assert_eq!(buffer_stored_bytes(&buf), b"foo x Foo");
+        let (beg, end) = buf.selection_range().unwrap();
+        assert_eq!((beg.offset, end.offset), (0, 3));
+    }
 }
diff --git a/crates/edit/src/icu.rs b/crates/edit/src/icu.rs
index 5b34d7b06a8..c7aeb37b03b 100644
--- a/crates/edit/src/icu.rs
+++ b/crates/edit/src/icu.rs
@@ -401,6 +401,9 @@ extern "C" fn utext_native_length(ut: &mut icu_ffi::UText) -> i64 {
 }
 
 extern "C" fn utext_access(ut: &mut icu_ffi::UText, native_index: i64, forward: bool) -> bool {
+    let tb = text_buffer_from_utext(ut);
+    let native_index = native_index.clamp(0, tb.text_length() as i64);
+
     if let Some(cache) = utext_access_impl(ut, native_index, forward) {
         let native_off = native_index as usize - cache.utf8_range.start;
         ut.chunk_contents = cache.utf16.as_ptr();
@@ -411,6 +414,21 @@ extern "C" fn utext_access(ut: &mut icu_ffi::UText, native_index: i64, forward:
         ut.native_indexing_limit = cache.native_indexing_limit as i32;
         true
     } else {
+        // There's no text in the requested direction: `native_index` is at the end of the text
+        // (forward) or at its start (backward). The `UTextAccess` contract says that the iteration
+        // position must then be left at that end of the string. If we leave the previous chunk
+        // in place, ICU keeps iterating from wherever that chunk was (for instance `uregex_reset`
+        // would resume the search at offset 0 instead of the end of the text).
+        //
+        // An empty chunk at `native_index` does that. Since `native_indexing_limit` is 0, ICU won't
+        // call `utext_map_offset_to_native` or `utext_map_native_index_to_utf16` for this chunk.
+        let double_cache = double_cache_from_utext(ut);
+        ut.chunk_contents = double_cache.cache[0].utf16.as_ptr();
+        ut.chunk_length = 0;
+        ut.chunk_offset = 0;
+        ut.chunk_native_start = native_index;
+        ut.chunk_native_limit = native_index;
+        ut.native_indexing_limit = 0;
         false
     }
 }