diff --git a/tests/tools/test_file_operations_edge_cases.py b/tests/tools/test_file_operations_edge_cases.py
index ca5fa188f63c9..93d506c754aaa 100644
--- a/tests/tools/test_file_operations_edge_cases.py
+++ b/tests/tools/test_file_operations_edge_cases.py
@@ -51,6 +51,31 @@ def test_content_sample_longer_than_1000(self, ops):
         sample = "\x00" * 200 + "a" * 800 + "\x00" * 1000
         assert ops._is_likely_binary("file.xyz", content_sample=sample) is False
 
+    def test_single_trailing_replacement_is_sampling_artifact(self, ops):
+        """A lone trailing U+FFFD is a `head -c 1000` truncation artifact.
+
+        When the 1000-byte sample boundary splits a multi-byte UTF-8 char,
+        the terminal's errors=replace decode emits exactly one trailing
+        U+FFFD. That is not binary content and must not block reading a
+        valid UTF-8 text file (regression for #81480).
+        """
+        sample = "# Instru\u00e7\u00f5es de review \u2014 valida\u00e7\u00e3o\n" * 60 + "\ufffd"
+        assert sample.endswith("\ufffd")
+        assert sample.count("\ufffd") == 1
+        assert ops._is_likely_binary("notes.md", content_sample=sample) is False
+
+    def test_multiple_scattered_replacements_are_binary(self, ops):
+        """Genuine binary data produces many scattered U+FFFD chars."""
+        # Random non-UTF-8 bytes decoded with errors=replace
+        sample = "\ufffd" * 40 + "abc" * 100 + "\ufffd" * 40
+        assert ops._is_likely_binary("data.bin", content_sample=sample) is True
+
+    def test_trailing_replacement_with_other_garbage_is_binary(self, ops):
+        """A trailing U+FFFD plus any other replacement char is binary."""
+        sample = "text" + "\ufffd" + "more" + "\ufffd"
+        assert sample.count("\ufffd") == 2
+        assert ops._is_likely_binary("f.xyz", content_sample=sample) is True
+
 
 # =========================================================================
 # _check_lint edge cases
diff --git a/tools/file_operations.py b/tools/file_operations.py
index 2c1d1a62c9675..9c765a513569a 100644
--- a/tools/file_operations.py
+++ b/tools/file_operations.py
@@ -903,11 +903,22 @@ def _is_likely_binary(self, path: str, content_sample: str = None) -> bool:
             # sample carries the replacement char as binary (read-only) so the
             # agent can't corrupt it. Legitimate UTF-8 text effectively never
             # contains U+FFFD.
-            if "\ufffd" in content_sample[:1000]:
-                return True
-            non_printable = sum(1 for c in content_sample[:1000]
+            sample = content_sample[:1000]
+            # One exception to the U+FFFD rule: read_file samples with
+            # `head -c 1000`, which truncates at a byte boundary. When that
+            # boundary splits a multi-byte UTF-8 character, the dangling
+            # continuation bytes decode to exactly one *trailing* U+FFFD —
+            # a sampling artifact, not evidence of binary content. Genuine
+            # binary data yields many scattered replacement chars (well above
+            # the non-printable ratio below too), so a lone trailing U+FFFD
+            # is treated as text, not binary. (#81480)
+            if "\ufffd" in sample:
+                if sample.count("\ufffd") > 1 or not sample.endswith("\ufffd"):
+                    return True
+                return False
+            non_printable = sum(1 for c in sample
                                if ord(c) < 32 and c not in '\n\r\t')
-            return non_printable / min(len(content_sample), 1000) > 0.30
+            return non_printable / min(len(sample), 1000) > 0.30
         
         return False