diff --git a/tests/test-dirs/saved-parts.t b/tests/test-dirs/saved-parts.t
new file mode 100644
index 0000000000..c4d9f7cbdf
--- /dev/null
+++ b/tests/test-dirs/saved-parts.t
@@ -0,0 +1,53 @@
+# Small test
+
+A file with 2 type errors:
+- arg passed to "maybe_gen" should be unit, it is an int
+- return type of "maybe_gen" is int option, in a context where an int is expected
+
+  $ cat >foo.ml <<EOF
+  > let maybe_gen () = Some 0 in
+  > (maybe_gen 3) + 2
+  > EOF
+
+  $ $MERLIN single errors -filename foo.ml < foo.ml | grep "message" | wc -l
+  2
+
+The number of "saved parts" corresponds to the number of errors:
+
+  $ $MERLIN single dump -what browse -filename foo.ml < foo.ml | \
+  > grep "saved-parts" | wc -l
+  2
+
+
+# Bigger test
+
+Basically the same test with one level of nesting, we go from 2 to 4 errors:
+
+  $ cat >foo.ml <<EOF
+  > let maybe_gen () = Some 0 in
+  > (maybe_gen ((maybe_gen 3) + 2)) + 3
+  > EOF
+
+  $ $MERLIN single errors -filename foo.ml < foo.ml | grep "message" | wc -l
+  4
+
+And the number of saved parts stays in sync with the number of errors:
+
+  $ $MERLIN single dump -what browse -filename foo.ml < foo.ml | \
+  > grep "saved-parts" | wc -l
+  4
+
+And we could nest once more, and everything is fine and linear:
+
+  $ cat >foo.ml <<EOF
+  > let maybe_gen () = Some 0 in
+  > (maybe_gen ((maybe_gen ((maybe_gen 3) + 2)) + 3)) + 4
+  > EOF
+
+  $ $MERLIN single errors -filename foo.ml < foo.ml | grep "message" | wc -l
+  6
+
+  $ $MERLIN single dump -what browse -filename foo.ml < foo.ml | \
+  > grep "saved-parts" | wc -l
+  6
+