diff --git a/src/common.cpp b/src/common.cpp
index ac2492805..d7182e6b6 100644
--- a/src/common.cpp
+++ b/src/common.cpp
@@ -1363,7 +1363,11 @@ static bool unescape_string_internal(const wchar_t *const input, const size_t in
                 }
                 case L'}': {
                     if (unescape_special) {
-                        assert(brace_count > 0 && "imbalanced brackets are a tokenizer error, we shouldn't be able to get here");
+                        // HACK: The completion machinery sometimes hands us partial tokens.
+                        // We can't parse them properly, but it shouldn't hurt,
+                        // so we don't assert here.
+                        // See #4954.
+                        // assert(brace_count > 0 && "imbalanced brackets are a tokenizer error, we shouldn't be able to get here");
                         brace_count--;
                         brace_text_start = brace_text_start && brace_count > 0;
                         to_append_or_none = BRACE_END;
diff --git a/src/complete.cpp b/src/complete.cpp
index 1df55e5a6..c256a0558 100644
--- a/src/complete.cpp
+++ b/src/complete.cpp
@@ -1078,6 +1078,9 @@ void completer_t::complete_param_expand(const wcstring &str, bool do_file,
     bool complete_from_start = !complete_from_separator || !string_prefixes_string(L"-", str);
 
     if (complete_from_separator) {
+        // FIXME: This just cuts the token,
+        // so any quoting or braces gets lost.
+        // See #4954.
         const wcstring sep_string = wcstring(str, sep_index + 1);
         std::vector<completion_t> local_completions;
         if (expand_string(sep_string, &local_completions, flags, NULL) == EXPAND_ERROR) {