DEBUG trap confused if trap action contains a function definition while a command triggering the trap wants to eval-assign a compound variable

[jghub]

I have tried to reduce the issue to a small (if not quite minimal) example. consider this setup:

1. a script named dbg defined as

function help {
   :
}
while true; do
   read -r __cmd?"dbg> "
   case "$__cmd" in
      '')
         break
         ;;
      [h?])
         help;;
      l)
         set|awk '!/.*=/ {exit}; /^[a-z]/' 1>&2;;
      L)
         set|awk '!/.*=/ {exit}; /^[A-Z_]/' 1>&2;;
      Q)
         exit;;
      *)
         trap 'print "\n\"$__cmd\" interrupted!" 1>&2; : ' INT
         set|grep "^$__cmd=" 1>&2
         (($?)) && eval $__cmd 1>&2
         trap - INT
         ;;
      esac
done

2. a script (whatever name, let's call it dbgtest) defined as

trap '. dbg' DEBUG
a='(v1=1 v2=2)'
eval b=$a
:

dbgtest wants to define a compound variable b by evaluating a text representation a of that variable (this is stupid in this example but in actual case, text representation is read from a file etc.). dbg is intended as minimal step-through "debugger". for the problem at hand only the initial break case (break REPL upon CR) is relevant: running ksh dbgtest puts you in the REPL of dbg. upon second CR at the dbg> prompt an error results:

dbgtest[4]: eval[1]: .: dbg: line 1: b.help: invalid discipline function

the error is triggered by the very presence of a function definition preceding the REPL (help in the example). if that function definition is removed the error does not happen (you can hit l after each CR to see the intermediate state of defined variables where I note that intermittendely there are variables called b.v1 and b.v2 which apparently appear temporarily during the eval based construction of final compound variable b).

I do not use or know anything about discipline functions but I seem to get the fact that the error happens because the eval tries to access b.help which is indeed an undefined/non-existent discipline function -- however I do not see that this would not be a bug (or what am I missing?).

seen with Version AJM 93u+m/1.1.0-alpha+c222d956 2025-06-22