Does your format ressemble XML enough to use an XML parser? That would be best. The remainder of the post assumes you can't use an XML parser.

If I follow correctly, you're using the lookahead so that the < of <div> doesn't get absorbed by the regexp. You can avoid that by breaking codetext into smaller tokens instead of treating it as a single token.

    entry        : chunk(s?) EOF { join('', @{$item[1]}) }

    chunk        : text { $item[1] }
                 | code { $item[1] }

    text         : TEXT                      { join('', '<div class="t
+ext">', $item[1],                  '</div>') }
    code         : CODE_OP CODE_TEXT CODE_CL { join('', '<div class="c
+ode">', CGI::escapeHTML($item[2]), '</div>') }

    # Tokens
    EOF          : m{^\Z}
    TEXT         : m{[\w ]+}                             { $item[1] }
    CODE_OP      : m{<xxx>}                              { $item[1] }
    CODE_CL      : m{</xxx>}                             { $item[1] }
    CODE_CHARS   : m{[\w $]+}                            { $item[1] }
    CODE_SPECIAL : m{<\w+>} { $item[1] ne '<xxx>' }      { $item[1] }

    # Pseudo-token
    CODE_TEXT    : CODE_TEXT_(s?)                        { join('', @{
+$item[1]}) }
    CODE_TEXT_   : CODE_CHARS                            { $item[1] }
                 | CODE_SPECIAL                          { $item[1] }
[download]

UNTESTED.