All of this is on 5.8.3, ActiveState build 809

Yes, the bug is that an error within eval should not affect other evals. I'm not sure where the problem lies, because the following code still works as expected and raises no error, no matter whether $string is a global or a lexical variable:

use strict;
use Test::More tests => 12;

sub second_eval_unaffected($) {
    my ($code) = @_;
    eval $code;
    diag $@ if $@;
    my $string = '';
    ok eval(q($string="new eval done";)), "Eval after >>$code<< still 
+works";
    is $@, '', "... and raises no error either.";
    is $string, 'new eval done', "... and sets the variable correctly"
+;
};

second_eval_unaffected 'my $var_without_accent;';
second_eval_unaffected 'my $var_with = "é_accent";';
second_eval_unaffected 'my statement_with_error;';
second_eval_unaffected 'my $var_with_é_accent";';
[download]

This program exhibits the problem as a self-contained Test::More program. The first sequence of steps leads to a failure, while the second run succeeds. This should not happen, but I don't know why :)

use strict;
use Test::More tests => 10;

sub gauntlet {
    my $code1=shift;

    my $string;
    my $expected = "new eval done";
    my $code2='$string="new eval done";';

    ok eval($code2), "Good code evals correctly";
    is $@, '', "... and raises no error on its own.";

    diag "Eval trial code\n";
    diag $code1;
    eval $code1; 
    diag '$@ is ', $@;

    undef $string;
    diag "Eval good code ($code2)\n";
    ok eval $code2;
    is $@, '', "No error raised";
    is $string, $expected;
};

gauntlet('my $var_with_é_accent;');
gauntlet('my $var_without_accent;');
[download]