I tested on 3.2, Win 7. sys.path starts with '' on startup, after restart, and after ^c interrupt of 'while True: pass'. It start with absolute path when running a file from the editor. I believe this is as should be.

I simplified the patch a bit. Its basic idea is to add a parameter with_cwd to PyShell.ModifiedInterpreter.transfer_path. That method is called by methods start_subprocess and restart_subprocess. start_process is only called by PyShell.PyShell.begin. So I think there is no need to propagate the new parameter back past the call in start_subprocess.

.restart_subprocess is called in .poll_subprocess, .runcode, PyShell.cancel_process, and PyShell.restart_shell. (I am a little puzzled why ModifiedInterpreter.runcode calls self.interp.restart_subprocess, as in the PyShell methods, instead of self.restart_subprocess, as in poll_subprocess, but moving on...)

restart_process needs the new parameter to pass to transfer_path because it should be True for F6 Restart Shell and not for F5 Run Module. Roger's patch also left the default False for the other calls. I do not know what they do so I cannot judge them. Roger, do you know, and did you consider each? (Leaving the current behavior for those will at worst not fix something that could be, so it is safe.)

restart_shell is *bound* to F6 Restart Shell (which passes the ignored even), so the default with_cwd must be True for that binding. It is *called* by a function ScriptBinding.py that gets bound, so the call can be altered to pass the correct False. I prefer that to making two functions. That call could be replaces by a call directly to .interp.restart_subprocess(False) (and restart_shell slightly simplified).

I retested after these changes to the patch. It is from TortoiseHG, so should apply (after I modified the file paths). Since hacking on idlelib is new for me, I would like other eyes and tests before I apply.