This is caused by https://github.com/python/cpython/blob/caa1280d1ee5f828f346b585169a7592371d3faa/Lib/pdb.py#L1096 which always passes our own current stdin and stdout, and this triggers the conditional in https://github.com/python/cpython/blob/caa1280d1ee5f828f346b585169a7592371d3faa/Lib/pdb.py#L144-L145.

self.stdin and self.stdout are initialized to sys.stdin and sys.stdout respectively when not passed explicitly: https://github.com/python/cpython/blob/caa1280d1ee5f828f346b585169a7592371d3faa/Lib/cmd.py#L87-L94.

See Also: https://github.com/ipython/ipython/pull/10721

I can confirm that this fixes cursor keys not working properly after "debug foo()" (recursive debugging) with pdb.

It was added in 477c8d5e702 (a huge svn merge commit), with this reference:


      r45955 | georg.brandl | 2006-05-10 19:13:20 +0200 (Wed, 10 May 2006) | 4 lines

      Patch #721464: pdb.Pdb instances can now be given explicit stdin and
      stdout arguments, making it possible to redirect input and output
      for remote debugging.

I think a good alternative patch might be:

```diff
 Lib/pdb.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git i/Lib/pdb.py w/Lib/pdb.py
index f5d33c27fc..daf49b3629 100755
--- i/Lib/pdb.py
+++ w/Lib/pdb.py
@@ -141,7 +141,9 @@ def __init__(self, completekey='tab', stdin=None, stdout=None, skip=None,
                  nosigint=False, readrc=True):
         bdb.Bdb.__init__(self, skip=skip)
         cmd.Cmd.__init__(self, completekey, stdin, stdout)
-        if stdout:
+        if stdout and stdout is not sys.stdout:
+            # stdout gets passed with do_debug for example, but should usually
+            # not disable using raw input then.
             self.use_rawinput = 0
         self.prompt = '(Pdb) '
         self.aliases = {}
```

Just for reference and searchability: this causes tab completion to not work with `debug foo()` also.

> I think a good alternative patch might be:

This however makes it behave different in tests, where stdout might be mocked/wrapped intentionally.

Therefore I think using the parent's `use_rawinput` is the better fix for this (the originally proposed patch).