First things first: what you are describing there is not a one-time-pad!
As I explained in another answer of mine:
Per definition, OTP requires the “key“ to be…
- a truly random one-time pad value,
- generated and exchanged in a secure way,
- at least as long as the message, and
- only to be used once.
What you describe (eg: using a key going
Orange Fish) can merely be compared to a simple, keyed substitution cipher with security boundries comparable to a regular Ceasar cipher. Differently worded: your encryption technique is no where near the cryptographic security a real one-time-pad can provide.
Now, to the core of your question: if Bob would base his key on the plaintext of what Alice send him, you would practically be creating a ciphertext by combining two plaintexts (which should actually remain secret).
To attack such a simple “cipher“, an attacker could simply exploit the statistical bias coming from the fact that both the key as well as the ciphertext are regular, human language. Letter frequencies et al would quickly help differring the key from the plaintext. The fact that the space character doesn’t even get encrypted, makes it even easier to recover the plaintexts. The result of using such an “encryption“ would be that an opponent would quickly be able to successfully attack it, and recover both the message Alice tried to secretly send to Bob, as well as the message Bob secretly tried to send to Alice as a reply.
Simpler said: your idea incorporates more than the usual danger of leaking a single secret… because it practically risks the leaking of two secrets: the plaintext which is being send, and the previously received plaintext which is
reused abused as key to encrypt the plaintext to be send.
As for an example of the cryptanalytic side of things, I’ll refer you to this question and its accepted answer, as they provide a lengthly and indeep insight on how one could extract a keyword from a keyword cipher – including a practical example. When reading that Q&A, keep in mind that your “algorithm” may seem similar, but actually is much simpler and incorporates more insecurities than what they are talking about there.
Long story short: the fact that Bob and Alice are risking to leak both their secrets twice is pretty inacceptable from a cryptographic point of view. Yet, I am sure every potential attacker will be more than happy when you allow him to gain a complete transcript of your communication in such an simple way; especially since you are including all those nice carbon-copies of the plaintext, which are (ab)used as keys.
Honestly, if I were a cryptanalytic opponent facing your idea in a practical situation, I’ld probably put you on my Christmas-card list for providing me with one of the easiest and most amusing jobs ever.