Can't modify user_defined context with Watson Assistant API v2 - ibm-cloud

Question

I'm trying to migrate chatbot to use newly introduced Assistant API v2.
My chatbot infrastructure includes Middleware services which modifies the context after getting the response from Watson. In some case I used to remove particular properties from the context and it worked fine. However I noticed that after migration to API v2 this approach does not work anymore as the deleted properties are somehow stored on Watson side.
For example I received following context from Watson:
{
"assistantId": "---",
"sessionId": "---",
"messageInput": {
"Text": "Some text",
"Options": {
"Debug": "true",
"ReturnContext": "true",
"Restart": "false"
}
},
"context":
{
"Global": "null",
"Skills": {
"AdditionalProperties": {
"main skill": {
"user_defined": {
"id": "23",
"description": "Dont know"
},
"system": {---}
}
}
}
}
}
Then I remove 'description' from the context and send request to Watson once more. Surprisingly 'description' is still there with the same value ('Dont know').
Possible solution would be not to remove a property but to set it's value to empty string. But even in this case my dialog does not work correctly as Watson somehow stores the point in dialog it was visiting previously (or not, these are my guesses). I assume it might be related to system.state property which stores an encoded state of the dialog (again - or not).
My question is why is dialog behaving this way?
How does it store the context information so I can't remove properties from user_defined context?
And how can I reset dialog state to initial keeping the same conversation_id (session_id)?
P.S. I'm using Watson API v2: 2020-04-01