Get rid of agent_input and only rely on working memory

[AlessandroSpallina]

When a plugin developer wants to use the hook before_agent_starts for tasks like modifying the chat history, acting on agent_input.chat_history has no effect.

This code isn't working:

@hook
def before_agent_starts(agent_input, cat):
    # When the LLM reads the chat history, it will lie generating the string "Generated in x seconds"
    # this is not an intended behavour, so we delete each occurrence of that string
    
    agent_input['chat_history'] = "\n".join([line for line in agent_input["chat_history"].splitlines() if '✨ _Generated' not in line])
   
    return agent_input

Instead, acting on cat.working_memory.history is the right way to modify the conversation history. The following code is working properly

@hook
def before_agent_starts(agent_input, cat):
    # When the LLM reads the chat history, it will lie generating the string "Generated in x seconds"
    # this is not an intended behavour, so we delete each occurrence of that string

    if len(cat.working_memory.history) >= 2:
        cat.working_memory.history[-2]['message'] = "\n".join([line for line in cat.working_memory.history[-2]['message'].splitlines() if '✨ _Generated' not in line])

    return agent_input

[pieroit]

Nasty bug, I'll see to it thanks

[alfredotoselli]

https://discord.com/channels/1092359754917089350/1092360210351398963/1288557418820800514

Probably this issue could be closed

[pieroit]

Thanks @AlessandroSpallina @alfredotoselli @matteocacciola for checking this out

Please put and take stateful info from / to working memory

Changed the name of the issue so we remember when and why ;)