[BUG] 400 Bad Request from Mistral reflection router example in Flo-ai

[joseph225]

Describe the bug
A clear and concise description of what the bug is with stack trace

400 Bad request when 3 agents try to resolve a reflection problem. Attached is the error log. Also attaching the sample code.

rag_reflection_router.py

error.log

To Reproduce
Steps to reproduce the behavior:

1. A minimal sample code to reproduce the bug is appreciated

Expected behavior
A clear and concise description of what you expected to happen.

Screenshots
If applicable, could you add screenshots to help explain your problem?

Environment (please complete the following information):

• python version: [e.g. 3.11]
• OS [e.g. windows, mac]
• flo-ai version [e.g. 0.0.1]

Additional context
Add any other context about the problem here.
``

[vizsatiz]

🧠 Why this happens in your reflection setup

In the A → B → A → C reflection pattern, each agent is represented as a separate LLM "turn".
When the router bounces messages between agents (main → critic → main → final), it keeps appending messages to the shared MessageMemory.

If each agent internally sends messages as "role": "assistant" (which is true for most flo_ai.models.Agent), then when the critic replies, and then main_agent replies again, the conversation will contain two "assistant" messages in a row.

That violates Mistrals expected alternation pattern.

🛠️ The possible fix:
We have to add a new field in agentic class named act_as maybe, and it can support 2 values, Assistant or User. Based on this act as if given we can be alter the memory to have role assigned accordingly

main_agent = Agent(
    name="main_agent",
    act_as="user",  # <-- Treat main agent as "user"
    system_prompt="You are the main agent...",
    llm=llm,
)

critic = Agent(
    name="critic",
    act_as="assistant",  # <-- Critic responds as assistant
    system_prompt="You are a critic...",
    llm=llm,
)

final_agent = Agent(
    name="final_agent",
    act_as="assistant",
    system_prompt="You are the final agent...",
    llm=llm,
)

That way, the memory alternates:

system → user (main) → assistant (critic) → user (main) → assistant (final)

This should fix it, and make the framework future proof too.

[vizsatiz]

Fix PR #147

[vizsatiz]

@joseph225 Would be ideal if you can checkout to this branch and test this issue with your model

Now you can pass act_as parameter into agents, both in yaml and code. This will add the agent message as whatever is passed into act_as. It should either be assistant or user. (Defaulted to assistant)

[vizsatiz]

Fixed for now in #147 ,