Basic checks

• I searched existing issues - this hasn't been reported
• I can reproduce this consistently
• This is a RubyLLM bug, not my application code

What's broken?

Structured schema responses cause BadRequestError: Invalid request in multi-turn conversations when messages are persisted. OpenAI API receives a Hash object instead of a JSON string for the message content.

How to reproduce

class PersonSchema < RubyLLM::Schema
  string :name, description: "Person's full name"
  integer :age, description: "Person's age in years"
  string :city, description: "City where they live"
end

chat_record = Chat.create!(model: 'gpt-4.1-nano')
chat_record.with_schema(PersonSchema)
response = chat_record.ask("Generate a person from Paris")
chat_record.with_schema(nil)
chat_record.ask("Generate a fictional bio for this person")

Expected behavior

Structured response from the first message should be serialized as a JSON string when included in the message history for subsequent API calls.

What actually happened

Structured response is sent as a Hash:

request -- POST https://api.openai.com/v1/chat/completions
request -- {:model=>"gpt-4.1-nano",
 :messages=>
  [{:role=>"user", :content=>"Generate a person from Paris"},
   {:role=>"assistant",
    :content=>{"name"=>"Sophie Laurent", "age"=>29, "city"=>"Paris"}},
   {:role=>"user", :content=>"Generate a fictional bio for this person"}],
 :stream=>false}
RubyLLM: API call failed, destroying message: 2322
Invalid type for 'messages[1].content': expected one of a string or array of objects, but got an object instead. (RubyLLM::BadRequestError)

Environment

• RubyLLM version: 1.9.1 (working in 1.8.2)
• Provider: OpenAI