Swapping models mid-session in AgentCore harness made the new model impersonate the old one

Amazon Bedrock AgentCore harness lets you swap model providers on the same session by overriding model on InvokeHarness. While testing Claude (default) → OpenAI gpt-oss-120b, I hit something odd. I asked "are you Claude or GPT?" in turn 1, then switched to gpt-oss in turn 2 and asked again — and it answered "I'm Claude," even though it was actually OpenAI under the hood.

Python

sid = (uuid.uuid4().hex * 2)[:40]
# Turn 1: ask the default Claude "Claude or GPT?" -> answers "I'm Claude"
invoke(sid, "... by the way, are you Claude or GPT?")
# Turn 2: override to gpt-oss on the same session
invoke(sid, "Are you Claude or GPT?",
       model={"bedrockModelConfig": {"modelId": "openai.gpt-oss-120b-1:0"}})
# => "I am a Claude model."  <- but it's actually OpenAI

The harness reloads the conversation history each turn and replays it to the new model. So the gpt-oss model apparently inherited the persona from the earlier "I am Claude" assistant message in history. Indeed, if you don't ask for identity in turn 1 and only ask gpt-oss, it correctly says it was built by OpenAI.

The takeaway: don't rely on a model's self-report to tell which model is actually running after a swap. To be sure, ask in a clean session, or pass a bogus modelId and let ConverseStream reject it with The provided model identifier is invalid. I wrote up the full hands-on verification in a blog post.