add missing .with_recipient('assistant') to tool message in demo code #62

[amirhosseinghanipour]

The demo was missing the required .with_recipient("assistant") call on tool messages. We can see in harmony format specification that tool messages should include to=assistant in their header because the tool response requires the assistant to process. It's currently producing this format:

<|start|>functions.lookup_weather<|channel|>commentary<|message|>{ "temperature": 20, "sunny": true }<|end|>

Here's our format specification in harmony:

<|start|>{toolname} to=assistant<|channel|>commentary<|message|>{output}<|end|>

So with that we can correctly produce this:

<|start|>functions.lookup_weather to=assistant<|channel|>commentary<|message|>{ "temperature": 20, "sunny": true }<|end|>

The issue #62 suggested doing that and I also created a simple test that compared both situations. I found out that the original had 241 tokens, the update has 244 tokens now, the first difference found at token 221, and the decoded text showed the missing to=assistant part in the tool message header.

So now 3 additional tokens when .with_recipient("assistant") is added.

While I was probing into this issue, I also discovered that there's a separate test failure in test_does_not_drop_if_ongoing_analysis that appears to be related to tokenization differences in the <|constrain|>json part, but this is unrelated to the missing .with_recipient("assistant") issue and should be addressed separately.

[scott-oai]

thanks!