How should prompt-chaining work?

[ErikBjare]

I thought that gptme "first prompt" - "second prompt" would wait for the first prompt to completely finish (exhaust without a command given), but turns out it runs the second prompt as soon as tools for the first one are given.

Example:

$ gptme 'what is 2+2' - 'what is the last answer times 23'
User: what is 2+2
Assistant: To calculate 2+2, I can use Python:

\```python
2 + 2
\```
Out[1]: System:
Executed code block.

Result:
\```
4
\```
User: what is the last answer times 23
Assistant: To calculate the last answer (4) times 23, I'll use Python again:

\```python
4 * 23
\```

What could be a good test case so we can test for this in CI? Something where it needs both a tool result and to comment/summarize or run more tools before it can generate an answer for the second prompt (the result of which should contain an expected flag to test for the presence of).