pub fn render_for_summary(
messages: &[Message],
max_result_chars: usize,
) -> StringExpand description
Flatten messages into plain text for the summariser.
Deliberately not a replay of the structured transcript. Sending the real
messages means sending tool_results with no tools declared on the request,
and llama-server answers that with an empty completion — found by running it,
not by reading the spec. Prose has no such failure mode on any provider, and
it also removes any chance of the summariser deciding to call something.