cognition.ai/wholeUserMessages — client capability flag indicating that
the client reads cognition.ai/quotedArtifacts from a text block’s _meta
and keeps pre-uploaded attachment URIs when a v2 user_message replaces an
optimistic prompt. When present, the proxy sends every multi-block prompt as
one user_message; otherwise prompts with quotes or private attachments go
out as user_message_chunks.