v0.47.4

Release Notes — v0.47.4

Patch: Compaction keeps the prompt you started the turn with

When a turn folded its context with compact_context, the message the turn
opened with — what you asked for, and any image you attached — could fall out of
the model's view. The summary took its place. Everything the agent did after
that point was working from a digest of your request instead of the request
itself.

The fold now keeps your original prompt, images included, at the head of what
is sent, with the summary folded in beneath it. If the turn is compacted again
later, the summary is stripped back out first, so your words stay the words you
wrote rather than being summarised a second time.

Other changes

• The guidance you send mid-run still reaches the model on an image-only
  prompt, where there is no text to append it to.