Skip to content

Review

View Markdown~828 tokens

Review marks passages of a reply and attaches notes to them, then sends every note back as one prompt. Use it when part of an answer is wrong and the rest is fine, so you can point at the exact lines instead of describing them.

For approving tool calls before they run, see Permissions. That is a different thing with a similar name.

Three ways in:

  • Ctrl+X r opens the last assistant reply.
  • /review does the same from the command palette.
  • Click ⋮ beside any message, then pick Review passages. Right-clicking the message is an optional shortcut when the terminal forwards it. This reaches thinking blocks, tool results, and your own earlier messages.

The modal shows the message with a caret.

KeyAction
↑ ↓Move the caret one row
← →Move the caret one character, carrying on to the next row
PageUp PageDownMove half a page
Home EndStart or end of the row
Ctrl+Home Ctrl+EndStart or end of the message
Shift with any motionExtend the selection. Moving up or down takes whole rows
Ctrl+ASelect the whole message
Ctrl+CCopy the selection
EnterWrite a note on the selection, or on the caret's row
n pJump to the next or previous note
e dEdit or delete the note under the caret
Ctrl+SSend every note to the prompt
EscClear the selection, or close

The mouse selects too. Drag across a passage, or double-click a word. Releasing the button copies what you selected.

A ▌ marks the rows you are about to annotate. A ● marks rows that already carry a note.

The quote a note captures is the markdown that produced those rows, not the glyphs on screen, so lists and emphasis arrive intact.

Enter opens an editor with the passage above it. Type the comment and press Ctrl+S to save, or Esc to go back without saving. An empty comment does not save.

You return to the passage view, where you can mark another range and repeat.

Ctrl+S compiles every note into one block and drops it in the prompt editor as a collapsed paste. Add any extra context, then press Enter to send. Nothing reaches the model until you do, so you can still delete the paste and start over.

The block looks like this:

<review>
Address each note on my previous message.
<note>
> HNSW always outperforms IVF-PQ for recall at any scale.
Overstated. Add the memory-budget caveat and cite the benchmark.
</note>
<note surface="tool result">
> {"latency_ms": 1240, "recall@10": 0.91}
Where did this come from? It contradicts the first section.
</note>
</review>

The surface attribute appears only when the passage came from somewhere other than assistant text.

Once sent, the transcript draws the block as a review card rather than raw tags. Selecting and copying that card gives you the block above, the same way copying rendered markdown gives you its source.

Notes collect until you send them. Press Esc to leave the modal, open a different message, and keep adding. Caudra reminds you how many are waiting.

Starting a new session or loading another one drops pending notes, because the messages they point at are gone.

Website privacy and analytics