Skip to content

Write feedback

Audience: anyone commenting on a plan on caret's review page.

Comment on lines

Pick the lines a comment is about, then type it. In either view of the plan:

  • Click a line to comment on it.
  • Drag across lines to comment on all of them.
  • With the keyboard, press V to select whole lines, then c to comment on the lines the cursor or selection covers.

After you submit a comment with ⌘↵, the cursor rests on it.

With the cursor on a comment:

  • e or i edits it.
  • Enter expands a collapsed comment and edits an expanded one. On a Resume marker it resumes the draft.
  • h collapses it and l expands it.
  • Shift+D asks to discard it and discards on confirm.
  • c comments on its lines, adding to its thread.
  • V starts a selection on its last line.

Accepted and Expired comments only fold, with h, l, or Enter while collapsed. Key caps show beside the controls of the comment the keyboard moved to, when Shortcut hints in Settings → Appearance is on. A mouse click hides them until the next key.

In the preview, hovering anywhere out to the plan's right edge shades the wrapped line under the pointer and shows a + beside it. Clicking the line or its + comments on every source line with text on it. V shades its selection the same way, with the + on the cursor's line.

The lines stay shaded while you write. Clicking and dragging pick lines, not text, so hold shift to select text instead. Selecting text opens no comment.

In the preview, comments and unsent drafts show right under the last wrapped line they cover, and hovering or focusing one highlights the text it covers. If you switch between the preview and the source view with a comment open, caret keeps its text as a draft, which you resume in either view.

Cite a skill or a file

In any feedback editor (the comment box beside a line, a comment card, or the Request changes dialog), two characters open a list:

  • / at the start of a word lists the skills the agent that wrote this plan can use.
  • @ lists the files under the project folder the agent was working in.

To use either list:

  • Move through it with the arrow keys, or with tab and shift+tab.
  • Press enter to take the highlighted row. caret leaves a space after it, ready for the next word.
  • Press esc to close the list without closing the dialog around it.

A / inside ordinary text, such as the path src/lib/api.ts, opens no list.

The / list filters as you type. Choosing a skill inserts its name in the exact form that agent needs, so a plugin skill arrives with its plugin's name in front, as /superpowers:brainstorming. Each row shows where its skill came from, which tells apart two sources that offer the same name. Which skills the list offers depends on the agent host:

  • Claude Code offers your own skills, the reviewed project's, and each enabled plugin's.
  • OpenCode offers its slash commands.
  • pi offers its skills.

caret reads the skill list once when the review opens, so a skill you add mid-review shows up after you reload the page. The name becomes part of the feedback the agent reads; caret never runs the skill itself.

The @ list matches loosely, so @srlbfoo finds src/lib/foo.ts, and each row marks the characters your query matched. Choosing a file leaves @ and the path, relative to the project folder, as ordinary text, so the reference still resolves when the agent picks the plan up in a later session. Type :42 after the name to cite line 42 with it. In a project too large to search in full, the list says how much of the result it shows.

Preview what you cite

With a list open, press ctrl+space to open a panel beside it. The panel shows more about the highlighted row: a skill's own description, or the opening lines of a file, colored in the theme you are reading in. It follows the arrow keys, so you can tell src/lib/api.ts from src/lib/api.test.ts, or brainstorming from linear-plan, without leaving the editor. A :42 after a filename moves the panel to that line and marks it. A line past the end of the file says where the file ends.

The panel sits to the right of the list, or to its left near the window's right edge. In a window too narrow for both side by side, the two stack on whichever side of your line has room, and stay together on screen.

Opening the panel keeps your place in the list. Press ctrl+space again to close it. caret remembers whether you left it open, so the next list you open, in this review or after a reload, opens the same way. A skill with no description, or a file caret cannot read, says so in a sentence, and the list keeps working.

The top of the list names the shortcut. Turning off Shortcut hints in Settings → Appearance removes that line; the shortcut keeps working.

Check that a reference resolves

As you type, a reference caret can find shows as a chip, a small tinted label around the text. One it cannot find stays plain text. So a missing chip catches a misremembered skill name or a path that does not exist while you can still fix it, rather than when the agent sends back its revised plan.

caret checks each kind against the same source its list completes from: a /name against the skills the agent that wrote this plan has, and a path against the project folder the agent was working in. Both re-check as you edit, so fixing a typo brings the chip back.

A file or folder chip is green, and a skill chip is blue. A path you completed with @ keeps its @ inside the chip, because the agent reads it too.

In plain text, caret looks a path up only when it has a / or a ., so an ordinary word like test never becomes a chip, even when the project has a test/ folder. Wrap a name in backticks to mark it as a path: `Makefile` at the top of the project becomes a chip, and a bare Makefile does not.

The chip only changes how the comment looks. The agent receives the exact text you typed, chip or not.

Send your feedback

Once your comments are in, decide on the plan from the top bar:

  • Request changes (r) opens Send the plan back for revision. It lists your comments and adds a General comment box for the plan as a whole. Click Send for revision. The agent revises the plan and presents it again as a new version.
  • Approve (a) asks Approve this plan?, with an optional Notes for the agent box. Confirm it, and the agent starts the work. Your comments and notes go with the approval, and the agent applies them without another planning round. The arrow beside Approve lists your agent host's other ways to approve, when it has any.
  • Reject (shift+r) asks Reject this plan?. Confirm it, and the agent is told the plan was rejected and to wait for your next message. Your comments are not sent.

When you have comments queued, the Approve and Reject dialogs list them and offer Request changes instead. Choose it when a comment changes the plan's structure, so you review the revised plan before the work starts.

Drafts you kept for later show in these dialogs under Unsent comments, and caret does not send them. To send one, click Cancel, click Resume on the draft in the plan, then click Comment.

In a window under 960px wide, Reject and Request changes move into the … menu beside Approve. Under 640px, Approve joins them.

Next steps