Agent tools

Three tools, and each one is a thing a reviewer does: read what is waiting, answer it, or say something of your own. Every per-thread argument takes an id or enough of the file and line to be unambiguous, because an agent that has just read config.ts:41 should not have to carry an opaque id back to say something about it.

review_list

What is waiting, with the code each comment was written against.

{ "file": "src/core/config.ts", "status": "waiting" }

• file narrows to one file. A path suffix is enough.
• status is waiting (open and answered) or all, which includes resolved threads — the history of a file, for when that is the question.

Each thread comes back with its id, where it is, what was said by whom, its status, the lines as they were when it was written, and whether the code has changed since.

review_reply

Answers one comment, and says whether it is done.

{ "id": "rv_0mgk2x1f4a", "text": "Named the file in the error and kept the original message.", "resolved": true }

One tool rather than a reply and a resolve, because they are the same act with a different ending: the agent is always saying something, and sometimes that something closes the thread. A separate resolve would invite resolving in silence, which is the one outcome nobody wants.

resolved: true is a claim that gets checked. A thread remembers the lines it was written against. If they still read exactly as they did, the reply is kept, the thread stays open for you, and the agent is told:

Replied, but not resolved: src/config.ts still reads exactly as it did when the comment was written, so nothing was changed. The thread is waiting on the person now.

Disagreeing is a reply, not a silence — an agent that thinks a comment is wrong should say so and leave it open, which is a different act from fixing it.

review_open

A comment of the agent’s own, on a line, a range, or a whole file.

{ "file": "src/core/config.ts", "from": 40, "to": 52, "text": "Worth a test for the empty case." }

For things worth saying that are not what it was asked to do right now: a bug noticed on the way past, a decision that wants a second opinion, work it is deliberately leaving behind. It writes through the same store the panel reads, so a note the agent leaves is a note you see — and it is waiting on you, exactly as a reply from it would be.

A note here outlives the conversation. It is on the branch, not in the chat, which is the point: write the ones a future reader needs, not a running commentary.

What the agent is told, once

Every conversation gets a few lines of guidance and, when something is actually waiting, one line saying how much and in which files. Never the threads themselves — the list is a tool call away, and a system prompt is not the place to put a review.

Without the server half

The tools come from @opencode-cockpit/review itself, so a normal install has them. If the plugin is running as a panel only, submit notices and sends the whole review as prose instead of naming tools the agent does not have.