# Plan review

> When an agent proposes a plan before writing code, AgentsRoom opens it in the agent's pane as a readable document instead of a raw terminal prompt.

- Area: Desktop app
- Plans: All plans
- Last checked against the product: 2026-11-03
- Web page: https://agentsroom.dev/docs/plan-review

## What it does

When an agent proposes a plan before writing code, AgentsRoom opens it in the agent's pane as a readable document instead of a raw terminal prompt. You select passages to comment on them, replace them, delete them or add text after them, write a general note, then approve the plan or send it back with your changes as one structured message. Every version of the plan in the conversation is kept, with a diff between the previous and the new one. The same card can annotate the agent's last message, or a markdown file such as a spec, and send your notes to an agent.

**Video (1 min 17 s):** [Plan review, shown in the real app](https://youtu.be/2ryEw-oJbYQ). Claude Code proposes a plan in plan mode; in the card above its input box, step 3 is selected and deleted, a test line is replaced, a general note is added and the plan is sent back with Request changes; version 2 arrives in the same card with its Changes tab, then Approve, and the agent codes the fix.

## Where to find it

- A plan: the card appears by itself above the agent's input box when the agent proposes a plan (title **Proposed plan**, with its version, **v1**, **v2**...). When the input box is empty, the card takes its place and shows two tabs instead of the title: **Plan** (the document) and **Message to the agent** (your usual input box). From **v2** on, a **Plan** / **Changes** switch shows what changed since the previous version.
- The two answer buttons sit at the bottom right of the card, as icons: **Request changes** (curved arrow) and **Approve** (check mark, also Cmd/Ctrl+Enter); hover them to read their label. **Answer in the terminal** is the terminal icon at the bottom left, next to the dictation mic.
- The arrow at the right of the card's header folds it to one line. Click anywhere on that line to open it again, on the computer or the phone (the line's buttons keep their own action).
- The agent's last message: right-click in the agent's terminal (or the selection panel) > **Annotate the last message**.
- A file: open a markdown file in the file preview > **Annotate** > pick the agent that receives the notes.
- Video tutorial: a small play icon at the right of the card's header, before the expand, collapse and close buttons (tooltip "Watch the video tutorial on YouTube"), opens the video on YouTube in your browser. It is there for a plan, a message or a file alike.

## How to use it

1. Select a passage in the plan (drag across it with the mouse). A small menu of four icons opens under it: **Comment**, **Replace**, **Insert after**, **Delete** (hover an icon to read its name). **Delete** is saved at once; the three others open a field: type your text and press Enter. The passage stays highlighted and the annotation is listed under the plan. Since 2026-10-31 this also holds for a passage that runs over a line break of the plan's text, which used to lose its highlight.
2. Optionally write a **General note** at the bottom.
3. Click **Approve** (or **Approve with notes** when you wrote some), or **Request changes** to send the plan back. **Answer in the terminal** hands the decision back to the CLI's own prompt.

On a message or a file, the button is **Send to the agent**: your notes go as one message, like a message typed in the agent's input box.

## Settings

- `planReviewEnabled` (global, Settings > Agent behavior > **Review plans in AgentsRoom**): on by default. Off, the CLI shows its own approval prompt as before. Pilotable with `settings_set`.

## Agent tools (MCP)

None.

## Providers

- **Claude Code**: plan mode. The agent waits for your answer in the card; approving leaves plan mode, requesting changes keeps the agent in plan mode with your feedback.
- **Codex** (Plan mode), **OpenCode** (Plan agent) and **Antigravity** (implementation plan review): AgentsRoom detects the plan at the end of the turn and opens the same card. Your answer is typed in the terminal for you (the menu choice, then your feedback as a message). No hook is installed in these CLIs.
- Every other CLI: **Annotate the last message** works with any agent.

## Mobile

Yes. The card appears in the agent's console on the phone: read the plan, tap a paragraph to comment on it, replace it, delete it or add text after it, then approve or request changes. The answer is applied by the paired desktop.

## Limits

- The plan card waits as long as you need (up to 24 hours for Claude Code). If AgentsRoom is closed or restarted meanwhile, the CLI shows its own prompt instead.
- Pressing Esc in the terminal closes the card.
- With Claude Code, the CLI's own approval menu ("Would you like to proceed?") stays visible in the terminal above the card while it waits; answer in the card, not in that menu. After **Request changes**, the terminal shows Claude Code's line "Denied by PermissionRequest hook" with your feedback: that is how Claude Code displays a plan sent back, not an error.
- On the phone, very long plans are cut at 60,000 characters (the desktop keeps all of it).

## Common questions

- **How do I send Claude Code's plan back with changes?** When the plan card opens above the agent's input box, select the passage you disagree with and pick Delete, Replace or Comment in the small menu, add a general note if you want, then click **Request changes** (curved arrow, bottom right). Claude Code stays in plan mode and proposes version 2 in the same card. Video: https://youtu.be/2ryEw-oJbYQ
- **Is there a video that shows it?** Yes, a 1 min 17 s demo in the real app: https://youtu.be/2ryEw-oJbYQ. The play icon in the header of the plan card ("Watch the video tutorial on YouTube") opens it too.
- **Can I still answer in the terminal?** Yes: **Answer in the terminal**, or turn the setting off.
- **Where are older versions of the plan?** In the **Changes** tab of the card, and on every computer signed in to the same account.
- **Does it work with Codex or OpenCode?** Yes, by detection: the card opens when the plan appears, and your answer is typed in the terminal.

## Related

- [Structured Questions](https://agentsroom.dev/docs/structured-questions.md): the agent's question cards, docked at the same place.
