Project assistant

The assistant is a chat beside every project page. Ask it about the project, or ask it to change something: a scenario, a prototype's wording, a file in a feature. You do not need an editor or a terminal, and you can take any change back.

Turning it on

The assistant is off in a new project. A project owner turns it on in Settings → Agents, by ticking Turn on the project assistant in the Project assistant box and saving. Until then there is no Assistant button and no Assistant conversations page.

Opening it

Click the Assistant button in the top bar of any project page, right of Help. The panel opens beside the page you are on; click the button again, or ✕, to close it. The panel stays open as you move between the project's pages, with your conversation and anything you had started typing.

Type and press Enter. Your first message starts the conversation. The suggestions in an empty panel are examples: clicking one puts it in the box for you to edit, it does not send.

The panel knows what you are looking at. Above the box it shows Viewing: and the page, for example Viewing: Course enrolment · Gherkin Scenario, and that goes with your message, so "this feature" and "these scenarios" mean the ones on screen. If you highlight text on the page first, it is offered too, as a quoted chip you can remove with ✕.

Choosing the model

Beside the message box, pick Elite, Pro or Speed: the same tiers as in Settings → Agents. A new conversation starts on the tier set there for the Project assistant. You can switch at any time; the next message is answered on the tier you picked, without starting over.

Sizes

The panel opens docked to the side. Its header switches it to Split screen, half the width beside the page, or Maximize, the whole content area. Restore puts it back. The size you choose is kept as you move between pages.

What it can change, and where

The assistant can read the whole project: the legacy code, the modernized code, every feature's branch and the history of the project's agent runs ("why did the last run of Checkout fail?").

It changes only what the rest of Repave picks up:

• scenarios and Gherkin, through the same commands the other agents use;
• prototypes;
• files in a feature's branch: code, tests, docs and .feature files.

It does not change the legacy code or the modernized code's main branch. Those are not where a change would be kept, so it says so and offers to make the change in the right feature instead.

Pulling, pushing and pull requests

Ask it to and it does the git work on a feature's branch the feature page's buttons do: pull in commits made outside Repave, merge the base in, push, and open or update the pull requests — or tell you where they stand. It does these only when you ask.

• Your edits are saved first. Before it pushes or pulls, what it has changed so far in that reply is saved on the branch under your name, as a change you can undo. A reply that pushes can show two changes: the edits before the push, and any after it.
• It never overwrites someone else's commits on its own. If a source repository's branch has commits that were not made in Repave, it pulls them in first. If the branch still differs, Repave asks you, in the conversation, which repositories to overwrite; only the ones you tick are.
• It never merges a feature. Merging stays with you, on the feature page or in your repository provider.

Seeing and undoing a change

After a reply that changed something, a card shows what changed, feature by feature, with the scenarios and files by name. The change is already saved on the feature's branch, under your name. Undo takes it back on the branch in Repave; if it was already pushed, push again to publish the undo.

Undo takes the whole change back, scenarios and files together. It is not offered while a job is working on that feature (Can't undo while Login is being implemented.), or when something changed after it: a later commit to the same lines, or someone editing the same scenarios or prototype. The card says which. Undo leaves alone anything that was not the assistant's, such as changes someone had not committed yet.

Behaviour changes go through the scenarios

When you ask for the application to behave differently, the assistant changes the feature's scenarios, not its code, and then offers Start implementation for that feature. Approve the new scenarios on the feature page first, as for any implementation; the button then starts it as if you had pressed Implement yourself, with the same checks.

When a feature is busy

While a job is working on a feature (an implementation, a prototype build, a push or a merge), the assistant leaves it alone and tells you. In the other direction, while the assistant is in the middle of changing a feature, starting an implementation or a reset there is refused for those few moments, and a push or merge waits until it is done.

Questions and approvals

The assistant asks you when it needs a decision, and asks your approval before anything it may not do on its own. Both appear above the message box. While one is waiting, the Assistant button carries a dot.

What needs your approval depends on how the project reaches Claude. Where the model's own safety check is available (a project with its own Anthropic key, on Elite or Pro), it approves routine work itself and asks you about the rest. Elsewhere, the assistant changes files without asking (only in a feature's branch, and you can undo any change) and asks you before running a command, except one that only reads. For a command you can choose Allow (this once), Allow in this conversation, Always allow (for the whole project, as with any agent) or Deny.

Conversations

Management → Assistant conversations lists every conversation in the project, running ones first, with who started each and how many changes it made; click one to open it in the panel. The Conversations button at the top left of the panel shows the same list. Open any of them and write to carry it on. + starts a new one. Assistant conversations are listed only here, not under running agents or in Agent Logs; when one is waiting for you, the assistant icon in the top bar shows a dot.

A conversation you leave for 30 minutes goes to sleep, and so does one whose question has waited 30 minutes for an answer. Your next message wakes it, with everything it knew. The time it spends working from a message you sent counts against your usage.

When the assistant is not available

Besides Turn on the project assistant, the assistant needs Agents can ask you and hear from you turned on in Settings → Agents, since it asks you questions and approvals; a project owner can turn it on from the panel. It is also not available while the project is at its budget, or when five conversations in the project are already running.

Viewers cannot use the assistant; the button is greyed out and says why.