Review and multi-step tools

Reviewing changes

review_changes (or /review) reviews uncommitted changes for correctness, security, and quality.

1. An independent read-only reviewer runs per dimension, each in a fresh context.

2. A skeptical verifier re-checks every candidate against the real code and drops false positives.

3. You get the confirmed findings.

• Live. Findings appear under the review as each reviewer reports them, marked unverified, and update in place as the verifier confirms or rejects them. Once the review ends, rejected findings fold into a "dropped as false positives" line you can expand.
• Click a finding to see why it matters and the suggested fix.
• Options: scope it to a base branch or specific paths; raise effort to high to verify each finding with several independent skeptics; pass model to run the reviewers and verifiers on a cheaper sibling model.

Showing images

show_images shows image files the agent produced, such as screenshots of an app or rendered mockups.

• Desktop app: a captioned gallery. Click any image to open it full size, step through the set with the arrow keys, compare it with a "before" image using a slider when the agent paired one, copy it, save it, show it in the Files panel, or open it in its default app.
• Terminal: each image is a clickable link to its file.
• No approval needed. It only reads image files (PNG, JPEG, GIF, or WebP, up to 5 MB each, 12 per gallery) inside the project.
• Images go to the model scaled down to at most 1568 pixels on their long side; the chat keeps the full-size originals.

Rendering a mockup

render_mockup (desktop app) turns an HTML mockup file in the project into screenshots without a dev server, shown as a gallery marked "Mockup from code".

• Sizes and themes: desktop, tablet, or mobile; light or dark; or a custom size.
• Locked down. The page runs in a hidden browser that can load its own files from the project folder, but nothing from the network or elsewhere on your computer.
• No approval needed. It writes nothing.

Viewing your running app

view_localhost (desktop app) loads a page from a dev server on your computer in a hidden browser, and shows the agent a screenshot plus the page's visible text and console output.

• Steps. It can run steps first: fill a field, click or double-click, press a key such as Enter or Escape, wait for an element, check that some text is or isn't shown. So the agent can try a flow such as logging in or submitting a form, and get a pass or fail for each step.
• Several views at once. It can capture the page in several viewports or color schemes in one call (desktop, tablet, or mobile; light or dark; or the full page height), shown as a gallery marked "Running app". Steps then run with only one of them.
• One call, one flow. Each call starts from a fresh page load, so a flow runs within one call.

Safety:

• Every call asks for approval. A call with steps is approved separately from plain views, and the approval lists each step (never the text being typed).
• Text typed into a password field is never shown back to the agent.
• Steps can't download or upload files, or use your camera, microphone, or location.
• The page can't leave localhost.