Code changes
With a repository connected, the assistant can change your site's code the same way a developer would: read the file, make a targeted edit, commit to a branch, and let the rebuild deploy it.
What it can change
Styling, components, page templates, configuration, content-schema definitions — anything in the repo, within the paths the requesting user's role allows. Typical asks:
How a change lands
- Read before edit. The assistant always reads the current file (from the branch it will write to) before changing it, and makes targeted edits rather than wholesale rewrites.
- Branch confirmed up front. The typical flow is code to the staging branch first, production when you want it live — the assistant suggests staging by default when one is connected and confirms before the first write of a task.
- Commit → automatic rebuild. Each write is a real commit (you'll see the commit reference in the reply). Your host's git-triggered build deploys it, typically in 30–60 seconds.
- Verified. The assistant re-reads the file after committing and confirms the change is in place before reporting done.
- Build checked. Where deploy-status access is configured (Netlify, Vercel, Cloudflare Pages, or GitHub Actions), the assistant also waits for the build and confirms it actually deployed — and if the build failed, it reads the error, fixes the file and recommits rather than telling you a broken change is live.
- Previewed. For visual changes it then offers a look — opening the page in the preview pane or sharing a screenshot — before anything is promoted to production.
Path scopes
Code access is governed by the same role rules as content: a role grants
code update on path patterns (for example src/**), optionally restricted to an environment — so you
can give a developer staging-only access, or scope an agency to one part of the codebase. Sensitive files (env
files, credentials) are blocked outright.
Schema changes
Changes to CMS schema definitions are higher-risk — they can affect every document of a type. The assistant explains the impact and asks for explicit confirmation before committing them, and prefers a content-level fix whenever one would achieve the same goal.