Web interface

Serves on http://localhost:8080, bound to your own machine only. --port changes the port. Light and dark, fully offline.

The project picker. The Projects parts follow a picker. It lists the projects on flanner's Projects page, and starts on All projects. With all projects, each table has a Project column. With one project picked, Reach shows that project in detail.

Add a project. The button beside the picker says how. Curb checks the projects on your Projects page, so add the project there with New project, or run flanner init inside the repository. Then come back and pick it. Curb reads the project's agent settings, workflows and code. It never runs them.

“This machine” means each agent started outside any project. Curb assesses it from an empty folder of its own, not from your home folder. The leak scan is the one machine part that reads projects: it covers the agents' own files and every project on the Projects page.

The pages hide credential names and secret locations, as the terminal does. You can show them, for a while, in one browser:

1. Press Show names and locations. The page shows a four-digit code.
2. Your operating system asks whether it is you. Its prompt shows a code too. Approve only if the two match.
3. That browser then shows names and locations for 5 minutes. The time left is always in view, and Hide names ends it at once. So does restarting the web interface.

Nothing is written to disk. The server keeps the reveal in memory, and sends every response that carries a name marked not to be stored. A second browser sees nothing. A prompt you refuse or ignore counts toward the pause after three refusals.

On a Linux desktop the names stay hidden. Its prompt shows its own text and cannot show the code. The page says so, and points to flanner curb show, which opens them in a window on your screen.

A secret's value is never shown, with or without this.

A button with a shield changes something. It opens a review first: what will change, what to expect, and how to undo it. Then your operating system asks you, as it does for the commands. While it asks, the button says it is waiting. If you refuse, nothing changes, and the page says so.

Nothing in an agent's settings, a project or a credential store changes from a page without that prompt. Deleting what Curb has stored needs it too. The one exception is the same as in the terminal: under the delegation, org policy changes that only tighten apply on their own.

• A fix is planned when the page reads the settings. It is written only if they have not changed since. If they have, nothing is written, nobody is asked, and the page checks again.
• Removing a secret from its file needs names showing, so you see which file changes.
• One approval is asked for at a time. A second request is refused, and says why.
• Anything that leaves the machine, or costs money, is said before it starts. Checking which secrets still work sends each to its issuer. Running tests uses tokens.

A check that only reads, such as Check again or Scan again, runs at once and changes nothing. A slow one runs in the background: the page shows the last answer, says it is checking, and updates itself.