claude -p "/usage" returns a model reply, so usage must be scraped from an interactive session

The `claude` CLI's `/usage` is a TUI-only slash command. There is no `claude usage` subcommand, and print mode does not render the panel.

### Repro

1. Authenticate the `claude` CLI as usual. 2. Run `claude -p "/usage"`. 3. Observe the output: a normal model reply, not the usage panel.

In print mode the string `/usage` reaches the model as ordinary prompt text. The panel is drawn by the interactive UI, and the non-interactive path never invokes it.

### What this rules out

- A `claude usage` subcommand: it does not exist. - `claude -p "/usage"`: it gives a model answer, so there is nothing to parse.

The model's answer to that prompt is not a usage readout, so do not treat it as one.

### What works

The only route I have found is to run an interactive `claude` process under a real terminal (a pty) and read the rendered screen:

- Spawn `claude` attached to a pseudo-terminal, not plain stdin/stdout pipes. The interactive UI expects a real terminal. - Wait for the REPL to finish starting up. - Send `/usage` as keystrokes, followed by Enter. - Read the terminal output once the panel has drawn. - Strip the ANSI escape sequences and parse the text that is left.

Because the UI repaints the screen, you are reading rendered output rather than a stream of structured data. On Windows I do this from Node with node-pty. Expect to handle timing: you have to wait until the panel has drawn before you capture it, and a fixed sleep is the simplest way to start.

### Consequences for an agent harness

- Usage is not available through a stable, documented machine interface that I could find. Scraping depends on the panel's current layout, so a UI change can break the parser. - Keep the scraper isolated behind one function and fail loudly when the expected text is missing, rather than returning zeros. - Spawning a full interactive session just to read usage is heavier than a subcommand would be. Do it on a schedule, not on every request.

If you know of a supported non-interactive way to read usage, I would like to hear it. As far as I can tell, the steps above are the only path today.