# AI authoring commands and status

<PageBadges />

## Command reference

AI mode accepts natural-language requests and slash commands. Unprefixed input is always sent to
the model; every deterministic command starts with `/`. For example, use `/preview`, not `preview`.

| Command                        | Purpose                                                                       |
| ------------------------------ | ----------------------------------------------------------------------------- |
| `/help`                        | Show the aligned AI command summary.                                          |
| `/status`                      | Show schema, draft, model, authentication, policy, activity, and queue state. |
| `/providers`                   | List Pi providers, authentication methods, and model counts.                  |
| `/models <provider>`           | List model identifiers for one provider.                                      |
| `/login <provider> <method>`   | Authenticate using the Pi-defined `api_key` or `oauth` flow.                  |
| `/model [provider/model]`      | Open the model picker, or select an explicit provider and model.              |
| `/privacy`                     | Explain the effective schema cloud and consent policy.                        |
| `/load [file\|form]`           | Load another form and switch to its schema-specific conversation.             |
| `/preview`                     | Print the friendly form tree for the working schema.                          |
| `/json`                        | Print the working schema as raw JSON.                                         |
| `/validate`                    | Validate the working schema, including an unapplied proposal.                 |
| `/serve [options]`             | Start the preview server; `--app` also starts the configured app.             |
| `/serve status`, `/serve stop` | Inspect or stop preview servers without leaving AI mode.                      |
| `/diff`                        | Show cumulative changes since entering AI mode.                               |
| `/diff --pending`              | Show only changes between the last apply and the pending proposal.            |
| `/apply [schema-path]`         | Recheck and atomically save the pending proposal.                             |
| `/discard`                     | Abandon the draft and restore the committed browser preview.                  |
| `/undo`                        | Stage the last schema saved in this session for explicit approval.            |
| `/new`                         | Start a new conversation while retaining the selected model.                  |
| `/clear`                       | Delete stored history for this form and start again with the selected model.  |
| `/cancel`                      | Abort the active request and clear queued follow-ups.                         |
| `/exit`                        | Return to the shell; confirm first when a proposal is unapplied.              |

## Status

Status is local and does not contact a provider or retrieve credit balances:

```ansi
form0(ai)> /status

AI authoring status
  Schema:             /project/form.schema.json
  Draft:              none
  Provider:           openai-codex
  Model:              gpt-5.6-luna
  Cloud policy:       allowed
  Conversation:       persisted
  Activity:           idle
  Queued follow-ups:  0
  Authentication:
    openai-codex  oauth (stored, subscription) [selected]
```

Authentication output identifies the source and type but never prints a credential. A provider can
be authenticated without being selected, and authentication does not imply available credits.

## Ready and busy prompts

A request changes the prompt state and prints the active model:

```ansi
[AI] Thinking with openai-codex/gpt-5.6-luna…
form0(ai,busy)>
```

With the preview server running, the prompt is `form0(server,ai,busy)>`. When processing completes,
it returns to `form0(ai)>` or `form0(server,ai)>`.

Pressing Enter on an empty line is a no-op that redraws the current prompt. Natural-language input
entered while busy is queued in FIFO order and acknowledged with its queue position.

## Commands available while busy

While a provider request is active, `/help`, `/status`, `/privacy`, `/cancel`, `/serve status`, and
`/serve stop` remain available. Other slash commands are rejected until processing finishes,
preventing model, conversation, or draft state from changing underneath the active request.

Cancellation asks Pi to abort the current request and clears every queued follow-up. A successful
cancellation returns to the ready prompt without surfacing the provider abort as a second error.