# AI authoring limitations and troubleshooting

<PageBadges />

## Current preview boundaries

AI authoring is intentionally limited to interactive creation and editing of form0 schemas. This
preview does not include:

- unattended or headless generation and automatic application
- record completion, extraction, inference, or submission analysis
- arbitrary filesystem, shell, web, MCP, extension, or third-party skill access
- image, PDF, or other multimodal form generation
- general custom-helper generation outside calculations and form-event code
- embeddings, vector databases, or a hosted retrieval service
- automatic truncation, compaction, or sectioning of oversized forms
- a form0-hosted provider gateway, billing, telemetry, or credit reporting
- AI interfaces in form0 bindings, playgrounds, landing pages, or reform

Commands, provider availability, prompts, and mutation behavior do not yet have stable compatibility
guarantees.

## No model is selected

Authentication and model selection are separate. Inspect providers, authenticate if necessary, and
select an exact model identifier:

```ansi
form0(ai)> /providers
form0(ai)> /login <provider> <api_key|oauth>
form0(ai)> /models <provider>
form0(ai)> /model <provider>/<model>
```

Use `/status` or `/model` without arguments to verify the current selection.

## A saved model changes after restart

The selected model is restored from the per-schema Pi conversation. If it no longer exists or its
credential is unavailable, the CLI displays Pi's fallback warning. Reauthenticate the original
provider or select another model. The schema itself is unchanged.

## Authentication succeeds but requests fail

Successful authentication does not prove that an account has remaining credits, an active
subscription entitlement, model access, or a reachable endpoint. Provider failures and usage-limit
messages are surfaced by the CLI. Resolve the account or provider issue, then retry; `/status` does
not contact the provider or report balances.

For OAuth in a headless environment, choose a device-code method when the provider offers one. A
browser flow returns to the AI prompt automatically after successful authentication.

## The complete form does not fit

The CLI estimates whether the complete schema, installed authoring catalog, request, and response
allowance fit the selected model context window. If they do not, the request is refused before data
is sent. The preview does not truncate or divide the form. Choose a model with a larger context window
or reduce the schema outside AI mode.

## Apply reports a stale revision

The schema changed on disk after the proposal was created. Apply is deliberately blocked. Discard the
draft, reload the current schema, repeat or revise the request, and inspect the new diff. The CLI does
not automatically merge external edits into an AI proposal.

## Remote documentation is unavailable

Remote form0 documentation is supplementary. Timeouts, redirects, unsupported responses, offline
operation, or an unavailable documentation host do not grant broader web access and do not weaken
validation. The agent continues with the installed core catalog.

## A request appears stuck

Use `/status` to inspect activity and the follow-up queue. Use `/cancel` to abort the current provider
request and clear queued messages. If cancellation itself fails, the CLI reports that error and keeps
the session available.

## A proposal is repeatedly invalid

Read the returned diagnostics and make the intent more explicit: identify fields by label or key,
state expected types, clarify repeatable scope, and describe calculation or event behavior. Installed
schema validation remains authoritative; do not bypass diagnostics by asking for raw file output.

See [Commands and status](/cli/ai-authoring/commands-status) and
[Authoring workflow](/cli/ai-authoring/authoring-workflow) for the associated recovery commands.