@mrclrchtr/supi-prompt-suggestions

Advisory ghost-text prompt suggestions for the pi coding agent.

Install

pi install npm:@mrclrchtr/supi-prompt-suggestions

What it does

After each assistant response, the extension can suggest a concise next prompt as dim ghost text in the editor.

Suggestions only appear when:

• you selected a suggestion model in /supi-settings
• the editor is empty
• the model finds a useful follow-up

Using suggestions

Key	Action
→ / Right Arrow / Tab	Accept the suggestion into the editor without submitting it
Esc	Dismiss the visible suggestion
Any text input	Hide the suggestion and continue typing
Delete all editor text	Restore the hidden suggestion

Accepted suggestions are inserted into the editor so you can edit them before sending. A hidden suggestion remains available until you accept it, dismiss it, or start new agent work.

Configuration

Open /supi-settings → Prompt suggestions.

• Suggestion model — choose disabled or one of PI's scoped enabled models

The default is disabled. Pick a cheap, fast model if you want lightweight suggestions after assistant turns.

Settings follow SuPi's normal scoped config behavior: set a global default, then override it per project when needed.

Suggestion requests use PI's model registry. PI owns authentication and provider routing; the extension does not resolve or copy credentials.

Privacy

When suggestions are enabled, the suggestion model receives only the last assistant message, trimmed to the final 8,000 characters.

The extension does not send:

• the full conversation transcript
• tool outputs
• file contents
• project metadata
• session metadata in the prompt

The request can include a separate opaque provider-routing identity. It is not part of the prompt and is derived from the PI session, provider, model, and prompt-suggestion stream.

Troubleshooting

If no suggestion appears:

• confirm Suggestion model is not disabled
• confirm the selected model is still enabled in PI for the current scope
• if a warning reports authentication failed, check the selected model's authentication in PI
• make sure the editor is empty after the assistant finishes
• wait for the suggestion spinner to finish; generation times out after about 20 seconds

Request failures show one warning for the active model and session. The warning contains only the provider/model identifier, a safe failure category, and an optional HTTP status. Categories include authentication failed, billing failed, quota exceeded, rate limit exceeded, timeout, and provider request failed. Repeated failures stay quiet until a request succeeds, settings change, or a new session starts. Warning state is held in memory, so /reload also starts a new warning cycle. The separate routing identity remains stable for the same PI session.