Feature: /btw command for quick ephemeral side questions

[northaxosky]

Describe the feature or problem you'd like to solve

During long coding sessions, I often want to ask a quick side question ("what does --frozen-lockfile do?", "what's the difference between let and const?") without polluting my conversation context or burning extra tokens on context growth.

Claude Code has a /btw command for exactly this — ephemeral, tool-free, one-shot Q&A that doesn't enter the conversation history.

Proposed solution

Add a built-in /btw slash command that:

1. Makes a separate, lightweight model call outside the main conversation — the question and answer are not appended to session history
2. Displays the answer in a temporary overlay or inline block (dismissed with Escape)
3. Uses no tools — just a direct model response
4. Does not increase token usage for subsequent turns (zero context pollution)

This would require exposing a "side-channel" or "ephemeral" model inference API, which extensions currently can't access. The extension SDK's session.send() and session.sendAndWait() both write to the main conversation. A new session.complete() or session.ask() method that returns a one-shot model response without touching conversation history would make this possible natively and also unlock a powerful primitive for extension authors.

Example prompts or workflows

> /btw What does the --frozen-lockfile flag do in npm?
It tells npm ci to fail if package-lock.json would need updating, ensuring reproducible installs.

> /btw What's the syntax for a TypeScript discriminated union?
type Shape = { kind: "circle"; radius: number } | { kind: "square"; side: number }

> /btw How do I revert the last git commit but keep changes?
git reset --soft HEAD~1

> /btw What HTTP status code means "rate limited"?
429 Too Many Requests

Additional context

I built a proof-of-concept extension that approximates this behavior using the onUserPromptSubmitted hook: northaxosky/copilot-btw

It intercepts /btw prompts and wraps them with instructions for brief, tool-free answers. However, because the SDK has no out-of-band model inference API, the Q&A still lands in conversation history. A first-party implementation with a true ephemeral model call would be the ideal solution.