UX: way to attach a question/annotation to a specific fragment of agent output instead of typing it as a new prompt

Problem

In long-running CLI sessions with rich on-screen output (tables, ledgers, status snapshots, code blocks, monitor events), there is no way for the user to point at a specific piece of data and attach a question or comment to it.

Everything has to go through the next-prompt text channel. If the agent has just printed a 17-row table, the user has to describe the row in words ("the 5/7 row" / "the second LAPACK entry") or paste a fragment back. That's slow, and the slowness often causes the user to defer the question — which then queues up behind a long-running background task instead of being asked immediately while the context is fresh.

Concrete example (from today's session)

Agent printed a status table with 17 rows. User wanted to ask about one specific row's pass rate. The friction of describing-by-content was high enough that the user typed a one-letter placeholder (a) intending to come back to it, then had to explain that he was selecting option a from an earlier multi-choice prompt that had since scrolled past.

What would help

Some affordance to attach a comment/question to a specific fragment of agent output:

• Click (or keyboard-navigate) to a span of output and pin a question to it
• Auto-generated stable references ([R5], [L12], etc.) the user can quote in their next prompt
• Or — even simpler — an "interrupt" / "annotate this" affordance that lets the user inject a question into the current context with a visible anchor, rather than having it queue up at the end of the prompt stream

The bigger pattern this surfaces: in agentic CLI sessions, the user's cognitive bandwidth is the bottleneck, and forcing them to describe-by-content rather than point-by-position taxes that bandwidth in proportion to how rich the agent's output has become.

Filed on behalf of a real long-session user via the agent itself; happy to provide a longer transcript if useful.