[FEATURE] Priority-aware compaction: let users mark sidebar questions as low-priority so they get compressed first

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

A long session is rarely one uniform conversation. In practice it's one main task
plus a scattering of tangential exchanges — quick brainstorm questions, side
curiosities, one-off "while you're here" lookups. These are useful in the moment
but irrelevant to the main task's global state.

Compaction currently treats the entire history as one flat sequence. Every
exchange competes for the same summary budget, so a throwaway sidebar discussion
can survive compaction at the same fidelity as a critical architectural decision,
and vice versa. After a couple of auto-compacts, main-task decisions get lossy
while summarized side threads are still taking up space.

Proposed Solution

Give messages (or exchange segments) a priority level, and make compaction
consume low-priority segments first:

1. Explicit marking — a lightweight way to flag a question as a sidebar at

the moment I ask it, e.g. /aside <question> or a prefix. The whole exchange
it spawns inherits the low-priority tag.

1. Compaction order — when compacting, collapse low-priority segments to a

one-line note (or drop them) first, and keep the main thread verbatim for
as long as the budget allows. Only start summarizing the main thread once the
sidebars are gone.

1. (Optional, later) Automatic segmentation — infer topical side-branches

and propose them as low-priority before compacting.

Even just (1) + (2) would help a lot: the user is the one who knows which
questions were brainstorming noise, and today there is no way to tell the
compactor.

Alternative Solutions

• /compact <instructions> works, but requires me to remember and describe every

side thread at compaction time — and does nothing for auto-compact.

• Rewind (double-Esc) prunes the branch entirely, but destroys it — sometimes I

still want the one-line conclusion.

• Subagents keep exploration out of the main context, but ad-hoc brainstorm

questions naturally happen inline in the main thread; nobody spawns an agent
for a 30-second curiosity.

Priority

Medium - Would be very helpful

Feature Category

CLI commands and flags

Use Case Example

_No response_

Additional Context

#7359 (selective /compact options) and #34299 (preserve user inputs during
compaction) point at the same underlying gap: compaction has no notion of
importance. This proposal adds the missing signal — message-level priority —
that both would benefit from.