Feature: collapsible/pinned output sections for large assistant responses

Problem

When Claude Code produces large outputs — multi-agent fan-out results, architecture design syntheses, long code diffs — they scroll off the visible area immediately. Once scrolled past, there's no way to keep them visible while continuing to work in the same terminal session. You're forced to choose between reviewing the output and continuing the conversation.

This becomes acute when:

• Running parallel subagent fan-outs (e.g., 5 agents each returning 500+ word design specs)
• Getting a synthesis/decision matrix that you need to reference while answering follow-up questions
• Long code reviews or plans that the next prompt should respond to

Desired behavior

Option A — Collapsible/accordion sections
Large assistant outputs (above a configurable line threshold, e.g. 50 lines) render with a collapsed summary header and an expand toggle. The summary stays visible in the scroll-back buffer; the full content is accessible on demand.

Example:

▶ [Flynn — Ingress Design] (47 lines) — click to expand
▶ [Rinzler — Isolation Design] (62 lines) — click to expand  
▼ [Synthesis — 6 decisions, 3 open questions]
   ... full content visible ...

Option B — Pinned output panel
A secondary scrollable pane (like a split terminal) that retains the last N lines of assistant output independently of the chat scroll position. User can scroll the pinned pane while the chat continues.

Option C — "Stash to file" shortcut
A keyboard shortcut or inline action on any assistant response block to write it to a temp file and open it in the editor (VSCode split pane). Simpler than a full UI change; works with existing editor split-pane workflows.

Workaround currently used

Manually writing synthesis outputs to vault/scratch files and opening them with code -r <path> in a VSCode split. Works but requires explicit delegation — the output gets printed to chat AND written to a file, which is redundant.

Context

Running Claude Code in VSCode extension mode. Parallel subagent fan-outs (using the Agent tool with multiple simultaneous calls) produce the most acute version of this problem — all results return at once as a large block that immediately fills and exceeds the visible terminal area.