pi-sticky-last-prompt

Pins the most recent user message you have scrolled completely past as a one-line bar at the very top of pi's fullscreen TUI. Left-click the bar and the transcript scrolls so that message sits right below it — a quick way to jump back to what you asked.

Install

pi install ./extensions/sticky-last-prompt          # local
pi install npm:pi-sticky-last-prompt                # npm (if published)

pi's git sources clone a whole repository and install what its root package.json declares; the pi-collections root is never published and declares no pi manifest, so there is no git:.../pi-collections/... form. To install just this extension from the repo, clone it and pi install the subdirectory locally, or publish it to npm.

Or add it to the packages array in ~/.pi/agent/settings.json:

{
  "packages": ["./extensions/sticky-last-prompt"]
}

Behavior

• The bar shows the latest prompt that has scrolled completely above the top of the viewport — the newest message with not a single row left on screen, so the pinned jump always targets something invisible. While a prompt is crossing the top edge (head gone, tail still visible) the bar hides rather than duplicate text sitting directly beneath it; scroll down until the message is entirely out of view and it becomes the pin. While a newer prompt is crossing, older fully-hidden prompts do not keep the pin — the bar just stays blank until the newer one clears the edge. Text is whitespace-collapsed to one line, ellipsized if too long, and themed with your active theme (accent icon on a selectedBg strip). Skill invocations count as prompts too — pi renders them as collapsible [skill] name blocks rather than user messages — and are pinned under that same [skill] <name> label.
• Left-click anywhere on the bar to scroll the transcript to the message currently shown. The view lands just below the bar, and follow-tail is disabled so new output doesn't yank you back — exactly like pi's built-in search jump.
• Everything else keeps stock behavior: mouse wheel, text selection, scrollbar, right-click paste, middle click, drags. The extension swallows precisely one gesture — a left-button press inside the bar row.
• On /reload or when starting on an existing session, the bar re-derives from the rendered transcript — no seeding step; it just tracks wherever the viewport opens.

Requires fullscreen TUI mode (pi ≥ 0.84); in regular mode the extension stays dormant.

How it works

The pin is a non-capturing full-width overlay anchored top-left; its text is resolved from the live transcript tree on every paint (message offsets cached per width + content height), with no polling timers. Click interception wraps the renderer instance's internal selection handler — in pi 0.85.x, click events are consumed centrally before any public extension API can see them, and that handler is the only seam left. The same instance patching teaches hasOverlay() to ignore our non-capturing bar (and only ours), so the stock scrollbar and text selection stay live while it is shown. All internal access is defensive: if a future pi renames these pieces, clicking stops working but nothing else breaks.

Known tradeoffs

• While the bar is visible, scrollbar dragging and mouse text selection keep working: the extension reports "no overlay" to pi while its own non-capturing bar is the only overlay shown. When another overlay is on top (search box, dialogs), stock pi suppression applies.
• The bar covers the top transcript row while shown.
• Message offsets are re-measured whenever content height changes (e.g. while a response streams in), which adds one extra render pass of the transcript per such frame; between those frames everything is cached.
• Transcript rebuilds (/tree navigation, compaction, session load) swap the transcript container's children array wholesale; the offset cache keys on that identity, so the bar re-derives immediately from whatever messages the new branch renders — no stale labels even at identical heights.