feat: paginate the eval --rich dashboard with left/right arrow keys

[mikasenghaas]

Summary

• When rollout rows overflow the eval --rich dashboard, the left/right arrow keys now page through them.
• Auto-advance is preserved: the 5s timer drives paging until the first arrow press, then hands off seamlessly (continuing from the page currently on screen) and stays where the user leaves it. The page count only grows as rollouts start, and the manual page is clamped so a terminal resize can't leave it out of range.
• Arrows are inert until there's more than one page: the Pager tracks the page count (set each render by _paginate) and ignores presses while a single page fits, so a stray arrow before rollouts overflow can't disable auto-advance or offset the starting page once paging begins.
• Key reading lives in the shared live_view engine (dashboard/base.py) behind an optional on_key handler: stdin is put in cbreak mode and read via the asyncio event loop (no extra thread), buffering a partial escape across reads (cbreak can split a key's 3 bytes over callbacks). Each press redraws immediately (rather than waiting for the next 0.25s refresh tick) so there's no lag. Because updates go through the same live.update(..., refresh=True) path as the timer, there's no extra flicker. setcbreak runs inside the try, so the terminal is always restored on exit even if setup raises.
• No-op when stdin isn't a terminal (POSIX-only), so non-interactive/legacy runs are unaffected. validate_dashboard, which shares live_view, is untouched (it passes no on_key).
• The page indicator gains a discreet ◄ ► hint ((page 2/6 ◄ ►)) so the controls are discoverable.

Verification

Drove the real dashboard/live_view pipeline inside a pseudo-terminal (real cbreak, real asyncio add_reader, real arrow-sequence parsing, Pager + _paginate, rich Live) with real arrow-key bytes:

• initial (no keypress): auto-advance active (manual=False), page rotates on the timer
• → read within 0.12s (< the 0.25s tick): page already advanced — confirms the immediate redraw (no lag)
• → →, ←: step correctly through pages, continuing from the on-screen auto page
• ← ×9 and → ×20: clamp at page 1 and the last page (no wrap, no out-of-range)
• split escape: ESC and [C sent in separate writes with a 0.35s gap (past a refresh tick) — still pages
• single-page gate: with rows that fit one page, right/right/left leave manual=False and the page at 1/1
• restore-after-raise: forcing add_reader to raise after setcbreak in a real pty — ICANON and ECHO are restored afterward
• stdin = /dev/null (non-tty): key reader no-ops, auto-advance still runs, clean exit — no crash

ruff check, ruff format, and ty check verifiers all pass.

🤖 Generated with Claude Code

Note

Add left/right arrow key pagination to the eval rich dashboard

• Adds a Pager class in eval.py that tracks pagination state: auto-advances pages every _PAGE_SECONDS until the user presses an arrow key, after which it follows manual input and clamps within bounds.
• Adds a _key_reader context manager in base.py that, on POSIX TTY stdin, uses cbreak mode and an asyncio reader to detect left/right arrow escape sequences and dispatch them to an on_key callback with an immediate redraw.
• The page indicator in the stats line now shows (page X/Y ◄ ►) when pagination is active.
• Behavioral Change: key reading is a no-op when on_key is not provided, termios is unavailable, or stdin is not a TTY, so non-interactive environments are unaffected.

^{Macroscope summarized 54acea6.}

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want reviews to match your repository better? Bugbot Learning can learn team-specific rules from PR activity. A team admin can enable Learning in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 8a43f69. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved

Self-contained CLI UX enhancement adding arrow key pagination to the eval dashboard. Changes are isolated to display logic, don't affect eval processing, and the author is the original maintainer of this dashboard code.

^{You can customize Macroscope's approvability policy. Learn more.}