tui: render PTY scroll-up from the structured message log

[edwin-zvs]

Problem

Scrolling up an interactive harness session in the TUI showed a broken history:

• codex pins its input box with a DECSTBM scroll region. vt100 0.16 only saves a scrolled-off line to scrollback when no region is active (grid.rs::scroll_up → !scroll_region_active()), so codex's conversation lines are dropped — scrollback is empty, and the viewport-snapshot fallback fragments it (duplicated banner, spliced status, only a few lines recoverable).
• claude sets no region but rewrites lines with cursor-up + a bordered input box. The snapshot fallback then duplicates answer lines and splices chrome (spinner, borders, input box) into them.

The web client doesn't have this — xterm.js keeps a real native scrollback for the same byte stream.

Fix

For PTY sessions that emit structured Message/Reasoning events and have no tool blocks, render scroll-up from the structured message log — rebuilt and reflowed to the current width. The live view stays the faithful PTY; a raw shell (no messages) keeps the snapshot path; tool-block sessions keep replay_full. Messages are fed into the shadow on both the hydration/replay and live event paths.

Rather than vendor/patch vt100 to match xterm.js's scrollback semantics, this uses the clean, complete, reflowable source the harness already provides.

Design rationale: specs/0017-pty-scrollback-from-messages.md.

Evidence

Reproduced claude's break by feeding a real interactive-claude capture through the shadow path:

distinct=69/80  duplicated=[49×2 … 76×3, 77×3, 78×3, 79×3]
❯ Output the integers 1 through 80 …
✶ Scampering…                          ← spinner spliced into history
⏺ 1
──2─────────────────────────────  3    ← "2"/"3" mangled into a border line

Tests (224 green):

• app::tests::pty_session_scrollback_uses_message_log — full hydration path → scroll-up renders the clean message log (❯ prompt, no PTY chrome), and the live view stays the PTY.
• pty_render::tests::codex_scrollback_renders_clean_message_log — codex scroll-region case.
• pty_render::tests::claude_style_redraw_scrollback_uses_messages — claude cursor-up redraw case: no chrome, no run-together, all answers scrollable.
• pty_render::tests::shell_session_without_messages_keeps_natural_scrollback — message-less sessions keep the snapshot path.

Binary

Code is entirely in crates/cli → the relevant binary is agent (the TUI). agentd / agentd-mcp / agentd-adapter-* build but are unchanged from main.

No live screenshot: interactive sessions in a fresh isolated daemon stall on first-run folder-trust / sandbox prompts that one-shot input can't fully drive. Engagement was confirmed via temporary debug instrumentation (render_shadow … msg_mode=true) on an interactive codex session, then removed.

🤖 Generated with Claude Code

[edwin-zvs]

Closing — this approach (rendering scroll-up from the structured message log) is a net regression in practice: it shows only the prose messages and drops the tool calls / command output / diffs that make up most of a coding session. The current main (snapshot scrollback) keeps the full terminal history, which is more useful even with its imperfections.

Pivoting to fixing the scrollback itself: patching the vt100 backend so codex's scroll-region-scrolled lines actually land in scrollback — the root cause (grid.rs::scroll_up only pushes to scrollback when no scroll region is active, whereas xterm/xterm.js save when the region's top margin is row 0). New PR to follow.