tui: vendor vt100 with a scroll-region scrollback fix

Cargo.toml

@@ -75,3 +75,12 @@ strip = "symbols"
 # (fast incremental rebuilds, full debuggability).
 [profile.dev.package."*"]
 opt-level = 3
+
+# Vendored `vt100` fork (see `vendor/vt100`). Upstream `Grid::scroll_up`
+# only pushes scrolled-off lines to scrollback when no scroll region is
+# active, which drops the entire conversation for children that pin a
+# footer with a top-anchored DECSTBM region (codex: `ESC[1;Nr`). The fork
+# saves whenever the region's top margin is row 0, matching xterm /
+# xterm.js, so PTY scroll-up shows the real, complete terminal history.
+[patch.crates-io]
+vt100 = { path = "vendor/vt100" }

specs/0017-pty-scrollback-is-real-terminal-history.md

@@ -0,0 +1,66 @@
+# 0017-pty-scrollback-is-real-terminal-history
+
+Status: accepted
+Date: 2026-06-01
+Area: tui
+Scope: What the TUI shows when the user scrolls back a PTY-backed session.
+
+## Decision
+
+Scrolling back a PTY session shows the **real terminal history** — every
+line the child actually painted (prose, tool calls, command output,
+diffs), in order — not a semantic substitute. The history comes from a
+faithful terminal emulation of the child's byte stream.
+
+To make that possible for children that pin a footer with a scroll
+region, agentd **vendors the `vt100` crate** (`vendor/vt100`, wired via
+`[patch.crates-io]`) with one change to `Grid::scroll_up`: a scrolled-off
+line is saved to scrollback whenever the scroll region's **top margin is
+row 0** (`scroll_top == 0`), instead of only when no region is active.
+
+## Reason
+
+The full terminal history is what's useful: a coding session is mostly
+tool calls and output, so any approach that scrolls only the conversation
+prose drops most of the content and is a net regression (we tried it).
+
+But the emulator must actually retain that history. Upstream `vt100`
+0.16.2 only pushes scrolled-off lines to scrollback when there is no
+active scroll region. Codex pins its input box with a top-anchored
+DECSTBM region (`ESC[1;Nr`), so its entire conversation scrolled off into
+nothing — `maxsb == 0`, and scroll-up showed an empty or
+snapshot-reconstructed mess. Real terminals don't behave that way: xterm
+(and xterm.js, which is why the web client never had this bug) save lines
+that scroll off the top of the screen whenever the region's top margin is
+the top line. The fork brings `vt100` in line with that.
+
+Vendoring is accepted over the alternatives: there is no upstream
+callback to hook the scroll event, and reconstructing scrollback from
+viewport snapshots is lossy. The fork is one line in one function.
+
+## Consequences
+
+- PTY scroll-up shows the complete, faithful terminal history for
+  scroll-region TUIs (codex) the same way it already did for
+  line-oriented children (claude/shell) — no per-harness handling.
+- The vendored `vt100` must be kept patched across upstream bumps; the
+  change is isolated to `Grid::scroll_up` and marked in-source as an
+  `agentd fork patch`. Prefer upstreaming it so the fork can be dropped.
+- No agentd code changes are required: the existing scrollback path fills
+  correctly once the emulator retains the lines.
+
+## Non-Goals
+
+- Not altering the live (non-scrolled) render.
+- Not changing alt-screen apps (vim/htop), which have no scrollback by
+  design.
+- Not a general rewrite of the scrollback emulation; only the
+  region-top-row-0 rule is corrected.
+
+## Examples
+
+- codex prints a long answer above a footer pinned by `ESC[1;37r`. Before:
+  scroll-up was empty / fragmented (`maxsb == 0`). After: the full
+  conversation is in scrollback, in order.
+- claude appends lines (full-screen scroll, no region) — unchanged; it
+  already populated scrollback.

vendor/vt100/.cargo-ok

@@ -0,0 +1 @@
+{"v":1}

vendor/vt100/.cargo_vcs_info.json

@@ -0,0 +1,6 @@
+{
+  "git": {
+    "sha1": "eb66ffaf7d771c13303ef73b29f6f2a56fdacecf"
+  },
+  "path_in_vcs": ""
+}