A tile's refresh (a {t:'r'} or {t:'c'} frame, every reconnect) wiped the
pane with a synchronous xterm clear() at the load's turn, BEFORE its fetch,
and wrote live frames straight through the fetch and the replay. That is
the replay clear CLAUDE.md "Terminal resilience" forbids: bytes still
queued in xterm are parsed after a synchronous clear and fuse into the
snapshot, and clear() keeps the cursor's row, column, SGR and margins, so
the capture (raw rows, no home) started wherever the cursor sat. A failed
or empty fetch left the tile blank.
The refresh now runs in the primary pane's order (_onSessionNeedsRefresh,
_resetTerminalForReplay):
- fetch first, so the tile keeps its last frame through the round trip and
through a grid tile's wait in the load queue;
- from the response on, live frames are held in _liveQueue with their
arrival time, as _pullHistory already did, and the body read of a bounded
window (grid tile, shell) gets the pull's 10 s budget, while Pane B's
unbounded full=1 keeps the request's own budget;
- then the queued in-stream \x1bc immediately before the replay;
- then the held frames that arrived after the response (_flushLiveQueue,
now shared with _pullHistory), then the owed marker.
A failed, aborted or empty fetch writes nothing and resets nothing.
The _stampMarkerIfOwed guard for a pending trailing refresh stays (that
refresh settles the marker itself either way); only its rationale changed.
The fake xterm now treats an in-stream RIS like clear() in its row
emulation.
Tests: the ones that counted clear() calls on the refresh path now count
the in-stream reset instead, assert it sits right before the replay and
that clear() is never called (unit single-flight block, the marker
ordering tests, the reconnect test, the grid {t:'r'} and marker tests, and
the scroll test's server-clear overflow case, which now goes through a
refresh). New: the screen is untouched on a failed or empty fetch and on a
failed body read (held frames written in order), frames before the
response are written through and later ones held behind the replay, the
cutoff drops frames the capture covers, the body budgets, and a grid tile
keeps its last frame through its own capture's round trip.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The server sends {t:'c'} from one place only: a fresh Claude pane's first
prompt (Session.startInteractive), meaning "refresh after startup". The
primary pane answers it with a refetch and replay (_onSessionClearTerminal),
and stands aside while the grid is open, so the tile's own handling was the
only one that ran. That handling was a bare xterm clear(), which keeps only
the cursor's row and drops the banner, a resumed transcript and all
scrollback. An idle Claude never repaints static rows, so a Claude session
Run into the grid, or Attached in a tile, came up as a near-empty tile.
_onLiveClear() now calls _refreshBuffer(), the {t:'r'} path: single-flight,
coalesced into one trailing refresh behind a load already running (a shell
pull's held frames included), and paced by the grid's TileLoadQueue. The
queued {clear:true} entry and its branch in _pullHistory's flush are gone,
along with the _clearTerminal helper they used.
Tests: two unit tests pinned the bare clear (a clear frame queued in order
during a pull, and one applied at once before the capture); they are
replaced by tests that the frame coalesces behind the pull and refetches,
plus a socket-level {t:'c'} test, a coalescing test, and a grid test that
the frame waits its turn in the load queue.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
#555 made the primary pane page opencode's transcript with PageUp/PageDown
from the wheel, because opencode draws in place on the alternate screen and
leaves the browser's buffer with no scrollback. A TerminalTile (a grid tile,
the split's Pane B) left every wheel to xterm, so in an opencode tile the
wheel scrolled nothing, or only stale rows.
The tile now runs the primary pane's own gates aimed at itself (its terminal,
its session, never the active one): xterm's tracking mode, the Claude
forwarding gate, then the hollow-buffer test. A wheel that passes them is
consumed in the capture phase and turned into PageUp/PageDown through the
shared pageKeysForTravel math, coalesced per tile (40 ms, 512 bytes, the twin
of the primary pane's queue) and sent ephemeral on the tile's own socket.
Every other wheel stays with xterm as before, the shell history pull
included. The file names no CLI: the mode rules stay in terminal-ui.js, and
terminal-tile.js joins the frontend no-id-branching guard.
A plain port of the primary's baseY === 0 test would almost never fire in a
grid. A tile's first capture is taken at the PTY's previous size (usually the
taller primary pane's) and written into a shorter xterm, and its own
row-shrinking fits (zoom-out, divider drags, tile count changes) push more
rows above the screen. The tile counts those rows as its own overflow: all of
them after a load whose capture held a single screen (the server's
captureRows), plus whatever a local fit or a PTY geometry report pushes up,
reset by a clear and clamped to baseY. The paging gate gets baseY minus that
count. Output that scrolls real lines still counts as history, so the tile
stops paging there.
#555's other half, stripping opencode's mouse DECSETs so a drag selects text,
is server-side and already reached tile sockets. It also left the tile's
xterm unable to encode opencode's clicks, so the tile now installs the
primary pane's desktop click report (bubble phase, gated on the session's
cliMouseTracking, the tile's own link hover and selection). Both listeners,
the flush timer and the page-key state are torn down in destroy().
Still out of scope, as the fileoverview now says: touch paging (tiles have
no touch path) and SGR wheel forwarding to Claude's fullscreen renderer
(tile-grid-plan follow-up 4), so a fullscreen Claude tile keeps leaving the
wheel to xterm.
Tests: test/terminal-tile-scroll.test.ts drives a real tile in the vm
harness (session targeting, every no-page case, accumulation, the cap,
coalescing, byte parity with the primary pane, the overflow discount through
a load, a fit, a geometry report and a clear, the click report and destroy);
the discount cases fail with it removed. The fake xterm gains opt-in row
emulation. test/terminal-tile-scroll.browser.test.ts checks the same model
against a real xterm with trusted wheel events (browser suite, not the gate).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
terminal-tile-unit slices connect() out of terminal-tile.js up to
'async _loadBuffer()'. The load queue commit (fca7acd0) gave _loadBuffer a
`{ refresh }` parameter, so that anchor stopped matching, indexOf returned
-1 and the slice ran to the end of the file: every check in the test
passed against code outside connect(). The anchor is now
'async _loadBuffer(' and the test asserts both anchors resolve, so a rename
fails it instead of widening it (checked by putting the old anchor back).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A tile's replay (writeChunked) wrote its capture 32 KB per animation
frame, so a 1 MiB load took about a second of frames, and in the grid the
load queue's slot was held across all of it: tile N+1's capture waited for
tile N's last frame. xterm 6 already parses its write queue in 12 ms
slices and yields between them, so the slices now all go in at once (up
to a 1 MiB window, since xterm's queue throws past 50 MB and Pane B's
unbounded full=1 capture can reach the server's 32 MB) and the replay
resolves on the callback of an empty write queued behind them, i.e. once
xterm has parsed the last slice. The single-flight flag is still held for
the whole replay. A disposed xterm never runs that callback, so destroy()
now settles a replay in progress: a removed tile can no longer hold its
flag or the grid's one load queue. Queued up front, the capture also stays
in one piece during a refresh: live output written meanwhile lands after
it, not between two of its slices.
Measured (tileperf, 6 printing shells with 1 MiB histories, headless,
n=3 interleaved A/B against the starting file, load 8.6 to 11.8):
- grid fresh open, 6 tiles, all painted: 5.10 s -> 2.57 s (-50%);
restore after reload: 6.49 s -> 3.86 s (-41%); per-tile replay
669 to 734 ms -> 298 to 321 ms (median).
- Same work in half the time: frames over 20 ms 26% -> 40% of the
(shorter) load window, about 86 -> 62 slow frames in all; longest long
task on restore 304 -> 227 ms; server event-loop delay unchanged
(max 111 to 122 -> 122 to 134 ms, one capture in flight throughout).
- Split Pane B (the other TerminalTile) with the main terminal on WebGL
and its long-task guard armed: load 1.6 to 3.8 s -> 0.8 to 1.7 s over
15 loads each; 0 long tasks of 200 ms or more either way, the guard
never tripped. With an unbounded full=1 capture (about 21k lines):
2.5 to 3.4 s -> 1.9 to 2.6 s, 0 long tasks of 200 ms or more.
- At checkpoint 1 (equivalent patch, n=3 to 6): fresh 6.4 -> 2.7 s,
restore 8.9 -> 4.2 s, TUI-style reconnect 11.2 to 11.8 -> 6.2 s.
Tests: the replay queues every slice at once and holds the flag until
xterm has parsed it; a replay larger than the window goes one window at a
time; a pane destroyed mid-parse settles at once; in the grid, a tile
destroyed while xterm still parses its replay releases the queue and the
next tile loads (fake xterm whose callbacks never run). The rAF-driven
tests now hold the parse callbacks instead. All mutation-checked (no
settle in destroy, settle before the parse, no window). Browser
split-pane-terminal: same 1 failed / 2 passed as at the starting HEAD
(the failure is in the test's own setup, before connect).
Scope: PR 1 (terminal-tile.js writeChunked and destroy(); Pane B replays
the same way). Moving it onto PR 1 needs its two call sites adapted
(PR 1 has no _runLoad yet) and leaves the tile-grid-load-queue.test.ts
hunk with PR 2.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
GET /api/sessions/:id/terminal runs synchronous tmux calls on the server, so
N tiles loading at once would stall every WebSocket and SSE stream back to
back (and after a deploy restart all N reopen within the same second).
TerminalTile takes the options PR 1 deferred to the grid:
- scheduleLoad(tile, kind, run): every capture the tile fetches (initial
load, reconnect refresh, server {t:'r'} refresh, shell history pull) runs
when its owner says so. Absent (the split's Pane B), a load runs at once.
- scrollback (the grid passes TILE_SCROLLBACK) and fontSize.
- boundedLoad: a TUI tile loads the bounded full=1&tail= window, never its
whole history.
TileLoadQueue (terminal-tile.js, DOM-free) is that one queue: concurrency 1,
a history pull ahead of background refreshes, then the owner's rank (the grid
ranks the focused tile first, then reading order). A destroyed tile's waiting
loads are dropped unrun, and destroy() aborts the running fetch so the queue
moves on.
Also, for Pane B as well: the load now has a deadline covering the body
(CodemanFetchDeadline), so a capture that never answers cannot hold the
single-flight flag (or the queue) forever; a refresh clears the screen at its
turn rather than when it is asked for; and a close while a load only waits in
the queue writes the disconnected marker at once.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A Codeman restart (every deploy) or a network blip used to leave Pane B
dead, with a marker asking the user to close and reopen the split.
TerminalTile now reconnects:
- A transient close reconnects on the primary pane's backoff ladder
(CodemanWsReconnect) plus jitter; the attempt count resets only on a
successful open. The redelivery sweep's forced close (1005) counts as
transient.
- On reopen the closed state is cleared before the buffer refresh that
closes the output gap, so no stale marker lands under a healthy pane.
- 4003/4004/4009/4010 stop the pane for good and report once through a
new onExit(code) callback; the marker says why.
- Sockets are replaced race-free: the old one is detached before a new
one opens, and every handler ignores events from a socket that is no
longer current. destroy() cancels a pending reconnect.
- reconnectNow() lets an owner skip the backoff.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Pure move and rename, no behavior change. The split pane's second
terminal (SplitTerminalPane) moves out of terminal-split.js into its own
terminal-tile.js (load order 7.4) as TerminalTile, so the tile grid can
reuse it. terminal-split.js keeps the split orchestration (picker,
divider, auto-collapse) and constructs a TerminalTile for Pane B.
Tests follow the class: split-pane-terminal-unit becomes
terminal-tile-unit, and the Shift+Enter guard and the two browser suites
read terminal-tile.js / window.TerminalTile. The browser suites match
master (one pre-existing environmental failure in both).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Codeman maintainer