fix(tiles): cap each tile's live-output backlog and recover dropped output

Browse Source

The server applies no WebSocket backpressure (16 KB / 8 ms batches, no
bufferedAmount check), and a tile wrote every live frame straight into
xterm. A flood a tile could not parse as fast (a shell tile running cat on
a huge log) piled up in xterm's own write queue without bound, on a main
thread up to six tiles share, until xterm's WriteBuffer threw past 50M
code units; onmessage's empty catch then dropped every frame silently and
nothing recaptured the screen. The primary pane caps its queues and drops
then recaptures (_onSessionTerminal, _scheduleDroppedOutputRecovery).

Each tile now writes live output through _writeLive:
- unparsed code units are counted, each write's callback counting its own
  back down; frames held behind a replay (_liveQueue) count too;
- the budget is TerminalTile.LIVE_BACKLOG_BUDGET, 4 MiB, deliberately not
  the primary pane's 128 KB: that caps its own rAF-paced queues, while
  xterm itself paces a tile, and a tight cap would trip on ordinary bursts
  and blank-and-reload the tile over and over;
- past it a frame is dropped, the tile stops writing onto the hole, and one
  refresh is scheduled, debounced and bounded by the primary pane's own
  rule (CodemanDroppedOutput: 2 s, DROP_RECOVERY_MAX_ATTEMPTS, never retried
  after a deadline abort). It is an ordinary refresh, so single-flight,
  bounded by lines=/tail= and paced by the grid's TileLoadQueue. The flag
  clears once a capture taken after the last dropped frame has replayed;
- a write that throws is the same drop, never a "malformed frame";
- past the bound the flag is released, so a tile is never left frozen;
- a reconnect starts the accounting over (an epoch makes callbacks from
  before it count nothing) and drops a pending recovery, since its own
  refresh replaces the screen; destroy() cancels it.
The live-queue flush after a pull or a refresh goes through the same path,
so a throwing write there cannot skip the load's marker and trailing
refresh either.

Tests (input harness, real constants and fake timers): the default budget
lets a 1 MiB unparsed burst through, parsed bytes stop counting, a trip
stops writing and ONE debounced refresh recaptures, a write throw takes the
same recovery, a hole in the held queue is recovered by another refresh,
bounded retries then release, no retry after a deadline, a reconnect resets
the count, and destroy cancels. The fake xterm can now hold and release
parses and throw on a write. Live writes now carry a callback, so the unit
tests match them on the data argument (a `.not.toHaveBeenCalledWith(data)`
would otherwise pass for nothing).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

...

This commit is contained in:

Codeman maintainer

2026-10-09 09:26:51 +02:00

parent cdb3c34ae0

commit 1def7de146

5 changed files with 444 additions and 19 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download Patch File Download Diff File Expand all files Collapse all files

docs/architecture-invariants.md

+1 -1

File diff suppressed because one or more lines are too long