When Your Terminals Stop Lying
The terminal display had been corrupting since we first wired tmux output through WebSocket to xterm.js. Claude Code’s Ink TUI would render for a moment, then text would scatter — fragments of the status bar appearing mid-screen, cursor positions drifting. The kind of bug where every screenshot looks different.
The format mismatch
I’d been using capture-pane to replay terminal history when a WebSocket connected,
then pipe-pane for live output. These two produce fundamentally different data:
capture-pane gives you a screen snapshot (what the terminal looks like right now),
while pipe-pane gives you the raw output stream (what the application is sending).
Mixing them corrupts any TUI that uses cursor positioning — which is exactly what
Claude Code’s Ink renderer does.
The fix was to drop capture-pane entirely. We start pipe-pane first, then force a full screen redraw by cycling the tmux pane size — shrink one column, wait 50ms, restore. The resize triggers SIGWINCH, which makes Ink re-render its entire screen cleanly through the pipe. No history replay, no format mismatch, no corruption.
We also found that the FIFO reader was doing new String(buf, 0, n) on raw byte
chunks — splitting multi-byte UTF-8 characters at 4KB boundaries. Extracted that
into a FifoRelay class with InputStreamReader for proper charset handling.
Memory management
The second piece was a memory management panel. When you’re running a dozen Claude agents across slots and repos, each spawning MCP servers and language servers, memory adds up fast. I wanted visibility and control in one place.
The panel has two tables: active terminals at the top (with pause, resume, terminate), and available repos/slots below (with bulk start). Click a row to see the process tree — claude at 448MB, playwright-mcp at 131MB, intellij-mcp at 37MB. That breakdown is what lets you understand where the memory actually goes.
We used pages-data-table for the tables instead of raw HTML — gets selection,
sorting, and efficient DOM updates for free. Had to work around a couple of
pages quirks: paginated mode fills the parent height regardless of content (filed
as casehub-pages#288), and the column settings pill feature silently replaces
custom renderers.
Process lifecycle
The MCP server orphan problem turned out to be well-documented in Claude Code’s
issue tracker. When Claude is killed, StdioClientTransport.close() only kills the
direct child — wrapper commands like npx mean the actual server is a grandchild
that becomes an orphan. Trellis’s treeKill() already handles this correctly
(walks the process tree, kills children first), so pause kills everything. Claude
respawns fresh MCP servers on resume anyway — there’s no benefit to keeping orphans alive.
We also found a race condition in pauseAgent(): the poll cycle could fire between
treeKill() and agents.put(PAUSED), seeing shell + RUNNING and removing the agent
from the map. Moving the state update before the kill eliminates the window.
What shipped
Thirty-two commits across 24 files. Terminal rendering that doesn’t corrupt. A memory panel that shows what’s eating RAM and lets you do something about it. And two upstream issues filed on casehub-pages for things the platform should handle better.