perf(state): batch tmux metadata in collect_state — ~13× faster /api/state

[tdody]

Summary

Fixes the real cause of the "app gets laggy over time" reports: /api/state was slow, and the dashboard polls it every 500 ms while a terminal modal is open.

collect_state() iterated libtmux objects — srv.sessions, s.windows, w.active_pane, pane.pane_current_command, pane.pane_current_path — and libtmux forks a separate tmux subprocess for each lazy property access (~150 ms per window). At the modal-open cadence that spawned dozens of tmux subprocesses per second, flooding the tmux server so the user's keystrokes — each a tmux send-keys — queued behind the storm. Result: laggy typing + slow modal opens that scaled with pane count and got dramatically worse under system memory pressure (subprocess forks stall on paging → multi-second /api/state spikes).

Change

Replace the per-window/per-pane libtmux fan-out with two batched tmux commands:

• tmux list-panes -a -f '#{pane_active}' -F <fmt> — all active-pane metadata (session + window + pane) in one subprocess.
• tmux list-clients -F <fmt> — all clients, grouped onto their session.

Parsed by pure, unit-tested helpers (_parse_pane_rows, _parse_clients_grouped); sessions/windows are derived from the rows. capture-pane stays per-pane (already cached, THI-181 — now keyed by session:window target). The StateResponse schema is unchanged, so the frontend is untouched.

Measured (live, 6 panes)

	median	max
before	380 ms	810 ms (3.4 s under memory pressure)
after	29 ms	84 ms

~13× faster, and the multi-second spikes are gone. The win grows with pane count (old cost was linear in windows; new cost is ~flat).

Tests (TDD)

• New pure-parser tests (_parse_pane_rows, _parse_clients_grouped): delimited-field parsing, blank/short-row skipping, spaces in names/paths, client grouping. Watched them fail (helpers didn't exist), then pass.
• Rewrote the collect_state tests to drive the new batched srv.cmd data source — covering sb-usage filtering, one-card-per-session dedup, capture-pane TTL cache (hit within TTL / re-capture after), client population/attached, malformed-row skip, and graceful degradation when list-panes errors mid-scan (no 500).
• Full backend suite: 393 passing; ruff (lint + format) + ty clean.

Notes

• Independent of fix(pane-stream): release FIFO read fd on pane WebSocket disconnect #116 (backend fd leak) and fix(terminal): don't close pane WebSocket mid-handshake on teardown #117 (dev WebSocket warning).
• The earlier "memory leak" reports were a red herring — the tab's JS heap/DOM are flat (verified over 100 min). The lag was this backend subprocess fan-out, amplified by system swap pressure.

🤖 Generated with Claude Code