perf(web): grouped low-impact findings — cold-start manifest fetch, bundle budget, editor drafts, mounted panes [low]

[thedancingdeveloper]

Impact: low (grouped) · Confidence: confirmed

Grouped low-impact PWA performance findings from the holistic review.

1. Production cold start blocks on a serialized /demo-manifest.json fetch

web/src/index.tsx:8 — await initializeRuntimeTransport() (runtimeTransport.ts:87-93, cache: "no-store") completes before Promise.all([import("./App"), …]) even starts. In a non-demo deployment this adds one guaranteed round trip (a 404) to first paint, in series with the App/Board chunk fetches. Fix: start the manifest fetch and import("./App") in parallel; install the transport before render.

2. scripts/check_bundle.py's "first screen" budget under-reports the real first screen

It reports 47.3KB JS, but index.tsx:19-27 unconditionally await import("./App") — the 145KB App chunk statically includes the eager Board, CommandPalette, Sessions, and FileTree (deliberate, per the comment at App.tsx:125-135) — so real pre-paint JS is ~200KB. The budget passes honestly today, but the dynamic-import indirection means the gate cannot see growth in the chunks that actually gate first paint. Fix: have the budget count App + its static imports as first-screen.

3. Editor captures a full-document draft on every keystroke

web/src/Editor.tsx:194-208 — onDidChangeModelContent calls editor.getValue() (O(document) string build), compares against savedContent (O(document)), and saveViewState() per keystroke. Fine for source files; janks on multi-MB files. Fix: debounce ~300ms; use model.getValueLength()/version-id as a cheap pre-filter before materializing the string.

4. All terminal panes stay mounted and attached forever on desktop

web/src/tabLifecycle.ts:8 (terminal → retention "always") + App.tsx:1809 (display: none when inactive): each hidden pane keeps a live WebSocket receiving and parsing PTY output into its xterm buffer (scrollback 5000). Deliberate and documented (Terminal.tsx:106-112); with many busy sessions the background parse + WS bandwidth adds up. If it shows in profiles: detach hidden panes after N minutes idle (close WS, keep the cache) and rely on the existing snapshot-resume path.

5. filteredSessions parses dates inside a sort comparator

web/src/History.tsx:481-490 (WIP code) — Date.parse in the comparator is O(n log n) parses; bounded by the 200/page load, so a micro-nit. Precompute the timestamp once per row when convenient.

(Verified well-done, for contrast: xterm.js with capped scrollback and binary WS frames; one shared SSE stream with backoff; Fenwick-tree MeasuredWindow virtualization on Board/Backlog; process-wide taxonomy cache; route-lazy heavy surfaces with a budget-gated 47KB first screen; no-cache service worker; historyReplay.ts linear and bounded.)

---

From a holistic security & performance review, 2026-09-02, branch feat/product-version @ 3933a09; line numbers refer to that tree.

[thedancingdeveloper]

Status: still open. None of the grouped low-impact findings have landed — on dev, index.tsx still serializes the runtime-transport fetch ahead of the App-chunk imports (item 1), Editor.tsx still materializes a full-document string per keystroke with no debounce (item 3), and History.tsx still parses dates inside the sort comparator (item 5). These are low-impact by design; keeping this as a grouped cleanup. (Sibling perf group #536 is now resolved.)