lite-render-perf

The recurring render-performance failure in this app is accidental over-re-rendering: one interaction re-renders every row instead of the few affected ones. Repositories can have hundreds of branch rows, so this turns single-digit-millisecond interactions into triple-digit ones. The compiler makes it easy to cause by accident because it silently stops memoizing exactly where you assume it won't.

The three mechanisms behind past regressions

1. Whether the compiler memoizes a derivation is not predictable from the source

The compiler only caches what it can prove safe to cache, and it declines more often than you would guess. Two derivations in this app are both module-level functions that call imported helpers, and they compile differently: buildOutlineNavigationIndex in WorkspacePage.tsx gets a memo block keyed on its inputs, while getDiffView in Details.tsx got none and re-ran on every render of a component that re-renders whenever the file selection changes.

Unmemoized inputs are contagious. getDiffView was called with values built by Match pipes that the compiler had not memoized either; feeding it stable values instead was enough to make the memo block appear. So the churn spreads outward from whichever value first fails to get a scope, which is why one unmemoized derivation can quietly cost far more than itself.

Either way the effect is the same: the derivation re-runs and returns fresh identities even when its data never changed, and everything keyed on those identities — context values, props, effect deps — churns with it.

What to do instead:

• Derive query data inside the query's select. react-query caches the result on (data, select function identity), and the compiler DOES memoize the closure on its captures — so the closure's identity becomes exactly the dependency key you want. Extra inputs are simply captured by the closure. Precedent: useBranchesOutline.ts. useQueries / useSuspenseQueries give you the same lever in combine — precedent: lineStats in OutlineTree/UncommittedChangesRow.tsx, which folds the results down inside combine instead of mapping them out and deriving in render the way Details.tsx does. Check what the closure captures: capture a value that is itself rebuilt each render and you have only moved the churn one level up.
• For redux-derived collections, use createSelector (see projects/project.ts), never a plain function returning a fresh array or object.
• Never conclude from the source that something is or is not memoized — read the compiler's output. A compiled function imports react/compiler-runtime and allocates slots with const $ = _c(n); a memoized value is assigned inside an if ($[i] !== dep) block, and its dependencies are exactly the values that block tests. A derivation sitting outside every such block runs on every render. Two ways to look: fetch the vite-transformed module from the running app (fetch('/src/path/to/file.ts').then(r => r.text()) over CDP — see below), or run the plugin offline on any file, which needs no app:

// Resolve both from node_modules/.pnpm — pnpm does not hoist them. There is
// no @babel/preset-typescript here, so strip types via parserOpts instead.
await babel.transformAsync(source, {
	filename: file,
	parserOpts: { plugins: ["typescript", "jsx"] },
	plugins: [["babel-plugin-react-compiler", { target: "19" }]],
	configFile: false,
	babelrc: false,
});

2. Context re-renders every consumer; it has no selector

use(SomeContext) re-renders the component whenever the context value's identity changes — there is no way to subscribe to a slice, and neither the compiler nor React.memo can bail a context read out. Note this is a different failure from mechanism 1: when the value's content legitimately changed, fixing identities cannot help — the only cure is consumers subscribing to something narrower than the whole value.

What to do instead:

• Context is for values that change rarely relative to their consumers.
• Per-row state must be a narrow useAppSelector returning a primitive (usually a boolean). Redux subscriptions only re-render when the selected value changes, so a selection change re-renders two rows, not all of them:

const isSelected = useAppSelector((state) => {
	const stored = projectSlice.selectors.selectPrimaryOutlineSelection(state, projectId);
	return stored !== null && operandEquals(stored, operand);
});

For the current shape, see useIsSelected in BranchesList.tsx and the selectIsSelectedOutline subscription in OutlineTree/ItemRow.tsx.

• If rows need a value that requires resolution against a big object, resolve it once at the top and store the resolved form, rather than letting every row resolve independently.

3. Identity discipline in the store and in query results

• Reducers must early-return when the update is a no-op, otherwise they produce a new state identity and wake every subscriber.
• Reading store data inside an event callback needs no subscription at all — take const store = useAppStore() at the top and call store.getState() in the callback, instead of subscribing with useAppSelector to a value only the callback reads.
• Destructure useMutation results (const { isPending, mutate } = ...) — the top-level result object is a new identity every render (@tanstack/query no-unstable-deps).
• Hook call order can change what the compiler memoizes; moving plain useQuery calls above a useQueries call has fixed memoization here before. If the compiled output shows a value outside every memo slot for no visible reason, try reordering the hooks around it.

Driving the dev app over CDP

pnpm dev in apps/lite starts vite (port 5173) and Electron with --remote-debugging-port=9222. The first run compiles a helper binary with cargo and takes minutes; Port 5173 is already in use means an instance is already running — reuse it, or lsof -ti tcp:5173 | xargs kill. React and Redux DevTools are auto-installed in dev builds.

curl http://127.0.0.1:9222/json lists debug targets. Pick the entry with "type": "page" whose URL is on 5173 — the others are extension service workers. Its webSocketDebuggerUrl speaks the Chrome DevTools Protocol; a few lines of Node (global WebSocket) are enough to drive it:

// Usage: node cdp.mjs eval "<expr>"   or:   node cdp.mjs shot out.png
const page = (await (await fetch("http://127.0.0.1:9222/json")).json()).find(
	(t) => t.type === "page" && t.url.includes("5173"),
);
const ws = new WebSocket(page.webSocketDebuggerUrl);
let id = 0;
const pending = new Map();
const send = (method, params) =>
	new Promise((res) => {
		pending.set(++id, res);
		ws.send(JSON.stringify({ id, method, params }));
	});
ws.onmessage = (e) => {
	const m = JSON.parse(e.data);
	pending.get(m.id)?.(m.result);
};
await new Promise((res) => (ws.onopen = res));
const [mode, arg] = process.argv.slice(2);
if (mode === "eval")
	console.log(
		JSON.stringify(
			await send("Runtime.evaluate", {
				expression: arg,
				awaitPromise: true,
				returnByValue: true,
			}),
		),
	);
else if (mode === "shot") {
	const { data } = await send("Page.captureScreenshot", { format: "png" });
	(await import("node:fs")).writeFileSync(arg, Buffer.from(data, "base64"));
}
process.exit(0);

Useful from Runtime.evaluate:

• The whole preload API is callable: window.lite.branchList(projectId) etc. — the fastest way to inspect real API payloads and sizes.
• Interact by dispatching events on DOM nodes found via aria-labels; screenshot to verify visually.

Gotchas:

• Repeated evaluates share one top-level scope: wrap every expression in a { ... } block or const x collides with the previous call ("Identifier 'x' has already been declared").
• Synthetic MouseEvents do not move real focus; behavior wired to onFocus (selection scopes) needs an explicit .focus() or real input.
• Renderer edits arrive via HMR, but edits to store/redux modules trigger a full page reload that wipes window.* state — re-arm any probes after. Electron main-process edits are only compiled at startup: restart pnpm dev.

Measuring: prove it, in both directions

React DevTools being installed means its hook is available in-page:

1. Arm a commit logger and interact:

const hook = window.__REACT_DEVTOOLS_GLOBAL_HOOK__;
window.__log = [];
const orig = hook.onCommitFiberRoot.bind(hook);
hook.onCommitFiberRoot = (id, root, ...rest) => {
	window.__log.push({ dur: root.current.actualDuration });
	return orig(id, root, ...rest);
};

2. Judge by actualDuration per commit: an interaction should cost single-digit milliseconds and re-render O(affected rows). A commit whose duration scales with total row count is the signature of a blast.
3. To find WHY, read the compiled module (mechanism 1 above). To watch a specific context value's identity across commits, extend the logger to walk root.current for the provider fiber (its type is the context object, so match on type.displayName) and compare memoizedProps.value against the previous commit's. Do NOT add window.* probe writes inside components or hooks to trace identities — that mutation makes the compiler bail out of the whole function and destroys the very memoization you are measuring.
4. Restore onCommitFiberRoot when done; measure before and after the fix.

Review checklist

• Any non-trivial derivation in render: is its result's identity load-bearing (context value, prop to many children, effect dep)? Move it into select/combine/createSelector, or read the compiled output to prove it harmless — do not assume the compiler handled it.
• Any new context: how often does its value change, and how many consumers?
• Any useAppSelector: does it return a primitive or a stored reference, not a fresh object?
• Any new reducer: does it no-op cleanly when the value is unchanged?