React DevTools Profiler Export Analysis

Parse React DevTools Profiler exports (version 4/5 JSON, top-level dataForRoots) with Python scripts in scripts/ — no DevTools UI needed. Works for React Native and web/Electron profiles alike.

Distinguish from a Chrome Performance trace: profiler export has {"version": N, "dataForRoots": [...]}; a trace has traceEvents (use the electron-profile skill for those).

Workflow

1. Overview — commit totals, top commits with updaters, component ranking by aggregate self duration, timeline buckets:

   python3 scripts/profile-overview.py <profile.json> [--top 25] [--bucket-ms 500]
   

2. Per-commit detail — full updater list, fiber counts, top fibers with ancestor paths, changeDescriptions if recorded:

   python3 scripts/profile-commits.py <profile.json> [--commit N] [--top 10] [--root 0]
   

Run from the skill directory (or use absolute script paths); both scripts need scripts/rdtlib.py alongside. The overview's top-commit list prints #N — that is the index to pass as --commit N (commits are chronological).

Interpreting results

Signal	Meaning
updaters on a commit	Fibers that scheduled the update — setState / store subscription (useSyncExternalStore) firing. The render entry point, not necessarily where time went.
Same broad updater set repeated across several commits	Shared-store fan-out: many components' selectors woke for one logical change. Find the common denominator subscription (here it was conversation meta) and fix identity churn there.
Unnamed fibers #1234	Fiber absent from the end-of-profiling snapshot = unmounted before profiling stopped (virtualized-list churn, pending rows replaced on server ack). Cross-reference the same id in updaters, which carries displayName.
prio=Immediate	Sync flush — store setState outside a React event (zustand, RPC callbacks). Each one blocks the frame; many small Immediate commits = stacked sync flushes worth batching.
High x<count> in component ranking	Component rendered in many commits — subscription too broad or unstable selector output (fresh object/closure identity per run defeats useShallow).
Large single-fiber self time in a big-fiber-count commit	Usually a subtree mount (new list row). Mount cost is mostly inherent; look at whether it mounts twice (key churn).
changeDescriptions: null	Profile recorded without "Record why each component rendered". Re-record with it enabled to get props/state/hooks/context attribution per fiber.

Scope caveat: durations are React render/effect time only. Store selector runs, native layout, bridge, and app JS outside render are invisible here — a small total does NOT mean no perf problem; it means the problem isn't React commits.

Fan-out diagnosis recipe: overview → spot repeated updater sets → per-commit to list exact updater components → read their store selectors → find the shared subscription whose selected values change identity without changing content → preserve identity at the producer (deep-equal reuse on store write), not with deeper memo at consumers.

Gotchas

• Timestamps are ms relative to profiling start; scripts print t= relative to first commit.
• fiberActualDurations = inclusive per fiber; fiberSelfDurations = self. Ranking uses self.
• Multiple roots possible (dataForRoots); overview iterates all, per-commit takes --root.
• Updater lists carry displayNames even for fibers missing from the snapshot — best way to name unmounted fibers.