React Scan
Use targeted captures to optimize measured Studio interactions. Do not infer a performance problem from source alone when the workflow can be reproduced.
Record a capture
Start from the repository root:
bun run react-scan:capture
For the Brand Studio, run:
bun run react-scan:capture:brand
Wait for Studio to finish building, open the printed Studio URL with the browser, and let the page settle. The React Scan WebMCP tools are only registered in this development capture mode.
Call start_react_scan_recording, then use browser-use to perform one named
interaction three times. Call stop_react_scan_recording immediately
afterwards. Keep the recording short so startup, HMR, and unrelated interactions
do not dominate the data. Call get_react_scan_recording to return the raw
events and an agent-oriented summary.
After retrieving the recording, stop Studio with Ctrl+C. Recordings only live in the WebMCP result; React Scan does not send them to a separate HTTP collector.
React Scan replaces the React DevTools profiling-hook channel while active. Do not record with the React DevTools Timeline Profiler at the same time.
Analyze the evidence
Read the summary returned by get_react_scan_recording first. Confirm the
capture contains commit events and inspect profilingHooksStatuses; do not
interpret a missing profiling channel as an idle application.
Prioritize components with high total self duration, repeated expensive renders, or a clear user-visible slow commit. Inclusive duration contains child work and can double-count a subtree, so use it to locate an expensive boundary, then use self duration and the raw tree to identify the actual work.
Use render causes as evidence, not automatic prescriptions:
- High
parentRenderCountwith no relevant prop, state, context, or hook change suggests a stable memoization boundary. - Repeated changed props suggest checking referential stability at the source location before adding memoization.
- Hook, state, or context changes require following the owning update; wrapping
the component in
memo()will not block them. - Source locations identify JSX call sites. Search for the component definition before editing.
For timeline-position findings, also read ../studio-perf/SKILL.md before
changing subscriptions or memoization boundaries.
Verify an optimization
Preserve the baseline capture. After changing code, record the same interaction the same number of times with the same viewport and compare a new capture. Require an improvement in the relevant commit/component metrics without a regression in behavior. Do not optimize every re-render: cheap necessary renders are often preferable to added memoization complexity.
Report the measured before/after change and the checks performed to verify behavior.