| name | web-performance |
| description | Use when the user reports jank, dropped frames, janky scrolling, slow click or typing response, poor INP, slow LCP, layout shifts (CLS), unresponsive pages, or asks to debug rendering, animation, or interaction performance in a browser app. |
Web Performance
Overview
Browser performance debugging via PerformanceObserver, with LoAF (long-animation-frame) as the primary signal. LoAF is the only entry type that, in one record, attributes a slow frame to a specific sourceURL + sourceFunctionName + sourceCharPosition + invokerType, with per-script forcedStyleAndLayoutDuration (sync reflow), pauseDuration (sync XHR / alert), and blockingDuration. Code inspection and performance.now() cannot reach this. Start with LoAFs, conclude from LoAFs.
When to use
Symptoms:
- Jank, dropped frames, janky scroll/swipe, complaints about frame rate
- Slow click / keypress / touch response, "unresponsive" complaints, poor INP
- Slow LCP, layout shifts (CLS)
- Animation stutter, transition jank, expensive renders during interaction
Do NOT use for:
- Backend / non-browser perf — use raw NDJSON file appends from your server runtime
- Memory leaks, bundle-size regressions — heap snapshots / bundle analyzers
- Logic bugs unrelated to timing — use raw fetch instrumentation
Core pattern
Before — manual performance.now() (wrong):
const t0 = performance.now();
drawSeries(data);
console.log("drawSeries took", performance.now() - t0);
Tells you a number. Doesn't tell you the function caused a long frame, what scheduled it, or whether it forced sync layout. Requires you to already suspect drawSeries.
After — LoAF observer (right):
new PerformanceObserver((list) => {
for (const loaf of list.getEntries()) send(loaf);
}).observe({ type: "long-animation-frame", buffered: true });
Reports every frame > 50ms across the whole page, with scripts[].sourceURL + sourceFunctionName + sourceCharPosition + invokerType + forcedStyleAndLayoutDuration for each script that ran in the frame. You don't need to know where the bug is in advance.
Workflow
- Generate 3-5 hypotheses about what's slow and where.
- Start the logging server (Implementation → STEP 0).
- Inject the LoAF observer as the first script in
<head> or top of SPA entry.
- Reproduce — automate via Playwright/Puppeteer if possible; otherwise give numbered steps and ask the user to confirm in their UI (do NOT ask them to type "done").
- Clear the log file before each run via the deletion tool (NOT
rm).
- Analyze LoAFs first; consult secondary signals only if LoAF is silent. Mark hypotheses CONFIRMED / REJECTED / INCONCLUSIVE with cited entries.
- Fix only with 100% confidence. Keep instrumentation in place; tag post-fix runs with
runId="post-fix".
- Verify by re-running and comparing before/after LoAFs with cited lines. If failed, revert rejected-hypothesis code (keep instrumentation), generate new hypotheses, iterate.
- Cleanup — remove the
#region debug log block only after verified success + explicit user confirmation.
Implementation
STEP 0: Start the logging server (background-only)
npx debug-agent@latest --json --daemon
--daemon forks the server into a detached process and exits immediately (your shell unblocks instantly — no need for &/nohup). --json makes the parent emit one machine-readable JSON line (without it you get a colored spinner). It prints one JSON line on startup:
{
"sessionId": "a1b2c3",
"endpoint": "http://127.0.0.1:54321/ingest/a1b2c3",
"logPath": "/tmp/debug-agent/debug-a1b2c3.log"
}
Capture endpoint (POST traces here), logPath (NDJSON written here on macOS at /var/folders/.../T/debug-agent/debug-<sessionId>.log), sessionId (in every payload). The log file is auto-created on first write — do NOT pre-create. Server is idempotent: re-running --json --daemon returns the same sessionId/port/logPath of the existing server, so it's safe to call at the start of every session. If startup fails, STOP and inform the user.
To clear the log file via HTTP without deleting/recreating it: curl -X DELETE <endpoint> returns {"ok":true,"cleared":true}.
STEP 1: Inject the LoAF observer
Replace __ENDPOINT__ and __SESSION_ID__. Paste once as the very first script in <head> (above framework bootstrap), or top of the SPA entry module (main.tsx, index.ts, app.tsx). One IIFE in one #region debug log block.
<script>
(() => {
const ENDPOINT = "__ENDPOINT__";
const SESSION_ID = "__SESSION_ID__";
const send = (kind, payload) =>
fetch(ENDPOINT, {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({
sessionId: SESSION_ID,
location: "PerformanceObserver:" + kind,
message: kind,
data: payload,
timestamp: Date.now(),
}),
keepalive: true,
}).catch(() => {});
const observe = (type, mapEntry) => {
try {
new PerformanceObserver((list) => {
for (const entry of list.getEntries()) send(type, mapEntry(entry));
}).observe({ type, buffered: true });
} catch {}
};
observe("long-animation-frame", (loaf) => ({
startTime: loaf.startTime,
duration: loaf.duration,
renderStart: loaf.renderStart,
styleAndLayoutStart: loaf.styleAndLayoutStart,
blockingDuration: loaf.blockingDuration,
firstUIEventTimestamp: loaf.firstUIEventTimestamp,
scripts: (loaf.scripts || []).map((scriptTiming) => ({
invoker: scriptTiming.invoker,
invokerType: scriptTiming.invokerType,
sourceURL: scriptTiming.sourceURL,
sourceFunctionName: scriptTiming.sourceFunctionName,
sourceCharPosition: scriptTiming.sourceCharPosition,
executionStart: scriptTiming.executionStart,
duration: scriptTiming.duration,
forcedStyleAndLayoutDuration: scriptTiming.forcedStyleAndLayoutDuration,
pauseDuration: scriptTiming.pauseDuration,
})),
}));
})();
</script>
buffered: true and keepalive: true are required (early entries + survival across navigation). The try/catch lets the observer no-op on browsers without LoAF support.
FORBIDDEN: logging DOM text, form values, cookies, tokens, PII-bearing URLs — emit only tag names, URLs, durations, rects.
Secondary observers (add inside the same IIFE only when LoAF can't see it)
| Add this observer | When |
|---|
event | INP investigation (input delay / handler / presentation breakdown) |
layout-shift | CLS investigation |
largest-contentful-paint | Slow LCP investigation |
longtask | Safari/Firefox fallback (no script attribution) |
paint | First Paint / FCP investigation |
observe("event", (e) => ({
name: e.name,
startTime: e.startTime,
duration: e.duration,
processingStart: e.processingStart,
processingEnd: e.processingEnd,
interactionId: e.interactionId,
targetTag: e.target?.tagName ?? null,
}));
observe("layout-shift", (s) => ({
startTime: s.startTime,
value: s.value,
hadRecentInput: s.hadRecentInput,
sources: (s.sources || []).map((x) => ({
nodeName: x.node?.nodeName ?? null,
previousRect: x.previousRect,
currentRect: x.currentRect,
})),
}));
observe("largest-contentful-paint", (p) => ({
startTime: p.startTime,
renderTime: p.renderTime,
loadTime: p.loadTime,
size: p.size,
url: p.url,
elementTag: p.element?.tagName ?? null,
}));
observe("longtask", (t) => ({
startTime: t.startTime,
duration: t.duration,
attribution: (t.attribution || []).map((a) => ({
name: a.name,
containerType: a.containerType,
containerSrc: a.containerSrc,
})),
}));
observe("paint", (p) => ({ name: p.name, startTime: p.startTime }));
STEPs 2-5: Log lifecycle
- Clear
logPath before each run. Two equivalent options: file-deletion tool (NOT rm) on the path, or curl -X DELETE <endpoint>. Only your session's file — never another session's. Clearing ≠ removing instrumentation.
- Read
logPath after the user confirms reproduction. Each line is { sessionId, id, timestamp, location, message, data }. Empty/missing → reproduction failed; clear and retry.
- Keep all instrumentation through fixes. Tag verification runs with
runId="post-fix". Removing logs before verification is FORBIDDEN.
- Cleanup after verified success + user confirmation: grep
#region debug log, delete each region (line-inclusive, including the wrapping <script> if injected via HTML), re-grep to confirm zero markers, git diff review.
Quick reference: decoding LoAF entries
Sort entries by duration (or blockingDuration for input-blocking jank), worst first. The script with the largest duration inside the worst LoAF's scripts[] is your culprit. If forcedStyleAndLayoutDuration > 10ms or > 25% of script duration, it's layout thrashing — look for sync offsetHeight / getBoundingClientRect / scrollTop reads interleaved with DOM writes.
| Field | Meaning | Diagnostic use |
|---|
duration | total frame time (ms), only fires > 50ms | severity |
blockingDuration | input-blocking portion | INP impact |
renderStart − startTime | JS work before render | "long script" share |
styleAndLayoutStart − renderStart | rAF / ResizeObserver / IntersectionObserver callbacks | pre-paint callback cost |
(startTime + duration) − styleAndLayoutStart | style + layout + paint | CSS / layout cost |
scripts[].sourceURL + sourceFunctionName + sourceCharPosition | exact code site | culprit |
scripts[].invoker | precise scheduler (e.g. BUTTON#btn-foo.onclick, TimerHandler:setTimeout, IMG#hero.onload) | which DOM node / timer / promise scheduled it |
scripts[].invokerType | category — event-listener / user-callback (covers setTimeout/setInterval/requestAnimationFrame/requestIdleCallback) / resolve-promise / reject-promise / classic-script / module-script | how it ran |
scripts[].forcedStyleAndLayoutDuration | sync reflow inside the script | layout thrashing |
scripts[].pauseDuration | sync XHR / alert time | hard blocks |
invoker vs invokerType: invokerType is a fixed-vocabulary category; invoker is the precise instance. Use invoker to tell setTimeout from requestAnimationFrame (both are user-callback), or to identify which button's click handler ran (e.g. BUTTON#btn-thrash.onclick).
Validated thrash fingerprint: when forcedStyleAndLayoutDuration / scripts[].duration > 0.5, the script is dominated by sync layout — almost always reads of offsetHeight / getBoundingClientRect / scrollTop / getComputedStyle interleaved with style/DOM writes. Example seen in practice: duration=252ms, forcedStyleAndLayoutDuration=247ms (98%) for a 1500-iteration read/write loop.
Cite the specific entry when concluding:
CONFIRMED hypothesis B: LoAF startTime=12483ms, duration=128ms, scripts[2]: sourceURL=https://app.com/static/chart.bundle.js, sourceFunctionName=drawSeries, invokerType=event-listener, duration=84ms, forcedStyleAndLayoutDuration=42ms — chart redraw triggers sync layout in scroll handler.
Secondary signal decoding (only when LoAF is silent):
event — input delay = processingStart − startTime; handler = processingEnd − processingStart; presentation = (startTime + duration) − processingEnd; > 200ms is poor INP. Filter out noise: a single click emits 10+ entries (pointerover, pointerenter, pointerdown, mousedown, pointerup, mouseup, click, mouseover, pointerout, mouseout, …). Most have interactionId: 0 (not real interactions) and share the same duration (they were all queued behind the same long task). Only entries with interactionId !== 0 represent distinct interactions for INP purposes — and within those, name === "click" (or keydown/pointerdown) is the canonical one.
layout-shift — CLS = sum of value where hadRecentInput === false; sources[] pinpoints node + rects.
longtask — > 50ms main-thread tasks; attribution[0].containerSrc often points at a third-party script. Coarser than LoAF; no sourceFunctionName.
Common mistakes
| Mistake | Why it fails | Fix |
|---|
| Snippet pasted after framework bootstrap | Misses early LoAFs and pre-bootstrap LCP | First <script> in <head> or top of entry module before any side-effecting import |
Forgetting buffered: true | Drops entries that fired before observation started | Always { type, buffered: true } |
| Logging server in foreground | Stalls the agent forever | Use & / nohup / block_until_ms: 0 |
Manual fetch logs alongside the observer | Pollutes traces, breaks one-region cleanup | Use only the IIFE; add observers, not separate logs |
Concluding from duration only | Ignores attribution; you'll guess wrong | Cite scripts[].sourceURL + sourceFunctionName + forcedStyleAndLayoutDuration |
Counting every event entry as a real interaction | One click emits 10+ entries (pointer*, mouse*, click); most are not "interactions" | Filter interactionId !== 0; usually the click/keydown/pointerdown row is canonical |
Reading invokerType and stopping there | setTimeout, setInterval, rAF, rIC all share invokerType=user-callback | Use invoker (e.g. TimerHandler:setTimeout) to disambiguate |
"Fix" by wrapping in setTimeout | Defers work; jank just shifts to a later frame | Verify post-fix LoAFs show the script duration is gone, not relocated |
Reading bundled sourceURL literally | Minified path is meaningless | Resolve via sourcemap; sourceCharPosition disambiguates collisions |
| Testing on Safari/Firefox first | LoAF is Chromium-only; you'll think the snippet is broken | Reproduce on Chrome/Edge ≥ 123 first |
| Removing instrumentation before verification | Can't prove fix worked; no traces if it regressed | Keep #region debug log until verified + user confirms |
| Deleting another session's log file | Corrupts an unrelated debug session | Only touch the logPath from YOUR STEP 0 |
Red flags — STOP and re-instrument
- "I'll just
setTimeout(..., 0) and call it fixed" → prove the script work is gone, not deferred
- "The
duration looks fine, must be CSS" → check forcedStyleAndLayoutDuration and styleAndLayoutStart first
- "No LoAF entries appeared" → wrong browser, snippet too late, or no frame
> 50ms
- "All hypotheses rejected" → never fix without runtime evidence; generate new hypotheses
- "I'll claim success based on the manual fix" → no, cite before/after LoAF lines