/argent-native-profiler

name: argent-native-profiler
description: Native profiling for CPU hotspots, UI hangs, memory issues. iOS via xctrace; Android via Perfetto. Use when diagnosing native-level performance issues.

1. Tools and platforms

• `native-profiler-start` → `native-profiler-stop` → `native-profiler-analyze`: record, export, and report.
• `profiler-stack-query`: drill into the analyzed data (hang stacks, function callers, thread breakdown, leaks).
• `profiler-load`: reload an earlier session from disk.

**iOS** (Instruments via `xctrace`): simulator only; a physical iPhone is not supported. Needs full Xcode (`xcrun xctrace version`). Reports CPU hotspots, UI hangs and memory leaks.

**Android** (Perfetto): emulator or physical device. Reports CPU hotspots and UI hangs (with jank reasons and the main thread's blocking function), plus an RSS-growth signal for memory pressure: treat it as a hint to confirm, not a confirmed leak. The app must be debuggable or declare `<profileable android:shell="true"/>` in its manifest, or no `perf_sample` callstacks are captured.

`argent-react-native-profiler` starts native profiling alongside React profiling by default; this skill's workflow and investigation patterns apply in both cases.

2. Workflow

**Complete all steps in order; do not break mid-flow.**

1. **Start.** Make sure the app is running (`launch-app` if not); `native-profiler-start` detects it. Pass `app_process` only if it reports several running user apps (e.g. dev and prod flavors) or picks the wrong one. Then let the user interact, or drive the app yourself (`argent-device-interact` skill). While a profiler records, interaction tools return no automatic screenshot or element tree, so the capture does not load the app. Call `screenshot` or `describe` when you need to see the screen; interaction tools still wait for the screen to settle before they return, so add no wait of your own; on an Android emulator each screenshot during a profile takes one frame in about 100 ms and does not wake the screen capture; on a physical Android device it takes about 150–200 ms and briefly wakes the device screen capture, so take them sparingly there. 2. **Stop.** `native-profiler-stop`. On iOS, check `exportDiagnostics` for export warnings. 3. **Analyze.** `native-profiler-analyze`. Findings are 🔴 RED (act on it) or 🟡 YELLOW (worth investigating, may be acceptable). 4. **Present and ask.** Summarize the key findings, then ask the user whether to investigate further, implement fixes, or stop. Do not move on silently. 5. **Drill down** with `profiler-stack-query` (patterns below). 6. **Fix and re-profile** the same scenario; compare with the earlier session via `profiler-load` (`list`, then `load_native` with its `session_id` and the `device_id`). Report honestly whether the target metric improved, regressed, or stayed flat; if there is no net benefit or something else regressed, say so and reconsider. For reproducible before/after runs, record the interaction as an e2e test (`argent-create-e2e` skill) before the first run and replay it with `e2e-execute`.

Investigation patterns

• **Hang** → mode `hang_stacks` for the native call chain → `function_callers` for the suspected function → read the native source.
• **CPU hotspot** → `thread_breakdown` for the per-thread split → `function_callers` for the dominant function.
• **Memory leak** (iOS) → `leak_stacks`, filtered by `object_type`, for the responsible frames and libraries. To attribute leaks shown as `<Call stack limit reached>`, re-run `native-profiler-start` with `malloc_stack_logging: true`; it restarts the app and adds overhead, so not for CPU or hang passes.

3. Caveats

• **Simulator and emulator timings reflect the host machine**, not device hardware. Compare runs against each other rather than trusting absolute numbers; a physical Android device gives real-hardware numbers.
• **Profiler overhead:** if `JSLexer`, `JSONEmitter` or Hermes internals dominate the JS thread, that is instrumentation cost, not app work. Discount those entries.
• **Run-to-run variance:** treat only consistent, directional changes (across 2+ runs, or a >15% delta) as signal.
• **Live data:** if the app fetches live API data, different responses change the workload independently of code changes. Note when data-dependent screens vary.