Argent Native Profiler

1. Tools

• native-profiler-start — start profiling on a booted device. iOS: xctrace recording for CPU, hangs, and leaks.
• native-profiler-stop — stop the profiler and export trace data to timestamped XML files.
• native-profiler-analyze — parse exported trace data and return a structured bottleneck payload.
• profiler-stack-query — drill into parsed data: hang stacks, function callers, thread breakdown, leak details.
• profiler-load — list and reload previous trace sessions from disk for re-investigation.

---

2. Platform Support

• iOS: Backend: Xcode Instruments via xctrace on a booted simulator or connected device. Requires Xcode command-line tools on PATH. Surfaces CPU hotspots, UI hangs, and memory leaks (instruments Leaks table).
• Android: Backend: Perfetto via adb shell perfetto + an in-process WASM trace-processor engine. Surfaces CPU hotspots and UI hangs, with per-hang jank reason codes, a main-thread state breakdown with blocked_function attribution, and a GC overlap annotation. Also reports an RSS-growth signal for memory pressure; treat it as a hint to confirm manually, not a confirmed leak. The target app must be debuggable or include <profileable android:shell="true"/> in its manifest for perf_sample callstacks to be captured.

---

3. Investigation Patterns

After native-profiler-analyze surfaces findings, use profiler-stack-query to drill into root causes:

• Hang detected → profiler-stack-query mode=hang_stacks for full native call chains → mode=function_callers for the suspected function → read native source.
• CPU hotspot → profiler-stack-query mode=thread_breakdown for per-thread distribution → mode=function_callers for the dominant function.
• Memory leak → profiler-stack-query mode=leak_stacks filtered by object_type for responsible frames and libraries.

After presenting findings, ask the user whether to investigate further, implement fixes, or stop. After applying fixes, always re-profile the same scenario and compare with profiler-load. Report honestly whether the target metric improved, regressed, or stayed flat. If the fix showed no net benefit or introduced regressions elsewhere, say so and reconsider.

Tip: For reproducible before/after comparisons, record the interaction sequence as a flow using the argent-create-flow skill before the first profiling run. Replay with flow-execute on subsequent runs to eliminate interaction variance.

Note: The argent-react-native-profiler instructs to start native profiling automatically alongside React profiling. This skill's workflow and investigation patterns apply in both cases.

---

4. Workflow

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

Step 0: Ensure the target app is running

The native-profiler-start tool auto-detects the running app on the device. You do not need to derive app_process manually — just make sure the app is launched.

1. If the app is already running on the device, skip to Step 1 (do not pass app_process). 2. If the app is not running, use launch-app with the correct bundle ID first. 3. Only pass app_process explicitly if the tool reports multiple running user apps and you need to disambiguate.

Note: If multiple build flavors are installed (dev, staging, prod), the tool will detect whichever one is currently running. If both are running, it will ask you to specify.

Step 1: Start recording

Call native-profiler-start with device_id (iOS UDID or Android serial). The tool auto-detects the running app and saves the trace to /tmp/argent-profiler-cwd/ with a timestamped filename. Let the user interact with the app or drive interaction via simulator tools (see argent-device-interact skill).

Step 2: Stop and export

Call native-profiler-stop with device_id. iOS sends SIGINT to xctrace, waits for trace packaging, and exports CPU, hangs, and leaks data to XML — check exportDiagnostics for any export warnings. Android sends SIGTERM to the on-device perfetto daemon, polls /proc/<pid> until it exits, then adb pulls the .pftrace to the host.

Step 3: Analyze

Call native-profiler-analyze with device_id. Returns a markdown report with bottlenecks categorized as CPU hotspots, UI hangs, or memory leaks, sorted by severity.

Step 4: Present findings and ask about next steps

Present a concise summary of the key findings. Then follow the "After analysis" guideline — ask whether to investigate further with query tools, implement fixes, or stop.

Step 5: Drill-down investigation

Use profiler-stack-query to investigate specific findings. See §3 Investigation Patterns for chaining guidance.

Step 6: Reload previous sessions

To revisit a previous trace:

1. Call profiler-load mode=list to see available sessions. 2. Call profiler-load mode=load_native session_id=<timestamp> device_id=<UDID> to re-parse the XML files. 3. Use profiler-stack-query to investigate the reloaded data.

---

5. Understanding Results

Bottlenecks are categorized by severity:

• RED: CPU functions taking >15% of total time, all UI hangs, and attributed memory leaks (those with a resolved responsible frame). These require immediate attention.
• YELLOW: CPU functions taking 3-15% of total time, and unattributed memory leaks (<Call stack limit reached>, no library — see the memory-leaks caveat below). Worth investigating but may be acceptable.

Each bottleneck type indicates a different class of problem:

• CPU hotspots: Native functions consuming excessive CPU time. Look for tight loops, expensive computations, or redundant work.
• UI hangs: Main thread blocked long enough to cause visible jank or unresponsiveness. Often caused by synchronous I/O, heavy layout passes, or lock contention.
• Memory leaks: Objects allocated but never freed. Common causes include retain cycles, unclosed resources, or forgotten observers. Argent records via xctrace --attach, which has no malloc-stack history, so on the simulator most leaks come back unattributed (<Call stack limit reached>, no library) and are dominated by benign system allocations — these are reported as a low-confidence YELLOW summary, not confirmed RED leaks. For attributed stacks, capture with malloc stack logging enabled at launch.

---

6. Important Caveats

• Simulator vs device: Simulator profiling reflects host Mac performance, not real device hardware. Use device profiling for accurate CPU timings and memory behavior.
• xctrace availability (iOS): Requires Xcode command-line tools installed. Verify with xcrun xctrace version.
• Profiler overhead: xctrace instrumentation adds CPU load. If JSLexer, JSONEmitter, or Hermes runtime internals dominate the JS thread in CPU hotspot results, those reflect profiler overhead — not app work. Discount those entries when evaluating findings.
• Run-to-run variance: Small fluctuations in CPU percentages between runs are normal. Treat only consistent directional changes (across 2+ runs or >15% delta) as actionable signal.
• Live data variability: If the app fetches live API data, different responses between runs change rendering workload independently of code changes. Note when data-dependent screens show variance.