How It Works

How It Works

Snitch is built to be honest: it never shows a number it can't stand behind.

The layers

1. Frame time + GC (load-bearing). A rolling window of frame times yields mean/median/p95/p99 and fps, plus GC gen-0/gen-1 collections per 1000 frames. This is build-independent and always reliable - it is the truth everything else is checked against.
2. Section timing (the "what is slow"). Stopwatch accumulators time named sections - both the ones a mod registers via the API and (with snitch vanilla on) vanilla hot paths wrapped by Harmony. These are self-measured: they only see the methods Snitch explicitly wraps, and they include a small patch overhead. They are not a sampling profiler.
3. State distributions. Cheap, read-only tallies of entities by state (NPCs by movement/visibility, trash by physics state, quests by state, plus any modder provider), polled a few times a second.
4. Ablation (causal cost). snitch ablate <lever> toggles a subsystem off, lets the frame time settle under a stability gate, and reports the delta - the build-independent "total cost" of that subsystem, including native cost the section timers can't see.

Honesty rules

• ProfilerRecorder engine counters are inert in Schedule I's IL2CPP build, so Snitch does not rely on them - frame-time + GC are the load-bearing truth.
• Section costs are self-measured (only wrapped methods). For "total subsystem cost" including native work (e.g. the NavMeshAgent), use ablation.
• Snitch reports its own overhead as the Snitch.Self section, so you can see what the profiler costs.
• Everything is idle until snitch start. Profiling is read-only and safe in multiplayer; state-mutating features (ablation levers) run host-only.

A worked finding

NPCMovement.Update is cheap (~0.06 ms/frame for ~200 NPCs); NPCMovement.FixedUpdate costs more (it runs at the physics rate). But the dominant per-NPC cost is native (the NavMeshAgent), which only ablation reveals - exactly the kind of thing Snitch is built to surface.