Nsight Graphics 2026 Skill

Wrapper around NVIDIA Nsight Graphics 2026.1+. The agent captures GPU frames, gets 3 small JSON artifacts, and uses drill-down subcommands to answer follow-up questions without ever loading the raw 300+ MB TSV bundle.

When to use

User intents that should trigger this skill:

• "Why is this frame slow on the GPU?" / "Profile GPU performance" / "Grab a frame to inspect"
• "Capture a frame from this game" / "frame trace" / "GPU Trace"
• "What draw calls does this frame make?" / "API stream"
• "How fast does this .ngfx-capture replay?" / "replay perf"
• "Inspect markers / stages / metrics in this .ngfx-gputrace"
• "Generate a C++ capture so I can edit and replay it"

Trigger keywords: nsight, ngfx, nsight graphics, GPU trace, frame capture, frame trace, gpu profiling, frame profiling.

Quick start (90% case)

# 1. capture a 200 ms GPU Trace 30 s after launch (game must already be loaded)
python "<SKILL_DIR>/scripts/nsight.py" gputrace-capture \
  --exe "C:\Game\game.exe" --wd "C:\Game" \
  --start-after-ms 30000 --max-duration-ms 200 \
  --architecture Ada --metric-set-name "Throughput Metrics" \
  --time-every-action \
  --out "D:\captures\game.ngfx-gputrace"

# 2. the wrapper writes 3 small JSON next to the .ngfx-gputrace:
#    <session>/game.gputrace.summary.json  (~50 KB)
#    <session>/game.gputrace.stages.json   (~3  KB)
#    <session>/game.gputrace.actions.json  (~15 KB)

# 3. drill into the dominant stage from summary.json's analysis.hotspots
python "<SKILL_DIR>/scripts/nsight.py" gputrace-stages \
  "<session>/game.ngfx-gputrace" --parent "Render Camera" --top 10

Replace <SKILL_DIR> with the directory containing this SKILL.md. For any subcommand's full flag list run python scripts/nsight.py <cmd> --help.

Three rules to never break:

• GPUTRACE_REGIMES.xls is 300+ MB — never Read it directly. Drill through gputrace-stages / gputrace-actions / gputrace-metric.
• Anti-cheat games (EAC/BattlEye/Vanguard) refuse capture and exit with code 4. Try cpp-capture or disable the anti-cheat per game policy.
• If a capture fails with a permission error, re-run from an elevated PowerShell. doctor reports is_admin.

Tool layout

A single Python entry point dispatches to 25 subcommands. Full flag list for any subcommand: python "<SKILL_DIR>/scripts/nsight.py" <cmd> --help.

Group	Subcommands	Purpose
Env	locate / doctor / capabilities / kill	Detect install, self-check, dump per-binary flags, kill residuals
Run no-capture	launch / attach	Run game under ngfx with no capture taken
Capture	capture / cpp-capture / gputrace-capture	Graphics Capture / C++ Capture / GPU Trace + 3 JSON
Trigger	trigger-hotkey	Synthesize F11 (or another F-key) into a target window so the agent — not a human — fires a --start-after-hotkey capture
Post-process	gputrace	Rebuild 3 JSON from existing trace
Drill	gputrace-stages / gputrace-actions / gputrace-metric	Stage tree / leaf-marker top-N / metric aggregate
Diagnose (7)	gputrace-{stalls,bandwidth,shader-bound,texture-cache,overdraw,geometry,draws}	Focused verdict[]-producing diagnostic commands
Replay	export-{metadata,functions,screenshot} / replay-perf / replay-analyze	Metadata / API stream / PNG extraction; replay timing; combined

Exit codes: 0 success, 2 user error, 3 Nsight not found, 4 underlying tool failed, 5 wrapper-side timeout. stdout is JSON for query commands; pass --out FILE to redirect.

Parameter Decision Guide

For decisions beyond the Quick Start path — picking a start trigger, tuning trace quality vs cost, navigating drill-down after capture, or reaching for an advanced ngfx flag — Read references/parameters.md. That file contains five decision tables:

• A. User goal → command → required flags → what to read
• B. GPU Trace start trigger choice (--start-after-ms vs --start-after-frames vs --start-after-hotkey vs SDK)
• C. Trace quality vs cost trade-off (--time-every-action, --set-gpu-clocks, --real-time-shader-profiler, etc.)
• D. Drill-down workflow once the three JSON artifacts are written
• E. Advanced 1:1 ngfx flags (rare; skip unless asked)

⚠️ One red-line warning that belongs in the body, not the reference: --multi-pass-metrics is unusable through this wrapper (Nsight 2025.3 + 2026.1, verified May 2026). Combined with the mandatory --auto-export it deterministically writes an unloadable .ngfx-gputrace. The wrapper still accepts the flag for forensic / bug-report purposes but emits a runtime WARNING and returns bundle_complete=False. Full investigation in DESIGN.md → Investigations.

Capabilities-driven feature gating

The wrapper checks every conditional flag against the live ngfx install and exits with code 4 + a clear message if your build lacks that flag — preventing cryptic ngfx parse errors. Run python scripts/nsight.py capabilities to see what's available locally; wrapper_features.<key> maps each wrapper flag to its underlying ngfx flag.

Subcommand reference

Full flag list for any subcommand: python scripts/nsight.py <cmd> --help. This section is the agent-facing purpose only; flag detail lives in --help.

gputrace-capture

Wraps ngfx.exe --activity "GPU Trace Profiler" --auto-export .... Always writes <session>/<file>.ngfx-gputrace + BASE/ + 3 JSON artifacts. Use --dry-run to preview the ngfx command line.

trigger-hotkey

Synthesize a function-key press (default F11) into a target window via Win32 SendInput. Pair with gputrace-capture --start-after-hotkey run in the background to let the agent — not a human — fire the trigger once an external workflow has reached the desired scene. See the Agent-triggered capture workflow pattern below.

gputrace

Re-runs the parser pipeline against an existing trace + BASE/. Use after a skill upgrade to regenerate JSON without re-launching the game.

gputrace-stages

Stage-tree drill. --parent REGEX returns children of all matching parents grouped by name. --depth N restricts to a specific depth.

gputrace-actions

Top-N leaf markers (deepest D3DPERF_EVENTS nodes). --filter matches the leaf's own name; --in-marker matches any ancestor in the path. --with-metrics adds the headline block via one streaming REGIMES pass (~1-2 s). Important: read the definition field — actions are leaf markers, not raw draw calls.

gputrace-metric

Aggregate one metric. --name PATTERN is a regex over the metric catalog; --all-matches allows multi-match output. --in-marker REGEX returns per-marker windows in addition to the global aggregate.

Diagnostic command family (7 commands)

gputrace-stalls, gputrace-bandwidth, gputrace-shader-bound, gputrace-overdraw, gputrace-geometry, gputrace-texture-cache, gputrace-draws — focused diagnostic commands. Each returns a verdict[] array of {tag, severity, message}, severity ∈ {info, medium, high}. Thresholds are heuristics; treat verdicts as investigation starters, not absolute judgments. Most accept --in-marker REGEX to scope to a subtree (gputrace-stalls and gputrace-draws are whole-frame only).

Recommended investigation order on an unknown frame:

1. gputrace-stalls — rule out CPU-bound / pipeline-bubble first; if the GPU isn't busy, the rest of the analysis doesn't matter.
2. gputrace-bandwidth — memory-bound vs compute-bound axis.
3. Branch:
   • memory-bound → gputrace-texture-cache to localize
   • compute-bound → gputrace-shader-bound for SM utilization detail
4. gputrace-overdraw — opaque-pass quality.
5. gputrace-geometry — vertex/primitive frontend.
6. gputrace-draws — CPU-side state-churn / small-batch signals.

replay-perf

Replays a .ngfx-capture N times via ngfx-replay -n N --perf-report-dir and parses iteration_times.csv. Useful for replay-cost regression; not for original-app GPU performance (that's gputrace-capture).

How to read the artifacts

Read summary.json first. Four fields that matter:

• analysis.frame_budget.verdict → 60fps / 30fps / below_30fps
• analysis.throughput.dominant → most loaded subsystem
• analysis.hotspots.slowest_stage → where to drill
• analysis.warnings → natural-language flags

Then pick a drill direction: stage > 50% GPU → drill it; throughput dominated by dram/pcie → memory-heavy; sm dominant → shaders. Always drill via subcommands, never by reading BASE/*.xls directly (REGIMES is 300+ MB).

Top-level JSON shapes:

summary.json:  source, session, summary{frame/marker/metric counts},
               analysis{frame_budget, throughput, hotspots, warnings},
               headline_metrics, metrics[], hardware_context
stages.json:   source, headline_metrics, roots[], top_stages[]
               each: name, depth, total_duration_ns, fraction_of_gpu, headline
actions.json:  source, definition, top_20_slowest_actions[]
               each: name, path, instance_count, duration fields, headline

Diagnostic verdict thresholds — use these to interpret verdict[] beyond just severity:

Command	Key signal	Healthy	Red
overdraw	overdraw_ratio	~1.0 (1.5-3 typical)	> 3
overdraw	zcull_rejection_rate	> 0.3	< 0.3 (ZCull defeated)
overdraw	late_z_attrition_rate	< 0.3	high (PS work thrown away)
bandwidth	per-tier % of peak	< 60%	≥ 80% saturated (60-80% pressure)
bandwidth	dominant tier vs SM	balanced	tier ≥ SM by 15pp → memory-bound
bandwidth	PCIe sustained	< 30%	≥ 30% (host↔device thrash)
shader-bound	sm_stall_ratio	low	high (memory latency / dep chains)
shader-bound	async-compute use	balanced	compute ≥ 30% but < 10% async (underused)
geometry	pixels_per_prim	≥ 16	< 4 (micro-triangles)
texture-cache	l1tex__t_sector_hit_rate.pct	≥ 70%	< 50% (thrashing)
texture-cache	miss_to_dram	low	≥ 10% (DRAM burned on misses)
stalls	gr_idle_pct	low	high (GPU not fed)
stalls	marker_coverage_pct	high	low (idle BETWEEN markers → CPU bound / waits)
draws	small_leaf_pct (< 5 μs)	low	high (batching candidates)

Workflow patterns

First-pass perf analysis

gputrace-capture → read summary.json → 1-2 drill rounds → write Markdown report. Stop drilling once the bottleneck is at the engine-marker level (e.g. "GPUDriven.RenderMesh(Bush)") or attributable to a metric pattern (e.g. dram throughput > 80%).

Deep bottleneck dive (single stage > 80% of frame time)

gputrace-stages --parent "<stage>" --top 20
gputrace-actions --in-marker "<stage>" --top 20 --with-metrics
gputrace-metric --name "dramc__throughput" --in-marker "<stage>"
gputrace-metric --name "sm__throughput" --in-marker "<stage>"

Capture-once-then-iterate

gputrace-capture --exe ... --out trace.ngfx-gputrace
# now drill repeatedly; no re-launch needed
gputrace-stages "<session>/trace.ngfx-gputrace" ...
gputrace-actions "<session>/trace.ngfx-gputrace" ...

Reanalyze an old trace

gputrace "<session>/old-trace.ngfx-gputrace"   # rebuilds 3 JSON

cpp-capture for repro (deterministic C++ replay code)

cpp-capture --exe ... --wait-seconds 10 --out D:\repro

Attach to a running game (can't relaunch, e.g. multiplayer)

attach --activity "GPU Trace Profiler" --pid 12345
# user presses F11 in the running app

Agent-triggered capture in a larger workflow

Use this when: the agent is orchestrating a wider workflow — launching the game, running automated tests / bot actions / other skills, then deciding when to capture — and only the agent (not a human at the keyboard) knows when the target scene is on screen. Timer-based triggers (--start-after-ms / --start-after-frames) don't fit because the runtime of the surrounding workflow is unpredictable.

Why F11 simulation is the only option: ngfx's GPU Trace activity exposes exactly five start triggers (--start-after-ms / --start-after-frames / --start-after-submits / --start-after-hotkey / --start-with-ngfx-sdk / --start-on-replay-begin). ngfx-rpc.exe is the UI↔replayer IPC, not a trace-trigger RPC, and --start-with-ngfx-sdk requires the game to link the NGFX SDK and call NGFX_GPUTrace_StartTrace. For an unmodified target, hotkey is the only externally addressable trigger, and trigger-hotkey synthesizes it without a human.

Pattern:

# 1. Launch ngfx + game in the background, armed for F11 trigger.
#    DO NOT wait on this command — it blocks until the trace is written.
python scripts/nsight.py gputrace-capture \
  --exe "C:\Game\game.exe" --wd "C:\Game" \
  --start-after-hotkey \
  --max-duration-ms 200 \
  --architecture <arch> --metric-set-name "Throughput Metrics" \
  --time-every-action \
  --out "D:\captures\game.ngfx-gputrace"
# (Run via your harness's background-task primitive; capture the task ID.)

# 2. Attach a streaming watcher to the background task's output so the
#    agent is notified when the bundle is written. The wrapper prints
#    "GPU Trace report saved to ..." then "bundle_complete=True". Watch
#    for those lines (and also for "error", "fatal", "denied" to fail fast).

# 3. The agent now runs the rest of the workflow (other skills, automated
#    tests, scripted gameplay) — no polling. The capture is dormant in
#    the game process; F11 has not been pressed yet.

# 4. When the workflow signals "ready to capture":
python scripts/nsight.py trigger-hotkey --process game
# Returns JSON with foreground_ok / sent fields. exit 0 = key delivered.

# 5. ngfx hook receives F11, runs the 200 ms trace, writes the bundle,
#    wrapper post-processes into 3 JSON. The streaming watcher from step 2
#    fires; the agent reads summary.json and drills as usual.

Notes:

• Background launch is mandatory. gputrace-capture blocks until the bundle is written, which is "forever" from the agent's point of view (it depends on when step 4 fires). The agent must run step 1 asynchronously.
• Watch for completion, do not poll. The wrapper writes bundle_complete=True
  • the saved path to its stdout/stderr; pipe that into your harness's streaming notification primitive so the agent reacts to the event rather than guessing.
• Windowed/borderless mode preferred. SendInput is reliable against windowed and borderless-window games. Exclusive-fullscreen + DRM combinations can swallow synthetic keystrokes; if trigger-hotkey reports sent=true but no trace appears, switch the game to borderless-window and retry.
• Foreground stealing is restricted on Windows. trigger-hotkey calls AllowSetForegroundWindow(ASFW_ANY) + SetForegroundWindow, which works from an elevated shell (doctor reports is_admin=true) and usually from a non-elevated shell that recently saw user input. If it can't bring the window forward it still sends the key (ngfx's hook is global) and emits a stderr WARNING. Pass --no-foreground to skip the foreground step entirely if you'd rather not steal focus.
• Use --use-scancode if the default fails. Some engines / DRM shims only honour hardware scancodes (KEYEVENTF_SCANCODE). The flag toggles the wParam encoding without other changes.
• ngfx 2026.1.x cleanup-phase crash is benign. After the trace is written, ngfx exits with rc 0xC0000409 after trace_written=True and bundle_complete=True. The wrapper detects this and still produces the 3 JSON. Don't treat the non-zero rc as failure; check bundle_complete instead.
• Disambiguating multiple instances. If tasklist finds more than one process matching --process, trigger-hotkey refuses and lists the PIDs; re-run with --pid <N>. Always use --pid when ngfx itself launched the game (you have the PID from the capture-task output).

Out of scope

• Per-API-call timing (no vkCmdDraw durations — Nsight 2026.1 removed GPUTrace.pyd; actions in this skill are leaf markers, not raw draw calls). Use RenderDoc/PIX for per-API timing.
• HTML/PDF report generation — agent writes Markdown directly.
• Capture diff (compare two runs) — future work.
• Cross-platform — Windows only.
• Auto-installing Nsight — doctor only detects.
• Remote capture (ngfx-rpc) — use ngfx-rpc.exe manually if needed.
• NVTX marker injection — game-side, not skill-side.
• GUI launching — open ngfx-ui.exe manually for visual inspection.