diff --git a/docs/renderer-profiler-plan.md b/docs/renderer-profiler-plan.md
new file mode 100644
index 000000000..a3aaf3219
--- /dev/null
+++ b/docs/renderer-profiler-plan.md
@@ -0,0 +1,352 @@
+# Renderer Profiler — Design Plan
+
+Production-grade instrumentation for OpenPencil's CanvasKit/Skia renderer, inspired by game engine profiling tools (Unreal Insights, Unity Frame Timing Manager) and browser render engine internals (Chrome Compositor, WebKit/Skia tracing).
+
+## What Skia gives us
+
+### 1. Profiling build of CanvasKit
+
+`canvaskit-wasm` ships three bundles:
+- **default** — stripped WASM names
+- **full** — includes Skottie etc.
+- **profiling** — same as full but with **full internal function names** preserved in WASM
+
+```ts
+import InitCanvasKit from 'canvaskit-wasm/profiling'
+```
+
+This means Chrome DevTools Performance panel flame charts will show real Skia C++ function names (`GrDrawingManager::internalFlush`, `SkCanvas::drawRect`, etc.) instead of mangled `$func123`. We should use this build behind a `?profiling` URL param or a dev-mode toggle.
+
+### 2. SkPicture recording & replay
+
+Skia's `SkPictureRecorder` records draw commands into an immutable `SkPicture`. We already use this for caching (`recordScenePicture`). Key profiling insight from Skia's own benchmark tool:
+
+> `CanvasKit.flush()` returns after it has sent all instructions to the GPU, but we don't know the GPU is done until the **next frame is requested**. Thus, we need to keep track of time **between frames** to accurately calculate draw time.
+
+- `drawPicture + flush` duration tells us CPU-side WASM/JS time
+- `total_frame_ms` (time between rAFs) tells us the real GPU-inclusive cost
+- If `total_frame_ms ≈ with_flush_ms` → CPU-bound
+- If `total_frame_ms >> with_flush_ms` → GPU-bound
+
+### 3. Skia Debugger (external tool)
+
+Skia has an online debugger at https://debugger.skia.org that can load `.skp` files and show:
+- Draw command playback
+- GPU op bounds visualization
+- Overdraw visualization
+- Clip/matrix state at any step
+- GPU operation batching (colored op IDs)
+
+We could add an "Export .skp" debug feature to save the current `scenePicture` for analysis.
+
+### 4. EXT_disjoint_timer_query (GPU timing)
+
+The `EXT_disjoint_timer_query_webgl2` WebGL extension provides **actual GPU-side timing** in nanoseconds. Figma built [figma/webgl-profiler](https://github.com/figma/webgl-profiler) on top of this exact extension.
+
+**Limitations:**
+- Only works in Desktop Chrome ≥ 70
+- `TIMESTAMP_EXT` was removed for security; only `TIME_ELAPSED_EXT` (begin/end query) works
+- Results are asynchronous (available next frame or later)
+- Non-draw commands often report 0ns (but totals are accurate)
+- Inconsistent on non-Apple-Silicon Intel GPUs
+
+**Usage pattern:**
+```ts
+const ext = gl.getExtension('EXT_disjoint_timer_query_webgl2')
+const query = gl.createQuery()
+gl.beginQuery(ext.TIME_ELAPSED_EXT, query)
+// ... draw calls ...
+gl.endQuery(ext.TIME_ELAPSED_EXT)
+// next frame: poll gl.getQueryParameter(query, gl.QUERY_RESULT_AVAILABLE)
+```
+
+### 5. Chrome Performance Panel extensibility
+
+Chrome now supports custom tracks via `performance.mark`/`performance.measure` with `detail.devtools` metadata, and the lower-overhead `console.timeStamp` API:
+
+```ts
+performance.measure("Scene Render", {
+  start: startMark,
+  detail: {
+    devtools: {
+      dataType: "track-entry",
+      track: "OpenPencil Renderer",
+      trackGroup: "OpenPencil",
+      color: "primary",
+      properties: [["Nodes", "142"], ["Cached", "true"]],
+      tooltipText: "Full scene render with SkPicture cache hit"
+    }
+  }
+})
+```
+
+This shows up as a dedicated "OpenPencil Renderer" track in the DevTools Performance panel alongside browser internals. Extremely powerful for production debugging.
+
+---
+
+## Architecture
+
+### Layers of profiling (from light to heavy)
+
+| Layer | Overhead | Always-on | What it measures |
+|-------|----------|-----------|-----------------|
+| **Frame Stats HUD** | ~0.1ms | Optional toggle | FPS, frame time, CPU/GPU split, node count, cache hits |
+| **User Timing marks** | ~0.01ms each | Dev builds | Phase timings visible in DevTools Performance panel |
+| **GPU Timer Queries** | ~0.2ms setup | On demand | Actual GPU execution time per frame (async readback) |
+| **Detailed profiler** | ~1-3ms | Manual trigger | Per-node timings, draw call counts, overdraw heatmap |
+
+### 1. Frame Stats HUD (always-available, minimal overhead)
+
+An in-canvas overlay rendered by SkiaRenderer itself, togglable via keyboard shortcut (Shift+P or similar). Modeled after game engine stat overlays.
+
+**Metrics displayed:**
+```
+ FPS: 60 (16.7ms)       ← smoothed rAF delta
+ CPU: 4.2ms             ← JS/WASM time (start of render → after flush)
+ GPU: 8.1ms             ← GPU time (async, from EXT_disjoint_timer_query or inferred)
+ ─────────────────
+ Nodes: 342 (47 culled)  ← total vs viewport-culled count
+ Draw calls: 128         ← WebGL draw call count
+ Cache: SkPicture HIT    ← scene picture reuse status
+ Textures: 12 (48MB)     ← image cache stats
+ WASM heap: 64MB         ← CanvasKit memory
+```
+
+**Frame time graph** — rolling 120-frame history bar chart (like Unreal's `stat unit`), color-coded:
+- Green: within budget (16.7ms for 60fps)
+- Yellow: over budget but < 2× (16.7–33.3ms)
+- Red: severe (> 33.3ms)
+
+The graph draws CPU bars from bottom and GPU bars from top, making CPU/GPU overlap visible at a glance (exactly how Unity Frame Timing Manager visualizes it).
+
+### 2. Phase instrumentation (User Timing API)
+
+Instrument the `render()` method with `performance.mark`/`performance.measure` using the Chrome DevTools extensibility API to create a custom "OpenPencil" track group:
+
+**Phases to instrument:**
+
+| Phase | What | Where in code |
+|-------|------|---------------|
+| `frame` | Total rAF-to-rAF | `useCanvas.ts` rAF callback |
+| `render:scene` | Scene drawing (picture replay OR live render) | `render()` |
+| `render:recordPicture` | SkPicture recording | `recordScenePicture()` |
+| `render:drawPicture` | SkPicture replay | `canvas.drawPicture()` |
+| `render:volatile` | Live render (hover/drag/text edit) | volatile path |
+| `render:sectionTitles` | Section title drawing | `drawSectionTitles()` |
+| `render:componentLabels` | Component labels | `drawComponentLabels()` |
+| `render:selection` | Selection borders/handles | `drawSelection()` |
+| `render:rulers` | Ruler drawing | `drawRulers()` |
+| `render:flush` | GPU command submission | `surface.flush()` |
+| `layout:compute` | Yoga layout | `computeAllLayouts()` |
+
+Replace the existing `console.time`/`console.timeEnd` calls (which already exist in the render method) with the structured User Timing API:
+
+```ts
+// Before (current)
+console.time('render:flush')
+this.surface.flush()
+console.timeEnd('render:flush')
+
+// After (with DevTools custom tracks)
+const start = performance.now()
+this.surface.flush()
+performance.measure('flush', {
+  start,
+  detail: {
+    devtools: {
+      dataType: 'track-entry',
+      track: 'Renderer',
+      trackGroup: 'OpenPencil',
+      color: 'tertiary'
+    }
+  }
+})
+```
+
+### 3. GPU Timer Queries
+
+Wrap the WebGL context to insert `EXT_disjoint_timer_query_webgl2` queries around the flush boundary:
+
+```ts
+class GpuTimer {
+  private ext: EXT_disjoint_timer_query_webgl2 | null
+  private pending: WebGLQuery[]
+  private results: number[] // rolling buffer of GPU times in ms
+
+  beginFrame(gl: WebGL2RenderingContext) { ... }
+  endFrame(gl: WebGL2RenderingContext) { ... }
+  pollResults(gl: WebGL2RenderingContext): number | null { ... }
+}
+```
+
+Since results come back asynchronously (typically 1-4 frames later), the HUD shows the most recently completed GPU measurement, not the current frame's. This is the same approach Unity and Unreal use.
+
+For accurate CPU vs GPU bottleneck detection (Figma's approach):
+- `perceived_fps` = 1000 / rAF delta
+- `cpu_fps` = 1000 / (JS render end − JS render start)  
+- `gpu_fps` = 1000 / GPU timer query result
+
+If `gpu_fps ≈ perceived_fps` and `cpu_fps >> perceived_fps` → **GPU-bound**  
+If `cpu_fps ≈ perceived_fps` and `gpu_fps >> perceived_fps` → **CPU-bound**
+
+### 4. Detailed Profiler (on-demand, higher overhead)
+
+Triggered via debug menu or keyboard shortcut. Captures detailed per-node data for a single frame or short burst:
+
+**Per-node metrics:**
+- Time spent in `renderNode()` (cumulative, including children)
+- Self time (excluding children)
+- Number of draw calls generated
+- Whether the node was culled
+- Fill/stroke/effect complexity (number of paints, blur radii, etc.)
+
+**Implementation:** A profiling wrapper around `renderNode()` that pushes/pops a stack:
+
+```ts
+private renderNodeProfiled(canvas, graph, nodeId, ...) {
+  const entry = this.profilerStack.push(nodeId)
+  entry.startTime = performance.now()
+  this.renderNode(canvas, graph, nodeId, ...)
+  entry.endTime = performance.now()
+  this.profilerStack.pop()
+}
+```
+
+**Overdraw visualization:** Skia has a built-in overdraw mode (`SkOverdrawCanvas` / `OverdrawColorFilter`). In CanvasKit, we can achieve this by rendering with `CanvasKit.MakeMatrix` color filter that maps draw count → heat colors. Alternative: render each node's bounds with additive blend to accumulate overdraw count as brightness.
+
+**Output formats:**
+- **In-canvas heatmap** — overlay showing per-pixel render cost (warm = slow)
+- **Speedscope-compatible JSON** — export frame profile for viewing in https://www.speedscope.app/ (same format Figma's webgl-profiler uses)
+- **Node table** — sorted list of most expensive nodes with self-time
+
+### 5. Draw Call Counter
+
+Proxy the WebGL context to count actual GL draw calls per frame:
+
+```ts
+function instrumentGL(gl: WebGL2RenderingContext) {
+  let drawCalls = 0
+  const origDrawArrays = gl.drawArrays.bind(gl)
+  const origDrawElements = gl.drawElements.bind(gl)
+  gl.drawArrays = (...args) => { drawCalls++; return origDrawArrays(...args) }
+  gl.drawElements = (...args) => { drawCalls++; return origDrawElements(...args) }
+  return {
+    getAndReset() { const c = drawCalls; drawCalls = 0; return c }
+  }
+}
+```
+
+This is the standard technique used by Spector.js, WebGL Inspector, and every game engine's stat display.
+
+---
+
+## Implementation plan
+
+### Phase 1: Foundation (`packages/core/src/profiler.ts`)
+
+Create a `RenderProfiler` class that all layers use:
+
+```ts
+export class RenderProfiler {
+  enabled = false
+  hudVisible = false
+  
+  // Frame stats (always tracked when enabled, near-zero cost)
+  readonly frameStats: FrameStats
+  
+  // Phase timing (User Timing API integration)
+  beginPhase(name: string): void
+  endPhase(name: string): void
+  
+  // GPU timer (optional, requires extension)
+  readonly gpuTimer: GpuTimer | null
+  
+  // Draw call counter
+  readonly drawCallCounter: DrawCallCounter | null
+  
+  // Detailed per-node profiling (on-demand)
+  beginDetailedCapture(): void
+  endDetailedCapture(): FrameCapture
+  
+  // HUD rendering (called by SkiaRenderer)
+  drawHUD(canvas: Canvas, ck: CanvasKit): void
+}
+```
+
+### Phase 2: Integrate into renderer
+
+1. Add `profiler: RenderProfiler` to `SkiaRenderer`
+2. Replace all `console.time`/`console.timeEnd` with `profiler.beginPhase`/`endPhase`
+3. Add phase markers for every render stage
+4. Wrap `surface.flush()` with GPU timing queries
+5. Add HUD drawing at end of `render()` (in screen space, after everything else)
+
+### Phase 3: HUD overlay
+
+Render directly on the Skia canvas as the very last step (after rulers). Uses small monospace font, semi-transparent dark background. Positioned top-left or bottom-left, not interfering with rulers.
+
+### Phase 4: GPU timing & bottleneck detection
+
+1. Implement `GpuTimer` class with `EXT_disjoint_timer_query_webgl2`
+2. Add CPU vs GPU bottleneck indicator to HUD
+3. Handle graceful degradation when extension unavailable
+
+### Phase 5: Detailed profiler & export
+
+1. Per-node timing capture
+2. Speedscope JSON export
+3. Overdraw visualization mode
+4. Node cost table (could display in dev panel or console)
+
+---
+
+## File structure
+
+```
+packages/core/src/
+  profiler/
+    index.ts              — re-exports
+    render-profiler.ts    — main RenderProfiler class
+    frame-stats.ts        — FPS/frame-time tracking with rolling averages
+    gpu-timer.ts          — EXT_disjoint_timer_query wrapper
+    draw-call-counter.ts  — WebGL proxy for draw call counting
+    phase-timer.ts        — User Timing API / console.timeStamp integration
+    hud-renderer.ts       — in-canvas HUD overlay drawing
+    frame-capture.ts      — detailed per-node capture data structures
+    speedscope-export.ts  — export to speedscope JSON format
+```
+
+## Activation
+
+| Mechanism | What activates |
+|-----------|---------------|
+| `?profiling` URL param | Use profiling CanvasKit build with full WASM names |
+| `Shift+P` keyboard shortcut | Toggle HUD overlay |
+| `Shift+Alt+P` | Start/stop detailed frame capture |
+| Dev menu → "Export frame profile" | Save speedscope JSON |
+| Dev menu → "Show overdraw" | Toggle overdraw heatmap |
+| `store.debug.profiler` | Programmatic control from Vue devtools |
+
+## Key design principles
+
+1. **Zero cost when off** — no allocations, no timing calls, no proxy overhead when profiler is disabled
+2. **Minimal cost when HUD-only** — < 0.2ms overhead for the basic stats display  
+3. **Render on canvas, not DOM** — HUD is drawn by Skia itself, no DOM overlays that could interfere with WebGL
+4. **Async GPU results** — never stall the pipeline waiting for GPU query results; show N-2 frame's GPU time
+5. **Production-safe** — the basic HUD and User Timing marks can ship in production behind a flag
+6. **DevTools-native** — leverage Chrome's Performance panel extensibility for zero-overhead external profiling
+7. **Export-friendly** — speedscope JSON format is the standard for flame chart sharing
+
+## References
+
+- [Skia Tracing](https://skia.org/docs/dev/tools/tracing) — Skia's internal `TRACE_EVENT` macros, Perfetto integration
+- [Skia Debugger](https://debugger.skia.org) — SKP visual debugger with GPU op bounds
+- [Skia Perf](https://skia.org/docs/dev/testing/skiaperf) — Skia's continuous perf monitoring with 400k+ metrics per commit
+- [CanvasKit profiling build](https://www.npmjs.com/package/canvaskit-wasm) — `canvaskit-wasm/profiling` with full WASM function names
+- [Figma webgl-profiler](https://github.com/figma/webgl-profiler) — Figma's GPU profiler using EXT_disjoint_timer_query, speedscope output
+- [Skia render-skp.html](https://skia.googlesource.com/skia/+/052566d8ccb7/tools/perf-canvaskit-puppeteer/render-skp.html) — Skia's own CanvasKit benchmark: frame-to-frame timing methodology
+- [EXT_disjoint_timer_query_webgl2 spec](https://registry.khronos.org/webgl/extensions/EXT_disjoint_timer_query_webgl2/) — WebGL GPU timing extension
+- [Chrome Performance extensibility API](https://developer.chrome.com/docs/devtools/performance/extension) — Custom tracks via `performance.measure` detail.devtools
+- [Unity Frame Timing Manager](https://docs.unity3d.com/2022.3/Documentation/Manual/frame-timing-manager.html) — CPU main thread / render thread / GPU split
+- [Unreal Insights](https://dev.epicgames.com/documentation/en-us/unreal-engine/introduction-to-performance-profiling-and-configuration-in-unreal-engine) — Per-thread, per-GPU frame profiling
+- Skia Milestone 132-133 release notes — `GrGLInterface` timer query support, `GpuStats.elapsedTime` for GPU time reporting
diff --git a/packages/core/src/index.ts b/packages/core/src/index.ts
index 27477f142..ee54fa06d 100644
--- a/packages/core/src/index.ts
+++ b/packages/core/src/index.ts
@@ -48,6 +48,8 @@ export { FigmaAPI, FigmaNodeProxy, type FigmaFontName } from './figma-api'
 export { ALL_TOOLS, defineTool, toolsToAI } from './tools'
 export type { ToolDef, ParamDef, ParamType } from './tools'
 export { SkiaRenderer, type RenderOverlays } from './renderer'
+export { RenderProfiler } from './profiler'
+export type { FrameCapture, NodeProfile } from './profiler'
 export { computeLayout, computeAllLayouts, setTextMeasurer } from './layout'
 export type { TextMeasurer } from './layout'
 export { getCanvasKit, getGpuBackend, type CanvasKitOptions, type GpuBackend } from './canvaskit'
diff --git a/packages/core/src/profiler/draw-call-counter.ts b/packages/core/src/profiler/draw-call-counter.ts
new file mode 100644
index 000000000..becc5be2e
--- /dev/null
+++ b/packages/core/src/profiler/draw-call-counter.ts
@@ -0,0 +1,45 @@
+const DRAW_METHODS = [
+  'drawArrays',
+  'drawElements',
+  'drawArraysInstanced',
+  'drawElementsInstanced'
+] as const
+
+type DrawMethod = (typeof DRAW_METHODS)[number]
+
+export class DrawCallCounter {
+  count = 0
+
+  private originals = new Map<DrawMethod, (...args: unknown[]) => void>()
+  private gl: WebGL2RenderingContext | null
+
+  constructor(gl: WebGL2RenderingContext | null) {
+    this.gl = gl
+    if (!gl) return
+
+    for (const method of DRAW_METHODS) {
+      const original = gl[method].bind(gl) as (...args: unknown[]) => void
+      this.originals.set(method, original)
+      ;(gl as unknown as Record<string, unknown>)[method] = (...args: unknown[]) => {
+        this.count++
+        original(...args)
+      }
+    }
+  }
+
+  reset(): number {
+    const prev = this.count
+    this.count = 0
+    return prev
+  }
+
+  destroy(): void {
+    const gl = this.gl
+    if (!gl) return
+
+    for (const [method, fn] of this.originals) {
+      ;(gl as unknown as Record<string, unknown>)[method] = fn
+    }
+    this.originals.clear()
+  }
+}
diff --git a/packages/core/src/profiler/frame-capture.ts b/packages/core/src/profiler/frame-capture.ts
new file mode 100644
index 000000000..2ee8a6059
--- /dev/null
+++ b/packages/core/src/profiler/frame-capture.ts
@@ -0,0 +1,138 @@
+const hasPerformance = typeof performance !== "undefined"
+
+export interface NodeProfile {
+  nodeId: string
+  name: string
+  type: string
+  depth: number
+  startTime: number
+  endTime: number
+  selfTime: number
+  drawCalls: number
+  culled: boolean
+  children: NodeProfile[]
+}
+
+export interface FrameCapture {
+  timestamp: number
+  totalTimeMs: number
+  cpuTimeMs: number
+  gpuTimeMs: number
+  totalNodes: number
+  culledNodes: number
+  drawCalls: number
+  scenePictureCacheHit: boolean
+  rootProfiles: NodeProfile[]
+}
+
+export class CaptureStack {
+  private stack: NodeProfile[] = []
+  private roots: NodeProfile[] = []
+  private frameStart = 0
+
+  begin(nodeId: string, name: string, type: string, culled: boolean): void {
+    const profile: NodeProfile = {
+      nodeId,
+      name,
+      type,
+      depth: this.stack.length,
+      startTime: hasPerformance ? performance.now() - this.frameStart : 0,
+      endTime: 0,
+      selfTime: 0,
+      drawCalls: 0,
+      culled,
+      children: [],
+    }
+    this.stack.push(profile)
+  }
+
+  end(drawCallsDelta: number): void {
+    const profile = this.stack.pop()
+    if (!profile) return
+
+    profile.endTime = hasPerformance ? performance.now() - this.frameStart : 0
+    profile.drawCalls = drawCallsDelta
+
+    let childrenTime = 0
+    for (const child of profile.children) {
+      childrenTime += child.endTime - child.startTime
+    }
+    profile.selfTime = profile.endTime - profile.startTime - childrenTime
+
+    const parent = this.stack[this.stack.length - 1]
+    if (parent) {
+      parent.children.push(profile)
+    } else {
+      this.roots.push(profile)
+    }
+  }
+
+  reset(frameStart: number): void {
+    this.stack.length = 0
+    this.roots.length = 0
+    this.frameStart = frameStart
+  }
+
+  getRootProfiles(): NodeProfile[] {
+    return this.roots
+  }
+}
+
+interface SpeedscopeFrame {
+  name: string
+}
+
+interface SpeedscopeEvent {
+  type: "O" | "C"
+  at: number
+  frame: number
+}
+
+export function toSpeedscopeJSON(capture: FrameCapture): string {
+  const frames: SpeedscopeFrame[] = []
+  const frameIndex = new Map<string, number>()
+  const events: SpeedscopeEvent[] = []
+
+  function getFrameIdx(nodeId: string, name: string): number {
+    const key = `${nodeId}\0${name}`
+    let idx = frameIndex.get(key)
+    if (idx === undefined) {
+      idx = frames.length
+      frames.push({ name: `${name} (${nodeId})` })
+      frameIndex.set(key, idx)
+    }
+    return idx
+  }
+
+  function walk(profile: NodeProfile): void {
+    const idx = getFrameIdx(profile.nodeId, profile.name)
+    events.push({ type: "O", at: profile.startTime, frame: idx })
+    for (const child of profile.children) {
+      walk(child)
+    }
+    events.push({ type: "C", at: profile.endTime, frame: idx })
+  }
+
+  for (const root of capture.rootProfiles) {
+    walk(root)
+  }
+
+  return JSON.stringify(
+    {
+      $schema: "https://www.speedscope.app/file-format-schema.json",
+      shared: { frames },
+      profiles: [
+        {
+          type: "evented",
+          name: "Frame Render",
+          unit: "milliseconds",
+          startValue: 0,
+          endValue: capture.totalTimeMs,
+          events,
+        },
+      ],
+    },
+    null,
+    2,
+  )
+}
diff --git a/packages/core/src/profiler/frame-stats.ts b/packages/core/src/profiler/frame-stats.ts
new file mode 100644
index 000000000..bc5bcce5f
--- /dev/null
+++ b/packages/core/src/profiler/frame-stats.ts
@@ -0,0 +1,122 @@
+const BUFFER_SIZE = 120
+
+const hasPerformance = typeof performance !== 'undefined'
+
+export class FrameStats {
+  frameTime = 0
+  cpuTime = 0
+  gpuTime = 0
+
+  minFrameTime = Infinity
+  maxFrameTime = 0
+  avgFrameTime = 0
+
+  minCpuTime = Infinity
+  maxCpuTime = 0
+  avgCpuTime = 0
+
+  minGpuTime = Infinity
+  maxGpuTime = 0
+  avgGpuTime = 0
+
+  smoothedFps = 0
+
+  totalNodes = 0
+  culledNodes = 0
+  drawCalls = 0
+  scenePictureCacheHit = false
+
+  private frameTimeBuffer = new Float64Array(BUFFER_SIZE)
+  private cpuTimeBuffer = new Float64Array(BUFFER_SIZE)
+  private gpuTimeBuffer = new Float64Array(BUFFER_SIZE)
+  private bufferIndex = 0
+  private bufferCount = 0
+  private lastTimestamp = 0
+
+  recordFrame(cpuTimeMs: number): void {
+    const now = hasPerformance ? performance.now() : 0
+
+    if (this.lastTimestamp > 0) {
+      this.frameTime = now - this.lastTimestamp
+    }
+    this.lastTimestamp = now
+
+    this.cpuTime = cpuTimeMs
+
+    const i = this.bufferIndex
+    this.frameTimeBuffer[i] = this.frameTime
+    this.cpuTimeBuffer[i] = this.cpuTime
+    this.gpuTimeBuffer[i] = this.gpuTime
+
+    this.bufferIndex = (i + 1) % BUFFER_SIZE
+    if (this.bufferCount < BUFFER_SIZE) this.bufferCount++
+
+    this.computeStats()
+  }
+
+  getFrameTimeHistory(): Float64Array {
+    return this.frameTimeBuffer
+  }
+
+  getCpuTimeHistory(): Float64Array {
+    return this.cpuTimeBuffer
+  }
+
+  getGpuTimeHistory(): Float64Array {
+    return this.gpuTimeBuffer
+  }
+
+  getBufferIndex(): number {
+    return this.bufferIndex
+  }
+
+  getBufferCount(): number {
+    return this.bufferCount
+  }
+
+  private computeStats(): void {
+    const n = this.bufferCount
+    if (n === 0) return
+
+    let ftSum = 0
+    let ftMin = Infinity
+    let ftMax = 0
+    let cpuSum = 0
+    let cpuMin = Infinity
+    let cpuMax = 0
+    let gpuSum = 0
+    let gpuMin = Infinity
+    let gpuMax = 0
+
+    for (let j = 0; j < n; j++) {
+      const ft = this.frameTimeBuffer[j]
+      ftSum += ft
+      if (ft < ftMin) ftMin = ft
+      if (ft > ftMax) ftMax = ft
+
+      const cpu = this.cpuTimeBuffer[j]
+      cpuSum += cpu
+      if (cpu < cpuMin) cpuMin = cpu
+      if (cpu > cpuMax) cpuMax = cpu
+
+      const gpu = this.gpuTimeBuffer[j]
+      gpuSum += gpu
+      if (gpu < gpuMin) gpuMin = gpu
+      if (gpu > gpuMax) gpuMax = gpu
+    }
+
+    this.minFrameTime = ftMin
+    this.maxFrameTime = ftMax
+    this.avgFrameTime = ftSum / n
+
+    this.minCpuTime = cpuMin
+    this.maxCpuTime = cpuMax
+    this.avgCpuTime = cpuSum / n
+
+    this.minGpuTime = gpuMin
+    this.maxGpuTime = gpuMax
+    this.avgGpuTime = gpuSum / n
+
+    this.smoothedFps = this.avgFrameTime > 0 ? 1000 / this.avgFrameTime : 0
+  }
+}
diff --git a/packages/core/src/profiler/gpu-timer.ts b/packages/core/src/profiler/gpu-timer.ts
new file mode 100644
index 000000000..6fb0f4640
--- /dev/null
+++ b/packages/core/src/profiler/gpu-timer.ts
@@ -0,0 +1,94 @@
+interface EXTDisjointTimerQuery {
+  TIME_ELAPSED_EXT: 0x88bf
+  GPU_DISJOINT_EXT: 0x8fbb
+}
+
+const MAX_PENDING_QUERIES = 4
+
+export class GPUTimer {
+  private gl: WebGL2RenderingContext | null
+  private ext: EXTDisjointTimerQuery | null = null
+  private pending: WebGLQuery[] = []
+  private activeQuery: WebGLQuery | null = null
+  private _lastGpuTimeMs = NaN
+
+  get available(): boolean {
+    return this.ext !== null
+  }
+
+  get lastGpuTimeMs(): number {
+    return this._lastGpuTimeMs
+  }
+
+  constructor(gl: WebGL2RenderingContext | null) {
+    this.gl = gl
+    if (gl) {
+      this.ext =
+        (gl.getExtension('EXT_disjoint_timer_query_webgl2') as EXTDisjointTimerQuery | null) ??
+        null
+    }
+  }
+
+  beginFrame(): void {
+    if (!this.gl || !this.ext) return
+    if (this.pending.length >= MAX_PENDING_QUERIES) return
+
+    const query = this.gl.createQuery()
+    if (!query) return
+
+    this.gl.beginQuery(this.ext.TIME_ELAPSED_EXT, query)
+    this.activeQuery = query
+  }
+
+  endFrame(): void {
+    if (!this.gl || !this.ext || !this.activeQuery) return
+
+    this.gl.endQuery(this.ext.TIME_ELAPSED_EXT)
+    this.pending.push(this.activeQuery)
+    this.activeQuery = null
+  }
+
+  pollResults(): number | null {
+    if (!this.gl || !this.ext) return null
+
+    const disjoint = this.gl.getParameter(this.ext.GPU_DISJOINT_EXT) as boolean
+
+    let result: number | null = null
+    const remaining: WebGLQuery[] = []
+
+    for (const query of this.pending) {
+      const ready = this.gl.getQueryParameter(query, this.gl.QUERY_RESULT_AVAILABLE) as boolean
+
+      if (ready) {
+        if (!disjoint) {
+          const ns = this.gl.getQueryParameter(query, this.gl.QUERY_RESULT) as number
+          this._lastGpuTimeMs = ns / 1_000_000
+          result = this._lastGpuTimeMs
+        }
+        this.gl.deleteQuery(query)
+      } else {
+        remaining.push(query)
+      }
+    }
+
+    this.pending = remaining
+    return result
+  }
+
+  destroy(): void {
+    if (!this.gl) return
+
+    if (this.activeQuery) {
+      if (this.ext) {
+        this.gl.endQuery(this.ext.TIME_ELAPSED_EXT)
+      }
+      this.gl.deleteQuery(this.activeQuery)
+      this.activeQuery = null
+    }
+
+    for (const query of this.pending) {
+      this.gl.deleteQuery(query)
+    }
+    this.pending = []
+  }
+}
diff --git a/packages/core/src/profiler/hud-renderer.ts b/packages/core/src/profiler/hud-renderer.ts
new file mode 100644
index 000000000..355f3ea03
--- /dev/null
+++ b/packages/core/src/profiler/hud-renderer.ts
@@ -0,0 +1,242 @@
+import type { FrameStats } from './frame-stats'
+import type { CanvasKit, Canvas, Paint, Font, Typeface } from 'canvaskit-wasm'
+
+const BUFFER_SIZE = 120
+const LINE_HEIGHT = 13
+const PADDING = 6
+const BAR_WIDTH = 2
+const BAR_GAP = 0.5
+const GRAPH_HEIGHT = 40
+const MAX_SCALE_MS = 50
+const BUDGET_MS = 16.67
+const FAST_MS = 16.7
+const SLOW_MS = 33.3
+const CORNER_RADIUS = 4
+const RULER_SIZE = 20
+const SWATCH_SIZE = 6
+const SWATCH_GAP = 3
+const LEGEND_ITEM_GAP = 10
+const GRAPH_WIDTH = BUFFER_SIZE * (BAR_WIDTH + BAR_GAP)
+const COL_WIDTH = 130
+const PANEL_WIDTH = Math.max(COL_WIDTH * 2, GRAPH_WIDTH) + PADDING * 2
+
+export class HudRenderer {
+  private bgPaint: Paint
+  private textPaint: Paint
+  private dimTextPaint: Paint
+  private greenPaint: Paint
+  private yellowPaint: Paint
+  private redPaint: Paint
+  private gpuPaint: Paint
+  private budgetLinePaint: Paint
+  private graphBgPaint: Paint
+  private hudFont: Font
+
+  constructor(private ck: CanvasKit) {
+    this.bgPaint = new ck.Paint()
+    this.bgPaint.setStyle(ck.PaintStyle.Fill)
+    this.bgPaint.setColor(ck.Color4f(0.1, 0.1, 0.1, 0.85))
+    this.bgPaint.setAntiAlias(true)
+
+    this.textPaint = new ck.Paint()
+    this.textPaint.setStyle(ck.PaintStyle.Fill)
+    this.textPaint.setColor(ck.Color4f(0.9, 0.9, 0.9, 1))
+    this.textPaint.setAntiAlias(true)
+
+    this.dimTextPaint = new ck.Paint()
+    this.dimTextPaint.setStyle(ck.PaintStyle.Fill)
+    this.dimTextPaint.setColor(ck.Color4f(0.55, 0.55, 0.55, 1))
+    this.dimTextPaint.setAntiAlias(true)
+
+    this.greenPaint = new ck.Paint()
+    this.greenPaint.setStyle(ck.PaintStyle.Fill)
+    this.greenPaint.setColor(ck.Color4f(0.3, 0.85, 0.4, 1))
+    this.greenPaint.setAntiAlias(true)
+
+    this.yellowPaint = new ck.Paint()
+    this.yellowPaint.setStyle(ck.PaintStyle.Fill)
+    this.yellowPaint.setColor(ck.Color4f(1.0, 0.85, 0.2, 1))
+    this.yellowPaint.setAntiAlias(true)
+
+    this.redPaint = new ck.Paint()
+    this.redPaint.setStyle(ck.PaintStyle.Fill)
+    this.redPaint.setColor(ck.Color4f(1.0, 0.3, 0.3, 1))
+    this.redPaint.setAntiAlias(true)
+
+    this.gpuPaint = new ck.Paint()
+    this.gpuPaint.setStyle(ck.PaintStyle.Fill)
+    this.gpuPaint.setColor(ck.Color4f(0.4, 0.6, 1.0, 1))
+    this.gpuPaint.setAntiAlias(true)
+
+    this.budgetLinePaint = new ck.Paint()
+    this.budgetLinePaint.setStyle(ck.PaintStyle.Stroke)
+    this.budgetLinePaint.setStrokeWidth(1)
+    this.budgetLinePaint.setColor(ck.Color4f(1, 1, 1, 0.3))
+    this.budgetLinePaint.setAntiAlias(true)
+    this.budgetLinePaint.setPathEffect(ck.PathEffect.MakeDash([3, 3], 0))
+
+    this.graphBgPaint = new ck.Paint()
+    this.graphBgPaint.setStyle(ck.PaintStyle.Fill)
+    this.graphBgPaint.setColor(ck.Color4f(0.05, 0.05, 0.05, 0.5))
+    this.graphBgPaint.setAntiAlias(true)
+
+    this.hudFont = new ck.Font(null, 10)
+  }
+
+  setTypeface(typeface: Typeface): void {
+    this.hudFont.delete()
+    this.hudFont = new this.ck.Font(typeface, 10)
+  }
+
+  draw(canvas: Canvas, stats: FrameStats, phases: Map<string, number>, showRulers: boolean): void {
+    const rulerOffset = showRulers ? RULER_SIZE : 0
+    const hasGraph = stats.getBufferCount() > 0
+
+    const phaseNames = [
+      'render:scene',
+      'render:drawPicture',
+      'render:recordPicture',
+      'render:volatile',
+      'render:sectionTitles',
+      'render:componentLabels',
+      'render:selection',
+      'render:rulers',
+      'render:flush'
+    ]
+    const visiblePhases = phaseNames.filter((n) => (phases.get(n) ?? 0) > 0.01)
+
+    const statsRows = 3
+    const phaseRows = visiblePhases.length > 0 ? 1 + visiblePhases.length : 0
+    const statsHeight = (statsRows + phaseRows) * LINE_HEIGHT
+    const graphSection = hasGraph ? GRAPH_HEIGHT + PADDING + LINE_HEIGHT : 0
+    const contentHeight = statsHeight + PADDING * 2 + graphSection
+
+    const bgX = rulerOffset + PADDING
+    const bgY = rulerOffset + PADDING
+
+    const bgRect = this.ck.LTRBRect(bgX, bgY, bgX + PANEL_WIDTH, bgY + contentHeight)
+    canvas.drawRRect(this.ck.RRectXY(bgRect, CORNER_RADIUS, CORNER_RADIUS), this.bgPaint)
+
+    const col1 = bgX + PADDING
+    const col2 = col1 + COL_WIDTH
+    let y = bgY + PADDING + LINE_HEIGHT
+
+    const fps = Math.round(stats.smoothedFps)
+    const avgFrame = stats.avgFrameTime.toFixed(1)
+    const avgCpu = stats.avgCpuTime.toFixed(1)
+    const gpuAvailable = !Number.isNaN(stats.avgGpuTime) && stats.avgGpuTime > 0
+    const avgGpu = gpuAvailable ? stats.avgGpuTime.toFixed(1) : 'n/a'
+    const cacheStatus = stats.scenePictureCacheHit ? 'HIT' : 'MISS'
+
+    canvas.drawText(`FPS: ${fps} (${avgFrame}ms)`, col1, y, this.textPaint, this.hudFont)
+    canvas.drawText(`Nodes: ${stats.totalNodes} (${stats.culledNodes} culled)`, col2, y, this.textPaint, this.hudFont)
+    y += LINE_HEIGHT
+
+    canvas.drawText(`CPU: ${avgCpu}ms`, col1, y, this.textPaint, this.hudFont)
+    canvas.drawText(`Draws: ${stats.drawCalls}`, col2, y, this.textPaint, this.hudFont)
+    y += LINE_HEIGHT
+
+    canvas.drawText(`GPU: ${avgGpu}${gpuAvailable ? 'ms' : ''}`, col1, y, this.textPaint, this.hudFont)
+    canvas.drawText(`Cache: ${cacheStatus}`, col2, y, this.textPaint, this.hudFont)
+
+    if (visiblePhases.length > 0) {
+      y += LINE_HEIGHT
+      canvas.drawText('Phases:', col1, y, this.dimTextPaint, this.hudFont)
+      for (const name of visiblePhases) {
+        y += LINE_HEIGHT
+        const ms = (phases.get(name) ?? 0).toFixed(2)
+        const label = name.replace('render:', '')
+        canvas.drawText(`  ${label}: ${ms}ms`, col1, y, this.dimTextPaint, this.hudFont)
+      }
+    }
+
+    if (hasGraph) {
+      const graphX = bgX + PADDING
+      const graphY = y + PADDING
+      this.drawBarGraph(canvas, stats, graphX, graphY)
+      this.drawLegendRow(canvas, graphX, graphY + GRAPH_HEIGHT + LINE_HEIGHT - 2)
+    }
+  }
+
+  private drawLegendRow(canvas: Canvas, x: number, y: number): void {
+    const items: [Paint, string][] = [
+      [this.greenPaint, '60fps'],
+      [this.yellowPaint, '30fps'],
+      [this.redPaint, 'slow'],
+      [this.gpuPaint, 'GPU']
+    ]
+    let cx = x
+    for (const [paint, label] of items) {
+      const swatchY = y - SWATCH_SIZE + 1
+      canvas.drawRect(
+        this.ck.LTRBRect(cx, swatchY, cx + SWATCH_SIZE, swatchY + SWATCH_SIZE),
+        paint
+      )
+      cx += SWATCH_SIZE + SWATCH_GAP
+      canvas.drawText(label, cx, y, this.dimTextPaint, this.hudFont)
+      cx += label.length * 5.5 + LEGEND_ITEM_GAP
+    }
+  }
+
+  private drawBarGraph(canvas: Canvas, stats: FrameStats, graphX: number, graphY: number): void {
+    const graphRect = this.ck.LTRBRect(graphX, graphY, graphX + GRAPH_WIDTH, graphY + GRAPH_HEIGHT)
+    canvas.drawRRect(this.ck.RRectXY(graphRect, 2, 2), this.graphBgPaint)
+
+    const cpuHistory = stats.getCpuTimeHistory()
+    const gpuHistory = stats.getGpuTimeHistory()
+    const bufferCount = stats.getBufferCount()
+    const bufferIndex = stats.getBufferIndex()
+
+    const barOffset = BUFFER_SIZE - bufferCount
+
+    for (let i = 0; i < bufferCount; i++) {
+      const histIndex = (bufferIndex + i) % BUFFER_SIZE
+
+      const cpuTime = cpuHistory[histIndex]
+      if (cpuTime <= 0) continue
+
+      const barHeight = Math.min((cpuTime / MAX_SCALE_MS) * GRAPH_HEIGHT, GRAPH_HEIGHT)
+      const barX = graphX + (i + barOffset) * (BAR_WIDTH + BAR_GAP)
+      const barY = graphY + GRAPH_HEIGHT - barHeight
+
+      let paint: Paint
+      if (cpuTime < FAST_MS) {
+        paint = this.greenPaint
+      } else if (cpuTime < SLOW_MS) {
+        paint = this.yellowPaint
+      } else {
+        paint = this.redPaint
+      }
+
+      canvas.drawRect(
+        this.ck.LTRBRect(barX, barY, barX + BAR_WIDTH, graphY + GRAPH_HEIGHT),
+        paint
+      )
+
+      const gpuTime = gpuHistory[histIndex]
+      if (!Number.isNaN(gpuTime) && gpuTime > 0) {
+        const gpuBarHeight = Math.min((gpuTime / MAX_SCALE_MS) * GRAPH_HEIGHT, GRAPH_HEIGHT)
+        canvas.drawRect(
+          this.ck.LTRBRect(barX, graphY + GRAPH_HEIGHT - gpuBarHeight, barX + BAR_WIDTH, graphY + GRAPH_HEIGHT),
+          this.gpuPaint
+        )
+      }
+    }
+
+    const budgetY = graphY + GRAPH_HEIGHT - (BUDGET_MS / MAX_SCALE_MS) * GRAPH_HEIGHT
+    canvas.drawLine(graphX, budgetY, graphX + GRAPH_WIDTH, budgetY, this.budgetLinePaint)
+  }
+
+  destroy(): void {
+    this.bgPaint.delete()
+    this.textPaint.delete()
+    this.dimTextPaint.delete()
+    this.greenPaint.delete()
+    this.yellowPaint.delete()
+    this.redPaint.delete()
+    this.gpuPaint.delete()
+    this.budgetLinePaint.delete()
+    this.graphBgPaint.delete()
+    this.hudFont.delete()
+  }
+}
diff --git a/packages/core/src/profiler/index.ts b/packages/core/src/profiler/index.ts
new file mode 100644
index 000000000..f2eb6daa3
--- /dev/null
+++ b/packages/core/src/profiler/index.ts
@@ -0,0 +1,8 @@
+export { RenderProfiler } from './render-profiler'
+export { FrameStats } from './frame-stats'
+export { GPUTimer } from './gpu-timer'
+export { DrawCallCounter } from './draw-call-counter'
+export { PhaseTimer } from './phase-timer'
+export { HudRenderer } from './hud-renderer'
+export { CaptureStack, toSpeedscopeJSON } from './frame-capture'
+export type { NodeProfile, FrameCapture } from './frame-capture'
diff --git a/packages/core/src/profiler/phase-timer.ts b/packages/core/src/profiler/phase-timer.ts
new file mode 100644
index 000000000..d2cceb700
--- /dev/null
+++ b/packages/core/src/profiler/phase-timer.ts
@@ -0,0 +1,62 @@
+type DevToolsColor =
+  | 'primary'
+  | 'secondary'
+  | 'tertiary'
+  | 'tertiary-dark'
+  | 'secondary-dark'
+  | 'secondary-light'
+
+function colorForPhase(name: string): DevToolsColor {
+  if (name === 'frame') return 'primary'
+  if (name === 'render:flush') return 'tertiary-dark'
+  if (name === 'render:recordPicture') return 'tertiary'
+  if (name.startsWith('render:')) return 'secondary'
+  if (name.startsWith('layout:')) return 'secondary-dark'
+  return 'secondary-light'
+}
+
+const SMOOTH = 0.05
+
+export class PhaseTimer {
+  enabled = false
+  readonly averages = new Map<string, number>()
+
+  private starts = new Map<string, number>()
+
+  beginPhase(name: string): void {
+    if (!this.enabled || typeof performance === 'undefined') return
+    this.starts.set(name, performance.now())
+  }
+
+  endPhase(name: string): void {
+    if (!this.enabled || typeof performance === 'undefined') return
+
+    const startTime = this.starts.get(name)
+    if (startTime === undefined) return
+    this.starts.delete(name)
+
+    const duration = performance.now() - startTime
+    const prev = this.averages.get(name)
+    this.averages.set(
+      name,
+      prev === undefined ? duration : prev + (duration - prev) * SMOOTH
+    )
+
+    performance.measure(name, {
+      start: startTime,
+      detail: {
+        devtools: {
+          dataType: 'track-entry',
+          track: 'Renderer',
+          trackGroup: 'OpenPencil',
+          color: colorForPhase(name)
+        }
+      }
+    })
+  }
+
+  clearPhases(): void {
+    this.starts.clear()
+    this.averages.clear()
+  }
+}
diff --git a/packages/core/src/profiler/render-profiler.ts b/packages/core/src/profiler/render-profiler.ts
new file mode 100644
index 000000000..ea00b3ad3
--- /dev/null
+++ b/packages/core/src/profiler/render-profiler.ts
@@ -0,0 +1,163 @@
+import { FrameStats } from './frame-stats'
+import { GPUTimer } from './gpu-timer'
+import { DrawCallCounter } from './draw-call-counter'
+import { PhaseTimer } from './phase-timer'
+import { HudRenderer } from './hud-renderer'
+import { CaptureStack, toSpeedscopeJSON } from './frame-capture'
+
+import type { FrameCapture } from './frame-capture'
+import type { CanvasKit, Canvas, Typeface } from 'canvaskit-wasm'
+
+const now = typeof performance !== 'undefined' ? () => performance.now() : () => 0
+
+export class RenderProfiler {
+  enabled = false
+  hudVisible = false
+  capturing = false
+
+  readonly stats = new FrameStats()
+  readonly phases = new PhaseTimer()
+  readonly gpuTimer: GPUTimer
+  readonly drawCallCounter: DrawCallCounter
+
+  private hud: HudRenderer | null = null
+  private typeface: Typeface | null = null
+  private captureStack: CaptureStack | null = null
+  private captureFrameStart = 0
+  private lastCapture: FrameCapture | null = null
+  private renderStartTime = 0
+
+  constructor(
+    private ck: CanvasKit,
+    gl: WebGL2RenderingContext | null
+  ) {
+    this.gpuTimer = new GPUTimer(gl)
+    this.drawCallCounter = new DrawCallCounter(gl)
+  }
+
+  toggle(): void {
+    this.hudVisible = !this.hudVisible
+    this.enabled = this.hudVisible
+    this.phases.enabled = this.enabled
+  }
+
+  beginFrame(): void {
+    if (!this.enabled) return
+    this.renderStartTime = now()
+    this.phases.beginPhase('frame')
+    this.gpuTimer.beginFrame()
+    this.drawCallCounter.reset()
+  }
+
+  endFrame(): void {
+    if (!this.enabled) return
+
+    this.gpuTimer.endFrame()
+    this.gpuTimer.pollResults()
+
+    const cpuTime = now() - this.renderStartTime
+    this.stats.gpuTime = this.gpuTimer.lastGpuTimeMs
+    this.stats.drawCalls = this.drawCallCounter.count
+    this.stats.recordFrame(cpuTime)
+
+    this.phases.endPhase('frame')
+  }
+
+  beginPhase(name: string): void {
+    if (!this.enabled) return
+    this.phases.beginPhase(name)
+  }
+
+  endPhase(name: string): void {
+    if (!this.enabled) return
+    this.phases.endPhase(name)
+  }
+
+  setNodeCounts(total: number, culled: number): void {
+    this.stats.totalNodes = total
+    this.stats.culledNodes = culled
+  }
+
+  setCacheHit(hit: boolean): void {
+    this.stats.scenePictureCacheHit = hit
+  }
+
+  beginCapture(): void {
+    this.capturing = true
+    this.captureStack = new CaptureStack()
+    this.captureFrameStart = now()
+    this.captureStack.reset(this.captureFrameStart)
+  }
+
+  endCapture(): FrameCapture | null {
+    if (!this.capturing || !this.captureStack) return null
+    this.capturing = false
+
+    const capture: FrameCapture = {
+      timestamp: this.captureFrameStart,
+      totalTimeMs: now() - this.captureFrameStart,
+      cpuTimeMs: this.stats.cpuTime,
+      gpuTimeMs: this.gpuTimer.lastGpuTimeMs,
+      totalNodes: this.stats.totalNodes,
+      culledNodes: this.stats.culledNodes,
+      drawCalls: this.stats.drawCalls,
+      scenePictureCacheHit: this.stats.scenePictureCacheHit,
+      rootProfiles: this.captureStack.getRootProfiles()
+    }
+
+    this.lastCapture = capture
+    this.captureStack = null
+    return capture
+  }
+
+  beginNode(nodeId: string, name: string, type: string, culled: boolean): void {
+    this.captureStack?.begin(nodeId, name, type, culled)
+  }
+
+  endNode(drawCallsBefore: number): void {
+    this.captureStack?.end(this.drawCallCounter.count - drawCallsBefore)
+  }
+
+  getLastCapture(): FrameCapture | null {
+    return this.lastCapture
+  }
+
+  exportSpeedscope(): string | null {
+    if (!this.lastCapture) return null
+    return toSpeedscopeJSON(this.lastCapture)
+  }
+
+  downloadSpeedscope(): void {
+    const json = this.exportSpeedscope()
+    if (!json || typeof document === 'undefined') return
+    const blob = new Blob([json], { type: 'application/json' })
+    const url = URL.createObjectURL(blob)
+    const a = document.createElement('a')
+    a.href = url
+    a.download = `openpencil-frame-${Date.now()}.speedscope.json`
+    a.click()
+    URL.revokeObjectURL(url)
+  }
+
+  setTypeface(typeface: Typeface): void {
+    this.typeface = typeface
+    this.hud?.setTypeface(typeface)
+  }
+
+  drawHUD(canvas: Canvas, showRulers: boolean): void {
+    if (!this.hudVisible) return
+    if (!this.hud) {
+      this.hud = new HudRenderer(this.ck)
+      if (this.typeface) this.hud.setTypeface(this.typeface)
+    }
+    this.hud.draw(canvas, this.stats, this.phases.averages, showRulers)
+  }
+
+  destroy(): void {
+    this.gpuTimer.destroy()
+    this.drawCallCounter.destroy()
+    this.hud?.destroy()
+    this.hud = null
+    this.phases.clearPhases()
+  }
+}
diff --git a/packages/core/src/renderer.ts b/packages/core/src/renderer.ts
index 435f1be54..eeb19b5fb 100644
--- a/packages/core/src/renderer.ts
+++ b/packages/core/src/renderer.ts
@@ -60,6 +60,7 @@ import {
 } from './constants'
 import { isFontLoaded } from './fonts'
 import { vectorNetworkToPath } from './vector'
+import { RenderProfiler } from './profiler'
 
 import type { SceneNode, SceneGraph, Fill, Stroke } from './scene-graph'
 import type { SnapGuide } from './snap'
@@ -147,6 +148,7 @@ export class SkiaRenderer {
   private scenePictureVersion = -1
   private scenePicturePageId: string | null = null
   private nodePictureCache = new Map<string, SkPicture>()
+  readonly profiler: RenderProfiler
 
   private rulerBgPaint: Paint
   private rulerTickPaint: Paint
@@ -170,6 +172,8 @@ export class SkiaRenderer {
   pageId: string | null = null
 
   private worldViewport = { x: 0, y: 0, w: 0, h: 0 }
+  private _nodeCount = 0
+  private _culledCount = 0
 
   private color4f(r: number, g: number, b: number, a: number): Float32Array {
     const c = this._tmpColor
@@ -201,9 +205,10 @@ export class SkiaRenderer {
     return type === 'COMPONENT' || type === 'COMPONENT_SET' || type === 'INSTANCE'
   }
 
-  constructor(ck: CanvasKit, surface: Surface) {
+  constructor(ck: CanvasKit, surface: Surface, gl?: WebGL2RenderingContext | null) {
     this.ck = ck
     this.surface = surface
+    this.profiler = new RenderProfiler(ck, gl ?? null)
 
     this.fillPaint = new ck.Paint()
     this.fillPaint.setStyle(ck.PaintStyle.Fill)
@@ -321,6 +326,7 @@ export class SkiaRenderer {
         this.sizeFont = new this.ck.Font(typeface, SIZE_FONT_SIZE)
         this.sectionTitleFont = new this.ck.Font(typeface, SECTION_TITLE_FONT_SIZE)
         this.componentLabelFont = new this.ck.Font(typeface, COMPONENT_LABEL_FONT_SIZE)
+        this.profiler.setTypeface(typeface)
       }
       this.fontMgr = this.ck.FontMgr.FromData(fontData) ?? null
     }
@@ -491,6 +497,9 @@ export class SkiaRenderer {
     overlays: RenderOverlays = {},
     sceneVersion = -1
   ): void {
+    const p = this.profiler
+    p.beginFrame()
+
     graph.clearAbsPosCache()
 
     const canvas = this.surface.getCanvas()
@@ -505,7 +514,6 @@ export class SkiaRenderer {
     }
 
     const hasVolatileOverlays =
-      overlays.hoveredNodeId != null ||
       overlays.dropTargetId != null ||
       overlays.rotationPreview != null ||
       overlays.editingTextId != null
@@ -516,48 +524,80 @@ export class SkiaRenderer {
       sceneVersion === this.scenePictureVersion &&
       this.pageId === this.scenePicturePageId
 
+    p.setCacheHit(!!canUsePicture)
+
     // Scene layer (world coordinates)
     canvas.save()
     canvas.scale(this.dpr, this.dpr)
     canvas.translate(this.panX, this.panY)
     canvas.scale(this.zoom, this.zoom)
 
+    p.beginPhase('render:scene')
     if (canUsePicture) {
+      p.beginPhase('render:drawPicture')
       canvas.drawPicture(this.scenePicture!)
+      p.endPhase('render:drawPicture')
     } else if (hasVolatileOverlays) {
+      this._nodeCount = 0
+      this._culledCount = 0
+      p.beginPhase('render:volatile')
       const pageNode = graph.getNode(this.pageId ?? graph.rootId)
       if (pageNode) {
         for (const childId of pageNode.childIds) {
           this.renderNode(canvas, graph, childId, overlays, 0, 0)
         }
       }
+      p.endPhase('render:volatile')
     } else {
+      this._nodeCount = 0
+      this._culledCount = 0
+      p.beginPhase('render:recordPicture')
       this.recordScenePicture(canvas, graph, sceneVersion)
+      p.endPhase('render:recordPicture')
     }
+    p.endPhase('render:scene')
 
     canvas.restore()
 
     // Section titles + component labels (screen coordinates, zoom-independent)
     canvas.save()
     canvas.scale(this.dpr, this.dpr)
+    p.beginPhase('render:sectionTitles')
     this.drawSectionTitles(canvas, graph)
+    p.endPhase('render:sectionTitles')
+    p.beginPhase('render:componentLabels')
     this.drawComponentLabels(canvas, graph)
+    p.endPhase('render:componentLabels')
     canvas.restore()
 
     // UI overlay layer (screen coordinates, zoom-independent)
     canvas.save()
     canvas.scale(this.dpr, this.dpr)
 
+    this.drawHoverHighlight(canvas, graph, overlays.hoveredNodeId)
+    p.beginPhase('render:selection')
     this.drawSelection(canvas, graph, selectedIds, overlays)
+    p.endPhase('render:selection')
     this.drawSnapGuides(canvas, overlays.snapGuides)
     this.drawMarquee(canvas, overlays.marquee)
     this.drawLayoutInsertIndicator(canvas, overlays.layoutInsertIndicator)
     this.drawPenOverlay(canvas, overlays.penState)
     this.drawRemoteCursors(canvas, graph, overlays.remoteCursors)
+    p.beginPhase('render:rulers')
     if (this.showRulers) this.drawRulers(canvas, graph, selectedIds)
+    p.endPhase('render:rulers')
+
+    // Profiler HUD (drawn last, on top of everything)
+    p.drawHUD(canvas, this.showRulers)
 
     canvas.restore()
+
+    p.beginPhase('render:flush')
     this.surface.flush()
+    p.endPhase('render:flush')
+
+    p.setNodeCounts(this._nodeCount, this._culledCount)
+    p.endFrame()
   }
 
   private recordScenePicture(canvas: Canvas, graph: SceneGraph, sceneVersion: number): void {
@@ -583,6 +623,37 @@ export class SkiaRenderer {
 
   // --- Selection UI ---
 
+  private drawHoverHighlight(
+    canvas: Canvas,
+    graph: SceneGraph,
+    hoveredNodeId?: string | null
+  ): void {
+    if (!hoveredNodeId) return
+    const node = graph.getNode(hoveredNodeId)
+    if (!node) return
+
+    const abs = graph.getAbsolutePosition(node.id)
+    const sx = abs.x * this.zoom + this.panX
+    const sy = abs.y * this.zoom + this.panY
+
+    this.auxStroke.setStrokeWidth(1 / this.zoom)
+    this.auxStroke.setColor(
+      this.isComponentType(node.type) ? this.compColor() : this.selColor()
+    )
+    this.auxStroke.setPathEffect(null)
+
+    canvas.save()
+    canvas.translate(sx, sy)
+    if (node.rotation !== 0) {
+      const cx = (node.width / 2) * this.zoom
+      const cy = (node.height / 2) * this.zoom
+      canvas.rotate(node.rotation, cx, cy)
+    }
+    canvas.scale(this.zoom, this.zoom)
+    this.strokeNodeShape(canvas, node, this.auxStroke)
+    canvas.restore()
+  }
+
   private drawSelection(
     canvas: Canvas,
     graph: SceneGraph,
@@ -969,6 +1040,8 @@ export class SkiaRenderer {
     const node = graph.getNode(nodeId)
     if (!node || !node.visible) return
 
+    this._nodeCount++
+
     const absX = parentAbsX + node.x
     const absY = parentAbsY + node.y
 
@@ -994,9 +1067,11 @@ export class SkiaRenderer {
           cx + diag / 2 < vp.x ||
           cy + diag / 2 < vp.y
         ) {
+          this._culledCount++
           return
         }
       } else if (absX > vp.x + vp.w || absY > vp.y + vp.h || absX + bw < vp.x || absY + bh < vp.y) {
+        this._culledCount++
         return
       }
     }
@@ -1049,14 +1124,6 @@ export class SkiaRenderer {
       canvas.drawRect(this.ck.LTRBRect(0, 0, node.width, node.height), this.auxStroke)
     }
 
-    // Hover highlight — shape-aware outline
-    if (overlays.hoveredNodeId === nodeId) {
-      this.auxStroke.setStrokeWidth(1 / this.zoom)
-      this.auxStroke.setColor(this.isComponentType(node.type) ? this.compColor() : this.selColor())
-      this.auxStroke.setPathEffect(null)
-      this.strokeNodeShape(canvas, node, this.auxStroke)
-    }
-
     // Clip + render children for containers
     const isClippableContainer =
       node.type === 'FRAME' || node.type === 'COMPONENT' || node.type === 'INSTANCE'
@@ -2942,6 +3009,7 @@ export class SkiaRenderer {
     for (const pic of this.nodePictureCache.values()) pic?.delete()
     this.nodePictureCache.clear()
     this.scenePicture?.delete()
+    this.profiler.destroy()
     this.surface.delete()
   }
 }
diff --git a/src/components/AppMenu.vue b/src/components/AppMenu.vue
index 67d71f752..736c816dc 100644
--- a/src/components/AppMenu.vue
+++ b/src/components/AppMenu.vue
@@ -111,6 +111,16 @@ const viewMenu: MenuItem[] = [
     label: 'Zoom out',
     shortcut: `${mod}-`,
     action: () => store.applyZoom(100, window.innerWidth / 2, window.innerHeight / 2)
+  },
+  { separator: true },
+  {
+    label: 'Performance profiler',
+    get checked() {
+      return store.renderer?.profiler.hudVisible ?? false
+    },
+    onCheckedChange: () => {
+      store.toggleProfiler()
+    }
   }
 ]
 
diff --git a/src/composables/use-canvas.ts b/src/composables/use-canvas.ts
index 56f1a68eb..b7bec0542 100644
--- a/src/composables/use-canvas.ts
+++ b/src/composables/use-canvas.ts
@@ -118,7 +118,8 @@ export function useCanvas(canvasRef: Ref<HTMLCanvasElement | null>, store: Edito
       }
     }
 
-    renderer = new SkiaRenderer(ck, surface)
+    const glCtx = (canvas.getContext('webgl2') ?? null) as WebGL2RenderingContext | null
+    renderer = new SkiaRenderer(ck, surface, glCtx)
     store.setCanvasKit(ck, renderer)
     renderer.loadFonts().then(() => renderNow())
     renderNow()
diff --git a/src/stores/editor.ts b/src/stores/editor.ts
index 14540c5c2..c22a28ae5 100644
--- a/src/stores/editor.ts
+++ b/src/stores/editor.ts
@@ -1477,6 +1477,11 @@ export function createEditorStore() {
     requestRender()
   }
 
+  function toggleProfiler() {
+    _renderer?.profiler.toggle()
+    requestRepaint()
+  }
+
   function toggleVisibility() {
     for (const id of state.selectedIds) {
       const node = graph.getNode(id)
@@ -2059,6 +2064,7 @@ export function createEditorStore() {
     goToMainComponent,
     bringToFront,
     sendToBack,
+    toggleProfiler,
     toggleVisibility,
     toggleLock,
     moveToPage,
diff --git a/tests/engine/profiler.test.ts b/tests/engine/profiler.test.ts
new file mode 100644
index 000000000..3c8e54696
--- /dev/null
+++ b/tests/engine/profiler.test.ts
@@ -0,0 +1,153 @@
+import { describe, it, expect } from 'bun:test'
+import { FrameStats } from '../../packages/core/src/profiler/frame-stats'
+import { DrawCallCounter } from '../../packages/core/src/profiler/draw-call-counter'
+import { PhaseTimer } from '../../packages/core/src/profiler/phase-timer'
+import { GPUTimer } from '../../packages/core/src/profiler/gpu-timer'
+import { CaptureStack, toSpeedscopeJSON } from '../../packages/core/src/profiler/frame-capture'
+
+describe('FrameStats', () => {
+  it('records frames and computes rolling averages', () => {
+    const stats = new FrameStats()
+
+    for (let i = 0; i < 10; i++) {
+      stats.recordFrame(5)
+    }
+
+    expect(stats.avgCpuTime).toBe(5)
+    expect(stats.smoothedFps).toBeGreaterThan(0)
+    expect(stats.getFrameTimeHistory()).toBeInstanceOf(Float64Array)
+    expect(stats.getBufferCount()).toBe(10)
+  })
+
+  it('tracks external fields', () => {
+    const stats = new FrameStats()
+    stats.totalNodes = 100
+    stats.culledNodes = 20
+    stats.drawCalls = 50
+    stats.scenePictureCacheHit = true
+
+    expect(stats.totalNodes).toBe(100)
+    expect(stats.culledNodes).toBe(20)
+    expect(stats.drawCalls).toBe(50)
+    expect(stats.scenePictureCacheHit).toBe(true)
+  })
+
+  it('handles GPU time as NaN initially', () => {
+    const stats = new FrameStats()
+    stats.recordFrame(1)
+    expect(stats.gpuTime).toBe(0)
+  })
+})
+
+describe('DrawCallCounter', () => {
+  it('handles null GL context', () => {
+    const counter = new DrawCallCounter(null)
+    expect(counter.count).toBe(0)
+    expect(counter.reset()).toBe(0)
+    counter.destroy()
+  })
+})
+
+describe('GPUTimer', () => {
+  it('handles null GL context', () => {
+    const timer = new GPUTimer(null)
+    expect(timer.available).toBe(false)
+    expect(timer.lastGpuTimeMs).toBeNaN()
+    timer.beginFrame()
+    timer.endFrame()
+    expect(timer.pollResults()).toBeNull()
+    timer.destroy()
+  })
+})
+
+describe('PhaseTimer', () => {
+  it('is disabled by default', () => {
+    const timer = new PhaseTimer()
+    expect(timer.enabled).toBe(false)
+  })
+
+  it('records measures when enabled', () => {
+    const timer = new PhaseTimer()
+    timer.enabled = true
+    timer.beginPhase('test')
+    timer.endPhase('test')
+    timer.clearPhases()
+  })
+
+  it('is a no-op when disabled', () => {
+    const timer = new PhaseTimer()
+    timer.beginPhase('test')
+    timer.endPhase('test')
+  })
+})
+
+describe('CaptureStack', () => {
+  it('builds a tree of node profiles', () => {
+    const stack = new CaptureStack()
+    stack.reset(performance.now())
+
+    stack.begin('node-1', 'Frame 1', 'FRAME', false)
+    stack.begin('node-2', 'Rect', 'RECTANGLE', false)
+    stack.end(2)
+    stack.end(3)
+
+    const roots = stack.getRootProfiles()
+    expect(roots).toHaveLength(1)
+    expect(roots[0].nodeId).toBe('node-1')
+    expect(roots[0].children).toHaveLength(1)
+    expect(roots[0].children[0].nodeId).toBe('node-2')
+    expect(roots[0].children[0].drawCalls).toBe(2)
+  })
+})
+
+describe('toSpeedscopeJSON', () => {
+  it('produces valid speedscope JSON', () => {
+    const capture = {
+      timestamp: 0,
+      totalTimeMs: 10,
+      cpuTimeMs: 8,
+      gpuTimeMs: 6,
+      totalNodes: 5,
+      culledNodes: 1,
+      drawCalls: 10,
+      scenePictureCacheHit: false,
+      rootProfiles: [
+        {
+          nodeId: 'n1',
+          name: 'Frame',
+          type: 'FRAME',
+          depth: 0,
+          startTime: 0,
+          endTime: 10,
+          selfTime: 5,
+          drawCalls: 3,
+          culled: false,
+          children: [
+            {
+              nodeId: 'n2',
+              name: 'Rect',
+              type: 'RECTANGLE',
+              depth: 1,
+              startTime: 2,
+              endTime: 7,
+              selfTime: 5,
+              drawCalls: 2,
+              culled: false,
+              children: []
+            }
+          ]
+        }
+      ]
+    }
+
+    const json = toSpeedscopeJSON(capture)
+    const parsed = JSON.parse(json)
+
+    expect(parsed.$schema).toBe('https://www.speedscope.app/file-format-schema.json')
+    expect(parsed.profiles).toHaveLength(1)
+    expect(parsed.profiles[0].type).toBe('evented')
+    expect(parsed.profiles[0].unit).toBe('milliseconds')
+    expect(parsed.shared.frames).toHaveLength(2)
+    expect(parsed.profiles[0].events).toHaveLength(4)
+  })
+})