[Hona]

Every benchmark in the repo now runs with strobe. The app's Playwright harness, core's six benchmark-* scripts, the UI motion script, the desktop startup script and the dev-startup suite are gone: −13,712 / +3,396 lines. 57 cases in 14 *.bench.ts files cover their workloads. Each case compares two builds in pairs at normal speed, decides with an interval, refuses setups it cannot defend, and explains a move from one profiled run per side.

Before → after

flowchart LR
  subgraph before["before: one harness per package"]
    A[app: Playwright fixture, BENCHMARK JSON,<br/>11 rAF probes, 12 percentile copies,<br/>7 CDP memory probes, 3 Vite fixture pages]
    C[core: 6 benchmark-* scripts,<br/>own timers, medians and JSON]
    U[ui: motion-bench.ts]
    D[desktop: bench-startup.ts log scraper,<br/>e2e/devex dev-start timer]
  end
  subgraph after["after: *.bench.ts beside the code"]
    AB[app/perf: 9 files] --> S
    CB[core/perf: location, snapshot, session] --> S
    UB[ui/perf: motion] --> S
    DB[desktop/perf: startup] --> S
    S[strobe: isolate, pair A/B, trace,<br/>decide, attribute, report]
  end
  before -. workloads ported .-> after

Loading

File	Cases	Replaces
app/perf/timeline.bench.ts	scroll history · type in composer · find in session · expand an edit	search benchmark, patch-groups page
app/perf/stream.bench.ts	stream a reply · into a tool-heavy history · with review open	streaming benchmark (3 variants), text projection
app/perf/tabs.bench.ts	switch tab (warm, review open, after a resize, 1,200 models) · open an unvisited tab (± review)	tab switch (cold/warm/review/resized), provider memory, laggy session
app/perf/navigation.bench.ts	Home → session · close the last tab · new session ×2 · new session page in/out · child session · restore 8 tabs · activate a restored tab · palette ×2	first navigation, Home tab navigation, session entry, inactive-tab prefetch, command palette
app/perf/home.bench.ts	load home · update a session in the background, at 500 / 5,000 / 10,000 sessions	Home session index
app/perf/review.bench.ts	open review · filter files · next file, at 100 / 10,000 files	review pane scaling, review pane
app/perf/diff.bench.ts	open a long diff · a huge diff · return to a long diff	session-ui highlighting page
app/perf/composer.bench.ts	type in a large draft · move the cursor · recall from history	composer write batch, composer-history page
app/perf/desktop.bench.ts	idle	—
desktop/perf/startup.bench.ts	launch to Home · to a restored tab · with the service stopped · first launch · dev start to Home	script/bench-startup.ts, e2e/devex
core/perf/location.bench.ts	load a location, cached and dropped · keep a location graph · keep a models.dev plugin instance	benchmark-location.ts, benchmark-location-memory.ts
core/perf/snapshot.bench.ts	capture ×4 · 8 captures at once · list and diff 100 changed files · restore 100 files · agent step loop · pack the store	benchmark-snapshot.ts, -pack.ts, -worker.ts
core/perf/session.bench.ts	an agent session of 15 steps, snapshots on and off	benchmark-snapshot-session.ts
ui/perf/motion.bench.ts	1,000 springs for 60 frames	script/motion-bench.ts

Three kinds of case

// A page: the production build on the e2e mock server. A cached transcript is not fetched again, and the
// switch moves nothing that was already on screen.
const warm: Case<{}> = {
  primary: "main.work",
  metrics: ["run.duration", "interaction.p50", "net.requests{/api/session/*/message}", "layout.shift"],
  run: ({ page, prepared }) => page.click(tab(prepared.target), { until: prepared.targetReady }),
  async reset({ page, prepared }) {
    await page.click(tab(prepared.source), { until: prepared.sourceReady })
    await page.idle()
  },
}

// A launch: timed from process spawn to a selector, with the main process's own epoch marks as milestones.
"launch to a restored tab": {
  primary: "launch.ready",
  launch: { until: composer, milestones: packaged, profile: "warm", before: warmService },
}

// A script: a fresh Bun process per block loads the package's harness and calls its exports.
// In a git ref's worktree strobe places this checkout's harness, so its imports reach that build's code.
"capture after adding a file": { ...timed, run: (ctx) => git(ctx, "capture"), reset: (ctx) => ctx.call("add") }

• until ends on what a person sees: strobe ignores visibility: hidden, and >>> reaches into the diff viewer's shadow root.
• Page mocks go through strobe's mock (fixtures/mock.ts), so the browser's HTTP cache stays on and a case that reloads does not refetch every script. The e2e mock server runs unchanged.
• Streaming cases get events from a real SSE endpoint (fixtures/events.ts) that never waits for the renderer.
• Nothing is CPU- or network-throttled. Cases that should pay a first visit reload in reset.

Counted, compared exactly

Count	Case	Expected
net.requests{/api/session/*/message}	switch tab (warm)	0: a cached transcript stays
net.requests{/api/session/*/message}	restore 8 tabs	1: only the active tab's
count.draft.encode	type in a large draft	about 25 for 51 keystrokes: batched
count.draft.encode	move the cursor in a large draft	0: a caret move saves no text
layout.shift	switch tab, open an unvisited tab	0: nothing on screen jumps
count.git	every snapshot case	the processes the operation spawned

Checked against merged PRs

PR	Case	Base (dev before)	Head (PR)	Change [interval]	Verdict
#60513	scroll history, main.busy	15.3%	14.2%	−7.3% [−9.9 … −4.2]	less, 8 pairs
#57016	stream a reply, main.work	489 ms	462 ms	−6.0% [−11.7 … −0.1]	less (small), 8 pairs; main.script −15%
#60755	scroll history, main.busy	13.9%	13.5%	−3.5% [−5.6 … −1.1]	less (small), 6 pairs
#60755	type in composer, interaction.p95	12.8 ms	12.8 ms	−0.4% [−2.0 … +1.0]	same (±2%)
A/A	switch tab, main.work	27.9 ms	27.6 ms	−1.0% [−1.9 … +0.7]	same (±2%), 19 pairs

#59419 ("coalesce snapshot captures and skip repeat ignore work"), on core

strobe bench packages/core/perf/snapshot.bench.ts --base cdaba26fa5~1 --head cdaba26fa5, on the generated 20,000-file repo:

Case	Base (dev before)	Head (PR)	Change [interval]	Verdict	count.git (exact)
8 captures at once, nothing changed	649 ms	103 ms	−84% [−84 … −84]	faster, 8 pairs	24 → 4
capture, nothing changed	84.3 ms	53.1 ms	−36% [−39 … −35]	faster, 8 pairs	3 → 2
agent step loop (20 steps)	4,187 ms	2,938 ms	−28% [−45 … −26]	faster, 8 pairs	136 → 100
capture after editing a file / adding one / editing 100	184 / 178 / 722 ms	184 / 179 / 728 ms	within ±2%	same	same
list / diff / restore 100 files	73.7 / 97.6 / 130 ms	73.9 / 97.5 / 129 ms	within ±2%	same	same
pack the store	383 ms	387 ms	+0.8% [−0.9 … +2.4]	underpowered (~43 pairs)	same

• The PR's own A/B promised that N concurrent callers cost at most 2 captures: 8 captures start 4 git processes after it, 24 before.
• Primary intervals are 98.4% at the planned looks (8, 12, 18, 26, 40 pairs), 95% overall.

Why, for #60513 ("compute highlight styles once per subtree in Chromium")

Without reading the diff, strobe pins the win on style recalculation. Timed pairs, re-run with 8 pairs:

Per scroll run	Base	Head	Change [95% interval]
main.busy (primary)	16.9%	15.6%	−8.3% [−14 … −2.0], less
main.style	110 ms	66.7 ms	−38% [−45 … −29]
main.script	86.2 ms	89.7 ms	−2.1% [−8.1 … +6.4]
main.layout	14.1 ms	14.4 ms	−1.3% [−9.5 … +9.0]

The profiled launches say where: the same ~500 forced style recalcs per run, each cheaper.

strobe query <rec> layout          base                       head
  forced style per run             495 (115 ms)               498 (84.1 ms)
  measureScrollState               425 ms                     294 ms
    packages/app/src/session/timeline/interaction.ts:122:25
    ← (anonymous) interaction.ts:140:26 ← tick packages/ui/src/motion/loop.ts:72:20
  handleListScroll                 140 ms                     132 ms
    packages/app/src/session/timeline/virtualizer.tsx:900:22 ← onScroll scroll-view.tsx:488:61

Removed

• App: e2e/performance, session-ui/performance and their fixture pages; the dev debug bar's HUD (FPS, jank, long tasks, INP, CLS, heap, nav time; bar.tsx −480, 63 locales −1,350), which strobe record app:opencode measures from presented frames instead of rAF gaps. The bar keeps provider stats, the dev toggles, memory diagnostics and the Knobs.
• Core: script/benchmark-{location,location-memory,snapshot,snapshot-pack,snapshot-session,snapshot-worker}.ts.
• UI: script/motion-bench.ts. Desktop: script/bench-startup.ts. App: e2e/devex.
• Their bench:* package scripts, a search event only the benchmark dispatched (search-controller.ts), messageDelay in the mock server, a profile script that imported from another machine's worktree, and a stale test-speed log.

Kept elsewhere

Old check	Owner now
Layout continuity under CPU stress	e2e/stability (bun run test:stability), a correctness suite
Theme preload before first paint (devex)	test-browser/theme-preload.test.ts, electron.vite.config.test.ts
Snapshot robustness probes	core/test/snapshot.test.ts
Pack correctness	core/test/git.test.ts

Setup

• @hona/strobe (catalog, 0.3.0) is a devDependency of the root, app, core, ui and desktop, for the bench files' types: typecheck and typecheck:e2e check every *.bench.ts. strobe runs its own embedded copy at run time. The package is also strobe's CLI launcher: bunx strobe … from the repository root runs the pinned build, one optional dependency per platform: Windows x64, macOS arm64 and x64, Linux x64 and arm64.
• bunfig.toml exempts @hona/strobe and its five platform packages from the 3-day minimum release age, as it does other first-party packages.
• Script cases (core, ui) run on every OS; on Linux and macOS they were checked from the npm package (bunx strobe bench … --check: snapshot 10/10, location 4/4, session 2/2, motion 1/1 on Linux; snapshot 10/10, motion 1/1 on macOS). Page and desktop cases measure the most on Windows; packages/app/perf/README.md lists what each OS gives.
• The snapshot fixtures are generated once: bun packages/core/perf/snapshot.ts medium opencode.

Strobe gaps found here, all built

Issues	Feature
#1, #2	Launch cases: spawn to a selector with milestones, through a command
#9	Script cases for Bun and Node code
#10, #11	Exact counters (requests, app counters), layout.shift
#3, #4	--check, published types
#5–#8	Default parameter subsets, hover and focus-aware waits, mocks that keep the HTTP cache, teardown
#12–#15	Case name in prepare, CSS visibility, shadow DOM, main.work across a page load

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[opencodejr]

The port looks clean: type checks pass for the bench files in app (e2e tsconfig), core, ui, desktop and gui-plugins; the stability and playwright-config/theme-preload tests pass; and the core snapshot harness runs under plain Bun with the git process counts the description reports (2 for a no-change capture, 4 for 8 at once). I could not run any *.bench.ts end to end, because strobe 0.2.0 only ships a Windows x64 binary (strobe on Linux: "no build for linux-x64 is installed").

Public API: unchanged. Only dev-only package scripts (bench:*, test:bench*, root bench:devex), internal e2e helpers and the dev-only debug bar performance readouts are removed.

1. macOS and Linux lose every benchmark. The removed tools (Playwright perf suite, core benchmark-*, motion-bench.ts, bench-startup.ts) ran anywhere Bun or Playwright ran. Now none of the 57 cases runs off Windows, including the pure-Bun core and ui cases. Neither the perf README nor the desktop README says so, and strobe.json lists a macOS exe path that can't be used yet. Please either keep a cross-platform way to run the script cases until strobe ships those builds, or state clearly that benchmarks are Windows-only for now (see inline).
2. Stale skill line. .opencode/skills/app-e2e-tests/SKILL.md:57 still says messageDelay exists for benchmarks (mock-server.ts:82), but this PR removes it. Agents following the skill will look for an option that no longer exists.

[opencodejr]

strobe 0.2.0 only ships @hona/strobe-win32-x64, so every command below fails on macOS and Linux (no build for linux-x64 is installed), including the pure-Bun script cases in core and ui that used to run anywhere as bun script/benchmark-*.ts. Please say here, and in packages/desktop/README.md, that benchmarks need Windows x64 for now, or keep a way to run the script harnesses (snapshot.ts, location.ts, motion.ts) without strobe until it ships other platforms.

[Hona]

Resolved by strobe 0.3.0, which this PR now pins. 0.3.0 ships macOS (arm64, x64) and Linux (x64, arm64) builds beside Windows x64 (@hona/strobe-darwin-arm64, -darwin-x64, -linux-x64, -linux-arm64). The root devDependencies list @hona/strobe too, so bunx strobe from the repository root runs the pinned build on every OS.

The script cases ran from the npm package on Linux and macOS (bun install --frozen-lockfile, then bunx strobe, no local strobe build):

Linux x64 (Debian 13 VM, 2 vCPU), e67bc51370:

$ bun packages/core/perf/snapshot.ts medium
$ bunx strobe bench packages/core/perf/snapshot.bench.ts --head . --check
10 of 10 cases passed
$ bunx strobe bench packages/core/perf/location.bench.ts --head . --check
4 of 4 cases passed
$ bunx strobe bench packages/core/perf/session.bench.ts --head . --check --params fixture=medium
2 of 2 cases passed
$ bunx strobe bench packages/ui/perf/motion.bench.ts --head . --check
1 of 1 cases passed

macOS arm64 (M4 MacBook Air, macOS 27), e67bc51370, fixture in the default temp folder:

$ bun packages/core/perf/snapshot.ts small
small: /private/var/folders/…/T/opencode-snapshot-bench/fixtures/small (2 s)
$ bunx strobe bench packages/core/perf/snapshot.bench.ts --head . --check --params fixture=small
10 of 10 cases passed
$ bunx strobe bench packages/ui/perf/motion.bench.ts --head . --check
1 of 1 cases passed

The macOS run found a bug in this PR, now fixed in a3719bc480. macOS's temp folder is a symlink (/var/folders is /private/var/folders). Snapshot compares a location with its worktree's real path, so a fixture generated under that symlink read as outside its own project, and every capture returned undefined. The harness now resolves the fixture root's real path.

Review point 2 (the stale messageDelay line in the e2e skill) was fixed in e825d74cc5.

What still differs by OS is now in packages/app/perf/README.md and packages/desktop/README.md. Script cases run the same everywhere. Page and desktop cases measure the most on Windows:

• On macOS and Linux, frames come from Chromium's trace only. OS present events (--os-frames) are Windows only.
• A software GPU is refused, so a VM or a server without a GPU cannot run page cases.
• macOS runs are not pinned to CPUs (macOS has no affinity API).
• On macOS, window pixels need the Screen Recording permission.
• On macOS, a desktop case launches only a build whose bundle id this account has never run. An isolated launch cannot move a Cocoa app's ~/Library, so strobe refuses rather than write into the user's own.
• Linux finds windows through X11.

[opencodejr]

Agreed, this is resolved at pr/head (e67bc5137).

• 0.3.0 is pinned: the root catalog in package.json pins @hona/strobe to 0.3.0. In bun.lock, 0.3.0 lists optional builds for darwin-arm64, darwin-x64, linux-arm64, linux-x64 and win32-x64.

• It runs on Linux x64: I ran it here in a separate worktree of pr/head (a Cloud Hypervisor VM). bun install --frozen-lockfile installed @hona/strobe-linux-x64@0.3.0. Then from the repository root:

  • bunx strobe bench packages/ui/perf/motion.bench.ts --head . --check passed 1 of 1.
  • bunx strobe bench packages/core/perf/location.bench.ts --head . --check passed 4 of 4.
  • strobe only warned that it was running in a VM, and the check passed anyway.

  I didn't run snapshot or session here, and I didn't test on macOS.

• The macOS fix looks right: packages/core/perf/snapshot.ts:28 now builds the fixture root with realpathSync(base). location.ts:30 already used fs.realpath, so both harnesses get the real path for /var/folders → /private/var/folders.

• The docs cover the OS limits: packages/app/perf/README.md:77-86 lists them: Windows-only --os-frames, a software GPU refused, no CPU pinning on macOS, Screen Recording permission, X11 on Linux. packages/desktop/README.md:73 covers the macOS bundle-id limit.

I didn't change anything in repo/. The shared checkout is a detached HEAD at e825d74cc, which is behind pr/head, and it has a bun.lock edit that isn't mine. I left both alone. The temporary worktree I used is removed.

[opencodejr]

The new commit fixes the stale messageDelay reference in the e2e skill I raised last time; its new timeline.ts:326 pointer is correct. The Windows-only point (thread on packages/app/perf/README.md) is still open. Public API unchanged. Nothing new to raise.

[opencodejr]

Benchmarks now run off Windows: on Linux, with strobe 0.3.0, the ui motion bench (1/1) and core snapshot bench (10/10) pass with --check, and the READMEs say what each OS supports. That covers my earlier point. The snapshot harness's real-path fix is correct. Public API unchanged. Nothing new to raise.

[opencodejr]

This push only merges dev, and none of the PR's files change. The core snapshot bench still passes on Linux after #61546 (--check, 10/10). Thread 1 (Windows-only benchmarks) was already addressed by the strobe 0.3.0 bump and the per-OS notes. Public API unchanged. Nothing new to raise.