Skip to content

Performance fix verification, 2026-10-18

Choose a tag to compare

@luke2x luke2x released this 18 Oct 19:55
· 992 commits to perf/subagent-bounded-preview since this release

Performance helper evidence (2026-10-18)

These are synthetic CPU benchmarks of the production helpers, not an application FPS, native Hermes, or heap benchmark. The script extracts the production helper source from the supplied Git refs and strips TypeScript. It does not execute caller components or parsing libraries outside the isolated helper.

Environment: {'node': 'v26.10.0', 'platform': 'darwin', 'arch': 'arm64', 'cpu': 'Apple M5'}

Method: five warmup rounds; 30 alternating before/after samples. GC runs outside timed sections when available. Input objects/strings are reused. Each large-preview sample batches 20 calls; short-preview samples batch 2,000. Output equivalence is asserted before timing. Raw per-call timings are in results.json.

Run from a repository checkout containing the refs (fetch the corresponding luke2x/t3code branches first):

node --expose-gc performance-bench.mjs d9eb2b6 98cc2b4 374b83c 59090b8

tree: median 3.652625 ms before; 0.333959 ms after.
preview: median 0.990725 ms before; 0.003665 ms after.
previewShort: median 0.000103 ms before; 0.000123 ms after.
html: median 64.355666 ms before; 2.788375 ms after.

Workloads:

  • Tree: 30,000 files under 21 expanded ancestors; measures flattening only, not tree construction or native rendering. A separate 200,000-file directory case throws RangeError before and returns 200,001 rows afterward. That wide case is a stress test, not a claimed field report.
  • Preview: a 262,144-character completed result, normalized for the existing 280-character display limit; plus a short, common-case progress message.
  • HTML: 3,000 inline code ranges with 6,000 HTML image tags (literal tags inside code and converted tags outside). Measures markdownHtmlContent, not the full markdown parser or layout.

The independent mobile visibility PR has React/jsdom behavior tests for text-only rows, sharing one observer across multiple videos, cleanup, and explicit visibility outside a feed. Its LegendList viewability boundary is mocked. No native scrolling recording, native frame-time measurement, or Hermes timing was collected because this host has no connected device, Android SDK/emulator, or full Xcode simulator tooling. It is a subscription-lifetime fix; these helper benchmarks are not evidence of its frame-rate impact.