Files
hyperframes/packages
Vance IngallsandClaude Opus 4.7 d20e28a623 fix(engine): gate software-GPU warn + thread resolved mode (staff review)
Three fixes from PR #822 self-review:

1. Gate the "Software GPU detected" console.warn so it only fires when the
   guard actually changed the outcome. Previously the warn condition
   (captureMode==="screenshot" && isSoftwareRenderer && headlessShell &&
   isLinux) was satisfied even when the caller had explicitly set
   forceScreenshot=true — misleading operators into thinking the guard
   kicked in when they had already picked screenshot mode. Add
   `!forceScreenshot` to the condition and one-shot the warn with a
   process-level latch to avoid spam across multi-worker renders.

2. Thread the already-resolved browserGpuMode from frameCapture.ts into
   acquireBrowser via a new AcquireBrowserOptions.resolvedBrowserGpuMode.
   The Promise cache made the duplicate resolution cheap, but threading
   the value removes the smell of two parallel resolutions of the same
   thing with no static guarantee they agree under future refactors.

3. Add fallback-launch-retry test: simulate probeBeginFrameSupport
   failure via a CDP mock whose `send` rejects, assert the second launch
   strips beginframe-only flags. Also add positive-control tests for
   "Linux + hardware + headless-shell does NOT warn" and "forceScreenshot
   suppresses the software-GPU warn".

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-14 04:33:29 +00:00
..
2026-05-13 17:01:15 -07:00
2026-05-13 17:01:15 -07:00
2026-05-13 17:01:15 -07:00