agents/skills/crossbench/SKILL.md
Crossbench is the central chrome runner harness for press benchmarks like JetStream, Speedometer or Motionmark, as well as for loading live pages.
~/crossbench or in
CHROMIUM_CHECKOUT/src/tools/perf/cb, but could be anywhere.
If not found in standard locations, ask the user for the path.Navigate to the crossbench directory and use ./cb.py directly (or via poetry
if vpython3 is not available).
# Run the latest stable JetStream with a specific d8 binary
./cb.py jetstream --browser=/path/to/d8 --env-validation=warn
And for chrome:
./cb.py jetstream --browser=out/x64.release/chrome --env-validation=warn
--separate for instance to
generate individual profiles for each sub test--story=XXX to only run a specific testUse ./cb.py mcp to launch a local MCP to interact with benchmarks.
./cb.py describe probes to list all probes../cb.py describe probe v8.log to show detailed help for a specific probe.Important probes for v8 investigation:
--probe='v8.log' to generate a v8.log file (see v8-log skill for
more information on how to analyse the log file)--probe='profiling' to generate a Linux perf trace.--probe=perfetto to collected perfetto tracesResults:
cb.results.json file listing all probe results--env-validation=warn to bypass environment input prompts../cb.py describe to understand subcommands and probes.poetry run cb_validate_hjson -- file.hjson.