Why it mattersA reviewer sees CPU time, memory and frame rate change on each pull request. The numbers come with the method. They are rejected when the machine cannot measure them well.
A team argues about whether a branch made the page faster. One person ran it on their laptop. Another ran it on a build box. Nobody wrote down which was warm or which had other apps running. The numbers never match.
Strobe is a new command-line tool that tries to fix this. It was first pushed to GitHub on 31 October 2026. The author is a developer called Hona. It is pre-release. It has 105 stars. The README says Strobe "records everything on one clock" and "refuses measurements it cannot defend".
What it measures and when it says no
Strobe measures CPU time, memory, and disk and network input and output. It also measures GPU time, frame times on screen, and native stack samples. The README lists what it can measure on each operating system. It also lists what it will refuse. On macOS it cannot count wakes per thread without root. So it gives per-process counts only. On Linux it needs a real display without a compositor for frame times. So it refuses them on Wayland.
The author wrote those limits into the tool. If the operating system cannot give a reliable number, Strobe returns nothing for that number. It does not return zero.
Each app under test keeps its own strobe.json in its repository. That file holds how to build the app, how to serve it, and how to launch it on its own. Strobe itself does not know those steps.
How it joins a pull request
One command posts a block into the pull request on GitHub. The example block in the README reads: "44% less CPU time · 78% less app memory". The block says CPU time per 8 second run fell 44 percent per pair. The medians went from 7.05 to 3.90 milliseconds. The branch was better in 8 of 8 pairs. These are the author's own numbers from the README. They are not a real test result.
The block gives the medians, the number of pairs, and the range the author calls "likely". It also says the screen was locked and the display was off during the run. The next run can be set up the same way.
A caution line sits under the headline for any metric that got worse. The block is marked with a hash of its own text. A hand edit changes the hash. The next run will notice.
Where it fits
The tool runs on Windows x64, macOS on both arm64 and x64, and Linux on both x64 and arm64. It comes with an MCP server. The CLI and the MCP server use the same JSON output and the same names. An AI agent can call it the same way a person does. An agent can ask "did this branch make scrolling faster" and get the numbers a reviewer would read.
A benchmark is a TypeScript file. The README shows one example. It scrolls a timeline element down 4,000 pixels and back up. It then reads the 95th percentile of frames presented. The author compares this shape to BenchmarkDotNet, the .NET performance tool.
The tool is five days old. The star count is 105. No independent test of its measurements exists yet. The claims about what it refuses are the author's own. A team that needs a defensible number in a pull request will see soon whether the refusals are too strict.
Source
Strobe on GitHub, by Hona.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
