Releases: Makio64/gpu-perf-agent
Release list
v2.1.1 — Simpler agent onboarding
GPU Performance Agent now leads with one skill-install command for Codex, Claude Code, Gemini CLI, and other supported agents:
npx skills add Makio64/gpu-perf-agent -gChoose your agent, start a new session in your project, and ask:
Use gpu-perf-agent to optimize the webgpu performance
The README includes this setup flow. The website adds a compact profiling diagram, clear technical explanations, larger text, and corrected header alignment. Profiling behavior is unchanged.
Validation: 147 tests passed, 14 browser integration tests skipped; package-install smoke check passed. Website layout and interactions were checked at five desktop and mobile widths.
GPU Performance Agent 2.1.0
GPU Performance Agent now has a simpler introduction and a clean website that gets you from “what is this?” to your first useful capture.
- A concise README leads with the value, one-command profiling, and the before/after workflow, with a matching visual cover.
- The complete technical reference is preserved in
docs/reference.mdand included in the npm package. - A responsive website includes interactive report examples, copyable commands, keyboard-accessible tabs, and agent/MCP integration guidance.
- The website builds with Node.js alone, uses no third-party runtime assets, and deploys automatically through GitHub Pages.
- Profiling behavior and defaults are unchanged from 2.0.0.
Website: https://makio64.github.io/gpu-perf-agent/
Validation: 147 unit/CLI tests passed (14 browser-only tests skipped); website checked at 320, 390, 768, 1024, and 1440px with no horizontal overflow or browser errors. Report tabs, keyboard navigation, clipboard copying and fallback, image loading, anchors, and JavaScript-disabled content were checked in Chrome. The exact release tarball passed a clean install, CLI doctor, and packaged-documentation checks.
Install: npm install -D gpu-perf-agent@2.1.0
GPU Performance Agent 2.0.0
GPU Performance Agent 2.0.0 reduces profiling overhead and makes repeated captures and regression checks more reliable for developers and agents.
- Zero required npm dependencies for the fast runner, incremental allocation tracking, reusable GPU timers, and compact agent output.
- Isolated persistent-browser jobs, bounded capture and trace lifecycles, queue limits, and robust cleanup.
- Interleaved A/B captures, Welch intervals for independent reports, validated budgets, and explicit invalid-measurement results.
- Existing MCP tools, HTML reports, detailed recommendations, adaptive sampling, screenshots, and CDP attachment remain available.
Migration
Requires Node.js 22+. Persistent jobs now use isolated contexts; choose --context shared for intentional state reuse. The agent CLI defaults to six 250 ms samples and two warmups. Adaptive sampling is opt-in on the CLI. Independent comparisons use Welch intervals; select --paired only for matched observations. Exit code 2 means invalid measurements or incompatible comparison settings. Reports use schema version 2 with structured warnings. See CHANGELOG.md for API changes.
Validation
- 161 tests passed with real Chrome/localhost tests enabled; zero failures or skips.
- Clean installation of the exact release tarball pulled in only gpu-perf-agent, with no required dependencies.
- Published entry points loaded; packaged CLI captured WebGPU measurements, a GPU trace, HTML output, and a screenshot.
- Optional Playwright runner passed a separate real-browser smoke check.
- Bundled agent skill validated.
Install: npm install -D gpu-perf-agent@2.0.0
The attached tarball is the artifact validated and submitted to npm. Snapshot and startup microbenchmarks in OPTIMIZATION.md measure tool overhead, not application GPU speedups.