git:20260906.ede7325 to git:20260909.17296ea

+ updated: "2026-09-08"

+ Use Chrome DevTools MCP or its experimental CLI for live browser diagnostics. [playwright](../playwright/SKILL.md) owns repeatable end-to-end journeys; [modern-web](../modern-web/SKILL.md) owns platform implementation guidance.

+ ## Setup

+ Require Node.js LTS, `npm`/`npx`, and supported Chrome. Reuse an available DevTools connection; otherwise follow [agent-mcp](../agent-mcp/SKILL.md) for the active harness. The launch command below uses a reviewed package pin; recheck installed help before updating it.

+ npx --yes chrome-devtools-mcp@1.9.0 --isolated --headless --no-usage-statistics --no-performance-crux

+ The MCP client starts that process over stdio. Use a dedicated browser profile. Keep remote debugging on loopback; browser content, network bodies, cookies, screenshots, and traces may expose private data. The two opt-out flags disable MCP usage statistics and CrUX URL lookups respectively.

+ For shell workflows, the same package exposes `chrome-devtools`; resolve it with `npm exec --yes --package=chrome-devtools-mcp@1.9.0 -- chrome-devtools <command>`. Inspect `status` before starting a daemon, then explicitly start the task's session with `start --workspace="$PWD" --no-usage-statistics --no-performance-crux`. The CLI otherwise starts a persistent daemon automatically and enables unrestricted file access by default. Read `start --help` for the installed flags; do not stop or repurpose another task's daemon.

+ ## Workflow

+

+ 1. **Identify the target**: list pages and select the intended page ID, URL, viewport, and browser mode. Take a fresh accessibility snapshot before using element UIDs; navigation and rerenders can invalidate them.

+ 1. **Reproduce the symptom**: capture console errors and relevant failed requests, then reduce to the smallest repeatable action. Redact credentials and private request data from artifacts.

+ 1. **Measure performance**: record a bounded trace of the same action before and after a change; preserve CPU/network throttling, cache conditions, viewport, and tool versions. Stop traces you started and save artifacts in the authorized workspace.

+ 1. **Check accessibility**: inspect roles, names, focus order, keyboard operation, and visible contrast. Combine automated checks with manual interaction; an accessibility tree or Lighthouse score alone does not establish WCAG conformance.

+ 1. **Investigate memory and cookies**: compare repeated lifecycle actions and heap snapshots when those tools are available; inspect `HttpOnly`, `Secure`, `SameSite`, and partitioning in the request's actual context. A single heap size or cookie attribute is not a diagnosis.

+ 1. **Verify the fix**: repeat the reproduction and relevant measurements, then add a regression test through the owning project workflow. Close task-owned pages and stop only a daemon created for this task.

+

+ - **Tool availability**: inspect the live tool schema or CLI help before calling a capability; optional categories, page routing, and CLI arguments vary by version.

+ - **Browser mode**: set headed or headless explicitly for reproducibility; the CLI defaults to headless, which need not match a separately configured MCP connection.

+ - **Lab versus field**: a local trace measures this run. CrUX field data and lab measurements describe different populations and time windows.

+ Upstream: [ChromeDevTools/chrome-devtools-mcp skills](https://github.com/ChromeDevTools/chrome-devtools-mcp/tree/main/skills). Use `skills add ChromeDevTools/chrome-devtools-mcp --list`, then follow the shared [vendor-skill policy](../agent-project/references/vendor-skills.md) for the required guidance.

+ - [Chrome DevTools for agents](https://github.com/ChromeDevTools/chrome-devtools-mcp) · [CLI](https://github.com/ChromeDevTools/chrome-devtools-mcp/blob/main/docs/cli.md) · [Tool reference](https://github.com/ChromeDevTools/chrome-devtools-mcp/blob/main/docs/tool-reference.md)