Chrome DevTools for coding agents
feat: add a skill for using chrome-devtools-mcp (#830)
Alex Rudenkoaa0a3679f59a
diff --git a/scripts/eval_gemini.ts b/scripts/eval_gemini.tsindex 6a3e9662..d75a2a53 100644--- a/scripts/eval_gemini.ts+++ b/scripts/eval_gemini.ts@@ -16,6 +16,7 @@ import {TestServer} from '../build/tests/server.js';const ROOT_DIR = path.resolve(import.meta.dirname, '..');const SCENARIOS_DIR = path.join(import.meta.dirname, 'eval_scenarios');+const SKILL_PATH = path.join(ROOT_DIR, 'skills', 'chrome-devtools', 'SKILL.md');// Define schema for our test scenariosexport interface CapturedFunctionCall {@@ -49,6 +50,7 @@ async function runSingleScenario(server: TestServer,modelId: string,debug: boolean,+ includeSkill: boolean,): Promise<void> {const debugLog = (...args: unknown[]) => {if (debug) {@@ -67,6 +69,17 @@ async function runSingleScenario(const loadedScenario = await loadScenario(absolutePath);const scenario = {...loadedScenario};+ // Prepend skill content if requested+ if (includeSkill) {+ if (!fs.existsSync(SKILL_PATH)) {+ throw new Error(+ `Skill file not found at ${SKILL_PATH}. Please ensure the skill file exists.`,+ );+ }+ const skillContent = fs.readFileSync(SKILL_PATH, 'utf-8');+ scenario.prompt = `${skillContent}\n\n---\n\n${scenario.prompt}`;+ }+// Append random queryid to avoid caching issues and test distinct runsconst randomId = Math.floor(Math.random() * 1000000);scenario.prompt = `${scenario.prompt}\nqueryid=${randomId}`;@@ -180,6 +193,10 @@ async function main() {type: 'boolean',default: false,},+ 'include-skill': {+ type: 'boolean',+ default: false,+ },},allowPositionals: true,});@@ -187,6 +204,7 @@ async function main() {const modelId = values.model;const debug = values.debug;const repeat = values.repeat;+ const includeSkill = values['include-skill'];const scenarioFiles =positionals.length > 0@@ -211,7 +229,14 @@ async function main() {`Running scenario: ${path.relative(ROOT_DIR, scenarioPath)} (Run ${i}/3)`,);}- await runSingleScenario(scenarioPath, apiKey, server, modelId, debug);+ await runSingleScenario(+ scenarioPath,+ apiKey,+ server,+ modelId,+ debug,+ includeSkill,+ );console.log(`✔ ${path.relative(ROOT_DIR, scenarioPath)} (Run ${i})`);successCount++;} catch (e) {diff --git a/skills/chrome-devtools/SKILL.md b/skills/chrome-devtools/SKILL.mdnew file mode 100644index 00000000..551c03be--- /dev/null+++ b/skills/chrome-devtools/SKILL.md@@ -0,0 +1,44 @@+---+name: chrome-devtools+description: Uses Chrome DevTools via MCP for efficient debugging, troubleshooting and browser automation. Use when debugging web pages, automating browser interactions, analyzing performance, or inspecting network requests.+---++## Core Concepts++**Browser lifecycle**: Browser starts automatically on first tool call using a persistent Chrome profile. Configure via CLI args in the MCP server configuration: `npx chrome-devtools-mcp@latest --help`.++**Page selection**: Tools operate on the currently selected page. Use `list_pages` to see available pages, then `select_page` to switch context.++**Element interaction**: Use `take_snapshot` to get page structure with element `uid`s. Each element has a unique `uid` for interaction. If an element isn't found, take a fresh snapshot - the element may have been removed or the page changed.++## Workflow Patterns++### Before interacting with a page++1. Navigate: `navigate_page` or `new_page`+2. Wait: `wait_for` to ensure content is loaded if you know what you look for.+3. Snapshot: `take_snapshot` to understand page structure+4. Interact: Use element `uid`s from snapshot for `click`, `fill`, etc.++### Efficient data retrieval++- Use `filePath` parameter for large outputs (screenshots, snapshots, traces)+- Use pagination (`pageIdx`, `pageSize`) and filtering (`types`) to minimize data+- Set `includeSnapshot: false` on input actions unless you need updated page state++### Tool selection++- **Automation/interaction**: `take_snapshot` (text-based, faster, better for automation)+- **Visual inspection**: `take_screenshot` (when user needs to see visual state)+- **Additional details**: `evaluate_script` for data not in accessibility tree++### Parallel execution++You can send multiple tool calls in parallel, but maintain correct order: navigate → wait → snapshot → interact.++## Troubleshooting++If `chrome-devtools-mcp` is insufficient, guide users to use Chrome DevTools UI:++- https://developer.chrome.com/docs/devtools+- https://developer.chrome.com/docs/devtools/ai-assistance