test: add extensible browser evaluation harness

evals/browser/cases/README.md

@@ -0,0 +1,9 @@
+# Browser evaluation cases
+
+Cases are discovered recursively from `*.case.json`; colocated `*.fixture.mjs` files are discovered
+as fixtures. Use `core` for small deterministic acceptance checks, `matrix` for seeded variations,
+and `regression/<case-id>/` for privacy-safe minimized user badcases.
+
+Run `pnpm eval:browser validate` after every case change. See the parent
+[`README.md`](../README.md) for the manifest contract, workflow actions, scaffolding, and acceptance
+criteria.

evals/browser/cases/core/diagnostics.case.json

@@ -0,0 +1,48 @@
+{
+  "$schema": "../../schemas/case.schema.json",
+  "schemaVersion": 1,
+  "id": "diagnostics",
+  "title": "Console and network inspection",
+  "order": 50,
+  "suite": "core",
+  "tags": ["console", "network", "diagnostics"],
+  "fixture": { "startPath": "/diagnostics" },
+  "prompts": {
+    "en": "Open {url}. Use browser diagnostics to inspect the page console and recent network activity. Report the console marker and the token returned by the ping request, then close the browser session.",
+    "zh-CN": "打开 {url}。使用浏览器诊断能力检查页面 console 和最近的 network 活动，报告 console marker 以及 ping 请求返回的 token，最后关闭浏览器会话。"
+  },
+  "coverage": [
+    "session.start",
+    "session.stop",
+    "page.navigate",
+    "inspect.console",
+    "inspect.network"
+  ],
+  "assertions": {
+    "site": [
+      {
+        "label": "diagnostic ping completed",
+        "type": "network.ping",
+        "minCount": 1
+      }
+    ],
+    "response": [
+      { "label": "reported console marker", "includes": "CONSOLE-61" },
+      { "label": "reported network token", "includes": "NETWORK-73" }
+    ],
+    "adapter": [
+      { "label": "console inspection was executed", "key": "consoleRead" },
+      { "label": "network inspection was executed", "key": "networkRead" },
+      { "label": "browser session was closed", "key": "sessionStopped" }
+    ]
+  },
+  "smoke": {
+    "steps": [
+      { "action": "navigate" },
+      { "action": "wait", "duration": "300ms" },
+      { "action": "console", "evidence": "consoleRead" },
+      { "action": "network", "evidence": "networkRead" },
+      { "action": "observe" }
+    ]
+  }
+}

evals/browser/cases/core/form-controls.case.json

@@ -0,0 +1,82 @@
+{
+  "$schema": "../../schemas/case.schema.json",
+  "schemaVersion": 1,
+  "id": "form-controls",
+  "title": "Form, select, keyboard, and screenshot",
+  "order": 20,
+  "suite": "core",
+  "tags": ["form", "keyboard", "screenshot", "select"],
+  "fixture": { "startPath": "/form" },
+  "prompts": {
+    "en": "Open {url}. Fill Text input with \"agent-parity\" and Textarea with \"BrowserSkill works\". Select the visible option \"Two\" from Dropdown. Capture a screenshot of the form, focus Submit, and press Enter to submit. Confirm that the result says \"Received!\", then close the browser session.",
+    "zh-CN": "打开 {url}。把 Text input 填成“agent-parity”，Textarea 填成“BrowserSkill works”，Dropdown 选择可见选项“Two”。截取表单截图，然后聚焦 Submit 并按 Enter 提交。确认结果页出现“Received!”，最后关闭浏览器会话。"
+  },
+  "coverage": [
+    "session.start",
+    "session.stop",
+    "page.navigate",
+    "interact.fill",
+    "interact.select",
+    "interact.press",
+    "inspect.screenshot"
+  ],
+  "assertions": {
+    "site": [
+      {
+        "label": "Enter was pressed on Submit",
+        "type": "form.enter_pressed",
+        "minCount": 1
+      },
+      {
+        "label": "form values were submitted",
+        "type": "form.submitted",
+        "where": {
+          "data.text": "agent-parity",
+          "data.notes": "BrowserSkill works",
+          "data.choice": "two"
+        },
+        "minCount": 1
+      },
+      {
+        "label": "result page was displayed",
+        "type": "page.shown",
+        "where": { "data.path": "/result" },
+        "minCount": 1
+      }
+    ],
+    "response": [{ "label": "confirmed result", "includes": "Received!" }],
+    "adapter": [
+      { "label": "screenshot artifact was created", "key": "screenshotCreated" },
+      { "label": "browser session was closed", "key": "sessionStopped" }
+    ]
+  },
+  "smoke": {
+    "steps": [
+      { "action": "navigate" },
+      {
+        "action": "fill",
+        "selector": "#text-input",
+        "value": "agent-parity"
+      },
+      {
+        "action": "fill",
+        "selector": "#notes",
+        "value": "BrowserSkill works"
+      },
+      { "action": "select", "selector": "#choice", "values": ["two"] },
+      {
+        "action": "screenshot",
+        "artifactName": "form.png",
+        "evidence": "screenshotCreated"
+      },
+      { "action": "press", "key": "Enter", "selector": "#submit" },
+      {
+        "action": "wait-site-event",
+        "type": "page.shown",
+        "where": { "path": "/result" }
+      },
+      { "action": "wait", "duration": "200ms" },
+      { "action": "observe" }
+    ]
+  }
+}

evals/browser/cases/core/hover-inspect.case.json

@@ -0,0 +1,59 @@
+{
+  "$schema": "../../schemas/case.schema.json",
+  "schemaVersion": 1,
+  "id": "hover-inspect",
+  "title": "Hover, local HTML, and screenshot",
+  "order": 30,
+  "suite": "core",
+  "tags": ["hover", "html", "screenshot"],
+  "fixture": { "startPath": "/hover-inspect" },
+  "prompts": {
+    "en": "Open {url}. Hover over Account, then choose Profile beta from the revealed menu. Inspect the local HTML around the Inspection card and capture a screenshot. Report both marker codes you find, then close the browser session.",
+    "zh-CN": "打开 {url}。把鼠标悬停在“Account”上，然后点击展开菜单中的“Profile beta”。检查 Inspection card 附近的局部 HTML，并截取页面截图。报告找到的两个标记代码，最后关闭浏览器会话。"
+  },
+  "coverage": [
+    "session.start",
+    "session.stop",
+    "page.navigate",
+    "interact.hover",
+    "interact.click",
+    "inspect.html",
+    "inspect.screenshot"
+  ],
+  "assertions": {
+    "site": [
+      {
+        "label": "Account menu was hovered",
+        "type": "hover.menu_opened",
+        "minCount": 1
+      },
+      {
+        "label": "Profile beta was selected",
+        "type": "hover.profile_clicked",
+        "minCount": 1
+      }
+    ],
+    "response": [
+      { "label": "reported hover marker", "includes": "HOVER-17" },
+      { "label": "reported HTML marker", "includes": "INSPECT-27" }
+    ],
+    "adapter": [
+      { "label": "HTML inspection was executed", "key": "htmlInspected" },
+      { "label": "screenshot artifact was created", "key": "screenshotCreated" },
+      { "label": "browser session was closed", "key": "sessionStopped" }
+    ]
+  },
+  "smoke": {
+    "steps": [
+      { "action": "navigate" },
+      { "action": "hover", "selector": "#account-menu" },
+      { "action": "click", "selector": "#profile-link" },
+      { "action": "html", "evidence": "htmlInspected" },
+      {
+        "action": "screenshot",
+        "artifactName": "hover-inspect.png",
+        "evidence": "screenshotCreated"
+      }
+    ]
+  }
+}

evals/browser/cases/core/mobile-emulation.case.json

@@ -0,0 +1,43 @@
+{
+  "$schema": "../../schemas/case.schema.json",
+  "schemaVersion": 1,
+  "id": "mobile-emulation",
+  "title": "Mobile emulation",
+  "order": 60,
+  "suite": "core",
+  "tags": ["emulation", "mobile", "viewport"],
+  "fixture": { "startPath": "/responsive" },
+  "prompts": {
+    "en": "Open {url}, emulate an iPhone 14, and report the viewport marker shown on the page. Close the browser session when finished.",
+    "zh-CN": "打开 {url}，模拟 iPhone 14，并报告页面显示的 viewport marker。完成后关闭浏览器会话。"
+  },
+  "coverage": ["session.start", "session.stop", "page.navigate", "assist.emulate"],
+  "assertions": {
+    "site": [
+      {
+        "label": "iPhone-sized viewport was observed",
+        "type": "viewport.sample",
+        "where": { "data.width": 390, "data.height": 844 },
+        "minCount": 1
+      }
+    ],
+    "response": [{ "label": "reported viewport marker", "includes": "VIEWPORT-84" }],
+    "adapter": [
+      { "label": "device emulation was executed", "key": "emulatedDevice" },
+      { "label": "browser session was closed", "key": "sessionStopped" }
+    ]
+  },
+  "smoke": {
+    "steps": [
+      { "action": "navigate" },
+      { "action": "resize", "width": 900, "height": 700 },
+      {
+        "action": "emulate",
+        "device": "iphone-14",
+        "evidence": "emulatedDevice"
+      },
+      { "action": "wait", "duration": "300ms" },
+      { "action": "html" }
+    ]
+  }
+}

evals/browser/cases/core/navigation-history.case.json

@@ -0,0 +1,67 @@
+{
+  "$schema": "../../schemas/case.schema.json",
+  "schemaVersion": 1,
+  "id": "navigation-history",
+  "title": "Navigation, history, and reload",
+  "order": 10,
+  "suite": "core",
+  "tags": ["navigation", "history", "reload"],
+  "fixture": { "startPath": "/navigation/start" },
+  "prompts": {
+    "en": "Open {url}. Read the main heading, follow the Details link, and report the code shown there. Then go back, go forward, reload the page, and close the browser session. Use only the browser capability for page operations.",
+    "zh-CN": "打开 {url}。读取页面主标题，点击“详情”，并报告详情页上的代码。然后依次后退、前进、刷新页面，最后关闭浏览器会话。页面操作只使用浏览器能力。"
+  },
+  "coverage": [
+    "session.start",
+    "session.stop",
+    "page.navigate",
+    "page.back",
+    "page.forward",
+    "page.reload",
+    "interact.click"
+  ],
+  "assertions": {
+    "site": [
+      {
+        "label": "details link was clicked",
+        "type": "navigation.detail.clicked",
+        "minCount": 1
+      },
+      {
+        "label": "back returned to the start page",
+        "type": "page.shown",
+        "where": { "data.path": "/navigation/start" },
+        "minCount": 2
+      },
+      {
+        "label": "forward returned to the detail page",
+        "type": "page.shown",
+        "where": { "data.path": "/navigation/detail" },
+        "minCount": 2
+      },
+      {
+        "label": "detail page was reloaded",
+        "type": "page.shown",
+        "where": {
+          "data.path": "/navigation/detail",
+          "data.navigationType": "reload"
+        },
+        "minCount": 1
+      }
+    ],
+    "response": [{ "label": "reported detail code", "includes": "NAV-42" }],
+    "adapter": [{ "label": "browser session was closed", "key": "sessionStopped" }]
+  },
+  "smoke": {
+    "steps": [
+      { "action": "navigate" },
+      { "action": "observe" },
+      { "action": "snapshot" },
+      { "action": "click", "selector": "#details-link" },
+      { "action": "html" },
+      { "action": "back" },
+      { "action": "forward" },
+      { "action": "reload" }
+    ]
+  }
+}

evals/browser/cases/core/tabs.case.json

@@ -0,0 +1,71 @@
+{
+  "$schema": "../../schemas/case.schema.json",
+  "schemaVersion": 1,
+  "id": "tabs",
+  "title": "Create, select, and close a tab",
+  "order": 40,
+  "suite": "core",
+  "tags": ["tabs", "ownership", "lifecycle"],
+  "fixture": { "startPath": "/tabs" },
+  "prompts": {
+    "en": "Open {url}. Create a new browser tab at {baseUrl}/tabs/child?run={runId}, read its token, click \"Mark child visited\", close the child tab, return to the original tab, and close the browser session.",
+    "zh-CN": "打开 {url}。新建一个浏览器标签页并访问 {baseUrl}/tabs/child?run={runId}，读取其中的 token，点击“Mark child visited”，关闭子标签页，回到原标签页，最后关闭浏览器会话。"
+  },
+  "coverage": [
+    "session.start",
+    "session.stop",
+    "page.navigate",
+    "interact.click",
+    "tabs.create",
+    "tabs.select",
+    "tabs.close"
+  ],
+  "assertions": {
+    "site": [
+      {
+        "label": "child tab page was displayed",
+        "type": "page.shown",
+        "where": { "data.path": "/tabs/child" },
+        "minCount": 1
+      },
+      {
+        "label": "child tab action ran",
+        "type": "tabs.child_marked",
+        "minCount": 1
+      }
+    ],
+    "response": [{ "label": "reported child token", "includes": "TAB-29" }],
+    "adapter": [
+      { "label": "child tab was closed", "key": "childTabClosed" },
+      { "label": "browser session was closed", "key": "sessionStopped" }
+    ]
+  },
+  "smoke": {
+    "steps": [
+      { "action": "navigate", "saveAs": "original" },
+      { "action": "tab-list" },
+      {
+        "action": "tab-create",
+        "url": "{baseUrl}/tabs/child?run={runId}",
+        "saveAs": "child"
+      },
+      {
+        "action": "navigate",
+        "url": "{baseUrl}/tabs/child?run={runId}",
+        "tabId": "{child.tab_id}"
+      },
+      {
+        "action": "click",
+        "selector": "#mark-child",
+        "tabId": "{child.tab_id}"
+      },
+      { "action": "html", "tabId": "{child.tab_id}" },
+      {
+        "action": "tab-close",
+        "tabId": "{child.tab_id}",
+        "evidence": "childTabClosed"
+      },
+      { "action": "tab-select", "tabId": "{original.tab_id}" }
+    ]
+  }
+}

evals/browser/cases/matrix/README.md

@@ -0,0 +1,5 @@
+# Seeded matrix cases
+
+Matrix cases vary page structure or timing deterministically from a recorded seed. The task,
+assertions, and success marker remain stable so a failed seed can be replayed exactly. A seed that
+exposes a distinct product bug should be minimized and promoted to a named regression case.

evals/browser/cases/matrix/generated-form.case.json

@@ -0,0 +1,77 @@
+{
+  "$schema": "../../schemas/case.schema.json",
+  "schemaVersion": 1,
+  "id": "generated-form",
+  "title": "Seeded form structure and hydration variants",
+  "order": 10,
+  "suite": "matrix",
+  "tags": ["dynamic-dom", "form", "hydration", "seeded"],
+  "seed": 20260827,
+  "fixture": { "startPath": "/matrix/generated-form" },
+  "prompts": {
+    "en": "Open {url}. Wait for the seeded form to finish hydrating. Fill Text input with \"seeded-agent\", select the visible option \"Two\", submit with Enter, and report the result marker and seed. Close the browser session.",
+    "zh-CN": "打开 {url}。等待 seeded form 完成 hydration，把 Text input 填成“seeded-agent”，选择可见选项“Two”，用 Enter 提交，并报告结果 marker 和 seed。最后关闭浏览器会话。"
+  },
+  "coverage": [
+    "session.start",
+    "session.stop",
+    "page.navigate",
+    "page.wait",
+    "inspect.observe",
+    "interact.fill",
+    "interact.select",
+    "interact.press"
+  ],
+  "assertions": {
+    "site": [
+      { "label": "generated form hydrated", "type": "matrix.ready", "minCount": 1 },
+      {
+        "label": "Enter was pressed on generated form",
+        "type": "matrix.enter_pressed",
+        "minCount": 1
+      },
+      {
+        "label": "generated values were submitted",
+        "type": "matrix.submitted",
+        "where": { "data.text": "seeded-agent", "data.choice": "two" },
+        "minCount": 1
+      },
+      {
+        "label": "matrix result page was displayed",
+        "type": "page.shown",
+        "where": { "data.path": "/matrix/generated-result" },
+        "minCount": 1
+      }
+    ],
+    "response": [
+      { "label": "reported matrix result", "includes": "MATRIX-OK" },
+      { "label": "reported selected seed", "includes": "SEED-" }
+    ],
+    "adapter": [{ "label": "browser session was closed", "key": "sessionStopped" }]
+  },
+  "smoke": {
+    "steps": [
+      { "action": "navigate" },
+      {
+        "action": "wait-site-event",
+        "type": "matrix.ready",
+        "timeoutMs": 5000
+      },
+      { "action": "observe" },
+      { "action": "fill", "selector": "[name=text]", "value": "seeded-agent" },
+      { "action": "select", "selector": "[name=choice]", "values": ["two"] },
+      {
+        "action": "press",
+        "key": "Enter",
+        "selector": "button[type=submit]"
+      },
+      {
+        "action": "wait-site-event",
+        "type": "page.shown",
+        "where": { "path": "/matrix/generated-result" }
+      },
+      { "action": "wait", "duration": "200ms" },
+      { "action": "observe" }
+    ]
+  }
+}

evals/browser/cases/regression/README.md

@@ -0,0 +1,12 @@
+# Regression cases
+
+Each user-reported badcase should live in its own directory with a manifest, synthetic fixture, and
+README describing the symptom, minimal reproduction, expected behavior, source reference, and fix.
+Never commit credentials, cookies, HAR files, production HTML, private screenshots, or proprietary
+assets.
+
+Create a starting point with:
+
+```sh
+pnpm eval:browser scaffold <case-id> --title "..." --source "issue-or-pr"
+```