Add Jev browser CLI mode with saved OpenRouter settings

.codex/skills/cate-cli/SKILL.md

@@ -38,6 +38,35 @@ panel. If a selected panel was closed, select another panel before continuing.
 
 ## Browser workflow
 
+### Jev mode
+
+Give Jev a natural-language task for the selected browser panel:
+
+```bash
+cate browser jev 'Set the Greeting field to Hi and click Save' --panel <id>
+cate browser jev 'Open the Settings tab' --max-steps 10 --json
+```
+
+Save your OpenRouter API key in **Cate Settings → CLI → OpenRouter API key**.
+No terminal export is needed. Enable Browser Read and Control in the same settings.
+This mode sends the prompt and
+page accessibility text to Jev (`typesafe/jev-1.13`) through OpenRouter's Decisions
+API. Cate makes provider requests using the saved key; the key is not sent to the CLI.
+Clearing the setting disables Jev. No other model is used: Jev selects browser
+operations, elements, and complete destination URLs supplied in the prompt.
+For text input, the prompt is split on whitespace and Jev selects one word as the
+entire field value. Duplicate words are offered once; punctuation is preserved.
+Words cannot be combined and new text cannot be generated. Include the exact field
+value as a single word and supply absolute HTTP/HTTPS destination URLs explicitly.
+
+Supported operations are HTTP/HTTPS navigation, clicks, field replacement, keys, page scrolling, and
+waits. Runs pin the panel/tab and stop on user takeover, uncertainty, errors,
+180 seconds, or the step limit (20 by default, up to 100). `--json` returns
+the status, action trace, model-call count, and final URL; only `done` exits zero.
+Completion is a model judgment based on the page; verify important outcomes.
+
+### JavaScript mode
+
 Browser control uses persistent JavaScript with the `cua` tab API. The old argv
 actions, selectors, page evaluation, and revisioned string refs have been removed.
 Start by binding a tab to get its accessibility state, then request a screenshot

e2e/browser-jev.spec.ts

@@ -0,0 +1,99 @@
+import { expect, test } from '@playwright/test'
+import type { ElectronApplication, Page } from 'playwright'
+import { closeApp, launchApp, seedTerminal } from './fixtures/electron-app'
+import { fixtureEvaluate, target } from './fixtures/browser-control'
+
+async function terminalCommand(page: Page, nodeId: string, command: string, marker: string) {
+  const wrapped = process.platform === 'win32'
+    ? `${command}; Write-Output ("${marker}:{0}" -f $LASTEXITCODE)\r`
+    : `${command}; cate_status=$?; printf '\\n${marker}:%s\\n' "$cate_status"\r`
+  expect(await page.evaluate(({ nodeId, wrapped }) => window.__cateE2E!.writeTerminal(nodeId, wrapped), { nodeId, wrapped })).toBe(true)
+  await expect.poll(() => page.evaluate(({ nodeId, marker }) =>
+    new RegExp(`${marker}:\\d+`).test(window.__cateE2E!.terminalText(nodeId) ?? ''), { nodeId, marker }), { timeout: 200_000 }).toBe(true)
+  const text = await page.evaluate(nodeId => window.__cateE2E!.terminalText(nodeId), nodeId)
+  return { code: Number(text!.match(new RegExp(`${marker}:(\\d+)`))![1]), text }
+}
+
+async function readyTerminal(page: Page) {
+  const nodeId = await seedTerminal(page, { x: 100, y: 100 })
+  await expect.poll(() => page.evaluate(id => window.__cateE2E!.terminalPtyId(id), nodeId), { timeout: 60_000 }).not.toBeNull()
+  return nodeId
+}
+
+/** Only replace the external provider. CLI, host permissions, settings, and browser actions are real. */
+async function mockDecisions(app: ElectronApplication, decisions: string[]) {
+  await app.evaluate((_electron, decisions) => {
+    const originalFetch = globalThis.fetch
+    globalThis.fetch = async (input, init) => {
+      if (String(input) !== 'https://openrouter.ai/api/alpha/decisions') return originalFetch(input, init)
+      if (new Headers(init?.headers).get('Authorization') !== 'Bearer test-saved-key') throw new Error('Saved key was not used')
+      const request = JSON.parse(String(init?.body))
+      const criteria = request.questions.decision.criteria
+      const next = decisions.shift()!
+      const choice = Object.hasOwn(criteria, next) ? next : Object.keys(criteria).find(key => criteria[key] === next)!
+      return new Response(JSON.stringify({ answers: { decision: { type: 'choice', choice, confidence: 1, probabilities: { [choice]: 1 } } } }))
+    }
+  }, decisions)
+}
+
+for (const word of ['Hi', 'Hello!']) test(`Jev saves ${word} using the Cate setting without an exported key`, async () => {
+  test.setTimeout(240_000)
+  const live = process.env.CATE_LIVE_JEV === '1'
+  if (live) expect(process.env.OPENROUTER_API_KEY).toBeTruthy()
+  const { electronApp: app, mainWindow: page } = await launchApp({ env: { OPENROUTER_API_KEY: '' } })
+  try {
+    if (live) {
+      await page.evaluate(key => window.electronAPI.settingsSet('cliOpenRouterApiKey', key), process.env.OPENROUTER_API_KEY!)
+    } else {
+      await page.evaluate(() => window.__cateE2E!.openSettings('cli'))
+      const keyInput = page.getByLabel('OpenRouter API key', { exact: true })
+      await expect(keyInput).toHaveAttribute('type', 'password')
+      await keyInput.fill(' test-saved-key ')
+      await keyInput.press('Enter')
+      await expect.poll(() => page.evaluate(() => window.electronAPI.settingsGet('cliOpenRouterApiKey'))).toBe('test-saved-key')
+      await page.keyboard.press('Escape')
+      await mockDecisions(app, ['setValue', 'c0', word, 'click', 'c0', 'done'])
+    }
+    const html = `<title>Greeting form</title><label>Greeting <input id="greeting"></label>
+      <button onclick="document.getElementById('result').textContent='Saved greeting: '+document.getElementById('greeting').value">Save</button><p id="result">Not saved yet</p>`
+    const browser = await page.evaluate(url => window.__cateE2E!.createBrowser(url, { x: 120, y: 120 }), `data:text/html,${encodeURIComponent(html)}`)
+    await target(page, browser, 'Greeting')
+    const terminal = await readyTerminal(page)
+    const result = await terminalCommand(page, terminal,
+      `cate browser jev 'Set Greeting to ${word} then click Save. Finish when the saved greeting shows ${word}' --panel ${browser.panelId} --max-steps 5 --json`, '__JEV_SAVED')
+    expect(result.code, result.text ?? '').toBe(0)
+    expect(result.text).toContain('"status":"done"')
+    expect(result.text).not.toContain('test-saved-key')
+    expect(await fixtureEvaluate(app, page, browser, 'document.getElementById("greeting").value')).toBe(word)
+    expect(await fixtureEvaluate(app, page, browser, 'document.getElementById("result").textContent')).toBe(`Saved greeting: ${word}`)
+    await page.evaluate(() => window.electronAPI.settingsSet('cliOpenRouterApiKey', ''))
+    const missing = await terminalCommand(page, terminal,
+      `cate browser jev 'Verify the saved greeting' --panel ${browser.panelId} --json`, '__JEV_CLEARED')
+    expect(missing.code).toBe(1)
+    expect(missing.text).toContain('Set the OpenRouter API key in Cate Settings')
+  } finally {
+    await page.evaluate(() => window.electronAPI.settingsSet('cliOpenRouterApiKey', '')).catch(() => {})
+    await closeApp(app)
+  }
+})
+
+test('Jev completes Wikipedia search through the CLI with a saved key', async () => {
+  test.skip(process.env.CATE_LIVE_JEV !== '1', 'Requires OpenRouter and public Wikipedia')
+  test.setTimeout(240_000)
+  expect(process.env.OPENROUTER_API_KEY).toBeTruthy()
+  const { electronApp: app, mainWindow: page } = await launchApp({ env: { OPENROUTER_API_KEY: '' } })
+  try {
+    await page.evaluate(key => window.electronAPI.settingsSet('cliOpenRouterApiKey', key), process.env.OPENROUTER_API_KEY!)
+    const browser = await page.evaluate(() => window.__cateE2E!.createBrowser('about:blank', { x: 120, y: 120 }))
+    const terminal = await readyTerminal(page)
+    const result = await terminalCommand(page, terminal,
+      `cate browser jev 'Go to https://www.wikipedia.org and search for Berlin then submit the search. Finish when the Berlin article heading is visible.' --panel ${browser.panelId} --max-steps 15 --json`, '__JEV_WIKIPEDIA')
+    expect(result.code, result.text ?? '').toBe(0)
+    expect(result.text).toContain('"status":"done"')
+    expect(await fixtureEvaluate(app, page, browser, 'document.querySelector("h1").textContent')).toBe('Berlin')
+    expect(await fixtureEvaluate(app, page, browser, 'location.pathname')).toBe('/wiki/Berlin')
+  } finally {
+    await page.evaluate(() => window.electronAPI.settingsSet('cliOpenRouterApiKey', '')).catch(() => {})
+    await closeApp(app)
+  }
+})

skills/cate-cli/SKILL.md

@@ -38,6 +38,35 @@ panel. If a selected panel was closed, select another panel before continuing.
 
 ## Browser workflow
 
+### Jev mode
+
+Give Jev a natural-language task for the selected browser panel:
+
+```bash
+cate browser jev 'Set the Greeting field to Hi and click Save' --panel <id>
+cate browser jev 'Open the Settings tab' --max-steps 10 --json
+```
+
+Save your OpenRouter API key in **Cate Settings → CLI → OpenRouter API key**.
+No terminal export is needed. Enable Browser Read and Control in the same settings.
+This mode sends the prompt and
+page accessibility text to Jev (`typesafe/jev-1.13`) through OpenRouter's Decisions
+API. Cate makes provider requests using the saved key; the key is not sent to the CLI.
+Clearing the setting disables Jev. No other model is used: Jev selects browser
+operations, elements, and complete destination URLs supplied in the prompt.
+For text input, the prompt is split on whitespace and Jev selects one word as the
+entire field value. Duplicate words are offered once; punctuation is preserved.
+Words cannot be combined and new text cannot be generated. Include the exact field
+value as a single word and supply absolute HTTP/HTTPS destination URLs explicitly.
+
+Supported operations are HTTP/HTTPS navigation, clicks, field replacement, keys, page scrolling, and
+waits. Runs pin the panel/tab and stop on user takeover, uncertainty, errors,
+180 seconds, or the step limit (20 by default, up to 100). `--json` returns
+the status, action trace, model-call count, and final URL; only `done` exits zero.
+Completion is a model judgment based on the page; verify important outcomes.
+
+### JavaScript mode
+
 Browser control uses persistent JavaScript with the `cua` tab API. The old argv
 actions, selectors, page evaluation, and revisioned string refs have been removed.
 Start by binding a tab to get its accessibility state, then request a screenshot