Seed library

Coverage of new patterns grows by adding worked examples, not by changing core. A seed is a human-written example — "when the captured steps look like this, the Playwright code should look like that" — that the optimization pass generalizes from as few-shot.

Writing a seed

Drop a JSON file in <projectRoot>/.hover/rules/:

{
  "name": "oauth-popup",
  "signature": ["browser_click", "browser_tabs:select"],
  "note": "sign in through a provider popup that opens a new tab",
  "example": {
    "steps": [
      { "tool": "browser_click", "element": "Sign in with Google button" },
      { "tool": "browser_tabs", "action": "select", "idx": 1 }
    ],
    "code": "const [popup] = await Promise.all([\n  context.waitForEvent('page'),\n  page.getByRole('button', { name: 'Sign in with Google' }).click(),\n]);\nawait popup.getByLabel('Email').fill('user@example.com');"
  }
}

• signature is a cheap relevance filter — Hover only feeds a seed to the pass when one of its base tools appears in the spec. It is not exact-matched.
• code must obey the same rules as generated specs: getByRole / getByLabel / getByText selectors, no XPath, no waitForTimeout.

No fork, no plugin code — just an example file. Share seeds with your team by committing .hover/rules/, or with the community as a package.

Built-in seeds

Hover ships a small built-in set for high-certainty, app-agnostic patterns (e.g. download → waitForEvent('download') pairing). The bar to be built-in is high: only deterministic, can't-mislead patterns qualify. High-frequency fixed-structure shapes like popup are hardcoded straight into the translator instead; semantic judgements (which feedback text to assert) stay as standing instructions in the pass, not seeds.