Vision fallback

Sometimes an element has nothing useful to match on by text — an icon-only button with no label, or several visually distinct controls that look identical in the accessibility tree.

When the normal text-based attempt fails and your configured model accepts image input, tamash-playwright automatically takes a screenshot and asks the model to point at the element visually. Same provider, same .env — no separate setup.

If it finds one, that on-screen point is translated back to a real element and a durable locator is derived from it, exactly as with the text path.

Which models support it

Any vision-capable model: gpt-4o / gpt-4.1, claude-haiku-4-5 and other recent Claude models, gemini-2.5-flash / gemini-2.5-pro, vision-capable Ollama models — and claude-subscription / copilot-subscription when pointed at a vision-capable model.

tamash (rule-based) has no model, so it never does vision. Check yours:

npx tamash-playwright doctor

The Vision capability row reports whether your configured model is expected to support it (by name — not a live check).

In the report

A vision heal shows vision=yes in the console line and the report attachment. It's strictly a second attempt — the cheaper text path always runs first.