tamash — rule-based, no AI

HEALER_PROVIDER=tamash

No API key, no subscription, no network call, no tokens. Available in every language.

How it resolves a broken locator

It searches the page's accessibility tree for the same text your .describe() label (or a decoded named-variable, e.g. txtLastName → "Last Name") already gives it, then applies the same structural-widening logic — nearby-label anchoring — that the AI providers use once they've found the right anchor.

doctor always reports it connected — there's nothing external for it to reach.

The honest tradeoff

It never guesses. These all get an honest decline (stage=ai_declined) rather than a guessed heal:

• a paraphrased description that doesn't closely match on-page text
• a raw CSS/XPath selector with no human-readable text at all
• two equally-plausible elements near the same anchor

Those are exactly the cases an AI's real reasoning can resolve and a text search can't. tamash also can't do the vision fallback (no model to look at an image) or action-recovery tactics (both need inference it deliberately doesn't attempt).

When to use it

A fast, free, fully deterministic first line of defense — especially for well-.describe()d, Page-Object-style suites, where its success rate is genuinely close to an AI's. Fall back to one of the AI providers for anything it declines.