ai-json-safe-parse repairs broken LLM JSON

This is the right kind of boring utility: narrow, cheap, and likely to save more time than it takes to add. It doesn’t try to replace schema validation; it just gets you from messy model text to something parseable fast.

• –The recovery order matters: direct parse first, then extraction and normalization, then increasingly aggressive fixups, which keeps the happy path predictable.
• –The API trio covers most call sites without forcing one error-handling style on every project.
• –Zero dependencies and browser/edge support make it easy to drop into both backend agents and frontend AI flows.
• –The package is especially useful for local models, which often need a last-mile cleanup layer before their output is safe to consume.
• –It’s a good fallback, but anything user-facing should still validate the parsed object against a schema.