perf(ai): speed up partial JSON streaming by 20-27% by skipping a parse that always fails

[deadstrobe5]

Background

parsePartialJson runs on every delta of streamText with Output.object, of streamObject, and on every tool-input-delta in processUIMessageStream (so every useChat tool call). It starts with JSON.parse on the whole accumulated text. While streaming, that text almost always ends inside an unclosed string, object or array, so this parse fails, builds a JSONParseError, and only then does fixJson run and the repaired text get parsed again.

Summary

fixJson already tracks the parser state, so this adds repairJson, which also returns isTruncated. When it is set, parsePartialJson skips the parse that is bound to fail. The results do not change. A differential test compared old and new on every prefix of 613 random, pretty-printed and mutated documents (28,305 inputs), and every result matched.

End-to-End Verification

Full streamText / readUIMessageStream pipelines with a mock model streaming 8-char deltas, Node 26. Each run is a separate process, 5 runs per side:

scenario	before	after
streamText + Output.object, 20 KB	296.6 ms	217.6 ms
readUIMessageStream tool input, 20 KB	404.7 ms	323.0 ms
streamText + Output.object, 50 KB	1722 ms	1292 ms

Checklist

• All commits are signed (PRs with unsigned commits cannot be merged)
• Tests have been added / updated (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features)
• A patch changeset for relevant packages has been added (for bug fixes / features - run pnpm changeset in the project root)
• I have reviewed this pull request (self-review)

Future Work

processUIMessageStream also structuredClones the message on every chunk, which is the next per-chunk cost on the useChat path.