[sanchitrk]

Summary

On every tool-input-delta, aikit appended the delta to the raw text and re-parsed the whole thing, so streaming large arguments (a big write or edit) cost quadratic CPU and memory. The raw text also rode on every partial message as partialJson.

• New utils/streamjson.ts: an incremental, tolerant JSON parser. It reads each delta once and returns snapshots that are never mutated after being handed out; finished strings and containers are shared, and only the open path is copied. Copy credit keeps total work linear even for a huge open container.
• stream.ts: a per-run Map<callID, ToolCallArgumentsStream> replaces partialJson; each delta sets block.arguments = args.append(delta). Final arguments still come from the AI SDK's parsed tool-call input.
• transform.ts: StreamingToolCallBlock, updateToolCallFromInput and the parseStreamingJson call are removed. jsonparse.ts is kept.

No wire or persisted format change.

Args (64-char deltas)	Before	After
100 KB	1.2 s	0.9 ms
200 KB	5.2 s	1.4 ms
500 KB	36 s	4.2 ms

Partial snapshots differ from the old parser only on edge cases: a trailing lone backslash is held back until its escape completes, __proto__ stays an own key, a missing comma is tolerated, and an array root shows as {}.

Test

test/e2e/streaming.tool.args.e2e.test.ts replays an Anthropic Messages SSE stream from a local server (no API key): about 600 KB of write arguments in 9,220 deltas. It asserts the final arguments, the delta and snapshot counts, that no snapshot is mutated, no partialJson on parts, and a 2 s bound (about 170 ms locally), then writes __artifacts__/streaming-tool-args.json.

vp check and vp run typecheck pass.

🤖 Generated with Claude Code

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}