PSJ: runnable original-stream identity checks before caller-owned tool dispatch

[canblmz1]

I maintain prefix-safe-json, a guard for original streamed tool arguments and termination evidence before caller-owned execution. There is now a runnable native-versus-PSJ comparison for developers evaluating that boundary.

For a trial without cloning the repository, download the small standalone ZIP and choose the AI SDK, OpenAI, Anthropic or LangChain directory in examples/README.md. All four launchers passed with published PSJ 0.5.3; the public ZIP was downloaded again, verified against source, and its OpenAI consumer passed in a clean temporary project.

From a PSJ checkout with Node 22+ and npm:

node scripts/try-npm-demo.mjs

It installs published PSJ 0.5.3 and pinned AI SDK 7.0.77 in a separate temporary project, with no API key. The new npm run identity example checks complete, schema-valid calls under successful termination:

Authored SDK fixture	Native effects	Locked tools + PSJ caller effects
Two distinct call IDs	2	2
Two calls sharing an ID	2	0

Native controls deliberately execute harmless in-memory counters. The PSJ path keeps SDK callbacks at zero and dispatches only after one-shot authority. The same comparison passes with installed AI SDK 7.0.140, and packed-candidate CI checks four exact SDK versions. Fixture source and verification limits.

For next-ai-draw-io specifically, the current route intentionally repairs JSON and its tool handler handles incomplete XML. Our external fit audit ran 57 upstream tests and two comparisons against the actual helpers: PSJ does not replace that XML recovery policy. This SDK comparison is not evidence that the application needs PSJ, a WorkflowAgent fix, or a live-provider/project integration result. The application's AI SDK 6.0.300 range remains a separate integration question.

If you want to evaluate caller-owned dispatch in your own development branch, start with a harmless counter and retain original deltas and terminal events. Trial/report guide. Please report the exact versions and whether this adds a useful check at your boundary; a standalone mock run is separate from an own-project result.