AI SDK calls in Next.js: auto-record via next.config or opt in per call?

I've been working on capturing every streamText / generateText call in a Next.js app on Vercel (prompt timing, tokens, failures) together with the request that triggered it, and I'm stuck on a design choice I'd like other opinions on.

Two ways I see:

Patch next.config at install time so every AI SDK call is recorded with zero code changes. Nothing gets missed, but it's "magic": you don't see in your code that it's happening.
Opt in per call (the AI SDK's own telemetry option), so it's explicit in the code, but any call someone forgets to flag goes unrecorded.

A few things I learned on the request side, in case they help anyone doing similar work:

instrumentation.ts is the clean entry point for server-side recording. If the app lives in src/, the file has to go there too.
Naming requests by Next's own route (GET /api/orders/[id]) instead of the raw URL keeps the data groupable.
Never overwrite an existing instrumentation.ts; many apps already have one for something else.
On Next.js 13/14 you may need experimental.instrumentationHook: true in next.config.
Stack traces from the compiled .next output only map back to your source if the source maps sit next to the build.

Questions:

Would you accept a tool patching next.config for this, or would that be a dealbreaker?
Anyone combining AI SDK telemetry with request traces in production? How do you link an AI call to the user request that caused it?

Disclosure: I'm building Relia, a monitoring tool, and this is for its Next.js setup. I'm asking for opinions on the approach, not trying to sell anything here.

Two ways I see:

Patch next.config at install time so every AI SDK call is recorded with zero code changes. Nothing gets missed, but it's "magic": you don't see in your code that it's happening.
Opt in per call (the AI SDK's own telemetry option), so it's explicit in the code, but any call someone forgets to flag goes unrecorded.

A few things I learned on the request side, in case they help anyone doing similar work:

instrumentation.ts is the clean entry point for server-side recording. If the app lives in src/, the file has to go there too.
Naming requests by Next's own route (GET /api/orders/[id]) instead of the raw URL keeps the data groupable.
Never overwrite an existing instrumentation.ts; many apps already have one for something else.
On Next.js 13/14 you may need experimental.instrumentationHook: true in next.config.
Stack traces from the compiled .next output only map back to your source if the source maps sit next to the build.

Questions:

Would you accept a tool patching next.config for this, or would that be a dealbreaker?
Anyone combining AI SDK telemetry with request traces in production? How do you link an AI call to the user request that caused it?

Disclosure: I'm building Relia, a monitoring tool, and this is for its Next.js setup. I'm asking for opinions on the approach, not trying to sell anything here.