fix(assistant): use internal fetch for MCP on workers

[onmax]

Follow-up to #1304.
related vercel/ai#13449

Summary

The assistant handler creates an MCP client that fetches the same worker's /mcp endpoint. On Cloudflare Workers this triggers error 1042 (self-fetch recursion protection), so /__docus__/assistant always returns 500.

The fix creates a custom MCP transport that uses event.fetch (Nitro's localFetch) instead of event.$fetch (which maps to globalThis.$fetch → global fetch() → CF self-fetch error). On CF Workers:

• event.fetch → useNitroApp().localFetch (internal routing, no HTTP)
• event.$fetch → globalThis.$fetch → fetch() (external HTTP, blocked by CF)

External MCP servers and local dev still use the standard HTTP transport.

Reproduction

Variant	Source	Deployed worker	Result
Bug	docus-assistant-worker-repro	docus-assistant-repro.je-cf9.workers.dev	500
Fix	docus-assistant-worker-repro-fix	docus-assistant-repro-fix.je-cf9.workers.dev	✅ streaming response

Verify

Bug worker:

curl -X POST https://docus-assistant-repro.je-cf9.workers.dev/__docus__/assistant \
  -H 'content-type: application/json' \
  -d '{"id":"t","messages":[{"parts":[{"type":"text","text":"hello"}],"id":"t1","role":"user"}],"trigger":"submit-message"}'

Fixed worker:

curl -X POST https://docus-assistant-repro-fix.je-cf9.workers.dev/__docus__/assistant \
  -H 'content-type: application/json' \
  -d '{"id":"t","messages":[{"parts":[{"type":"text","text":"hello"}],"id":"t1","role":"user"}],"trigger":"submit-message"}'

Local verification:

pnpm typecheck

[vercel]

@onmax is attempting to deploy a commit to the NuxtLabs Team on Vercel.

A member of the Team first needs to authorize it.

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/create-docus@1309

npm i https://pkg.pr.new/docus@1309

commit: fa6d475