feat(agent-cache/ai): cache streaming LLM responses via wrapStream

[amitkojha05]

Summary

The agent-cache Vercel AI SDK adapter currently implements wrapGenerate for non-streaming calls. Streaming responses are not cached, requiring callers to manually accumulate streamed content and invoke store() themselves.

This PR adds support for wrapStream.

On a cache miss, the upstream stream is tee'd:

• One branch flows to the caller unchanged.
• The second branch accumulates text-delta chunks and stores the final response via cache.llm.store() on finish.

On a cache hit:

• doStream is skipped entirely.
• The cached response is replayed as a single text-delta event followed by a finish event.
• The finish event includes:

providerMetadata: {
  agentCache: {
    hit: true,
  },
}

Tool-call streams are not cached, matching existing wrapGenerate behavior.

Additional guarantees:

• Store failures never affect the caller-facing stream.
• check() failures fall through to the upstream provider.
• Fully backward compatible; existing wrapGenerate consumers are unaffected.

Changes

adapters/ai.ts

• Added synthesizeStreamFromCached()
• Added accumulateAndStore()
• Added wrapStream to the middleware return object
• No changes to existing wrapGenerate behavior

adapters/__tests__/ai.test.ts

Added 5 new tests:

1. Cache miss → cache hit flow
2. providerMetadata included on cache hits
3. Tool-call stream bypass
4. Upstream error propagation
5. Store failure does not break caller stream

CHANGELOG.md

• Added v0.7.0 entry

README.md

• Updated Vercel AI SDK documentation section

Testing

pnpm --filter @betterdb/agent-cache test
pnpm --filter @betterdb/agent-cache build

[amitkojha05]

Hey @KIvanow and @jamby77 ,Please review this PR ;would love to work on your feedback