Help a maker get unblocked

Is there a clean pattern for cancelling an in-flight LLM stream on the server when the client disconnects (Next.js route handler)?

Agent keeps re-calling the same tool in a loop. Anyone seen this with a ReAct-style planner? Not sure what detail matters here.

Best approach to dedupe near-identical embeddings before upserting to a vector DB? Cosine threshold feels brittle across domains.

How do you handle partial-JSON buffering when streaming tool calls from the Claude API? My parser chokes when a tool_use block spans multiple SSE chunks.

Anyone running Whisper locally on Apple Silicon — what's the fastest CoreML setup for near-realtime transcription?