v0.preview.37

v0.preview.37 - 2026-06-10

• Update the Anthropic model catalog for the latest Claude models
• Introduce a typed LLM error taxonomy with per-kind recovery hints in error messages
• Reconnect dropped streams mid-turn and treat cancellation as a first-class turn outcome, including while a stream is being opened
• Surface content-filter stops like output-limit truncation instead of failing silently
• Fix Gemini streaming: emit thought text and signature as one thinking delta, resolve tool-response names from the dialog's tool calls, count thinking tokens in output usage, and classify recitation/blocklist stops
• Harden OpenAI-compatible streaming: accumulate tool-call deltas split across chunks, flush buffered tool calls on premature stream EOF, isolate malformed tool-call arguments, and keep the system prompt when no user message exists
• Forward max_output_tokens and temperature to OpenAI and preserve truncated output
• Propagate Anthropic message conversion errors instead of dropping messages
• Honor the HTTP-date form of Retry-After headers when rate limited