v0.preview.55

v0.preview.55 - 2026-07-09

• Add offline-mode attachment to existing local llama-compatible servers, with broader localhost detection, /v1/models validation, retryable attach errors, UTF-8-safe input, and corrected external-server state handling
• Add Antix Claude Fable 5 support
• Add Grok 4.5 to the xAI and OpenRouter model catalogs
• Fix OAuth provider login started from /providers so the hidden picker no longer captures input or makes the TUI appear frozen
• Fix Kimi thinking request compatibility by omitting unsupported thinking.type = "enabled" while preserving explicit disabled requests
• Improve chat composer wrapping so English words wrap at boundaries while long tokens, CJK text, cursor movement, and resizes stay aligned
• Normalize ambient thinking phrases to sentence case instead of shouting in all caps
• Compact sessions before the next turn starts to reduce turn-lifecycle and context-management edge cases