v0.7.3

Added

• API Gateway Server Auto-Start (#40): Support RELAY_AI_SERVER_AUTOSTART=true (or server.autostart: true in ~/.relay-ai/config.json) to automatically launch the in-process API Gateway server (port 17645) on container/system boot without requiring manual user interaction in the Web UI. Added a toggle control in the Web UI Server tab.
• Provider Model Catalog Filters & Favorites Action Button: Added a minimum context window size filter (≥ 128k, ≥ 200k, ≥ 500k, ≥ 1M) and a Free models only checkbox to the Provider Model Catalog browser (#provider/{id}). Added an interactive + action button for each model row in full UI mode to quickly add models to Global Favorites (max 20/20) or Antigravity Favorites (max 6/6) with real-time room capacity checks and duplicate protection (hidden in API Gateway Server mode).
• Diagnostic log line (sdk: dropped user turn with unrecognized block types: [...], visible under --trace in ~/.relay-ai/logs/proxy-debug.log) for the case where a translated user turn would otherwise silently vanish instead of reaching the provider.

Fixed

• Container boot reboot loops from stale /data/ui.lock (#40): checkExistingServer now uses a 3-tier health check combining same-PID matching, process existence checking, and an HTTP health probe against /api/config. Stale lockfiles from container restarts or hard crashes are cleared automatically instead of causing infinite Docker PID 1 restart loops.
• Qwen (@ai-sdk/alibaba) no longer derails mid-session with a hallucinated "empty message"/translation tangent. Every occurrence traced back to the same trigger: a request ending on a tool result (i.e. every step of a normal agentic tool loop) — a documented qwen/DashScope function-calling quirk where the model hallucinates an empty user turn instead of summarizing the tool output. Relay now appends a minimal neutral continuation turn when a qwen request ends on a tool message, scoped narrowly to that provider and invisible to Claude Code.
• Status-line context meter no longer flickers empty→full on every turn for SDK-routed (non-Anthropic) models. Relay was seeding message_start with input_tokens: 0 and only reporting the real count once the stream finished; it now seeds a same-request estimate upfront so the meter shows a stable number from the first token.
• Hardened serializeToolResultContent against a missing tool-result content field, which previously could ship a JS undefined (not a string) into the upstream SDK call instead of an empty string.