Prompts queued during a turn now survive being interrupted: they are preserved across user-initiated aborts, drained after a turn aborts itself, and edits made to the queue inside the abort window are applied rather than lost. Stopping a session now has consistent full-stop semantics across hosts Session context stays durable across aborts and hub restarts, so an interrupted session resumes with the state it had rather than a reset one Queued turns that fail are now reported as `run.failed` instead of completing silently A hung MCP server no longer takes down session creation, and stdio servers that were never configured get a 30-second initialize budget instead of blocking indefinitely Remote SSE MCP servers now surface an OAuth authorization prompt on a 401 instead of failing outright, and remote MCP supports pre-registered OAuth clients for setups where dynamic client registration isn't available LiteLLM requests route through Chat Completions instead of the Responses API, fixing calls against LiteLLM proxies Network interruptions that happen mid-stream but before any model output are retried instead of failing the turn Vertex ADC token refreshes use the configured fetch, so…