v16.3.0

Ordinary scoping questions now go through each host's built-in question control — AskUserQuestion in Claude Code, native questions in Codex — as one batched ask with the recommendation first. The Attune widget is reserved for the constructs those controls cannot carry (ranking, triage boards, number and date fields), on widget-capable hosts only. Automatic prompt refinement asks for material missing details up front; attune config set prompt_refinement false opts out user-wide and a leading [no-refine] skips one turn.

The memory layer stops being silent about what it is. attune memory status and a new attune doctor line name the resolved backend, transport, and reachability, and a first-run choice (attune memory use auto|file|redis) is persisted and honored, so the file tier never probes for a Redis it was told not to use.

Also in 16.3.0: the premium tier moves to Claude Fable 5.1; attune doctor becomes a full install diagnostic; cross-review runs on a verified Claude subscription without API spend; surface-parity inventory and registry gates (Task 1B increments 1–2); attune-forms floor 0.15.0; the MCP server's PyPI version check is one thread per process (ATTUNE_VERSION_CHECK=0 disables it); and register classes C3 and C4b are closed.

Pending, stated plainly: prompt refinement's installed-desktop validation and the context-routed form runtime's production activation remain outstanding. The MCP route stays closed until a verified runtime is installed.

Full changelog: https://github.com/Smart-AI-Memory/attune-ai/blob/v16.3.0/CHANGELOG.md