Skip to content
dbx tools

v0.9.63

Release Notes — Since v0.9.61 Includes HEAD and working-tree changes; generated churn and test-only changes excluded.

Operator Changes

  • Genie and model gateway accept mutually exclusive --model / --model-class selectors. Gateway requests can omit model or use x-dbx-tools-model-class.
  • Mastra defaults to on-demand skill search with user-scoped workspace caching and a five-minute TTL. Configure workspaceSkillRefreshTtlMs; set workspaceSkillSearch: false to restore eager catalog loading.
  • Release automation accepts --release-notes-instructions; published docs include retained release notes.

Breaking Changes

  • Graphiti removes --model, --embedder-model, and --embedder-dimensions, plus their environment settings. Use --model-class / MODEL_CLASS (default: chat-fast); embedding models and dimensions come from discovery.
  • Genie homes become profile-scoped, sharing state across models. Existing profile-model homes are not automatically reused.
  • Default model selection no longer favors GPT. Search terms such as “summarize” and “embedding” no longer imply capability classes; use explicit classes or model names.

Fixes

  • Default ranking prefers newer models within the leading family; chat searches no longer impose a balanced-class ceiling.
  • Discovery probes missing embedding dimensions; Graphiti fails explicitly when dimensions remain unavailable.
  • Shared skill caches survive request cleanup. Chat Thoughts omit Markdown tables and collapse excess whitespace.

Validation

  • 98 focused tests passed across nine files; zero failures. Version consistency and diff whitespace checks passed.
  • Full builds, Python runtime tests, and live deployment checks were not run. No files were modified.