AI 0.1.95 / Operations
Context, tools, and recommendations
Understand how AI receives conversation context and interacts with commands from official add-ons.
Conversation context
Short context is optional. When enabled, AI receives only the configured number of recent turns for the current privacy-preserving actor. The data expires automatically and is not a durable transcript or cross-site memory. ai-reset removes the current actor’s context immediately.
Command awareness
Command awareness supplies a sanitized inventory of commands already available to the current user. It does not include hidden administrator commands, credentials, arbitrary shortcode execution, or permission to call a handler.
Official add-ons may register reviewed AI tools. The owning add-on keeps responsibility for capability checks, data minimization, and its Free/Pro boundary. Context contributed by an add-on is marked as untrusted reference data and cannot become a system instruction.
Recommendations and guarded runs
Free recommendations insert a proposed command for the user to review. Pro may run a classified read-only recommendation only when AI Pro is active, guarded runs are enabled, the optional allowlist permits it, the current user may access the command, and any required one-time browser approval succeeds.
Unknown, mutating, wallet-signing, browser-only, or capability-mismatched commands fail closed. The model’s wording is never treated as authorization.