Artifact tool gains gated inherited type capabilities and prompt text variants

What

The Artifact tool, which Claude uses to publish pages, gained two experiments. Both are controlled by server-side gates (switches Anthropic can flip remotely) or by environment variables. Nothing has been read about any of these gates.

Inherited capabilities from a type:

• The publish description can append " [capabilities: declared by the Anthropic-published type this Artifact was made from; this publish cannot change them]". While that note is shown, the usual connector-grant note is left out.
• The note appears only for a type-locked Artifact with no explicit capabilities or contract, and only when the setting CLAUDE_CODE_ARTIFACT_INHERITED_TYPE_GRANT is on or, failing that, the gate tengu_swirling_squid, which falls back to false. The result is fixed once per session.
• A schema that already has watch_rail, auto_react and db_str_replace gains an optional inherited_type_grant field, and the Artifact configuration report includes inherited_type_grant: true when the feature is on. New report fields type_first_party and caps_read_not_found follow the same switch.

Text variants:

• The tool's guidance is now assembled from a table of variants, v0, v1 and v2, covering the head, deliverable, capabilities, storage, update, unread-files and never-publish sections. A variant can reorder them, for example putting capabilities last.
• v1 and v2 carry rewritten text, such as "Always author the page as .html", a browser-storage section and new rules about updating and reading before publishing.
• The variant is chosen once per session: CLAUDE_CODE_ARTIFACT_TEXT_VARIANT first; otherwise v0 when CLAUDE_CODE_SESSION_ORIGIN is claude_ai_chat; otherwise the gate tengu_cobalt_plinth_sloe, which falls back to "v0". The interactive tool now builds its prompt with this textVariant.

Why

The first experiment would let Artifacts made from an Anthropic-published type take their capabilities from that type, so a publish cannot change them. The second is a test of different wordings of the tool's instructions. The fallback for both is the existing behaviour, and the rendered prompts did not change.