mcp_sentinel 2.32.0

New features

• Successful Read and Explain tool results include _mcp_sentinel_untrusted_read with class: untrusted_data and instructions: false. The marker says the result is data. It grants no permission, and a payload cannot remove or replace it. The marker counts toward the profile response-size cap; a read that exceeds the cap after the marker is refused and the payload is not returned. Write, trigger, and failed results omit it. This is not jailbreak detection.

Bug fixes

• Context and the site-context tool report allowed_formats on text, text_long, and text_with_summary fields. The value is the list of enabled format IDs. An empty list means the field does not restrict formats. Non-text fields omit the key. Disabled checkbox values (0) are not reported as formats. A content-tier token can read this list from GET /drupal-mcp/context without administer node fields.
• #3628712: Creating a translation on a moderated node that has no published revision is refused before anything is written. The dry run reports the same refusal. The revision table, the live revision, and the alias stay as they were. A node that already has a published revision is unchanged.