Core 0.7.21

App AI review now processes bounded source parts after the required machine checks. Submission reviews and administrator policy trials share the same planner and execution path. Completed part receipts are retained between worker claims; changed source or review evidence cannot reuse them. Findings and unresolved risks remain in the final review, and all required parts must complete before an administrator can approve. AI never approves or publishes Apps.

The Console shows review progress. Documentation, dependency lock data and validated JSON translation dictionaries are inventoried separately from executable-source review. Related source context is bounded and missing context is recorded; this is not a claim that automated review proves an App safe. Oversized individual source units or unsupported input fail explicitly instead of being truncated.

Deploy Core and the matching Agent Gateway together. No SDK, Composer lock, environment or database migration change is required. Previously published review history is retained; pending approvals require the current review revision.

App write callbacks now reference the original Core request receipt instead of repeating its verification hashes. Core still rechecks that receipt and current authority before applying a callback. Failed signing rolls back undelivered request reservations. Existing SDKs and still-valid tokens issued immediately before deployment remain compatible.