A cross-app agent is only action-ready when order, authority, human approval, duplicate protection, and recovery are explicit.
The sample manifest declares six actions: read customer records, research a website, draft a summary, request approval, send email, and log the result. Two actions mark side effects, so at sixty pixels per declared action the bars span three hundred sixty and one hundred twenty. These are declarations rather than executions. This offline page cannot load its schema validator, reports Ajv unavailable, and disables Compile. Loading the sample still changes the editor, but no validation or external action is proved. The explanation that follows describes the written compiler rather than a successful local run. Consider a hypothetical dependency chain A then B then C. The written ordering algorithm starts with zero indegree A, removes its outgoing edge, and releases B, then C. It orders all three actions. If C also becomes a prerequisite of A, every node initially has an incoming edge. The starting queue is empty and zero actions are ordered. At one hundred pixels per ordered action the bars show three hundred and zero. The source marks an incomplete ordering as a cycle. Missing dependency identifiers instead return a separate missing-reference report. Neither outcome is a real agent execution. The sample lists four capabilities. Removing mail send leaves three, changing the capability bars from three hundred twenty to two hundred forty pixels at eighty pixels per grant. The send-email write still requests that missing capability, so the written capability gate would fail even though other grants remain. Irreversible actions also require an approval ancestor, and side effects require nonempty idempotency and rollback strings. The compiler does not prove that approval was fresh, that keys are unique, or that a rollback succeeds. A ready manifest would therefore mean passing these structural gates, not evidence of human consent or successful execution.