[rami-hatoum]

Records said who did something by the caller's id alone, an opaque identifier where a runtime signs people in, and nothing told an agent whom its connection acts as. A person who had just joined an org asked their agent what the org held and was told "you have one playbook, created this morning, and a connection you added an hour ago", when another member had created both.

A caller now has a name beside its id: a signed-in person's as the name claim of their token, which a token must now hold; a key's id; the local developer's account name on the computer, or local where the computer names none; and a playbook's id where a playbook acts as itself. Every field that says who did something names the caller by id and name: a playbook's and a definition's created_by, a connection's added_by and connected_by, a run's started_by, an answer's answered_by, and a prediction version's trained_by, approved_by and refused_by. The fact that makes each record carries the caller in its data, as recorded at the time; the ledger's own by stays the id. A run a workflow's call starts names the caller who started the run.

The instructions of an MCP connection open with whom it acts as, by name and id, and in which org, so an agent reading records tells "you" from the others. The rest of the instructions is worded tighter, every instruction kept, so the whole stays within its 2,000 characters, which startup checks. The OpenAPI document, the READMEs and the reference pages describe each field as the caller who did it, by id and name. A ledger written before this change holds facts of the old shape and is not read; the product has no stored data to carry, and a hosted runtime moves its objects when it takes this release.