Skip to main content

Built-in skills are copies that follow the server until an organization edits them

How agents work in Mission Control (where they work, handing a task on, asking for input) was hardcoded into the opening prompt, so an organization could not make it its own. We decided that the server ships such guidance as built-in skills, and that every organization gets its own copy of each: an ordinary skill that it can edit, turn off or delete like any other. A copy starts with the built-in's mode and reach: guidance every agent needs (working-in-mission-control) is global and ambient, while know-how only some roles need (writing-roles, writing-skills) is triggered and given to those roles by the organization. A copy records which built-in it is and which version of it it was taken from. While the organization has not edited it, the server keeps it at its own version, syncing when the server starts and when an organization is created; once edited, it is the organization's and is left alone, showing when the server's version has moved on, until someone resets it. Built-ins are templates (ADR 0008), so one skill covers a task with or without a repository.

Considered Options​

  • Built-ins live only in code; organizations override one with a skill of the same slug: rejected; turning one off or editing it would need state of its own, separate from skills, and everything that already works on skills (revisions, enabling, global, the editors) would need a second path.
  • Copy the defaults into each organization once, at creation: rejected; copies would never get fixes, and agents' knowledge of the tools would drift from the tools.
  • Keep the harness guidance in code, out of reach: rejected; organizations should be able to shape how their agents work, and the copy model keeps the server's version one reset away.

Consequences​

  • Each sync change is a revision with no actor (Mission Control); a reset is a revision by whoever reset it, and the organization's edits stay in the history.
  • A deleted copy is not made again, and a skill the organization made with a built-in's slug is theirs: no copy is made beside it.
  • Only what "How you work" said has moved so far, as the skill working-in-mission-control. Facts about the task (its history, handoff, pull requests, who can take it, submodules) stay in the prompt, since they are data rather than guidance. Turning the skill off leaves agents without instructions for handing tasks on, so its description says to edit it instead.
  • In the opening prompt the guidance now sits with the other ambient skills, before the task, rather than at the end.