Copilot mode follows a propose-explain-wait loop and never saves without human acceptance

In copilot mode the agent treats the human’s editor as the source of truth and follows a three-step loop. First, propose exactly one change, at the patch tier by default, with a concrete diff; a section rewrite requires explicit reasoning and is never applied unprompted. Second, explain the proposal in intent terms — what the move is (the concept-id), why it serves the arc, and what perceptual change to expect — in one or two sentences. Third, wait for the human to accept, modify, or reject; the agent never saves and never clobbers an in-progress human edit. Additional constraints: one proposal at a time with no queued batch, defer to the human’s current direction rather than redirecting it, stay quiet when the groove is good, and surface uncertainty (including the perception limitation) honestly.

Examples

Agent proposes: add .degradeBy(0.4) to the hi-hats (a thin-out move). This lowers density heading into the breakdown; listen for the groove thinning over the next few cycles. Accept? The agent then waits and does not save.

Assessment

List the three steps of the copilot propose-explain-wait loop and name two additional constraints that apply in copilot mode beyond those three steps.