--- status: accepted date: 2026-10-22 --- # 0005 — A definition has one shape: no prompt mode, no locks A definition carries a body, a tool selection, and a handful of settings. It does not carry a prompt mode, and it cannot lock a field against its callers. ## Context Upstream's agent files could name two things this specification does not have: - `prompt_mode: append | replace` — whether the child's system prompt embedded the parent's prompt in an `` wrapper or replaced it. - `locked: model | thinking | ...` — a declaration that a caller may not override the named fields. Both are mechanisms for the same question — "who wins?" — and the specification answers it once: the request wins, and the definition fills what the request does not name. A second mechanism would make the answer depend on which file the reader happened to open, and the specification's own precedence rules are unconditional ("a tool-call model override wins over an agent-file model"). The prompt mode goes with them: with no mode to select, the child's prompt has one shape — the inherited identity, then the project context, then the header, then the standing instructions in the standard wrapper. ## Decision Remove `prompt_mode` and `locked`. A definition's body is its standing instructions; an empty body inherits the instructions of the layer below, and a nonempty body replaces them. ## Consequences - The "twin" label, which marked an append-mode child in the widget and the transcript heading, is gone: every child adopts the parent's identity prefix now, so the label would mark everything. - Upstream's `` wrapper is the only shape. A specialist's body is now always wrapped, so it reads as the specialist's instructions rather than as the whole prompt. - The parent's *orchestration* instructions are still excluded, because upstream's inherited-identity cut already removes the session-resolved tail (ADR 0006, ADR 0008). What a child inherits is the operator's and the project's own words, not the parent's coordinator role.