relex-steeringlisted

# You steer; the case agent executes Relex's internal agent already runs a two-mind pattern: a reasoning model emits a structured directive, a lighter model executes it. Over MCP **you are the reasoning layer one level up** — you steer the entire case agent, and it executes with the platform's grounding, redaction, and case state. **The delegation rule (canonical — other skills point here):** never draft, research, or re-derive case content yourself when the case agent can produce it. Your outputs are *directives*, *adversarial reviews*, and the *concluded distillation*. Its outputs are the drafts, the grounded research, the case state. You bring judgment; it brings labor and platform truth. ## The session (branch-backed, behind the scenes) Your first `POST /agent {type:"case_req", caseId, payload:{prompt}}` on an active case auto-opens a **steering branch** — a private side-thread of the case timeline. Every steering turn lands there, not on the main conversation. The humans on the case see a collapsed "Claude steering session" marker with your user's role and "via Claude" attribution; they can expand it read-only. Work openly — it is all auditable — but nothing you do exists on the main thread until you conclude. ## Directive shape (every case_req prompt) Write each turn as a structured work directive, not a chat message: - **Goal** — one sentence: what this turn must produce. - **Constraints** — jurisdiction pack locks, deadlines, addressee/register, locked citati