rpcs1.dev

Use this server for three categories of work. Category 1 — Agent tuning: Call recommend_agent_configuration when designing, tuning, or diagnosing an AI agent. Category 2 — Translation: Call interpret when a user says something ambiguous. Call normalize when input is fragmented. Call rewrite when text needs tone adjustment. Category 3 — Translation Bridge (per-user adaptation): Call calibrate_profile once to build the user’s ReceiverProfile from five quick questions (store the returned JSON in the user’s notes/memory for reuse). Then, on EVERY reply to that user, call render_reply with your draft and their profile, and apply the returned instructions before answering. When the user’s request is ambiguous, call prepare_prompt with their text and profile to recover intent before acting. This is the loop: prepare_prompt on the way in, render_reply on the way out. When the user’s request could mean several different things, call route_intent — and YOU propose the candidate readings: 3–7 short hypotheses covering the plausible interpretations of THIS message (including likely-typo readings and idiom-vs-literal readings), with your own likelihood estimates. The built-in default set is a generic product-intent starter, not an interpretation engine — passing it for arbitrary sentences yields uniform confusion over irrelevant options. Division of labor: the model GENERATES the branches; the deterministic router HOLDS them, scores the ambiguity, and decides commit / present options / clarify. route_intent is the AUTHORITY on commit-vs-clarify; prepare_prompt supplies the canonical translation and recovered entities — when the two disagree on whether to ask, follow route_intent.