semif-api-rocm

planner.js (3231B)

---

 "use strict";
 
 /* Shared client-side helpers for the /plan learning loop.
  *
  * Game-agnostic by design: a demo records each completed action as a transcript
  * turn, and Planner.triggered() decides when a decision result demands new
  * rules (a confident "insufficient" choice). The no-rules platformer uses this
  * today; the puzzle room can adopt the same trigger policy and wire format.
  */
 const Planner = (() => {
   const INSUFFICIENT_ID = "insufficient";
   // Plan only on a confident "I cannot decide"; a weak insufficient just loses
   // the argmax to the best real action (self-game.js handles that fallback).
   const INSUFFICIENT_THRESHOLD = 0.99;
   const TRANSCRIPT_KEEP = 24; // completed actions retained for planning
 
   const fresh = () => [];
 
   // One completed action + observed outcome, as the user turn the planner reads.
   const record = (transcript, content) => {
     transcript.push({ role: "user", content });
     while (transcript.length > TRANSCRIPT_KEEP) transcript.shift();
   };
 
   // [optionId, probability] of the server's choice, mirroring its argmax.
   const best = (result) => {
     let top = 0;
     result.probabilities.forEach((p, i) => { if (p > result.probabilities[top]) top = i; });
     return [result.option_ids[top], result.probabilities[top]];
   };
 
   // A confident "I cannot decide" demands new rules. The threshold is the
   // caller's policy: the platformer wants real confidence (0.99) before it
   // spends a planner run, while the puzzle room passes 0 — any plurality win
   // for "insufficient" plans, since it has no failure signal until the step
   // budget runs out and a wandering model must be able to ask for rules early.
   const triggered = (result, threshold = INSUFFICIENT_THRESHOLD) => {
     const [id, p] = best(result);
     return id === INSUFFICIENT_ID && p >= threshold;
   };
 
   // Composes the /plan user prompt from facts the orchestrator owns: the
   // simulation's one-line goal, a note on what triggered planning, and the
   // rules currently in effect ("" on the first plan). Game-agnostic — every
   // simulation supplies the same three facts, so no game needs its own prompt.
   const context = (goal, trigger, rules) => {
     const parts = [`Objective: ${goal}`, `Trigger: ${trigger}`];
     if (rules) {
       parts.push(
         "Previous rules — the actor failed while these were in force. " +
         "That failure indicts the rules (their facts, priorities, or " +
         "framing), not the actor's comprehension of them. Never resubmit " +
         "a reworded or lightly edited version: change the substance, or " +
         "discard the set and write a fresh one.\n" + rules);
     }
     return parts.join("\n");
   };
 
   async function request(body) {
     const response = await fetch("/plan", {
       method: "POST", headers: { "Content-Type": "application/json" },
       body: JSON.stringify(body),
     });
     const payload = JSON.parse(await response.text());
     if (!response.ok) {
       throw new Error(typeof payload.detail === "string" ? payload.detail : response.statusText);
     }
     return payload;
   }
 
   return { INSUFFICIENT_ID, INSUFFICIENT_THRESHOLD, TRANSCRIPT_KEEP,
     fresh, record, best, triggered, context, request };
 })();