import { clamp01, jaccard, tokenize, type ActionScore, type AgentMode, type ObservedEntity, type PageState, type SemanticAction } from "@src/shared";
import type { WorldModel } from "./WorldModel.ts";

/**
 * Information Gain Engine (§26–§27):
 *   gain = novelty × relevance × expected_entity_yield × confidence × source_quality ÷ cost
 * with penalties for already-seen / duplicate / loop / low-quality / low-relevance.
 */
export interface GainContext {
  goal: string;
  mode: AgentMode;
  world: WorldModel;
  state: PageState;
  recentActionTypes: string[]; // last N action types
  recentUrls: string[];
  stepsWithoutNewEntities: number;
  /** Learned expectations: action type → average new entities produced (from the platform model). */
  learnedYield?: Record<string, number>;
  /** Labels (type + target) of actions that failed in the last few steps — retrying them blindly is a loop in disguise. */
  recentFailures?: Set<string>;
}

export function actionKey(a: SemanticAction): string {
  return `${a.type}:${a.target_url ?? a.query ?? a.label}`;
}

const TYPE_YIELD: Record<string, number> = {
  OPEN_CHANNEL: 0.8,
  OPEN_PROFILE: 0.8,
  OPEN_PAGE: 0.75,
  OPEN_VIDEO: 0.65,
  OPEN_POST: 0.65,
  OPEN_COMMENTS: 0.55,
  OPEN_ENTITY: 0.6,
  SCROLL_DOWN: 0.6,
  SEARCH: 0.85,
  EXPAND: 0.4,
  PLAY_VIDEO: 0.3,
  SCROLL_UP: 0.05,
  BACK: 0.2,
  RETURN_TO_FEED: 0.3,
  FILTER: 0.4,
  WAIT_FOR_CONTENT: 0.1,
  END_SESSION: 0.0,
  FORWARD: 0.1,
  PAUSE_VIDEO: 0.0,
  COLLAPSE: 0.0,
};

export function relevanceOf(text: string | undefined, goal: string): number {
  if (!goal.trim()) return 0.6;
  if (!text) return 0.35;
  const g = tokenize(goal);
  const t = tokenize(text);
  if (g.size === 0 || t.size === 0) return 0.4;
  let hits = 0;
  for (const tok of g) if (t.has(tok) || [...t].some((x) => x.startsWith(tok.slice(0, 5)) && tok.length > 4)) hits++;
  const coverage = hits / g.size;
  return clamp01(0.3 + 0.7 * coverage + 0.2 * jaccard(g, t));
}

function entityOf(ctx: GainContext, a: SemanticAction): ObservedEntity | undefined {
  return a.target_ref ? ctx.state.entities.find((e) => e.ref === a.target_ref) : undefined;
}

export function scoreActions(ctx: GainContext, actions: SemanticAction[]): ActionScore[] {
  const scores: ActionScore[] = [];
  const currentRel = relevanceOf(ctx.state.summary_text.slice(0, 1500), ctx.goal);
  for (const a of actions) {
    const penalties: string[] = [];
    const e = entityOf(ctx, a);
    let novelty = 0.5;
    let relevance = 0.5;
    let confidence = 0.7;
    let source_quality = 0.7;
    let yieldExp = ctx.learnedYield?.[a.type] !== undefined ? clamp01(ctx.learnedYield[a.type]! / 10) : TYPE_YIELD[a.type] ?? 0.3;

    if (e) {
      novelty = ctx.world.novelty(e);
      relevance = relevanceOf(`${e.name ?? ""} ${e.text ?? ""} ${e.author ?? ""}`, ctx.goal);
      confidence = Math.max(0, ...e.provenance.map((p) => p.confidence));
      source_quality = e.provenance.some((p) => p.surface === "network") && e.provenance.some((p) => p.surface === "dom") ? 0.95 : 0.75;
      if (ctx.world.isVisited(e)) penalties.push("already_seen");
      if (/^(navigation|header|sidebar)/.test(e.context ?? "") && ctx.mode !== "learn") penalties.push("low_quality_source");
      if (!e.name && !e.text) penalties.push("low_confidence_entity");
      // Profiles/channels are hubs: raise yield when the goal is about people/orgs.
      if ((e.type === "channel" || e.type === "profile") && /personalit|people|person|creator|influenc|chaîne|channel|profil/i.test(ctx.goal)) yieldExp = Math.min(1, yieldExp + 0.15);
    } else {
      switch (a.type) {
        case "SCROLL_DOWN":
          relevance = currentRel;
          novelty = ctx.stepsWithoutNewEntities > 2 ? 0.2 : 0.6;
          if (ctx.recentActionTypes.slice(-4).every((t) => t === "SCROLL_DOWN") && ctx.recentActionTypes.length >= 4) penalties.push("navigation_loop");
          break;
        case "SEARCH":
          relevance = 0.9;
          novelty = ctx.recentActionTypes.includes("SEARCH") ? 0.15 : 0.9;
          if (ctx.state.classification.page_type === "SEARCH_RESULTS") penalties.push("already_seen");
          break;
        case "EXPAND":
          relevance = currentRel;
          novelty = 0.5;
          break;
        case "PLAY_VIDEO":
          relevance = currentRel;
          novelty = ctx.state.classification.page_type === "VIDEO_DETAIL" ? 0.6 : 0.3;
          break;
        case "BACK":
        case "RETURN_TO_FEED":
          relevance = 0.4;
          novelty = ctx.stepsWithoutNewEntities > 1 ? 0.6 : 0.2;
          if (ctx.recentActionTypes.slice(-2).includes(a.type)) penalties.push("navigation_loop");
          break;
        case "END_SESSION":
          relevance = 0.1;
          novelty = ctx.stepsWithoutNewEntities > 6 ? 0.9 : 0.05;
          break;
        default:
          break;
      }
    }
    if (ctx.mode === "observe" && a.type !== "SCROLL_DOWN" && a.type !== "WAIT_FOR_CONTENT" && a.type !== "END_SESSION") penalties.push("mode_observe_no_navigation");
    if (relevance < 0.35 && e) penalties.push("low_relevance");
    if (ctx.recentFailures?.has(actionKey(a))) penalties.push("recently_failed");

    let gain = (novelty * relevance * yieldExp * confidence * source_quality) / Math.max(0.25, a.cost);
    for (const p of penalties) gain *= p === "already_seen" ? 0.05 : p === "navigation_loop" ? 0.15 : p === "mode_observe_no_navigation" ? 0.02 : p === "recently_failed" ? 0.1 : 0.5;

    scores.push({ action_id: a.id, novelty, relevance, expected_entity_yield: yieldExp, confidence, source_quality, cost: a.cost, penalties, information_gain: gain });
  }
  return scores.sort((x, y) => y.information_gain - x.information_gain);
}