#!/usr/bin/env node
// detect-skill-lifecycle.mjs — Spor B / lag-1-analog at SKILL granularity.
//
// PRODUCES ONLY REPORTS — never writes to skills/. Mirrors the lag-1 invariant:
// detection surfaces candidates; any skill-lifecycle op (merge/sanitize/retire/
// create) goes through decisions.json + operator-gate (later phases B3).
//
// Sesjon 13 (B1, first detector): OVERLAP. Two deterministic signals, combined:
//   (1) boundary-tension graph — operator-curated k1-trigger-prompts.json:
//       each out_of_domain entry is a sibling prompt tagged belongs_to=<skill>,
//       i.e. a hand-labelled "confusable neighbour". Symmetric counts = how much
//       two skills sit on each other's trigger boundary.
//   (2) df-weighted lexical trigger-surface overlap — shared content tokens
//       between two descriptions, each weighted 1/df so domain-common vocabulary
//       ("azure") contributes little and distinctive shared tokens contribute more.
//
// The eng<->infra pair (Azure-deployment boundary) is surfaced as focusPair —
// the operator-designated first target for B1.
//
// Usage:
//   node scripts/kb-eval/detect-skill-lifecycle.mjs          # human summary
//   node scripts/kb-eval/detect-skill-lifecycle.mjs --json   # machine output
//   node scripts/kb-eval/detect-skill-lifecycle.mjs --write  # persist report JSON
//
// Zero dependencies. Reuses eval.mjs extractors + kb-update atomic-write.

import { readFileSync, readdirSync, existsSync, mkdirSync } from 'node:fs';
import { join, dirname } from 'node:path';
import { fileURLToPath } from 'node:url';
import { splitFrontmatter, extractDescription } from './eval.mjs';
import { atomicWriteJson } from '../kb-update/lib/atomic-write.mjs';

const __dirname = dirname(fileURLToPath(import.meta.url));
const PLUGIN_ROOT = join(__dirname, '..', '..');
const SKILLS_DIR = join(PLUGIN_ROOT, 'skills');
const DATA_DIR = join(__dirname, 'data');
const PROMPTS_FILE = join(DATA_DIR, 'k1-trigger-prompts.json');
const OUT_FILE = join(DATA_DIR, 'skill-lifecycle-report.json');

// Operator-designated B1 focus boundary.
const FOCUS_PAIR = ['ms-ai-engineering', 'ms-ai-infrastructure'];

// Function words (no + en) of length >= 3. Tokens < 3 chars are dropped anyway,
// so this list only needs the longer connectives. Domain nouns are NOT here —
// df-weighting handles common domain vocabulary instead.
const STOPWORDS = new Set([
  'the', 'and', 'for', 'with', 'between', 'before', 'not', 'are', 'that', 'this',
  'into', 'over', 'per', 'use', 'used', 'when', 'which', 'how', 'via', 'from',
  'eller', 'som', 'til', 'med', 'mot', 'ved', 'for', 'har', 'kan', 'ikke', 'der',
  'det', 'den', 'ein', 'eit', 'sin',
  // description-format boilerplate (every skill ends with "Triggers on:")
  'triggers', 'trigger',
]);

/** Lowercase, split on non-alphanumeric, drop stopwords + tokens < 3 chars. */
export function tokenize(text) {
  return (text.toLowerCase().match(/[a-z0-9æøå]+/g) || [])
    .filter((w) => w.length >= 3 && !STOPWORDS.has(w));
}

/** Trigger surface of a description: quoted phrases + content-token set. */
export function extractTriggerSurface(description) {
  const phrases = (description.match(/"([^"]+)"/g) || []).map((p) => p.slice(1, -1));
  return { phrases, tokens: new Set(tokenize(description)) };
}

/** token -> number of skill-surfaces that contain it. */
export function buildDocumentFrequency(surfaces) {
  const df = new Map();
  for (const s of surfaces) {
    for (const t of s.tokens) df.set(t, (df.get(t) || 0) + 1);
  }
  return df;
}

/** Order-independent pair key. */
export function pairKey(a, b) {
  return [a, b].sort().join('|');
}

/** Lexical overlap between two surfaces, df-weighted. */
export function lexicalOverlap(surfaceA, surfaceB, df) {
  const shared = [];
  for (const t of surfaceA.tokens) if (surfaceB.tokens.has(t)) shared.push(t);
  shared.sort();
  const union = new Set([...surfaceA.tokens, ...surfaceB.tokens]).size;
  const jaccard = union > 0 ? shared.length / union : 0;
  let weightedScore = 0;
  for (const t of shared) weightedScore += 1 / (df.get(t) || 1);
  return {
    shared,
    jaccard: Number(jaccard.toFixed(4)),
    weightedScore: Number(weightedScore.toFixed(4)),
  };
}

/**
 * Symmetric boundary-tension matrix from the curated prompt set.
 * Counts out_of_domain entries whose belongs_to is one of the real skills
 * (controls / out-of-stack entries are ignored). Keyed by pairKey.
 */
export function boundaryTensionMatrix(promptSet) {
  const skills = Object.keys(promptSet).filter((k) => k !== '_meta');
  const skillSet = new Set(skills);
  const m = {};
  for (const s of skills) {
    for (const e of promptSet[s].out_of_domain || []) {
      const b = e && e.belongs_to;
      if (!skillSet.has(b) || b === s) continue;
      const k = pairKey(s, b);
      m[k] = (m[k] || 0) + 1;
    }
  }
  return m;
}

/**
 * Pure core: given { skill -> description } and the curated prompt set, compute
 * the overlap report section. combined = boundaryTension + weightedScore
 * (operator-grounded primary signal + distinctive-lexical corroboration).
 */
export function computeOverlapFromInputs(descriptionsBySkill, promptSet) {
  const skills = Object.keys(descriptionsBySkill).sort();
  const surfaces = {};
  for (const s of skills) surfaces[s] = extractTriggerSurface(descriptionsBySkill[s]);
  const df = buildDocumentFrequency(Object.values(surfaces));
  const tension = boundaryTensionMatrix(promptSet);

  const pairs = [];
  for (let i = 0; i < skills.length; i++) {
    for (let j = i + 1; j < skills.length; j++) {
      const a = skills[i];
      const b = skills[j];
      const key = pairKey(a, b);
      const lexical = lexicalOverlap(surfaces[a], surfaces[b], df);
      const boundaryTension = tension[key] || 0;
      const combined = Number((boundaryTension + lexical.weightedScore).toFixed(4));
      pairs.push({ pair: [a, b], key, boundaryTension, lexical, combined });
    }
  }
  // sort by combined desc, then key asc for stable ties
  pairs.sort((x, y) => y.combined - x.combined || x.key.localeCompare(y.key));

  const focusKey = pairKey(...FOCUS_PAIR);
  const focusPair = pairs.find((p) => p.key === focusKey) || null;

  return {
    method:
      'deterministic: (1) operator-curated boundary-tension (k1-trigger-prompts belongs_to), ' +
      '(2) df-weighted lexical trigger-surface overlap. combined = boundaryTension + weightedScore.',
    focusPairReason:
      'Azure-deployment boundary engineering(build) <-> infrastructure(operate) — operator-designated B1 target.',
    pairs,
    focusPair,
  };
}

/** Read the five SKILL.md descriptions from disk. */
function loadDescriptions() {
  const out = {};
  for (const e of readdirSync(SKILLS_DIR, { withFileTypes: true })) {
    if (!e.isDirectory()) continue;
    const md = join(SKILLS_DIR, e.name, 'SKILL.md');
    if (!existsSync(md)) continue;
    out[e.name] = extractDescription(splitFrontmatter(readFileSync(md, 'utf8')).frontmatter);
  }
  return out;
}

function buildReport() {
  const promptSet = JSON.parse(readFileSync(PROMPTS_FILE, 'utf8'));
  const descriptions = loadDescriptions();
  return {
    rubric: 'skill-lifecycle',
    phase: 'B1',
    note: 'Detection only — never writes to skills/. Candidates feed decisions.json + operator-gate (B3).',
    overlap: computeOverlapFromInputs(descriptions, promptSet),
  };
}

function main() {
  const args = process.argv.slice(2);
  const jsonOut = args.includes('--json');
  const doWrite = args.includes('--write');

  const report = buildReport();

  if (doWrite) {
    mkdirSync(DATA_DIR, { recursive: true });
    atomicWriteJson(OUT_FILE, report);
  }
  if (jsonOut) {
    process.stdout.write(JSON.stringify(report) + '\n');
    return;
  }

  const o = report.overlap;
  console.log(`\nSkill-livssyklus — B1 overlap-detektor (${o.pairs.length} par)\n`);
  console.log('Par (sortert på combined = grensetension + df-vektet leksikalsk):\n');
  for (const p of o.pairs) {
    const lex = p.lexical.shared.length ? p.lexical.shared.join(', ') : '—';
    const focus = p.key === pairKey(...FOCUS_PAIR) ? '  ◀ FOCUS (Azure-deployment)' : '';
    console.log(`  ${p.combined.toFixed(2).padStart(6)}  ${p.pair.join(' / ')}${focus}`);
    console.log(`          tension=${p.boundaryTension}  lex(w=${p.lexical.weightedScore}, jac=${p.lexical.jaccard})  delt: ${lex}`);
  }
  console.log(`\nFocus-par: ${o.focusPair ? o.focusPair.pair.join(' <-> ') : '(ingen)'} — ${o.focusPairReason}`);
  console.log('\n(Rapport skrives med --write til data/skill-lifecycle-report.json; aldri til skills/.)\n');
}

if (process.argv[1] && fileURLToPath(import.meta.url) === process.argv[1]) {
  main();
}