EVCS/tests/advisor/run.mjs

#!/usr/bin/env node
// Tests de l'EVCS Advisor Shadow V1 : exécute le VRAI code de BUILD_PROMPT et
// VALIDATE_PARSE (sandbox vm) contre des réponses LLM simulées.
// Vérifie aussi la GARANTIE MODE OMBRE : BUILD_HA_CALLS ne touche que advisor_*/logbook.
// Intégré au push (tools/n8n-push.mjs) pour le workflow advisor.

import fs from 'fs';
import vm from 'vm';
import path from 'path';
import { fileURLToPath } from 'url';

const __dirname = path.dirname(fileURLToPath(import.meta.url));
const ROOT = path.resolve(__dirname, '../..');
const src = f => fs.readFileSync(path.join(ROOT, 'src/advisor', f), 'utf8');

const BUILD_PROMPT = src('BUILD_PROMPT.js');
const VALIDATE_PARSE = src('VALIDATE_PARSE.js');
const BUILD_HA_CALLS = src('BUILD_HA_CALLS.js');

function run(code, inputJson, nodeOutputs = {}) {
  const ctx = {
    $input: { first: () => ({ json: inputJson }) },
    $: name => ({ first: () => ({ json: nodeOutputs[name] }) }),
    console, Date, Intl, Math, Number, Object, String, JSON, Array, Boolean,
  };
  vm.createContext(ctx);
  return vm.runInContext(`(function(){\n${code}\n})()`, ctx);
}

// Snapshot simulé (soir d'été, VW branchée, SOC 80)
const snapshot = {
  data: JSON.stringify({
    advisor_enabled: 'on', advisor_agree_count: '10', advisor_diverge_count: '2',
    advisor_prev_strategy: 'auto_solar_plus',
    soc_percent: '80', battery_current_a: '-5', pv_power_w: '500', pv_power_10m_avg_w: '600',
    house_ac_power_w: '900', grid_state: 'off', acin1_available: '0',
    evcs_vehicle_connected: 'on', evcs_charging: 'off', evcs_current_a: '0',
    evcs_surplus_state: 'idle_wait_conditions', evcs_source_mode: 'Solaire',
    user_strategy: 'auto_solar_plus', allow_grid_hc_night: 'on', allow_grid_hc_midday: 'on',
    night_target_soc_actual: '35', current_max_auto_a: '22', plafond_edf_on_a: '20',
    departure_time: '08:50:00', departure_priority_enabled: 'on', required_energy_kwh: '10.5',
    today_remaining_kwh: '1.2', tomorrow_kwh: '45',
  }),
};

let failures = 0;
function check(label, cond, detail = '') {
  if (cond) console.log(`✅ ${label}`);
  else { failures++; console.log(`❌ ${label} ${detail}`); }
}

// ---------- BUILD_PROMPT ----------
const promptOut = run(BUILD_PROMPT, snapshot)[0].json;
check('BUILD_PROMPT: enabled=true', promptOut.enabled === true);
check('BUILD_PROMPT: 2 messages system+user', promptOut.messages?.length === 2 && promptOut.messages[0].role === 'system');
check('BUILD_PROMPT: contexte transporté', promptOut.context?.soc === 80 && promptOut.context?.userStrategy === 'auto_solar_plus');
check('BUILD_PROMPT: prompt < 6000 chars (modèle local)', (promptOut.messages[0].content.length + promptOut.messages[1].content.length) < 6000);

const killSwitchOut = run(BUILD_PROMPT, { data: JSON.stringify({ ...JSON.parse(snapshot.data), advisor_enabled: 'off' }) })[0].json;
check('BUILD_PROMPT: kill-switch OFF → enabled=false', killSwitchOut.enabled === false);

// ---------- VALIDATE_PARSE : 7 réponses LLM simulées ----------
const ctxNode = { BUILD_PROMPT: promptOut };
const validAdvice = '{"strategy":"auto_solar_plus","night_target_soc":35,"current_max_a":22,"ev_charge_tonight":"oui","confidence":0.85,"reason":"Forecast fort demain, HC nuit pour la VW."}';

const cases = [
  {
    label: 'réponse valide → ok',
    resp: { model: 'qwen3.6-35b', choices: [{ message: { content: validAdvice } }] },
    expect: o => o.status === 'ok' && o.advice.strategy === 'auto_solar_plus' && o.advice.nightTargetSoc === 35,
  },
  {
    label: 'balises <think> + fences markdown → ok',
    resp: { model: 'qwen3-vl', choices: [{ message: { content: `<think>je réfléchis...\nblabla</think>\n\`\`\`json\n${validAdvice}\n\`\`\`` } }] },
    expect: o => o.status === 'ok' && o.advice.evChargeTonight === 'oui',
  },
  {
    label: 'stratégie hors enum → invalid (rejet)',
    resp: { model: 'x', choices: [{ message: { content: validAdvice.replace('auto_solar_plus', 'charge_maximale') } }] },
    expect: o => o.status === 'invalid' && o.advice === null,
  },
  {
    label: 'night_target_soc 15 (< plancher 22) → invalid',
    resp: { model: 'x', choices: [{ message: { content: validAdvice.replace('"night_target_soc":35', '"night_target_soc":15') } }] },
    expect: o => o.status === 'invalid' && o.advice === null,
  },
  {
    label: 'current_max_a 40 (> 28) → invalid',
    resp: { model: 'x', choices: [{ message: { content: validAdvice.replace('"current_max_a":22', '"current_max_a":40') } }] },
    expect: o => o.status === 'invalid' && o.advice === null,
  },
  {
    label: 'texte sans JSON → invalid',
    resp: { model: 'x', choices: [{ message: { content: 'Je recommande le mode solaire.' } }] },
    expect: o => o.status === 'invalid',
  },
  {
    label: 'les deux LLM en échec → llm_offline',
    resp: { error: 'connect ECONNREFUSED 10.10.10.8:8000' },
    expect: o => o.status === 'llm_offline' && o.advice === null,
  },
];

for (const c of cases) {
  const out = run(VALIDATE_PARSE, c.resp, ctxNode)[0].json;
  check(`VALIDATE: ${c.label}`, c.expect(out), `(obtenu: ${JSON.stringify(out).slice(0, 120)})`);
}

// ---------- BUILD_HA_CALLS : garantie mode ombre ----------
const okValidated = run(VALIDATE_PARSE, cases[0].resp, ctxNode)[0].json;
const calls = run(BUILD_HA_CALLS, okValidated, ctxNode).map(i => i.json);
const ALLOWED = /^(input_text|input_number)\.evcs_advisor_/;
const violations = calls.filter(c => !(c.domain === 'logbook' || ALLOWED.test(c.data.entity_id || '')));
check('OMBRE: aucune écriture hors advisor_*/logbook', violations.length === 0, JSON.stringify(violations));
check('OMBRE: écrit strategy + reason', calls.some(c => (c.data.entity_id || '').includes('advisor_strategy')) && calls.some(c => (c.data.entity_id || '').includes('advisor_reason')));

// ── LE COMPTEUR MESURE LA CIBLE, CONTRE LE BILAN (12/09/2026) ──────────
// Il comparait la STRATEGIE, qui vaut auto_solar_plus la quasi-totalite du
// temps : 517/23, soit 95,7 % — un score qui mesurait une constante. La cible,
// seul champ cable, divergeait de 13 points du bilan sans etre comptee.
function compteur(ctx) {
  const node = { BUILD_PROMPT: { ...promptOut, context: ctx } };
  const validated = run(VALIDATE_PARSE, cases[0].resp, node)[0].json;
  const out = run(BUILD_HA_CALLS, validated, node).map(i => i.json);
  const id = k => out.some(c => (c.data.entity_id || '').includes(k));
  return { accord: id('agree_count'), divergence: id('diverge_count'), ecart: id('advisor_divergence') };
}
const base = promptOut.context;

// L'agent dit 35 (le validAdvice des cas ci-dessus).
const proche = compteur({ ...base, balanceStatus: 'ok', balanceRequiredSoc: 33 });
check('COMPTEUR: cible a 2 points du bilan → accord', proche.accord && !proche.divergence);

const loin = compteur({ ...base, balanceStatus: 'ok', balanceRequiredSoc: 48.4 });
check('COMPTEUR: cible a 13 points du bilan → divergence (le cas reel du 12/09)', loin.divergence && !loin.accord);

const sansBilan = compteur({ ...base, balanceStatus: 'degrade', balanceRequiredSoc: 48.4 });
check('COMPTEUR: bilan indisponible → on ne compte RIEN plutot que de compter faux',
  !sansBilan.accord && !sansBilan.divergence);

const offlineValidated = run(VALIDATE_PARSE, cases[6].resp, ctxNode)[0].json;
const offlineCalls = run(BUILD_HA_CALLS, offlineValidated, ctxNode).map(i => i.json);
check('OMBRE: llm_offline → statut écrit, pas de recommandation', offlineCalls.every(c => !(c.data.entity_id || '').includes('advisor_strategy')) && offlineCalls.some(c => (c.data.value || '').toString().startsWith('llm_offline')));

console.log(failures ? `\n${failures} ÉCHEC(S) — NE PAS POUSSER` : '\nADVISOR OK — tous les tests passent');
process.exit(failures ? 1 : 0);