宜博 / lemonboss

/**
 * Arena — 进化竞技场
 *
 * 核心选择压力来源。
 * 把基因组编译为 prompt → 用 prompt 在文本场景中推理 → Judge 评分。
 * 不需要文件系统，不需要沙盒目录，纯 LLM 推理。
 */
const evolutionQueries = require('../queries').evolutionQueries;
const { evaluateScenario } = require('../evaluator');

// ════════════════════════════════════════════════════════
// Genome → Prompt 编译
// ════════════════════════════════════════════════════════

/**
 * 将 6 维基因组编译为自然语言 prompt
 *
 * 这个 prompt 代表基因组的"表型"——它如何在环境中表现。
 */
function compileGenomeToPrompt(genome) {
  const dims = {};
  for (const key of ['mental_models', 'pattern_library', 'judgment', 'tool_fluency', 'taste', 'scar_tissue']) {
    const val = genome[key];
    dims[key] = typeof val === 'string' ? JSON.parse(val) : (val || {});
  }

  const sections = [];

  // 心智模型 → 第一性原理
  const mm = dims.mental_models;
  if (mm.ironLaws?.length) sections.push(`## 铁律\n${mm.ironLaws.map(l => `- ${typeof l === 'string' ? l : l.insight || JSON.stringify(l)}`).join('\n')}`);
  if (mm.causalChains?.length) sections.push(`## 因果推理\n${mm.causalChains.map(c => `- ${typeof c === 'string' ? c : c.model || JSON.stringify(c)}`).join('\n')}`);

  // 模式库 → 识别模式
  const pl = dims.pattern_library;
  if (pl.problemPatterns?.length) sections.push(`## 问题识别模式\n${pl.problemPatterns.map(p => `- **${p.id || p.signature || ''}**: ${p.signature || p.solution || JSON.stringify(p)}`).join('\n')}`);
  if (pl.solutionPatterns?.length) sections.push(`## 解决方案模式\n${pl.solutionPatterns.map(p => `- **${p.id || p.name || ''}**: ${p.applicableWhen || JSON.stringify(p)}`).join('\n')}`);

  // 判断力 → 决策规则
  const jd = dims.judgment;
  if (jd.decisionRules?.length) sections.push(`## 决策规则\n${jd.decisionRules.map(r => `- ${typeof r === 'string' ? r : r.decision || JSON.stringify(r)}`).join('\n')}`);
  if (jd.exclusions?.length) sections.push(`## 不适用场景\n${jd.exclusions.map(e => `- ${typeof e === 'string' ? e : JSON.stringify(e)}`).join('\n')}`);

  // 伤疤记忆 → 反模式
  const st = dims.scar_tissue;
  if (st.antiPatterns?.length) sections.push(`## 反模式（必须避开）\n${st.antiPatterns.map(a => `- ❌ ${a.pattern || JSON.stringify(a)}`).join('\n')}`);
  if (st.failures?.length) sections.push(`## 历史教训\n${st.failures.map(f => `- ${f.lesson || f.preventionRule || JSON.stringify(f)}`).join('\n')}`);

  // 品味 → 质量标准
  const ta = dims.taste;
  if (ta.qualitySignals?.length) sections.push(`## 质量标准\n${ta.qualitySignals.map(s => `- ${s.signal || (typeof s === 'string' ? s : JSON.stringify(s))}`).join('\n')}`);

  // 工具流利度 → 工具使用
  const tf = dims.tool_fluency;
  if (tf.idioms?.length) sections.push(`## 工具使用惯用法\n${tf.idioms.map(i => `- ${i.step || ''}: ${i.tools?.join(', ') || JSON.stringify(i)}`).join('\n')}`);

  return `你是一个 ${genome.skill_name} 专家 Agent。以下是你的能力配置：\n\n${sections.join('\n\n')}\n\n请基于以上能力完成后续任务。`;
}

// ════════════════════════════════════════════════════════
// 竞技场评估
// ════════════════════════════════════════════════════════

/**
 * 在 Arena 中评估单个基因组
 *
 * @param {object} genome - 基因组记录
 * @param {object[]} scenarios - Benchmark 场景数组
 * @returns {{ overall, scores }}
 */
async function evaluateGenomeInArena(genome, scenarios) {
  const prompt = compileGenomeToPrompt(genome);
  const scores = [];

  for (const scenario of scenarios) {
    try {
      const result = await evaluateScenario(
        prompt,
        scenario.scenario || scenario.description || '',
        scenario.evaluationCriteria || {},
      );
      scores.push({
        scenarioId: scenario.id,
        overall: result.overall || 0,
        dimensionScores: result.dimensionScores || {},
        reasoning: result.reasoning || '',
      });
    } catch (err) {
      console.error(`  ⚠️  Arena 场景 ${scenario.id} 评估失败: ${err.message}`);
      scores.push({ scenarioId: scenario.id, overall: 0, error: err.message });
    }
  }

  const avgOverall = scores.length > 0
    ? scores.reduce((sum, s) => sum + (s.overall || 0), 0) / scores.length
    : 0;

  return { genomeId: genome.id, overall: avgOverall, scores };
}

/**
 * 评估 Run 中所有存活基因组
 */
async function evaluateRunInArena(runId) {
  const genomes = await evolutionQueries.getGenomesByRun(runId, { aliveOnly: true });
  if (genomes.length === 0) return { ranked: [] };

  const skillName = genomes[0].skill_name;
  const benchmarks = await evolutionQueries.getBenchmarks(skillName);

  // 优先使用新的 scenario 字段，兼容旧的 description-only 格式
  const scenarios = benchmarks.map(b => {
    const criteria = typeof b.evaluation_criteria === 'string'
      ? JSON.parse(b.evaluation_criteria) : (b.evaluation_criteria || {});
    return {
      id: b.id,
      scenario: b.scenario || b.description || '',
      evaluationCriteria: criteria,
    };
  });

  if (scenarios.length === 0) {
    console.log(`  ⚠️  ${skillName} 没有 Benchmark 场景，跳过 Arena`);
    return { ranked: [] };
  }

  console.log(`🏟️  Arena: ${genomes.length} 基因组 × ${scenarios.length} 场景\n`);

  const results = [];
  for (const genome of genomes) {
    const result = await evaluateGenomeInArena(genome, scenarios);
    results.push(result);

    // 更新 overall_fitness
    await evolutionQueries.updateGenomeFitness(genome.id, {}, result.overall);
    console.log(`  ${genome.skill_name}[${genome.id.slice(0, 6)}]: ${(result.overall * 100).toFixed(1)}%`);
  }

  // 按得分排序
  results.sort((a, b) => b.overall - a.overall);
  return { ranked: results };
}

module.exports = {
  compileGenomeToPrompt,
  evaluateGenomeInArena,
  evaluateRunInArena,
};