portal/tools/router-classifier.mjs

#!/usr/bin/env node
/**
 * Router classifier — pure regex Layer 1 + LLM Layer 2 (escalation).
 * Stage 3 of router discipline overhaul.
 *
 * Layer 1: regex по реестровым keyword/classification триггерам активных узлов.
 * Возвращает { taskType, micro, recommendedNode, confidence, source: 'regex' }.
 *
 * Layer 2 (см. classifyByLLM): Sonnet с реестром в prompt'е.
 *
 * Pure (Layer 1): read-only, никакого fs/exec/net. Caller передаёт registry.
 */

const TASK_TYPE_KEYWORDS = {
  feature: ['фич', 'feature', 'новый функционал', 'add feature'],
  planning: ['план', 'plan', 'спека', 'spec', 'roadmap'],
  bugfix: ['баг', 'bug', 'дебаг', 'debug', 'почини', 'fix', 'ошибк', 'не работает'],
  refactor: ['рефактор', 'refactor', 'почисти код', 'упрости'],
  cleanup: ['уберём', 'удали', 'remove', 'cleanup', 'dead code'],
  marketing: ['маркетинг', 'marketing', 'кампани', 'лендинг'],
  security: ['безопасност', 'security', 'уязвимост', 'vulnerability'],
  analysis: ['проанализируй', 'analysis', 'разбер', 'investigate'],
  monitoring: ['мониторинг', 'monitor', 'трейс', 'observability'],
  'memory-sync': ['запомни', 'обнови память', 'memory', 'CLAUDE.md', 'MEMORY.md'],
  question: ['что такое', 'как работает', 'почему', 'объясни', 'расскажи'],
};

const MICRO_KEYWORDS = [
  'опечатк', 'typo',
  'переименуй', 'rename',
  'удали мёртв', 'dead code',
  'формат', 'format',
  'константу', 'one constant',
];

function lower(s) { return String(s || '').toLowerCase(); }

function detectTaskType(prompt) {
  const p = lower(prompt);
  for (const [t, kws] of Object.entries(TASK_TYPE_KEYWORDS)) {
    for (const kw of kws) {
      if (p.includes(kw)) return t;
    }
  }
  return 'unknown';
}

function detectMicro(prompt) {
  const p = lower(prompt);
  return MICRO_KEYWORDS.some((kw) => p.includes(kw));
}

/**
 * Flexible keyword matching: handles RU morphology by checking if
 * - prompt contains the keyword (exact), OR
 * - keyword contains the prompt fragment (keyword starts with what's in prompt), OR
 * - prompt fragment starts with the keyword stem (first 6+ chars of keyword)
 */
function keywordMatches(promptLower, keywordLower) {
  if (promptLower.includes(keywordLower)) return true;
  // Stem match: use first 6 chars of keyword as stem (handles inflections like рассылку vs рассылка)
  if (keywordLower.length >= 6) {
    const stem = keywordLower.slice(0, -1); // drop last char for RU inflection tolerance
    if (promptLower.includes(stem)) return true;
  }
  return false;
}

function detectRecommendedNode(prompt, registry) {
  const p = lower(prompt);
  let best = { id: null, weight: 0 };

  for (const node of registry.nodes || []) {
    if (node.status !== 'active') continue;
    for (const t of node.triggers || []) {
      const w = t.weight ?? 1.0;
      if (t.keyword) {
        if (keywordMatches(p, lower(t.keyword)) && w > best.weight) {
          best = { id: node.id, weight: w };
        }
      } else if (t.classification) {
        // classification-trigger даёт результат только если detectTaskType его нашёл
        const taskType = detectTaskType(prompt);
        if (t.classification === taskType && w > best.weight) {
          best = { id: node.id, weight: w };
        }
      }
    }
  }
  return best.id;
}

// Hard keyword stems that signal a high-confidence match
const HARD_KEYWORD_STEMS = [
  'списан', 'биллинг', 'маркетинг', 'email-рассылк',
  '152-фз', 'go-live', 'фич', 'план', 'баг',
];

function computeConfidence(taskType, recommendedNode, prompt) {
  if (recommendedNode === null && taskType === 'unknown') return 0.1;
  if (recommendedNode === null) return 0.4;
  // Keyword match даёт high confidence; classification-only — medium.
  const p = lower(prompt);
  const hasHardKeyword = HARD_KEYWORD_STEMS.some((stem) => p.includes(stem));
  if (hasHardKeyword) return 0.9;
  if (taskType === 'unknown') return 0.5;
  return 0.7;
}

export function classifyByRegex(prompt, registry) {
  const taskType = detectTaskType(prompt);
  const micro = detectMicro(prompt);
  const recommendedNode = detectRecommendedNode(prompt, registry);
  const confidence = computeConfidence(taskType, recommendedNode, prompt);
  return { taskType, micro, recommendedNode, confidence, source: 'regex' };
}

// ─── Layer 2: LLM escalation ────────────────────────────────────────────────

const LLM_SYSTEM_PROMPT = `You are a router classifier for an AI coding assistant. Given a user prompt and a registry of available skills/tools (nodes), choose:
- taskType: one of {feature, planning, bugfix, refactor, cleanup, marketing, security, analysis, monitoring, memory-sync, question, unknown}
- micro: true if the task is a tiny edit (≤2 files, ≤20 lines, e.g. typo / rename / single constant)
- recommendedNode: id of the single best-matching active node, or null if nothing matches
- confidence: 0.0-1.0
- recommendedChain: id of the chain (L1-L16) if the task fits a known chain, else null
- reasoning: 1-2 sentences why

Reply with ONLY a JSON object, no prose. Example:
{"taskType":"bugfix","micro":false,"recommendedNode":"#62","confidence":0.9,"recommendedChain":null,"reasoning":"keyword 'списание' matches #62 billing-audit"}`;

export function buildLLMPrompt(prompt, registry) {
  const nodes = (registry.nodes || []).filter((n) => n.status === 'active');
  const nodeLines = nodes.map((n) => {
    const triggers = (n.triggers || [])
      .slice(0, 3)
      .map((t) => t.keyword || `cls:${t.classification}`)
      .filter(Boolean)
      .join(', ');
    return `- ${n.id} ${n.name} [${triggers}]`;
  }).join('\n');

  const chains = Object.entries(registry.chains || {})
    .map(([id, c]) => `- ${id}: ${c.name} [${(c.sequence || []).join(' → ')}]`)
    .join('\n');

  return `${LLM_SYSTEM_PROMPT}

## Available nodes
${nodeLines}

## Available chains
${chains}

## User prompt
${prompt}

Reply with JSON object only.`;
}

export function parseLLMResponse(text) {
  if (!text) return null;
  const trimmed = String(text).trim();
  // Strip ```json``` wrapper if present
  const stripped = trimmed.replace(/^```(?:json)?\s*\n?/, '').replace(/\n?```$/, '').trim();
  try {
    const parsed = JSON.parse(stripped);
    if (typeof parsed.taskType !== 'string') return null;
    return parsed;
  } catch {
    return null;
  }
}

export function shouldEscalate(regexResult) {
  if (regexResult.micro) return false;
  if (regexResult.confidence >= 0.7) return false;
  return true;
}

export async function callAnthropicAPI(prompt, { apiKey, model = 'claude-haiku-4-5-20251001', fetchImpl = fetch }) {
  const r = await fetchImpl('https://api.anthropic.com/v1/messages', {
    method: 'POST',
    headers: {
      'x-api-key': apiKey,
      'anthropic-version': '2023-06-01',
      'content-type': 'application/json',
    },
    body: JSON.stringify({
      model,
      max_tokens: 300,
      messages: [{ role: 'user', content: prompt }],
    }),
  });
  if (!r.ok) {
    throw new Error(`Anthropic API ${r.status}: ${await r.text()}`);
  }
  const data = await r.json();
  return data.content?.[0]?.text || '';
}

function hashPrompt(s) {
  let h = 0;
  for (let i = 0; i < s.length; i++) {
    h = ((h << 5) - h) + s.charCodeAt(i);
    h |= 0;
  }
  return String(h);
}

export async function classify(prompt, registry, options = {}) {
  const regexResult = classifyByRegex(prompt, registry);
  if (!shouldEscalate(regexResult)) return regexResult;

  const cache = options.cache;
  const key = hashPrompt(prompt);
  if (cache && cache.has(key)) {
    return { ...cache.get(key), source: 'cache' };
  }

  const llmCall = options.llmCall || (async () => {
    const llmPrompt = buildLLMPrompt(prompt, registry);
    const text = await callAnthropicAPI(llmPrompt, { apiKey: process.env.ANTHROPIC_API_KEY });
    return parseLLMResponse(text);
  });

  let llmResult;
  try {
    llmResult = await llmCall();
  } catch (err) {
    // LLM-down — fallback to regex result with diagnostic flag
    return { ...regexResult, llmError: err.message };
  }

  if (!llmResult) return regexResult; // unparseable — fallback

  const finalResult = { ...llmResult, source: 'llm' };
  if (cache) cache.set(key, finalResult);
  return finalResult;
}