thumbgate 1.4.3 → 1.4.4

package/scripts/multi-hop-recall.js

@@ -1,240 +0,0 @@
-#!/usr/bin/env node
-'use strict';
-
-/**
- * Multi-Hop Recall — chains related lessons across hops for deeper context.
- *
- * Single-hop: query → top N lessons (current behavior)
- * Multi-hop:  query → hop-1 lessons → extract tags/domains/rootCauses →
- *             hop-2 lessons (related) → deduplicate → ranked chain
- *
- * Inspired by Chroma Context-1's multi-hop retrieval pattern.
- * Pro-only feature — gated via requirePro('multi-hop-recall').
- *
- * @module multi-hop-recall
- */
-
-const { requirePro } = require('./pro-features');
-
-/**
- * Extract expansion terms from a set of lessons for the next hop.
- * Pulls tags, domains, rootCauses, and key phrases from whatToChange.
- */
-function extractExpansionTerms(lessons) {
-  const terms = new Set();
-
-  for (const lesson of lessons) {
-    // Tags
-    const tags = Array.isArray(lesson.tags) ? lesson.tags : safeParseTags(lesson.tags);
-    for (const tag of tags) {
-      if (tag && tag.length > 2) terms.add(tag);
-    }
-
-    // Domain
-    if (lesson.domain && lesson.domain !== 'general') {
-      terms.add(lesson.domain);
-    }
-
-    // Root cause category
-    if (lesson.rootCause) {
-      terms.add(lesson.rootCause);
-    }
-
-    // Key phrases from whatToChange (3+ char words, skip stopwords)
-    if (lesson.whatToChange) {
-      const words = lesson.whatToChange
-        .toLowerCase()
-        .replace(/[^a-z0-9\s]/g, '')
-        .split(/\s+/)
-        .filter((w) => w.length > 3 && !STOPWORDS.has(w));
-      for (const word of words.slice(0, 5)) {
-        terms.add(word);
-      }
-    }
-  }
-
-  return Array.from(terms);
-}
-
-const STOPWORDS = new Set([
-  'this', 'that', 'with', 'from', 'have', 'been', 'were', 'will',
-  'should', 'would', 'could', 'about', 'their', 'there', 'which',
-  'when', 'what', 'than', 'then', 'them', 'they', 'into', 'some',
-  'also', 'more', 'very', 'just', 'does', 'done', 'make', 'made',
-]);
-
-function safeParseTags(tags) {
-  if (Array.isArray(tags)) return tags;
-  if (typeof tags === 'string') {
-    try { return JSON.parse(tags); } catch { return []; }
-  }
-  return [];
-}
-
-/**
- * Deduplicate lessons by ID, keeping first occurrence (higher-ranked).
- */
-function deduplicateById(lessons) {
-  const seen = new Set();
-  return lessons.filter((l) => {
-    const id = l.id;
-    if (seen.has(id)) return false;
-    seen.add(id);
-    return true;
-  });
-}
-
-/**
- * Score a lesson's relevance to a set of expansion terms.
- * Higher = more relevant to the chain.
- */
-function scoreRelevance(lesson, expansionTerms) {
-  let score = 0;
-  const termSet = new Set(expansionTerms.map((t) => t.toLowerCase()));
-
-  // Tag overlap: 3 points per matching tag
-  const tags = Array.isArray(lesson.tags) ? lesson.tags : safeParseTags(lesson.tags);
-  for (const tag of tags) {
-    if (termSet.has(tag.toLowerCase())) score += 3;
-  }
-
-  // Domain match: 2 points
-  if (lesson.domain && termSet.has(lesson.domain.toLowerCase())) score += 2;
-
-  // Root cause match: 2 points
-  if (lesson.rootCause && termSet.has(lesson.rootCause.toLowerCase())) score += 2;
-
-  // Content overlap: 1 point per matching word in whatToChange
-  if (lesson.whatToChange) {
-    const words = lesson.whatToChange.toLowerCase().split(/\s+/);
-    for (const word of words) {
-      if (word.length > 3 && termSet.has(word)) score += 1;
-    }
-  }
-
-  return score;
-}
-
-/**
- * Perform multi-hop recall over a lesson database.
- *
- * @param {Function} searchFn - Search function: (query, options) => lesson[]
- *   Must accept (query, { limit, signal, domain }) and return lesson rows.
- * @param {string} query - Initial search query
- * @param {object} [options]
- * @param {number} [options.maxHops=2] - Maximum chain depth (1 = single-hop, 2+ = multi-hop)
- * @param {number} [options.hopLimit=10] - Max results per hop
- * @param {number} [options.totalLimit=15] - Max total results across all hops
- * @param {string} [options.signal] - Filter by 'positive' or 'negative'
- * @param {boolean} [options.skipProCheck=false] - Skip Pro license check (for testing)
- * @param {Function} [options.requireProFn=requirePro] - Injectable Pro gate helper for testing
- * @returns {{ results: object[], hops: object[], totalHops: number, expansionTerms: string[] }}
- */
-function multiHopRecall(searchFn, query, options = {}) {
-  const {
-    maxHops = 2,
-    hopLimit = 10,
-    totalLimit = 15,
-    signal,
-    skipProCheck = false,
-    requireProFn = requirePro,
-  } = options;
-
-  // Pro gate (unless testing)
-  if (!skipProCheck && !requireProFn('multi-hop-recall')) {
-    return { results: [], hops: [], totalHops: 0, expansionTerms: [], proRequired: true };
-  }
-
-  // Clamp hops to [1, 3] to prevent runaway chains
-  const hops = Math.max(1, Math.min(maxHops, 3));
-  const allResults = [];
-  const hopMetadata = [];
-  let expansionTerms = [];
-
-  // Hop 1: direct query search
-  const hop1Results = searchFn(query, { limit: hopLimit, signal });
-  allResults.push(...hop1Results);
-  hopMetadata.push({
-    hop: 1,
-    query,
-    resultsCount: hop1Results.length,
-    type: 'direct',
-  });
-
-  if (hops < 2 || hop1Results.length === 0) {
-    return {
-      results: deduplicateById(allResults).slice(0, totalLimit),
-      hops: hopMetadata,
-      totalHops: 1,
-      expansionTerms: [],
-    };
-  }
-
-  // Extract expansion terms from hop 1
-  expansionTerms = extractExpansionTerms(hop1Results);
-
-  if (expansionTerms.length === 0) {
-    return {
-      results: deduplicateById(allResults).slice(0, totalLimit),
-      hops: hopMetadata,
-      totalHops: 1,
-      expansionTerms: [],
-    };
-  }
-
-  // Hop 2+: search using expansion terms
-  for (let hop = 2; hop <= hops; hop++) {
-    // Build expansion query from top terms (limit to 5 to keep FTS manageable)
-    const queryTerms = expansionTerms.slice(0, 5).join(' OR ');
-    const hopResults = searchFn(queryTerms, { limit: hopLimit, signal });
-
-    // Score and sort by relevance to expansion terms
-    const scored = hopResults
-      .map((lesson) => ({ ...lesson, _hopScore: scoreRelevance(lesson, expansionTerms) }))
-      .filter((l) => l._hopScore > 0)
-      .sort((a, b) => b._hopScore - a._hopScore);
-
-    allResults.push(...scored);
-    hopMetadata.push({
-      hop,
-      query: queryTerms,
-      resultsCount: scored.length,
-      type: 'expansion',
-      termsUsed: expansionTerms.slice(0, 5),
-    });
-
-    // Extract new terms for next hop (if any)
-    if (hop < hops && scored.length > 0) {
-      const newTerms = extractExpansionTerms(scored);
-      // Only add truly new terms
-      const existingSet = new Set(expansionTerms);
-      const novel = newTerms.filter((t) => !existingSet.has(t));
-      expansionTerms = [...expansionTerms, ...novel];
-    }
-  }
-
-  // Deduplicate and cap at totalLimit
-  const deduplicated = deduplicateById(allResults).slice(0, totalLimit);
-
-  // Tag each result with its hop number
-  const hop1Ids = new Set(hop1Results.map((l) => l.id));
-  const tagged = deduplicated.map((l) => ({
-    ...l,
-    _hop: hop1Ids.has(l.id) ? 1 : 2,
-  }));
-
-  return {
-    results: tagged,
-    hops: hopMetadata,
-    totalHops: hopMetadata.length,
-    expansionTerms,
-  };
-}
-
-module.exports = {
-  multiHopRecall,
-  extractExpansionTerms,
-  scoreRelevance,
-  deduplicateById,
-  STOPWORDS,
-};

package/scripts/per-step-scoring.js

@@ -1,163 +0,0 @@
-#!/usr/bin/env node
-'use strict';
-
-/**
- * Per-Step Scoring — OpenClaw-RL inspired real-time learning signals.
- *
- * Converts audit trail entries (allow/deny/warn per tool call) into
- * +1/-1 binary scoring for DPO/KTO training data export.
- *
- * OpenClaw-RL: "Binary RLA scores +1/-1 from next-state feedback after every step."
- * ThumbGate: audit trail already captures allow/deny/warn per step.
- * This module bridges the gap — turns gate decisions into training signals.
- */
-
-const fs = require('fs');
-const path = require('path');
-const { resolveFeedbackDir } = require('./feedback-paths');
-const { ensureParentDir, readJsonl } = require('./fs-utils');
-
-function getFeedbackDir() { return resolveFeedbackDir(); }
-const SCORES_FILE = 'step-scores.jsonl';
-function getScoresPath() { return path.join(getFeedbackDir(), SCORES_FILE); }
-
-// ---------------------------------------------------------------------------
-// Score Conversion
-// ---------------------------------------------------------------------------
-
-/**
- * Convert an audit trail decision to a binary score.
- * allow = +1, warn = -0.5, deny = -1
- */
-function decisionToScore(decision) {
-  switch (decision) {
-    case 'allow': return 1;
-    case 'warn': return -0.5;
-    case 'deny': return -1;
-    default: return 0;
-  }
-}
-
-/**
- * Score a single tool call step from an audit trail entry.
- */
-function scoreStep(auditEntry) {
-  return {
-    id: `step_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`,
-    timestamp: auditEntry.timestamp || new Date().toISOString(),
-    toolName: auditEntry.toolName || 'unknown',
-    agentId: auditEntry.agentId || auditEntry.source || 'unknown',
-    decision: auditEntry.decision || 'allow',
-    score: decisionToScore(auditEntry.decision),
-    gateId: auditEntry.gateId || null,
-    context: auditEntry.message || auditEntry.context || '',
-  };
-}
-
-/**
- * Score all audit trail entries and persist to step-scores.jsonl.
- */
-function scoreAuditTrail(auditEntries) {
-  const scores = auditEntries.map(scoreStep);
-  const scoresPath = getScoresPath();
-  ensureParentDir(scoresPath);
-  for (const s of scores) fs.appendFileSync(scoresPath, JSON.stringify(s) + '\n');
-  return { scored: scores.length, scores };
-}
-
-// ---------------------------------------------------------------------------
-// DPO Pair Generation from Per-Step Scores
-// ---------------------------------------------------------------------------
-
-/**
- * Generate DPO pairs from per-step scores.
- * Groups consecutive tool calls by agent session, pairs +1 (chosen) with -1 (rejected).
- */
-function generateStepDpoPairs({ periodHours = 24 } = {}) {
-  const allScores = readJsonl(getScoresPath());
-  const cutoff = Date.now() - periodHours * 60 * 60 * 1000;
-  const recent = allScores.filter((s) => new Date(s.timestamp).getTime() > cutoff);
-
-  const positives = recent.filter((s) => s.score > 0);
-  const negatives = recent.filter((s) => s.score < 0);
-
-  const pairs = [];
-  const maxPairs = Math.min(positives.length, negatives.length);
-
-  for (let i = 0; i < maxPairs; i++) {
-    const pos = positives[i];
-    const neg = negatives[i];
-    pairs.push({
-      prompt: `Tool call: ${neg.toolName}`,
-      chosen: `${pos.toolName}: ${pos.context || 'Action allowed by gate policy'}`.slice(0, 500),
-      rejected: `${neg.toolName}: ${neg.context || 'Action blocked/warned by gate policy'}`.slice(0, 500),
-      chosenScore: pos.score,
-      rejectedScore: neg.score,
-      metadata: { chosenGate: pos.gateId, rejectedGate: neg.gateId },
-    });
-  }
-
-  return {
-    pairs,
-    pairCount: pairs.length,
-    totalPositive: positives.length,
-    totalNegative: negatives.length,
-    periodHours,
-  };
-}
-
-// ---------------------------------------------------------------------------
-// KTO (Kahneman-Tversky Optimization) Export
-// ---------------------------------------------------------------------------
-
-/**
- * Export per-step scores in KTO format.
- * KTO uses individual (prompt, completion, score) tuples — not pairs.
- */
-function exportStepKto({ periodHours = 24 } = {}) {
-  const allScores = readJsonl(getScoresPath());
-  const cutoff = Date.now() - periodHours * 60 * 60 * 1000;
-  const recent = allScores.filter((s) => new Date(s.timestamp).getTime() > cutoff);
-
-  const ktoEntries = recent.map((s) => ({
-    prompt: `Tool call: ${s.toolName}`,
-    completion: `${s.decision}: ${s.context || s.toolName}`.slice(0, 500),
-    label: s.score > 0,
-    score: s.score,
-    metadata: { gateId: s.gateId, agentId: s.agentId },
-  }));
-
-  return { entries: ktoEntries, count: ktoEntries.length, periodHours };
-}
-
-// ---------------------------------------------------------------------------
-// Stats
-// ---------------------------------------------------------------------------
-
-/**
- * Get scoring stats for a time period.
- */
-function getStepScoringStats({ periodHours = 24 } = {}) {
-  const allScores = readJsonl(getScoresPath());
-  const cutoff = Date.now() - periodHours * 60 * 60 * 1000;
-  const recent = allScores.filter((s) => new Date(s.timestamp).getTime() > cutoff);
-
-  const positive = recent.filter((s) => s.score > 0).length;
-  const negative = recent.filter((s) => s.score < 0).length;
-  const neutral = recent.filter((s) => s.score === 0).length;
-
-  const byTool = {};
-  for (const s of recent) {
-    if (!byTool[s.toolName]) byTool[s.toolName] = { positive: 0, negative: 0, total: 0 };
-    byTool[s.toolName].total++;
-    if (s.score > 0) byTool[s.toolName].positive++;
-    else if (s.score < 0) byTool[s.toolName].negative++;
-  }
-
-  return { total: recent.length, positive, negative, neutral, byTool, periodHours };
-}
-
-module.exports = {
-  decisionToScore, scoreStep, scoreAuditTrail,
-  generateStepDpoPairs, exportStepKto, getStepScoringStats, getScoresPath,
-};