thumbgate 1.28.2 → 1.28.4

package/scripts/lesson-retrieval.js

@@ -0,0 +1,485 @@
+#!/usr/bin/env node
+'use strict';
+
+/**
+ * Per-action lesson retrieval.
+ * v3: bi-encoder retrieval → cross-encoder reranking
+ *
+ * Stage 1 (bi-encoder): score all memories independently using token overlap,
+ *   bigram Jaccard, tool-name matching, and recency decay.  Retrieve top-50.
+ * Stage 2 (cross-encoder): rerank the top-50 candidates by computing a
+ *   field-weighted BM25 score that processes (query, lesson) jointly, then
+ *   blend with the original bi-encoder score.  Return top-maxResults.
+ */
+
+const RECENCY_DECAY_DAYS = 30;
+const RERANK_CANDIDATE_POOL = 50; // bi-encoder retrieves this many; reranker picks topK
+
+function retrieveRelevantLessons(toolName, actionContext, options = {}) {
+  const { maxResults = 5, feedbackDir } = options;
+  const { getFeedbackPaths, readJSONL } = require('./feedback-loop');
+  const { rerankLessons } = require('./lesson-reranker');
+  const pathMod = require('path');
+  const paths = feedbackDir
+    ? { MEMORY_LOG_PATH: pathMod.join(feedbackDir, 'memory-log.jsonl') }
+    : getFeedbackPaths();
+
+  const memories = readJSONL(paths.MEMORY_LOG_PATH, { maxLines: 200 });
+  if (memories.length === 0) return [];
+
+  const actionSig = buildActionSignature(toolName, actionContext);
+
+  // Stage 1 — bi-encoder: score all memories independently, take top-50 candidates
+  const candidates = memories
+    .map((mem) => ({
+      ...mem,
+      relevanceScore: scoreRelevance(mem, toolName, actionContext, actionSig),
+    }))
+    .filter((m) => m.relevanceScore > 0.1)
+    .sort((a, b) => b.relevanceScore - a.relevanceScore)
+    .slice(0, RERANK_CANDIDATE_POOL);
+
+  if (candidates.length === 0) return [];
+
+  // Stage 2 — cross-encoder reranker: rerank candidates by joint (query, lesson) score
+  const reranked = rerankLessons(actionContext, candidates, {
+    topK: maxResults,
+    toolName,
+  });
+
+  // Stage 3 (opt-in) — Memora-style nucleus stop: trim the low-mass tail so a
+  // dominant lesson isn't padded out to maxResults. No-op unless topP < 1.
+  const deduped = dedupeSupersededLessons(reranked);
+  const selected = filterTopP(deduped, resolveTopP(options), { minKeep: options.minKeep });
+
+  return selected.map((m) => ({
+    id: m.id,
+    title: m.title,
+    content: m.content,
+    signal: m.tags?.includes('negative') ? 'negative' : 'positive',
+    rule: m.structuredRule || null,
+    relevanceScore: m.rerankedScore ?? m.relevanceScore,
+    timestamp: m.timestamp,
+  }));
+}
+
+/**
+ * Reciprocal Rank Fusion — merge several ranked id-lists into one ranking.
+ *
+ * RRF is scale-free: it fuses on rank position, not raw scores, so the lexical
+ * (BM25-ish, 0..~1.5) and dense (cosine, -1..1) rankers combine without any
+ * normalization. score(id) = Σ 1/(k + rank), rank starting at 1. k=60 is the
+ * value from the original Cormack et al. paper and the de-facto standard.
+ *
+ * @param {string[][]} rankedLists - each inner array is ids in descending relevance
+ * @param {object} [options]
+ * @param {number} [options.k=60]
+ * @returns {Array<{id:string, score:number}>} fused ids, descending
+ */
+function reciprocalRankFusion(rankedLists = [], options = {}) {
+  const k = Number.isFinite(options.k) ? options.k : 60;
+  const scores = new Map();
+  for (const list of rankedLists) {
+    if (!Array.isArray(list)) continue;
+    list.forEach((id, index) => {
+      if (id === undefined || id === null) return;
+      const rank = index + 1;
+      scores.set(id, (scores.get(id) || 0) + 1 / (k + rank));
+    });
+  }
+  return [...scores.entries()]
+    .map(([id, score]) => ({ id, score }))
+    .sort((a, b) => b.score - a.score);
+}
+
+function loadMemories(feedbackDir) {
+  const { getFeedbackPaths, readJSONL } = require('./feedback-loop');
+  const pathMod = require('path');
+  const paths = feedbackDir
+    ? { MEMORY_LOG_PATH: pathMod.join(feedbackDir, 'memory-log.jsonl') }
+    : getFeedbackPaths();
+  return readJSONL(paths.MEMORY_LOG_PATH, { maxLines: 200 });
+}
+
+function shapeLesson(m) {
+  return {
+    id: m.id,
+    title: m.title,
+    content: m.content,
+    signal: m.tags?.includes('negative') ? 'negative' : 'positive',
+    rule: m.structuredRule || null,
+    relevanceScore: m.rerankedScore ?? m.relevanceScore,
+    timestamp: m.timestamp,
+  };
+}
+
+/**
+ * Hybrid (dense + sparse) per-action lesson retrieval — the async counterpart of
+ * retrieveRelevantLessons. Used by the async gate path (gates-engine runAsync).
+ *
+ * Pipeline: lexical ranking ⊕ dense (embedding) ranking → Reciprocal Rank Fusion
+ * → cross-encoder rerank → top-K. Dense recall surfaces past mistakes that share
+ * no keywords with the action (paraphrase/synonym) — the value lexical alone misses.
+ *
+ * HONEST DEGRADATION: if no real embedder is available, or embedding errors, this
+ * returns the pure-lexical result (identical to retrieveRelevantLessons). Never
+ * fabricates semantics.
+ */
+async function retrieveRelevantLessonsAsync(toolName, actionContext, options = {}) {
+  const { maxResults = 5, feedbackDir } = options;
+
+  let embeddingIndex;
+  try {
+    embeddingIndex = require('./lesson-embedding-index');
+  } catch {
+    return retrieveRelevantLessons(toolName, actionContext, options);
+  }
+
+  // No real embedder → degrade to lexical (no fake vectors, no regression).
+  if (!options.embedder && !embeddingIndex.isEmbedderAvailable()) {
+    return retrieveRelevantLessons(toolName, actionContext, options);
+  }
+
+  const memories = loadMemories(feedbackDir);
+  if (memories.length === 0) return [];
+
+  const actionSig = buildActionSignature(toolName, actionContext);
+
+  // Sparse: score every memory, keep those with any lexical signal.
+  const lexicalScored = memories
+    .map((mem) => ({ ...mem, relevanceScore: scoreRelevance(mem, toolName, actionContext, actionSig) }))
+    .filter((m) => m.relevanceScore > 0.1)
+    .sort((a, b) => b.relevanceScore - a.relevanceScore);
+  const lexicalRanked = lexicalScored.slice(0, RERANK_CANDIDATE_POOL).map((m) => m.id);
+
+  // Check if any lexical match is conclusive (exact/regex match on structured rule or high relevance)
+  let conclusive = false;
+  for (const candidate of lexicalScored) {
+    if (candidate.relevanceScore >= 0.85) {
+      conclusive = true;
+      break;
+    }
+    const rule = candidate.structuredRule;
+    if (rule) {
+      const cond = String(rule.trigger?.condition || rule.if || '').trim().toLowerCase();
+      if (cond.length >= 3) {
+        if (actionContext.toLowerCase().includes(cond)) {
+          conclusive = true;
+          break;
+        }
+        try {
+          const regex = new RegExp(cond, 'i');
+          if (regex.test(actionContext)) {
+            conclusive = true;
+            break;
+          }
+        } catch {}
+      }
+    }
+  }
+
+  if (conclusive) {
+    // Short-circuit: skip embedding/dense search completely
+    const { rerankLessons } = require('./lesson-reranker');
+    const reranked = rerankLessons(actionContext, lexicalScored.slice(0, RERANK_CANDIDATE_POOL), { topK: maxResults, toolName });
+    return filterTopP(dedupeSupersededLessons(reranked), resolveTopP(options), { minKeep: options.minKeep }).map(shapeLesson);
+  }
+
+  // WHERE-clause pruning: filter memories before vector search to only include
+  // memories relevant to the current toolName or context.
+  const prunedMemories = memories.filter((mem) => {
+    // 1. If lexical score is non-trivial, keep it.
+    const score = scoreRelevance(mem, toolName, actionContext, actionSig);
+    if (score > 0.1) return true;
+
+    // 2. Otherwise, check tool compatibility: if toolsUsed is specified, it must contain our tool
+    const memTools = mem.metadata?.toolsUsed || [];
+    if (memTools.length > 0 && !memTools.some(t => t.toLowerCase() === toolName.toLowerCase())) {
+      return false;
+    }
+    const ruleTools = mem.structuredRule?.metadata?.toolsUsed || [];
+    if (ruleTools.length > 0 && !ruleTools.some(t => t.toLowerCase() === toolName.toLowerCase())) {
+      return false;
+    }
+    return true;
+  });
+
+  // Dense: rank the pruned corpus by embedding similarity (cached vectors).
+  let semanticRanked = [];
+  if (prunedMemories.length > 0) {
+    try {
+      const dense = await embeddingIndex.semanticRank(actionContext, prunedMemories, {
+        feedbackDir,
+        embedder: options.embedder,
+      });
+      semanticRanked = dense.slice(0, RERANK_CANDIDATE_POOL).map((d) => d.id);
+    } catch {
+      // Embedding failed at runtime → fall back to pure lexical.
+      return retrieveRelevantLessons(toolName, actionContext, options);
+    }
+  }
+
+  // Fuse. Candidate pool is the union — dense can introduce lessons lexical missed.
+  const fused = reciprocalRankFusion([lexicalRanked, semanticRanked]);
+  if (fused.length === 0) return [];
+
+  const byId = new Map(memories.map((m) => [m.id, m]));
+  const lexById = new Map(lexicalScored.map((m) => [m.id, m.relevanceScore]));
+  const topFusedScore = fused[0].score || 1;
+
+  const candidates = fused
+    .slice(0, RERANK_CANDIDATE_POOL)
+    .map((entry) => {
+      const mem = byId.get(entry.id);
+      if (!mem) return null;
+      // Carry a relevanceScore the cross-encoder can blend against. Prefer the
+      // lexical score when present; otherwise use the normalized fusion score so
+      // dense-only candidates still rank sensibly.
+      const relevanceScore = lexById.has(entry.id)
+        ? lexById.get(entry.id)
+        : entry.score / topFusedScore;
+      return { ...mem, relevanceScore };
+    })
+    .filter(Boolean);
+
+  if (candidates.length === 0) return [];
+
+  const { rerankLessons } = require('./lesson-reranker');
+  const reranked = rerankLessons(actionContext, candidates, { topK: maxResults, toolName });
+  return filterTopP(dedupeSupersededLessons(reranked), resolveTopP(options), { minKeep: options.minKeep }).map(shapeLesson);
+}
+
+function buildActionSignature(toolName, actionContext) {
+  const toolLower = (toolName || '').toLowerCase();
+  const contextLower = (actionContext || '').toLowerCase();
+  const sigPaths = extractPaths(actionContext);
+  const tokens = tokenize(contextLower);
+  const ngramSet = textBigrams(contextLower);
+  return { toolLower, contextLower, paths: sigPaths, tokens, ngramSet };
+}
+
+function textBigrams(text) {
+  const normalized = (text || '')
+    .toLowerCase()
+    .replace(/[^a-z0-9\s]/g, ' ')
+    .replace(/\s+/g, ' ')
+    .trim();
+  const set = new Set();
+  for (let i = 0; i < normalized.length - 1; i++) {
+    set.add(normalized.slice(i, i + 2));
+  }
+  return set;
+}
+
+function bigramJaccard(setA, setB) {
+  if (setA.size === 0 && setB.size === 0) return 1;
+  if (setA.size === 0 || setB.size === 0) return 0;
+  let intersection = 0;
+  for (const item of setA) {
+    if (setB.has(item)) intersection++;
+  }
+  const union = setA.size + setB.size - intersection;
+  return union === 0 ? 0 : intersection / union;
+}
+
+function scoreRelevance(memory, toolName, actionContext, actionSig) {
+  const sig = actionSig || buildActionSignature(toolName, actionContext);
+  let score = 0;
+
+  const memText = ((memory.title || '') + ' ' + (memory.content || '') + ' ' + (memory.tags || []).join(' ')).toLowerCase();
+
+  if (memory.metadata?.toolsUsed?.some((t) => t.toLowerCase() === sig.toolLower)) score += 0.4;
+  if (memText.includes(sig.toolLower)) score += 0.2;
+
+  const memPaths = memory.metadata?.filesInvolved || extractPaths(memText);
+  const pathOverlap = sig.paths.filter((p) =>
+    memPaths.some((mp) => mp.includes(p) || p.includes(mp)),
+  );
+  if (pathOverlap.length > 0) score += 0.3;
+
+  const memTokens = tokenize(memText);
+  const overlap = sig.tokens.filter((t) => memTokens.includes(t));
+  score += Math.min(overlap.length * 0.05, 0.3);
+
+  // Fuzzy n-gram matching (only when there is already signal)
+  if (score > 0) {
+    const memBigrams = textBigrams(memText);
+    const fuzzyScore = bigramJaccard(sig.ngramSet, memBigrams);
+    score += fuzzyScore * 0.2;
+  }
+
+  if (memory.tags?.includes('negative')) score += 0.1;
+
+  if (memory.timestamp) {
+    const ageMs = Date.now() - new Date(memory.timestamp).getTime();
+    const ageDays = ageMs / (1000 * 60 * 60 * 24);
+    const decay = Math.max(0, 1 - ageDays / RECENCY_DECAY_DAYS);
+    score *= 0.5 + 0.5 * decay;
+  }
+
+  if (memory.structuredRule) score += 0.15;
+
+  return score;
+}
+
+function extractPaths(text) {
+  return [...new Set((text || '').match(/(?:src\/|scripts\/|tests\/)[^\s,)'"<>]+/g) || [])];
+}
+
+function tokenize(text) {
+  return (text || '').split(/[\s.,;:!?()\[\]{}"'`]+/).filter((t) => t.length > 3);
+}
+
+/**
+ * Resolve the nucleus (top-P) cutoff for a retrieval call. Precedence:
+ *   options.topP (finite, 0 < p <= 1)  →  env THUMBGATE_RETRIEVAL_TOP_P  →  1.0 (off).
+ * 1.0 is a no-op (returns the full reranked top-K), so retrieval behaviour is
+ * unchanged unless an operator explicitly opts in.
+ */
+function resolveTopP(options = {}) {
+  const fromOption = Number(options && options.topP);
+  if (Number.isFinite(fromOption) && fromOption > 0 && fromOption <= 1) return fromOption;
+  const fromEnv = Number(process.env.THUMBGATE_RETRIEVAL_TOP_P);
+  if (Number.isFinite(fromEnv) && fromEnv > 0 && fromEnv <= 1) return fromEnv;
+  return 1.0;
+}
+
+/**
+ * Nucleus (top-P) filter over a ranked lesson list — the lightweight realization of
+ * Memora's "policy-guided retriever decides when to stop" (ICML 2026). Instead of a
+ * fixed top-K, keep the smallest prefix whose NORMALIZED relevance mass covers `topP`.
+ *
+ * Scores are normalized into a probability distribution first, so the cutoff is
+ * scale-free: it behaves the same on raw bi-encoder scores (0..~1.5) and on blended
+ * rerank scores. When one or two lessons dominate the distribution, fewer lessons are
+ * returned (less context stuffed into the prompt, fewer tokens); when relevance is
+ * spread out, more are kept.
+ *
+ * Guarantees:
+ *   - never returns more lessons than the input (purely subtractive);
+ *   - always returns at least `minKeep` (default 1) when the input is non-empty;
+ *   - topP >= 1 (or non-positive/NaN) is a no-op: returns the input, order preserved.
+ *
+ * @param {Array<object>} lessons - ranked lessons (each with rerankedScore or relevanceScore)
+ * @param {number} [topP=1.0] - cumulative normalized mass to cover, in (0, 1]
+ * @param {object} [options]
+ * @param {number} [options.minKeep=1] - hard floor on lessons returned
+ * @returns {Array<object>} the retained prefix, highest score first
+ */
+function filterTopP(lessons, topP = 1.0, options = {}) {
+  if (!Array.isArray(lessons) || lessons.length === 0) return [];
+  const minKeep = Math.max(1, Math.floor(options && options.minKeep) || 1);
+  if (!(topP > 0) || topP >= 1) return lessons.slice();
+
+  const scoreOf = (l) => {
+    const s = l && (l.rerankedScore ?? l.relevanceScore);
+    return Number.isFinite(s) && s > 0 ? s : 0;
+  };
+  const sorted = [...lessons].sort((a, b) => scoreOf(b) - scoreOf(a));
+  const total = sorted.reduce((sum, l) => sum + scoreOf(l), 0);
+  if (total <= 0) return sorted.slice(0, Math.min(minKeep, sorted.length));
+
+  let cumulative = 0;
+  const kept = [];
+  for (const lesson of sorted) {
+    kept.push(lesson);
+    cumulative += scoreOf(lesson) / total;
+    if (kept.length >= minKeep && cumulative >= topP) break;
+  }
+  return kept;
+}
+
+/**
+ * Drop superseded / contradictory lessons before final selection.
+ *
+ * Retrieval can surface two lessons on the SAME topic at near-equal relevance —
+ * a stale one and a newer correction. Handing an agent contradictory guidance
+ * ("never force-push" AND "force-push is fine") poisons its decision and wastes
+ * tokens — the "context poisoning" failure mode. This collapses same-topic
+ * lessons on the reranked list:
+ *   - same topic + SAME signal     → duplicate: keep the higher-ranked, drop rest
+ *   - same topic + OPPOSITE signal → contradiction: keep the most RECENT (it
+ *                                    supersedes), drop the older
+ *
+ * "Same topic" = identical/near-identical structured-rule trigger (the strong
+ * signal), else high character-bigram similarity of title+content. Conservative
+ * by design: distinct lessons are never merged, and survivor order is preserved.
+ * Affects only lesson RETRIEVAL — never the hard gate rules.
+ *
+ * @param {Array<object>} lessons - reranked lessons/memories, best-first
+ * @param {object} [options]
+ * @param {number} [options.similarityThreshold=0.82] - bigram-Jaccard cutoff for "same topic"
+ * @returns {Array<object>} deduped list, order preserved
+ */
+function dedupeSupersededLessons(lessons, options = {}) {
+  if (!Array.isArray(lessons) || lessons.length <= 1) {
+    return Array.isArray(lessons) ? lessons.slice() : [];
+  }
+  const threshold = Number.isFinite(options.similarityThreshold) ? options.similarityThreshold : 0.82;
+  const signalOf = (x) => x.signal || (x.tags?.includes('negative') ? 'negative' : 'positive');
+  const topicSignature = (x) => {
+    const rule = x.rule || x.structuredRule;
+    const cond = rule && (rule.trigger?.condition || rule.if);
+    if (cond && String(cond).trim().length >= 3) return String(cond).toLowerCase().trim();
+    return `${x.title || ''} ${x.content || ''}`.toLowerCase().trim();
+  };
+  const timeOf = (x) => {
+    const t = x.timestamp ? new Date(x.timestamp).getTime() : NaN;
+    return Number.isFinite(t) ? t : -Infinity;
+  };
+  const sigs = lessons.map(topicSignature);
+  const grams = sigs.map((s) => textBigrams(s));
+  const sameTopic = (i, j) => {
+    if (!sigs[i] || !sigs[j]) return false;
+    if (sigs[i] === sigs[j]) return true;
+    return bigramJaccard(grams[i], grams[j]) >= threshold;
+  };
+
+  const kept = []; // indices into `lessons`, best-first
+  for (let i = 0; i < lessons.length; i++) {
+    let collapsed = false;
+    for (let k = 0; k < kept.length; k++) {
+      const j = kept[k];
+      if (!sameTopic(i, j)) continue;
+      if (signalOf(lessons[i]) === signalOf(lessons[j])) {
+        collapsed = true; // duplicate → keep the already-kept (higher-ranked)
+      } else if (timeOf(lessons[i]) > timeOf(lessons[j])) {
+        kept[k] = i; // contradiction → newer supersedes, take the slot
+        collapsed = true;
+      } else {
+        collapsed = true; // contradiction, incoming is older → drop it
+      }
+      break;
+    }
+    if (!collapsed) kept.push(i);
+  }
+  return kept.map((idx) => lessons[idx]);
+}
+
+function calculateRetrievalEntropy(lessons) {
+  if (!Array.isArray(lessons) || lessons.length === 0) return 0;
+  let pW = 0, nW = 0, tW = 0;
+  for (const l of lessons) {
+    const w = l.relevanceScore || 0.1;
+    if (l.signal === "positive") pW += w; else nW += w;
+    tW += w;
+  }
+  if (tW === 0) return 0;
+  const pPos = pW / tW, pNeg = nW / tW;
+  const entropy = (pPos > 0 ? -pPos * Math.log2(pPos) : 0) + (pNeg > 0 ? -pNeg * Math.log2(pNeg) : 0);
+  return Number(entropy.toFixed(4));
+}
+
+module.exports = {
+  retrieveRelevantLessons,
+  retrieveRelevantLessonsAsync,
+  reciprocalRankFusion,
+  scoreRelevance,
+  buildActionSignature,
+  textBigrams,
+  bigramJaccard,
+  calculateRetrievalEntropy,
+  filterTopP,
+  resolveTopP,
+  dedupeSupersededLessons,
+};

package/scripts/llm-client.js

@@ -2,6 +2,7 @@
 'use strict';
 
 const { runStep } = require('./durability/step');
+const { redactSecrets } = require('./secret-redaction');
 
 const MODELS = {
   FAST: 'claude-haiku-4-5-20251001',
@@ -141,6 +142,19 @@ function parseClaudeJson(text) {
   }
 }
 
+function buildSafeProviderError(error) {
+  const rawMessage = error && error.message ? String(error.message) : 'provider request failed';
+  const summary = {
+    name: String(error && error.name || 'Error').slice(0, 80),
+    message: redactSecrets(rawMessage).split('\n')[0].slice(0, 500),
+  };
+  const status = Number(error && (error.status || (error.response && error.response.status)));
+  if (Number.isFinite(status)) summary.status = status;
+  const code = error && error.code;
+  if (typeof code === 'string' || typeof code === 'number') summary.code = code;
+  return summary;
+}
+
 function getZaiApiKey(env = process.env) {
   return env.ZAI_API_KEY || env.THUMBGATE_ZAI_API_KEY || '';
 }
@@ -198,6 +212,8 @@ async function callGeminiInternal(options = {}) {
   const { detectInferenceBackend } = require('./local-model-profile');
   const providerMode = detectInferenceBackend(env).providerMode;
 
+  if (providerMode !== 'vertex' && !env.GEMINI_API_KEY) return null;
+
   try {
     const { GoogleGenAI } = require('@google/genai');
     if (!_geminiClient) {
@@ -246,7 +262,7 @@ async function callGeminiInternal(options = {}) {
       model: options.model,
     };
   } catch (err) {
-    console.error('Gemini/Vertex AI execution error:', err);
+    console.error('Gemini/Vertex AI execution error:', JSON.stringify(buildSafeProviderError(err)));
     return null;
   }
 }
@@ -366,5 +382,6 @@ module.exports = {
   parseClaudeJson,
   normalizeCacheOptions,
   buildClaudeRequest,
+  buildSafeProviderError,
   MODELS,
 };

package/scripts/mcp-policy.js

@@ -30,6 +30,11 @@ function loadSubagentProfiles() {
   return parsed;
 }
 
+function getConfiguredDefaultProfile() {
+  const { getSetting } = require('./settings-hierarchy');
+  return String(getSetting('mcp.defaultProfile') || 'essential');
+}
+
 function getActiveMcpProfile(toolName) {
   const explicitProfile = process.env.THUMBGATE_MCP_PROFILE || null;
   const runtimeSubagentProfile = process.env.THUMBGATE_SUBAGENT_PROFILE || null;
@@ -42,7 +47,7 @@ function getActiveMcpProfile(toolName) {
       const routing = routeProfile({ toolName });
       return routing.profile;
     }
-    return explicitProfile || 'default';
+    return explicitProfile || getConfiguredDefaultProfile();
   }
 
   const config = loadSubagentProfiles();
@@ -86,6 +91,7 @@ module.exports = {
   loadMcpPolicy,
   loadSubagentProfiles,
   getActiveMcpProfile,
+  getConfiguredDefaultProfile,
   getAllowedTools,
   isToolAllowed,
   assertToolAllowed,

package/scripts/workflow-sentinel.js

@@ -3,7 +3,7 @@
 
 const fs = require('fs');
 const path = require('path');
-const { execFileSync } = require('child_process');
+const { execFileSync } = require('node:child_process');
 
 const {
   DEFAULT_BASE_BRANCH,
@@ -90,7 +90,9 @@ function safeExecFileLines(binary, args, cwd) {
 }
 
 function normalizeGlob(glob) {
-  return normalizePosix(glob).replace(/\/+$/, '');
+  let normalized = normalizePosix(glob);
+  while (normalized.endsWith('/')) normalized = normalized.slice(0, -1);
+  return normalized;
 }
 
 function sanitizeGlobList(globs) {
@@ -282,17 +284,27 @@ function isProtectedApprovalRelevant(toolName, toolInput = {}) {
 function normalizeMemoryGuardForSentinel(memoryGuard, isHighRisk) {
   if (!memoryGuard || memoryGuard.mode === 'allow') return memoryGuard;
   const reason = String(memoryGuard.reason || '');
-  const broadToolOnlySignal = /^Tool "[^"]+" has \d+ attributed negative\(s\), \d+ total negative\(s\)$/i.test(reason);
-  if (!isHighRisk && broadToolOnlySignal) {
+  const broadToolOnlySignal = /^Tool "[^"]+" (?:has \d+ attributed negative\(s\), \d+ total negative\(s\)|has recurring negative patterns \(count: \d+\))$/i.test(reason);
+  if (broadToolOnlySignal) {
     return {
       ...memoryGuard,
-      mode: 'warn',
-      reason: `${reason}. Treating this as advisory because the current action is not in the high-risk command set.`,
+      mode: 'allow',
+      reason: `${reason}. Ignored for this action because no contextual pattern matched the current input.`,
     };
   }
   return memoryGuard;
 }
 
+function normalizeLearnedPolicyForSentinel(learnedPolicy, actionable) {
+  if (!learnedPolicy?.enabled || actionable) return learnedPolicy;
+  return {
+    ...learnedPolicy,
+    enabled: false,
+    advisoryOnly: true,
+    reason: 'non_execution_action',
+  };
+}
+
 function buildTaskScopeViolation(taskScope, affectedFiles) {
   if (!Array.isArray(affectedFiles) || affectedFiles.length === 0) return null;
   if (!taskScope || !Array.isArray(taskScope.allowedPaths) || taskScope.allowedPaths.length === 0) {
@@ -1465,7 +1477,12 @@ function evaluateWorkflowSentinel(toolName, toolInput = {}, options = {}) {
       options,
     }
   );
-  const taskScopeViolation = buildTaskScopeViolation(governanceState.taskScope, affectedFiles);
+  // Scope is an execution boundary, not a read boundary. Treating Read/Glob/Grep
+  // paths as writes generated a warning on almost every inspection and trained
+  // callers to ignore the sentinel.
+  const taskScopeViolation = highRiskAction
+    ? buildTaskScopeViolation(governanceState.taskScope, affectedFiles)
+    : null;
   const protectedSurface = buildProtectedSurface(governanceState, affectedFiles);
   const protectedSurfaceForRisk = isProtectedApprovalRelevant(normalizedToolName, normalizedToolInput)
     ? protectedSurface
@@ -1481,7 +1498,7 @@ function evaluateWorkflowSentinel(toolName, toolInput = {}, options = {}) {
     affectedFiles,
   }), options.feedbackOptions || {});
   const memoryGuard = normalizeMemoryGuardForSentinel(rawMemoryGuard, highRiskAction);
-  const learnedPolicy = getInterventionRecommendation({
+  const rawLearnedPolicy = getInterventionRecommendation({
     toolName: normalizedToolName,
     command: normalizedToolInput.command || '',
     affectedFiles,
@@ -1495,6 +1512,18 @@ function evaluateWorkflowSentinel(toolName, toolInput = {}, options = {}) {
       || process.env.THUMBGATE_FEEDBACK_DIR
       || (repoRoot ? path.join(repoRoot, '.thumbgate') : null),
   });
+  // Learned deny/warn models exist to constrain execution, not observation.
+  // Keeping their prediction in the report preserves diagnostics while making
+  // read-only inspection immune to stale or unrelated training state.
+  const learnedPolicy = normalizeLearnedPolicyForSentinel(
+    rawLearnedPolicy,
+    normalizedToolName === 'Bash'
+      || EDIT_LIKE_TOOLS.has(normalizedToolName)
+      || highRiskAction
+      || actionProfile.backgroundAgent
+      || actionProfile.economicAction
+      || actionProfile.customerSystemAction
+  );
   const blastRadius = buildBlastRadius({
     affectedFiles,
     integrity,
@@ -1638,6 +1667,7 @@ module.exports = {
   evaluateWorkflowSentinel,
   isHighRiskAction,
   loadGovernanceState,
+  normalizeLearnedPolicyForSentinel,
   scoreRisk,
 };