thumbgate 1.2.0 → 1.3.0

package/scripts/workflow-sentinel.js

@@ -16,6 +16,7 @@ const {
 } = require('./operational-integrity');
 const { buildDockerSandboxPlan } = require('./docker-sandbox-planner');
 const { evaluatePretool } = require('./hybrid-feedback-context');
+const { getInterventionRecommendation } = require('./intervention-policy');
 
 const GOVERNANCE_STATE_PATH = path.join(process.env.HOME || '/tmp', '.thumbgate', 'governance-state.json');
 const DEFAULT_PROTECTED_FILE_GLOBS = [
@@ -387,6 +388,7 @@ function scoreRisk({
   affectedFiles,
   integrity,
   memoryGuard,
+  learnedPolicy,
   blastRadius,
   taskScopeViolation,
   protectedSurface,
@@ -472,6 +474,43 @@ function scoreRisk({
       { mode: memoryGuard.mode }
     );
   }
+  if (learnedPolicy && learnedPolicy.enabled && learnedPolicy.prediction) {
+    const confidence = learnedPolicy.prediction.confidence || 0;
+    const label = learnedPolicy.prediction.label;
+    if (label === 'deny' && confidence >= 0.6) {
+      addDriver(
+        drivers,
+        'learned_policy_deny',
+        Math.min(0.26, 0.16 + (confidence * 0.12)),
+        'Learned intervention policy predicts a deny-worthy failure pattern.',
+        { confidence, label }
+      );
+    } else if (label === 'warn' && confidence >= 0.3) {
+      addDriver(
+        drivers,
+        'learned_policy_warn',
+        Math.min(0.18, 0.1 + (confidence * 0.08)),
+        'Learned intervention policy predicts elevated execution risk.',
+        { confidence, label }
+      );
+    } else if (label === 'verify' && confidence >= 0.3) {
+      addDriver(
+        drivers,
+        'learned_policy_verify',
+        Math.min(0.16, 0.08 + (confidence * 0.06)),
+        'Learned intervention policy predicts a verification gap before close-out.',
+        { confidence, label }
+      );
+    } else if (label === 'recall' && confidence >= 0.3) {
+      addDriver(
+        drivers,
+        'learned_policy_recall',
+        Math.min(0.14, 0.06 + (confidence * 0.05)),
+        'Learned intervention policy predicts prior lessons are needed before execution.',
+        { confidence, label }
+      );
+    }
+  }
 
   const score = Math.min(1, drivers.reduce((sum, driver) => sum + driver.weight, 0));
   return {
@@ -492,6 +531,7 @@ function scoreRisk({
 function buildEvidence({
   integrity,
   memoryGuard,
+  learnedPolicy,
   blastRadius,
   taskScopeViolation,
   protectedSurface,
@@ -500,6 +540,16 @@ function buildEvidence({
   if (memoryGuard && memoryGuard.mode && memoryGuard.mode !== 'allow') {
     evidence.push(`Memory guard predicted ${memoryGuard.mode}: ${memoryGuard.reason}`);
   }
+  if (learnedPolicy && learnedPolicy.enabled && learnedPolicy.prediction) {
+    const topTokens = Array.isArray(learnedPolicy.topTokens)
+      ? learnedPolicy.topTokens.map((entry) => entry.token).slice(0, 3)
+      : [];
+    evidence.push(
+      `Learned policy predicted ${learnedPolicy.prediction.label} (${Math.round((learnedPolicy.prediction.confidence || 0) * 100)}% confidence)`
+      + (topTokens.length ? ` from ${topTokens.join(', ')}` : '')
+      + '.'
+    );
+  }
   if (taskScopeViolation) {
     evidence.push(
       taskScopeViolation.reasonCode === 'missing_task_scope'
@@ -575,6 +625,7 @@ function buildRemediations({
   protectedSurface,
   blastRadius,
   memoryGuard,
+  learnedPolicy,
   executionSurface,
 }) {
   const remediations = [];
@@ -611,6 +662,24 @@ function buildRemediations({
       'The system already has evidence that this action pattern failed before.'
     );
   }
+  if (learnedPolicy && learnedPolicy.enabled && learnedPolicy.prediction) {
+    if (learnedPolicy.prediction.label === 'verify' && learnedPolicy.prediction.confidence >= 0.3) {
+      push(
+        'verify_before_closeout',
+        'Raise verification before claiming success',
+        'Run the relevant proof or test command and confirm the exact output before retrying or closing out.',
+        'The learned policy predicts this path tends to fail at verification time.'
+      );
+    }
+    if (learnedPolicy.prediction.label === 'recall' && learnedPolicy.prediction.confidence >= 0.3) {
+      push(
+        'retrieve_lessons',
+        'Inspect prior lessons',
+        'Call retrieve_lessons or search_lessons for this tool context before retrying.',
+        'The learned policy predicts this action needs prior lessons and corrective context.'
+      );
+    }
+  }
   if (blastRadius.fileCount >= 4 || blastRadius.surfaceCount >= 3) {
     push(
       'split_blast_radius',
@@ -636,6 +705,11 @@ function buildReasoning(report) {
     `Workflow sentinel risk ${report.band} (${report.riskScore}) for ${report.toolName}.`,
     `Blast radius: ${report.blastRadius.summary}.`,
   ];
+  if (report.learnedPolicy && report.learnedPolicy.enabled && report.learnedPolicy.prediction) {
+    lines.push(
+      `Learned policy predicted ${report.learnedPolicy.prediction.label} (${report.learnedPolicy.prediction.confidence}).`
+    );
+  }
   if (report.executionSurface?.shouldSandbox) {
     lines.push(`Execution surface: ${report.executionSurface.summary}`);
   }
@@ -658,15 +732,32 @@ function getSentinelActionType(toolName) {
   return '';
 }
 
-function chooseDecision({ riskScore, integrity, memoryGuard, blastRadius, command }) {
+function chooseDecision({ riskScore, integrity, memoryGuard, learnedPolicy, blastRadius, command }) {
   const hasOperationalBlockers = Boolean(integrity && Array.isArray(integrity.blockers) && integrity.blockers.length > 0);
   const destructiveBypass = /\bgit\s+push\b.*(?:--force|-f)\b/i.test(command) || /\bgh\s+pr\s+merge\b.*--admin\b/i.test(command);
+  const learnedPrediction = learnedPolicy && learnedPolicy.enabled ? learnedPolicy.prediction : null;
+  const learnedHardStop = Boolean(
+    learnedPrediction
+      && learnedPrediction.label === 'deny'
+      && learnedPrediction.confidence >= 0.7
+  );
+  const learnedWarning = Boolean(
+    learnedPrediction
+      && ['warn', 'verify', 'deny'].includes(learnedPrediction.label)
+      && learnedPrediction.confidence >= 0.3
+  );
+  const learnedRecall = Boolean(
+    learnedPrediction
+      && learnedPrediction.label === 'recall'
+      && learnedPrediction.confidence >= 0.3
+  );
   const lowBlastRadius = blastRadius.fileCount <= 1
     && blastRadius.surfaceCount <= 1
     && blastRadius.releaseSensitiveFiles.length === 0
     && blastRadius.unapprovedProtectedFiles === 0;
   const lowRiskHandoff = /\bgit\s+push\b|\bgh\s+pr\s+(?:create|merge)\b/i.test(command)
     && !destructiveBypass
+    && !learnedHardStop
     && lowBlastRadius
     && !hasOperationalBlockers
     && memoryGuard
@@ -686,10 +777,10 @@ function chooseDecision({ riskScore, integrity, memoryGuard, blastRadius, comman
   if (lowRiskHandoff) {
     return 'allow';
   }
-  if (destructiveBypass || repeatedHighBlast || (hasOperationalBlockers && riskScore >= 0.72) || riskScore >= 0.86) {
+  if (destructiveBypass || learnedHardStop || repeatedHighBlast || (hasOperationalBlockers && riskScore >= 0.72) || riskScore >= 0.86) {
     return 'deny';
   }
-  if (riskScore >= 0.45) {
+  if (riskScore >= 0.45 || (learnedWarning && riskScore >= 0.3) || (learnedRecall && riskScore >= 0.34)) {
     return 'warn';
   }
   return 'allow';
@@ -732,6 +823,20 @@ function evaluateWorkflowSentinel(toolName, toolInput = {}, options = {}) {
     affectedFiles,
   }), options.feedbackOptions || {});
   const memoryGuard = normalizeMemoryGuardForSentinel(rawMemoryGuard, highRiskAction);
+  const learnedPolicy = getInterventionRecommendation({
+    toolName,
+    command: toolInput.command || '',
+    affectedFiles,
+    integrity,
+    memoryGuard,
+    riskBand: highRiskAction ? 'high' : 'low',
+    taskScopeViolation,
+    protectedSurface: protectedSurfaceForRisk,
+  }, {
+    feedbackDir: options.feedbackDir
+      || process.env.THUMBGATE_FEEDBACK_DIR
+      || (repoRoot ? path.join(repoRoot, '.thumbgate') : null),
+  });
   const blastRadius = buildBlastRadius({
     affectedFiles,
     integrity,
@@ -743,6 +848,7 @@ function evaluateWorkflowSentinel(toolName, toolInput = {}, options = {}) {
     affectedFiles,
     integrity,
     memoryGuard,
+    learnedPolicy,
     blastRadius,
     taskScopeViolation,
     protectedSurface: protectedSurfaceForRisk,
@@ -763,6 +869,7 @@ function evaluateWorkflowSentinel(toolName, toolInput = {}, options = {}) {
     riskScore: risk.score,
     integrity,
     memoryGuard,
+    learnedPolicy,
     blastRadius: {
       ...blastRadius,
       unapprovedProtectedFiles: protectedSurfaceForRisk.unapprovedProtectedFiles.length,
@@ -772,6 +879,7 @@ function evaluateWorkflowSentinel(toolName, toolInput = {}, options = {}) {
   const evidence = buildEvidence({
     integrity,
     memoryGuard,
+    learnedPolicy,
     blastRadius,
     taskScopeViolation,
     protectedSurface: protectedSurfaceForRisk,
@@ -782,6 +890,7 @@ function evaluateWorkflowSentinel(toolName, toolInput = {}, options = {}) {
     protectedSurface: protectedSurfaceForRisk,
     blastRadius,
     memoryGuard,
+    learnedPolicy,
     executionSurface,
   });
   const summary = decision === 'allow'
@@ -790,7 +899,7 @@ function evaluateWorkflowSentinel(toolName, toolInput = {}, options = {}) {
       ? 'Predicted workflow risk is elevated before execution.'
       : 'Predicted workflow failure before execution.';
   const report = {
-    sentinelVersion: 'workflow-sentinel-v1',
+    sentinelVersion: 'workflow-sentinel-v2',
     toolName,
     decision,
     riskScore: risk.score,
@@ -802,6 +911,7 @@ function evaluateWorkflowSentinel(toolName, toolInput = {}, options = {}) {
     remediations,
     executionSurface,
     memoryGuard,
+    learnedPolicy,
     taskScopeViolation,
     operationalIntegrity: {
       ok: integrity.ok,

package/skills/thumbgate/SKILL.md

@@ -86,7 +86,7 @@ Bounded retrieval of relevant feedback history for the current task. The agent g
 
 | | Free | Pro | Team |
 |---|---|---|---|
-| Feedback capture | Unlimited | Unlimited | Unlimited |
+| Feedback capture | 3/day | Unlimited | Unlimited |
 | Lesson search | 5/day | Unlimited | Unlimited |
 | Active gates | 5 | Unlimited | Unlimited |
 | Dashboard | - | Yes | Yes |