@buildaharness/aielia 0.3.2 → 0.3.3

package/dist/{cli-session-BiX2DOni.js → cli-session-DnSc6Hhx.js}

@@ -934,13 +934,19 @@ function isNearDuplicateText(a, b) {
   const lb = b.toLowerCase();
   return la.includes(lb) || lb.includes(la);
 }
+function modelInferredFactsEnabled(env) {
+  const source = typeof process !== "undefined" ? process.env : {};
+  const raw = String(source.AUDIT_MODEL_INFERRED_FACTS ?? "").trim().toLowerCase();
+  if (raw === "") return true;
+  return !["0", "false", "off", "no", "disabled"].includes(raw);
+}
 function mergeTurnFacts(lexicalFacts, llmFacts) {
   const novelLlmFacts = llmFacts.filter((llmFact) => !lexicalFacts.some((lexFact) => isNearDuplicateText(lexFact.text, llmFact.text)));
   return [...lexicalFacts, ...novelLlmFacts];
 }
 function buildTurnFacts(sessionId, userMessage, statedFacts) {
   const lexicalFacts = extractFactsFromTurn(userMessage, `turn:${sessionId}`);
-  const llmFacts = statedFacts.map((fact) => ({
+  const llmFacts = modelInferredFactsEnabled() ? statedFacts.map((fact) => ({
     text: fact.text,
     extractedAt: (/* @__PURE__ */ new Date()).toISOString(),
     sourceTurn: `turn:${sessionId}`,
@@ -948,7 +954,7 @@ function buildTurnFacts(sessionId, userMessage, statedFacts) {
     durable: fact.durable,
     confidence: fact.confidence,
     category: fact.category
-  }));
+  })) : [];
   return mergeTurnFacts(lexicalFacts, llmFacts);
 }
 function shouldAutoPromote(fact) {
@@ -6137,6 +6143,12 @@ function toTaskRiskLevel(riskLevel) {
 function planTaskRiskLevel(description) {
   return classifyRisk(description).riskLevel;
 }
+function decompositionEnabled(env) {
+  const source = typeof process !== "undefined" ? process.env : {};
+  const raw = String(source.AUDIT_DECOMPOSITION ?? "").trim().toLowerCase();
+  if (raw === "") return true;
+  return !["0", "false", "off", "no", "disabled"].includes(raw);
+}
 class TurnInterpreter {
   constructor(llmClient, model, planService, reminderStore) {
     this.llmClient = llmClient;
@@ -6198,8 +6210,9 @@ class TurnInterpreter {
    */
   async resolveTasks(params) {
     const { userMessage, sessionId, classification, planForCancelCheck, onUsage } = params;
+    const decompose = decompositionEnabled();
     let initialTasks = toHarnessTasks([{ id: "respond", description: userMessage, depends_on: [] }], toTaskRiskLevel(classification.riskLevel));
-    const decomposed = classification.decomposedTasks;
+    const decomposed = decompose ? classification.decomposedTasks : null;
     if (decomposed) {
       initialTasks = toHarnessTasks(decomposed, toTaskRiskLevel(classification.riskLevel));
     }
@@ -6214,7 +6227,7 @@ class TurnInterpreter {
     } else {
       planClassifiedTrace = { isCandidate: classification.matchedPlanTemplate !== null, matchedTemplate: classification.matchedPlanTemplate };
     }
-    if (initialTasks.length === 1 && initialTasks[0].id === "respond" && classification.riskLevel !== "LOW" && looksLikeCodingFact(userMessage)) {
+    if (decompose && initialTasks.length === 1 && initialTasks[0].id === "respond" && classification.riskLevel !== "LOW" && looksLikeCodingFact(userMessage)) {
       const reframed = await reframeTaskDescriptionWithLLM(userMessage, this.llmClient, this.model(), onUsage);
       if (reframed) {
         initialTasks = toHarnessTasks([{ id: "respond", description: reframed, depends_on: [] }], toTaskRiskLevel(classification.riskLevel));
@@ -6285,6 +6298,12 @@ const REVIEW_SCHEMA = {
   },
   required: ["conflict"]
 };
+function semanticChangeReviewEnabled(env) {
+  const source = typeof process !== "undefined" ? process.env : {};
+  const raw = String(source.AUDIT_SEMANTIC_CHANGE_REVIEW ?? "").trim().toLowerCase();
+  if (raw === "") return true;
+  return !["0", "false", "off", "no", "disabled"].includes(raw);
+}
 const SYSTEM_PROMPT$2 = `You check whether a proposed action genuinely conflicts with something already known to be true (a high-confidence belief) or predicted (an active hypothesis's predicted observation) — a real logical conflict, not just a superficially related topic (e.g. proposing to remove something a belief says is required, or an action that presumes the opposite of what's predicted). You are given "changeDescription", "highConfidenceBeliefs", and "hypothesisPredictions" as JSON. Respond with JSON only: {"conflict": boolean, "reason": string}. reason only needs to be set when conflict is true.`;
 async function checkSemanticReviewConflict(changeDescription, highConfidenceBeliefs, hypothesisPredictions, llmClient, model, onUsage) {
   if (looksLikeCodingFact(changeDescription)) return { conflict: false };
@@ -6342,6 +6361,12 @@ async function checkSemanticFailureMatch(symptoms, libraryEntries, llmClient, mo
   }
 }
 const NON_CHECKABLE_DEFAULT_CRITERION = "Respond helpfully, accurately, and safely to the user request.";
+function semanticCriterionCoverageEnabled(env) {
+  const source = typeof process !== "undefined" ? process.env : {};
+  const raw = String(source.AUDIT_SEMANTIC_CRITERION_COVERAGE ?? "").trim().toLowerCase();
+  if (raw === "") return true;
+  return !["0", "false", "off", "no", "disabled"].includes(raw);
+}
 const COVERAGE_SCHEMA = {
   type: "object",
   properties: {
@@ -6367,6 +6392,18 @@ async function checkSemanticCriterionCoverage(criterion, beliefs, llmClient, mod
     return false;
   }
 }
+function verificationEnabled(env) {
+  const source = typeof process !== "undefined" ? process.env : {};
+  const raw = String(source.AUDIT_VERIFICATION ?? "").trim().toLowerCase();
+  if (raw === "") return true;
+  return !["0", "false", "off", "no", "disabled"].includes(raw);
+}
+function reviewerPassEnabled(env) {
+  const source = typeof process !== "undefined" ? process.env : {};
+  const raw = String(source.AUDIT_REVIEWER_PASS ?? "").trim().toLowerCase();
+  if (raw === "") return true;
+  return !["0", "false", "off", "no", "disabled"].includes(raw);
+}
 const PLAN_AUTO_ADVANCE_TASK_CEILING = 10;
 class HarnessBridge {
   constructor(memory, experienceStore, checkpointStore, llmClient, model, maxSteps, planService, assistantSession, onTrace, oneLoopMode = DEFAULT_ONE_LOOP_MODE) {
@@ -6448,6 +6485,13 @@ class HarnessBridge {
         onVerification: (result) => {
           lastVerification = result;
         },
+        // AUDIT_VERIFICATION (feature-value audit, Phase C5, eval-only) → skip verify() entirely.
+        // Default ON — unchanged shipped behaviour; skipVerification stays undefined.
+        skipVerification: verificationEnabled() ? void 0 : true,
+        // AUDIT_REVIEWER_PASS (feature-value audit, Phase C6, eval-only) → skip reviewerPass()
+        // (and any reviewer_pass_2 re-run) entirely, including its C1/C2 sub-mechanisms. Default
+        // ON — unchanged shipped behaviour; skipReviewerPass stays undefined.
+        skipReviewerPass: reviewerPassEnabled() ? void 0 : true,
         // Trajectory Supervisor GATHER_EVIDENCE host (S5). Inert unless a supervisorDecider is
         // also wired and returns a GATHER_EVIDENCE directive at a stall edge; absent → the
         // harness degrades GATHER_EVIDENCE to CONTINUE.
@@ -6494,7 +6538,11 @@ class HarnessBridge {
         // Layered on top of review-proposed-change.ts's lexical isNegation check — same "skip
         // when it reads like a coding fact" gate contradictionChecker uses, since that's the
         // domain the fixed-phrase check already covers reasonably well.
-        semanticChangeReviewer: (input) => checkSemanticReviewConflict(input.changeDescription, input.highConfidenceBeliefs, input.hypothesisPredictions, this.llmClient, this.model(), onUsage),
+        // AUDIT_SEMANTIC_CHANGE_REVIEW (feature-value audit, Phase C2) gates the whole hook: OFF →
+        // no host semanticChangeReviewer is wired at all, so the harness's mechanical
+        // reviewProposedChange (lexical isNegation) is the only conflict check. Default ON —
+        // unchanged shipped behaviour.
+        semanticChangeReviewer: semanticChangeReviewEnabled() ? (input) => checkSemanticReviewConflict(input.changeDescription, input.highConfidenceBeliefs, input.hypothesisPredictions, this.llmClient, this.model(), onUsage) : void 0,
         // Layered on top of FailureModeLibrary's own exact-string-overlap match() — see
         // failure-mode-matcher.ts's doc comment for why exact equality against a curated symptom
         // list almost never happens for free-text observations in practice.
@@ -6505,7 +6553,10 @@ class HarnessBridge {
         // Layered on top of reviewerPass's implementerLens's own `.includes()` substring check —
         // called only for a success criterion that substring check found no coverage for. See
         // semantic-criterion-coverage.ts's doc comment.
-        semanticCriterionCoverage: (criterion, beliefs) => checkSemanticCriterionCoverage(criterion, beliefs, this.llmClient, this.model(), onUsage),
+        // AUDIT_SEMANTIC_CRITERION_COVERAGE (feature-value audit, Phase C1) gates the whole hook:
+        // OFF → no host semanticCriterionCoverage is wired at all, so the reviewer's implementer
+        // lens runs its `.includes()` substring check alone. Default ON — unchanged shipped behaviour.
+        semanticCriterionCoverage: semanticCriterionCoverageEnabled() ? (criterion, beliefs) => checkSemanticCriterionCoverage(criterion, beliefs, this.llmClient, this.model(), onUsage) : void 0,
         // Stop right after a MEDIUM/HIGH-risk plan step resolves (COMPLETE or FAILED), before
         // the loop would go pick the next one — undefined for a non-plan turn, so shouldPause is
         // simply never checked and behavior is unchanged.
@@ -8289,4 +8340,4 @@ export {
   buildWhyChain as y,
   classifyError as z
 };
-//# sourceMappingURL=cli-session-BiX2DOni.js.map
+//# sourceMappingURL=cli-session-DnSc6Hhx.js.map