@remnic/core 9.3.701 → 9.3.702

package/dist/{chunk-ODTWHSY2.js → chunk-64XXOQRW.js}

@@ -4,7 +4,7 @@ import {
 } from "./chunk-ZFXCQPNO.js";
 import {
   StorageManager
-} from "./chunk-SFMRLXIV.js";
+} from "./chunk-GFIARMA7.js";
 import {
   MutationSerializer
 } from "./chunk-LTJAMRGI.js";
@@ -336,4 +336,4 @@ export {
   resolveNamespaceStorageRoot,
   NamespaceStorageRouter
 };
-//# sourceMappingURL=chunk-ODTWHSY2.js.map
+//# sourceMappingURL=chunk-64XXOQRW.js.map

package/dist/{chunk-ZT7B64BE.js → chunk-6TAETM63.js}

@@ -10,7 +10,7 @@ import {
 import {
   planNamespaceMaintenance,
   runNamespaceMaintenanceBatchPlan
-} from "./chunk-JKOKX3PS.js";
+} from "./chunk-JMA4RYRN.js";
 import {
   resolveHomeDir
 } from "./chunk-JUC24CTX.js";
@@ -441,4 +441,4 @@ var MaintenanceScheduler = class {
 export {
   MaintenanceScheduler
 };
-//# sourceMappingURL=chunk-ZT7B64BE.js.map
+//# sourceMappingURL=chunk-6TAETM63.js.map

package/dist/{chunk-HF4N43Q7.js → chunk-6VVP6NK7.js}

@@ -1,7 +1,7 @@
 import {
   StorageManager,
   normalizeEntityName
-} from "./chunk-SFMRLXIV.js";
+} from "./chunk-GFIARMA7.js";
 import {
   extractJsonCandidates
 } from "./chunk-RGMVMVMF.js";
@@ -870,4 +870,4 @@ export {
   resolveBriefingSaveDir,
   briefingFilename
 };
-//# sourceMappingURL=chunk-HF4N43Q7.js.map
+//# sourceMappingURL=chunk-6VVP6NK7.js.map

package/dist/{chunk-FN2SM5SN.js → chunk-ABCGDW3A.js}

@@ -9,7 +9,7 @@ import {
   ProactiveExtractionResultSchema,
   ProactiveQuestionsResultSchema,
   buildProfileConsolidationResultSchema
-} from "./chunk-ROZJACKP.js";
+} from "./chunk-KKK7YTYN.js";
 import {
   normalizeReasoningTrace
 } from "./chunk-54V4BZWP.js";
@@ -36,9 +36,6 @@ import {
 import {
   normalizeProcedureSteps
 } from "./chunk-QDW3E4RD.js";
-import {
-  sanitizeMemoryContent
-} from "./chunk-FVQJYWH7.js";
 import {
   FallbackLlmClient,
   fallbackLlmRuntimeContextFromConfig,
@@ -51,6 +48,12 @@ import {
 import {
   extractJsonCandidates
 } from "./chunk-RGMVMVMF.js";
+import {
+  buildFactProvenance
+} from "./chunk-SK2CR6MW.js";
+import {
+  sanitizeMemoryContent
+} from "./chunk-FVQJYWH7.js";
 import {
   log
 } from "./chunk-NU3CSQ4H.js";
@@ -206,7 +209,13 @@ var ExtractionEngine = class {
       reasoningTrace: (() => {
         const candidate = f?.reasoningTrace && typeof f.reasoningTrace === "object" && !Array.isArray(f.reasoningTrace) ? f.reasoningTrace : f?.reasoning_trace && typeof f.reasoning_trace === "object" && !Array.isArray(f.reasoning_trace) ? f.reasoning_trace : null;
         return candidate ? normalizeReasoningTrace(candidate) ?? void 0 : void 0;
-      })()
+      })(),
+      // Issue #1575 PR 2: the LLM provides a verbatim supporting quote
+      // per fact. Optional/nullable per repo gotcha 6 (OpenAI Responses
+      // API emits null for absent optional fields). The post-parse
+      // validator (buildFactProvenance) locates this quote in the
+      // buffered turns and builds verified ProvenanceSource[] entries.
+      quote: typeof f?.quote === "string" && f.quote.trim().length > 0 ? f.quote : void 0
     })).filter((f) => f.content.length > 0) : [];
     const questions = Array.isArray(parsed?.questions) ? parsed.questions.map((q) => {
       if (typeof q === "string") return { question: q, context: "", priority: 0.5 };
@@ -480,7 +489,8 @@ var ExtractionEngine = class {
       `Return at most ${maxAdditional} additional high-confidence memory candidates that were omitted from the base extraction.`,
       "Only include information directly supported by the conversation. Do not speculate. Do not repeat the base extraction.",
       "Return only valid JSON with this shape:",
-      '{"facts":[{"category":"fact","content":"...","confidence":0.0,"tags":["..."],"entityRef":"optional","promptedByQuestion":"optional"}],"profileUpdates":["..."],"entities":[{"name":"...","type":"person","facts":["..."],"structuredSections":[{"key":"beliefs","title":"Beliefs","facts":["..."]}],"promptedByQuestion":"optional"}],"relationships":[{"source":"...","target":"...","label":"...","promptedByQuestion":"optional"}]}',
+      '{"facts":[{"category":"fact","content":"...","confidence":0.0,"tags":["..."],"entityRef":"optional","promptedByQuestion":"optional","quote":"optional verbatim span from a single turn"}],"profileUpdates":["..."],"entities":[{"name":"...","type":"person","facts":["..."],"structuredSections":[{"key":"beliefs","title":"Beliefs","facts":["..."]}],"promptedByQuestion":"optional"}],"relationships":[{"source":"...","target":"...","label":"...","promptedByQuestion":"optional"}]}',
+      this.config.provenance?.enabled ? '- Source quotes: For each fact, include a "quote" field with the EXACT verbatim words from the conversation that support the fact (a contiguous span from a single turn, not a paraphrase). Cap at ~300 chars.' : "",
       "",
       "Base extracted facts (do not repeat):",
       factsPreview || "(none)",
@@ -683,6 +693,54 @@ var ExtractionEngine = class {
     }
     return null;
   }
+  /**
+   * Attach claim-level provenance spans to each fact in the extraction result
+   * (issue #1575 PR 2). Runs once at write time, after sanitize + proactive
+   * pass so ALL facts (base + proactive additions) get verified spans before
+   * the result is returned for persistence.
+   *
+   * The validator locates each fact's LLM-provided `quote` in the buffered
+   * turns and builds a `ProvenanceSource[]` with verified offsets. Never
+   * throws, never drops a fact — an unverifiable span is a tagged state, not
+   * a silent failure (rule 34 spirit). When `provenance.enabled` is false,
+   * this is a no-op (byte-identical to pre-feature behavior, rule 39).
+   */
+  attachProvenanceToResult(result, turns) {
+    if (!this.config.provenance?.enabled) {
+      if (result.facts.length === 0) return result;
+      return {
+        ...result,
+        facts: result.facts.map((fact) => {
+          if (fact.quote === void 0) return fact;
+          const { quote: _stripped, ...rest } = fact;
+          return rest;
+        })
+      };
+    }
+    if (result.facts.length === 0) return result;
+    const provenanceTurns = turns.map((t) => ({
+      content: t.content,
+      sessionKey: t.sessionKey,
+      logicalSessionKey: t.logicalSessionKey,
+      timestamp: t.timestamp,
+      turnId: t.turnFingerprint
+    }));
+    const facts = result.facts.map((fact) => {
+      const built = buildFactProvenance(
+        fact.quote,
+        provenanceTurns,
+        this.config.provenance
+      );
+      const { quote: _stripped, ...factWithoutQuote } = fact;
+      return {
+        ...factWithoutQuote,
+        ...built.sources && built.sources.length > 0 ? { sources: built.sources } : {},
+        ...built.provenance !== "none" ? { provenance: built.provenance } : {},
+        ...built.requireSpansPending ? { requireSpansPending: true } : {}
+      };
+    });
+    return { ...result, facts };
+  }
   async extract(turns, existingEntities) {
     const substantiveTurns = turns.filter((t) => t.content.trim().length > 0);
     if (substantiveTurns.length === 0) {
@@ -735,7 +793,8 @@ var ExtractionEngine = class {
             this.emit({ kind: "llm_end", traceId, model: this.config.localLlmModel, operation: "extraction", durationMs });
             log.debug(`extraction: used local LLM \u2014 ${localResult.facts.length} facts, ${localResult.entities.length} entities`);
             const sanitized = this.sanitizeExtractionResult(localResult, messageTimestamp);
-            return await this.applyProactiveQuestionPass(conversation, sanitized);
+            const finalResult = await this.applyProactiveQuestionPass(conversation, sanitized);
+            return this.attachProvenanceToResult(finalResult, boundedTurns);
           }
           if (!this.config.localLlmFallback) {
             log.warn("extraction: local LLM failed and fallback disabled");
@@ -777,7 +836,8 @@ var ExtractionEngine = class {
             this.emit({ kind: "llm_end", traceId, model: this.config.model, operation: "extraction", durationMs });
             log.debug(`extraction: used direct client (${this.config.model}) \u2014 ${directResult.facts.length} facts, ${directResult.entities.length} entities`);
             const sanitized = this.sanitizeExtractionResult(directResult, messageTimestamp);
-            return await this.applyProactiveQuestionPass(conversation, sanitized);
+            const finalResult = await this.applyProactiveQuestionPass(conversation, sanitized);
+            return this.attachProvenanceToResult(finalResult, boundedTurns);
           }
           try {
             this.emit({
@@ -878,7 +938,8 @@ var ExtractionEngine = class {
             questions: result.questions ?? [],
             identityReflection: result.identityReflection ?? void 0
           }, messageTimestamp);
-          return await this.applyProactiveQuestionPass(conversation, sanitized);
+          const finalResult = await this.applyProactiveQuestionPass(conversation, sanitized);
+          return this.attachProvenanceToResult(finalResult, boundedTurns);
         }
         this.emit({
           kind: "llm_error",
@@ -990,7 +1051,8 @@ These are durable insights - capture them:
 === Rules ===
 - Extract only NEW information worth remembering across sessions
 - Skip transient details (file paths, current errors, temporary states, agent actions)
-- Confidence: Explicit (0.95-1.0), Implied (0.70-0.94), Inferred (0.40-0.69), Speculative (0.00-0.39)
+- Confidence: Explicit (0.95-1.0), Implied (0.70-0.94), Inferred (0.40-0.69), Speculative (0.00-0.39)${this.config.provenance?.enabled ? `
+- Source quotes: For each fact, include a "quote" field with the EXACT verbatim words from the conversation that support the fact (copy a contiguous span from a single turn, not a paraphrase). Cap at ~300 chars.` : ""}
 - Corrections get highest confidence (0.95+)
 - Each fact should be standalone and self-contained
 - Lines labelled [context user] or [context assistant] are reference context only. Use them to resolve pronouns and adjacent question/answer pairs, but do not extract a memory stated only in context lines unless a normal [user] or [assistant] line confirms or completes it.
@@ -1216,7 +1278,8 @@ Rules:
   * Explicit (0.95-1.0): Direct user statements \u2014 "I prefer X", "my name is Y"
   * Implied (0.70-0.94): Strong contextual inference \u2014 user consistently does X, clear from conversation flow
   * Inferred (0.40-0.69): Pattern recognition \u2014 reasonable guess from limited evidence
-  * Speculative (0.00-0.39): Tentative hypothesis \u2014 weak signal, needs future confirmation. Speculative memories auto-expire after 30 days if not confirmed.
+  * Speculative (0.00-0.39): Tentative hypothesis \u2014 weak signal, needs future confirmation. Speculative memories auto-expire after 30 days if not confirmed.${this.config.provenance?.enabled ? `
+- Source quotes: For each fact, include a "quote" field containing the EXACT verbatim words from the conversation that support the fact. Copy a contiguous span from a single speaker turn (not a paraphrase, not a summary). Cap at ~300 characters. This grounds every memory in the literal utterance that created it.` : ""}
 - For commitments: include any deadline or timeframe mentioned${this.config.extractionScopeClassificationEnabled ? `
 
 Scope classification:
@@ -2328,4 +2391,4 @@ ${memoryList}` }
 export {
   ExtractionEngine
 };
-//# sourceMappingURL=chunk-FN2SM5SN.js.map
+//# sourceMappingURL=chunk-ABCGDW3A.js.map