fix: wrap extraction transcript turns in speaker tags so assistant context cannot be misattributed

dist/index.js

@@ -39,7 +39,7 @@ import { createReflectionEventId } from "./src/reflection-event-store.js";
 import { buildReflectionMappedMetadata } from "./src/reflection-mapped-metadata.js";
 import { createMemoryCLI } from "./cli.js";
 import { isNoise } from "./src/noise-filter.js";
-import { normalizeAutoCaptureText } from "./src/auto-capture-cleanup.js";
+import { normalizeAutoCaptureText, buildConversationTurnsForExtraction, trimTurnsToUserCap, dedupePairWindow, } from "./src/auto-capture-cleanup.js";
 // Import smart extraction & lifecycle components
 import { SmartExtractor, createExtractionRateLimiter } from "./src/smart-extractor.js";
 import { compressTexts, estimateConversationValue } from "./src/session-compressor.js";
@@ -1845,6 +1845,7 @@ function _initPluginState(api) {
             noiseBank.init(embedder).catch((err) => api.logger.debug(`memory-lancedb-pro: noise bank init: ${String(err)}`));
             const admissionRejectionAuditWriter = createAdmissionRejectionAuditWriter(config, resolvedDbPath, api);
             smartExtractor = new SmartExtractor(store, embedder, llmClient, {
+                captureAssistantEligible: config.captureAssistant === true,
                 user: "User",
                 extractMinMessages: config.extractMinMessages ?? 4,
                 extractMaxChars: config.extractMaxChars ?? 8000,
@@ -1885,6 +1886,7 @@ function _initPluginState(api) {
     const autoCaptureSeenTextCount = new Map();
     const autoCapturePendingIngressTexts = new Map();
     const autoCaptureRecentTexts = new Map();
+    const autoCaptureRecentPairTurns = new Map();
     return {
         config,
         resolvedDbPath,
@@ -1912,6 +1914,7 @@ function _initPluginState(api) {
         autoCaptureSeenTextCount,
         autoCapturePendingIngressTexts,
         autoCaptureRecentTexts,
+        autoCaptureRecentPairTurns,
     };
 }
 export function isAgentOrSessionExcluded(agentId, sessionKey, patterns) {
@@ -2020,7 +2023,7 @@ const memoryLanceDBProPlugin = {
             _registeredApisMap.delete(api); // dual-track rollback: Map un-claim
             throw err;
         }
-        const { config, resolvedDbPath, vectorDim, store, embedder, retriever, canonicalCorpusIndexer, dreamingEngine, dreamingScheduler, scopeManager, migrator, smartExtractor, mdMirror, decayEngine, tierManager, extractionRateLimiter, reflectionErrorStateBySession, reflectionDerivedBySession, reflectionDerivedSuppressionBySession, reflectionByAgentCache, reflectionByAgentCacheGeneration, recallHistory, turnCounter, autoCaptureSeenTextCount, autoCapturePendingIngressTexts, autoCaptureRecentTexts, } = singleton;
+        const { config, resolvedDbPath, vectorDim, store, embedder, retriever, canonicalCorpusIndexer, dreamingEngine, dreamingScheduler, scopeManager, migrator, smartExtractor, mdMirror, decayEngine, tierManager, extractionRateLimiter, reflectionErrorStateBySession, reflectionDerivedBySession, reflectionDerivedSuppressionBySession, reflectionByAgentCache, reflectionByAgentCacheGeneration, recallHistory, turnCounter, autoCaptureSeenTextCount, autoCapturePendingIngressTexts, autoCaptureRecentTexts, autoCaptureRecentPairTurns, } = singleton;
         warnForDisabledChannelPlugin(api.config, api.logger);
         async function sleep(ms, signal) {
             if (signal?.aborted) {
@@ -2900,29 +2903,36 @@ const memoryLanceDBProPlugin = {
                             ? config.scopes?.default ?? "global"
                             : scopeManager.getDefaultScope(agentId);
                         const sessionKey = ctx?.sessionKey || event.sessionKey || "unknown";
-                        api.logger.debug(`memory-lancedb-pro: auto-capture agent_end payload for agent ${agentId} (sessionKey=${sessionKey}, captureAssistant=${config.captureAssistant === true}, ${summarizeAgentEndMessages(event.messages)})`);
+                        api.logger.debug(`memory-lancedb-pro: auto-capture agent_end payload for agent ${agentId} (sessionKey=${sessionKey}, captureAssistant=${JSON.stringify(config.captureAssistant)}, ${summarizeAgentEndMessages(event.messages)})`);
                         // Extract text content from messages
                         const eligibleTexts = [];
+                        const assistantContextTexts = [];
+                        const conversationTurns = [];
                         let skippedAutoCaptureTexts = 0;
+                        const captureAssistantValue = config.captureAssistant;
+                        const captureAssistantEligible = captureAssistantValue === true;
+                        const captureAssistantAsContext = captureAssistantValue === "context";
                         for (const msg of event.messages) {
                             if (!msg || typeof msg !== "object") {
                                 continue;
                             }
                             const msgObj = msg;
                             const role = msgObj.role;
-                            const captureAssistant = config.captureAssistant === true;
-                            if (role !== "user" &&
-                                !(captureAssistant && role === "assistant")) {
+                            const isEligibleRole = role === "user" || (captureAssistantEligible && role === "assistant");
+                            const isContextOnlyRole = captureAssistantAsContext && role === "assistant";
+                            if (!isEligibleRole && !isContextOnlyRole) {
                                 continue;
                             }
+                            const targetTexts = isEligibleRole ? eligibleTexts : assistantContextTexts;
                             const content = msgObj.content;
                             if (typeof content === "string") {
                                 const normalized = normalizeAutoCaptureText(role, content, shouldSkipReflectionMessage);
                                 if (!normalized) {
                                     skippedAutoCaptureTexts++;
                                 }
                                 else {
-                                    eligibleTexts.push(normalized);
+                                    targetTexts.push(normalized);
+                                    conversationTurns.push({ role: role, text: normalized });
                                 }
                                 continue;
                             }
@@ -2940,7 +2950,8 @@ const memoryLanceDBProPlugin = {
                                             skippedAutoCaptureTexts++;
                                         }
                                         else {
-                                            eligibleTexts.push(normalized);
+                                            targetTexts.push(normalized);
+                                            conversationTurns.push({ role: role, text: normalized });
                                         }
                                     }
                                 }
@@ -2978,6 +2989,24 @@ const memoryLanceDBProPlugin = {
                             pruneMapIfOver(autoCaptureRecentTexts, AUTO_CAPTURE_MAP_MAX_ENTRIES);
                         }
                         const minMessages = config.extractMinMessages ?? 4;
+                        // Rolling PAIR window (operator spec: extractMinMessages counts
+                        // user<->assistant pairs, and captureAssistant context rides the
+                        // same window). This call's new pairs -- kept user turns with the
+                        // assistant replies interleaved in true order -- extend what
+                        // earlier calls buffered, bounded to extractMinMessages user turns
+                        // (or this call's own new-user count when larger, so unextracted
+                        // user turns are never trimmed out of their own transcript).
+                        const thisCallPairTurns = buildConversationTurnsForExtraction({
+                            messageLoopTurns: conversationTurns,
+                            eligibleTexts,
+                            newUserTexts: newTexts,
+                        });
+                        const priorPairTurns = autoCaptureRecentPairTurns.get(sessionKey) || [];
+                        const pairWindowTurns = trimTurnsToUserCap(dedupePairWindow([...priorPairTurns, ...thisCallPairTurns]), Math.max(minMessages, thisCallPairTurns.filter((turn) => turn.role === "user").length));
+                        if (thisCallPairTurns.length > 0) {
+                            autoCaptureRecentPairTurns.set(sessionKey, pairWindowTurns);
+                            pruneMapIfOver(autoCaptureRecentPairTurns, AUTO_CAPTURE_MAP_MAX_ENTRIES);
+                        }
                         if (skippedAutoCaptureTexts > 0) {
                             api.logger.debug(`memory-lancedb-pro: auto-capture skipped ${skippedAutoCaptureTexts} injected/system text block(s) for agent ${agentId}`);
                         }
@@ -3033,10 +3062,17 @@ const memoryLanceDBProPlugin = {
                             if (cumulativeCount >= minMessages) {
                                 api.logger.debug(`memory-lancedb-pro: auto-capture running smart extraction for agent ${agentId} (cumulative=${cumulativeCount} >= minMessages=${minMessages}, cleanTexts=${cleanTexts.length})`);
                                 const conversationText = cleanTexts.join("\n");
+                                // The pair window is the transcript; user turns the noise
+                                // filter dropped stay out of it so they cannot become sources.
+                                const noiseDroppedTexts = new Set(texts.filter((text) => !cleanTexts.includes(text)));
+                                const finalConversationTurns = pairWindowTurns.filter((turn) => !(turn.role === "user" && noiseDroppedTexts.has(turn.text)));
+                                const assistantWindowTexts = finalConversationTurns
+                                    .filter((turn) => turn.role === "assistant")
+                                    .map((turn) => turn.text);
                                 // issue #417 Fix #10: prevent hook crash on LLM API errors / network timeouts
                                 let stats = null;
                                 try {
-                                    stats = await smartExtractor.extractAndPersist(conversationText, sessionKey, { scope: defaultScope, scopeFilter: accessibleScopes, agentId });
+                                    stats = await smartExtractor.extractAndPersist(conversationText, sessionKey, { scope: defaultScope, scopeFilter: accessibleScopes, agentId, assistantContextTexts: assistantWindowTexts, conversationTurns: finalConversationTurns });
                                 }
                                 catch (err) {
                                     api.logger.error(`memory-lancedb-pro: smart-extract failed for agent ${agentId}: ${String(err)}`);
@@ -3056,6 +3092,7 @@ const memoryLanceDBProPlugin = {
                                     // consumed history length there instead, so the next turn
                                     // only sees the delta.
                                     autoCaptureSeenTextCount.set(sessionKey, pendingIngressTexts.length > 0 ? 0 : eligibleTexts.length);
+                                    autoCaptureRecentPairTurns.delete(sessionKey);
                                     return; // Smart extraction handled everything
                                 }
                                 if ((stats.boundarySkipped ?? 0) === 0) {
@@ -4066,10 +4103,15 @@ const memoryLanceDBProPlugin = {
                 const now = new Date(params.timestampMs ?? Date.now());
                 const dateStr = now.toISOString().split("T")[0];
                 const timeStr = now.toISOString().split("T")[1].split(".")[0];
+                // Session key/id stay out of `text`: it is the FTS index surface, and
+                // the `simple` tokenizer splits a key like
+                // `agent:main:cron:<uuid>:run:<uuid>` on its punctuation — so every session
+                // summary ends up indexed under `agent`, `main`, `cron`, `run`. A query
+                // mentioning any of those then BM25-matches every session summary in the
+                // store regardless of content. Both ids are already recorded structurally
+                // in metadata below, so provenance is unaffected.
                 const memoryText = [
                     `Session: ${dateStr} ${timeStr} UTC`,
-                    `Session Key: ${params.sessionKey}`,
-                    `Session ID: ${params.sessionId}`,
                     `Source: ${params.source}`,
                     "",
                     "Conversation Summary:",
@@ -4685,7 +4727,7 @@ export function parsePluginConfig(value) {
             }
             return s;
         })(),
-        captureAssistant: cfg.captureAssistant === true,
+        captureAssistant: cfg.captureAssistant === "context" ? "context" : cfg.captureAssistant === true,
         retrieval: typeof cfg.retrieval === "object" && cfg.retrieval !== null
             ? (() => {
                 const retrieval = { ...cfg.retrieval };