agenr 0.8.8 → 0.8.9

package/CHANGELOG.md

@@ -1,5 +1,18 @@
 # Changelog
 
+## [0.8.9]
+
+### Added
+- feat(extractor): broadened extraction prompt to capture personal user context (health, diet, family, occupation, location, values) even from casual or passing mentions; added 6-month durability test heuristic to distinguish durable personal facts from transient states (issue #173)
+- feat(extractor): new few-shot examples for RELATIONSHIP, PREFERENCE, FACT, and EVENT types covering personal context scenarios with scoring rationale
+
+### Fixed
+- fix(ingest): suppress redundant whole-file ignored-params warning; now fires once per ingest run via shared ExtractRunOnceFlags object instead of once per file (issue #168)
+- fix(ingest): silence SQLITE_ERROR vector-index-not-found pre-fetch error during bulk ingest when vector index is intentionally absent; all other pre-fetch errors still log (issue #168)
+- fix(ingest): detect .jsonl.reset.TIMESTAMP session files as JSONL adapter by extending suffix-stripping regex to handle both .deleted and .reset suffixes (issue #169)
+- fix(consolidate): added merge system prompt constraint that expiry must be exactly permanent or temporary, never a date or timestamp; complements existing runtime fallback (issue #172)
+- fix(daemon): daemon install plist now uses the runtime CLI path resolved from argv[1] via the injected argvFn, preventing hardcoded npm global paths from breaking pnpm installs (issue #174)
+
 ## [0.8.8]
 
 ### Fixed

package/dist/cli-main.d.ts

@@ -170,6 +170,16 @@ interface ExtractChunkCompleteResult {
     durationMs?: number;
     warnings: string[];
 }
+/**
+ * Shared mutable state object passed across multiple extractKnowledgeFromChunks
+ * calls (e.g., across files in a single ingest run) to suppress duplicate
+ * one-time warnings. When undefined, all warnings fire on every call,
+ * preserving backward-compatible behavior for callers that do not share state.
+ */
+interface ExtractRunOnceFlags {
+    /** Set to true after the whole-file ignored-params warning fires once. */
+    hasWarnedWholeFileIgnoredParams?: boolean;
+}
 declare function extractKnowledgeFromChunks(params: {
     file: string;
     chunks: TranscriptChunk[];
@@ -192,6 +202,8 @@ declare function extractKnowledgeFromChunks(params: {
     embeddingApiKey?: string;
     noPreFetch?: boolean;
     embedFn?: (texts: string[], apiKey: string) => Promise<number[][]>;
+    /** Shared warning state across calls; pass one object per ingest run to dedupe one-time logs. */
+    onceFlags?: ExtractRunOnceFlags;
 }): Promise<ExtractChunksResult>;
 
 interface ResolveLlmClientInput {

package/dist/cli-main.js

@@ -7088,7 +7088,8 @@ function buildMergeContext(cluster) {
     "Merge the provided related entries into one canonical entry.",
     "Only include information explicitly stated in the source entries. Do not infer or add details not present.",
     "Prefer preserving temporal changes in the merged narrative.",
-    "Call merge_entries with your final merged result."
+    "Call merge_entries with your final merged result.",
+    'expiry must be exactly the string "permanent" or "temporary" -- never a date, timestamp, or other value.'
   ].join("\n");
   let contentLimit;
   let payload = formatClusterEntries(cluster, contentLimit);
@@ -11164,7 +11165,7 @@ var vscodeCopilotAdapter = {
 
 // src/adapters/registry.ts
 async function detectAdapter(filePath) {
-  const strippedPath = filePath.replace(/\.deleted\.[^/\\]+$/, "");
+  const strippedPath = filePath.replace(/\.(deleted|reset)\.[^/\\]+$/, "");
   const ext = path23.extname(strippedPath).toLowerCase();
   if (ext === ".jsonl") {
     const firstLine = await readFirstNonEmptyLine(filePath);
@@ -11725,15 +11726,15 @@ var SUBMIT_DEDUPED_KNOWLEDGE_TOOL = {
 // src/extractor.ts
 var SYSTEM_PROMPT = `You are a selective memory extraction engine. Extract only knowledge worth remembering beyond the immediate step.
 
-Default action: SKIP. Most chunks should produce zero entries.
+Default action: SKIP. Most chunks should produce zero entries. Personal disclosures (health, diet, family structure, pets, occupation, location, values, relationships) are high-signal -- do not skip them just because the mention is brief or incidental. Apply the durability gate below to all entries including personal ones.
 
 ## Types
 
-FACT \u2014 Verifiable information about a system, project, person, or concept.
+FACT \u2014 Verifiable information about a system, project, person, or concept. Personal facts count as first-class entries: health conditions, family members, pets, where the user lives or works, occupation, hobbies, recurring habits, and lifestyle circumstances.
 DECISION \u2014 A choice that constrains future options. Requires BOTH the choice AND the rationale. If rationale is missing, use fact or event instead.
 PREFERENCE \u2014 A stated or demonstrated preference that should influence future behavior.
 LESSON \u2014 An insight from experience that should change future behavior.
-EVENT \u2014 A significant milestone, launch, or completion. NOT "the assistant ran git status."
+EVENT \u2014 A significant milestone, launch, completion, or one-time life moment. Includes project launches, deployments, merges; and personal milestones like starting a new job, relocating, major health events, or notable personal experiences. NOT recurring habits or routines (those are FACT or PREFERENCE). NOT "the assistant ran git status." NOT vague references with no anchoring detail. If a personal behavior is described as newly begun, prefer EVENT for the initiation; use FACT if the stable ongoing state is also established in the same context.
 RELATIONSHIP \u2014 A connection between named entities. Content must include both entities and the relation.
 TODO \u2014 A persistent future action not completed in this chunk and not a one-step session instruction.
 
@@ -11756,7 +11757,7 @@ Do NOT emit a completion event for:
 ## Durability Gate
 
 Only extract if useful in future conversations/tasks after the current immediate execution.
-If uncertain whether durable, skip.
+If uncertain whether durable, apply the 6-month test: would knowing this fact 6 months from now still help an AI assist this person? Health conditions, family structure, location, diet, pets, occupation, values -> YES, extract. Current mood, today's plans, this week's weather -> NO, skip. Personal disclosures are durable by default -- if a user reveals something true about themselves, capture it even if mentioned casually or as an aside.
 
 ## Importance (1-10)
 
@@ -11843,8 +11844,9 @@ If you cannot name a concrete topic, skip the entry.
 5. Code-level implementation details likely to churn (unless architecture-level decision)
 6. One workflow split into multiple near-duplicate entries \u2014 merge into one
 7. Minor rephrases/duplicates of another extracted entry
-8. Greetings, acknowledgments, small talk
+8. Pure pleasantries with no personal content: greetings, acknowledgments, filler phrases that reveal nothing about the user ("how are you", "thanks", "sounds good", "got it"). Do NOT suppress personal disclosures just because they appear inside casual phrasing or as asides. "I follow keto so no carbs -- anyway, back to the project" contains an extractable preference even though the user pivoted away. The pivot does not cancel the fact. Test: does the underlying disclosure reveal something true and durable about the user? If yes, extract it.
 9. Transient implementation status unless it represents a milestone, decision, or lesson
+10. Transient personal states: current mood, today's fatigue, this week's busyness, temporary travel plans, passing weather references. These are not personal facts. Extract only if the condition is recurring or structural ("I'm always exhausted" may indicate a chronic health pattern worth capturing; "I'm tired today" does not).
 
 ## Explicit Memory Requests
 
@@ -11949,6 +11951,17 @@ RELATIONSHIP:
   "source_context": "Architecture discussion about memory integration"
 }
 
+RELATIONSHIP:
+{
+  "type": "relationship",
+  "subject": "user and sister Sarah",
+  "content": "User's sister Sarah lives nearby and they see each other regularly. She helps with childcare.",
+  "importance": 7,
+  "expiry": "permanent",
+  "tags": ["family", "personal", "relationship"],
+  "source_context": "User mentioned sister while discussing weekend plans"
+}
+
 TODO:
 {
   "type": "todo",
@@ -12015,6 +12028,39 @@ PREFERENCE:
   "source_context": "User mentioned scheduling preference during calendar discussion -- scored 6 not 8 because no parallel session needs to act on this immediately; it is a low-urgency convenience preference"
 }
 
+PREFERENCE:
+{
+  "type": "preference",
+  "subject": "user dietary preference",
+  "content": "User follows a strict ketogenic diet and avoids carbohydrates. Do not suggest high-carb meals, recipes, or foods.",
+  "importance": 7,
+  "expiry": "permanent",
+  "tags": ["diet", "keto", "personal", "health"],
+  "source_context": "User mentioned diet as an aside mid-conversation -- scored 7 because future sessions about food, health, or restaurants need this; the casual phrasing does not reduce its durability"
+}
+
+FACT:
+{
+  "type": "fact",
+  "subject": "user morning routine",
+  "content": "User wakes at 6:15 AM and goes to the gym every weekday morning before work.",
+  "importance": 6,
+  "expiry": "permanent",
+  "tags": ["routine", "health", "personal"],
+  "source_context": "User described morning routine while discussing their schedule -- scored 6 not 8 because this is low-urgency biographical context; no parallel session needs to act on it immediately"
+}
+
+EVENT:
+{
+  "type": "event",
+  "subject": "user job change",
+  "content": "User started a new senior engineering role at a new company. This is a recent career change.",
+  "importance": 7,
+  "expiry": "permanent",
+  "tags": ["career", "work", "personal"],
+  "source_context": "User mentioned new job while discussing their schedule -- scored 7 not 9 because this is a significant life milestone worth preserving but does not require immediate cross-session action"
+}
+
 // NOTE: These examples are drawn from OpenClaw transcripts (agent role labels, tool-verified claims). They provide soft cross-platform guidance for hedged-claim handling; mechanical enforcement (importance cap + unverified tag) is applied for openclaw, codex, and claude-code via applyConfidenceCap().
 Example: hedged agent claim (correct handling):
 {
@@ -12459,6 +12505,10 @@ var ParseResponseError = class extends Error {
 function normalize3(value) {
   return value.trim().toLowerCase();
 }
+function isVectorIndexNotFoundMessage(message) {
+  const normalized = message.toLowerCase();
+  return normalized.includes("vector index") && normalized.includes("not found");
+}
 async function preFetchRelated(chunkText2, db, embeddingApiKey, embedFn = embed, onVerbose) {
   const run = async () => {
     try {
@@ -12490,7 +12540,11 @@ async function preFetchRelated(chunkText2, db, embeddingApiKey, embedFn = embed,
       onVerbose?.(`[pre-fetch] ${above.length} above threshold ${PREFETCH_SIMILARITY_THRESHOLD}`);
       return above.slice(0, MAX_PREFETCH_RESULTS).map((candidate) => candidate.entry);
     } catch (error) {
-      onVerbose?.(`[pre-fetch] skipped: ${error instanceof Error ? error.message : String(error)}`);
+      const message = error instanceof Error ? error.message : String(error);
+      if (isVectorIndexNotFoundMessage(message)) {
+        return [];
+      }
+      onVerbose?.(`[pre-fetch] skipped: ${message}`);
       return [];
     }
   };
@@ -13251,10 +13305,15 @@ async function extractKnowledgeFromChunks(params) {
   let effectiveNoPreFetch = wholeFileMode ? true : params.noPreFetch ?? false;
   if (wholeFileMode && params.verbose) {
     const ignoredParamsLine = "[whole-file] interChunkDelayMs and llmConcurrency have no effect in whole-file mode";
-    if (params.onVerbose) {
-      params.onVerbose(ignoredParamsLine);
-    } else {
-      console.warn(ignoredParamsLine);
+    if (!params.onceFlags?.hasWarnedWholeFileIgnoredParams) {
+      if (params.onVerbose) {
+        params.onVerbose(ignoredParamsLine);
+      } else {
+        console.warn(ignoredParamsLine);
+      }
+      if (params.onceFlags) {
+        params.onceFlags.hasWarnedWholeFileIgnoredParams = true;
+      }
     }
   }
   if (wholeFileMode) {
@@ -14545,6 +14604,7 @@ async function runIngestCommand(inputPaths, options, deps) {
   let totalChunksFailed = 0;
   let filesWithChunkFailures = 0;
   const chunkStatsByFile = /* @__PURE__ */ new Map();
+  const extractOnceFlags = { hasWarnedWholeFileIgnoredParams: false };
   let firstPassFailedIndexSet = /* @__PURE__ */ new Set();
   let bulkTeardownComplete = false;
   let bulkVectorRebuildDurationSeconds = null;
@@ -14699,6 +14759,7 @@ async function runIngestCommand(inputPaths, options, deps) {
         db: options.noPreFetch ? void 0 : db,
         embeddingApiKey: options.noPreFetch ? void 0 : embeddingApiKey ?? void 0,
         noPreFetch: options.noPreFetch === true,
+        onceFlags: extractOnceFlags,
         onVerbose: verbose ? (line) => {
           clack4.log.info(line, clackOutput);
         } : void 0,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agenr",
-  "version": "0.8.8",
+  "version": "0.8.9",
   "openclaw": {
     "extensions": [
       "dist/openclaw-plugin/index.js"