@anuma/sdk 6.18.0 → 6.19.1

package/dist/react/index.cjs

@@ -6545,7 +6545,16 @@ async function callPortalJsonCompletion(req) {
       ...reinforce && { reinforceJsonContract: true }
     } : reinforce ? { ...req, reinforceJsonContract: true } : req;
     const outcome = await attemptPortalJson(attemptReq, endpoint);
-    if (outcome.kind === "ok") return outcome.value;
+    if (outcome.kind === "ok") {
+      req.onAttempt?.({ attempt, ok: true });
+      return outcome.value;
+    }
+    req.onAttempt?.({
+      attempt,
+      ok: false,
+      reason: outcome.code,
+      ...outcome.httpStatus !== void 0 && { httpStatus: outcome.httpStatus }
+    });
     lastFailure = {
       reason: outcome.code,
       ...outcome.httpStatus !== void 0 && { httpStatus: outcome.httpStatus },
@@ -8568,7 +8577,7 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
   let tailSlice = [];
   const itemById = new Map(items.map((i) => [i.id, i]));
   if (options?.rerank) {
-    const rerankTopN = options.rerankTopN ?? 30;
+    const rerankTopN = options.rerankTopN ?? DEFAULT_RERANK_TOP_N;
     const ceWeight = options.ceWeight ?? 0.1;
     const headSlice = v2Ranked.slice(0, rerankTopN);
     const ceStart = nowMs2();
@@ -8614,9 +8623,12 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
   }
   if (sideLanesAsync.length > 0) {
     const headIds = combined.map((r) => r.uniqueId);
-    const fused = rrfFuse([headIds, headIds, ...sideLanesAsync], options?.rrfK);
-    const fusedHead = combined.map((r) => ({ ...r, similarity: fused.get(r.uniqueId) ?? r.similarity })).sort((a, b) => b.similarity - a.similarity);
-    const seen = new Set(combined.map((r) => r.uniqueId));
+    const tailIds = tailSlice.map((r) => r.uniqueId);
+    const rankedIds = tailIds.length > 0 ? [...headIds, ...tailIds] : headIds;
+    const fused = rrfFuse([rankedIds, rankedIds, ...sideLanesAsync], options?.rrfK);
+    const fusedHead = [...combined, ...tailSlice].map((r) => ({ ...r, similarity: fused.get(r.uniqueId) ?? r.similarity })).sort((a, b) => b.similarity - a.similarity);
+    tailSlice = [];
+    const seen = new Set(fusedHead.map((r) => r.uniqueId));
     const itemById2 = new Map(items.map((i) => [i.id, i]));
     for (const lane of sideLanesAsync) {
       for (const id of lane) {
@@ -8641,7 +8653,7 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
     const lambda = options.mmrLambda ?? 0.7;
     const mmrTopN = options.mmrTopN ?? 20;
     const itemById2 = new Map(items.map((i) => [i.id, i]));
-    const mmrCandidates = combined.slice(0, mmrTopN).map((r) => ({
+    const mmrCandidates = [...combined, ...tailSlice].slice(0, mmrTopN).map((r) => ({
       id: r.uniqueId,
       score: r.similarity,
       embedding: itemById2.get(r.uniqueId)?.embedding ?? [],
@@ -8649,7 +8661,7 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
     }));
     const picked = applyMMR(mmrCandidates, limit, lambda);
     const pickedIds = new Set(picked.map((p) => p.id));
-    const resultMap = new Map(combined.map((r) => [r.uniqueId, r]));
+    const resultMap = new Map([...combined, ...tailSlice].map((r) => [r.uniqueId, r]));
     const pickedResults = picked.map((p) => {
       const orig = resultMap.get(p.id);
       return {
@@ -8763,7 +8775,7 @@ async function rankComposite(originalQuery, originalQueryEmbedding, subQueries,
     }
   }
   if (options?.rerank && combined.length > 0) {
-    const rerankTopN = options.rerankTopN ?? 30;
+    const rerankTopN = options.rerankTopN ?? DEFAULT_RERANK_TOP_N;
     const ceWeight = options.ceWeight ?? 0.1;
     const headSlice = combined.slice(0, rerankTopN);
     const tailSlice = combined.slice(rerankTopN);
@@ -9648,7 +9660,7 @@ ${formatted}`;
     }
   };
 }
-var nowMs2, warnedProgrammaticDecomposeIgnored, SUPERSESSION_SIMILARITY_THRESHOLD, SUPERSESSION_MIN_AGE_GAP_MS, SUPERSESSION_BOOST_FACTOR, SUPERSESSION_MAX_WINDOW, EMBEDDINGS_DEGRADED_EMPTY_NO_LEXICAL;
+var DEFAULT_RERANK_TOP_N, nowMs2, warnedProgrammaticDecomposeIgnored, SUPERSESSION_SIMILARITY_THRESHOLD, SUPERSESSION_MIN_AGE_GAP_MS, SUPERSESSION_BOOST_FACTOR, SUPERSESSION_MAX_WINDOW, EMBEDDINGS_DEGRADED_EMPTY_NO_LEXICAL;
 var init_searchTool = __esm({
   "src/lib/memoryVault/searchTool.ts"() {
     "use strict";
@@ -9666,6 +9678,7 @@ var init_searchTool = __esm({
     init_bm25();
     init_decomposeQuery();
     init_lruCache();
+    DEFAULT_RERANK_TOP_N = 5;
     nowMs2 = () => typeof performance !== "undefined" && typeof performance.now === "function" ? performance.now() : Date.now();
     warnedProgrammaticDecomposeIgnored = false;
     SUPERSESSION_SIMILARITY_THRESHOLD = 0.7;
@@ -18159,7 +18172,8 @@ Extract durable user facts.`,
     ...options.maxAttempts !== void 0 && { maxAttempts: options.maxAttempts },
     ...options.timeoutMs !== void 0 && { timeoutMs: options.timeoutMs },
     ...options.totalTimeoutMs !== void 0 && { totalTimeoutMs: options.totalTimeoutMs },
-    ...options.backoffMs && { backoffMs: options.backoffMs }
+    ...options.backoffMs && { backoffMs: options.backoffMs },
+    ...options.onAttempt && { onAttempt: options.onAttempt }
   });
   if (parsed === null) {
     options.onExhaustedEmpty?.(failure ?? { reason: "empty-content", attempts: 1 });

package/dist/react/index.d.mts

@@ -6271,7 +6271,7 @@ declare function getEntitiesByMemoryIdsOp(ctx: EntityOperationsContext, memoryId
  *
  * @public
  */
-type TaskType = "title" | "commit_message" | "memory_extract" | "memory_dedup" | "memory_decay" | "memory_injection_check" | "memory_verify_support" | "memory_autosort" | "memory_topic" | "memory_consolidate" | "memory_decompose" | "memory_profile_synth" | "memory_graph" | "folder_context" | "selection_memory" | "summarize" | "style_analysis" | "classify_search" | "app_inspiration_generate" | "app_inspiration_trends" | "app_inspiration_remix" | "spotlight_image" | "slide_image" | "slide_video" | "slide_music" | "slide_sfx" | "media_video" | "media_audio" | "media_bg_removal" | "media_image_preset" | "tool_app_builder" | "tool_document_builder" | "tool_slides" | "tool_slides_image_gen" | "connector_guidance" | "cf_task" | "xbot_intent" | "xbot_council" | "scheduled_task";
+type TaskType = "title" | "commit_message" | "memory_extract" | "memory_import_extract" | "memory_dedup" | "memory_decay" | "memory_injection_check" | "memory_verify_support" | "memory_autosort" | "memory_topic" | "memory_consolidate" | "memory_decompose" | "memory_profile_synth" | "memory_graph" | "folder_context" | "selection_memory" | "summarize" | "style_analysis" | "classify_search" | "app_inspiration_generate" | "app_inspiration_trends" | "app_inspiration_remix" | "spotlight_image" | "slide_image" | "slide_video" | "slide_music" | "slide_sfx" | "media_video" | "media_audio" | "media_bg_removal" | "media_image_preset" | "tool_app_builder" | "tool_document_builder" | "tool_slides" | "tool_slides_image_gen" | "connector_guidance" | "cf_task" | "xbot_intent" | "xbot_council" | "scheduled_task";
 
 /**
  * Shared portal LLM call for the memory pipeline.
@@ -6368,6 +6368,26 @@ type PortalLlmFailureReason =
  | "null-completion"
 /** `totalTimeoutMs` was spent before the next attempt could run. */
  | "time-budget-exhausted";
+/**
+ * One attempt of a {@link callPortalJsonCompletion} call, reported to
+ * `onAttempt` as it settles — success included. The wire-level diagnostic the
+ * give-up hook cannot provide: a call that succeeds on its third try returns a
+ * value and fires no `onFailure`, yet it cost three completions and its first
+ * two answers were unusable. Extraction quality regressions have hidden in
+ * exactly that shape (a prompt change that made the model answer `NONE` first,
+ * JSON on retry), so the extraction eval gates on the first-attempt clean rate.
+ *
+ * @public
+ */
+interface PortalLlmAttempt {
+    /** 1-based attempt index. */
+    attempt: number;
+    /** Whether this attempt produced a parseable JSON value. */
+    ok: boolean;
+    /** Classification when `ok` is false. */
+    reason?: PortalLlmFailureReason;
+    httpStatus?: number;
+}
 /**
  * A give-up report: the classified {@link PortalLlmFailureReason} plus the
  * little context worth carrying into telemetry. Both extra fields are bounded
@@ -6624,7 +6644,7 @@ interface MemoryVaultSearchOptions {
      * When true, switches to the async pipeline (rankFusedVaultMemoriesAsync).
      */
     rerank?: boolean;
-    /** Number of CE rerank candidates. Default 30. */
+    /** Number of CE rerank candidates. Default 5 ({@link DEFAULT_RERANK_TOP_N}). */
     rerankTopN?: number;
     /** Multiplicative cross-encoder blend weight. Default 0.1. Only used when `rerank` is true. */
     ceWeight?: number;
@@ -6979,7 +6999,8 @@ interface RecallOptions {
      * swallowed (diagnostics must never break retrieval). Off unless provided.
      */
     onDiagnostics?: (diagnostics: RecallDiagnostics) => void;
-    /** Number of candidates fed to the cross-encoder rerank stage. Default: 30. */
+    /** Number of candidates fed to the cross-encoder rerank stage. Default: 5;
+     *  was 30 until 2026-08-13 — see anuma-ai/sdk#845. */
     rerankTopN?: number;
     /** Multiplicative cross-encoder blend weight. Default: 0.1. */
     ceWeight?: number;
@@ -7547,6 +7568,12 @@ interface ExtractFactsOptions extends PortalLlmAuth {
      * (i.e. redaction silently eating facts) is alarmable.
      */
     onCandidatesDropped?: () => void;
+    /**
+     * Per-attempt wire diagnostic, forwarded to {@link callPortalJsonCompletion}.
+     * The extraction eval gates on it (first-attempt clean rate); production
+     * callers may leave it unset.
+     */
+    onAttempt?: (attempt: PortalLlmAttempt) => void;
 }
 /**
  * Stage 1 — call the LLM to extract candidate facts from the recent

package/dist/react/index.d.ts

@@ -6271,7 +6271,7 @@ declare function getEntitiesByMemoryIdsOp(ctx: EntityOperationsContext, memoryId
  *
  * @public
  */
-type TaskType = "title" | "commit_message" | "memory_extract" | "memory_dedup" | "memory_decay" | "memory_injection_check" | "memory_verify_support" | "memory_autosort" | "memory_topic" | "memory_consolidate" | "memory_decompose" | "memory_profile_synth" | "memory_graph" | "folder_context" | "selection_memory" | "summarize" | "style_analysis" | "classify_search" | "app_inspiration_generate" | "app_inspiration_trends" | "app_inspiration_remix" | "spotlight_image" | "slide_image" | "slide_video" | "slide_music" | "slide_sfx" | "media_video" | "media_audio" | "media_bg_removal" | "media_image_preset" | "tool_app_builder" | "tool_document_builder" | "tool_slides" | "tool_slides_image_gen" | "connector_guidance" | "cf_task" | "xbot_intent" | "xbot_council" | "scheduled_task";
+type TaskType = "title" | "commit_message" | "memory_extract" | "memory_import_extract" | "memory_dedup" | "memory_decay" | "memory_injection_check" | "memory_verify_support" | "memory_autosort" | "memory_topic" | "memory_consolidate" | "memory_decompose" | "memory_profile_synth" | "memory_graph" | "folder_context" | "selection_memory" | "summarize" | "style_analysis" | "classify_search" | "app_inspiration_generate" | "app_inspiration_trends" | "app_inspiration_remix" | "spotlight_image" | "slide_image" | "slide_video" | "slide_music" | "slide_sfx" | "media_video" | "media_audio" | "media_bg_removal" | "media_image_preset" | "tool_app_builder" | "tool_document_builder" | "tool_slides" | "tool_slides_image_gen" | "connector_guidance" | "cf_task" | "xbot_intent" | "xbot_council" | "scheduled_task";
 
 /**
  * Shared portal LLM call for the memory pipeline.
@@ -6368,6 +6368,26 @@ type PortalLlmFailureReason =
  | "null-completion"
 /** `totalTimeoutMs` was spent before the next attempt could run. */
  | "time-budget-exhausted";
+/**
+ * One attempt of a {@link callPortalJsonCompletion} call, reported to
+ * `onAttempt` as it settles — success included. The wire-level diagnostic the
+ * give-up hook cannot provide: a call that succeeds on its third try returns a
+ * value and fires no `onFailure`, yet it cost three completions and its first
+ * two answers were unusable. Extraction quality regressions have hidden in
+ * exactly that shape (a prompt change that made the model answer `NONE` first,
+ * JSON on retry), so the extraction eval gates on the first-attempt clean rate.
+ *
+ * @public
+ */
+interface PortalLlmAttempt {
+    /** 1-based attempt index. */
+    attempt: number;
+    /** Whether this attempt produced a parseable JSON value. */
+    ok: boolean;
+    /** Classification when `ok` is false. */
+    reason?: PortalLlmFailureReason;
+    httpStatus?: number;
+}
 /**
  * A give-up report: the classified {@link PortalLlmFailureReason} plus the
  * little context worth carrying into telemetry. Both extra fields are bounded
@@ -6624,7 +6644,7 @@ interface MemoryVaultSearchOptions {
      * When true, switches to the async pipeline (rankFusedVaultMemoriesAsync).
      */
     rerank?: boolean;
-    /** Number of CE rerank candidates. Default 30. */
+    /** Number of CE rerank candidates. Default 5 ({@link DEFAULT_RERANK_TOP_N}). */
     rerankTopN?: number;
     /** Multiplicative cross-encoder blend weight. Default 0.1. Only used when `rerank` is true. */
     ceWeight?: number;
@@ -6979,7 +6999,8 @@ interface RecallOptions {
      * swallowed (diagnostics must never break retrieval). Off unless provided.
      */
     onDiagnostics?: (diagnostics: RecallDiagnostics) => void;
-    /** Number of candidates fed to the cross-encoder rerank stage. Default: 30. */
+    /** Number of candidates fed to the cross-encoder rerank stage. Default: 5;
+     *  was 30 until 2026-08-13 — see anuma-ai/sdk#845. */
     rerankTopN?: number;
     /** Multiplicative cross-encoder blend weight. Default: 0.1. */
     ceWeight?: number;
@@ -7547,6 +7568,12 @@ interface ExtractFactsOptions extends PortalLlmAuth {
      * (i.e. redaction silently eating facts) is alarmable.
      */
     onCandidatesDropped?: () => void;
+    /**
+     * Per-attempt wire diagnostic, forwarded to {@link callPortalJsonCompletion}.
+     * The extraction eval gates on it (first-attempt clean rate); production
+     * callers may leave it unset.
+     */
+    onAttempt?: (attempt: PortalLlmAttempt) => void;
 }
 /**
  * Stage 1 — call the LLM to extract candidate facts from the recent

package/dist/react/index.mjs

@@ -179,14 +179,14 @@ import {
   updateVaultMemoryOp,
   useEncryption,
   vaultMemoryToStored
-} from "./chunk-TVPCH7ZZ.mjs";
+} from "./chunk-2EEPAOHJ.mjs";
 import {
   PiiRedactor,
   createStreamingDeAnonymizer,
   isPiiRedactor,
   notifyConsolidationFallback,
   resolvePiiRedactor
-} from "./chunk-7X7BNMGS.mjs";
+} from "./chunk-AVYB5IXC.mjs";
 import {
   INTERNAL_FLOW_MARKER,
   callPortalJsonCompletion,
@@ -202,7 +202,7 @@ import {
   taskTypeHeader,
   validateEndpointOverride,
   withInternalFlowMarker
-} from "./chunk-4Z2X4UYQ.mjs";
+} from "./chunk-WUG6GTHO.mjs";
 import {
   BASE_URL,
   MCP_R2_DOMAIN,
@@ -7711,7 +7711,7 @@ async function tryConsolidate(trimmed, ctx, options, prepared) {
     content: m.content,
     similarity: m.similarity
   }));
-  const { consolidateMemory: doConsolidate } = await import("./consolidate-7NMKF7UC.mjs");
+  const { consolidateMemory: doConsolidate } = await import("./consolidate-URHGJT7I.mjs");
   const decision = await doConsolidate(trimmed, candidates, consolidateOptions);
   if (decision.action === "create") return null;
   if (decision.action === "supersede" && decision.content) {
@@ -7953,7 +7953,8 @@ Extract durable user facts.`,
     ...options.maxAttempts !== void 0 && { maxAttempts: options.maxAttempts },
     ...options.timeoutMs !== void 0 && { timeoutMs: options.timeoutMs },
     ...options.totalTimeoutMs !== void 0 && { totalTimeoutMs: options.totalTimeoutMs },
-    ...options.backoffMs && { backoffMs: options.backoffMs }
+    ...options.backoffMs && { backoffMs: options.backoffMs },
+    ...options.onAttempt && { onAttempt: options.onAttempt }
   });
   if (parsed === null) {
     options.onExhaustedEmpty?.(failure ?? { reason: "empty-content", attempts: 1 });

package/dist/react/{recall-OKNLU335.mjs → recall-D6ODL36R.mjs}

@@ -1,7 +1,7 @@
 import {
   recall
-} from "./chunk-TVPCH7ZZ.mjs";
-import "./chunk-4Z2X4UYQ.mjs";
+} from "./chunk-2EEPAOHJ.mjs";
+import "./chunk-WUG6GTHO.mjs";
 import "./chunk-GYKJOWEO.mjs";
 export {
   recall

package/dist/server/{chunk-4LPSEROP.mjs → chunk-RKQ4MRS2.mjs}

@@ -1,6 +1,6 @@
 import {
   getLogger
-} from "./chunk-BSCEZOIW.mjs";
+} from "./chunk-YXL5VKCO.mjs";
 
 // src/lib/pii/patterns.ts
 function luhnCheck(digits) {

package/dist/server/{chunk-K7H7ICPK.mjs → chunk-SPHKT76E.mjs}

@@ -1,7 +1,7 @@
 import {
   callPortalJsonCompletion,
   getLogger
-} from "./chunk-BSCEZOIW.mjs";
+} from "./chunk-YXL5VKCO.mjs";
 import {
   BASE_URL,
   __commonJS,
@@ -19658,6 +19658,7 @@ function createVaultEmbeddingCache(maxSize = DEFAULT_VAULT_CACHE_SIZE) {
 }
 
 // src/lib/memoryVault/searchTool.ts
+var DEFAULT_RERANK_TOP_N = 5;
 var nowMs = () => typeof performance !== "undefined" && typeof performance.now === "function" ? performance.now() : Date.now();
 var warnedProgrammaticDecomposeIgnored = false;
 function normalizeEventTimeKind(kind) {
@@ -19888,7 +19889,7 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
   let tailSlice = [];
   const itemById = new Map(items.map((i) => [i.id, i]));
   if (options?.rerank) {
-    const rerankTopN = options.rerankTopN ?? 30;
+    const rerankTopN = options.rerankTopN ?? DEFAULT_RERANK_TOP_N;
     const ceWeight = options.ceWeight ?? 0.1;
     const headSlice = v2Ranked.slice(0, rerankTopN);
     const ceStart = nowMs();
@@ -19934,9 +19935,12 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
   }
   if (sideLanesAsync.length > 0) {
     const headIds = combined.map((r) => r.uniqueId);
-    const fused = rrfFuse([headIds, headIds, ...sideLanesAsync], options?.rrfK);
-    const fusedHead = combined.map((r) => ({ ...r, similarity: fused.get(r.uniqueId) ?? r.similarity })).sort((a, b) => b.similarity - a.similarity);
-    const seen = new Set(combined.map((r) => r.uniqueId));
+    const tailIds = tailSlice.map((r) => r.uniqueId);
+    const rankedIds = tailIds.length > 0 ? [...headIds, ...tailIds] : headIds;
+    const fused = rrfFuse([rankedIds, rankedIds, ...sideLanesAsync], options?.rrfK);
+    const fusedHead = [...combined, ...tailSlice].map((r) => ({ ...r, similarity: fused.get(r.uniqueId) ?? r.similarity })).sort((a, b) => b.similarity - a.similarity);
+    tailSlice = [];
+    const seen = new Set(fusedHead.map((r) => r.uniqueId));
     const itemById2 = new Map(items.map((i) => [i.id, i]));
     for (const lane of sideLanesAsync) {
       for (const id of lane) {
@@ -19961,7 +19965,7 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
     const lambda = options.mmrLambda ?? 0.7;
     const mmrTopN = options.mmrTopN ?? 20;
     const itemById2 = new Map(items.map((i) => [i.id, i]));
-    const mmrCandidates = combined.slice(0, mmrTopN).map((r) => ({
+    const mmrCandidates = [...combined, ...tailSlice].slice(0, mmrTopN).map((r) => ({
       id: r.uniqueId,
       score: r.similarity,
       embedding: itemById2.get(r.uniqueId)?.embedding ?? [],
@@ -19969,7 +19973,7 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
     }));
     const picked = applyMMR(mmrCandidates, limit, lambda);
     const pickedIds = new Set(picked.map((p) => p.id));
-    const resultMap = new Map(combined.map((r) => [r.uniqueId, r]));
+    const resultMap = new Map([...combined, ...tailSlice].map((r) => [r.uniqueId, r]));
     const pickedResults = picked.map((p) => {
       const orig = resultMap.get(p.id);
       return {
@@ -20083,7 +20087,7 @@ async function rankComposite(originalQuery, originalQueryEmbedding, subQueries,
     }
   }
   if (options?.rerank && combined.length > 0) {
-    const rerankTopN = options.rerankTopN ?? 30;
+    const rerankTopN = options.rerankTopN ?? DEFAULT_RERANK_TOP_N;
     const ceWeight = options.ceWeight ?? 0.1;
     const headSlice = combined.slice(0, rerankTopN);
     const tailSlice = combined.slice(rerankTopN);
@@ -20923,7 +20927,7 @@ function createMemoryVaultSearchTool(vaultCtx, embeddingOptions, cache, searchOp
             subQueries = decomp.subQueries;
           }
         }
-        const { recall: recall2 } = await import("./recall-YXJATOSD.mjs");
+        const { recall: recall2 } = await import("./recall-NLCZUGXN.mjs");
         let recallDegraded = [];
         const result = await recall2(
           query,

package/dist/server/{chunk-BSCEZOIW.mjs → chunk-YXL5VKCO.mjs}

@@ -159,7 +159,16 @@ async function callPortalJsonCompletion(req) {
       ...reinforce && { reinforceJsonContract: true }
     } : reinforce ? { ...req, reinforceJsonContract: true } : req;
     const outcome = await attemptPortalJson(attemptReq, endpoint);
-    if (outcome.kind === "ok") return outcome.value;
+    if (outcome.kind === "ok") {
+      req.onAttempt?.({ attempt, ok: true });
+      return outcome.value;
+    }
+    req.onAttempt?.({
+      attempt,
+      ok: false,
+      reason: outcome.code,
+      ...outcome.httpStatus !== void 0 && { httpStatus: outcome.httpStatus }
+    });
     lastFailure = {
       reason: outcome.code,
       ...outcome.httpStatus !== void 0 && { httpStatus: outcome.httpStatus },

package/dist/server/{consolidate-MSAXASHL.mjs → consolidate-54P6QZKU.mjs}

@@ -1,10 +1,10 @@
 import {
   notifyConsolidationFallback,
   resolvePiiRedactor
-} from "./chunk-4LPSEROP.mjs";
+} from "./chunk-RKQ4MRS2.mjs";
 import {
   callPortalJsonCompletion
-} from "./chunk-BSCEZOIW.mjs";
+} from "./chunk-YXL5VKCO.mjs";
 import "./chunk-3RPGXXPZ.mjs";
 
 // src/lib/memory/consolidate.ts

package/dist/server/index.cjs

@@ -20704,7 +20704,16 @@ async function callPortalJsonCompletion(req) {
       ...reinforce && { reinforceJsonContract: true }
     } : reinforce ? { ...req, reinforceJsonContract: true } : req;
     const outcome = await attemptPortalJson(attemptReq, endpoint);
-    if (outcome.kind === "ok") return outcome.value;
+    if (outcome.kind === "ok") {
+      req.onAttempt?.({ attempt, ok: true });
+      return outcome.value;
+    }
+    req.onAttempt?.({
+      attempt,
+      ok: false,
+      reason: outcome.code,
+      ...outcome.httpStatus !== void 0 && { httpStatus: outcome.httpStatus }
+    });
     lastFailure = {
       reason: outcome.code,
       ...outcome.httpStatus !== void 0 && { httpStatus: outcome.httpStatus },
@@ -22274,7 +22283,7 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
   let tailSlice = [];
   const itemById = new Map(items.map((i) => [i.id, i]));
   if (options?.rerank) {
-    const rerankTopN = options.rerankTopN ?? 30;
+    const rerankTopN = options.rerankTopN ?? DEFAULT_RERANK_TOP_N;
     const ceWeight = options.ceWeight ?? 0.1;
     const headSlice = v2Ranked.slice(0, rerankTopN);
     const ceStart = nowMs2();
@@ -22320,9 +22329,12 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
   }
   if (sideLanesAsync.length > 0) {
     const headIds = combined.map((r) => r.uniqueId);
-    const fused = rrfFuse([headIds, headIds, ...sideLanesAsync], options?.rrfK);
-    const fusedHead = combined.map((r) => ({ ...r, similarity: fused.get(r.uniqueId) ?? r.similarity })).sort((a, b) => b.similarity - a.similarity);
-    const seen = new Set(combined.map((r) => r.uniqueId));
+    const tailIds = tailSlice.map((r) => r.uniqueId);
+    const rankedIds = tailIds.length > 0 ? [...headIds, ...tailIds] : headIds;
+    const fused = rrfFuse([rankedIds, rankedIds, ...sideLanesAsync], options?.rrfK);
+    const fusedHead = [...combined, ...tailSlice].map((r) => ({ ...r, similarity: fused.get(r.uniqueId) ?? r.similarity })).sort((a, b) => b.similarity - a.similarity);
+    tailSlice = [];
+    const seen = new Set(fusedHead.map((r) => r.uniqueId));
     const itemById2 = new Map(items.map((i) => [i.id, i]));
     for (const lane of sideLanesAsync) {
       for (const id of lane) {
@@ -22347,7 +22359,7 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
     const lambda = options.mmrLambda ?? 0.7;
     const mmrTopN = options.mmrTopN ?? 20;
     const itemById2 = new Map(items.map((i) => [i.id, i]));
-    const mmrCandidates = combined.slice(0, mmrTopN).map((r) => ({
+    const mmrCandidates = [...combined, ...tailSlice].slice(0, mmrTopN).map((r) => ({
       id: r.uniqueId,
       score: r.similarity,
       embedding: itemById2.get(r.uniqueId)?.embedding ?? [],
@@ -22355,7 +22367,7 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
     }));
     const picked = applyMMR(mmrCandidates, limit, lambda);
     const pickedIds = new Set(picked.map((p) => p.id));
-    const resultMap = new Map(combined.map((r) => [r.uniqueId, r]));
+    const resultMap = new Map([...combined, ...tailSlice].map((r) => [r.uniqueId, r]));
     const pickedResults = picked.map((p) => {
       const orig = resultMap.get(p.id);
       return {
@@ -22469,7 +22481,7 @@ async function rankComposite(originalQuery, originalQueryEmbedding, subQueries,
     }
   }
   if (options?.rerank && combined.length > 0) {
-    const rerankTopN = options.rerankTopN ?? 30;
+    const rerankTopN = options.rerankTopN ?? DEFAULT_RERANK_TOP_N;
     const ceWeight = options.ceWeight ?? 0.1;
     const headSlice = combined.slice(0, rerankTopN);
     const tailSlice = combined.slice(rerankTopN);
@@ -23354,7 +23366,7 @@ ${formatted}`;
     }
   };
 }
-var nowMs2, warnedProgrammaticDecomposeIgnored, SUPERSESSION_SIMILARITY_THRESHOLD, SUPERSESSION_MIN_AGE_GAP_MS, SUPERSESSION_BOOST_FACTOR, SUPERSESSION_MAX_WINDOW, EMBEDDINGS_DEGRADED_EMPTY_NO_LEXICAL;
+var DEFAULT_RERANK_TOP_N, nowMs2, warnedProgrammaticDecomposeIgnored, SUPERSESSION_SIMILARITY_THRESHOLD, SUPERSESSION_MIN_AGE_GAP_MS, SUPERSESSION_BOOST_FACTOR, SUPERSESSION_MAX_WINDOW, EMBEDDINGS_DEGRADED_EMPTY_NO_LEXICAL;
 var init_searchTool = __esm({
   "src/lib/memoryVault/searchTool.ts"() {
     "use strict";
@@ -23372,6 +23384,7 @@ var init_searchTool = __esm({
     init_bm25();
     init_decomposeQuery();
     init_lruCache();
+    DEFAULT_RERANK_TOP_N = 5;
     nowMs2 = () => typeof performance !== "undefined" && typeof performance.now === "function" ? performance.now() : Date.now();
     warnedProgrammaticDecomposeIgnored = false;
     SUPERSESSION_SIMILARITY_THRESHOLD = 0.7;
@@ -29772,7 +29785,8 @@ Extract durable user facts.`,
     ...options.maxAttempts !== void 0 && { maxAttempts: options.maxAttempts },
     ...options.timeoutMs !== void 0 && { timeoutMs: options.timeoutMs },
     ...options.totalTimeoutMs !== void 0 && { totalTimeoutMs: options.totalTimeoutMs },
-    ...options.backoffMs && { backoffMs: options.backoffMs }
+    ...options.backoffMs && { backoffMs: options.backoffMs },
+    ...options.onAttempt && { onAttempt: options.onAttempt }
   });
   if (parsed === null) {
     options.onExhaustedEmpty?.(failure ?? { reason: "empty-content", attempts: 1 });

package/dist/server/index.d.mts

@@ -4766,7 +4766,7 @@ interface EntityOperationsContext {
  *
  * @public
  */
-type TaskType = "title" | "commit_message" | "memory_extract" | "memory_dedup" | "memory_decay" | "memory_injection_check" | "memory_verify_support" | "memory_autosort" | "memory_topic" | "memory_consolidate" | "memory_decompose" | "memory_profile_synth" | "memory_graph" | "folder_context" | "selection_memory" | "summarize" | "style_analysis" | "classify_search" | "app_inspiration_generate" | "app_inspiration_trends" | "app_inspiration_remix" | "spotlight_image" | "slide_image" | "slide_video" | "slide_music" | "slide_sfx" | "media_video" | "media_audio" | "media_bg_removal" | "media_image_preset" | "tool_app_builder" | "tool_document_builder" | "tool_slides" | "tool_slides_image_gen" | "connector_guidance" | "cf_task" | "xbot_intent" | "xbot_council" | "scheduled_task";
+type TaskType = "title" | "commit_message" | "memory_extract" | "memory_import_extract" | "memory_dedup" | "memory_decay" | "memory_injection_check" | "memory_verify_support" | "memory_autosort" | "memory_topic" | "memory_consolidate" | "memory_decompose" | "memory_profile_synth" | "memory_graph" | "folder_context" | "selection_memory" | "summarize" | "style_analysis" | "classify_search" | "app_inspiration_generate" | "app_inspiration_trends" | "app_inspiration_remix" | "spotlight_image" | "slide_image" | "slide_video" | "slide_music" | "slide_sfx" | "media_video" | "media_audio" | "media_bg_removal" | "media_image_preset" | "tool_app_builder" | "tool_document_builder" | "tool_slides" | "tool_slides_image_gen" | "connector_guidance" | "cf_task" | "xbot_intent" | "xbot_council" | "scheduled_task";
 
 /**
  * Shared portal LLM call for the memory pipeline.
@@ -4863,6 +4863,26 @@ type PortalLlmFailureReason =
  | "null-completion"
 /** `totalTimeoutMs` was spent before the next attempt could run. */
  | "time-budget-exhausted";
+/**
+ * One attempt of a {@link callPortalJsonCompletion} call, reported to
+ * `onAttempt` as it settles — success included. The wire-level diagnostic the
+ * give-up hook cannot provide: a call that succeeds on its third try returns a
+ * value and fires no `onFailure`, yet it cost three completions and its first
+ * two answers were unusable. Extraction quality regressions have hidden in
+ * exactly that shape (a prompt change that made the model answer `NONE` first,
+ * JSON on retry), so the extraction eval gates on the first-attempt clean rate.
+ *
+ * @public
+ */
+interface PortalLlmAttempt {
+    /** 1-based attempt index. */
+    attempt: number;
+    /** Whether this attempt produced a parseable JSON value. */
+    ok: boolean;
+    /** Classification when `ok` is false. */
+    reason?: PortalLlmFailureReason;
+    httpStatus?: number;
+}
 /**
  * A give-up report: the classified {@link PortalLlmFailureReason} plus the
  * little context worth carrying into telemetry. Both extra fields are bounded
@@ -5210,7 +5230,7 @@ interface MemoryVaultSearchOptions {
      * When true, switches to the async pipeline (rankFusedVaultMemoriesAsync).
      */
     rerank?: boolean;
-    /** Number of CE rerank candidates. Default 30. */
+    /** Number of CE rerank candidates. Default 5 ({@link DEFAULT_RERANK_TOP_N}). */
     rerankTopN?: number;
     /** Multiplicative cross-encoder blend weight. Default 0.1. Only used when `rerank` is true. */
     ceWeight?: number;
@@ -5565,7 +5585,8 @@ interface RecallOptions {
      * swallowed (diagnostics must never break retrieval). Off unless provided.
      */
     onDiagnostics?: (diagnostics: RecallDiagnostics) => void;
-    /** Number of candidates fed to the cross-encoder rerank stage. Default: 30. */
+    /** Number of candidates fed to the cross-encoder rerank stage. Default: 5;
+     *  was 30 until 2026-08-13 — see anuma-ai/sdk#845. */
     rerankTopN?: number;
     /** Multiplicative cross-encoder blend weight. Default: 0.1. */
     ceWeight?: number;
@@ -6133,6 +6154,12 @@ interface ExtractFactsOptions extends PortalLlmAuth {
      * (i.e. redaction silently eating facts) is alarmable.
      */
     onCandidatesDropped?: () => void;
+    /**
+     * Per-attempt wire diagnostic, forwarded to {@link callPortalJsonCompletion}.
+     * The extraction eval gates on it (first-attempt clean rate); production
+     * callers may leave it unset.
+     */
+    onAttempt?: (attempt: PortalLlmAttempt) => void;
 }
 /**
  * Stage 1 — call the LLM to extract candidate facts from the recent

package/dist/server/index.d.ts

@@ -4766,7 +4766,7 @@ interface EntityOperationsContext {
  *
  * @public
  */
-type TaskType = "title" | "commit_message" | "memory_extract" | "memory_dedup" | "memory_decay" | "memory_injection_check" | "memory_verify_support" | "memory_autosort" | "memory_topic" | "memory_consolidate" | "memory_decompose" | "memory_profile_synth" | "memory_graph" | "folder_context" | "selection_memory" | "summarize" | "style_analysis" | "classify_search" | "app_inspiration_generate" | "app_inspiration_trends" | "app_inspiration_remix" | "spotlight_image" | "slide_image" | "slide_video" | "slide_music" | "slide_sfx" | "media_video" | "media_audio" | "media_bg_removal" | "media_image_preset" | "tool_app_builder" | "tool_document_builder" | "tool_slides" | "tool_slides_image_gen" | "connector_guidance" | "cf_task" | "xbot_intent" | "xbot_council" | "scheduled_task";
+type TaskType = "title" | "commit_message" | "memory_extract" | "memory_import_extract" | "memory_dedup" | "memory_decay" | "memory_injection_check" | "memory_verify_support" | "memory_autosort" | "memory_topic" | "memory_consolidate" | "memory_decompose" | "memory_profile_synth" | "memory_graph" | "folder_context" | "selection_memory" | "summarize" | "style_analysis" | "classify_search" | "app_inspiration_generate" | "app_inspiration_trends" | "app_inspiration_remix" | "spotlight_image" | "slide_image" | "slide_video" | "slide_music" | "slide_sfx" | "media_video" | "media_audio" | "media_bg_removal" | "media_image_preset" | "tool_app_builder" | "tool_document_builder" | "tool_slides" | "tool_slides_image_gen" | "connector_guidance" | "cf_task" | "xbot_intent" | "xbot_council" | "scheduled_task";
 
 /**
  * Shared portal LLM call for the memory pipeline.
@@ -4863,6 +4863,26 @@ type PortalLlmFailureReason =
  | "null-completion"
 /** `totalTimeoutMs` was spent before the next attempt could run. */
  | "time-budget-exhausted";
+/**
+ * One attempt of a {@link callPortalJsonCompletion} call, reported to
+ * `onAttempt` as it settles — success included. The wire-level diagnostic the
+ * give-up hook cannot provide: a call that succeeds on its third try returns a
+ * value and fires no `onFailure`, yet it cost three completions and its first
+ * two answers were unusable. Extraction quality regressions have hidden in
+ * exactly that shape (a prompt change that made the model answer `NONE` first,
+ * JSON on retry), so the extraction eval gates on the first-attempt clean rate.
+ *
+ * @public
+ */
+interface PortalLlmAttempt {
+    /** 1-based attempt index. */
+    attempt: number;
+    /** Whether this attempt produced a parseable JSON value. */
+    ok: boolean;
+    /** Classification when `ok` is false. */
+    reason?: PortalLlmFailureReason;
+    httpStatus?: number;
+}
 /**
  * A give-up report: the classified {@link PortalLlmFailureReason} plus the
  * little context worth carrying into telemetry. Both extra fields are bounded
@@ -5210,7 +5230,7 @@ interface MemoryVaultSearchOptions {
      * When true, switches to the async pipeline (rankFusedVaultMemoriesAsync).
      */
     rerank?: boolean;
-    /** Number of CE rerank candidates. Default 30. */
+    /** Number of CE rerank candidates. Default 5 ({@link DEFAULT_RERANK_TOP_N}). */
     rerankTopN?: number;
     /** Multiplicative cross-encoder blend weight. Default 0.1. Only used when `rerank` is true. */
     ceWeight?: number;
@@ -5565,7 +5585,8 @@ interface RecallOptions {
      * swallowed (diagnostics must never break retrieval). Off unless provided.
      */
     onDiagnostics?: (diagnostics: RecallDiagnostics) => void;
-    /** Number of candidates fed to the cross-encoder rerank stage. Default: 30. */
+    /** Number of candidates fed to the cross-encoder rerank stage. Default: 5;
+     *  was 30 until 2026-08-13 — see anuma-ai/sdk#845. */
     rerankTopN?: number;
     /** Multiplicative cross-encoder blend weight. Default: 0.1. */
     ceWeight?: number;
@@ -6133,6 +6154,12 @@ interface ExtractFactsOptions extends PortalLlmAuth {
      * (i.e. redaction silently eating facts) is alarmable.
      */
     onCandidatesDropped?: () => void;
+    /**
+     * Per-attempt wire diagnostic, forwarded to {@link callPortalJsonCompletion}.
+     * The extraction eval gates on it (first-attempt clean rate); production
+     * callers may leave it unset.
+     */
+    onAttempt?: (attempt: PortalLlmAttempt) => void;
 }
 /**
  * Stage 1 — call the LLM to extract candidate facts from the recent