@anuma/sdk 3.11.1 → 3.11.3

package/dist/expo/index.d.mts

@@ -5232,7 +5232,20 @@ type RecallDegradation =
 "rerank-unavailable"
 /** `budget: 'high'` requested but no `decomposeOptions`, so query
  *  decomposition was skipped and the budget downgraded to mid. */
- | "decompose-unavailable";
+ | "decompose-unavailable"
+/** There was no usable cosine lane, so results were ranked on BM25 (lexical)
+ *  alone: either the query embedding failed (or came back empty), or no
+ *  candidate had a vector to score against it because the row (re)embed failed.
+ *  Also set when the chunk lane — cosine-only, with no lexical equivalent — was
+ *  skipped for the same reason. The embeddings provider is a single upstream
+ *  with no fallback, so this is a whole-provider outage rather than a per-item
+ *  miss — expect a quality drop, not a total loss.
+ *
+ *  NOT set for a partial failure that leaves cosine running (a sub-query embed
+ *  failing back to the single-query path, or a row batch failing while other
+ *  rows keep usable vectors) — those are logged, not reported, so this stays a
+ *  reliable outage signal rather than a general embedding-error counter. */
+ | "embeddings-unavailable";
 /**
  * Per-call recall observability payload (see {@link RecallOptions.onDiagnostics}).
  * All timings are wall-clock milliseconds. Lane counts are post-dedupe,

package/dist/expo/index.d.ts

@@ -5232,7 +5232,20 @@ type RecallDegradation =
 "rerank-unavailable"
 /** `budget: 'high'` requested but no `decomposeOptions`, so query
  *  decomposition was skipped and the budget downgraded to mid. */
- | "decompose-unavailable";
+ | "decompose-unavailable"
+/** There was no usable cosine lane, so results were ranked on BM25 (lexical)
+ *  alone: either the query embedding failed (or came back empty), or no
+ *  candidate had a vector to score against it because the row (re)embed failed.
+ *  Also set when the chunk lane — cosine-only, with no lexical equivalent — was
+ *  skipped for the same reason. The embeddings provider is a single upstream
+ *  with no fallback, so this is a whole-provider outage rather than a per-item
+ *  miss — expect a quality drop, not a total loss.
+ *
+ *  NOT set for a partial failure that leaves cosine running (a sub-query embed
+ *  failing back to the single-query path, or a row batch failing while other
+ *  rows keep usable vectors) — those are logged, not reported, so this stays a
+ *  reliable outage signal rather than a general embedding-error counter. */
+ | "embeddings-unavailable";
 /**
  * Per-call recall observability payload (see {@link RecallOptions.onDiagnostics}).
  * All timings are wall-clock milliseconds. Lane counts are post-dedupe,

package/dist/expo/index.mjs

@@ -5,11 +5,14 @@ import {
   DEFAULT_MIN_CHUNK_SIZE,
   DEFAULT_MIN_CONTENT_LENGTH,
   DEFAULT_VAULT_CACHE_SIZE,
+  EMBEDDINGS_DEGRADED_EMPTY,
   ENTITY_FANOUT,
   ENTITY_KINDS,
   LRUCache,
   MAX_HOPS,
   NODE_BUDGET,
+  RECALL_MAX_LIMIT,
+  RECALL_TOOL_NAME,
   RerankerUnavailableError,
   TOPICS_EXTRACTION_VERSION,
   TREND_RECENT_WINDOW_DAYS,
@@ -111,7 +114,7 @@ import {
   upsertMessageOp,
   useEncryption,
   vaultMemoryToStored
-} from "./chunk-7CTI3XR6.mjs";
+} from "./chunk-7AW2VD5W.mjs";
 import {
   PII_PATTERNS,
   PiiRedactor,
@@ -1069,6 +1072,9 @@ var CompletionsStrategy = class {
           }
         }
       }
+      if (choice.finish_reason) {
+        accumulator.finishReason = choice.finish_reason;
+      }
       if (choice.finish_reason === "tool_calls" || choice.finish_reason === "stop") {
         if (accumulator.toolCalls.size > 0) {
           accumulator.content = mergeXaiInlineParameterTags(
@@ -1151,7 +1157,9 @@ var CompletionsStrategy = class {
             content: finalContent,
             ...toolCalls && { tool_calls: toolCalls }
           },
-          finish_reason: toolCalls ? "tool_calls" : "stop"
+          // Prefer what the provider actually said; fall back to the derived
+          // value only when the stream carried no finish_reason at all.
+          finish_reason: accumulator.finishReason ?? (toolCalls ? "tool_calls" : "stop")
         }
       ],
       ...usage && { usage },
@@ -1822,7 +1830,7 @@ var ResponsesStrategy = class {
       }
       return result;
     }
-    if (typedChunk.type === "response.completed") {
+    if (typedChunk.type === "response.completed" || typedChunk.type === "response.incomplete") {
       if (typedChunk.response?.usage) {
         const u = typedChunk.response.usage;
         const promptTokens = u.input_tokens ?? u.prompt_tokens ?? 0;
@@ -1838,6 +1846,11 @@ var ResponsesStrategy = class {
           ...creditsExhausted !== void 0 && { credits_exhausted: creditsExhausted }
         };
       }
+      const resp = typedChunk.response;
+      const looksIncomplete = typedChunk.type === "response.incomplete" || resp?.status === "incomplete";
+      if (looksIncomplete && resp?.incomplete_details?.reason === "max_output_tokens") {
+        accumulator.finishReason = "length";
+      }
       if (typedChunk.response?.tools_checksum && !accumulator.toolsChecksum) {
         accumulator.toolsChecksum = typedChunk.response.tools_checksum;
       }
@@ -3374,7 +3387,17 @@ Executing tool: ${toolInfo}
       flushDeAnon();
       currentAccumulator.content += tip;
     }
+    const truncatedToNothing = (acc) => acc.finishReason === "length" && acc.toolCalls.size === 0 && acc.content.trim() === "";
+    const TRUNCATION_ERROR = "Model response was truncated at the output-token limit before it produced any usable content or tool call. Retry with a higher `maxOutputTokens`, or prompt for fewer/smaller tool calls per turn.";
     if (toolIteration > 0) {
+      if (truncatedToNothing(currentAccumulator)) {
+        await fireRunError({ error: TRUNCATION_ERROR, stage: "model" });
+        return {
+          data: buildResponseFinal(currentAccumulator),
+          error: TRUNCATION_ERROR,
+          toolsChecksum: currentAccumulator.toolsChecksum
+        };
+      }
       const finalResponse = buildResponseFinal(currentAccumulator);
       if (onFinish) onFinish(finalResponse);
       await fireRunEnd({
@@ -3388,6 +3411,14 @@ Executing tool: ${toolInfo}
         autoExecutedToolResults: accumulatedToolResults
       };
     }
+    if (truncatedToNothing(accumulator)) {
+      await fireRunError({ error: TRUNCATION_ERROR, stage: "model" });
+      return {
+        data: response,
+        error: TRUNCATION_ERROR,
+        toolsChecksum: accumulator.toolsChecksum
+      };
+    }
     if (onFinish) onFinish(response);
     await fireRunEnd({
       finalContent: finalContentText(accumulator),
@@ -6396,6 +6427,11 @@ async function retain(content, ctx, options = {}) {
         ...options.folderId !== void 0 && { folderId: options.folderId }
       }
     );
+    if (prepared.embeddingFailure || prepared.embeddingsUnavailable) {
+      throw new Error(
+        "retain: embeddings unavailable \u2014 refusing to auto-merge against an inert cosine lane, which would create a duplicate instead of merging into the existing memory. Retry when embeddings recover, or pass enableAutoMerge: false to force a create."
+      );
+    }
     if (options.consolidateOptions) {
       const outcome = await tryConsolidate(trimmed, ctx, options, prepared);
       if (outcome) {
@@ -7628,10 +7664,6 @@ function rankProfileCandidates(inputs, now = Date.now(), options = {}) {
   return ranked;
 }
 
-// src/lib/memory/recallConstants.ts
-var RECALL_TOOL_NAME = "recall_memory";
-var RECALL_MAX_LIMIT = 50;
-
 // src/lib/memory/recallTool.ts
 var DEFAULT_LIMIT = 8;
 var DEFAULT_BUDGET = "low";
@@ -7782,7 +7814,11 @@ function createRecallTool(ctx, toolOptions, callbacks) {
       conversationMemories += effectiveLimit;
       let surfaced = 0;
       try {
+        let recallDegraded = [];
         const recallOpts = {
+          onDiagnostics: (d) => {
+            recallDegraded = d.degraded;
+          },
           types: defaultTypes,
           limit: effectiveLimit,
           budget: defaultBudget,
@@ -7814,6 +7850,9 @@ function createRecallTool(ctx, toolOptions, callbacks) {
           }
         }
         surfaced = result.memories.length;
+        if (result.memories.length === 0 && recallDegraded.includes("embeddings-unavailable")) {
+          return EMBEDDINGS_DEGRADED_EMPTY;
+        }
         const formatted = formatRecallResult(result.memories);
         return truncatedByBudget && surfaced > 0 ? `${formatted}
 

package/dist/expo/{recall-IBGUFY4P.mjs → recall-JWQ4IJX6.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-7CTI3XR6.mjs";
+} from "./chunk-7AW2VD5W.mjs";
 import "./chunk-ILAWDQDE.mjs";
 import "./chunk-GYKJOWEO.mjs";
 export {