@anuma/sdk 7.3.2 → 7.3.3

package/dist/expo/{chunk-AF37N5Z6.mjs → chunk-XW3LYSDA.mjs}

@@ -709,6 +709,13 @@ var FOREIGN_ENDPOINT_SUFFIX = {
   chat: "/responses",
   responses: "/chat/completions"
 };
+function endpointForTransport(endpoint, transport) {
+  const { path, rest } = splitEndpoint(endpoint);
+  const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
+  if (!path.endsWith(foreign)) return endpoint;
+  const own = FOREIGN_ENDPOINT_SUFFIX[transport === "chat" ? "responses" : "chat"];
+  return `${path.slice(0, -foreign.length)}${own}${rest}`;
+}
 function assertTransportMatchesEndpoint(transport, endpoint) {
   const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
   if (!splitEndpoint(endpoint).path.endsWith(foreign)) return;
@@ -766,10 +773,7 @@ async function callPortalJsonCompletion(req) {
       throw new Error(overrideValidation.message);
     }
     endpoint = overrideValidation.endpoint;
-    const { path, rest } = splitEndpoint(endpoint);
-    if (autoUpgraded && path.endsWith("/chat/completions")) {
-      endpoint = `${path.slice(0, -"/chat/completions".length)}/responses${rest}`;
-    }
+    if (autoUpgraded) endpoint = endpointForTransport(endpoint, transport);
     assertTransportMatchesEndpoint(transport, endpoint);
   }
   const maxAttempts = Math.max(1, req.maxAttempts ?? 3);
@@ -1074,6 +1078,7 @@ export {
   withInternalFlowMarker,
   supportsResponseFormat,
   requiresResponsesTransport,
+  endpointForTransport,
   parseRetryAfterMs,
   resolvePortalAuthHeaders,
   callPortalJsonCompletion,

package/dist/expo/{consolidate-TOITBD6W.mjs → consolidate-4CQJYKRK.mjs}

@@ -1,12 +1,14 @@
 import {
   callPortalJsonCompletion,
+  endpointForTransport,
   notifyConsolidationFallback,
+  requiresResponsesTransport,
   resolvePiiRedactor
-} from "./chunk-AF37N5Z6.mjs";
+} from "./chunk-XW3LYSDA.mjs";
 import "./chunk-GYKJOWEO.mjs";
 
 // src/lib/memory/consolidate.ts
-var DEFAULT_CONSOLIDATION_MODEL = "inclusionai/ling-2.6-flash";
+var DEFAULT_CONSOLIDATION_MODEL = "openai/gpt-6-luna";
 var DEFAULT_CONSOLIDATE_ATTEMPTS = 3;
 var DEFAULT_CONSOLIDATE_TOTAL_TIMEOUT_MS = 2e4;
 var SYSTEM_PROMPT = `You consolidate a new memory against existing memories from the same user.
@@ -46,7 +48,11 @@ OUTPUT \u2014 strict JSON, no prose:
 For "create": content is the new memory verbatim (or a slight refinement); omit targetId/targetIds.
 For "update": content is the merged richest-version, \u226480 words; targetId is the single memory to update.
 For "supersede": content is the NEW fact only (\u226480 words); "targetIds" lists EVERY stale memory being retired (all candidates describing the same now-changed attribute). You MUST also name the subjects: "newSubject" for the new memory, and one entry in "targetSubjects" for EACH memory you are retiring (use "targetSubject" as well when you retire a single memory). List a subject ONLY for the memories you are actually retiring, not for every candidate you were shown. Write the plain subject, not a sentence: "the user", "the user's sister", "the user's manager", "Biscuit the dog". Rule 1a means a supersede is only valid when every one of those subjects is the SAME as "newSubject"; name them and check them against each other before you commit to the action. If ANY of them differs, the answer is "create" \u2014 do not retire the ones that match and keep the rest, because a group that mixes subjects was grouped wrongly in the first place. A subjectless fact ("Lives in Denver", "Works at Riverbend") is about "the user" \u2014 the extractor omits the subject when it is the user, so treat the absence of a subject as the user, never as unknown.
-For "noop": no content (existing memory is already correct); targetId is that memory.`;
+For "noop": no content (existing memory is already correct); targetId is that memory.
+
+ADDING TO AN INTRODUCED ENTITY (check this before choosing "create"): when an existing memory only introduces or names someone or something in the user's life ("User has a cat named Mochi", "User's sister is called Ana") and the new memory gives an attribute of that SAME entity that the existing memory does not state yet (its age, breed, job, home city), choose "update": targetId is that memory and content folds both into one sentence ("User has a cat named Mochi, who is four years old."). The "dog's name vs dog's age" example above means two memories that each already state a different attribute stay separate; a memory that only introduces the entity absorbs the next attribute about it. If the existing memory already states a value and the new memory changes it, that is "supersede" (rule 4), not this.
+
+CARRY FORWARD UNCHANGED DETAILS (every "update" and "supersede"): a memory can hold several details at once, such as one folded from an introduction and an attribute ("User's sister Ana lives in Paris"). content replaces ONLY the value that changed and keeps every other detail of the memory it replaces, the subject's relation to the user and name above all: "Ana lives in London" against "User's sister Ana lives in Paris" is "supersede" with content "User's sister Ana lives in London", never just "Ana lives in London". Name the subject the same way on both sides: here "newSubject" and "targetSubjects" are both "the user's sister Ana".`;
 async function consolidateMemory(newContent, candidates, options) {
   const fallback = {
     action: "create",
@@ -69,14 +75,23 @@ async function consolidateMemory(newContent, candidates, options) {
 
 Existing memories (top ${candidates.length} by cosine):
 ${candidateText}`;
+  const model = options.model ?? DEFAULT_CONSOLIDATION_MODEL;
+  const viaResponses = requiresResponsesTransport(model);
+  const endpointOverride = options.endpointOverride === void 0 ? void 0 : endpointForTransport(options.endpointOverride, viaResponses ? "responses" : "chat");
   let parsed;
+  let failure;
   try {
     parsed = await callPortalJsonCompletion({
+      onFailure: (f) => {
+        failure = f;
+      },
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      ...endpointOverride !== void 0 && { endpointOverride },
       taskType: "memory_consolidate",
-      model: options.model ?? DEFAULT_CONSOLIDATION_MODEL,
+      model,
+      ...viaResponses ? { transport: "responses", reasoning: { effort: "low" } } : { transport: "chat" },
       systemPrompt: SYSTEM_PROMPT,
       userMessage,
       tag: "memory/consolidate",
@@ -88,7 +103,7 @@ ${candidateText}`;
   } catch (err) {
     return degrade("llm_error", fallback, options, err);
   }
-  if (parsed === null) return degrade("llm_error", fallback, options);
+  if (parsed === null) return degrade(fallbackReasonFor(failure), fallback, options, failure);
   const validIds = new Set(candidates.map((c) => c.id));
   const result = validate(parsed, trimmed, validIds);
   if (!result) return degrade("invalid_response", fallback, options);
@@ -101,6 +116,11 @@ ${candidateText}`;
   }
   return reportRefusal(result, options);
 }
+function fallbackReasonFor(failure) {
+  if (failure?.reason === "time-budget-exhausted") return "llm_timeout";
+  if (failure?.reason === "content-flagged") return "content_flagged";
+  return "llm_error";
+}
 function reportRefusal(result, options) {
   if (result.fallbackReason) {
     notifyConsolidationFallback(result.fallbackReason, options.onFallback);

package/dist/expo/index.cjs

@@ -807,6 +807,13 @@ function splitEndpoint(endpoint) {
     rest: cut === -1 ? "" : endpoint.slice(cut)
   };
 }
+function endpointForTransport(endpoint, transport) {
+  const { path, rest } = splitEndpoint(endpoint);
+  const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
+  if (!path.endsWith(foreign)) return endpoint;
+  const own = FOREIGN_ENDPOINT_SUFFIX[transport === "chat" ? "responses" : "chat"];
+  return `${path.slice(0, -foreign.length)}${own}${rest}`;
+}
 function assertTransportMatchesEndpoint(transport, endpoint) {
   const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
   if (!splitEndpoint(endpoint).path.endsWith(foreign)) return;
@@ -863,10 +870,7 @@ async function callPortalJsonCompletion(req) {
       throw new Error(overrideValidation.message);
     }
     endpoint = overrideValidation.endpoint;
-    const { path, rest } = splitEndpoint(endpoint);
-    if (autoUpgraded && path.endsWith("/chat/completions")) {
-      endpoint = `${path.slice(0, -"/chat/completions".length)}/responses${rest}`;
-    }
+    if (autoUpgraded) endpoint = endpointForTransport(endpoint, transport);
     assertTransportMatchesEndpoint(transport, endpoint);
   }
   const maxAttempts = Math.max(1, req.maxAttempts ?? 3);
@@ -1212,14 +1216,23 @@ async function consolidateMemory(newContent, candidates, options) {
 
 Existing memories (top ${candidates.length} by cosine):
 ${candidateText}`;
+  const model = options.model ?? DEFAULT_CONSOLIDATION_MODEL;
+  const viaResponses = requiresResponsesTransport(model);
+  const endpointOverride = options.endpointOverride === void 0 ? void 0 : endpointForTransport(options.endpointOverride, viaResponses ? "responses" : "chat");
   let parsed;
+  let failure;
   try {
     parsed = await callPortalJsonCompletion({
+      onFailure: (f) => {
+        failure = f;
+      },
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      ...endpointOverride !== void 0 && { endpointOverride },
       taskType: "memory_consolidate",
-      model: options.model ?? DEFAULT_CONSOLIDATION_MODEL,
+      model,
+      ...viaResponses ? { transport: "responses", reasoning: { effort: "low" } } : { transport: "chat" },
       systemPrompt: SYSTEM_PROMPT3,
       userMessage,
       tag: "memory/consolidate",
@@ -1231,7 +1244,7 @@ ${candidateText}`;
   } catch (err) {
     return degrade("llm_error", fallback, options, err);
   }
-  if (parsed === null) return degrade("llm_error", fallback, options);
+  if (parsed === null) return degrade(fallbackReasonFor(failure), fallback, options, failure);
   const validIds = new Set(candidates.map((c) => c.id));
   const result = validate2(parsed, trimmed, validIds);
   if (!result) return degrade("invalid_response", fallback, options);
@@ -1244,6 +1257,11 @@ ${candidateText}`;
   }
   return reportRefusal(result, options);
 }
+function fallbackReasonFor(failure) {
+  if (failure?.reason === "time-budget-exhausted") return "llm_timeout";
+  if (failure?.reason === "content-flagged") return "content_flagged";
+  return "llm_error";
+}
 function reportRefusal(result, options) {
   if (result.fallbackReason) {
     notifyConsolidationFallback(result.fallbackReason, options.onFallback);
@@ -1317,7 +1335,7 @@ var init_consolidate = __esm({
     init_redactor();
     init_consolidationFallback();
     init_portalLlm();
-    DEFAULT_CONSOLIDATION_MODEL = "inclusionai/ling-2.6-flash";
+    DEFAULT_CONSOLIDATION_MODEL = "openai/gpt-6-luna";
     DEFAULT_CONSOLIDATE_ATTEMPTS = 3;
     DEFAULT_CONSOLIDATE_TOTAL_TIMEOUT_MS = 2e4;
     SYSTEM_PROMPT3 = `You consolidate a new memory against existing memories from the same user.
@@ -1357,7 +1375,11 @@ OUTPUT \u2014 strict JSON, no prose:
 For "create": content is the new memory verbatim (or a slight refinement); omit targetId/targetIds.
 For "update": content is the merged richest-version, \u226480 words; targetId is the single memory to update.
 For "supersede": content is the NEW fact only (\u226480 words); "targetIds" lists EVERY stale memory being retired (all candidates describing the same now-changed attribute). You MUST also name the subjects: "newSubject" for the new memory, and one entry in "targetSubjects" for EACH memory you are retiring (use "targetSubject" as well when you retire a single memory). List a subject ONLY for the memories you are actually retiring, not for every candidate you were shown. Write the plain subject, not a sentence: "the user", "the user's sister", "the user's manager", "Biscuit the dog". Rule 1a means a supersede is only valid when every one of those subjects is the SAME as "newSubject"; name them and check them against each other before you commit to the action. If ANY of them differs, the answer is "create" \u2014 do not retire the ones that match and keep the rest, because a group that mixes subjects was grouped wrongly in the first place. A subjectless fact ("Lives in Denver", "Works at Riverbend") is about "the user" \u2014 the extractor omits the subject when it is the user, so treat the absence of a subject as the user, never as unknown.
-For "noop": no content (existing memory is already correct); targetId is that memory.`;
+For "noop": no content (existing memory is already correct); targetId is that memory.
+
+ADDING TO AN INTRODUCED ENTITY (check this before choosing "create"): when an existing memory only introduces or names someone or something in the user's life ("User has a cat named Mochi", "User's sister is called Ana") and the new memory gives an attribute of that SAME entity that the existing memory does not state yet (its age, breed, job, home city), choose "update": targetId is that memory and content folds both into one sentence ("User has a cat named Mochi, who is four years old."). The "dog's name vs dog's age" example above means two memories that each already state a different attribute stay separate; a memory that only introduces the entity absorbs the next attribute about it. If the existing memory already states a value and the new memory changes it, that is "supersede" (rule 4), not this.
+
+CARRY FORWARD UNCHANGED DETAILS (every "update" and "supersede"): a memory can hold several details at once, such as one folded from an introduction and an attribute ("User's sister Ana lives in Paris"). content replaces ONLY the value that changed and keeps every other detail of the memory it replaces, the subject's relation to the user and name above all: "Ana lives in London" against "User's sister Ana lives in Paris" is "supersede" with content "User's sister Ana lives in London", never just "Ana lives in London". Name the subject the same way on both sides: here "newSubject" and "targetSubjects" are both "the user's sister Ana".`;
     USER_SUBJECT_ALIASES = /* @__PURE__ */ new Set(["user", "themselves", "themself", "they", "me", "self"]);
   }
 });
@@ -18455,7 +18477,12 @@ function createAutoExtractor(options) {
       ...baseUrl !== void 0 && { baseUrl }
     };
   };
-  const consolidateOptions = options.consolidate ? { ...options.consolidate, ...subPassAuth(options.consolidate.baseUrl) } : void 0;
+  const consolidateEndpoint = options.consolidate?.endpointOverride ?? options.extract.endpointOverride;
+  const consolidateOptions = options.consolidate ? {
+    ...options.consolidate,
+    ...subPassAuth(options.consolidate.baseUrl),
+    ...consolidateEndpoint !== void 0 && { endpointOverride: consolidateEndpoint }
+  } : void 0;
   const injectionClassifierOptions = options.injectionClassifier ? { ...options.injectionClassifier, ...subPassAuth(options.injectionClassifier.baseUrl) } : void 0;
   if (!options.memoryStore && options.retainCtx.vaultCtx.entityCtx && !options.entityCtx) {
     getLogger().warn(

package/dist/expo/index.d.mts

@@ -4821,7 +4821,7 @@ type RetainSource = "manual" | "auto-extracted" | "capsule";
  *   rate of exactly zero is as likely to mean "the fields are being omitted" as
  *   "the mistake never happens". Check compliance before reading zero as health.
  */
-type ConsolidationFallbackReason = "llm_error" | "invalid_response" | "target_vanished"
+type ConsolidationFallbackReason = "llm_error" | "llm_timeout" | "content_flagged" | "invalid_response" | "target_vanished"
 /**
  * A `supersede` was downgraded to `create` because the model's own stated
  * subjects disagreed: the new fact is about someone other than the fact it
@@ -4860,6 +4860,7 @@ interface RetainOptions {
     consolidateOptions?: PortalLlmAuth & {
         baseUrl?: string;
         model?: string;
+        endpointOverride?: string;
         /**
          * Invoked when the consolidator degrades to its "create" fallback
          * instead of returning a real decision — `llm_error` for network /
@@ -8967,6 +8968,7 @@ interface CreateAutoExtractorOptions {
         baseUrl?: string;
         /** Override the consolidation model. Default: see `consolidate.ts`. */
         model?: string;
+        endpointOverride?: string;
         /**
          * Notified on each degraded create-fallback (LLM failure or
          * schema-violating response). See

package/dist/expo/index.d.ts

@@ -4821,7 +4821,7 @@ type RetainSource = "manual" | "auto-extracted" | "capsule";
  *   rate of exactly zero is as likely to mean "the fields are being omitted" as
  *   "the mistake never happens". Check compliance before reading zero as health.
  */
-type ConsolidationFallbackReason = "llm_error" | "invalid_response" | "target_vanished"
+type ConsolidationFallbackReason = "llm_error" | "llm_timeout" | "content_flagged" | "invalid_response" | "target_vanished"
 /**
  * A `supersede` was downgraded to `create` because the model's own stated
  * subjects disagreed: the new fact is about someone other than the fact it
@@ -4860,6 +4860,7 @@ interface RetainOptions {
     consolidateOptions?: PortalLlmAuth & {
         baseUrl?: string;
         model?: string;
+        endpointOverride?: string;
         /**
          * Invoked when the consolidator degrades to its "create" fallback
          * instead of returning a real decision — `llm_error` for network /
@@ -8967,6 +8968,7 @@ interface CreateAutoExtractorOptions {
         baseUrl?: string;
         /** Override the consolidation model. Default: see `consolidate.ts`. */
         model?: string;
+        endpointOverride?: string;
         /**
          * Notified on each degraded create-fallback (LLM failure or
          * schema-violating response). See

package/dist/expo/index.mjs

@@ -20,7 +20,7 @@ import {
   taskTypeHeader,
   validateEndpointOverride,
   withInternalFlowMarker
-} from "./chunk-AF37N5Z6.mjs";
+} from "./chunk-XW3LYSDA.mjs";
 import {
   BASE_URL,
   MCP_R2_DOMAIN,
@@ -16121,7 +16121,7 @@ async function tryConsolidate(trimmed, ctx, options, prepared) {
     content: m.content,
     similarity: m.similarity
   }));
-  const { consolidateMemory: doConsolidate } = await import("./consolidate-TOITBD6W.mjs");
+  const { consolidateMemory: doConsolidate } = await import("./consolidate-4CQJYKRK.mjs");
   const decision = await doConsolidate(trimmed, candidates, consolidateOptions);
   if (decision.action === "create") return decision.fallbackReason ? null : { create: true };
   if (decision.action === "supersede" && decision.content) {
@@ -16837,7 +16837,12 @@ function createAutoExtractor(options) {
       ...baseUrl !== void 0 && { baseUrl }
     };
   };
-  const consolidateOptions = options.consolidate ? { ...options.consolidate, ...subPassAuth(options.consolidate.baseUrl) } : void 0;
+  const consolidateEndpoint = options.consolidate?.endpointOverride ?? options.extract.endpointOverride;
+  const consolidateOptions = options.consolidate ? {
+    ...options.consolidate,
+    ...subPassAuth(options.consolidate.baseUrl),
+    ...consolidateEndpoint !== void 0 && { endpointOverride: consolidateEndpoint }
+  } : void 0;
   const injectionClassifierOptions = options.injectionClassifier ? { ...options.injectionClassifier, ...subPassAuth(options.injectionClassifier.baseUrl) } : void 0;
   if (!options.memoryStore && options.retainCtx.vaultCtx.entityCtx && !options.entityCtx) {
     getLogger().warn(

package/dist/react/{chunk-TGLAOFVA.mjs → chunk-7QX4J7KV.mjs}

@@ -709,6 +709,13 @@ var FOREIGN_ENDPOINT_SUFFIX = {
   chat: "/responses",
   responses: "/chat/completions"
 };
+function endpointForTransport(endpoint, transport) {
+  const { path, rest } = splitEndpoint(endpoint);
+  const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
+  if (!path.endsWith(foreign)) return endpoint;
+  const own = FOREIGN_ENDPOINT_SUFFIX[transport === "chat" ? "responses" : "chat"];
+  return `${path.slice(0, -foreign.length)}${own}${rest}`;
+}
 function assertTransportMatchesEndpoint(transport, endpoint) {
   const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
   if (!splitEndpoint(endpoint).path.endsWith(foreign)) return;
@@ -766,10 +773,7 @@ async function callPortalJsonCompletion(req) {
       throw new Error(overrideValidation.message);
     }
     endpoint = overrideValidation.endpoint;
-    const { path, rest } = splitEndpoint(endpoint);
-    if (autoUpgraded && path.endsWith("/chat/completions")) {
-      endpoint = `${path.slice(0, -"/chat/completions".length)}/responses${rest}`;
-    }
+    if (autoUpgraded) endpoint = endpointForTransport(endpoint, transport);
     assertTransportMatchesEndpoint(transport, endpoint);
   }
   const maxAttempts = Math.max(1, req.maxAttempts ?? 3);
@@ -1073,6 +1077,7 @@ export {
   withInternalFlowMarker,
   supportsResponseFormat,
   requiresResponsesTransport,
+  endpointForTransport,
   parseRetryAfterMs,
   resolvePortalAuthHeaders,
   callPortalJsonCompletion,

package/dist/react/{consolidate-ZEINV2LX.mjs → consolidate-WLDYJF6J.mjs}

@@ -1,12 +1,14 @@
 import {
   callPortalJsonCompletion,
+  endpointForTransport,
   notifyConsolidationFallback,
+  requiresResponsesTransport,
   resolvePiiRedactor
-} from "./chunk-TGLAOFVA.mjs";
+} from "./chunk-7QX4J7KV.mjs";
 import "./chunk-GYKJOWEO.mjs";
 
 // src/lib/memory/consolidate.ts
-var DEFAULT_CONSOLIDATION_MODEL = "inclusionai/ling-2.6-flash";
+var DEFAULT_CONSOLIDATION_MODEL = "openai/gpt-6-luna";
 var DEFAULT_CONSOLIDATE_ATTEMPTS = 3;
 var DEFAULT_CONSOLIDATE_TOTAL_TIMEOUT_MS = 2e4;
 var SYSTEM_PROMPT = `You consolidate a new memory against existing memories from the same user.
@@ -46,7 +48,11 @@ OUTPUT \u2014 strict JSON, no prose:
 For "create": content is the new memory verbatim (or a slight refinement); omit targetId/targetIds.
 For "update": content is the merged richest-version, \u226480 words; targetId is the single memory to update.
 For "supersede": content is the NEW fact only (\u226480 words); "targetIds" lists EVERY stale memory being retired (all candidates describing the same now-changed attribute). You MUST also name the subjects: "newSubject" for the new memory, and one entry in "targetSubjects" for EACH memory you are retiring (use "targetSubject" as well when you retire a single memory). List a subject ONLY for the memories you are actually retiring, not for every candidate you were shown. Write the plain subject, not a sentence: "the user", "the user's sister", "the user's manager", "Biscuit the dog". Rule 1a means a supersede is only valid when every one of those subjects is the SAME as "newSubject"; name them and check them against each other before you commit to the action. If ANY of them differs, the answer is "create" \u2014 do not retire the ones that match and keep the rest, because a group that mixes subjects was grouped wrongly in the first place. A subjectless fact ("Lives in Denver", "Works at Riverbend") is about "the user" \u2014 the extractor omits the subject when it is the user, so treat the absence of a subject as the user, never as unknown.
-For "noop": no content (existing memory is already correct); targetId is that memory.`;
+For "noop": no content (existing memory is already correct); targetId is that memory.
+
+ADDING TO AN INTRODUCED ENTITY (check this before choosing "create"): when an existing memory only introduces or names someone or something in the user's life ("User has a cat named Mochi", "User's sister is called Ana") and the new memory gives an attribute of that SAME entity that the existing memory does not state yet (its age, breed, job, home city), choose "update": targetId is that memory and content folds both into one sentence ("User has a cat named Mochi, who is four years old."). The "dog's name vs dog's age" example above means two memories that each already state a different attribute stay separate; a memory that only introduces the entity absorbs the next attribute about it. If the existing memory already states a value and the new memory changes it, that is "supersede" (rule 4), not this.
+
+CARRY FORWARD UNCHANGED DETAILS (every "update" and "supersede"): a memory can hold several details at once, such as one folded from an introduction and an attribute ("User's sister Ana lives in Paris"). content replaces ONLY the value that changed and keeps every other detail of the memory it replaces, the subject's relation to the user and name above all: "Ana lives in London" against "User's sister Ana lives in Paris" is "supersede" with content "User's sister Ana lives in London", never just "Ana lives in London". Name the subject the same way on both sides: here "newSubject" and "targetSubjects" are both "the user's sister Ana".`;
 async function consolidateMemory(newContent, candidates, options) {
   const fallback = {
     action: "create",
@@ -69,14 +75,23 @@ async function consolidateMemory(newContent, candidates, options) {
 
 Existing memories (top ${candidates.length} by cosine):
 ${candidateText}`;
+  const model = options.model ?? DEFAULT_CONSOLIDATION_MODEL;
+  const viaResponses = requiresResponsesTransport(model);
+  const endpointOverride = options.endpointOverride === void 0 ? void 0 : endpointForTransport(options.endpointOverride, viaResponses ? "responses" : "chat");
   let parsed;
+  let failure;
   try {
     parsed = await callPortalJsonCompletion({
+      onFailure: (f) => {
+        failure = f;
+      },
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      ...endpointOverride !== void 0 && { endpointOverride },
       taskType: "memory_consolidate",
-      model: options.model ?? DEFAULT_CONSOLIDATION_MODEL,
+      model,
+      ...viaResponses ? { transport: "responses", reasoning: { effort: "low" } } : { transport: "chat" },
       systemPrompt: SYSTEM_PROMPT,
       userMessage,
       tag: "memory/consolidate",
@@ -88,7 +103,7 @@ ${candidateText}`;
   } catch (err) {
     return degrade("llm_error", fallback, options, err);
   }
-  if (parsed === null) return degrade("llm_error", fallback, options);
+  if (parsed === null) return degrade(fallbackReasonFor(failure), fallback, options, failure);
   const validIds = new Set(candidates.map((c) => c.id));
   const result = validate(parsed, trimmed, validIds);
   if (!result) return degrade("invalid_response", fallback, options);
@@ -101,6 +116,11 @@ ${candidateText}`;
   }
   return reportRefusal(result, options);
 }
+function fallbackReasonFor(failure) {
+  if (failure?.reason === "time-budget-exhausted") return "llm_timeout";
+  if (failure?.reason === "content-flagged") return "content_flagged";
+  return "llm_error";
+}
 function reportRefusal(result, options) {
   if (result.fallbackReason) {
     notifyConsolidationFallback(result.fallbackReason, options.onFallback);

package/dist/react/index.cjs

@@ -807,6 +807,13 @@ function splitEndpoint(endpoint) {
     rest: cut === -1 ? "" : endpoint.slice(cut)
   };
 }
+function endpointForTransport(endpoint, transport) {
+  const { path, rest } = splitEndpoint(endpoint);
+  const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
+  if (!path.endsWith(foreign)) return endpoint;
+  const own = FOREIGN_ENDPOINT_SUFFIX[transport === "chat" ? "responses" : "chat"];
+  return `${path.slice(0, -foreign.length)}${own}${rest}`;
+}
 function assertTransportMatchesEndpoint(transport, endpoint) {
   const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
   if (!splitEndpoint(endpoint).path.endsWith(foreign)) return;
@@ -863,10 +870,7 @@ async function callPortalJsonCompletion(req) {
       throw new Error(overrideValidation.message);
     }
     endpoint = overrideValidation.endpoint;
-    const { path, rest } = splitEndpoint(endpoint);
-    if (autoUpgraded && path.endsWith("/chat/completions")) {
-      endpoint = `${path.slice(0, -"/chat/completions".length)}/responses${rest}`;
-    }
+    if (autoUpgraded) endpoint = endpointForTransport(endpoint, transport);
     assertTransportMatchesEndpoint(transport, endpoint);
   }
   const maxAttempts = Math.max(1, req.maxAttempts ?? 3);
@@ -1212,14 +1216,23 @@ async function consolidateMemory(newContent, candidates, options) {
 
 Existing memories (top ${candidates.length} by cosine):
 ${candidateText}`;
+  const model = options.model ?? DEFAULT_CONSOLIDATION_MODEL;
+  const viaResponses = requiresResponsesTransport(model);
+  const endpointOverride = options.endpointOverride === void 0 ? void 0 : endpointForTransport(options.endpointOverride, viaResponses ? "responses" : "chat");
   let parsed;
+  let failure;
   try {
     parsed = await callPortalJsonCompletion({
+      onFailure: (f) => {
+        failure = f;
+      },
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      ...endpointOverride !== void 0 && { endpointOverride },
       taskType: "memory_consolidate",
-      model: options.model ?? DEFAULT_CONSOLIDATION_MODEL,
+      model,
+      ...viaResponses ? { transport: "responses", reasoning: { effort: "low" } } : { transport: "chat" },
       systemPrompt: SYSTEM_PROMPT3,
       userMessage,
       tag: "memory/consolidate",
@@ -1231,7 +1244,7 @@ ${candidateText}`;
   } catch (err) {
     return degrade("llm_error", fallback, options, err);
   }
-  if (parsed === null) return degrade("llm_error", fallback, options);
+  if (parsed === null) return degrade(fallbackReasonFor(failure), fallback, options, failure);
   const validIds = new Set(candidates.map((c) => c.id));
   const result = validate2(parsed, trimmed, validIds);
   if (!result) return degrade("invalid_response", fallback, options);
@@ -1244,6 +1257,11 @@ ${candidateText}`;
   }
   return reportRefusal(result, options);
 }
+function fallbackReasonFor(failure) {
+  if (failure?.reason === "time-budget-exhausted") return "llm_timeout";
+  if (failure?.reason === "content-flagged") return "content_flagged";
+  return "llm_error";
+}
 function reportRefusal(result, options) {
   if (result.fallbackReason) {
     notifyConsolidationFallback(result.fallbackReason, options.onFallback);
@@ -1317,7 +1335,7 @@ var init_consolidate = __esm({
     init_redactor();
     init_consolidationFallback();
     init_portalLlm();
-    DEFAULT_CONSOLIDATION_MODEL = "inclusionai/ling-2.6-flash";
+    DEFAULT_CONSOLIDATION_MODEL = "openai/gpt-6-luna";
     DEFAULT_CONSOLIDATE_ATTEMPTS = 3;
     DEFAULT_CONSOLIDATE_TOTAL_TIMEOUT_MS = 2e4;
     SYSTEM_PROMPT3 = `You consolidate a new memory against existing memories from the same user.
@@ -1357,7 +1375,11 @@ OUTPUT \u2014 strict JSON, no prose:
 For "create": content is the new memory verbatim (or a slight refinement); omit targetId/targetIds.
 For "update": content is the merged richest-version, \u226480 words; targetId is the single memory to update.
 For "supersede": content is the NEW fact only (\u226480 words); "targetIds" lists EVERY stale memory being retired (all candidates describing the same now-changed attribute). You MUST also name the subjects: "newSubject" for the new memory, and one entry in "targetSubjects" for EACH memory you are retiring (use "targetSubject" as well when you retire a single memory). List a subject ONLY for the memories you are actually retiring, not for every candidate you were shown. Write the plain subject, not a sentence: "the user", "the user's sister", "the user's manager", "Biscuit the dog". Rule 1a means a supersede is only valid when every one of those subjects is the SAME as "newSubject"; name them and check them against each other before you commit to the action. If ANY of them differs, the answer is "create" \u2014 do not retire the ones that match and keep the rest, because a group that mixes subjects was grouped wrongly in the first place. A subjectless fact ("Lives in Denver", "Works at Riverbend") is about "the user" \u2014 the extractor omits the subject when it is the user, so treat the absence of a subject as the user, never as unknown.
-For "noop": no content (existing memory is already correct); targetId is that memory.`;
+For "noop": no content (existing memory is already correct); targetId is that memory.
+
+ADDING TO AN INTRODUCED ENTITY (check this before choosing "create"): when an existing memory only introduces or names someone or something in the user's life ("User has a cat named Mochi", "User's sister is called Ana") and the new memory gives an attribute of that SAME entity that the existing memory does not state yet (its age, breed, job, home city), choose "update": targetId is that memory and content folds both into one sentence ("User has a cat named Mochi, who is four years old."). The "dog's name vs dog's age" example above means two memories that each already state a different attribute stay separate; a memory that only introduces the entity absorbs the next attribute about it. If the existing memory already states a value and the new memory changes it, that is "supersede" (rule 4), not this.
+
+CARRY FORWARD UNCHANGED DETAILS (every "update" and "supersede"): a memory can hold several details at once, such as one folded from an introduction and an attribute ("User's sister Ana lives in Paris"). content replaces ONLY the value that changed and keeps every other detail of the memory it replaces, the subject's relation to the user and name above all: "Ana lives in London" against "User's sister Ana lives in Paris" is "supersede" with content "User's sister Ana lives in London", never just "Ana lives in London". Name the subject the same way on both sides: here "newSubject" and "targetSubjects" are both "the user's sister Ana".`;
     USER_SUBJECT_ALIASES = /* @__PURE__ */ new Set(["user", "themselves", "themself", "they", "me", "self"]);
   }
 });
@@ -20732,7 +20754,12 @@ function createAutoExtractor(options) {
       ...baseUrl !== void 0 && { baseUrl }
     };
   };
-  const consolidateOptions = options.consolidate ? { ...options.consolidate, ...subPassAuth(options.consolidate.baseUrl) } : void 0;
+  const consolidateEndpoint = options.consolidate?.endpointOverride ?? options.extract.endpointOverride;
+  const consolidateOptions = options.consolidate ? {
+    ...options.consolidate,
+    ...subPassAuth(options.consolidate.baseUrl),
+    ...consolidateEndpoint !== void 0 && { endpointOverride: consolidateEndpoint }
+  } : void 0;
   const injectionClassifierOptions = options.injectionClassifier ? { ...options.injectionClassifier, ...subPassAuth(options.injectionClassifier.baseUrl) } : void 0;
   if (!options.memoryStore && options.retainCtx.vaultCtx.entityCtx && !options.entityCtx) {
     getLogger().warn(

package/dist/react/index.d.mts

@@ -5337,7 +5337,7 @@ type RetainSource = "manual" | "auto-extracted" | "capsule";
  *   rate of exactly zero is as likely to mean "the fields are being omitted" as
  *   "the mistake never happens". Check compliance before reading zero as health.
  */
-type ConsolidationFallbackReason = "llm_error" | "invalid_response" | "target_vanished"
+type ConsolidationFallbackReason = "llm_error" | "llm_timeout" | "content_flagged" | "invalid_response" | "target_vanished"
 /**
  * A `supersede` was downgraded to `create` because the model's own stated
  * subjects disagreed: the new fact is about someone other than the fact it
@@ -5376,6 +5376,7 @@ interface RetainOptions {
     consolidateOptions?: PortalLlmAuth & {
         baseUrl?: string;
         model?: string;
+        endpointOverride?: string;
         /**
          * Invoked when the consolidator degrades to its "create" fallback
          * instead of returning a real decision — `llm_error` for network /
@@ -10700,6 +10701,7 @@ interface CreateAutoExtractorOptions {
         baseUrl?: string;
         /** Override the consolidation model. Default: see `consolidate.ts`. */
         model?: string;
+        endpointOverride?: string;
         /**
          * Notified on each degraded create-fallback (LLM failure or
          * schema-violating response). See

package/dist/react/index.d.ts

@@ -5337,7 +5337,7 @@ type RetainSource = "manual" | "auto-extracted" | "capsule";
  *   rate of exactly zero is as likely to mean "the fields are being omitted" as
  *   "the mistake never happens". Check compliance before reading zero as health.
  */
-type ConsolidationFallbackReason = "llm_error" | "invalid_response" | "target_vanished"
+type ConsolidationFallbackReason = "llm_error" | "llm_timeout" | "content_flagged" | "invalid_response" | "target_vanished"
 /**
  * A `supersede` was downgraded to `create` because the model's own stated
  * subjects disagreed: the new fact is about someone other than the fact it
@@ -5376,6 +5376,7 @@ interface RetainOptions {
     consolidateOptions?: PortalLlmAuth & {
         baseUrl?: string;
         model?: string;
+        endpointOverride?: string;
         /**
          * Invoked when the consolidator degrades to its "create" fallback
          * instead of returning a real decision — `llm_error` for network /
@@ -10700,6 +10701,7 @@ interface CreateAutoExtractorOptions {
         baseUrl?: string;
         /** Override the consolidation model. Default: see `consolidate.ts`. */
         model?: string;
+        endpointOverride?: string;
         /**
          * Notified on each degraded create-fallback (LLM failure or
          * schema-violating response). See

package/dist/react/index.mjs

@@ -19,7 +19,7 @@ import {
   taskTypeHeader,
   validateEndpointOverride,
   withInternalFlowMarker
-} from "./chunk-TGLAOFVA.mjs";
+} from "./chunk-7QX4J7KV.mjs";
 import {
   BASE_URL,
   MCP_R2_DOMAIN,
@@ -18190,7 +18190,7 @@ async function tryConsolidate(trimmed, ctx, options, prepared) {
     content: m.content,
     similarity: m.similarity
   }));
-  const { consolidateMemory: doConsolidate } = await import("./consolidate-ZEINV2LX.mjs");
+  const { consolidateMemory: doConsolidate } = await import("./consolidate-WLDYJF6J.mjs");
   const decision = await doConsolidate(trimmed, candidates, consolidateOptions);
   if (decision.action === "create") return decision.fallbackReason ? null : { create: true };
   if (decision.action === "supersede" && decision.content) {
@@ -18906,7 +18906,12 @@ function createAutoExtractor(options) {
       ...baseUrl !== void 0 && { baseUrl }
     };
   };
-  const consolidateOptions = options.consolidate ? { ...options.consolidate, ...subPassAuth(options.consolidate.baseUrl) } : void 0;
+  const consolidateEndpoint = options.consolidate?.endpointOverride ?? options.extract.endpointOverride;
+  const consolidateOptions = options.consolidate ? {
+    ...options.consolidate,
+    ...subPassAuth(options.consolidate.baseUrl),
+    ...consolidateEndpoint !== void 0 && { endpointOverride: consolidateEndpoint }
+  } : void 0;
   const injectionClassifierOptions = options.injectionClassifier ? { ...options.injectionClassifier, ...subPassAuth(options.injectionClassifier.baseUrl) } : void 0;
   if (!options.memoryStore && options.retainCtx.vaultCtx.entityCtx && !options.entityCtx) {
     getLogger().warn(

package/dist/server/{chunk-WCG5BFI6.mjs → chunk-AU3Z7NBC.mjs}

@@ -696,6 +696,13 @@ var FOREIGN_ENDPOINT_SUFFIX = {
   chat: "/responses",
   responses: "/chat/completions"
 };
+function endpointForTransport(endpoint, transport) {
+  const { path, rest } = splitEndpoint(endpoint);
+  const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
+  if (!path.endsWith(foreign)) return endpoint;
+  const own = FOREIGN_ENDPOINT_SUFFIX[transport === "chat" ? "responses" : "chat"];
+  return `${path.slice(0, -foreign.length)}${own}${rest}`;
+}
 function assertTransportMatchesEndpoint(transport, endpoint) {
   const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
   if (!splitEndpoint(endpoint).path.endsWith(foreign)) return;
@@ -753,10 +760,7 @@ async function callPortalJsonCompletion(req) {
       throw new Error(overrideValidation.message);
     }
     endpoint = overrideValidation.endpoint;
-    const { path, rest } = splitEndpoint(endpoint);
-    if (autoUpgraded && path.endsWith("/chat/completions")) {
-      endpoint = `${path.slice(0, -"/chat/completions".length)}/responses${rest}`;
-    }
+    if (autoUpgraded) endpoint = endpointForTransport(endpoint, transport);
     assertTransportMatchesEndpoint(transport, endpoint);
   }
   const maxAttempts = Math.max(1, req.maxAttempts ?? 3);
@@ -1055,6 +1059,7 @@ export {
   withInternalFlowMarker,
   supportsResponseFormat,
   requiresResponsesTransport,
+  endpointForTransport,
   resolvePortalAuthHeaders,
   callPortalJsonCompletion,
   extractJsonCandidate,

package/dist/server/{consolidate-WY6BHODK.mjs → consolidate-B6OSCZU4.mjs}

@@ -1,12 +1,14 @@
 import {
   callPortalJsonCompletion,
+  endpointForTransport,
   notifyConsolidationFallback,
+  requiresResponsesTransport,
   resolvePiiRedactor
-} from "./chunk-WCG5BFI6.mjs";
+} from "./chunk-AU3Z7NBC.mjs";
 import "./chunk-3RPGXXPZ.mjs";
 
 // src/lib/memory/consolidate.ts
-var DEFAULT_CONSOLIDATION_MODEL = "inclusionai/ling-2.6-flash";
+var DEFAULT_CONSOLIDATION_MODEL = "openai/gpt-6-luna";
 var DEFAULT_CONSOLIDATE_ATTEMPTS = 3;
 var DEFAULT_CONSOLIDATE_TOTAL_TIMEOUT_MS = 2e4;
 var SYSTEM_PROMPT = `You consolidate a new memory against existing memories from the same user.
@@ -46,7 +48,11 @@ OUTPUT \u2014 strict JSON, no prose:
 For "create": content is the new memory verbatim (or a slight refinement); omit targetId/targetIds.
 For "update": content is the merged richest-version, \u226480 words; targetId is the single memory to update.
 For "supersede": content is the NEW fact only (\u226480 words); "targetIds" lists EVERY stale memory being retired (all candidates describing the same now-changed attribute). You MUST also name the subjects: "newSubject" for the new memory, and one entry in "targetSubjects" for EACH memory you are retiring (use "targetSubject" as well when you retire a single memory). List a subject ONLY for the memories you are actually retiring, not for every candidate you were shown. Write the plain subject, not a sentence: "the user", "the user's sister", "the user's manager", "Biscuit the dog". Rule 1a means a supersede is only valid when every one of those subjects is the SAME as "newSubject"; name them and check them against each other before you commit to the action. If ANY of them differs, the answer is "create" \u2014 do not retire the ones that match and keep the rest, because a group that mixes subjects was grouped wrongly in the first place. A subjectless fact ("Lives in Denver", "Works at Riverbend") is about "the user" \u2014 the extractor omits the subject when it is the user, so treat the absence of a subject as the user, never as unknown.
-For "noop": no content (existing memory is already correct); targetId is that memory.`;
+For "noop": no content (existing memory is already correct); targetId is that memory.
+
+ADDING TO AN INTRODUCED ENTITY (check this before choosing "create"): when an existing memory only introduces or names someone or something in the user's life ("User has a cat named Mochi", "User's sister is called Ana") and the new memory gives an attribute of that SAME entity that the existing memory does not state yet (its age, breed, job, home city), choose "update": targetId is that memory and content folds both into one sentence ("User has a cat named Mochi, who is four years old."). The "dog's name vs dog's age" example above means two memories that each already state a different attribute stay separate; a memory that only introduces the entity absorbs the next attribute about it. If the existing memory already states a value and the new memory changes it, that is "supersede" (rule 4), not this.
+
+CARRY FORWARD UNCHANGED DETAILS (every "update" and "supersede"): a memory can hold several details at once, such as one folded from an introduction and an attribute ("User's sister Ana lives in Paris"). content replaces ONLY the value that changed and keeps every other detail of the memory it replaces, the subject's relation to the user and name above all: "Ana lives in London" against "User's sister Ana lives in Paris" is "supersede" with content "User's sister Ana lives in London", never just "Ana lives in London". Name the subject the same way on both sides: here "newSubject" and "targetSubjects" are both "the user's sister Ana".`;
 async function consolidateMemory(newContent, candidates, options) {
   const fallback = {
     action: "create",
@@ -69,14 +75,23 @@ async function consolidateMemory(newContent, candidates, options) {
 
 Existing memories (top ${candidates.length} by cosine):
 ${candidateText}`;
+  const model = options.model ?? DEFAULT_CONSOLIDATION_MODEL;
+  const viaResponses = requiresResponsesTransport(model);
+  const endpointOverride = options.endpointOverride === void 0 ? void 0 : endpointForTransport(options.endpointOverride, viaResponses ? "responses" : "chat");
   let parsed;
+  let failure;
   try {
     parsed = await callPortalJsonCompletion({
+      onFailure: (f) => {
+        failure = f;
+      },
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      ...endpointOverride !== void 0 && { endpointOverride },
       taskType: "memory_consolidate",
-      model: options.model ?? DEFAULT_CONSOLIDATION_MODEL,
+      model,
+      ...viaResponses ? { transport: "responses", reasoning: { effort: "low" } } : { transport: "chat" },
       systemPrompt: SYSTEM_PROMPT,
       userMessage,
       tag: "memory/consolidate",
@@ -88,7 +103,7 @@ ${candidateText}`;
   } catch (err) {
     return degrade("llm_error", fallback, options, err);
   }
-  if (parsed === null) return degrade("llm_error", fallback, options);
+  if (parsed === null) return degrade(fallbackReasonFor(failure), fallback, options, failure);
   const validIds = new Set(candidates.map((c) => c.id));
   const result = validate(parsed, trimmed, validIds);
   if (!result) return degrade("invalid_response", fallback, options);
@@ -101,6 +116,11 @@ ${candidateText}`;
   }
   return reportRefusal(result, options);
 }
+function fallbackReasonFor(failure) {
+  if (failure?.reason === "time-budget-exhausted") return "llm_timeout";
+  if (failure?.reason === "content-flagged") return "content_flagged";
+  return "llm_error";
+}
 function reportRefusal(result, options) {
   if (result.fallbackReason) {
     notifyConsolidationFallback(result.fallbackReason, options.onFallback);

package/dist/server/index.cjs

@@ -15674,6 +15674,13 @@ function splitEndpoint(endpoint) {
     rest: cut === -1 ? "" : endpoint.slice(cut)
   };
 }
+function endpointForTransport(endpoint, transport) {
+  const { path, rest } = splitEndpoint(endpoint);
+  const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
+  if (!path.endsWith(foreign)) return endpoint;
+  const own = FOREIGN_ENDPOINT_SUFFIX[transport === "chat" ? "responses" : "chat"];
+  return `${path.slice(0, -foreign.length)}${own}${rest}`;
+}
 function assertTransportMatchesEndpoint(transport, endpoint) {
   const foreign = FOREIGN_ENDPOINT_SUFFIX[transport];
   if (!splitEndpoint(endpoint).path.endsWith(foreign)) return;
@@ -15730,10 +15737,7 @@ async function callPortalJsonCompletion(req) {
       throw new Error(overrideValidation.message);
     }
     endpoint = overrideValidation.endpoint;
-    const { path, rest } = splitEndpoint(endpoint);
-    if (autoUpgraded && path.endsWith("/chat/completions")) {
-      endpoint = `${path.slice(0, -"/chat/completions".length)}/responses${rest}`;
-    }
+    if (autoUpgraded) endpoint = endpointForTransport(endpoint, transport);
     assertTransportMatchesEndpoint(transport, endpoint);
   }
   const maxAttempts = Math.max(1, req.maxAttempts ?? 3);
@@ -16079,14 +16083,23 @@ async function consolidateMemory(newContent, candidates, options) {
 
 Existing memories (top ${candidates.length} by cosine):
 ${candidateText}`;
+  const model = options.model ?? DEFAULT_CONSOLIDATION_MODEL;
+  const viaResponses = requiresResponsesTransport(model);
+  const endpointOverride = options.endpointOverride === void 0 ? void 0 : endpointForTransport(options.endpointOverride, viaResponses ? "responses" : "chat");
   let parsed;
+  let failure;
   try {
     parsed = await callPortalJsonCompletion({
+      onFailure: (f) => {
+        failure = f;
+      },
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      ...endpointOverride !== void 0 && { endpointOverride },
       taskType: "memory_consolidate",
-      model: options.model ?? DEFAULT_CONSOLIDATION_MODEL,
+      model,
+      ...viaResponses ? { transport: "responses", reasoning: { effort: "low" } } : { transport: "chat" },
       systemPrompt: SYSTEM_PROMPT3,
       userMessage,
       tag: "memory/consolidate",
@@ -16098,7 +16111,7 @@ ${candidateText}`;
   } catch (err) {
     return degrade("llm_error", fallback, options, err);
   }
-  if (parsed === null) return degrade("llm_error", fallback, options);
+  if (parsed === null) return degrade(fallbackReasonFor(failure), fallback, options, failure);
   const validIds = new Set(candidates.map((c) => c.id));
   const result = validate2(parsed, trimmed, validIds);
   if (!result) return degrade("invalid_response", fallback, options);
@@ -16111,6 +16124,11 @@ ${candidateText}`;
   }
   return reportRefusal(result, options);
 }
+function fallbackReasonFor(failure) {
+  if (failure?.reason === "time-budget-exhausted") return "llm_timeout";
+  if (failure?.reason === "content-flagged") return "content_flagged";
+  return "llm_error";
+}
 function reportRefusal(result, options) {
   if (result.fallbackReason) {
     notifyConsolidationFallback(result.fallbackReason, options.onFallback);
@@ -16184,7 +16202,7 @@ var init_consolidate = __esm({
     init_redactor();
     init_consolidationFallback();
     init_portalLlm();
-    DEFAULT_CONSOLIDATION_MODEL = "inclusionai/ling-2.6-flash";
+    DEFAULT_CONSOLIDATION_MODEL = "openai/gpt-6-luna";
     DEFAULT_CONSOLIDATE_ATTEMPTS = 3;
     DEFAULT_CONSOLIDATE_TOTAL_TIMEOUT_MS = 2e4;
     SYSTEM_PROMPT3 = `You consolidate a new memory against existing memories from the same user.
@@ -16224,7 +16242,11 @@ OUTPUT \u2014 strict JSON, no prose:
 For "create": content is the new memory verbatim (or a slight refinement); omit targetId/targetIds.
 For "update": content is the merged richest-version, \u226480 words; targetId is the single memory to update.
 For "supersede": content is the NEW fact only (\u226480 words); "targetIds" lists EVERY stale memory being retired (all candidates describing the same now-changed attribute). You MUST also name the subjects: "newSubject" for the new memory, and one entry in "targetSubjects" for EACH memory you are retiring (use "targetSubject" as well when you retire a single memory). List a subject ONLY for the memories you are actually retiring, not for every candidate you were shown. Write the plain subject, not a sentence: "the user", "the user's sister", "the user's manager", "Biscuit the dog". Rule 1a means a supersede is only valid when every one of those subjects is the SAME as "newSubject"; name them and check them against each other before you commit to the action. If ANY of them differs, the answer is "create" \u2014 do not retire the ones that match and keep the rest, because a group that mixes subjects was grouped wrongly in the first place. A subjectless fact ("Lives in Denver", "Works at Riverbend") is about "the user" \u2014 the extractor omits the subject when it is the user, so treat the absence of a subject as the user, never as unknown.
-For "noop": no content (existing memory is already correct); targetId is that memory.`;
+For "noop": no content (existing memory is already correct); targetId is that memory.
+
+ADDING TO AN INTRODUCED ENTITY (check this before choosing "create"): when an existing memory only introduces or names someone or something in the user's life ("User has a cat named Mochi", "User's sister is called Ana") and the new memory gives an attribute of that SAME entity that the existing memory does not state yet (its age, breed, job, home city), choose "update": targetId is that memory and content folds both into one sentence ("User has a cat named Mochi, who is four years old."). The "dog's name vs dog's age" example above means two memories that each already state a different attribute stay separate; a memory that only introduces the entity absorbs the next attribute about it. If the existing memory already states a value and the new memory changes it, that is "supersede" (rule 4), not this.
+
+CARRY FORWARD UNCHANGED DETAILS (every "update" and "supersede"): a memory can hold several details at once, such as one folded from an introduction and an attribute ("User's sister Ana lives in Paris"). content replaces ONLY the value that changed and keeps every other detail of the memory it replaces, the subject's relation to the user and name above all: "Ana lives in London" against "User's sister Ana lives in Paris" is "supersede" with content "User's sister Ana lives in London", never just "Ana lives in London". Name the subject the same way on both sides: here "newSubject" and "targetSubjects" are both "the user's sister Ana".`;
     USER_SUBJECT_ALIASES = /* @__PURE__ */ new Set(["user", "themselves", "themself", "they", "me", "self"]);
   }
 });
@@ -33697,7 +33719,12 @@ function createAutoExtractor(options) {
       ...baseUrl !== void 0 && { baseUrl }
     };
   };
-  const consolidateOptions = options.consolidate ? { ...options.consolidate, ...subPassAuth(options.consolidate.baseUrl) } : void 0;
+  const consolidateEndpoint = options.consolidate?.endpointOverride ?? options.extract.endpointOverride;
+  const consolidateOptions = options.consolidate ? {
+    ...options.consolidate,
+    ...subPassAuth(options.consolidate.baseUrl),
+    ...consolidateEndpoint !== void 0 && { endpointOverride: consolidateEndpoint }
+  } : void 0;
   const injectionClassifierOptions = options.injectionClassifier ? { ...options.injectionClassifier, ...subPassAuth(options.injectionClassifier.baseUrl) } : void 0;
   if (!options.memoryStore && options.retainCtx.vaultCtx.entityCtx && !options.entityCtx) {
     getLogger().warn(

package/dist/server/index.d.mts

@@ -4503,7 +4503,7 @@ type RetainSource = "manual" | "auto-extracted" | "capsule";
  *   rate of exactly zero is as likely to mean "the fields are being omitted" as
  *   "the mistake never happens". Check compliance before reading zero as health.
  */
-type ConsolidationFallbackReason = "llm_error" | "invalid_response" | "target_vanished"
+type ConsolidationFallbackReason = "llm_error" | "llm_timeout" | "content_flagged" | "invalid_response" | "target_vanished"
 /**
  * A `supersede` was downgraded to `create` because the model's own stated
  * subjects disagreed: the new fact is about someone other than the fact it
@@ -4542,6 +4542,7 @@ interface RetainOptions {
     consolidateOptions?: PortalLlmAuth & {
         baseUrl?: string;
         model?: string;
+        endpointOverride?: string;
         /**
          * Invoked when the consolidator degrades to its "create" fallback
          * instead of returning a real decision — `llm_error` for network /
@@ -9174,6 +9175,7 @@ interface CreateAutoExtractorOptions {
         baseUrl?: string;
         /** Override the consolidation model. Default: see `consolidate.ts`. */
         model?: string;
+        endpointOverride?: string;
         /**
          * Notified on each degraded create-fallback (LLM failure or
          * schema-violating response). See

package/dist/server/index.d.ts

@@ -4503,7 +4503,7 @@ type RetainSource = "manual" | "auto-extracted" | "capsule";
  *   rate of exactly zero is as likely to mean "the fields are being omitted" as
  *   "the mistake never happens". Check compliance before reading zero as health.
  */
-type ConsolidationFallbackReason = "llm_error" | "invalid_response" | "target_vanished"
+type ConsolidationFallbackReason = "llm_error" | "llm_timeout" | "content_flagged" | "invalid_response" | "target_vanished"
 /**
  * A `supersede` was downgraded to `create` because the model's own stated
  * subjects disagreed: the new fact is about someone other than the fact it
@@ -4542,6 +4542,7 @@ interface RetainOptions {
     consolidateOptions?: PortalLlmAuth & {
         baseUrl?: string;
         model?: string;
+        endpointOverride?: string;
         /**
          * Invoked when the consolidator degrades to its "create" fallback
          * instead of returning a real decision — `llm_error` for network /
@@ -9174,6 +9175,7 @@ interface CreateAutoExtractorOptions {
         baseUrl?: string;
         /** Override the consolidation model. Default: see `consolidate.ts`. */
         model?: string;
+        endpointOverride?: string;
         /**
          * Notified on each degraded create-fallback (LLM failure or
          * schema-violating response). See

package/dist/server/index.mjs

@@ -13,7 +13,7 @@ import {
   taskTypeHeader,
   validateEndpointOverride,
   withInternalFlowMarker
-} from "./chunk-WCG5BFI6.mjs";
+} from "./chunk-AU3Z7NBC.mjs";
 import {
   BASE_URL,
   __commonJS,
@@ -31273,7 +31273,7 @@ async function tryConsolidate(trimmed, ctx, options, prepared) {
     content: m.content,
     similarity: m.similarity
   }));
-  const { consolidateMemory: doConsolidate } = await import("./consolidate-WY6BHODK.mjs");
+  const { consolidateMemory: doConsolidate } = await import("./consolidate-B6OSCZU4.mjs");
   const decision = await doConsolidate(trimmed, candidates, consolidateOptions);
   if (decision.action === "create") return decision.fallbackReason ? null : { create: true };
   if (decision.action === "supersede" && decision.content) {
@@ -31989,7 +31989,12 @@ function createAutoExtractor(options) {
       ...baseUrl !== void 0 && { baseUrl }
     };
   };
-  const consolidateOptions = options.consolidate ? { ...options.consolidate, ...subPassAuth(options.consolidate.baseUrl) } : void 0;
+  const consolidateEndpoint = options.consolidate?.endpointOverride ?? options.extract.endpointOverride;
+  const consolidateOptions = options.consolidate ? {
+    ...options.consolidate,
+    ...subPassAuth(options.consolidate.baseUrl),
+    ...consolidateEndpoint !== void 0 && { endpointOverride: consolidateEndpoint }
+  } : void 0;
   const injectionClassifierOptions = options.injectionClassifier ? { ...options.injectionClassifier, ...subPassAuth(options.injectionClassifier.baseUrl) } : void 0;
   if (!options.memoryStore && options.retainCtx.vaultCtx.entityCtx && !options.entityCtx) {
     getLogger().warn(

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anuma/sdk",
-  "version": "7.3.2",
+  "version": "7.3.3",
   "description": "",
   "main": "./dist/index.cjs",
   "module": "./dist/index.mjs",