@anuma/sdk 6.7.3 → 6.8.1

package/dist/server/{chunk-I2U35Z3O.mjs → chunk-KF5OU4IN.mjs}

@@ -40,6 +40,12 @@ function withInternalFlowMarker(systemPrompt) {
 ${systemPrompt}`;
 }
 
+// src/lib/taskType.ts
+var TASK_TYPE_HEADER = "X-Anuma-Task-Type";
+function taskTypeHeader(taskType) {
+  return taskType ? { [TASK_TYPE_HEADER]: taskType } : {};
+}
+
 // src/lib/memory/portalLlm.ts
 function defaultBaseUrl() {
   return typeof process !== "undefined" && process.env?.ANUMA_PORTAL_BASE_URL || BASE_URL;
@@ -182,7 +188,13 @@ async function attemptPortalJson(req, endpoint) {
   try {
     response = await fetchImpl(`${baseUrl}${endpoint}`, {
       method: "POST",
-      headers: { ...authHeaders, "Content-Type": "application/json" },
+      // Attached HERE, at the one transport every background memory op shares, so
+      // a new flow gets the provenance by construction rather than by remembering.
+      headers: {
+        ...authHeaders,
+        ...taskTypeHeader(req.taskType),
+        "Content-Type": "application/json"
+      },
       body: JSON.stringify(requestBody),
       signal: controller.signal
     });
@@ -326,6 +338,7 @@ export {
   getLogger,
   INTERNAL_FLOW_MARKER,
   withInternalFlowMarker,
+  taskTypeHeader,
   validateEndpointOverride,
   supportsResponseFormat,
   resolvePortalAuthHeaders,

package/dist/server/{chunk-5KIUGHQZ.mjs → chunk-MEQ2PAEX.mjs}

@@ -1,6 +1,6 @@
 import {
   getLogger
-} from "./chunk-I2U35Z3O.mjs";
+} from "./chunk-KF5OU4IN.mjs";
 
 // src/lib/pii/patterns.ts
 function luhnCheck(digits) {

package/dist/server/{consolidate-S7QZ7V6T.mjs → consolidate-QDWEYFTZ.mjs}

@@ -1,10 +1,10 @@
 import {
   notifyConsolidationFallback,
   resolvePiiRedactor
-} from "./chunk-5KIUGHQZ.mjs";
+} from "./chunk-MEQ2PAEX.mjs";
 import {
   callPortalJsonCompletion
-} from "./chunk-I2U35Z3O.mjs";
+} from "./chunk-KF5OU4IN.mjs";
 import "./chunk-3RPGXXPZ.mjs";
 
 // src/lib/memory/consolidate.ts
@@ -77,6 +77,7 @@ ${candidateText}`;
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      taskType: "memory_consolidate",
       model: options.model ?? DEFAULT_CONSOLIDATION_MODEL,
       systemPrompt: SYSTEM_PROMPT,
       userMessage,

package/dist/server/index.cjs

@@ -17132,6 +17132,18 @@ var init_internalFlowMarker = __esm({
   }
 });
 
+// src/lib/taskType.ts
+function taskTypeHeader(taskType) {
+  return taskType ? { [TASK_TYPE_HEADER]: taskType } : {};
+}
+var TASK_TYPE_HEADER;
+var init_taskType = __esm({
+  "src/lib/taskType.ts"() {
+    "use strict";
+    TASK_TYPE_HEADER = "X-Anuma-Task-Type";
+  }
+});
+
 // src/lib/db/entities/types.ts
 function normalizeEntityName(name) {
   return name.trim().toLowerCase();
@@ -20697,7 +20709,13 @@ async function attemptPortalJson(req, endpoint) {
   try {
     response = await fetchImpl(`${baseUrl}${endpoint}`, {
       method: "POST",
-      headers: { ...authHeaders, "Content-Type": "application/json" },
+      // Attached HERE, at the one transport every background memory op shares, so
+      // a new flow gets the provenance by construction rather than by remembering.
+      headers: {
+        ...authHeaders,
+        ...taskTypeHeader(req.taskType),
+        "Content-Type": "application/json"
+      },
       body: JSON.stringify(requestBody),
       signal: controller.signal
     });
@@ -20844,6 +20862,7 @@ var init_portalLlm = __esm({
     init_endpointOverride();
     init_internalFlowMarker();
     init_logger();
+    init_taskType();
     RESPONSE_FORMAT_OK = /* @__PURE__ */ new Set(["openai", "inclusionai", "deepseek"]);
     RESPONSE_SCHEMA_OK = /* @__PURE__ */ new Set(["openai"]);
     RETRYABLE_HTTP = /* @__PURE__ */ new Set([408, 409, 425, 429]);
@@ -21177,6 +21196,7 @@ Output strict JSON, no prose: { "expand": [<entity names to expand, verbatim fro
         ...options.apiKey !== void 0 && { apiKey: options.apiKey },
         ...options.getToken !== void 0 && { getToken: options.getToken },
         ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+        taskType: "memory_graph",
         model: options.model ?? DEFAULT_REFINER_MODEL,
         systemPrompt,
         userMessage: `Question: ${query}
@@ -23912,6 +23932,7 @@ ${candidateText}`;
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      taskType: "memory_consolidate",
       model: options.model ?? DEFAULT_CONSOLIDATION_MODEL,
       systemPrompt: SYSTEM_PROMPT3,
       userMessage,
@@ -26786,6 +26807,7 @@ init_types2();
 init_clientConfig();
 init_internalFlowMarker();
 init_logger();
+init_taskType();
 var DEFAULT_SUMMARY_TOKEN_THRESHOLD = 4e3;
 var DEFAULT_SUMMARY_MIN_WINDOW_MESSAGES = 4;
 var DEFAULT_SUMMARY_MODEL = "cerebras/qwen-3-235b-a22b-instruct-2507";
@@ -26972,7 +26994,12 @@ async function callSummarizationLlm(prompt, model, token, baseUrl) {
       method: "POST",
       headers: {
         "Content-Type": "application/json",
-        Authorization: `Bearer ${token}`
+        Authorization: `Bearer ${token}`,
+        // Names the task so the portal can own this prompt server-side instead of
+        // trusting the one we build. Summarization is the archetypal Class-B flow:
+        // one fixed purpose, no conversation, and today its instructions are ours
+        // to reword.
+        ...taskTypeHeader("summarize")
       },
       body: JSON.stringify({
         model,
@@ -28815,6 +28842,7 @@ async function classifyInjectionCandidates(candidates, options) {
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      taskType: "memory_injection_check",
       model: options.model ?? DEFAULT_MODEL2,
       systemPrompt: SYSTEM_PROMPT2,
       userMessage: `Candidate facts:
@@ -29524,6 +29552,7 @@ async function extractFacts(messages, options) {
     ...options.apiKey !== void 0 && { apiKey: options.apiKey },
     ...options.getToken !== void 0 && { getToken: options.getToken },
     ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+    taskType: "memory_extract",
     ...options.endpointOverride !== void 0 && {
       endpointOverride: options.endpointOverride
     },
@@ -30092,6 +30121,7 @@ function createLlmDecayClassifier(options) {
           ...options.apiKey !== void 0 && { apiKey: options.apiKey },
           ...options.getToken !== void 0 && { getToken: options.getToken },
           ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+          taskType: "memory_decay",
           model: options.model ?? DEFAULT_MODEL3,
           systemPrompt: SYSTEM_PROMPT5,
           userMessage: `Memory content:
@@ -31249,6 +31279,7 @@ When a memory refers to one of these, reuse the EXACT existing name (same spelli
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      taskType: "memory_topic",
       ...options.endpointOverride !== void 0 && {
         endpointOverride: options.endpointOverride
       },
@@ -31442,6 +31473,7 @@ ${evidence}`);
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      taskType: "memory_verify_support",
       model: options.model ?? DEFAULT_MODEL5,
       systemPrompt: SYSTEM_PROMPT7,
       userMessage: `Stored facts and their sources:

package/dist/server/index.d.mts

@@ -4680,6 +4680,44 @@ interface EntityOperationsContext {
     allowUnscopedRows?: boolean;
 }
 
+/**
+ * Shared portal LLM call for the memory pipeline.
+ *
+ * The three LLM-driven steps (`extractFacts`, `consolidateMemory`,
+ * `decomposeQuery`) share the same shape: POST to `/api/v1/chat/completions`
+ * with a system+user prompt and `response_format: json_object`, parse
+ * `choices[0].message.content` as JSON, validate, return null on any failure.
+ * This helper owns the fetch+timeout+parse+log boilerplate so each caller
+ * stays focused on its prompt and validator.
+ *
+ * PROMPT CACHING. The portal inserts explicit prompt-cache breakpoints only for
+ * `anthropic/` models. Its request schema does carry OpenAI's
+ * `prompt_cache_key`, but that is a routing hint for OpenAI's own implicit
+ * cache rather than a switch that enables caching, and we deliberately don't
+ * send it: the memory pipeline runs open-weights models by default (gpt-oss for
+ * extraction/topics, ling for consolidation), and this module already has one
+ * hard lesson about OpenAI-standard fields a non-OpenAI provider hard-400s on
+ * (`response_format`, below). Revisit only with a verified per-provider gate.
+ *
+ * So on the models this pipeline actually uses, the only cache that can hit is
+ * the provider's own implicit prefix cache — which makes prefix stability the
+ * one lever a caller controls. Messages are assembled system-first (below), so
+ * every caller must pass a CONSTANT `systemPrompt` and keep all per-call
+ * material in `userMessage`: a single interpolated timestamp, id, or count in a
+ * system prompt shifts every byte after it and misses the cache on every
+ * request thereafter. The memory callers are regression-gated on that property
+ * in `promptPrefixStability.test.ts`.
+ *
+ * Whether those implicit caches hit is observable: each successful response
+ * logs the portal's reported token counts at debug level, including
+ * `portal.cached_tokens` (the prompt tokens the provider served from cache —
+ * the portal omits the field when there was no hit). See {@link logPortalUsage}.
+ *
+ * There is no batch/async-bulk route on the portal, so a backfill or reindex
+ * has no discounted-throughput path either; its only levers are the same stable
+ * prefix and caller-side batching (e.g. topicExtract's 10-memories-per-call).
+ */
+
 /**
  * Auth for portal LLM calls (extraction, consolidation, decomposition,
  * reflection). Mirrors `memoryEngine`'s `EmbeddingOptions` dual-auth:

package/dist/server/index.d.ts

@@ -4680,6 +4680,44 @@ interface EntityOperationsContext {
     allowUnscopedRows?: boolean;
 }
 
+/**
+ * Shared portal LLM call for the memory pipeline.
+ *
+ * The three LLM-driven steps (`extractFacts`, `consolidateMemory`,
+ * `decomposeQuery`) share the same shape: POST to `/api/v1/chat/completions`
+ * with a system+user prompt and `response_format: json_object`, parse
+ * `choices[0].message.content` as JSON, validate, return null on any failure.
+ * This helper owns the fetch+timeout+parse+log boilerplate so each caller
+ * stays focused on its prompt and validator.
+ *
+ * PROMPT CACHING. The portal inserts explicit prompt-cache breakpoints only for
+ * `anthropic/` models. Its request schema does carry OpenAI's
+ * `prompt_cache_key`, but that is a routing hint for OpenAI's own implicit
+ * cache rather than a switch that enables caching, and we deliberately don't
+ * send it: the memory pipeline runs open-weights models by default (gpt-oss for
+ * extraction/topics, ling for consolidation), and this module already has one
+ * hard lesson about OpenAI-standard fields a non-OpenAI provider hard-400s on
+ * (`response_format`, below). Revisit only with a verified per-provider gate.
+ *
+ * So on the models this pipeline actually uses, the only cache that can hit is
+ * the provider's own implicit prefix cache — which makes prefix stability the
+ * one lever a caller controls. Messages are assembled system-first (below), so
+ * every caller must pass a CONSTANT `systemPrompt` and keep all per-call
+ * material in `userMessage`: a single interpolated timestamp, id, or count in a
+ * system prompt shifts every byte after it and misses the cache on every
+ * request thereafter. The memory callers are regression-gated on that property
+ * in `promptPrefixStability.test.ts`.
+ *
+ * Whether those implicit caches hit is observable: each successful response
+ * logs the portal's reported token counts at debug level, including
+ * `portal.cached_tokens` (the prompt tokens the provider served from cache —
+ * the portal omits the field when there was no hit). See {@link logPortalUsage}.
+ *
+ * There is no batch/async-bulk route on the portal, so a backfill or reindex
+ * has no discounted-throughput path either; its only levers are the same stable
+ * prefix and caller-side batching (e.g. topicExtract's 10-memories-per-call).
+ */
+
 /**
  * Auth for portal LLM calls (extraction, consolidation, decomposition,
  * reflection). Mirrors `memoryEngine`'s `EmbeddingOptions` dual-auth:

package/dist/server/index.mjs

@@ -141,12 +141,12 @@ import {
   updateVaultMemoryOp,
   upsertMessageOp,
   vaultMemoryToStored
-} from "./chunk-G6IH6QTZ.mjs";
+} from "./chunk-FS6DNUNT.mjs";
 import {
   createStreamingDeAnonymizer,
   notifyConsolidationFallback,
   resolvePiiRedactor
-} from "./chunk-5KIUGHQZ.mjs";
+} from "./chunk-MEQ2PAEX.mjs";
 import {
   INTERNAL_FLOW_MARKER,
   callPortalJsonCompletion,
@@ -154,9 +154,10 @@ import {
   getLogger,
   resolvePortalAuthHeaders,
   supportsResponseFormat,
+  taskTypeHeader,
   validateEndpointOverride,
   withInternalFlowMarker
-} from "./chunk-I2U35Z3O.mjs";
+} from "./chunk-KF5OU4IN.mjs";
 import {
   BASE_URL,
   __commonJS,
@@ -3408,7 +3409,12 @@ async function callSummarizationLlm(prompt, model, token, baseUrl) {
       method: "POST",
       headers: {
         "Content-Type": "application/json",
-        Authorization: `Bearer ${token}`
+        Authorization: `Bearer ${token}`,
+        // Names the task so the portal can own this prompt server-side instead of
+        // trusting the one we build. Summarization is the archetypal Class-B flow:
+        // one fixed purpose, no conversation, and today its instructions are ours
+        // to reword.
+        ...taskTypeHeader("summarize")
       },
       body: JSON.stringify({
         model,
@@ -5219,6 +5225,7 @@ async function classifyInjectionCandidates(candidates, options) {
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      taskType: "memory_injection_check",
       model: options.model ?? DEFAULT_MODEL,
       systemPrompt: SYSTEM_PROMPT,
       userMessage: `Candidate facts:
@@ -5693,7 +5700,7 @@ async function tryConsolidate(trimmed, ctx, options, prepared) {
     content: m.content,
     similarity: m.similarity
   }));
-  const { consolidateMemory: doConsolidate } = await import("./consolidate-S7QZ7V6T.mjs");
+  const { consolidateMemory: doConsolidate } = await import("./consolidate-QDWEYFTZ.mjs");
   const decision = await doConsolidate(trimmed, candidates, consolidateOptions);
   if (decision.action === "create") return null;
   if (decision.action === "supersede" && decision.content) {
@@ -5918,6 +5925,7 @@ async function extractFacts(messages, options) {
     ...options.apiKey !== void 0 && { apiKey: options.apiKey },
     ...options.getToken !== void 0 && { getToken: options.getToken },
     ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+    taskType: "memory_extract",
     ...options.endpointOverride !== void 0 && {
       endpointOverride: options.endpointOverride
     },
@@ -6480,6 +6488,7 @@ function createLlmDecayClassifier(options) {
           ...options.apiKey !== void 0 && { apiKey: options.apiKey },
           ...options.getToken !== void 0 && { getToken: options.getToken },
           ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+          taskType: "memory_decay",
           model: options.model ?? DEFAULT_MODEL2,
           systemPrompt: SYSTEM_PROMPT3,
           userMessage: `Memory content:
@@ -7603,6 +7612,7 @@ When a memory refers to one of these, reuse the EXACT existing name (same spelli
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      taskType: "memory_topic",
       ...options.endpointOverride !== void 0 && {
         endpointOverride: options.endpointOverride
       },
@@ -7792,6 +7802,7 @@ ${evidence}`);
       ...options.apiKey !== void 0 && { apiKey: options.apiKey },
       ...options.getToken !== void 0 && { getToken: options.getToken },
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
+      taskType: "memory_verify_support",
       model: options.model ?? DEFAULT_MODEL4,
       systemPrompt: SYSTEM_PROMPT5,
       userMessage: `Stored facts and their sources:

package/dist/server/{recall-BIPVQMG4.mjs → recall-2GAPE2XH.mjs}

@@ -1,7 +1,7 @@
 import {
   recall
-} from "./chunk-G6IH6QTZ.mjs";
-import "./chunk-I2U35Z3O.mjs";
+} from "./chunk-FS6DNUNT.mjs";
+import "./chunk-KF5OU4IN.mjs";
 import "./chunk-3RPGXXPZ.mjs";
 export {
   recall

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anuma/sdk",
-  "version": "6.7.3",
+  "version": "6.8.1",
   "description": "",
   "main": "./dist/index.cjs",
   "module": "./dist/index.mjs",