@anuma/sdk 9.1.4 → 9.1.6

package/dist/memory/context.d.mts

@@ -183,6 +183,51 @@ interface EntityOperationsContext {
     canWrite?: () => Promise<boolean>;
 }
 
+/**
+ * Class-B task types — the vocabulary of the `X-Anuma-Task-Type` header.
+ *
+ * A Class-B flow has ONE fixed purpose and no conversation: extract memories
+ * from a snippet, pick a topic, consolidate duplicates, classify a query. Today
+ * each of those builds its own system prompt on the client and sends it as a
+ * `role:system` message, which makes the prompt a suggestion rather than a
+ * contract — a caller can reword it into "you are a general assistant" and get
+ * general chat at internal-endpoint prices.
+ *
+ * The fix is that the SERVER owns the text and the client only names the task.
+ * This header is the name. The portal allowlists it against its own copy of this
+ * enum (ai-portal `internal/systemprompt/tasks.go`) and DROPS anything it does
+ * not recognise, so an unknown value degrades to "no task declared" rather than
+ * failing the request.
+ *
+ * ⚠ These string values are a wire contract with ai-portal. Adding one here
+ * before the portal knows it is harmless (it is dropped); renaming or removing
+ * one silently stops that flow being counted. Keep them in step with
+ * `AllTaskTypes` on the portal side, and see ai-portal
+ * `docs/BACKEND_OWNED_TASK_PROMPTS.md` for the rollout.
+ *
+ * ⚠ RELEASE ORDER. This union is a PUBLIC type, so a split is breaking in both
+ * directions and has to be sequenced:
+ *
+ *   1. merge + publish this package;
+ *   2. bump `@anuma/sdk` in the consumer and switch its call sites to the new
+ *      names in the same commit.
+ *
+ * A consumer on the OLD version cannot compile the new names, and a consumer on
+ * the NEW version cannot compile the retired ones (`app_inspiration` is gone
+ * here, replaced by the three `app_inspiration_*` names). Both halves of that
+ * are type-only — the wire itself always degrades safely, because the portal
+ * drops any value it does not recognise. Concretely: ai-memoryless-client#5782
+ * is blocked on this publish for exactly that reason.
+ *
+ * Deliberately ABSENT: agent personas, `reflect()` when it answers the user, and
+ * the app-idea chat. Those carry an open conversation with the user, so a fixed
+ * server-owned prompt would be wrong rather than merely premature — the same
+ * boundary {@link INTERNAL_FLOW_MARKER} draws.
+ *
+ * @public
+ */
+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" | "memory_recall_embed" | "memory_retain_embed" | "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";
+
 /**
  * Auth for portal LLM calls (extraction, consolidation, decomposition,
  * reflection). Mirrors `memoryEngine`'s `EmbeddingOptions` dual-auth:
@@ -202,6 +247,76 @@ interface PortalLlmAuth {
     getToken?: () => Promise<string | null>;
 }
 
+/**
+ * Options for embedding generation
+ *
+ * Supports two auth methods:
+ * - `getToken`: For Privy identity tokens (uses Authorization: Bearer header)
+ * - `apiKey`: For direct API keys (uses X-API-Key header)
+ *
+ * At least one of `getToken` or `apiKey` must be provided.
+ */
+interface EmbeddingOptions {
+    taskType?: TaskType;
+    /** Function to get auth token (e.g., Privy's getIdentityToken). Uses Authorization: Bearer header. */
+    getToken?: () => Promise<string | null>;
+    /** Direct API key for server-side usage. Uses X-API-Key header. */
+    apiKey?: string;
+    /** Base URL for the API */
+    baseUrl?: string;
+    /** Embedding model to use */
+    model?: string;
+    /** Max texts per API call for batch embeddings (default: 100). Larger arrays are split into chunks. */
+    batchSize?: number;
+    /**
+     * Optional in-memory cache for embedding vectors. When provided, texts
+     * are looked up in this map before calling the API, and new embeddings
+     * are stored after generation. Useful when the same texts are embedded
+     * repeatedly (e.g., across eval iterations or re-indexing runs).
+     *
+     * Values are stored as `Float32Array` (the embedding model's native
+     * precision) rather than a float64 `number[]`, halving the resident RAM of
+     * the cache with no precision loss. `generateEmbedding(s)` still return
+     * `number[]` at the API boundary, so callers are unaffected.
+     */
+    cache?: Map<string, Float32Array>;
+    /** Called after each embedding API call with the token usage from the response. */
+    onUsage?: (usage: {
+        promptTokens: number;
+        totalTokens: number;
+    }) => void;
+    /**
+     * Optional transform applied to each text immediately before it is sent to
+     * the embeddings endpoint (e.g. `PiiRedactor.maskText`). The cache and result
+     * ordering still key on the original text — only the API request body is
+     * transformed — so callers can keep storing/displaying the original value
+     * while real PII never reaches the server. Used when PII redaction is active.
+     */
+    maskInput?: (text: string) => string;
+    /**
+     * Deadline, in ms, for EACH embeddings HTTP attempt (default 15000). An
+     * attempt that exceeds it is aborted and counts as a transient failure, so the
+     * bounded retry still applies. `0` disables the deadline.
+     */
+    timeoutMs?: number;
+    /**
+     * Deadline, in ms, for the `getToken()` read that precedes a request (default
+     * 10000). A provider that never settles rejects instead of hanging the
+     * embedding — and the recall waiting on it. `0` disables the deadline.
+     */
+    tokenTimeoutMs?: number;
+    /**
+     * Overall deadline, in ms, for one `generateEmbedding` or `generateEmbeddings` call — the token read,
+     * every retry attempt and the backoff between them. Unset (the default) means
+     * only the per-attempt deadlines apply, which is right for background/bulk
+     * embeds. The recall query path sets it (see
+     * `RecallOptions.queryEmbedTotalTimeoutMs`) so an outage degrades a turn to
+     * BM25 within a few seconds instead of ~4 x `timeoutMs`. Batch calls share
+     * one budget across authentication, all chunks, retries and backoff.
+     */
+    totalTimeoutMs?: number;
+}
+
 interface RecencyOptions {
     /** Override "now" — useful for deterministic tests and back-dated benchmarks. */
     now?: Date;
@@ -213,6 +328,13 @@ interface RecencyOptions {
     noDateMultiplier?: number;
 }
 
+type VaultCorpusCache = WeakMap<VaultMemoryOperationsContext, Map<string, Promise<StoredVaultMemory[]>>>;
+type RowRequests = Map<EmbeddingOptions["maskInput"], Map<string, Promise<number[]>>>;
+interface VaultPreparation {
+    corpora: VaultCorpusCache;
+    rows: WeakMap<VaultMemoryOperationsContext, RowRequests>;
+}
+
 /**
  * How a fact's evidence has been evolving.
  *
@@ -300,6 +422,7 @@ interface RankedMemory {
     updatedAt: Date;
 }
 interface RecallOptions {
+    /** @internal */ vaultPreparation?: VaultPreparation;
     /** Embedding cache shared with the rest of the turn, so the query is embedded once across lanes and the send. */
     embeddingCache?: Map<string, Float32Array>;
     /** Which kinds to search. Default: ['fact']. */

package/dist/memory/context.d.ts

@@ -183,6 +183,51 @@ interface EntityOperationsContext {
     canWrite?: () => Promise<boolean>;
 }
 
+/**
+ * Class-B task types — the vocabulary of the `X-Anuma-Task-Type` header.
+ *
+ * A Class-B flow has ONE fixed purpose and no conversation: extract memories
+ * from a snippet, pick a topic, consolidate duplicates, classify a query. Today
+ * each of those builds its own system prompt on the client and sends it as a
+ * `role:system` message, which makes the prompt a suggestion rather than a
+ * contract — a caller can reword it into "you are a general assistant" and get
+ * general chat at internal-endpoint prices.
+ *
+ * The fix is that the SERVER owns the text and the client only names the task.
+ * This header is the name. The portal allowlists it against its own copy of this
+ * enum (ai-portal `internal/systemprompt/tasks.go`) and DROPS anything it does
+ * not recognise, so an unknown value degrades to "no task declared" rather than
+ * failing the request.
+ *
+ * ⚠ These string values are a wire contract with ai-portal. Adding one here
+ * before the portal knows it is harmless (it is dropped); renaming or removing
+ * one silently stops that flow being counted. Keep them in step with
+ * `AllTaskTypes` on the portal side, and see ai-portal
+ * `docs/BACKEND_OWNED_TASK_PROMPTS.md` for the rollout.
+ *
+ * ⚠ RELEASE ORDER. This union is a PUBLIC type, so a split is breaking in both
+ * directions and has to be sequenced:
+ *
+ *   1. merge + publish this package;
+ *   2. bump `@anuma/sdk` in the consumer and switch its call sites to the new
+ *      names in the same commit.
+ *
+ * A consumer on the OLD version cannot compile the new names, and a consumer on
+ * the NEW version cannot compile the retired ones (`app_inspiration` is gone
+ * here, replaced by the three `app_inspiration_*` names). Both halves of that
+ * are type-only — the wire itself always degrades safely, because the portal
+ * drops any value it does not recognise. Concretely: ai-memoryless-client#5782
+ * is blocked on this publish for exactly that reason.
+ *
+ * Deliberately ABSENT: agent personas, `reflect()` when it answers the user, and
+ * the app-idea chat. Those carry an open conversation with the user, so a fixed
+ * server-owned prompt would be wrong rather than merely premature — the same
+ * boundary {@link INTERNAL_FLOW_MARKER} draws.
+ *
+ * @public
+ */
+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" | "memory_recall_embed" | "memory_retain_embed" | "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";
+
 /**
  * Auth for portal LLM calls (extraction, consolidation, decomposition,
  * reflection). Mirrors `memoryEngine`'s `EmbeddingOptions` dual-auth:
@@ -202,6 +247,76 @@ interface PortalLlmAuth {
     getToken?: () => Promise<string | null>;
 }
 
+/**
+ * Options for embedding generation
+ *
+ * Supports two auth methods:
+ * - `getToken`: For Privy identity tokens (uses Authorization: Bearer header)
+ * - `apiKey`: For direct API keys (uses X-API-Key header)
+ *
+ * At least one of `getToken` or `apiKey` must be provided.
+ */
+interface EmbeddingOptions {
+    taskType?: TaskType;
+    /** Function to get auth token (e.g., Privy's getIdentityToken). Uses Authorization: Bearer header. */
+    getToken?: () => Promise<string | null>;
+    /** Direct API key for server-side usage. Uses X-API-Key header. */
+    apiKey?: string;
+    /** Base URL for the API */
+    baseUrl?: string;
+    /** Embedding model to use */
+    model?: string;
+    /** Max texts per API call for batch embeddings (default: 100). Larger arrays are split into chunks. */
+    batchSize?: number;
+    /**
+     * Optional in-memory cache for embedding vectors. When provided, texts
+     * are looked up in this map before calling the API, and new embeddings
+     * are stored after generation. Useful when the same texts are embedded
+     * repeatedly (e.g., across eval iterations or re-indexing runs).
+     *
+     * Values are stored as `Float32Array` (the embedding model's native
+     * precision) rather than a float64 `number[]`, halving the resident RAM of
+     * the cache with no precision loss. `generateEmbedding(s)` still return
+     * `number[]` at the API boundary, so callers are unaffected.
+     */
+    cache?: Map<string, Float32Array>;
+    /** Called after each embedding API call with the token usage from the response. */
+    onUsage?: (usage: {
+        promptTokens: number;
+        totalTokens: number;
+    }) => void;
+    /**
+     * Optional transform applied to each text immediately before it is sent to
+     * the embeddings endpoint (e.g. `PiiRedactor.maskText`). The cache and result
+     * ordering still key on the original text — only the API request body is
+     * transformed — so callers can keep storing/displaying the original value
+     * while real PII never reaches the server. Used when PII redaction is active.
+     */
+    maskInput?: (text: string) => string;
+    /**
+     * Deadline, in ms, for EACH embeddings HTTP attempt (default 15000). An
+     * attempt that exceeds it is aborted and counts as a transient failure, so the
+     * bounded retry still applies. `0` disables the deadline.
+     */
+    timeoutMs?: number;
+    /**
+     * Deadline, in ms, for the `getToken()` read that precedes a request (default
+     * 10000). A provider that never settles rejects instead of hanging the
+     * embedding — and the recall waiting on it. `0` disables the deadline.
+     */
+    tokenTimeoutMs?: number;
+    /**
+     * Overall deadline, in ms, for one `generateEmbedding` or `generateEmbeddings` call — the token read,
+     * every retry attempt and the backoff between them. Unset (the default) means
+     * only the per-attempt deadlines apply, which is right for background/bulk
+     * embeds. The recall query path sets it (see
+     * `RecallOptions.queryEmbedTotalTimeoutMs`) so an outage degrades a turn to
+     * BM25 within a few seconds instead of ~4 x `timeoutMs`. Batch calls share
+     * one budget across authentication, all chunks, retries and backoff.
+     */
+    totalTimeoutMs?: number;
+}
+
 interface RecencyOptions {
     /** Override "now" — useful for deterministic tests and back-dated benchmarks. */
     now?: Date;
@@ -213,6 +328,13 @@ interface RecencyOptions {
     noDateMultiplier?: number;
 }
 
+type VaultCorpusCache = WeakMap<VaultMemoryOperationsContext, Map<string, Promise<StoredVaultMemory[]>>>;
+type RowRequests = Map<EmbeddingOptions["maskInput"], Map<string, Promise<number[]>>>;
+interface VaultPreparation {
+    corpora: VaultCorpusCache;
+    rows: WeakMap<VaultMemoryOperationsContext, RowRequests>;
+}
+
 /**
  * How a fact's evidence has been evolving.
  *
@@ -300,6 +422,7 @@ interface RankedMemory {
     updatedAt: Date;
 }
 interface RecallOptions {
+    /** @internal */ vaultPreparation?: VaultPreparation;
     /** Embedding cache shared with the rest of the turn, so the query is embedded once across lanes and the send. */
     embeddingCache?: Map<string, Float32Array>;
     /** Which kinds to search. Default: ['fact']. */

package/dist/react/{chunk-7QX4J7KV.mjs → chunk-RSA3S6OA.mjs}

@@ -723,6 +723,12 @@ function assertTransportMatchesEndpoint(transport, endpoint) {
     `endpointOverride "${endpoint}" is the other transport's endpoint, but transport is "${transport}". The request body is built for the transport, so this sends the wrong shape and 400s without retry.`
   );
 }
+var TIME_BUDGET_EXHAUSTED = {
+  kind: "terminal",
+  code: "time-budget-exhausted",
+  reason: "time budget exhausted"
+};
+var MAX_TIMER_DELAY_MS = 2147483647;
 var RETRYABLE_HTTP = /* @__PURE__ */ new Set([408, 409, 425, 429]);
 function isRetryableStatus(status) {
   return status >= 500 || RETRYABLE_HTTP.has(status);
@@ -778,7 +784,8 @@ async function callPortalJsonCompletion(req) {
   }
   const maxAttempts = Math.max(1, req.maxAttempts ?? 3);
   const startedAt = Date.now();
-  const overBudget = () => req.totalTimeoutMs !== void 0 && Date.now() - startedAt >= req.totalTimeoutMs;
+  const deadlineAt = req.totalTimeoutMs === void 0 ? void 0 : startedAt + req.totalTimeoutMs;
+  const overBudget = () => deadlineAt !== void 0 && Date.now() >= deadlineAt;
   let lastFailure;
   let reinforce = false;
   const giveUp = () => {
@@ -791,17 +798,11 @@ async function callPortalJsonCompletion(req) {
       lastFailure = { reason: "time-budget-exhausted", attempts: attempt - 1 };
       break;
     }
-    const attemptReq = req.totalTimeoutMs !== void 0 ? {
-      ...req,
-      timeoutMs: Math.max(
-        1,
-        Math.min(req.timeoutMs ?? 6e4, req.totalTimeoutMs - (Date.now() - startedAt))
-      ),
-      ...reinforce && { reinforceJsonContract: true }
-    } : reinforce ? { ...req, reinforceJsonContract: true } : req;
+    const attemptReq = reinforce ? { ...req, reinforceJsonContract: true } : req;
     const outcome = await attemptPortalJson(
       { ...attemptReq, transport },
-      endpoint
+      endpoint,
+      deadlineAt
     );
     if (outcome.kind === "ok") {
       req.onAttempt?.({ attempt, ok: true });
@@ -841,13 +842,38 @@ async function callPortalJsonCompletion(req) {
   }
   return giveUp();
 }
-async function attemptPortalJson(req, endpoint) {
+async function attemptPortalJson(req, endpoint, deadlineAt) {
+  const controller = new AbortController();
+  if (deadlineAt === void 0 || deadlineAt - Date.now() > MAX_TIMER_DELAY_MS) {
+    return requestPortalJson(req, endpoint, controller);
+  }
+  const remainingMs = deadlineAt - Date.now();
+  if (remainingMs <= 0) return TIME_BUDGET_EXHAUSTED;
+  let timer;
+  try {
+    const outcome = await Promise.race([
+      new Promise((resolve) => {
+        timer = setTimeout(() => {
+          resolve(TIME_BUDGET_EXHAUSTED);
+          controller.abort();
+        }, remainingMs);
+      }),
+      requestPortalJson(req, endpoint, controller, deadlineAt)
+    ]);
+    return Date.now() >= deadlineAt ? TIME_BUDGET_EXHAUSTED : outcome;
+  } finally {
+    clearTimeout(timer);
+  }
+}
+async function requestPortalJson(req, endpoint, controller, deadlineAt) {
   const baseUrl = req.baseUrl ?? defaultBaseUrl();
   const fetchImpl = req.fetchFn ?? fetch;
-  const timeoutMs = req.timeoutMs ?? 6e4;
   const authHeaders = await resolvePortalAuthHeaders(req, req.tag);
+  const remainingMs = deadlineAt === void 0 ? Infinity : deadlineAt - Date.now();
+  if (controller.signal.aborted || remainingMs <= 0) return TIME_BUDGET_EXHAUSTED;
   if (authHeaders === null)
     return { kind: "terminal", code: "auth-unavailable", reason: "auth unavailable (no token)" };
+  const timeoutMs = Math.min(req.timeoutMs ?? 6e4, remainingMs);
   const usesPrefill = req.transport !== "responses" && req.model.startsWith("anthropic/");
   const prefill = usesPrefill ? "{" : "";
   const messages = [
@@ -857,7 +883,6 @@ async function attemptPortalJson(req, endpoint) {
   ];
   if (prefill) messages.push({ role: "assistant", content: prefill });
   const supportsJsonObjectFormat = supportsResponseFormat(req.model);
-  const controller = new AbortController();
   const timer = setTimeout(() => controller.abort(), timeoutMs);
   const requestBody = req.transport === "responses" ? {
     model: req.model,

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

@@ -4,7 +4,7 @@ import {
   notifyConsolidationFallback,
   requiresResponsesTransport,
   resolvePiiRedactor
-} from "./chunk-7QX4J7KV.mjs";
+} from "./chunk-RSA3S6OA.mjs";
 import "./chunk-GYKJOWEO.mjs";
 
 // src/lib/memory/consolidate.ts