@anuma/sdk 6.4.2 → 6.6.0

package/dist/server/index.cjs

@@ -15911,6 +15911,90 @@ var init_vector = __esm({
   }
 });
 
+// src/lib/processors/encoding.ts
+async function dataUrlToArrayBuffer(dataUrl) {
+  if (dataUrl.startsWith("blob:") || dataUrl.startsWith("http://") || dataUrl.startsWith("https://")) {
+    const response = await fetch(dataUrl);
+    return response.arrayBuffer();
+  }
+  const base64 = dataUrl.split(",")[1];
+  if (typeof Buffer !== "undefined") {
+    const buf = Buffer.from(base64, "base64");
+    return buf.buffer.slice(buf.byteOffset, buf.byteOffset + buf.byteLength);
+  }
+  const binary = atob(base64);
+  const buffer = new ArrayBuffer(binary.length);
+  const view = new Uint8Array(buffer);
+  for (let i = 0; i < binary.length; i++) {
+    view[i] = binary.charCodeAt(i);
+  }
+  return buffer;
+}
+function uint8ArrayToBase64(data) {
+  if (typeof Buffer !== "undefined") {
+    return Buffer.from(data).toString("base64");
+  }
+  let base64 = "";
+  for (let i = 0; i < data.length; i += B64_ENCODE_CHUNK) {
+    const slice = data.subarray(i, i + B64_ENCODE_CHUNK);
+    let binary = "";
+    for (let j = 0; j < slice.length; j++) {
+      binary += String.fromCharCode(slice[j]);
+    }
+    base64 += btoa(binary);
+  }
+  return base64;
+}
+function base64ToUint8Array(base64) {
+  if (typeof Buffer !== "undefined") {
+    const buf = Buffer.from(base64, "base64");
+    return new Uint8Array(buf.buffer.slice(buf.byteOffset, buf.byteOffset + buf.byteLength));
+  }
+  const binary = atob(base64);
+  const bytes = new Uint8Array(binary.length);
+  for (let i = 0; i < binary.length; i++) {
+    bytes[i] = binary.charCodeAt(i);
+  }
+  return bytes;
+}
+var B64_ENCODE_CHUNK;
+var init_encoding = __esm({
+  "src/lib/processors/encoding.ts"() {
+    "use strict";
+    B64_ENCODE_CHUNK = 32768 * 3;
+  }
+});
+
+// src/lib/memoryEngine/vectorEncoding.ts
+function decodeChunkVector(vector, onMalformed) {
+  if (!vector || vector.length === 0) return new Float32Array(0);
+  if (typeof vector !== "string") return Float32Array.from(vector);
+  if (vector.length % 4 !== 0 || !CANONICAL_BASE64.test(vector)) {
+    onMalformed?.();
+    return new Float32Array(0);
+  }
+  try {
+    const bytes = base64ToUint8Array(vector);
+    if (bytes.byteLength === 0 || bytes.byteLength % BYTES_PER_FLOAT32 !== 0) {
+      onMalformed?.();
+      return new Float32Array(0);
+    }
+    return new Float32Array(bytes.buffer, bytes.byteOffset, bytes.byteLength / BYTES_PER_FLOAT32);
+  } catch {
+    onMalformed?.();
+    return new Float32Array(0);
+  }
+}
+var BYTES_PER_FLOAT32, CANONICAL_BASE64;
+var init_vectorEncoding = __esm({
+  "src/lib/memoryEngine/vectorEncoding.ts"() {
+    "use strict";
+    init_encoding();
+    BYTES_PER_FLOAT32 = 4;
+    CANONICAL_BASE64 = /^[A-Za-z0-9+/]+={0,2}$/;
+  }
+});
+
 // src/lib/db/chat/conversationEncryption.ts
 async function encryptConversationFields(conversation, address, signMessage, embeddedWalletSigner) {
   if (!address || !signMessage) {
@@ -16797,6 +16881,7 @@ async function searchChunksOp(ctx, queryVector, options) {
   const messages = await ctx.messagesCollection.query(...queryConditions).fetch();
   const candidates = [];
   let staleSkipped = 0;
+  let unreadableVectors = 0;
   for (const message of messages) {
     const msgConvId = String(message._getRaw("conversation_id"));
     if (!activeConversationIds.has(msgConvId)) continue;
@@ -16834,7 +16919,9 @@ async function searchChunksOp(ctx, queryVector, options) {
     const chunks = await readJsonField(message, "chunks", ctx.walletAddress);
     if (chunks && chunks.length > 0) {
       const vectors = chunks.map(
-        (chunk) => chunk.vector && chunk.vector.length > 0 ? Float32Array.from(chunk.vector) : new Float32Array(0)
+        (chunk) => decodeChunkVector(chunk.vector, () => {
+          unreadableVectors++;
+        })
       );
       chunkCache?.set(message.id, { version, chunks: vectors });
       for (let ci = 0; ci < chunks.length; ci++) {
@@ -16865,6 +16952,11 @@ async function searchChunksOp(ctx, queryVector, options) {
       `searchChunksOp: skipped ${staleSkipped} messages whose embedding model differs from the current model (${embeddingModel}) \u2014 re-embed via chunkAndEmbedAllMessages`
     );
   }
+  if (unreadableVectors > 0) {
+    getLogger().warn(
+      `searchChunksOp: could not read ${unreadableVectors} chunk vectors and left them out of ranking \u2014 the stored value is corrupt, so those chunks stay unsearchable until the message is re-embedded via chunkAndEmbedAllMessages`
+    );
+  }
   const topK = candidates.sort((a, b) => b.similarity - a.similarity).slice(0, limit);
   const uniqueMessages = /* @__PURE__ */ new Map();
   for (const c of topK) uniqueMessages.set(c.message.id, c.message);
@@ -16986,6 +17078,7 @@ var init_operations = __esm({
     import_uuid2 = require("uuid");
     init_logger();
     init_vector();
+    init_vectorEncoding();
     init_encryption_utils();
     init_conversationEncryption();
     init_encryption();
@@ -17025,6 +17118,20 @@ var init_clientConfig = __esm({
   }
 });
 
+// src/lib/internalFlowMarker.ts
+function withInternalFlowMarker(systemPrompt) {
+  if (systemPrompt.includes(INTERNAL_FLOW_MARKER)) return systemPrompt;
+  return `${INTERNAL_FLOW_MARKER}
+${systemPrompt}`;
+}
+var INTERNAL_FLOW_MARKER;
+var init_internalFlowMarker = __esm({
+  "src/lib/internalFlowMarker.ts"() {
+    "use strict";
+    INTERNAL_FLOW_MARKER = "Anuma internal first-party flow (not user chat).";
+  }
+});
+
 // src/lib/db/entities/types.ts
 function normalizeEntityName(name) {
   return name.trim().toLowerCase();
@@ -18657,6 +18764,15 @@ var init_chunking = __esm({
   }
 });
 
+// src/lib/chat/toolResults.ts
+var TOOL_RESULT_ORIGIN;
+var init_toolResults = __esm({
+  "src/lib/chat/toolResults.ts"() {
+    "use strict";
+    TOOL_RESULT_ORIGIN = "tool_result";
+  }
+});
+
 // src/lib/memoryEngine/constants.ts
 var DEFAULT_API_EMBEDDING_MODEL;
 var init_constants = __esm({
@@ -19742,6 +19858,11 @@ var init_generate = __esm({
 });
 
 // src/lib/memoryEngine/embeddings.ts
+function isNonEmbeddableOrigin(origin, reembedDiscarded = false) {
+  if (typeof origin !== "string") return false;
+  if (reembedDiscarded && origin === CHUNKS_DISCARDED_ORIGIN) return false;
+  return NON_EMBEDDABLE_ORIGINS.has(origin);
+}
 async function embedMessage(ctx, messageId, options) {
   const message = await getMessageOp(ctx, messageId);
   if (!message) {
@@ -19750,7 +19871,7 @@ async function embedMessage(ctx, messageId, options) {
   if (message.vector && message.vector.length > 0) {
     return message;
   }
-  if (message.origin === NON_EMBEDDABLE_ORIGIN) {
+  if (isNonEmbeddableOrigin(message.origin)) {
     return message;
   }
   if (isEncrypted(message.content)) {
@@ -19786,7 +19907,7 @@ async function embedAllMessages(ctx, options, filter) {
       if (message.role === "system") {
         continue;
       }
-      if (message.origin === NON_EMBEDDABLE_ORIGIN) {
+      if (isNonEmbeddableOrigin(message.origin, filter?.reembedDiscarded)) {
         continue;
       }
       considered++;
@@ -19826,7 +19947,7 @@ async function chunkAndEmbedMessage(ctx, messageId, options) {
   if (message.chunks && message.chunks.length > 0) {
     return message;
   }
-  if (message.origin === NON_EMBEDDABLE_ORIGIN) {
+  if (isNonEmbeddableOrigin(message.origin)) {
     return message;
   }
   if (isEncrypted(message.content)) {
@@ -19874,7 +19995,7 @@ async function chunkAndEmbedAllMessages(ctx, options, filter) {
       if (hasVector && !filter?.rechunkExisting && !isStale) continue;
       if (filter?.roles && !filter.roles.includes(message.role)) continue;
       if (message.role === "system") continue;
-      if (message.origin === NON_EMBEDDABLE_ORIGIN) continue;
+      if (isNonEmbeddableOrigin(message.origin, filter?.reembedDiscarded)) continue;
       considered++;
       if (isEncrypted(message.content)) {
         stillEncrypted++;
@@ -19943,10 +20064,11 @@ async function chunkAndEmbedAllMessages(ctx, options, filter) {
   }
   return embeddedCount;
 }
-var DEFAULT_MIN_CONTENT_LENGTH, NON_EMBEDDABLE_ORIGIN;
+var DEFAULT_MIN_CONTENT_LENGTH, CHUNKS_DISCARDED_ORIGIN, NON_EMBEDDABLE_ORIGINS;
 var init_embeddings = __esm({
   "src/lib/memoryEngine/embeddings.ts"() {
     "use strict";
+    init_toolResults();
     init_operations();
     init_encryption_utils();
     init_logger();
@@ -19955,7 +20077,8 @@ var init_embeddings = __esm({
     init_generate();
     init_generate();
     DEFAULT_MIN_CONTENT_LENGTH = 10;
-    NON_EMBEDDABLE_ORIGIN = "tool_result";
+    CHUNKS_DISCARDED_ORIGIN = "chunks_discarded";
+    NON_EMBEDDABLE_ORIGINS = /* @__PURE__ */ new Set([TOOL_RESULT_ORIGIN, CHUNKS_DISCARDED_ORIGIN]);
   }
 });
 
@@ -20538,7 +20661,12 @@ async function attemptPortalJson(req, endpoint) {
   const isAnthropic = req.model.startsWith("anthropic/");
   const prefill = isAnthropic ? "{" : "";
   const messages = [
-    { role: "system", content: req.systemPrompt },
+    // Every caller of this helper is a first-party BACKGROUND op (fact extraction,
+    // topic sweep, consolidation, classifiers, graph traversal, query
+    // decomposition). Marking here — the one place they all funnel through — is what
+    // keeps the portal's detector reading them as genuine rather than markerless.
+    // See ../internalFlowMarker.
+    { role: "system", content: withInternalFlowMarker(req.systemPrompt) },
     { role: "user", content: req.userMessage }
   ];
   if (prefill) messages.push({ role: "assistant", content: prefill });
@@ -20676,6 +20804,7 @@ var init_portalLlm = __esm({
     "use strict";
     init_clientConfig();
     init_endpointOverride();
+    init_internalFlowMarker();
     init_logger();
     RESPONSE_FORMAT_OK = /* @__PURE__ */ new Set(["openai", "inclusionai", "deepseek"]);
     RESPONSE_SCHEMA_OK = /* @__PURE__ */ new Set(["openai"]);
@@ -26617,6 +26746,7 @@ init_types2();
 
 // src/lib/chat/summarize.ts
 init_clientConfig();
+init_internalFlowMarker();
 init_logger();
 var DEFAULT_SUMMARY_TOKEN_THRESHOLD = 4e3;
 var DEFAULT_SUMMARY_MIN_WINDOW_MESSAGES = 4;
@@ -26809,7 +26939,17 @@ async function callSummarizationLlm(prompt, model, token, baseUrl) {
       body: JSON.stringify({
         model,
         stream: false,
-        messages: [{ role: "user", content: prompt }]
+        // The marker rides a system message because this call has no other
+        // provenance: it posts to the MAIN chat endpoint (not /utility/*), carries no
+        // conversationId, and its prompt is neither the app's chat base prompt nor any
+        // registered flow fingerprint — so without this it reads as markerless, i.e.
+        // as a scripted abuser. Progressive summarization runs for every free-tier
+        // long conversation, so at the PORTAL_DETECTION_REJECT_MARKERLESS flip it
+        // would 403 and silently fall back to verbatim history (context bloat + cost).
+        messages: [
+          { role: "system", content: INTERNAL_FLOW_MARKER },
+          { role: "user", content: prompt }
+        ]
       }),
       signal: controller.signal
     });
@@ -30496,6 +30636,7 @@ init_reranker();
 
 // src/lib/memory/synthesizeProfile.ts
 init_operations3();
+init_internalFlowMarker();
 init_logger();
 init_constants();
 init_embeddings();
@@ -30857,7 +30998,10 @@ async function synthesizeFacet(facet, ctx, options, prior) {
     limit,
     types: ["fact"],
     maxTokens: DEFAULT_FACET_MAX_TOKENS,
-    systemPrompt: buildFacetSystemPrompt(facet),
+    // Profile-facet synthesis is a background op, so it marks its own prompt.
+    // reflect() itself must NOT mark — it also serves the user's own questions,
+    // and that traffic is chat, not an internal flow. See ../internalFlowMarker.
+    systemPrompt: withInternalFlowMarker(buildFacetSystemPrompt(facet)),
     responseSchema: facetResponseSchema(facet.key),
     memories
   });
@@ -32619,44 +32763,7 @@ var WalletPoller = class {
 
 // src/lib/processors/ExcelProcessor.ts
 init_logger();
-
-// src/lib/processors/encoding.ts
-var B64_ENCODE_CHUNK = 32768 * 3;
-async function dataUrlToArrayBuffer(dataUrl) {
-  if (dataUrl.startsWith("blob:") || dataUrl.startsWith("http://") || dataUrl.startsWith("https://")) {
-    const response = await fetch(dataUrl);
-    return response.arrayBuffer();
-  }
-  const base64 = dataUrl.split(",")[1];
-  if (typeof Buffer !== "undefined") {
-    const buf = Buffer.from(base64, "base64");
-    return buf.buffer.slice(buf.byteOffset, buf.byteOffset + buf.byteLength);
-  }
-  const binary = atob(base64);
-  const buffer = new ArrayBuffer(binary.length);
-  const view = new Uint8Array(buffer);
-  for (let i = 0; i < binary.length; i++) {
-    view[i] = binary.charCodeAt(i);
-  }
-  return buffer;
-}
-function uint8ArrayToBase64(data) {
-  if (typeof Buffer !== "undefined") {
-    return Buffer.from(data).toString("base64");
-  }
-  let base64 = "";
-  for (let i = 0; i < data.length; i += B64_ENCODE_CHUNK) {
-    const slice = data.subarray(i, i + B64_ENCODE_CHUNK);
-    let binary = "";
-    for (let j = 0; j < slice.length; j++) {
-      binary += String.fromCharCode(slice[j]);
-    }
-    base64 += btoa(binary);
-  }
-  return base64;
-}
-
-// src/lib/processors/ExcelProcessor.ts
+init_encoding();
 if (typeof process !== "undefined" && typeof process.umask !== "function") {
   process.umask = (_mask) => 18;
 }
@@ -32835,6 +32942,7 @@ var PdfProcessor = class {
 
 // src/lib/processors/TextProcessor.ts
 init_logger();
+init_encoding();
 var MIME = {
   PLAIN: "text/plain",
   MARKDOWN: "text/markdown",
@@ -32918,6 +33026,7 @@ var TextProcessor = class {
 // src/lib/processors/WordProcessor.ts
 var import_mammoth = __toESM(require("mammoth"));
 init_logger();
+init_encoding();
 var WordProcessor = class {
   constructor() {
     this.name = "word";
@@ -32961,6 +33070,7 @@ var WordProcessor = class {
 // src/lib/processors/ZipProcessor.ts
 var import_jszip = __toESM(require("jszip"));
 init_logger();
+init_encoding();
 var ZipProcessor = class {
   constructor(options = {}) {
     this.name = "zip";

package/dist/server/index.d.mts

@@ -3423,13 +3423,24 @@ type ChatRole = "user" | "assistant" | "system";
  */
 type MessageFeedback = "like" | "dislike" | null;
 /**
- * Which producer synthesised a message, for rows the user did not type and the
- * UI does not render. Set at write time by that producer; a union so further
- * synthetic kinds can be added without another column.
+ * Provenance for a row that needs handling the content cannot justify. Set at
+ * write time by whoever produced or repaired the row; a union so further kinds
+ * can be added without another column.
+ *
+ * Read it as an enum, never as a boolean. The values do NOT share a meaning:
+ * `tool_result` marks a synthetic row that is both unindexed and hidden, while
+ * `chunks_discarded` marks an ordinary typed message that is unindexed and still
+ * rendered. Any predicate that collapses this to "has an origin" hides real
+ * messages (see `isToolResultsRow`).
  *
  * - `tool_result`: the hidden `[Tool Execution Results]` row built from a
  *   turn's auto-executed tool results. Skipped by the embedding sweep (see
- *   `memoryEngine/embeddings`).
+ *   `memoryEngine/embeddings`) AND hidden from the transcript.
+ * - `chunks_discarded`: an ordinary, still-rendered message whose chunk vectors
+ *   were built over `enc:v3:` ciphertext (sdk#864) and have been discarded
+ *   rather than re-embedded at the user's own expense (client#5618). Skipped by
+ *   the embedding sweep so nothing re-embeds it, but NOT hidden — only
+ *   `tool_result` hides a row, and `isToolResultsRow` is what draws that line.
  * - undefined/null: a normal message, or any row written before v44.
  *
  * Stored in the plaintext `origin` column — never encrypted. The deferred
@@ -3437,7 +3448,7 @@ type MessageFeedback = "like" | "dislike" | null;
  * encrypted flag would be unreadable exactly where it matters and the skip
  * would fail open.
  */
-type MessageOrigin = "tool_result";
+type MessageOrigin = "tool_result" | "chunks_discarded";
 /**
  * Metadata for files attached to messages.
  *
@@ -3633,7 +3644,18 @@ interface StoredConversationSummary {
 interface MessageChunk {
     /** The chunk text */
     text: string;
-    /** Embedding vector for this chunk */
+    /**
+     * Embedding vector for this chunk.
+     *
+     * Read it through `decodeChunkVector` rather than passing it straight to a
+     * numeric call site. This stays `number[]` only while the writer still emits
+     * JSON arrays; it widens to `number[] | string` in the release that flips the
+     * writer to base64 float32 (sdk#862), and that widening is the whole reason
+     * that release is a major. Readers already accept both, so code written
+     * against the decoder today needs no change when it happens — code written
+     * against the raw array degrades silently, because a string handed to
+     * `Float32Array.from` yields one NaN per character rather than an error.
+     */
     vector: number[];
     /** Character offset where this chunk starts in the original message */
     startOffset: number;
@@ -7373,6 +7395,17 @@ declare function embedAllMessages(ctx: StorageOperationsContext, options: Embedd
     roles?: ("user" | "assistant")[];
     /** Minimum content length to embed (default: 30). Shorter messages are skipped. */
     minContentLength?: number;
+    /**
+     * Re-index rows the ciphertext sweep marked {@link CHUNKS_DISCARDED_ORIGIN}.
+     * Off by default: this spends the user's own embedding credits, so it belongs
+     * to an explicit user action, never to a background pass. Opens that marker
+     * only — `tool_result` rows stay excluded.
+     *
+     * Per call, not a state change: a row that re-indexes successfully keeps its
+     * marker, so a later pass that omits this flag skips it again. See
+     * {@link CHUNKS_DISCARDED_ORIGIN}.
+     */
+    reembedDiscarded?: boolean;
 }): Promise<number>;
 /**
  * Chunk and embed a single message, storing chunk embeddings in the database.
@@ -7418,6 +7451,21 @@ declare function chunkAndEmbedAllMessages(ctx: StorageOperationsContext, options
     rechunkExisting?: boolean;
     /** Minimum content length to embed (default: 30). Shorter messages are skipped. */
     minContentLength?: number;
+    /**
+     * Re-index rows the ciphertext sweep marked {@link CHUNKS_DISCARDED_ORIGIN}.
+     * Off by default: this spends the user's own embedding credits, so it belongs
+     * to an explicit user action, never to a background pass. Opens that marker
+     * only — `tool_result` rows stay excluded.
+     *
+     * `rechunkExisting` cannot substitute for it: a discarded row has neither
+     * chunks nor vector, so it never reaches that check and falls straight to the
+     * origin gate.
+     *
+     * Per call, not a state change: a row that re-indexes successfully keeps its
+     * marker, so a later pass that omits this flag skips it again. See
+     * {@link CHUNKS_DISCARDED_ORIGIN}.
+     */
+    reembedDiscarded?: boolean;
 }): Promise<number>;
 
 /**

package/dist/server/index.d.ts

@@ -3423,13 +3423,24 @@ type ChatRole = "user" | "assistant" | "system";
  */
 type MessageFeedback = "like" | "dislike" | null;
 /**
- * Which producer synthesised a message, for rows the user did not type and the
- * UI does not render. Set at write time by that producer; a union so further
- * synthetic kinds can be added without another column.
+ * Provenance for a row that needs handling the content cannot justify. Set at
+ * write time by whoever produced or repaired the row; a union so further kinds
+ * can be added without another column.
+ *
+ * Read it as an enum, never as a boolean. The values do NOT share a meaning:
+ * `tool_result` marks a synthetic row that is both unindexed and hidden, while
+ * `chunks_discarded` marks an ordinary typed message that is unindexed and still
+ * rendered. Any predicate that collapses this to "has an origin" hides real
+ * messages (see `isToolResultsRow`).
  *
  * - `tool_result`: the hidden `[Tool Execution Results]` row built from a
  *   turn's auto-executed tool results. Skipped by the embedding sweep (see
- *   `memoryEngine/embeddings`).
+ *   `memoryEngine/embeddings`) AND hidden from the transcript.
+ * - `chunks_discarded`: an ordinary, still-rendered message whose chunk vectors
+ *   were built over `enc:v3:` ciphertext (sdk#864) and have been discarded
+ *   rather than re-embedded at the user's own expense (client#5618). Skipped by
+ *   the embedding sweep so nothing re-embeds it, but NOT hidden — only
+ *   `tool_result` hides a row, and `isToolResultsRow` is what draws that line.
  * - undefined/null: a normal message, or any row written before v44.
  *
  * Stored in the plaintext `origin` column — never encrypted. The deferred
@@ -3437,7 +3448,7 @@ type MessageFeedback = "like" | "dislike" | null;
  * encrypted flag would be unreadable exactly where it matters and the skip
  * would fail open.
  */
-type MessageOrigin = "tool_result";
+type MessageOrigin = "tool_result" | "chunks_discarded";
 /**
  * Metadata for files attached to messages.
  *
@@ -3633,7 +3644,18 @@ interface StoredConversationSummary {
 interface MessageChunk {
     /** The chunk text */
     text: string;
-    /** Embedding vector for this chunk */
+    /**
+     * Embedding vector for this chunk.
+     *
+     * Read it through `decodeChunkVector` rather than passing it straight to a
+     * numeric call site. This stays `number[]` only while the writer still emits
+     * JSON arrays; it widens to `number[] | string` in the release that flips the
+     * writer to base64 float32 (sdk#862), and that widening is the whole reason
+     * that release is a major. Readers already accept both, so code written
+     * against the decoder today needs no change when it happens — code written
+     * against the raw array degrades silently, because a string handed to
+     * `Float32Array.from` yields one NaN per character rather than an error.
+     */
     vector: number[];
     /** Character offset where this chunk starts in the original message */
     startOffset: number;
@@ -7373,6 +7395,17 @@ declare function embedAllMessages(ctx: StorageOperationsContext, options: Embedd
     roles?: ("user" | "assistant")[];
     /** Minimum content length to embed (default: 30). Shorter messages are skipped. */
     minContentLength?: number;
+    /**
+     * Re-index rows the ciphertext sweep marked {@link CHUNKS_DISCARDED_ORIGIN}.
+     * Off by default: this spends the user's own embedding credits, so it belongs
+     * to an explicit user action, never to a background pass. Opens that marker
+     * only — `tool_result` rows stay excluded.
+     *
+     * Per call, not a state change: a row that re-indexes successfully keeps its
+     * marker, so a later pass that omits this flag skips it again. See
+     * {@link CHUNKS_DISCARDED_ORIGIN}.
+     */
+    reembedDiscarded?: boolean;
 }): Promise<number>;
 /**
  * Chunk and embed a single message, storing chunk embeddings in the database.
@@ -7418,6 +7451,21 @@ declare function chunkAndEmbedAllMessages(ctx: StorageOperationsContext, options
     rechunkExisting?: boolean;
     /** Minimum content length to embed (default: 30). Shorter messages are skipped. */
     minContentLength?: number;
+    /**
+     * Re-index rows the ciphertext sweep marked {@link CHUNKS_DISCARDED_ORIGIN}.
+     * Off by default: this spends the user's own embedding credits, so it belongs
+     * to an explicit user action, never to a background pass. Opens that marker
+     * only — `tool_result` rows stay excluded.
+     *
+     * `rechunkExisting` cannot substitute for it: a discarded row has neither
+     * chunks nor vector, so it never reaches that check and falls straight to the
+     * origin gate.
+     *
+     * Per call, not a state change: a row that re-indexes successfully keeps its
+     * marker, so a later pass that omits this flag skips it again. See
+     * {@link CHUNKS_DISCARDED_ORIGIN}.
+     */
+    reembedDiscarded?: boolean;
 }): Promise<number>;
 
 /**