@anuma/sdk 6.11.0 → 6.12.1

package/dist/server/index.d.ts

@@ -3013,6 +3013,22 @@ interface DeferLoadingConfig {
     enabled: boolean;
     /** Server-tool names to keep non-deferred (besides the tool-search tool), in priority order. */
     hotToolNames: readonly string[];
+    /**
+     * Server-tool names to drop entirely while defer-loading is on.
+     *
+     * Defer widens the per-turn selection to the whole catalog, which also bypasses the `excludeTools`
+     * baked into a caller's semantic filter. A filter built by {@link createServerToolsFilter} tags itself
+     * with its own exclusions and those are honoured automatically — so most callers need nothing here.
+     *
+     * Set this when the filter's tag can't be seen: a hand-written filter, a filter wrapped in a plain
+     * closure (the wrapper drops the tag), or a static-array `serverToolsFilter`. Anything excluded
+     * UNCONDITIONALLY (a tool the app replaces with its own UI, a capability the model already has
+     * natively) belongs in one of the two places, or defer quietly hands it back to the model.
+     *
+     * This list is UNIONED with the filter's own tag, never a replacement for it — so adding one name
+     * here cannot re-admit the tools the filter already excludes.
+     */
+    excludeTools?: readonly string[];
 }
 /**
  * Result of tool matching with similarity score
@@ -3192,7 +3208,7 @@ interface CreateServerToolsFilterOptions {
  * });
  * ```
  */
-declare function createServerToolsFilter(options?: CreateServerToolsFilterOptions): (embeddings: number[] | number[][], tools: ServerTool[]) => string[];
+declare function createServerToolsFilter(options?: CreateServerToolsFilterOptions): ServerToolsFilterFunction;
 /**
  * Default exclusions baked into `defaultServerToolsFilter`.
  *
@@ -3258,13 +3274,21 @@ declare const SERVER_TOOL_DEPENDENCY_SETS: ToolSet[];
  * If you need to customize (extra excludes, different limits, opt into
  * tool-set expansion), call `createServerToolsFilter` directly.
  */
-declare const defaultServerToolsFilter: (embeddings: number[] | number[][], tools: ServerTool[]) => string[];
+declare const defaultServerToolsFilter: ServerToolsFilterFunction;
 /**
  * Type for a server-tools filter — a function that takes prompt embeddings
  * and the full server tool catalog and returns the names of tools to keep.
  * Matches `useChatStorage`'s `serverTools` callback signature.
  */
-type ServerToolsFilterFunction = (embeddings: number[] | number[][], tools: ServerTool[]) => string[];
+type ServerToolsFilterFunction = ((embeddings: number[] | number[][], tools: ServerTool[]) => string[]) & {
+    /**
+     * The unconditional exclusions this filter applies, exposed by
+     * {@link createServerToolsFilter} so defer-loading can honour them without the caller
+     * repeating the list — see {@link resolveDeferredServerTools}. Absent on a hand-written
+     * filter, or on one wrapped in a plain closure.
+     */
+    readonly excludeTools?: readonly string[];
+};
 /**
  * Options for `selectServerToolsForPrompt`.
  */
@@ -3292,9 +3316,10 @@ interface SelectServerToolsForPromptOptions {
      */
     cache?: ToolsCacheBackend;
     /**
-     * Phase 3 defer-loading. When `enabled`, this helper returns the FULL catalog (skipping semantic/
-     * static filtering) to mirror useChatStorage's responses send path, which swaps in the full catalog
-     * for mergeTools + tool-search. Omit/disabled → today's filtered selection.
+     * Phase 3 defer-loading. When `enabled`, this helper skips SEMANTIC filtering to mirror
+     * useChatStorage's responses send path, which hands the catalog to mergeTools + tool-search. The
+     * caller's unconditional constraints still apply — an explicit static array, and exclusions (see
+     * {@link resolveDeferredServerTools}). Omit/disabled → today's filtered selection.
      */
     deferLoading?: DeferLoadingConfig;
 }
@@ -4698,6 +4723,51 @@ interface EntityOperationsContext {
     allowUnscopedRows?: 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_dedup" | "memory_decay" | "memory_injection_check" | "memory_verify_support" | "memory_autosort" | "memory_topic" | "memory_consolidate" | "memory_decompose" | "memory_profile_synth" | "memory_graph" | "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";
+
 /**
  * Shared portal LLM call for the memory pipeline.
  *
@@ -8799,6 +8869,28 @@ interface ReflectOptions extends RecallOptions, PortalLlmAuth {
     maxTokens?: number;
     /** Override the grounding system prompt. */
     systemPrompt?: string;
+    /**
+     * Extra caller instruction to carry on the USER turn, between the question and
+     * the evidence block (see the `userMessage` assembly below). This is the slot a
+     * background caller uses to keep its per-request data OUT of the system message
+     * without colliding with the numbered evidence list — profile-facet synthesis
+     * puts its section label, guidance and response-field hint here so its system
+     * half can stay fixed and server-ownable.
+     */
+    userInstructions?: string;
+    /**
+     * Class-B task name for the `X-Anuma-Task-Type` header, or nothing.
+     *
+     * Deliberately OPTIONAL and unset by default. reflect() also answers the user's
+     * OWN question, and that traffic is chat, not an internal flow — declaring a
+     * task type unconditionally here would put an internal-flow name on real
+     * conversation, which is the same boundary `INTERNAL_FLOW_MARKER` draws
+     * (reflect is deliberately unmarked; its background caller marks its own
+     * prompt — see ../internalFlowMarker.ts). So the name is per call: only a
+     * caller with ONE fixed purpose passes
+     * one, and today that is profile-facet synthesis (`memory_profile_synth`).
+     */
+    taskType?: TaskType;
     /** Endpoint for the answer LLM. */
     baseUrl?: string;
     /** Override fetch (for tests). */

package/dist/server/index.mjs

@@ -141,7 +141,7 @@ import {
   updateVaultMemoryOp,
   upsertMessageOp,
   vaultMemoryToStored
-} from "./chunk-BKB6AT7Z.mjs";
+} from "./chunk-7YTEE6JT.mjs";
 import {
   createStreamingDeAnonymizer,
   notifyConsolidationFallback,
@@ -6963,11 +6963,13 @@ Respond with ONLY a single JSON object conforming to this JSON Schema, with no p
 ${JSON.stringify(
     options.responseSchema
   )}` : basePrompt;
-  const userMessage = `Question:
-${trimmed}
-
-Memories (use only these as evidence):
-${evidence}`;
+  const userMessage = [
+    `Question:
+${trimmed}`,
+    ...options.userInstructions ? [options.userInstructions] : [],
+    `Memories (use only these as evidence):
+${evidence}`
+  ].join("\n\n");
   const maxTokens = options.maxTokens ?? DEFAULT_MAX_TOKENS;
   const fetchImpl = options.fetchFn ?? fetch;
   const log = getLogger();
@@ -6981,7 +6983,13 @@ ${evidence}`;
   try {
     response = await fetchImpl(`${baseUrl}/api/v1/chat/completions`, {
       method: "POST",
-      headers: { ...authHeaders, "Content-Type": "application/json" },
+      // No task type unless the caller named one — an undeclared task adds no
+      // header at all, which is what keeps user-facing reflect() unlabelled.
+      headers: {
+        ...authHeaders,
+        ...taskTypeHeader(options.taskType),
+        "Content-Type": "application/json"
+      },
       body: JSON.stringify({
         model,
         // Modern OpenAI field; the portal reads only `max_completion_tokens`
@@ -7421,10 +7429,15 @@ async function synthesizeFacet(facet, ctx, options, prior) {
     limit,
     types: ["fact"],
     maxTokens: DEFAULT_FACET_MAX_TOKENS,
-    // 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)),
+    // Profile-facet synthesis is a background op, so it marks its own prompt AND
+    // names its own task. reflect() itself must do neither — it also serves the
+    // user's own questions, and that traffic is chat, not an internal flow. See
+    // ../internalFlowMarker and ReflectOptions.taskType.
+    taskType: "memory_profile_synth",
+    systemPrompt: withInternalFlowMarker(FACET_SYSTEM_PROMPT),
+    // The facet's label/guidance/response-fields, on the user turn — what keeps
+    // the system half fixed and therefore server-ownable.
+    userInstructions: buildFacetUserInstructions(facet),
     responseSchema: facetResponseSchema(facet.key),
     memories
   });
@@ -7482,19 +7495,22 @@ function fallbackSection(facet, prior) {
     stale: true
   };
 }
-function buildFacetSystemPrompt(facet) {
-  const structuredField = FACET_STRUCTURED_RESPONSE_HINT[facet.key] ?? "";
-  return `You are writing the "${facet.label}" section of a person's shareable profile, using their private memories (supplied as evidence) as the only source of truth.
-
-Task for this section:
-${facet.guidance}
+var FACET_SYSTEM_PROMPT = `You are writing one section of a person's shareable profile, using their private memories (supplied as evidence) as the only source of truth. The user turn names the section, states the task for it, and gives the exact JSON shape to respond in.
 
 Rules:
 - Ground every claim in the supplied memories \u2014 never invent, infer beyond, or embellish what they support.
 - If the memories don't cover this section, return an empty summary with hasEvidence=false. Do not pad or guess.
 - Write in third person about the person, in a natural voice suitable for a public profile (no "I"/"you", no name repetition).
 - Be concise and specific; no preamble, hedging, or meta-commentary.
-- Respond as JSON: { "summary": <the section text, or "">, "hasEvidence": <true|false>${structuredField} }.`;
+- Respond as JSON in exactly the shape the user turn states, with no extra fields.`;
+function buildFacetUserInstructions(facet) {
+  const structuredField = FACET_STRUCTURED_RESPONSE_HINT[facet.key] ?? "";
+  return `Section: "${facet.label}"
+
+Task for this section:
+${facet.guidance}
+
+Respond as JSON: { "summary": <the section text, or "">, "hasEvidence": <true|false>${structuredField} }.`;
 }
 function extractFacetText(structured) {
   if (structured && typeof structured === "object") {
@@ -8286,6 +8302,15 @@ function filterServerTools(serverTools, includeNames) {
   const includeSet = new Set(includeNames);
   return serverTools.filter((tool) => includeSet.has(tool.name));
 }
+function resolveDeferredServerTools(allServerTools, serverToolsFilter, config) {
+  const allowed = serverToolsFilter === void 0 || typeof serverToolsFilter === "function" ? allServerTools : filterServerTools(allServerTools, [...serverToolsFilter]);
+  const excluded = new Set(config.excludeTools ?? []);
+  if (typeof serverToolsFilter === "function") {
+    for (const name of serverToolsFilter.excludeTools ?? []) excluded.add(name);
+  }
+  if (excluded.size === 0) return allowed;
+  return allowed.filter((tool) => !excluded.has(tool.name));
+}
 var DEFAULT_TOOL_MATCH_OPTIONS = {
   limit: 5,
   minSimilarity: 0.3,
@@ -8562,7 +8587,7 @@ function createServerToolsFilter(options = {}) {
   const exclude = new Set(options.excludeTools ?? []);
   const sets = options.toolSets ?? [];
   const matchOpts = options.matchOptions;
-  return (embeddings, tools) => {
+  const filter = (embeddings, tools) => {
     const matches = findMatchingTools(embeddings, tools, matchOpts);
     if (matches.length === 0) return [];
     const matchedNames = new Set(matches.map((m) => m.tool.name));
@@ -8580,6 +8605,10 @@ function createServerToolsFilter(options = {}) {
     for (const name of exclude) finalNames.delete(name);
     return [...finalNames];
   };
+  return Object.defineProperty(filter, "excludeTools", {
+    value: Object.freeze([...exclude]),
+    enumerable: false
+  });
 }
 var DEFAULT_EXCLUDED_SERVER_TOOLS = [
   "AnumaVisionMCP-anuma_analyze_image",
@@ -8661,7 +8690,8 @@ async function selectServerToolsForPrompt(options) {
     return [];
   }
   if (allServerTools.length === 0) return [];
-  if (deferLoading?.enabled) return allServerTools;
+  if (deferLoading?.enabled)
+    return resolveDeferredServerTools(allServerTools, serverToolsFilter, deferLoading);
   if (typeof serverToolsFilter === "function") {
     if (prompt.length < MIN_CONTENT_LENGTH_FOR_TOOLS) return [];
     let promptEmbedding;

package/dist/server/{recall-HVIEAH7J.mjs → recall-HOVWMT4Q.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-BKB6AT7Z.mjs";
+} from "./chunk-7YTEE6JT.mjs";
 import "./chunk-KF5OU4IN.mjs";
 import "./chunk-3RPGXXPZ.mjs";
 export {

package/dist/tools/selection/index.cjs

@@ -86,6 +86,7 @@ __export(index_exports, {
   computeToolGuidance: () => computeToolGuidance,
   createServerToolsFilter: () => createServerToolsFilter,
   defaultServerToolsFilter: () => defaultServerToolsFilter,
+  deferFormattingConfig: () => deferFormattingConfig,
   deriveActiveToolSets: () => deriveActiveToolSets,
   expandToolSetsAdditive: () => expandToolSetsAdditive,
   filterAssembledClientTools: () => filterAssembledClientTools,
@@ -105,6 +106,7 @@ __export(index_exports, {
   mergeTools: () => mergeTools,
   resolveCouncilPlan: () => resolveCouncilPlan,
   resolveCouncilServerTools: () => resolveCouncilServerTools,
+  resolveDeferredServerTools: () => resolveDeferredServerTools,
   resolvePlan: () => resolvePlan,
   resolveServerTools: () => resolveServerTools,
   resolveToolChoice: () => resolveToolChoice,
@@ -1658,6 +1660,20 @@ function clientToolToCompletionsFormat(tool) {
 }
 var TOOL_SEARCH_TOOL_TYPE = "tool_search_tool_regex_20251119";
 var TOOL_SEARCH_TOOL_NAME = "tool_search_tool_regex";
+function resolveDeferredServerTools(allServerTools, serverToolsFilter, config) {
+  const allowed = serverToolsFilter === void 0 || typeof serverToolsFilter === "function" ? allServerTools : filterServerTools(allServerTools, [...serverToolsFilter]);
+  const excluded = new Set(config.excludeTools ?? []);
+  if (typeof serverToolsFilter === "function") {
+    for (const name of serverToolsFilter.excludeTools ?? []) excluded.add(name);
+  }
+  if (excluded.size === 0) return allowed;
+  return allowed.filter((tool) => !excluded.has(tool.name));
+}
+function deferFormattingConfig(serverToolsFilter, config) {
+  if (!config?.enabled) return config;
+  const isExplicitList = serverToolsFilter !== void 0 && typeof serverToolsFilter !== "function";
+  return isExplicitList ? void 0 : config;
+}
 function byNameAscending(a, b) {
   if (a.name < b.name) return -1;
   if (a.name > b.name) return 1;
@@ -2013,7 +2029,7 @@ function createServerToolsFilter(options = {}) {
   const exclude = new Set(options.excludeTools ?? []);
   const sets = options.toolSets ?? [];
   const matchOpts = options.matchOptions;
-  return (embeddings, tools) => {
+  const filter = (embeddings, tools) => {
     const matches = findMatchingTools(embeddings, tools, matchOpts);
     if (matches.length === 0) return [];
     const matchedNames = new Set(matches.map((m) => m.tool.name));
@@ -2031,6 +2047,10 @@ function createServerToolsFilter(options = {}) {
     for (const name of exclude) finalNames.delete(name);
     return [...finalNames];
   };
+  return Object.defineProperty(filter, "excludeTools", {
+    value: Object.freeze([...exclude]),
+    enumerable: false
+  });
 }
 var DEFAULT_EXCLUDED_SERVER_TOOLS = [
   "AnumaVisionMCP-anuma_analyze_image",
@@ -2112,7 +2132,8 @@ async function selectServerToolsForPrompt(options) {
     return [];
   }
   if (allServerTools.length === 0) return [];
-  if (deferLoading?.enabled) return allServerTools;
+  if (deferLoading?.enabled)
+    return resolveDeferredServerTools(allServerTools, serverToolsFilter, deferLoading);
   if (typeof serverToolsFilter === "function") {
     if (prompt.length < MIN_CONTENT_LENGTH_FOR_TOOLS) return [];
     let promptEmbedding;
@@ -2738,6 +2759,7 @@ async function selectCouncilClientTools(input, options) {
   computeToolGuidance,
   createServerToolsFilter,
   defaultServerToolsFilter,
+  deferFormattingConfig,
   deriveActiveToolSets,
   expandToolSetsAdditive,
   filterAssembledClientTools,
@@ -2757,6 +2779,7 @@ async function selectCouncilClientTools(input, options) {
   mergeTools,
   resolveCouncilPlan,
   resolveCouncilServerTools,
+  resolveDeferredServerTools,
   resolvePlan,
   resolveServerTools,
   resolveToolChoice,

package/dist/tools/selection/index.d.mts

@@ -272,7 +272,63 @@ interface DeferLoadingConfig {
     enabled: boolean;
     /** Server-tool names to keep non-deferred (besides the tool-search tool), in priority order. */
     hotToolNames: readonly string[];
+    /**
+     * Server-tool names to drop entirely while defer-loading is on.
+     *
+     * Defer widens the per-turn selection to the whole catalog, which also bypasses the `excludeTools`
+     * baked into a caller's semantic filter. A filter built by {@link createServerToolsFilter} tags itself
+     * with its own exclusions and those are honoured automatically — so most callers need nothing here.
+     *
+     * Set this when the filter's tag can't be seen: a hand-written filter, a filter wrapped in a plain
+     * closure (the wrapper drops the tag), or a static-array `serverToolsFilter`. Anything excluded
+     * UNCONDITIONALLY (a tool the app replaces with its own UI, a capability the model already has
+     * natively) belongs in one of the two places, or defer quietly hands it back to the model.
+     *
+     * This list is UNIONED with the filter's own tag, never a replacement for it — so adding one name
+     * here cannot re-admit the tools the filter already excludes.
+     */
+    excludeTools?: readonly string[];
 }
+/**
+ * The server-tool set defer-loading should format for one turn.
+ *
+ * Defer changes how tools are SENT (full definitions + `defer_loading` + a tool-search tool, for a
+ * byte-stable cacheable prefix). It must not change WHICH tools the caller allowed, so two things
+ * survive that a plain "just use the whole catalog" discards:
+ *
+ * 1. **an explicit static array** — `serverToolsFilter: ['x']` is a deliberate scoping, and callers
+ *    commonly pair it with `tool_choice:'required'`. Replacing it with the catalog turns "you must
+ *    call x" into "you must call something, here is everything". Defer still applies, within the array.
+ * 2. **unconditional exclusions** — see {@link DeferLoadingConfig.excludeTools}.
+ *
+ * A filter FUNCTION is deliberately NOT consulted: it is per-prompt semantic narrowing, and skipping it
+ * is the whole point of defer (the model reaches everything else through tool-search). Only the
+ * caller's unconditional constraints survive.
+ */
+declare function resolveDeferredServerTools(allServerTools: ServerTool[], serverToolsFilter: readonly string[] | ServerToolsFilterFunction | undefined, config: DeferLoadingConfig): ServerTool[];
+/**
+ * The defer config {@link mergeTools} should FORMAT this request with — `undefined` when defer
+ * formatting must be skipped even though defer-loading is enabled.
+ *
+ * Scoping the selection to a caller's explicit static array (see {@link resolveDeferredServerTools})
+ * is only half the job. `formatServerToolsWithDefer` still marks every non-hot tool `defer_loading`
+ * and prepends the tool-search tool, so for a one-tool array whose tool isn't hot — every creation
+ * mode: video, music, sfx, image, slides — the search tool becomes the ONLY directly callable entry
+ * in the request. Callers pair those arrays with `tool_choice:'required'`, so "you must call the video
+ * generator" silently degrades to "you must call tool-search", and a caller scanning tool-call events
+ * for the generator's name never sees it.
+ *
+ * An explicit array is a closed set the caller already narrowed: there is nothing left to discover, so
+ * defer has no work to do and its formatting is pure downside. Sending those tools normally is exactly
+ * today's (defer-off) behavior for that request.
+ *
+ * Note this is NOT the same as treating the array as `hotToolNames`: hot tools are non-deferred, but
+ * `formatServerToolsWithDefer` prepends the tool-search tool unconditionally, so a `required` turn
+ * could still satisfy the constraint by calling it. Skipping the formatting is what actually fixes it.
+ *
+ * The catalog path (a filter function, or no filter) is untouched — full tool-search + hot + deferred.
+ */
+declare function deferFormattingConfig(serverToolsFilter: readonly string[] | ServerToolsFilterFunction | undefined, config: DeferLoadingConfig | undefined): DeferLoadingConfig | undefined;
 declare function mergeTools(serverTools: ServerTool[], clientTools: Array<LlmapiChatCompletionTool | ToolConfig> | undefined, apiType?: "responses" | "completions", deferConfig?: DeferLoadingConfig): Array<Record<string, unknown>>;
 /**
  * Result of tool matching with similarity score
@@ -509,7 +565,7 @@ interface CreateServerToolsFilterOptions {
  * });
  * ```
  */
-declare function createServerToolsFilter(options?: CreateServerToolsFilterOptions): (embeddings: number[] | number[][], tools: ServerTool[]) => string[];
+declare function createServerToolsFilter(options?: CreateServerToolsFilterOptions): ServerToolsFilterFunction;
 /**
  * Default exclusions baked into `defaultServerToolsFilter`.
  *
@@ -575,13 +631,21 @@ declare const SERVER_TOOL_DEPENDENCY_SETS: ToolSet[];
  * If you need to customize (extra excludes, different limits, opt into
  * tool-set expansion), call `createServerToolsFilter` directly.
  */
-declare const defaultServerToolsFilter: (embeddings: number[] | number[][], tools: ServerTool[]) => string[];
+declare const defaultServerToolsFilter: ServerToolsFilterFunction;
 /**
  * Type for a server-tools filter — a function that takes prompt embeddings
  * and the full server tool catalog and returns the names of tools to keep.
  * Matches `useChatStorage`'s `serverTools` callback signature.
  */
-type ServerToolsFilterFunction = (embeddings: number[] | number[][], tools: ServerTool[]) => string[];
+type ServerToolsFilterFunction = ((embeddings: number[] | number[][], tools: ServerTool[]) => string[]) & {
+    /**
+     * The unconditional exclusions this filter applies, exposed by
+     * {@link createServerToolsFilter} so defer-loading can honour them without the caller
+     * repeating the list — see {@link resolveDeferredServerTools}. Absent on a hand-written
+     * filter, or on one wrapped in a plain closure.
+     */
+    readonly excludeTools?: readonly string[];
+};
 /**
  * Options for `selectServerToolsForPrompt`.
  */
@@ -609,9 +673,10 @@ interface SelectServerToolsForPromptOptions {
      */
     cache?: ToolsCacheBackend;
     /**
-     * Phase 3 defer-loading. When `enabled`, this helper returns the FULL catalog (skipping semantic/
-     * static filtering) to mirror useChatStorage's responses send path, which swaps in the full catalog
-     * for mergeTools + tool-search. Omit/disabled → today's filtered selection.
+     * Phase 3 defer-loading. When `enabled`, this helper skips SEMANTIC filtering to mirror
+     * useChatStorage's responses send path, which hands the catalog to mergeTools + tool-search. The
+     * caller's unconditional constraints still apply — an explicit static array, and exclusions (see
+     * {@link resolveDeferredServerTools}). Omit/disabled → today's filtered selection.
      */
     deferLoading?: DeferLoadingConfig;
 }
@@ -1377,4 +1442,4 @@ declare function resolveServerTools(descriptor: ToolIntentDescriptor, catalog: S
  */
 declare function isCoercingIntent(creation: CreationIntent, catalog: ServerToolCatalog): boolean;
 
-export { type AssembledToolsFilterFn, BUILT_IN_TOOL_SETS, CLIENT_TOOLS_MIN_SIMILARITY, CLIENT_TOOLS_RELEVANCE_RATIO, CREATION_INTENT_CLIENT_FACTORIES, CREATION_INTENT_CLIENT_FILTER, CREATION_INTENT_TOOL_SET, type CachedServerTools, type ClientFactoryBuilder, type ClientFactoryKey, type ClientToolsFilterMode, type ConnectorGuidance, type ConnectorGuidanceInput, type CouncilClientSelectionOptions, type CouncilClientToolsInput, type CreateServerToolsFilterOptions, type CreationIntent, DEFAULT_CACHE_EXPIRATION_MS, DEFAULT_EXCLUDED_SERVER_TOOLS, DEFAULT_SERVER_TOOLS_MATCH_OPTIONS, type DeferLoadingConfig, FULL_GENERATIVE_CLIENT_FACTORIES, MAX_CLIENT_TOOLS_AFTER_FILTER, MEMORY_SAVE_TOOL_NAME, MIN_CONTENT_LENGTH_FOR_TOOLS, type ParsedServerToolsResponse, type PostFilterKey, type ResolvePlanContext, SERVER_TOOL_DEPENDENCY_SETS, type SelectServerSideToolsOptions, type SelectServerToolsForPromptOptions, type SelectionLogger, type ServerTool, type ServerToolCatalog, type ServerToolCatalogEntry, type ServerToolsFilter, type ServerToolsFilterContext, type ServerToolsFilterFunction, type ServerToolsOptions, type ServerToolsResponse, TOOL_CATALOG, type ThinkingModeHint, type ToolActivationEvent, type ToolChoice, type ToolConfig, type ToolExecutor, type ToolIntentDescriptor, type ToolLane, type ToolMatchOptions, type ToolMatchResult, type ToolPlanSpec, type ToolSelectionAdapters, type ToolSet, type ToolsCacheBackend, activatedToolSetNames, applyToolSets, assembleClientTools, autoFilterClientTools, buildConnectorGuidance, buildDeniedToolsRider, clearServerToolsCache, composeCouncilClientTools, computeToolGuidance, createServerToolsFilter, defaultServerToolsFilter, deriveActiveToolSets, expandToolSetsAdditive, filterAssembledClientTools, filterServerTools, findMatchingTools, getCachedServerTools, getCatalogEntry, getMaxToolRounds, getServerTools, getThinkingMode, getToolDescription, getToolName, getToolsChecksum, isCoercingIntent, localStorageToolsCache, mergeActiveToolSets, mergeTools, resolveCouncilPlan, resolveCouncilServerTools, resolvePlan, resolveServerTools, resolveToolChoice, scoreTools, selectCouncilClientTools, selectServerSideTools, selectServerToolsForPrompt, shouldRefreshTools, stripCoercedMemoryWrite, toolSetSystemPrompts };
+export { type AssembledToolsFilterFn, BUILT_IN_TOOL_SETS, CLIENT_TOOLS_MIN_SIMILARITY, CLIENT_TOOLS_RELEVANCE_RATIO, CREATION_INTENT_CLIENT_FACTORIES, CREATION_INTENT_CLIENT_FILTER, CREATION_INTENT_TOOL_SET, type CachedServerTools, type ClientFactoryBuilder, type ClientFactoryKey, type ClientToolsFilterMode, type ConnectorGuidance, type ConnectorGuidanceInput, type CouncilClientSelectionOptions, type CouncilClientToolsInput, type CreateServerToolsFilterOptions, type CreationIntent, DEFAULT_CACHE_EXPIRATION_MS, DEFAULT_EXCLUDED_SERVER_TOOLS, DEFAULT_SERVER_TOOLS_MATCH_OPTIONS, type DeferLoadingConfig, FULL_GENERATIVE_CLIENT_FACTORIES, MAX_CLIENT_TOOLS_AFTER_FILTER, MEMORY_SAVE_TOOL_NAME, MIN_CONTENT_LENGTH_FOR_TOOLS, type ParsedServerToolsResponse, type PostFilterKey, type ResolvePlanContext, SERVER_TOOL_DEPENDENCY_SETS, type SelectServerSideToolsOptions, type SelectServerToolsForPromptOptions, type SelectionLogger, type ServerTool, type ServerToolCatalog, type ServerToolCatalogEntry, type ServerToolsFilter, type ServerToolsFilterContext, type ServerToolsFilterFunction, type ServerToolsOptions, type ServerToolsResponse, TOOL_CATALOG, type ThinkingModeHint, type ToolActivationEvent, type ToolChoice, type ToolConfig, type ToolExecutor, type ToolIntentDescriptor, type ToolLane, type ToolMatchOptions, type ToolMatchResult, type ToolPlanSpec, type ToolSelectionAdapters, type ToolSet, type ToolsCacheBackend, activatedToolSetNames, applyToolSets, assembleClientTools, autoFilterClientTools, buildConnectorGuidance, buildDeniedToolsRider, clearServerToolsCache, composeCouncilClientTools, computeToolGuidance, createServerToolsFilter, defaultServerToolsFilter, deferFormattingConfig, deriveActiveToolSets, expandToolSetsAdditive, filterAssembledClientTools, filterServerTools, findMatchingTools, getCachedServerTools, getCatalogEntry, getMaxToolRounds, getServerTools, getThinkingMode, getToolDescription, getToolName, getToolsChecksum, isCoercingIntent, localStorageToolsCache, mergeActiveToolSets, mergeTools, resolveCouncilPlan, resolveCouncilServerTools, resolveDeferredServerTools, resolvePlan, resolveServerTools, resolveToolChoice, scoreTools, selectCouncilClientTools, selectServerSideTools, selectServerToolsForPrompt, shouldRefreshTools, stripCoercedMemoryWrite, toolSetSystemPrompts };

package/dist/tools/selection/index.d.ts

@@ -272,7 +272,63 @@ interface DeferLoadingConfig {
     enabled: boolean;
     /** Server-tool names to keep non-deferred (besides the tool-search tool), in priority order. */
     hotToolNames: readonly string[];
+    /**
+     * Server-tool names to drop entirely while defer-loading is on.
+     *
+     * Defer widens the per-turn selection to the whole catalog, which also bypasses the `excludeTools`
+     * baked into a caller's semantic filter. A filter built by {@link createServerToolsFilter} tags itself
+     * with its own exclusions and those are honoured automatically — so most callers need nothing here.
+     *
+     * Set this when the filter's tag can't be seen: a hand-written filter, a filter wrapped in a plain
+     * closure (the wrapper drops the tag), or a static-array `serverToolsFilter`. Anything excluded
+     * UNCONDITIONALLY (a tool the app replaces with its own UI, a capability the model already has
+     * natively) belongs in one of the two places, or defer quietly hands it back to the model.
+     *
+     * This list is UNIONED with the filter's own tag, never a replacement for it — so adding one name
+     * here cannot re-admit the tools the filter already excludes.
+     */
+    excludeTools?: readonly string[];
 }
+/**
+ * The server-tool set defer-loading should format for one turn.
+ *
+ * Defer changes how tools are SENT (full definitions + `defer_loading` + a tool-search tool, for a
+ * byte-stable cacheable prefix). It must not change WHICH tools the caller allowed, so two things
+ * survive that a plain "just use the whole catalog" discards:
+ *
+ * 1. **an explicit static array** — `serverToolsFilter: ['x']` is a deliberate scoping, and callers
+ *    commonly pair it with `tool_choice:'required'`. Replacing it with the catalog turns "you must
+ *    call x" into "you must call something, here is everything". Defer still applies, within the array.
+ * 2. **unconditional exclusions** — see {@link DeferLoadingConfig.excludeTools}.
+ *
+ * A filter FUNCTION is deliberately NOT consulted: it is per-prompt semantic narrowing, and skipping it
+ * is the whole point of defer (the model reaches everything else through tool-search). Only the
+ * caller's unconditional constraints survive.
+ */
+declare function resolveDeferredServerTools(allServerTools: ServerTool[], serverToolsFilter: readonly string[] | ServerToolsFilterFunction | undefined, config: DeferLoadingConfig): ServerTool[];
+/**
+ * The defer config {@link mergeTools} should FORMAT this request with — `undefined` when defer
+ * formatting must be skipped even though defer-loading is enabled.
+ *
+ * Scoping the selection to a caller's explicit static array (see {@link resolveDeferredServerTools})
+ * is only half the job. `formatServerToolsWithDefer` still marks every non-hot tool `defer_loading`
+ * and prepends the tool-search tool, so for a one-tool array whose tool isn't hot — every creation
+ * mode: video, music, sfx, image, slides — the search tool becomes the ONLY directly callable entry
+ * in the request. Callers pair those arrays with `tool_choice:'required'`, so "you must call the video
+ * generator" silently degrades to "you must call tool-search", and a caller scanning tool-call events
+ * for the generator's name never sees it.
+ *
+ * An explicit array is a closed set the caller already narrowed: there is nothing left to discover, so
+ * defer has no work to do and its formatting is pure downside. Sending those tools normally is exactly
+ * today's (defer-off) behavior for that request.
+ *
+ * Note this is NOT the same as treating the array as `hotToolNames`: hot tools are non-deferred, but
+ * `formatServerToolsWithDefer` prepends the tool-search tool unconditionally, so a `required` turn
+ * could still satisfy the constraint by calling it. Skipping the formatting is what actually fixes it.
+ *
+ * The catalog path (a filter function, or no filter) is untouched — full tool-search + hot + deferred.
+ */
+declare function deferFormattingConfig(serverToolsFilter: readonly string[] | ServerToolsFilterFunction | undefined, config: DeferLoadingConfig | undefined): DeferLoadingConfig | undefined;
 declare function mergeTools(serverTools: ServerTool[], clientTools: Array<LlmapiChatCompletionTool | ToolConfig> | undefined, apiType?: "responses" | "completions", deferConfig?: DeferLoadingConfig): Array<Record<string, unknown>>;
 /**
  * Result of tool matching with similarity score
@@ -509,7 +565,7 @@ interface CreateServerToolsFilterOptions {
  * });
  * ```
  */
-declare function createServerToolsFilter(options?: CreateServerToolsFilterOptions): (embeddings: number[] | number[][], tools: ServerTool[]) => string[];
+declare function createServerToolsFilter(options?: CreateServerToolsFilterOptions): ServerToolsFilterFunction;
 /**
  * Default exclusions baked into `defaultServerToolsFilter`.
  *
@@ -575,13 +631,21 @@ declare const SERVER_TOOL_DEPENDENCY_SETS: ToolSet[];
  * If you need to customize (extra excludes, different limits, opt into
  * tool-set expansion), call `createServerToolsFilter` directly.
  */
-declare const defaultServerToolsFilter: (embeddings: number[] | number[][], tools: ServerTool[]) => string[];
+declare const defaultServerToolsFilter: ServerToolsFilterFunction;
 /**
  * Type for a server-tools filter — a function that takes prompt embeddings
  * and the full server tool catalog and returns the names of tools to keep.
  * Matches `useChatStorage`'s `serverTools` callback signature.
  */
-type ServerToolsFilterFunction = (embeddings: number[] | number[][], tools: ServerTool[]) => string[];
+type ServerToolsFilterFunction = ((embeddings: number[] | number[][], tools: ServerTool[]) => string[]) & {
+    /**
+     * The unconditional exclusions this filter applies, exposed by
+     * {@link createServerToolsFilter} so defer-loading can honour them without the caller
+     * repeating the list — see {@link resolveDeferredServerTools}. Absent on a hand-written
+     * filter, or on one wrapped in a plain closure.
+     */
+    readonly excludeTools?: readonly string[];
+};
 /**
  * Options for `selectServerToolsForPrompt`.
  */
@@ -609,9 +673,10 @@ interface SelectServerToolsForPromptOptions {
      */
     cache?: ToolsCacheBackend;
     /**
-     * Phase 3 defer-loading. When `enabled`, this helper returns the FULL catalog (skipping semantic/
-     * static filtering) to mirror useChatStorage's responses send path, which swaps in the full catalog
-     * for mergeTools + tool-search. Omit/disabled → today's filtered selection.
+     * Phase 3 defer-loading. When `enabled`, this helper skips SEMANTIC filtering to mirror
+     * useChatStorage's responses send path, which hands the catalog to mergeTools + tool-search. The
+     * caller's unconditional constraints still apply — an explicit static array, and exclusions (see
+     * {@link resolveDeferredServerTools}). Omit/disabled → today's filtered selection.
      */
     deferLoading?: DeferLoadingConfig;
 }
@@ -1377,4 +1442,4 @@ declare function resolveServerTools(descriptor: ToolIntentDescriptor, catalog: S
  */
 declare function isCoercingIntent(creation: CreationIntent, catalog: ServerToolCatalog): boolean;
 
-export { type AssembledToolsFilterFn, BUILT_IN_TOOL_SETS, CLIENT_TOOLS_MIN_SIMILARITY, CLIENT_TOOLS_RELEVANCE_RATIO, CREATION_INTENT_CLIENT_FACTORIES, CREATION_INTENT_CLIENT_FILTER, CREATION_INTENT_TOOL_SET, type CachedServerTools, type ClientFactoryBuilder, type ClientFactoryKey, type ClientToolsFilterMode, type ConnectorGuidance, type ConnectorGuidanceInput, type CouncilClientSelectionOptions, type CouncilClientToolsInput, type CreateServerToolsFilterOptions, type CreationIntent, DEFAULT_CACHE_EXPIRATION_MS, DEFAULT_EXCLUDED_SERVER_TOOLS, DEFAULT_SERVER_TOOLS_MATCH_OPTIONS, type DeferLoadingConfig, FULL_GENERATIVE_CLIENT_FACTORIES, MAX_CLIENT_TOOLS_AFTER_FILTER, MEMORY_SAVE_TOOL_NAME, MIN_CONTENT_LENGTH_FOR_TOOLS, type ParsedServerToolsResponse, type PostFilterKey, type ResolvePlanContext, SERVER_TOOL_DEPENDENCY_SETS, type SelectServerSideToolsOptions, type SelectServerToolsForPromptOptions, type SelectionLogger, type ServerTool, type ServerToolCatalog, type ServerToolCatalogEntry, type ServerToolsFilter, type ServerToolsFilterContext, type ServerToolsFilterFunction, type ServerToolsOptions, type ServerToolsResponse, TOOL_CATALOG, type ThinkingModeHint, type ToolActivationEvent, type ToolChoice, type ToolConfig, type ToolExecutor, type ToolIntentDescriptor, type ToolLane, type ToolMatchOptions, type ToolMatchResult, type ToolPlanSpec, type ToolSelectionAdapters, type ToolSet, type ToolsCacheBackend, activatedToolSetNames, applyToolSets, assembleClientTools, autoFilterClientTools, buildConnectorGuidance, buildDeniedToolsRider, clearServerToolsCache, composeCouncilClientTools, computeToolGuidance, createServerToolsFilter, defaultServerToolsFilter, deriveActiveToolSets, expandToolSetsAdditive, filterAssembledClientTools, filterServerTools, findMatchingTools, getCachedServerTools, getCatalogEntry, getMaxToolRounds, getServerTools, getThinkingMode, getToolDescription, getToolName, getToolsChecksum, isCoercingIntent, localStorageToolsCache, mergeActiveToolSets, mergeTools, resolveCouncilPlan, resolveCouncilServerTools, resolvePlan, resolveServerTools, resolveToolChoice, scoreTools, selectCouncilClientTools, selectServerSideTools, selectServerToolsForPrompt, shouldRefreshTools, stripCoercedMemoryWrite, toolSetSystemPrompts };
+export { type AssembledToolsFilterFn, BUILT_IN_TOOL_SETS, CLIENT_TOOLS_MIN_SIMILARITY, CLIENT_TOOLS_RELEVANCE_RATIO, CREATION_INTENT_CLIENT_FACTORIES, CREATION_INTENT_CLIENT_FILTER, CREATION_INTENT_TOOL_SET, type CachedServerTools, type ClientFactoryBuilder, type ClientFactoryKey, type ClientToolsFilterMode, type ConnectorGuidance, type ConnectorGuidanceInput, type CouncilClientSelectionOptions, type CouncilClientToolsInput, type CreateServerToolsFilterOptions, type CreationIntent, DEFAULT_CACHE_EXPIRATION_MS, DEFAULT_EXCLUDED_SERVER_TOOLS, DEFAULT_SERVER_TOOLS_MATCH_OPTIONS, type DeferLoadingConfig, FULL_GENERATIVE_CLIENT_FACTORIES, MAX_CLIENT_TOOLS_AFTER_FILTER, MEMORY_SAVE_TOOL_NAME, MIN_CONTENT_LENGTH_FOR_TOOLS, type ParsedServerToolsResponse, type PostFilterKey, type ResolvePlanContext, SERVER_TOOL_DEPENDENCY_SETS, type SelectServerSideToolsOptions, type SelectServerToolsForPromptOptions, type SelectionLogger, type ServerTool, type ServerToolCatalog, type ServerToolCatalogEntry, type ServerToolsFilter, type ServerToolsFilterContext, type ServerToolsFilterFunction, type ServerToolsOptions, type ServerToolsResponse, TOOL_CATALOG, type ThinkingModeHint, type ToolActivationEvent, type ToolChoice, type ToolConfig, type ToolExecutor, type ToolIntentDescriptor, type ToolLane, type ToolMatchOptions, type ToolMatchResult, type ToolPlanSpec, type ToolSelectionAdapters, type ToolSet, type ToolsCacheBackend, activatedToolSetNames, applyToolSets, assembleClientTools, autoFilterClientTools, buildConnectorGuidance, buildDeniedToolsRider, clearServerToolsCache, composeCouncilClientTools, computeToolGuidance, createServerToolsFilter, defaultServerToolsFilter, deferFormattingConfig, deriveActiveToolSets, expandToolSetsAdditive, filterAssembledClientTools, filterServerTools, findMatchingTools, getCachedServerTools, getCatalogEntry, getMaxToolRounds, getServerTools, getThinkingMode, getToolDescription, getToolName, getToolsChecksum, isCoercingIntent, localStorageToolsCache, mergeActiveToolSets, mergeTools, resolveCouncilPlan, resolveCouncilServerTools, resolveDeferredServerTools, resolvePlan, resolveServerTools, resolveToolChoice, scoreTools, selectCouncilClientTools, selectServerSideTools, selectServerToolsForPrompt, shouldRefreshTools, stripCoercedMemoryWrite, toolSetSystemPrompts };